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
1204.4621	1	1204	2012-04-20T13:19:03	On Almost-Invariant Subspaces and Approximate Commutation	[ "math.FA", "math.OA" ]	A closed subspace of a Banach space $\cX$ is almost-invariant for a collection $\cS$ of bounded linear operators on $\cX$ if for each $T \in \cS$ there exists a finite-dimensional subspace $\cF_T$ of $\cX$ such that $T \cY \subseteq \cY + \cF_T$. In this paper, we study the existence of almost-invariant subspaces of infinite dimension and codimension for various classes of Banach and Hilbert space operators. We also examine the structure of operators which admit a maximal commuting family of almost-invariant subspaces. In particular, we prove that if $T$ is an operator on a separable Hilbert space and if $TP-PT$ has finite rank for all projections $P$ in a given maximal abelian self-adjoint algebra $\fM$ then $T=M+F$ where $M\in\fM$ and $F$ is of finite rank.	math.FA	math	ON ALMOST-INVARIANT SUBSPACES AND APPROXIMATE COMMUTATION LAURENT W. MARCOUX1, ALEXEY I. POPOV1, AND HEYDAR RADJAVI1 Abstract. A closed subspace of a Banach space X is almost-invariant for a collection S of bounded linear operators on X if for each T ∈ S there exists a finite-dimensional subspace FT of X such that TY ⊆ Y + FT. In this paper, we study the existence of almost- invariant subspaces of infinite dimension and codimension for various classes of Banach and Hilbert space operators. We also examine the structure of operators which admit a maximal commuting family of almost-invariant subspaces. In particular, we prove that if T is an operator on a separable Hilbert space and if TP − PT has finite rank for all projections P in a given maximal abelian self-adjoint algebra M then T = M + F where M ∈ M and F is of finite rank. 1. Introduction One of the best-known problems in Operator Theory concerns the search for non- trivial, closed, invariant subspaces for an operator (or class of operators) acting on an infinite-dimensional, separable Banach space X . The existence or non-existence of such spaces has been proven for large classes of operators. Examples of operators without invariant subspaces have been found by Enflo [8, 9] and Read [27]. Moreover, Read constructed a number of operators without invariant subspaces. Among them are an op- erator acting on ℓ1 [28], a quasinilpotent operator [30], a strictly singular operator [31], or an operator without invariant closed sets [29]. The most recent construction is published by Sirotkin [32]. The question of the existence of invariant subspaces for an arbitrary bounded linear operator acting on a reflexive Banach space, and in particular on a Hilbert space, remains open. This is the Invariant Subspace Problem. We refer the reader to [1, Section 10] for a brief review of this topic. Another source is the monograph [25]. The problem which we shall examine in this paper is very closely related, but not equivalent, to this problem. For the purposes of this paper, all subspaces of a Banach space are assumed to be closed in the norm topology. Given a Banach space X and a bounded linear operator T on X , we shall say that a subspace Y of X is an almost- invariant subspace for T if there exists a finite-dimensional subspace M of X so that TY ⊆ Y + M. While the finite-dimensional space M appearing in this expression is not unique, nevertheless, if the subspace Y is almost-invariant for T, then the minimum dimension of a finite-dimensional subspace M for which TY ⊆ Y + M is well-defined, and is referred to as the defect of the subspace Y for T. Clearly, if Y is finite-dimensional or finite-codimensional then it is almost-invariant under every operator. This motivates 1 Research supported in part by NSERC (Canada). 2010 AMS Subject Classification: 47A15, 47A46, 47B07, 47L10 . November 7, 2018. 1 2 the notion of a half-space, that is, a subspace of X such that dim Y = dim (X /Y ) = ∞. One may then ask whether every operator T on a Banach space X has an almost-invariant half-space. These notions appeared in the papers [2] and [23]. In [2], it was shown that almost- invariant half-spaces exist for a class of operators which includes quasinilpotent weighted shift operators acting on Banach spaces with bases. Among other results, it was shown in [2] that a closed subspace Y of X is almost-invariant for T if and only if there exists a finite-rank perturbation F of T so that Y is invariant for the operator T + F. Also, if T has an almost-invariant half-space then so does T∗. The main thrust of [23] is the analysis of common almost-invariant half-spaces for algebras of operators. It was shown that if an algebra A of operators on a Banach space X is norm-closed then the dimensions of the defects corresponding to each operator in A are uniformly bounded. Another result of [23] is that if a norm-closed algebra A is finitely-generated and commutative then the existence of a common almost-invariant half-space for A implies the existence of an invariant half-space for A. Also, it was shown that for a norm-closed algebra A, the almost-invariant half-spaces of A and those of the closure of A in the weak operator topology, A , are the same. This is no longer true if A is not closed in the norm topology. WOT The main result of [2] is motivated by the study of almost-invariant half-spaces of Donoghue operators. Recall that a (backward) weighted shift D on ℓ2 is called a Donoghue operator if its weights (wi) are all non-zero and wi ↓ 0. See, e.g., [25, Section 4.4] for a discussion of Donoghue operators. The following property of Donoghue operators is very interesting from our point of view (see [6], [22] and [34]): 1.1. Theorem. The invariant subspaces of a Donoghue operator are all of the form span {ek : k = 1, . . . , n}, where (en) stands for the usual orthonormal basis of ℓ2. In particular, Donoghue operators do not have invariant half-spaces. The following result from [2] implies that Donoghue operators have almost-invariant half-spaces. 1.2. Theorem. [2, Theorem 3.2] Let T be an operator on a Banach space X . Suppose that the following three conditions are satisfied. (i) T has no eigenvalues; (ii) the unbounded component of the resolvent set ρ(T) contains {z ∈ C : 0 < z < ε} for some ε > 0; (iii) there exists e ∈ X such that for each k ∈ N, the vector Tke does not belong to the closed span of the set of vectors {Tie : i 6= k}. Then T has an almost-invariant half-space, Y. 1.3. Remark. Condition (i) in Theorem 1.2 is, in fact, redundant. To justify this claim, we need to outline the proof of this theorem. The half-space Y ⊆ X almost-invariant for T is constructed as a closed span of the form span {h(λ, e) : λ ∈ Λ} where Λ is a certain infinite subset of ρ(T) and h(λ, e) = (λI − T)−1(e), λ ∈ Λ. (We warn the reader that our definition of Λ and h(λ, e) differs slightly from the definition presented in [2].) Condition (i) is used in the following two steps: 3 (1) the set {h(λ, e) : λ ∈ Λ} is linearly independent, so Y is of infinite dimension; (2) a sequence of functionals ( fk)∞ k=1 annihilating Y is defined by assigning values to fk(Tie) (i > 0) which requires the sequence (Tie)i>0 to be linearly independent. We claim that both (1) and (2) follow from the condition (i') p(T)e 6= 0 for any non-zero polynomial p, which, clearly, follows from condition (iii) of Theorem 1.2. Indeed, it is evident that the linear independence of (Tie)i>0 follows from (i'). Let us show that (i') implies (1). Assuming (i'), we will show that the set {h(λ, e) : λ ∈ Λ} is linearly independent for any choice of Λ. Suppose that for some non-zero scalars a1, . . . , an and distinct λ1, . . . , λn in Λ, we have a1h(λ1, e) + a2h(λ2, e) + · · · + anh(λn, e) = 0 That is, n ∑ i=1 ai(λi I − T)−1(e) = 0. Multiplying both sides of this equation by (λ1 I − T)(λ2 I − T) · · · (λn I − T), we obtain: p(T)(e) = 0 for some (non-zero, by elementary algebra) polynomial p. This contra- dicts (i'). There are other closely related questions which have already been examined. For exam- ple, in the Hilbert space setting, Brown and Pearcy [4] showed that if T is a bounded linear operator on an infinite-dimensional Hilbert space H, then there exists an orthogonal pro- jection P of H onto a half-space L of H so that K = (I − P)TP is compact, i.e. so that L is invariant for the operator T − K. In fact, the work of Fillmore, Stampfli and Williams [11] shows that one can do this by considering points in the left-essential spectrum of T. A direct consequence of Voiculescu's non-commutative Weyl-von Neumann Theorem [33] is that given any bounded linear operator T acting on an infinite-dimensional Hilbert space H, there exists a compact operator K and a closed half-space L of H which is reducing for T − K; that is, L is invariant for both T − K and (T − K)∗. In Section 2 of this paper, we obtain almost-invariant subspaces for some classes of operators. First, we show that every quasinilpotent, triangularizable injective operator has an almost-invariant half-space (see Section 2 for the definition of a triangularizable operator). Next, we use this result to show amongst other things that every polynomially compact operator on a reflexive Banach space has an almost-invariant half-space. Observe that the class of Donoghue operators which served as motivation for Theorem 1.2 consists of compact operators. Finally, we study a special class of triangularizable operators: bitriangular operators. We show that every bitriangular operator is either of form λI + F, for some scalar λ and a finite-rank operator F, or it has a hyperinvariant half-space. In particular, every bitriangular operator has an invariant half-space. Section 3 is concerned with the study of common almost-invariant half-spaces for al- gebras of operators. We extend a result from [23] by showing that if a norm-closed algebra A of operators on a Banach space X has an almost-invariant half-space which is complemented in X , then A has a common invariant half-space. In particular, if X is a Hilbert space then the existence of an almost-invariant half-space for a norm-closed 4 algebra always implies the existence of an invariant half-space for that algebra. That is, if P is a projection onto a half-space in H such that TP − PTP is of finite rank for all T in the algebra then the algebra has an invariant half-space. We would like to bring to the reader's attention that the corresponding statement in which "finite rank" is replaced with "compact" is not true, as the following example shows: 1.4. Example. Let D be the C∗-algebra of all diagonal (with respect to a fixed orthonor- mal basis) operators in B(H). Let A = {D + K : D ∈ D, K ∈ K(H)} = D + K(H). It is well known (see, e.g., [21, Theorem 3.1.7]) that A is a norm closed subalgebra of B(H). Clearly, TP − PTP ∈ K(H) for all T ∈ A and every projection P ∈ D. On the other hand, A contains all compact operators in B(H), hense it is dense in B(H) in the weak operator topology. In particular, A has no invariant subspaces. In 1972, Johnson and Parrot [17] showed that if W is an abelian von Neumann algebra acting on a Hilbert space H, and if T is a bounded operator on H for which TW − WT is compact for all W ∈ W , then there exists an operator V in the commutant W ′ := {X ∈ B(H) : XW = WX for all W ∈ W } of W and K compact so that T = V + K. In particular, if W is a maximal abelian self-adjoint subalgebra (i.e. a masa) in B(H), then V ∈ W . The condition that TW − WT be compact for all W ∈ W is readily seen to be equivalent to the condition that TP − PT be compact for all projections in W , since the linear span of the projections is norm-dense in W by the Spectral Theorem for normal operators. In Section 4 of the present work we replace this condition with the related condition that TP − PT be finite-rank, or equivalently, that the range of P be an almost-invariant subspace for T. More specifically, we study operators with the following property: there exists a masa D in B(H) such that for every projection P ∈ D, the space PH is almost- invariant under the given operator. We show that every such operator can be written in the form D + F where D is in the masa and F is of finite rank. Unlike in the setting of the Johnson and Parrot result, the fact that the operator T has finite-rank commutators TM − MT with every element M in the masa (and not just with the projections P in the masa) is part of the conclusion of our result, not of the hypothesis. In the case of a continuous masa, this decomposition is unique. We also study norm-closed algebras of operators leaving all spaces of the form PH (for P in a masa D) almost-invariant. We show that every such algebra is contained in D + Fk(H) for some k ∈ N, where Fk(H) denotes the set of operator on H of rank at most k. We use these results to show that if an operator on H leaves every half-space of H almost-invariant then it can be written as λI + F for some scalar λ and finite-rank operator F. In the last section of the paper, we introduce the notion of an almost-reducing subspace. Recall that a space Y ⊆ H is reducing for a collection S of operators in B(H) if Y and Y ⊥ are both invariant for each T ∈ S. We say that Y is almost-reducing for S if Y and Y ⊥ are each almost-invariant for S. We show that an analogue of our result from Section 3 for almost-reducing subspaces does not hold. There exist norm-closed algebras with many common almost-reducing subspaces that do not have common reducing subspaces. In fact, one can find a singly generated norm-closed algebra with plenty of almost-reducing half-spaces whose generator does not have reducing subspaces. 5 Throughout this work, X , Y and Z will denote complex Banach spaces, while H, K and L will be reserved for (complex) Hilbert spaces. We shall consistently use the following notation: for X a Banach space, B(X ), K(X ), F (X ) and Fk(X ) are the algebras, respectively, of all bounded operators, the compact operators, the finite-rank operators, and those of finite-rank less than or equal to k for a given k ∈ N. For an operator T, σ(T), σp(T) and rank T are, respectively, the spectrum of T, the point spectrum of T (i.e. the set of eigenvalues of T), and the rank of T. For a subset S of X , S is the closure of S in the norm topology, and if S = {xn}∞ n=1 is a countable subset of S, then [xn]n is the closed linear span of S. If {Yi : i ∈ I} is an arbitrary collection of subspaces of X , then Wi∈I Yi is their closed linear span. We call an operator P ∈ B(H) a projection if P = P2 = P∗. An operator (on a Banach space) that only satisfies E = E2 will be referred to as idempotent. Finally, if Y ⊆ X is a subspace, then codimX Y := dim X /Y. We shall also use the notation codim Y if the space X is understood. The following (obvious) lemma will be used many times throughout the paper without further reference. 1.5. Lemma. Let X be a Banach space, Y a complemented subspace of X and EY : X → X be an idempotent with range Y. Then Y is almost-invariant under operator T on X if and only if the operator (I − EY )TEY is of finite rank. Moreover, the defect of Y for T is equal to rank(cid:0)(I − EY )TEY(cid:1). 2. Existence of almost-invariant half-spaces for single operators In this section we will establish the existence of almost-invariant half-spaces for several classes of operators. The following lemma proves the simple fact that half-spaces exist in all Banach spaces. 2.1. Lemma. Every infinite-dimensional Banach space contains a half-space. Proof. Let (xn) be a basic sequence in X (see, e.g., [18, Theorem 1.a.5]). Put Y = [x2n]∞ We claim that Y is a half-space. n=1. It is clear that dim Y = ∞. To see that codim Y = ∞, observe first that (x2n) is also i=n anx2n. Since (xn) is n=1 bnx2n−1, where m ∈ N. It follows (cid:3) a basic sequence, so that every member of Y can be written as ∑∞ basic, Y contains no nonzero elements of form ∑m that codim[xn]Y = ∞. Hence, codimX Y = ∞. We remind the reader that an operator T on a Banach space X is called triangularizable if there is a chain C of subspaces of X that is maximal with respect to inclusion and has the property that each member Y of C is T-invariant (see, e.g., [26, Definition 7.1.1]). We point out that the maximality of the chain C implies that C is complete; that is, it is closed under arbitrary intersections, as well as under closed spans of its members. Our first goal is to show that all injective, triangularizable quasinilpotent operators admit an almost-invariant half-space. We will start with two simple lemmas. 6 2.2. Lemma. Let (xn) be a sequence in a Banach space X . Then xn 6∈ [xk]k6=n holds for all n ∈ N if and only if xn 6∈ [xk]k>n holds for all n ∈ N. Proof. The "only if" part of the lemma is trivial. Let us establish the "if" part. Let (xn) be such that xn 6∈ [xk]k>n holds for all n ∈ N. Observe that the space span{xk}k<n + [xk]k>n is dense in [xk]k6=n. Since the sum of a finite-dimensional space and a closed space is again closed, this shows that [xk]k6=n = span{xk}k<n + [xk]k>n. Suppose that xn ∈ [xk]k6=n for some n ∈ N. Then we can write xn = a1x1 + · · · + an−1xn−1 + u where u ∈ [xk]k>n. Put an = −1 and let i be the smallest number such u. This shows that xi ∈ [xk]k>i, a that ai contradiction. (cid:3) 6= 0. Then xi = − ai+1 ai xi+1 − · · · − an ai xn − 1 ai 2.3. Lemma. Let T be a quasinilpotent operator on a Banach space X . If Z ⊆ Y are T-invariant subspaces of X and dim(Y /Z) = 1 then TY ⊆ Z. Proof. Write Y as a direct sum Y = Z ⊕ [y] for some y ∈ Y \ Z with kyk = 1. Let E : Y → Y be the idempotent with range [y] and null space Z. If TY 6⊆ Z then there exists a non-zero α ∈ C such that Ty = αy + z for some z ∈ Z. Then, since TZ ⊆ Z, we obtain ETny = αny, for all n ∈ N. However, this implies that kTnk > 1 kEk kETnk > 1 kEk αn, so that the spectral radius r(T) of T satisfies the inequality r(T) > α > 0, a contradiction. (cid:3) 2.4. Theorem. Every triangularizable quasinilpotent injective operator T on a Banach space X has an almost-invariant half-space. Proof. Let C be a triangularizing chain of invariant subspaces for T. Clearly, we may assume without loss of generality that no member of C is a half-space. Also, it follows from the injectivity of T that C has no elements having finite dimension, as the restriction of T to such a subspace would be nilpotent, thereby contradicting the injectivity of T. Therefore, every member of C is of finite codimension in X . Following [26, Definition 7.1.6], for each Y ∈ C, we define Y− = W{Z ∈ C : Z ⊆ Y, Z 6= Y }. It follows from maximality of C that for each Y ∈ C, either Y− = Y or dim(Y /Y−) = 1 (see [26, Theorem 7.1.9(iii)]). However, the condition Y− = Y implies that C contains members having infinite codimension in X . So, dim(Y /Y−) = 1 for all Y ∈ C. Let Y1 be an arbitrary nonzero element of C. Pick a non-zero vector e1 ∈ Y1 \ Y1−. Denote e2 = Te1. By Lemma 2.3, e2 ∈ Y1−. Also, since ker T = {0}, e2 6= 0. and Y2 ( Y1. Also, Y2− 6= Y2. In particular, e2 ∈ Y2 \ Y2−. Let Y2 = T{Y ∈ C : e2 ∈ Y }. Since C is a complete chain, Y2 ∈ C. Clearly, Y2 6= {0} Denote e3 = Te2. Again, e3 6= 0 and e3 ∈ Y2−. Put Y3 = T{Y ∈ C : e3 ∈ Y }. Then Y3 ∈ C is such that Y3 6= {0}, Y3 ( Y2, and e3 ∈ Y3 \ Y3−. Continuing inductively, we obtain a strictly decreasing chain of subspaces Y1 ) Y2 ) Y3 ) Y4 ) . . . in C such that if ek = Tk−1e1 then ek ∈ Yk \ Yk+1. The sequence (ek), in particular, satisfies the condition that ei 6∈ [ej]j>i for all i ∈ N. By Lemma 2.2, ei 6∈ [ej]j6=i for all i ∈ N. It follows from Theorem 1.2 that T has an almost-invariant half-space. (cid:3) 7 If the underlying Banach space in Theorem 2.4 is reflexive then we can drop the injec- tivity assumption from the hypothesis of the theorem. 2.5. Theorem. Every triangularizable quasinilpotent operator on a reflexive Banach space has an almost-invariant half-space. Proof. Let T ∈ B(X ) be a triangularizable quasinilpotent operator. Clearly, we may as- sume that the nullity of T is finite, for otherwise, any half-space of ker T is an invariant half-space for T. First, we will show that, without loss of generality, we may assume that T has dense range. Denote Yn = TnX , Y0 = X . Since each Yn is T-invariant, we may assume that each Yn is of finite codimension in X . Write X = Y1 ⊕ Z for some finite-dimensional space Z. We have: Y2 = T2X = TY1, so Y1 = TX = TY1 + TZ = TY1 + TZ = Y2 + TZ, since TZ is finite-dimensional. Similarly, we get Yn = Yn+1 + TnZ for all n ∈ N. Now write the space TnZ as Wn1 ⊕ Wn2 where Wn1 ⊆ Yn+1 and Wn2 ∩ Yn+1 = {0}. Then Tn+1Z = T(Wn1 ⊕ Wn2) = TWn1 + TWn2, with TWn1 ⊆ Yn+2. It follows that codimYnYn+1 is a decreasing sequence. Hence, it must stabilize at some n. Restricting T to Yn, we may assume without loss of generality that this happens for n = 1. Thus Yn = Yn+1 ⊕ TnZ, dim TnZ = dim Z for all n ∈ N. If Z = 0, then T has dense range. Otherwise, pick a non-zero z ∈ Z. Then Tnz ∈ TnZ ⊆ Yn \ Yn+1. It follows that the sequence (Tnz)n>1 has the property that Tnz 6∈ [Tiz]i>n. Since T is quasinilpotent, we get by Lemma 2.2, Theorem 1.2 and Remark 1.3 that T has an almost-invariant half-space. So, we may assume that T has dense range. Then, clearly, T∗ is injective and quasinilpo- tent. Also, if (Mα) is a triangularizing chain for T then (M⊥ α ) is a triangularizing chain for T∗. Thus, T∗ is triangularizable. By Theorem 2.4, T∗ has an almost-invariant half- space. Hence, by [2, Proposition 1.7], so does T = T∗∗. (cid:3) 2.6. Remark. The affirmative answer to the Invariant Subspace Problem for quasinilpo- tent operators on Hilbert spaces would imply that every quasiniplotent operator in B(H) is triangularizable. Hence, Theorem 2.5 shows that, when we restrict our attention to the quasinilpotent operators on Hilbert spaces, the problem of the existence of almost- invariant half-spaces is a weakening of the Invariant Subspace Problem. The same remark can be made about the reflexive spaces. Our next goal is finding almost-invariant half-spaces for polynomially compact oper- ators. Recall that an operator T is called polynomially compact if there is a non-zero polynomial p such that p(T) is compact. 8 2.7. Lemma. Let T be an operator on an infinite-dimensional Banach space X . If σ(T) has infinitely many connected components then T has an invariant half-space. Proof. The set σ(T) must contain infinitely many connected components that are relatively open sets in σ(T). Denote all such components by {σn}∞ n=1. For each n ∈ N, there is a Riesz projection for σn, that is, an idempotent En such that EnT = TEn, σ(TEnEnX ) = σn, and σ(T(I − En)(I−En)X ) = σ(T) \ σn. Observe that, since σ(T) \ σn is infinite for each n, (I − En)X is infinite-dimensional for each n. As such, if EnX is infinite-dimensional for some n, then EnX is a T-invariant half-space. Thus we may assume that EnX is finite-dimensional for each n ∈ N. It follows that each σn is a singleton, σn = {λn}, and λn is an eigenvalue. For each λn, pick a non-zero vector xn such that Txn = λnxn. Define Y = [x2n]∞ n=1. We claim that Y is a T-invariant half-space. The T-invariance is obvious. Also, Y is clearly infinite-dimensional. We need to prove that codim Y = ∞. Let m ∈ N be arbitrary. Let F be a Riesz projection corresponding to σ1 ∪ σ3 ∪ · · · ∪ σ2m−1. We claim that Y ⊆ (I − F)X . Clearly, it is enough to prove that each x2n belongs to (I − F)X . Denote X1 = FX , X2 = (I − F)X , T1 = TX1, T2 = TX2, so that X = X1 ⊕ X2 and T = T1 ⊕ T2. Write x2n as (x(1) 2n , x(2) 2n ), relative to this decomposition of X . Then Tx2n = (cid:0)T1x(1) 2n , x(2) 2n(cid:1). Since λ2n is not in the spectrum of T1, x(1) 2n , T2x(2) 2n = 0. It follows that x2n ∈ X2. Since m ∈ N was arbitrary, this shows that Y is a half-space. λ2n(cid:0)x(1) 2n(cid:1) = (cid:3) 2.8. Remark. The half-space constructed in Lemma 2.7 is spanned by the eigenvectors corresponding to even-numbered isolated eigenvalues of T. It should be noted that if we drop the condition that the eigenvalues are isolated, the construction may not work. For example, let T be the backward shift on ℓ2. Every element of the open unit disk is an eigenvalue for T. For each λ in the open unit disk, let xλ = (λ, λ2, λ3, . . . ), so that Txλ = λxλ. Then for any subset Λ of the open unit disk having an accumulation point also in the open unit disk, the span of {xλ : λ ∈ Λ} is dense in ℓ2. Indeed, let Y = W{xλ : λ ∈ Λ}. Let y = (y1, y2, y3, . . . ) ∈ Y ⊥ be arbitrary. Consider the function f (z) = ∑∞ f (λ) = ∑∞ point in the interior of the unit disk. Therefore, f and, hence, y must be equal to zero. i=1 yizi, which is analytic in the unit disk. For each λ ∈ Λ, we have i=1 yiλi = hxλ, yi = 0. Thus, the set of zeros of f is infinite, with an accumulation 2.9. Remark. It follows from Lemma 2.7 that, when looking to prove the existence of almost-invariant half-spaces for an operator T, one may assume that the spectrum of T Indeed, assume σ(T) is not connected. By Lemma 2.7, we may assume is connected. that σ(T) = σ1 ∪ σ2 ∪ · · · ∪ σn where each σi is connected. For each i = 1, . . . , n, consider a Riesz projection Ei corresponding to σi. That is, EiT = TEi, σ(TPiEiX ) = σi, and E1 + E2 + · · · + En = I. Denote by Xi the subspace EiX . It is clear that we may assume without loss of generality that only one of Xi's, say, X1, is infinite-dimensional. Consider the operator S = E1TE1 ∈ B(X1). If S has an almost-invariant half-space then so does T. Conversely, if Y ⊆ X is a T-almost-invariant half-space then Y ∩ X1 is an S-almost- invariant half-space. 9 2.10. Remark. Note that the operator S in Remark 2.9 is triangularizable if T is triangu- larizable. 2.11. Theorem. Every triangularizable operator T with countable spectrum acting on a reflexive Banach space X has an almost-invariant half-space. Proof. By Remarks 2.9 and 2.10, we may assume that σ(T) is a singleton. Subtracting a scalar multiple of identity, we may assume that T is quasinilpotent. The statement of the theorem now follows from Theorem 2.5. (cid:3) 2.12. Corollary. Every polynomially compact operator on a reflexive Banach space X admits an almost-invariant half-space. Proof. Let T be the polynomially compact operator. If p is a polynomial such that p(T) ∈ K(X ) and π : B(X ) → B(X )/K(X ) is the canonical quotient map, then p(π(T)) = 0, and so the essential spectrum σ(π(T)) of T is finite. From this and the Putnam-Schechter theorem (see [24] or [5, Theorem 6.8]), it follows that the spectrum of T consists of count- ably many points, and the finitely many elements of σ(π(T)) are the only possible accu- mulation points of σ(T). Since every polynomially compact operator is triangularizable (see [3, 12]), it follows from Theorem 2.11 that T has an almost-invariant half-space. (cid:3) 2.13. Remark. In view of Theorem 2.12 and the well-known theorem of Lomonosov (see [19]; see also, e.g., [1, Theorem 10.19]) that all operators commuting with a non-zero compact operator have hyperinvariant subspaces, one may ask if compact operators al- ways admit almost hyperinvariant half-spaces, that is, half-spaces which are almost-invariant under all operators commuting with a given compact operator. This question has a nega- tive answer. Indeed, let D be a Donoghue operator (see the Introduction for the definition and properties of Donoghue operators). Observe that D is compact. If D had an almost hyperinvariant half-space then this half-space would be almost-invariant for the norm- closed algebra A generated by D. However, it follows from [23, Theorem 3.6] that A has no almost-invariant half-spaces. For the remainder of this section, we shall restrict our attention to separable Hilbert spaces. Given the results of this section, it makes sense to pose the following question: does every triangularizable operator admit an almost-invariant half-space? We will consider a special case of triangularizable operators: bitriangular operators. Recall that an operator T on a separable Hilbert space H triangular if it has an upper triangular matrix with respect to some orthonormal basis indexed by the natural numbers. That is, there exists an orthonormal basis (en)∞ n=1 for H such that hTej, eii = 0 whenever i > j. We refer the reader to [14, Chapter 3], [7] and [15] for more information about the triangular operators. The operator T is called bitriangular if both T and T∗ are triangular, perhaps with respect to different orthonormal bases. We will use some definitions and notations from [7]. For an operator A ∈ B(H) and an integer k > 1, denote by ker(A, k) the space ker Ak ⊖ ker Ak−1. By nul(A, k) we will denote the dimension of ker(A, k), and finally, the symbol α(A, k) will stand for the difference nul(A, k) − nul(A, k + 1). Here the difference ∞ − ∞ is considered to be ∞. 10 Observe that if T is an operator acting on a finite-dimensional space, the number nul(T − λ, k) counts the number of Jordan blocks corresponding to λ of size at least k, while the number α(T − λ, k) counts the number of blocks for λ of size exactly k. This motivates the following definition from [7]: if T ∈ B(H), the canonical Jordan model for T is defined as J(T) = Mλ∈σp(T) J(T, λ) = Mλ∈σp(T)Mk>1 (λIk + Jk)(α(T−λ,k)). Here Ik is the identity k × k-matrix and Jk is the nilpotent k × k-matrix 0 1 0 . . . 0 0 1 . . . . . . 0 0 0 . . . 0 0 0 . . . 0 0 1 0 ,   Jk =   and the symbol A(j) means the direct sum of j copies of the operator A. The direct sum is taken in the Hilbert space sense: the underlying space is a Hilbert space direct sum of the corresponding summands (in particular, each summand is orthogonal to any other summand). Recall that an operator A ∈ B(H) is called a quasiaffinity if A is injective and has dense range. Two operators T and S in B(H) are called quasisimilar if there exist two quasiaffinities A and B such that AT = SA and TB = BS. It is known (see [16]) that quasisimilarity preserves the existence of hyperinvariant subspaces. However, it may not preserve the structure of the lattice of hyperinvariant subspaces (see [13]). It is not known whether or not quasisimilarity preserves the existence of invariant subspaces. The following statement about bitriangular operators is the main result of [7]. 2.14. Theorem. [7, Theorem 4.6] Every bitriangular operator is quasisimilar to its canonical Jordan model. We are now ready to state our result about bitrangular operators. 2.15. Theorem. If T is a bitriangular operator then either T can be written as λI + F where F is a finite rank operator or T has a hyperinvariant half-space. In particular, T has an invariant half-space. Proof. Let T be a bitriangular operator that cannot be written in the form λI + F where F is a finite rank operator. Following [7, Lemma 4.5], for a subset Γ ⊆ C, set H(T, Γ) = _{ker(T − λ)k : λ ∈ Γ, k ∈ N}. Since T is bitriangular, the point spectrum σp(T) is non-empty and countable (see [7, Theorem 3.1]). Write σp(T) as a (perhaps finite) sequence of distinct complex numbers λ1, λ2, λ3 . . . . For each n, define (λn Ik + Jk)(α(T−λn,k)). Tn = Mk>1 11 The canonical Jordan model of T is J(T) = Ln Tn. By Theorem 2.14, T is quasisimilar to J(T). Fix two quasiaffinities A and B such that AT = J(T)A and TB = BJ(T). Notice that the point spectrum of each summand in J(T) is a singleton: σp(Tn) = {λn}. Assume first that σp(T) is infinite. Consider Γ = {λ2n}∞ n=1. By [7, Lemma 4.5], the spaces H(T, Γ) and H(T, C \ Γ) are T-hyperinvariant, H = H(T, Γ) ∨ H(T, C \ Γ) and H(T, Γ) ∩ H(T, C \ Γ) = {0}. Also, AH(T, Γ) = H(J(T), Γ). It is clear that H(J(T), Γ) is infinite-dimensional. Therefore, so is H(T, Γ). Analogously, the space H(T, C \ Γ) is infinite-dimensional, too. It follows that H(T, Γ) is a T-hyperinvariant half-space. Assume now that σp(T) is finite: σp(T) = {λ1, . . . , λn}. Abbreviate H(T, {λi}) as H(T, λi). Then H = H(T, λ1) ∨ · · · ∨ H(T, λn), where each H(T, λi) is T-hyperinvariant, and each pair of the summands in this decomposition has trivial intersection. If there are a T-hyperinvariant half-space. two indices i 6= j such that dim(cid:0)H(T, λi)(cid:1) = dim(cid:0)H(T, λi)(cid:1) = ∞, then either of them is Observe that if all of H(T, λi) is finite-dimensional then J(T) is finite rank. Hence, by injectivity of A, we can conclude that T is finite rank, too. Therefore, we may assume without loss of generality that exactly one of the H(T, λi), i > 1 -- say H(T, λ1) -- is infinite- It follows from [7, Lemma 2.3] and [15, Proposition 2.9] that TH(T,λ1) is dimensional. a bitriangular operator. Therefore, we may assume that the point spectrum of T is a singleton. Denote this unique element of σp(T) by λ. Since scalar perturbations do not change hyperinvariant subspaces, we may assume that λ = 0. Also, since J(T) is not finite In particular, J(T) acts on an rank, it must have infinitely many blocks of size > 2. infinite-dimensional space and has infinite-dimensional kernel. By injectivity of B, ker T is infinite-dimensional, too. By [7, Proposition 3.2], H = Wk>1 ker Tk. It follows from the definition of J(T) that if ker T were finite codimensional, J(T) would have only finitely many blocks of size > 2, a contradiction. Since we assumed that T is not finite-rank at the outset, ker T is a half-space, which is easily seen to be hyperinvariant for T. (cid:3) 3. Operator algebras with common almost-invariant half-spaces 3.1. Our goal in this section is to show that if A is a norm-closed algebra of operators on a Banach space X , and if Y is an almost-invariant subspace for A, then A admits an invariant subspace. We begin with a Lemma which will also be useful in the next section. 3.2. Lemma. Let X and Y be Banach spaces. Suppose that T = (cid:20)A X C B(cid:21) ∈ B(X ⊕ Y ). If κ ∈ N and rank T = rank C = κ < ∞, then X is uniquely determined by A, C and B. If, in addition, C is invertible (which can only happen when dim(X ⊕ Y ) < ∞), then X = AC−1B. Proof. Since C : X → Y has rank κ < ∞, W1 := ran C is topologically complemented in Y. Thus we may decompose Y = W1 ⊕ W2 for some closed subspace W2 of Y. Similarly, V2 := ker C is finite-codimensional in X , and so we may decompose X = V1 ⊕ V2 for some closed subspace V1 of X . With respect to these decompositions of X and Y, we may 12 then write C = where C0 is invertible. From rank considerations, it then follows that T must be of the form: V1 V2 0 W1 C0 W2 0 (cid:18) 0 (cid:19), T =   ∗ ∗ ∗ 0 ∗ ∗ ∗ 0 − − − − − − ∗ ∗ C0 0 0 0 0 0   relative to the decomposition X ⊕ Y = V1 ⊕ V2 ⊕ W1 ⊕ W2. Let E0 : W1 → V1 denote the inverse of C0, so that E0C0 = IV1 and C0E0 = IW1. We can extend E0 to a continuous linear map E = (cid:20)E0 0 0(cid:21) , 0 with the block-matrix decomposition corresponding to that of an operator from Y = W1 ⊕ W2 to X = V1 ⊕ V2. Then EC ∈ B(X ) is simply the projection of X onto V1, and the operator matrix for T above shows that AEC = A. From this we conclude that (cid:20)IX −AE Since the operator(cid:20)IX −AE 0 0 B C (cid:21) . C B(cid:21) = (cid:20) 0 X − AEB IY (cid:21)(cid:20)A X IY (cid:21) is clearly invertible, it preserves rank. But rank C = κ = rank (cid:20) 0 X − AEB (cid:21) C B then implies that X − AEB = 0. Since E is clearly uniquely determined by C, the state- ment of the Lemma is proven. ✷ The following theorem is the main result of this section. 3.3. Theorem. Let X be a Banach space and A ⊆ B(X ) be a norm-closed algebra of operators. Suppose that Y is a closed, topologically complemented half-space of X and that for each T ∈ A, there exists a finite-dimensional subspace MT of X so that TY ⊆ (Y + MT). Then A admits an invariant half-space L. Moreover, there exists a finite-dimensional subspace F of X so that L is a finite-codimensional subspace of Y + F . Remark. Informally, one could think of L as lying "within a finite-dimensional subspace" of Y. Proof. Clearly we may assume without loss of generality that A is unital. Let E : X → X be a continuous idempotent map with range equal to Y. The kernel Z of E (equal to the range of I − E) serves as a topological complement to Y. Consider the set S := {(I − E)TE : T ∈ A} as a subspace of B(Y, Z). Observe that by Theorem 2.7 of [23], there exists 1 6 κ < ∞ such that 13 Fix an element T ∈ A such that rank((I − E)TE) = κ. max T∈A rank(cid:0)(I − E)TE(cid:1) = κ. Write T = (cid:20)T1 T2 T3 T4(cid:21) relative to the decomposition X = Y ⊕ Z, so that T3 = (I − E) T EEX . Now W1 := ran T3 is a finite-dimensional subspace of Z, and hence is complemented in that space, say Z = W1 ⊕ W2. Similarly, V1 := ker T3 is a finite- codimensional subspace of Y, and thus is complemented there, say Y = V1 ⊕ V2. With respect to these decompositions, we may write T3 = W1 W2 V1 V2 T32 0 0 0 (cid:19). (cid:18) Furthermore, T32 is an injective map from V2 onto W1, and thus is an invertible element of B(V2, W1). We claim that S ∈ S implies that SV1 ⊆ W1. Let L ∈ S, and write L = W1 W2 V1 V2 L1 L2 L3 L4(cid:19). (cid:18) Then, for all α ∈ C, L + αT3 ∈ S, and thus rank(L + αT3) 6 κ. In fact, this has rank equal to rank T32 = κ for all but finitely many values of α, since rank (L2 + αT32) = κ for all but finitely many values of α ∈ C. By Lemma 3.2, for all but finitely many values of α, L3 = L4(L2 + αT32)−1L1. By letting the absolute value of α tend to infinity, we see that k(L2 + αT32)−1k tends to zero, and thus L3 tends to zero. But kL3k is constant, a contradiction unless L3 = 0. This proves the claim. This allows us to write each element of the algebra A in the form V1 V2 W1 W2 ∗ ∗ ∗ ∗ ∗ ∗ ∗ 0 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ .   V1 V2 W1 W2   Now consider the space L := AV1 ⊆ X . Clearly, L is A-invariant. Also, since A is unital, V1 ⊆ L, so that L is infinite-dimensional. On the other hand, it follows from the above matrix form of operators in A and the fact that the space Y ⊕ W1 is closed that L = A(V1) ⊆ V1 ⊕ V2 ⊕ W1 = Y ⊕ W1. Since X = Y ⊕ W1 ⊕ W2 and dim W2 = ∞, the space L is of infinite codimension. Therefore, L is a half-space. The final statement of the theorem is easy to verify. ✷ 14 The next Corollary follows from Theorem 3.3 and the fact that every closed subspace of a Hilbert space is topologically complemented. Observe that the Corollary applies, in particular, to any C∗-subalgebra of B(H), whose invariant half-spaces are automatically reducing. 3.4. Corollary. Let A ⊆ B(H) be a norm-closed algebra of operators on H. Let Y be a closed half-space of H and suppose that for each T ∈ A, there exists a finite-dimensional subspace MT of H so that TY ⊆ Y + MT. Then A admits an invariant half-space. 4. almost-invariant subspaces in masas 4.1. Let us first establish some notation that will be used throughout this section. By H we shall denote a separable Hilbert space, and D ⊆ B(H) will denote a maximal abelian, selfadjoint subalgebra (i.e. a masa) of B(H). If (X, Σ, µ) is a measure space and H = L2(X, µ), then the separability of H implies that µ is σ-finite. Moreover, the set D = {M f : f ∈ L∞(X, µ)}, where M f : L2(X, µ) → L2(X, µ) is the operator M f g = f g, is a masa in B(L2(X, µ)). As is well-known, given any masa D in B(H), there exists a measure space (X, µ) so that H is unitarily equivalent to L2(X, µ) and D is unitarily equivalent to the masa {M f : f ∈ L∞(X, µ)} defined above. For E ⊆ X a measurable set, Ec will denote the complement of E in X, and the projec- tion L2(X, µ) → L2(X, µ) 7→ χE · f , f will be denoted by PE. Also, if f , g ∈ L2(X, µ) then the rank-one operator h 7→ hh, gi f in B(L2(X, µ)) will be denoted by f ⊗ g∗. Let D ⊆ B(H) be a masa. We shall denote by P (D) the set of orthogonal projections in D. Suppose that D ∈ D and that F ∈ F (H) has rank n < ∞. Let T = D + F. It is clear that if P ∈ P (D), then rank (TP − PTP) = rank(I − P)FP 6 n < ∞. Our goal in this section is to provide a converse to this result. More specifically, wiith D ⊆ B(H) a masa as above, suppose that T ∈ B(H) is such that every projection P ∈ P (D) is almost-invariant for T. We will show that this implies that T = D + F for some D ∈ D and F ∈ F (H). We begin with a technical lemma. 4.2. Lemma. Let (X, Σ, µ) be a measure space. Suppose that E ∈ Σ, T ∈ B(L2(X, µ)), and f1, . . . , fn ∈ L2(X, µ) are such that the set {PEc TPE fk}n k=1 is linearly independent. Then there exists ε > 0 such that for all A ⊆ Ec and B ⊆ E in Σ with µ(A), µ(B) < ε, the set {PEc\ATPE\B fk}n k=1 is linearly independent. Proof. Denote the space span{ f1, . . . , fn} by M. Assume that there are sequences (Ak), (Bk) of measurable sets such that µ(Ak), µ(Bk) → 0 and rank(PEc\Ak TPE\BkM) 6 n − 1. Observe that the sequence (PEc\Ak TPE\Bk) converges to PEc TPE in the strong operator SOT−→ PEc TPEM. However, M is finite-dimensional, topology. In particular, PEc\Ak TPE\BkM 15 hence PEc\Ak TPE\BkM that rank(PEc TPEM) 6 n − 1, contrary to the assumptions of the lemma. k·k −→ PEc TPEM. It follows from the lower semicontinuity of the rank (cid:3) 4.3. As pointed out above, a necessary condition for an operator T to be expressible in the form T = D + F with D ∈ D and F a finite-rank operator is the existence of an integer n (which may be chosen to be rank F) so that sup rank (I − P)TP 6 n < ∞. P∈P (D) Our first goal therefore is to show that such an integer always exists. 4.4. Theorem. Let D be a masa in B(H) and T ∈ B(H). Suppose that for every projection P ∈ D onto a half-space YP, the space YP is almost-invariant under T. Then there exists κ ∈ N such that rank(TP − PTP) 6 κ for all projections P ∈ D (including the projections onto finite- dimensional or finite-codimensional subspaces). : Proof. Let us choose a measure space (X, Σ, µ) such that B(H) = L2(X, µ) and D = f ∈ L∞(X, µ)} is the set of multiplication operators on L2(X, µ). We may next {M f decompose L2(X, µ) = L2(Xd, µd) ⊕ L2(Xc, µc) where the measure µd is purely atomic and µc is atomless. Relative to this decomposition of B(H), T can be written in the matrix form (cid:20)T1 T2 F T4(cid:21) , where rank(F) < ∞ (and rank T2 < ∞, though we shall not need this). The above decomposition of L2(X, µ) induces a decomposition of the algebra D as D = Dd + Dc, where Pd (resp. Pc) is the orthogonal projection of L2(X, µ) onto L2(Xd, µd) (resp. L2(Xc, µc)), Dd = PdD and Dc = PcD. If we can show that there are κ1, κ2 ∈ N such that rank(T1P − PT1P) 6 κ1 and rank(T4Q − QT4Q) 6 κ2 for all projections P ∈ L2(X, µd) and Q ∈ L2(X, µc), then κ = κ1 + κ2 + rank(F) will satisfy the conclusion of the theorem. So, we may assume without loss of generality that µ is either purely atomic or atomless. We consider three cases. Case 1: H = L2(X, µ) where µ is an atomless, finite measure. Suppose that {rank(TPE − PETPE) : E measurable} is unbounded. In particular, there TPE1 6= 0. That is, there exists f11 ∈ H such that Observe that, since µ is finite and atomless, the condition of unboundedness of the set {rank(TPE − PETPE) : E measurable} implies that, given N ∈ N and ε > 0 there exist F, G ∈ Σ such that F ∩ G = ∅, µ(F), µ(G) < ε, and rank(PFTPG) > N. Put Q1 = E1. By Lemma 4.2, there exists ε1 > 0 such that PQc 1\ATPQ1\B f11 6= 0 for all measurable sets A and B satisfying µ(A), µ(B) < ε1. By the observation above, we can find disjoint sets E2 and F2 such that µ(E2), µ(F2) < ε1 2 and rank(PF2 TPE2) > 2. That is, there are f21 and f22 in H such that supp f21, supp f22 ⊆ E2 and {PF2 T f21, PF2 T f22} is linearly independent. exists a measurable set E1 such that PEc supp f11 ⊆ E1 and PEc T f11 6= 0. 1 1 16 2 Put Q2 = (Q1 ∪ E2) \ F2. Then E2 ⊆ Q2 and F2 ⊆ Qc 2\ATPQ2\B f2i}2 i=1 is linearly independent. By Lemma 4.2, there exists 0 < ε2 < ε1 2 such that the set {PQc i=1 is linearly independent for all measurable sets A and B satisfying µ(A), µ(B) < ε2. Again, fix disjoint sets E3 and F3 such that µ(E3), µ(F3) < ε2 2 and rank(PF3TPE3) > 3. There exist f31, f32, and f33 such that supp f3i ⊆ E3 for i = 1, 2, 3 and the set {PF3 T f3i}3 i=1 is linearly independent. 2, hence the set {PQc TPQ2 f2i}2 ε n−2 such that the set {PQc n−1\ATPQn−1\B fn−1,i}n−1 Continue this process indefinitely. On the n-th step, use Lemma 4.2 to find 0 < εn−1 < i=1 is linearly independent whenever A and B are such that µ(A), µ(B) < εn−1. Fix disjoint sets En and Fn such that µ(En), µ(Fn) < ε n−1 i=1 such that supp fni ⊆ En for all i = 2 1, 2, . . . , n, and the set {PFn T fni}n i=1 is linearly independent. Finally, define Qn = (Qn−1 ∪ En) \ Fn. and rank(PFn TPEn) > n; then there are { fni}n 2 Observe that, since En ∩ Fn = ∅ for all n, we have Qn = ∪n Put Q = ∪∞ that is not T-almost-invariant. k=1(cid:0)Ek \ (∪∞ i=k+1Fi)(cid:1), n ∈ N. i=k+1Fi)(cid:1). We will prove that PQ is a projection onto a half-space k=1(cid:0)Ek \ (∪n Indeed, let us show that, for each n ∈ N, the set {PQc TPQ fni}n i=1 is linearly inde- k=n+1Ek and ε k−1 2 . The choice of εn guarantees for all 1 6 k 6 m − 1 and m ∈ N. Therefore µ(An), µ(Bn) < εn. It follows i=1 is linearly independent. Since supp fni ⊆ En ⊆ Qn for n \ An ⊆ Qc im- i=1 is linearly independent, too. It follows that the operator In particular, PQ is a half-space, and PQ is not T-almost- pendent. Observe that one can write Q = (Qn ∪ An) \ Bn where An ⊆ ∪∞ k=n+1Fk. In particular, µ(An), µ(Bn) < ∑∞ Bn ⊆ ∪∞ εm 6 ε m−k 2k that the set {PQc all i, we have PQ fni = P(Qn∪An)\Bn fni = PQn\Bn fni. Hence, the inclusion Qc plies that the set {PQc TPQ fni}n PQcTPQ is of infinite rank. invariant. n\An TPQn\Bn fni}n k=n+1 Case 2. H = L2(X, µ) where µ is an atomless σ-finite measure. Assume that the set {rank PEcTPE : E measurable} is unbounded. First, let us show that for each N ∈ N and each set A of finite measure, there exist disjoint subsets E and F of Ac such that µ(E), µ(F) < ∞ and rank(PFTPE) > N. Indeed, represent T as h T1 T2 T3 T4i relative to the decomposition H = PA ⊕ PAc. By the assumptions of the theorem, T2 and T3 are of finite rank. Also, by Case 1, there is κA such that rank(PEcTPE) 6 κA for all measurable E ⊆ A. Pick a measurable set G such that It follows from PGc TPG = PGc\Ac TPG\A + rank(PGc TPG) > N + rankT2 + rankT3 + κA. PGc\Ac TPG\Ac + PGc\ATPG\A + PGc\ATPG\Ac that PGc\ATPG\A has rank at least N. Since both Gc \ A and G \ A can be approximated by sets of finite measure, the existence of E and F with required properties follows. 0 TPE0) = ∞. Pick disjoint sets E1 and F1 of finite measure such that PEc We will repeatedly use the above observation to construct a set E0 such that TPE1 6= 0. rank(PEc Next, pick disjoint sets E2, F2 ⊆ (E1 ∪ F1)c of finite measure such that rank(PF2 TPE2) > 2. Continuing inductively, we construct two sequences (En) and (Fn) of sets of finite mea- sure such that every member of either sequence is disjoint from all other members of both sequences and rank(PFn TPEn) > n for all n ∈ N. Define E0 = ∪n∈NEn. Then 1 17 rank(PEc halfspace that is not T-almost-invariant. 0 TPE0) > rank(PFn TPEn) > n for all n ∈ N, so that PE0 is a projection onto a Case 3. H = L2(X, µ) where µ is a purely atomic measure. This case is proved by the same argument as Case 2. We simply approximate the measurable sets by finite sets instead of sets of finite measure, then repeat the argument almost verbatim. (cid:3) Before proceeding to the proof of the main result of this section, we pause for two more technical lemmas. The thrust of the second of these lemmas is to show that if, in the setting of a continuous measure space, an "off-diagonal corner" of an operator T has rank κ, then we can compress that corner into as small a "subcorner" as we wish (in the sense that both the range and initial space correspond to sets of small measure) and still keep the rank of the compression as great as the rank of the original corner. 4.5. Lemma. Let κ > 1 be an integer. Suppose that Y is a subset of a continuous, σ-finite, Borel measure space (X, µ), and that { f1, f2, ..., fκ} is a linearly independent family of measurable functions from Y to C. Then for all ε > 0 there exists W ⊆ Y with 0 < µ(W) < ε such that { f1W, f2W, ..., fκW} is also a linearly independent set. Proof. The proof will be by induction on κ. The case κ = 1 is clear, since any non-zero function f1 must still be non-zero on a set of arbitrarily small measure. (The measure is continuous!) Assume that the result holds for κ − 1. We shall prove that it holds for κ. Indeed, suppose otherwise. Then there exists ε > 0 such that for all W ⊆ Y with µ(W) < ε, the set { f1W, f2W, ..., fκW} is linearly dependent. Use the induction hypothesis to choose W0 ⊆ Y with µ(W0) < ε 2 such that { f1W0, f2W0, ..., fκ−1W0} is linearly independent. Since { f1W0, f2W0, ..., fκW0} is linearly dependent, there exist unique λ1, λ2, ..., λκ−1 ∈ C so that fκW0 = λ1 f1W0 + λ2 f2W0 + · · · + λκ−1 fκ−1W0. If W1 ⊆ Y and µ(W1) 6 ε 2 , then { f1W0∪W1, f2W0∪W1, ..., fκW0∪W1} is linearly dependent. Thus some linear combination of these functions must be zero. By restricting to the subset W0 of W0 ∪ W1, we see that the only possible linear combination which does this is: λ1 f1W0∪W1 + λ2 f2W0∪W1 + · · · + λκ−1 fκ−1W0∪W1 − fκW0∪W1 = 0. Now, since µ is σ-finite, we can write Y = ∪αWα with µ(Wα) < ε λ1 f1Y + λ2 f2Y + · · · + λκ−1 fκ−1Y − fκY = 0, 2 for all α. But then a contradiction. ✷ 18 4.6. Lemma. Let µ be a continuous, σ-finite Borel measure on a measure space X. Suppose that D = {M f : f ∈ L∞(X, µ)} is a masa acting on H = L2(X, µ), Z0 ⊆ X is a measurable set, and P0 := PZ0 ∈ D is a projection for which rank (P⊥ 0 TP0) = κ < ∞. Then for each m > 1, there exist Ym, Zm ⊆ X measurable subsets satisfying (i) max{µ(Ym), µ(Zm)} < 1 m ; (ii) Ym+1 ⊆ Ym ⊆ Zc 0, Zm+1 ⊆ Zm ⊆ Z0 for all m > 1; for which rank (PYm TPZm) = κ. Proof. Since rank(P⊥ Zc 0 → C and linearly independent functions g1, g2, ..., gκ : Z0 → C such that 0 TP0) = κ, there exist linearly independent functions f1, f2, ..., fκ : κ ∑ j=1 By Lemma 4.5, we can find Y1 ⊆ Zc 0 and Z1 ⊆ Z0 measurable so that max{µ(Y1), µ(Z1)} < 1 and both { f1Y1, f2Y1, ..., fκY1 } and {g1Z1, g2Z1, ..., gκZ1 } are lin- early independent. Then P⊥ 0 TP0 = fj ⊗ g∗ j . Pz1 TPY1 = κ ∑ j=1 fjY1 ⊗ gjZ1 ∗ has rank κ. The result then follows from a routine induction argument. ✷ The following is the main result of this section. 4.7. Theorem. Let H be a separable Hilbert space and let D be a masa in B(H). Suppose that T ∈ B(H) has the property that the range of every projection in D is almost-invariant for T. Then there exists F ∈ F (H) and D ∈ D so that T = D + F. Furthermore, if κ = sup{rank (I − P)TP : P ∈ P (D)} then κ < ∞ and (a) If D is an atomic masa, then F may be chosen so that rank F 6 3κ. (b) If D is a continuous masa, then F is uniquely determined and rank F 6 κ. (c) For D a general masa, F may be chosen so that rank F 6 6κ. Proof. First observe that by Theorem 4.4, κ = sup{rank (I − P)TP : P ∈ P (D)} < ∞ and κ = max{rank (I − P)TP : P ∈ P (D)}. (a) We begin with the case where D is a discrete masa. Let E := {ξα : α ∈ A} denote the orthonormal basis for H corresponding to characteristic functions onto the atoms A of the masa. Choose P0 ∈ D so that κ = rank (I − P0)TP0. It follows that the matrix of (I − P0)TP0 relative to {ξα}α admits κ linearly independent columns. We shall relabel the corresponding basis vectors as {e1, e2, ..., eκ}. In a similar way, we may find κ linearly independent rows of (I − P0)TP0, and we may relabel the basis vectors corresponding to those rows as { f1, f2, ..., fκ}. (Observe that since P0 is obviously perpendicular to (I − P0), it follows that {e1, e2, .., eκ} ∩ { f1, f2, ..., fκ} = ∅.) Let us define He = span{ej}κ j=1, H f = span{ fj}κ j=1, and H1 = H ⊖ (He ⊕ H f ). We may then write the matrix for T relative to H = He ⊕ H1 ⊕ H f as 19 Let Q denote the orthogonal projection of H onto He ⊕ H1. Then T1 T2 T3 C B T6 T0 A T9 T =     . κ > rank (Q⊥TQ) = rank(cid:2)T0 A(cid:3) > rank T0 = κ. It follows that A = T0W for some W ∈ B(H1, H f ). Similarly, C = XT0 for some X ∈ B(He, H1). Let g ∈ E \{e1, e2, ..., eκ, f1, f2, ..., fκ}, and let R denote the orthogonal projection of H onto span {g, f1, ..., fκ}⊥. Then rank(R⊥TR) = κ, since T0 is a compression of R⊥TR. Let H2 = H1 ⊖ Cg. Relative to the decomposition H = He ⊕ H2 ⊕ Cg ⊕ H f , we may write . T1 T2,1 T2,g T3 C1 B2,2 B2,g T6,1 C2 Bg,2 Bg,g T6,g T0 A1 Ag T9 T =   T0 A1(cid:21) = κ = rank(cid:2)T0 A1 Ag(cid:3) .   In particular, this means that the column of Ag must be a linear combination of the Here, Bg,g = hTg, gi. Thus rank (cid:20)C2 Bg,2 columns of T0. We can therefore choose a new entry zg ∈ C so that the column (cid:20) zg Ag(cid:21) is precisely the same linear combination of the columns of (cid:20)C2 T0(cid:21). In fact, by Lemma 3.2, zg = C2T−1 In particular, the diagonal operator D1 = diag(Bg,g − zg) : g ∈ E \{e1, e2, ..., eκ, f1, f2, ..., fκ} is bounded. Set D = 0 ⊕ D1 ⊕ 0 ∈ B(He ⊕ H1 ⊕ H f ), so that D ∈ D. It then follows that with 0 k · kAgk 6 kTk2 · kT−1 0 k. 0 Ag, so that (cid:12)(cid:12)zg(cid:12)(cid:12) 6 kC2k · kT−1 T − D =   T2 T1 T3 C B − D1 T6 T0 T9 A   , every row of (cid:2)C B − D1(cid:3) is a linear combination of the rows of(cid:2)T0 A(cid:3), so that Finally, rank (T − D) 6 rank(cid:20) C B − D1 T0 T0 A (cid:21) = κ. rank(cid:20) C B − D1 A (cid:21) + rank (cid:2)T1 T2 T3(cid:3) + rank   6 3κ. 0 T6 T9   20 (That the ranks of the last two operators are each at most κ follows from the fact that the range of the second operator is contained in span {e1, e2, ..., eκ}, while the domain of the third operator is span { f1, f2, ..., fκ}.) (b) Next we consider the case where the underlying measure space is atomless. Let Z0 ⊆ X be measurable and such that PZ0 ∈ D satisfies rank(P⊥ TPZ0 = Z0 ∑κ TPZ0 ) = κ. Write P⊥ Z0 j=1 fj ⊗ g∗ Consider Yn, Zn, n > 1 chosen as in Lemma 4.6. Then j for some fj ∈ L2(Zc 0, µ), gj ∈ L2(Z0, µ), 1 6 j 6 κ. rank(P⊥ Zn TPZn ) 6 κ = sup P∈P (D) rank(P⊥TP) for all n > 1. On the other hand, PYn PZn = 0 implies that κ = rank (PYn TPZn) 6 rank (P⊥ Zn TPZn) 6 κ, and hence rank (P⊥ Zn TPZn) = κ. This allows us to write P⊥ Zn 1 6 j 6 κ and each of {uj}κ TPZn = ∑κ j=1 uj ⊗ v∗ n → C, vj j=1 is linearly independent. But then j where uj : Zc j=1 and {vj}κ : Zn → C, PYn(P⊥ Zn TPZn) = PYn TPZn = κ ∑ j=1 fjYn ⊗ (gjZn )∗ = κ ∑ j=1 uj ⊗ v∗ j . From this it follows that span {v1, v2, ..., vκ} = span {g1Zn , g2Zn , ..., gκZn }. Hence we may rewrite P⊥ Zn TPZn = u(n) j ⊗ (gjZn )∗ κ ∑ j=1 for some choice of u(n) j : Zn c → C, {u(n) j }κ j=1 linearly independent. But then κ ∑ j=1 u(n) j Yn ⊗ (gjZn )∗ = κ ∑ j=1 fjYn ⊗ (gjZn )∗ implies that u(n) j Yn = fjYn. Now, for m > n, we may write P⊥ Zm TPZm = κ ∑ j=1 u(m) j ⊗ (gjZm )∗. Also, for m > n, observe that P⊥ Zn (P⊥ Zm TPZm )PZm = P⊥ Zn (P⊥ Zn TPZn )PZm, which implies that κ ∑ j=1 u(m) j Zn c ⊗ (gjZm )∗ = κ ∑ j=1 u(n) j Yn ⊗ (gjZm )∗, which in turn implies that u(m) j Zn c = u(n) j Yn. This allows us to define uj : X → C by x 7→ lim n→∞ u(n) j (x), which is well-defined on X\ ∩∞ defined almost everywhere on X. n=1 Zn. Note that ∩∞ n=1Zn has measure zero, so that uj is 21 Note that for any n > 1, P⊥ Zn TPZn = κ ∑ j=1 ujZn c ⊗ (gjZn )∗. In a similar manner, by considering PYn TP⊥ Yn C so that , we can construct functions v1, v2, ..., vκ : X → Yn = PYn TP⊥ κ ∑ j=1 j=1 uj ⊗ v∗ j . We claim that both {u1, u2, ..., uκ} and {v1, v2, ..., vκ} are j=1 is linearly independent, Now consider F = ∑κ linearly independent sets. Indeed, since {ujYn }κ so is {u1, u2, ..., uκ}; similarly {v1, v2, ..., vκ} is linearly independent. Hence rank F = κ. j=1 = { fjYn }κ fjPYn X ⊗ (vjP⊥ for all n > 1. Yn X)∗ Let D = T − F. We shall prove that D ∈ D by showing that D commutes with every projection in D. Let R ∈ P (D) be a projection satisfying PZn n. That is, R = PV for some n. We have rank (R⊥TR) 6 κ. Consider R⊥TR measurable set V satisfying Zn ⊆ V ⊆ Yc as an operator from RH to R⊥H. Relative to the decomposition RH = PZnH ⊕ (PV\ZnH) and R⊥H = (PVc\YnH) ⊕ PYnH, we may write 6 R 6 PYc 0 R⊥(P⊥ (PYn TP⊥ Zn TPZn )RH = (cid:20)A 0 C 0(cid:21) ; C B(cid:21) ; C 0(cid:21) ; PVc\Yn TPV\Zn RH = (cid:20)0 X 0(cid:21) , Yn)RH = (cid:20) 0 PYn TPZn RH = (cid:20) 0 0 0 and so that (R⊥TR)RH = (cid:0)R⊥(P⊥ = (cid:20)A X C B(cid:21) . Zn TPZn) + (PYn TP⊥ Yn)R − PYn TPZn + PVc\Yn TPV\Zn(cid:1)RH Since rank (T⊥TR)RH = κ = rank C = rank (PYn TPZn )RH, by Lemma 3.2, X is uniquely determined by A, B and C. Observe that R⊥P⊥ Zn TPZn = R⊥( κ ∑ j=1 ujZc n ⊗ (vj∗ Zn ), PYn TP⊥ Yn = ( κ ∑ j=1 ujYn ⊗ (vjYc n)∗)R, and 22 PYn TPZn = ( κ ∑ j=1 ujYn ⊗ (vjZn )∗)R. But X = ∑κ j=1 ujVc\Yn ⊗ (vjV\Zn )∗ is a particular choice which satisfies rank (cid:20)A X C B(cid:21) = rank C = κ, for κ ∑ j=1 j=1, {vjZn }κ has rank κ by virtue of the fact that {ujYn}κ R⊥TR = ujR⊥ ⊗ (vjR)∗ j=1 are linearly independent families. Thus R⊥TR = R⊥FR for any projection R ∈ D with PZn Since µ(∩∞ n=1Zn) = 0 = µ(∩∞ n=1Yn), it follows that Q⊥TQ = Q⊥FQ for all projections 6 R 6 PYn. Q ∈ D. But then QD − DQ = Q(T − F) − (T − F)Q = (QT − TQ) − (QF − FQ) = (QTQ⊥ + QTQ) − (Q⊥TQ + QTQ) − (QF − FQ) = QFQ⊥ − Q⊥FQ − QF + FQ = −QFQ + QFQ = 0 for all projections Q ∈ D, whence D = T − F ∈ D, completing the proof in this setting. Finally, let us show that the decomposition T = D + F is unique. Let D1, D2 ∈ D and F1, F2 ∈ F (H) be such that T = D1 + F1 = D2 + F2. Then F1 − F2 = D2 − D1 ∈ D. Since all non-zero multiplication operators have infinite rank, we have F1 = F2. But then D1 = T − F1 = T − F2 = D2. (c) Now consider the case of an arbitrary, σ-finite measure space (X, µ). We first partition the space into its discrete and continuous parts (Xd, µd) and (Xc, µc) (where µd = µXd and µc = µXc respectively) and consider the corresponding decomposition of the Hilbert space: H = L2(X, µ) = L2(Xd, µd) ⊕ L2(Xc, µc). Let Pd (resp. Pc) denote the orthogonal projection of H onto H, and note that Pd, Pc ∈ D. As such, this produces a decomposition of the masa D = Dd ⊕ Dc. Relative to the above It is not too difficult to see that the decomposition of H, let us write T = (cid:20)T1 T2 T3 T4(cid:21). condition that supP∈P (D) rank (P⊥TP) = κ implies that and sup Pd(P⊥T1P)Pd 6 κ, P∈P (Dd) sup Pc(P⊥T4P)Pc 6 κ. P∈P (Dc) Also, the estimates rank T2, rank T3 6 κ are clear. From part (a), we may write T1 = D1 + F1 with T1 ∈ Dd, F1 ∈ B(L2(Xd, µd)) and rank F1 6 3κ. In a similar way, by part (b), there exist D4 ∈ Dc, F4 ∈ B(L2(Xc, µc)) with rank F4 6 κ so that T4 = D4 + F4. 23 Finally, T = D + F, where D = D1 ⊕ D4 ∈ D, and F = (cid:20)F1 T2 T3 F4(cid:21) satisfies rankF 6 rank F1 + rank T2 + rank T3 + rank F4 6 (3 + 1 + 1 + 1)κ = 6κ. ✷ 4.8. Remark. It should be clear that in the case of a discrete masa, no uniqueness of the decomposition T = D + F can be expected. Indeed, let, for example, D = ℓ∞ be a masa in B(ℓ2) and T ∈ B(H) be arbitrary. By Theorem 4.7 we can find a decomposition T = D + F with D ∈ D, F ∈ F (H). Put D1 = D + e1 ⊗ e∗ 1 and F1 = F − e1 ⊗ e∗ 1, where e1 = (1, 0, 0, . . . ) is the first basic vector. It is clear that D1 ∈ D, F1 ∈ F (H), and T = D1 + F1. The following theorem is a version of [23, Theorem 2.7] in the context of masas. 4.9. Theorem. Suppose that A is a norm-closed algebra and that D ⊆ B(L2(X, µ)) is a masa. Suppose also that for all projections P ∈ D and all A ∈ A we have that rank (AP − PAP) < ∞. Then there exists κ > 0 so that rank (AP − PAP) 6 κ for all A ∈ A and P ∈ P (D). Proof. Let Ek := {T ∈ A : rank (I − P)TP 6 k for all P ∈ P (D)}. Since every member in A is of the form D + F for some D ∈ D and F ∈ F (H) by Theorem 4.7, it follows that We now claim that Ek is closed for each k > 1. To that end, let P ∈ P (D). If Tn = Dn + Fn ∈ Ek and limn Tn = T ∈ A, then A = ∪∞ k=1Ek. (I − P)TP = lim n = lim n = lim n (I − P)TnP (I − P)(Dn + Fn)P (I − P)FnP. But rank (I − P)FnP 6 k for all n > 1 implies by the lower-semicontinuity of the rank that rank (I − P)TP 6 k, whence T ∈ Ek and Ek is closed. By the Baire Category Theorem, we get that the interior of Ek is non-empty for some k > 0. Choose such a k, and suppose that T0 = D0 + F0 lie in the interior of Ek. Then Ek − T0 := {T − T0 : T ∈ Ek} contains a ball in A of positive radius centred at 0. But then for all A ∈ A, there exists t > 0 so that tA ∈ Ek − T0. That is, tA = T1 − T0 for some T1 ∈ Ek. 24 Hence for all projections P ∈ D, rank(I − P)AP = rank(I − P)tAP = rank(I − P)(T1 − T0)P 6 rank(I − P)T1P + rank(I − P)T0P 6 k + k. Setting κ = 2k completes the proof. ✷ 4.10. Corollary. Let A ⊆ B(H) be a norm-closed algebra and D ⊆ B(H) be a masa such that (I − P)TP ∈ F (H) for all T ∈ A and P ∈ P (D). Then there exists k ∈ N such that A ⊆ D + Fk(H) where Fk(H) = {T ∈ B(H) : rank T 6 k}. In the case of a continuous measure space, we are able to say more about the structure of an algebra A satisfying the condition above. We need an auxiliary lemma. 4.11. Lemma. Suppose that L ⊆ Fk(H) is a linear subspace. Then there exist two invertible operators S and T in B(H) and a projection P = P∗ = P2 ∈ B(H) of rank at most k such that SLT ⊆ PB(H) + B(H)P. Proof. Pick an operator A0 ∈ L of maximal rank. Let κ = rank (A0). There exist two 0 0 satisfies the conclusion of the lemma. invertible operators S and T such that SA0T = (cid:20)Iκ 0 Indeed, let A ∈ L be arbitrary. Write SAT = (cid:20)A11 A12 A22 = 0. For every scalar λ, we have S(A + λA0)T = (cid:20)λIκ + A11 A12 A11 is invertible for all large enough λ, so that rank (cid:20)λIκ + A11 A12 0(cid:21). We claim that P = (cid:20)Iκ 0 0(cid:21) A21 A22(cid:21). It is enough to prove that A22(cid:21). The matrix λIκ + A22(cid:21) = rank (λIκ + A11) for large λ. By the reasoning analogous to that of Lemma 3.2, we get A22 = A21(λIκ + A11)−1 A12. Taking λ → ∞, we find that A22 = 0. (cid:3) 4.12. Theorem. Let µ be a continuous measure on a measure space X. Suppose that H = L2(X, µ) and A ⊆ B(H) is a closed algebra such that rank (P⊥TP) 6 κ for some constant κ and for all projections P ∈ D and all T ∈ A. Then there exist projections Q, R ∈ B(H) so that A21 A21 A ⊆ D + QB(H) + B(H)R, with rank Q 6 κ and rank R 6 κ. Remark. In view of Corollary 4.10 the existence of κ from the hypothesis of Theorem 4.12 follows from the condition rank (P⊥TP) < ∞ for all projections P ∈ D and all T ∈ A. Proof. By Theorem 4.7 (b), each T ∈ A may be written in a unique way as T = DT + FT, where DT ∈ D, and rank FT < ∞. Since A is a vector space, so is the set FA := {FT : T ∈ A}. Indeed, for T1, T2 ∈ A and λ ∈ C, λ1T1 + T2 = (λDT1 + DT2) + (λFT1 + FT2) 25 is a particular decomposition of λT1 + T2 as the sum of an element of the masa with a finite rank operator. Since the decomposition is unique, it follows that Dλ1T1+T2 = λ1DT1 + DT2, and similarly, Fλ1T1+T2 = λ1FT1 + FT2. Moreover, as we saw in Theorem 4.7(b), rank FT 6 κ for all T ∈ A, and so FA is a vector space of operators having rank at most κ. By Lemma 4.11, there exist invertible operators S, T ∈ B(H) and a projection of rank at most k such that FA ⊆ S−1PB(H)T−1 + S−1B(H)PT−1 ⊆ S−1PSB(H) + B(H)TPT−1. Now, if Q is the orthogonal projection onto the range of S−1PS and R is the orthogonal projection onto ker(TPT−1)⊥ then rank Q, rank R 6 κ and FA ⊆ QB(H) + B(H)R. ✷ 4.13. Example. While Theorem 4.12 guarantees the existence of projections Q, R ∈ B(H) so that A ⊆ D + QB(H) + B(H)R, there may not exist Q, R ∈ D so that A ⊆ D + QB(H) + B(H)R. To see this, let H = L2([0, 1], dx) where "dx" denotes Lebesgue mea- sure and let F ∈ B(H) be a rank one operator without invariant subspaces of the form L2(X, dx), where X ⊆ [0, 1] measurable. For example, we may let f : [0, 1] → C be the constant function f (x) = 1 for all x ∈ [0, 1], and set F = f ⊗ f ∗. Set A = CI + CF. Let D denote the canonical masa D ≃ L∞([0, 1], dx) in B(H). Then A is a norm-closed algebra and clearly A ⊆ D + F1 where F1 denotes the set of all rank one operators on H. 4.14. Remark. As we have just seen, the projections Q and R obtained in Theorem 4.12 do not have to lie in D. We remark, however, that if one of Q and R is zero -- say R = 0 -- then Q must commute with every D ∈ D appearing as the diagonal of an element T ∈ A. That is, for all T = DT + FT ∈ A, QDT = DTQ. Indeed, choose T0 ∈ A and P0 a projection in D so that rank(P⊥ 0 T0P0) = κ = sup rank (P⊥TP). T∈A, P∈P (D) Since T0 = D0 + F0 for some D0 ∈ D and F0 ∈ QB(H) by Theorem 4.12, it is clear that rank P⊥ 0 F0P = κ, and in particular, rank F0 = κ = rank Q. In other words, ran F0 = QH. For any T = D + F ∈ A, T0T ∈ A, and TT0 = (DD0) + (FD0 + DF0 + FF0). Since DD0 ∈ D and FD0 + DF0 + FF0 is finite-rank, the uniqueness of this decomposition (it is here that we need the measure to be continuous) implies that FD0 + DF0 + FF0 ∈ QB(H). Hence Q(FD0 + DF0 + FF0) = FD0 + DF0 + FF0, which implies that QDF0 = DF0. But ran F0 = ran Q, and thus QDQ = DQ. Thus ran Q is a finite-dimensional invariant subspace for the normal operator D. But any such subspace must be reducing for D, and hence QD = DQ. The rest of this section is motivated by [2, Proposition 5.1] which states that if T ∈ B(X ) is such that every half-space in X is T-almost-invariant then T has many invariant subspaces. We show that, at least for the case of operators acting on a Hilbert space H, much more is true: the operator must be of the form αI + F where rank F < ∞. 26 4.15. Proposition. Let T ∈ B(H). Suppose that given any masa D for B(H), we can find D ∈ D and a finite-rank operator F so that T = D + F. Then there exists α ∈ C and H a finite-rank operator so that T = αI + H. Proof. Let 0 6= P be any projection in B(H). By a routine application of Zorn's Lemma, we can find a masa D with P ∈ D. Let T = D + F, where D ∈ D and F ∈ F (H). Then D, P ∈ D implies that DP = PD, and hence TP − PT = FP − PF ∈ F (H). That is, TP − PT ∈ F (H) for every projection P in B(H). By a result of Fillmore [10] (see also the result of Matsumoto [20]), every X ∈ B(H) is a finite linear combination of projections. It follows that TX − XT ∈ F (H) for every X ∈ B(H). Let (en)∞ n=1 be an orthonormal basis for H, and let D0 denote the set of diagonal operators on H relative to this basis. Since D0 is again a masa in B(H), we may write T = D0 + F0 for some D0 = diag (dn)∞ n=1 ∈ D0 and F0 ∈ F (H). Let S ∈ B(H) denote the unilateral forward shift operator relative to this basis, so that Sen = en+1 for all n > 1. Since ST − TS ∈ F (H), we also have SD0 − D0S ∈ F (H). But SD0 − D0S =   0 d1 − d2 0 d2 − d3 0 d3 − d4 0 . . . . . . .   This operator is finite-rank if and only if there exists N > 1 so that dn − dn+1 = 0 for all n > N; i.e. dn = dN for all n > N. Thus D0 ∈ CI + F (H), and hence T = D0 + F0 ∈ CI + F (H) as well. ✷ 4.16. Corollary. Suppose that T ∈ B(H) and that every half-space of H is almost-invariant for T. Then T = αI + H for some α ∈ C and H ∈ B(H) a finite-rank operator. Proof. By Theorem 4.7, we know that if D is a masa in H then T = D + F for some D ∈ D and some finite-rank operator F. Since this is true for all masas, the result then follows from Proposition 4.15. ✷ 5. Almost-reducing subspaces As we have already seen, if X is a Banach space and A is a norm-closed subalgebra of B(X ) admitting a complemented almost-invariant half-space, then A admits an invariant half-space. Let H be a Hilbert space and ∅ 6= S ⊆ B(H). We shall say that a half-space M of H is reducing for S if the orthogonal projection Q of H onto M lies in the commutant S ′ := {T ∈ B(H) : TS = ST for all S ∈ S} of S. Equivalently, M must be invariant for both S and for S ∗ := {S∗ : S ∈ S}. We shall say that M is almost reducing for S if both M and M⊥ are almost-invariant for S, or equivalently, if M is almost-invariant for both S and S ∗. In light of the results of Section 3, it is reasonable to ask whether, in the Hilbert space setting, a norm-closed subalgebra A of operators in B(H) admitting an almost-reducing half-space admits a reducing subspace. The following example shows that this need not be the case. 27 5.1. Example. Let H = ℓ2(Z) with orthonormal basis {en}n∈Z. Recall that for x, y ∈ H, x ⊗ y∗ represents the rank-one operator x ⊗ y∗(z) = hz, yix for all z ∈ H. Let A denote the algebra of operators A whose matrix [aij] relative to the basis {en}n∈Z satisfies: aij = 0 if (i − j > 2, or i − j = 1 and i is odd. In other words, every matrix A in A is of the following form: . . . . . . ∗ ... e−3 e−2 e−1 e0 e1 e2 e3 ...   e−3 e−2 e−1 e0 e1 e2 e3 . . . ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗   ∗ ∗ ∗ ∗ ∗ ∗ . . . Observe that if D denotes the diagonal (bounded) operators relative to {en}n∈Z, then D ⊆ A, and D is a masa in B(H). If M is a reducing subspace for A, and if Q de- notes the orthogonal projection of H onto M, then Q ∈ A′, whence Q ∈ D′ = D. Thus the only possible reducing subspaces must be standard subspaces. (Recall that a stan- dard subspace relative to this masa is a subspace densely spanned by the basis vectors it contains.) Since Q ∈ D, we can write Q = diag{qn}n∈Z. Since Q = Q2, we have that qn ∈ {0, 1} for all n ∈ Z. Suppose that Q 6= 0, so that there exists m ∈ Z for which qm = 1. Case 1: yields: If m is even, then Wm := em−1 ⊗ e∗ m ∈ A. Since Q ∈ A′, QWm = WmQ, which qm−1(em−1 ⊗ e∗ m) = qm(em−1 ⊗ e∗ m). Thus qm−1 = qm = 1. Similarly, Vm := em+1 ⊗ e∗ m ∈ A, and the equation QVm = VmQ implies qm+1(em+1 ⊗ e∗ m) = qm(em+1 ⊗ e∗ m). Thus qm+1 = qm = 1. Case 2: Wm = em ⊗ e∗ QWm = WmQ, which imply that qm−1 = qm+1 = qm = 1. m+1 and Vm = em ⊗ e∗ If m is odd, we can employ a similar argument, only this time setting m−1. Again, Vm, Wm ∈ A and so QVm = VmQ, while 28 From this it readily follows that qn = 1 for all n ∈ Z, and so Q = I. We have shown that A does not admit any proper reducing subspaces. On the other hand, A admits a plethora of almost-reducing half-spaces. For example, if for n ∈ Z we set Hn := span{ek : k 6 n}, then each Hn is easily seen to be an almost- reducing half-space for A (with defect equal to 1). 5.2. Example. With a bit more effort, it is possible to construct a single irreducible operator T ∈ B(ℓ2(Z)) such that AT := alg(T) admits a large number of almost-reducing half-spaces. We shall exhibit such an operator T which lies in the algebra A defined above, implying that AT ⊆ A. In particular, every almost-reducing subspace of A is automatically almost reducing for AT. With {en}n∈Z as above, consider the operator T whose matrix is given by T = ... e−3 e−2 e−1 e0 e1 e2 e3 ...   . . . . . . ∗ e−3 e−2 e−1 e0 e1 e2 e3 . . . d−3 v−2 d−2 w−2 d−1 v0 d0 w0 d1 v2 d2 w2 d3 ∗ . . . .   The values of {dn}n∈Z, {w2n}n∈Z and {v2n}n∈Z will be chosen in such a way as to guar- antee that T is irreducible. We first decompose where Ho = [{e2k+1 : k ∈ Z}] and He := [{e2k : k ∈ Z}]. Relative to this decomposition, we may write ℓ2(Z) = Ho ⊕ He, where D0 = diag{d2k+1}k∈Z and De = diag{d2k}k∈Z. The precise form of T2 need not concern us just yet. T = (cid:20)Do T2 0 De(cid:21) , Suppose that P = (cid:20) Po Q Q∗ Pe(cid:21) is a projection commuting with T. Then Q∗Do Q∗T2 + PeDe(cid:21) . (cid:20)DoPo + T2Q∗ DoQ + T2Pe (cid:21) = (cid:20) PoDo PoT2 + QDe DeQ∗ DePe In particular, we have DeQ∗ = Q∗Do. Recall that if H and K are Hilbert spaces, A ∈ B(H), B ∈ B(K), then the Rosenblum operator τA,B : B(H, K) → B(H, K) defined by τA,B(X) = AX − XB has spectrum con- tained in σ(A) − σ(B) := {α − β : α ∈ σ(A), β ∈ σ(B)} (see, e.g., [14, Corollary 3.2]). In particular, therefore, if σ(A) ∩ σ(B) = ∅, then τA,B is invertible, and hence is injective. As such, if we choose Do and De so that σ(Do) ∩ σ(De) = ∅, then τDe,Do injective combined with the equation DeQ∗ = Q∗Do implies that Q∗ = 0, and hence Q = 0. 29 o Po = PoD∗ Furthermore, this will then force DoPo = PoDo and DePe = PeDe. By considering adjoints, we obtain D∗ e , so that Po belongs to the commutant (W∗(Do))′ of the von Neumann algebra (W∗(Do)) generated by Do and similarly, Pe ∈ (W∗(De))′. If we choose all {d2k+1 : k ∈ Z} and {d2k : k ∈ Z} to be distinct, then (W∗(Do)′ = W∗(Do) = [{e2k+1 ⊗ e∗ 2k : k ∈ Z}], which are simply the diagonal masas of Ho and He respectively. 2k+1 : k ∈ Z}], while W∗(De)′ = W∗(De) = [{e2k ⊗ e∗ e Pe = PeD∗ o and D∗ To satisfy both of the above conditions, it suffices to choose {d2k : k ∈ Z} a dense subset 4 , 1] with d2k 6= d2j unless j = k and {d2k+1 : k ∈ Z} dense in [ 1 2 ] with d2k+1 6= d2j+1 of [ 3 unless j = k. In particular σ(Do) ∩ σ(De) = [ 1 4 , 1 Combined with the previous observations, we find that Po and Pe are diagonal oper- ators relative to the bases {e2k+1 : k ∈ Z} and {e2k : k ∈ Z} respectively, which shows that 4 , 1 2 ] ∩ [ 3 4 , 1] = ∅. P = Po ⊕ Pe ≃ diag{..., p−2, p−1, p0, p1, p2, ...} lies in the masa D = [{en ⊗ e∗ {v2k : k ∈ Z} and {w2k : k ∈ Z} so as to guarantee that pi = pj for all i, j ∈ Z. n : n ∈ Z}] of B(ℓ2(Z)). There remains only to choose Now PT = . . . . . . ∗ ... e−3 e−2 e−1 e0 e1 e2 e3 ...   e−3 e−2 e−1 e0 e1 e2 e3 . . . p−3d−3 p−3v−2 p−2d−2 p−1w−2 p−1d−1 p−1v0 p0d0 p1w0 ,   p1d1 p1v2 p2d2 p3w2 p3d3 ∗ . . . 30 while TP = ... e−3 e−2 e−1 e0 e1 e2 e3 ...   . . . . . . ∗ e−3 e−2 e−1 e0 e1 e2 e3 . . . p−3d−3 p−2v−2 p−2d−2 p−2w−2 p−1d−1 p0v0 p0d0 p0w0 p1d1 .   p2v2 p2d2 p2w2 p3d3 ∗ . . . But PT = TP, and so p2k−1v2k = p2kv2k p2k+1w2k = p2kw2k, for all k ∈ Z. As long as w2k 6= 0 6= v2k for all k ∈ Z, we find that p2k−1 = p2k = p2k+1 for all k, from which it clearly follows that pi = pj for all i, j ∈ Z. For example, choosing v2k = 1 = w2k for all k ∈ Z will do. With these (highly non-unique) choices of dk, v2k and w2k, we find that any projection P commuting with T must be scalar - i.e., P = 0 or P = I. It follows that the corresponding T is irreducible, as required. References [1] Y. Abramovich and C. Aliprantis, An invitation to operator theory, Graduate Studies in Mathematics, 50. American Mathematical Society, Providence, RI, 2002. [2] G. Androulakis, A.I. Popov, A. Tcaciuc, V.G. Troitsky, Almost-invariant half-spaces of operators on Banach spaces. Integral Equations and Operator Theory, 65 (2009), 473 -- 484. [3] A. Bernstein, A. Robinson, Solution of an invariant subspace problem of K.T. Smith and P.R. Halmos, Pacific J. Math. 16 (1966), 421 -- 431. [4] A. Brown, C. Pearcy, Compact restrictions of operators, Acta Sci. Math. (Szeged) 32 (1971), 271 -- 282. [5] J.B. Conway, A course in functional analysis, Second edition. Graduate Texts in Mathematics, 96. Springer- Verlag, New York, 1990. [6] W. F. Donoghue, Jr., The lattice of invariant subspaces of a completely continuous quasi-nilpotent transformation, Pacific J. Math. 7 (1957), 1031 -- 1035. [7] K. Davidson, D. Herrero, The Jordan form of a bitriangular operator. J. Funct. Anal. 94 (1990), 27 -- 73. [8] P. Enflo, On the invariant subspace problem in Banach spaces, Seminaire Maurey -- Schwartz (1975 -- 1976) Espaces Lp, applications radonifiantes et g`eom`etrie des espaces de Banach, Exp. Nos. 14-15, 7 pp. Centre Math., ´Ecole Polytech., Palaiseau, 1976. [9] P. Enflo, On the invariant subspace problem for Banach spaces, Acta Math. 158 (1987), 213 -- 313. [10] P. Fillmore, Sums of operators with square zero, Acta Sci. Math. (Szeged) 28 (1967), 285 -- 288. [11] P.A. Fillmore, J.G. Stampfli, and J.P. Williams, On the essential numerical range, the essential spectrum, and a problem of Halmos, Acta Sci. Math. (Szeged), 33 (1972), 179 -- 192. [12] P.R. Halmos, Invariant subspaces of polynomially compact operators. Pacific J. Math. 16 (1966), 433 -- 437. [13] D.A. Herrero, Quasisimilarity does not preserve the hyperlattice. Proc. Amer. Math. Soc. 65 (1977), 80 -- 84. 31 [14] D.A. Herrero, Approximation of Hilbert space operators. Vol. 1, Second edition. Pitman Research Notes in Mathematics Series, 224. Longman Scientific & Technical, Harlow; copublished in the United States with John Wiley & Sons, Inc., New York, 1989. [15] D.A. Herrero, Triangular operators. Bull. London Math. Soc. 23 (1991), 513 -- 554. [16] T. Hoover, Quasi-similarity of operators. Illinois J. Math. 16 (1972), 678 -- 686. [17] B.E. Johnson and S.K Parrot, Operators commuting with a von Neumann algebra modulo the set of compact operators. J. Funct. Anal. 11 (1972), 39 -- 61. [18] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces. I. Sequence spaces. Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 92. Springer-Verlag, Berlin-New York, 1977. [19] V. Lomonosov, Invariant subspaces of the family of operators that commute with a completely continuous opera- tor, Funkcional. Anal. i Prilozen., 7 (3) (1973), 55 -- 56. [20] K. Matsumoto, Selfadjoint operators as a real span of 5 projections, Math. Japon. 29 (1984), 291 -- 294. [21] Murphy, G.J., C∗-algebras and Operator Theory, 1990, Academic Press, San Diego, London. [22] N. K. Nikol'skiı, The invariant subspaces of certain completely continuous operators, Vestnik Leningrad. Univ. 20 1965, 68 -- 77 (Russian). [23] A.I. Popov, Almost-invariant half-spaces of algebras of operators, Integral Equations Operator Theory 67 (2010), 247 -- 256. [24] C.R. Putnam, The spectra of operators having resolvents of first-order growth, Trans. Amer. Math. Soc. 133 1968 505 -- 510. [25] H. Radjavi, P. Rosenthal, Invariant subspaces, Second edition. Dover Publications, Inc., Mineola, NY, 2003. [26] H. Radjavi, P. Rosenthal, Simultaneous triangularization. Springer-Verlag, New York, 2000. [27] C. J. Read, A solution to the invariant subspace problem, Bull. London Math. Soc. 16 (1984), 337 -- 401. [28] C. J. Read, A solution to the invariant subspace problem on the space l1, Bull. London Math. Soc. 17 (1985), 305 -- 317. [29] C. J. Read, The invariant subspace problem for a class of Banach spaces. II. Hypercyclic operators, Israel J. Math. 63 (1988), 1 -- 40. [30] C. J. Read, Quasinilpotent operators and the invariant subspace problem, J. London Math. Soc. (2) 56 (1997), 595 -- 606. [31] C. J. Read, Strictly singular operators and the invariant subspace problem, Studia Math. 132 (1999), 203 -- 226. [32] G. Sirotkin, Infinite matrices with "few" non-zero entries and without non-trivial invariant subspaces. J. Funct. Anal. 256 (2009), 1865 -- 1874. [33] D. Voiculescu, A non-commutative Weyl-von Neumann theorem, Rev. Roumaine Math. Pures Appl. 21 (1976), 97 -- 113. [34] D.V. Yakubovich, Invariant subspaces of weighted shift operators, Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 141 (1985), 100 -- 143, 189 -- 190. Department of Pure Mathematics, University of Waterloo, Waterloo, Ontario, Canada N2L 3G1 E-mail address: [email protected] E-mail address: [email protected] E-mail address: [email protected]
1707.01141	1	1707	2017-07-04T19:44:11	John-Nirenberg Inequalities and Weight Invariant BMO Spaces	[ "math.FA" ]	This work explores new deep connections between John-Nirenberg type inequalities and Muckenhoupt weight invariance for a large class of $BMO$-type spaces. The results are formulated in a very general framework in which $BMO$ spaces are constructed using a base of sets, used also to define weights with respect to a non-negative measure (not necessarily doubling), and an appropriate oscillation functional. This includes as particular cases many different function spaces on geometric settings of interest. As a consequence the weight invariance of several $BMO$ and Triebel-Lizorkin spaces considered in the literature is proved. Most of the invariance results obtained under this unifying approach are new even in the most classical settings.	math.FA	math	JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES JAROD HART AND RODOLFO H. TORRES Abstract. This work explores new deep connections between John- Nirenberg type inequalities and Muckenhoupt weight invariance for a large class of BM O-type spaces. The results are formulated in a very general framework in which BM O spaces are constructed using a base of sets, used also to define weights with respect to a non-negative measure (not necessarily doubling), and an appropriate oscillation functional. This includes as particular cases many different function spaces on geo- metric settings of interest. As a consequence the weight invariance of several BM O and Triebel-Lizorkin spaces considered in the literature is proved. Most of the invariance results obtained under this unifying approach are new even in the most classical settings. 7 1 0 2 l u J 4 ] . A F h t a m [ 1 v 1 4 1 1 0 . 7 0 7 1 : v i X r a 1. Introduction Spaces of Bounded Mean Oscillation (BM O) have been, and continue to be, of great interest and a subject of intense research in harmonic analysis. One of the most fascinating aspects of BM O spaces is their self-improvement properties, which go back to the work of John and Nirenberg in [23]. The space BM O can be defined to be the collection of locally integrable functions f such that kf kBM O = sup Q⊂Rn 1 Q Q f (x) − fQdx < ∞. Here, as usual, fQ = 1 Q ´Q f (x)dx denotes the average of f over the cube Q, and the supremum is taken over the collection Q of all cubes in Rn with sides parallel to the axes. The crucial property of BM O functions is the John-Nirenberg inequality {x ∈ Q : f (x) − fQ > λ} ≤ c1Qe − c2λ kf kBM O where c1 and c2 depend only on the dimension. A well-known immediate consequence of the John-Nirenberg inequality is the p-power integrability, which we also refer to as p-invariance, kf kBM O ≈ sup Q∈Q(cid:18) 1 Q Q f (x) − fQpdx(cid:19)1/p , for all 1 < p < ∞. Moreover, it can also be proved that the above equivalence also hold for 0 < p < 1 even though the right hand-side is not a norm in 1 2 JAROD HART AND RODOLFO H. TORRES such a case. See for example the work of Stromberg [35] (or Lemma 5.1 below). The John-Nirenberg inequality also implies that ef (x)/ρ is locally inte- grable for any f ∈ BM O and an appropriate constant ρ > 0. This exponen- tial integrability led to a deep connection to Muckenhoupt weight theory, which in a rough sense says that log(A2) = BM O; here Ap denotes the class of Muckenhoupt weights of index 1 ≤ p ≤ ∞. More precisely, for every weight w ∈ A2, log(w) ∈ BM O, and for every f ∈ BM O (real-valued), there exists λ > 0 such that ef /λ ∈ A2; see for example the book of Garc´ıa-Cuerva and Rubio de Francia [12]. Another deep connection was made between Muckenhoupt weights and BM O in the work of Muckenhoupt and Wheeden [29]. They proved that a function f is in BM O if and only if f it is of bounded mean oscillation with respect to w for all w ∈ A∞. That is, if for each w ∈ A∞, define BM Ow to be the collection of all w-locally integrable functions f such that 1 w(Q) Q f (x) − fw,Qw(x)dx < ∞. kf kBM Ow = sup Q⊂Rn Then, BM O = BM Ow and (1) kf kBM Ow ≈ kf kBM O. Here w(Q) = ´Q w(x)dx is the w-measure of Q, and fw,Q = f (x)w(x)dx. 1 w(Q) Q Quantitative refinements of the above weight invariant result, were re- cently obtained by Hytonen and P´erez [19] and Tsutsui [37], who gave pre- cise control of the constants appearing in (1). Other weight invariant results in the literature include, for example, the work of Bui and Taibleson [5], Harboure, Salinas, and Viviani [13] and an article by the first author and Oliveira [15]. In [5, Theorem 3], the authors show that the weighted endpoint Triebel-Lizorkin spaces F α,∞ ∞,w coincide for all w ∈ A∞. In [13, Proposition 4], the authors show that BM O, which agrees with the Triebel-Lizorkin space F 0,2 ∞ , continuously embeds into a weighted version F 0,2 ∞,w for all w ∈ A1 with comparable norms. Moreover, it is shown in [15, Theorem 2.11] that for any s > 0, ∞,w for all w ∈ A2 and F 0,2 ∞ = F 0,2 ∞ = Is(BM O) = Is(BM Ow) = F s,2 F s,2 ∞,w, for all w ∈ A∞ with comparable norms; here Is(BM O) denotes the Sobolev- BM O spaces defined by Neri [30] and studied in depth by Strichartz [33, 34]. The main purpose of this article is to further explore the connections be- tween John-Nirenberg type inequalities and Muckenhoupt weight invariance for generalized BM O spaces and apply them to obtain several new results in a diversity of contexts. These connections are made for BM O-type spaces formulated in a very general setting involving a non-negative measure µ, an JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 3 oscillation functional Λ, a base of sets B used to define weights, some class of "functions" F, and a space X(B, µ, F, Λ) = X(B, Λ) = X consisting of all the f ∈ F for which kf kX = sup B∈B 1 µ(B) B Λ(f, B)(x)dµ(x) < ∞. For each f ∈ F and B ∈ B, Λ(f, B)(x) is assumed to be a non-negative, µ-locally integrable function. In principle, these objects generalize the role of dµ = dx the Lebesgue measure, B = Q, and Λ(f, Q)(x) = f (x) − fQ for f ∈ F = L1 loc(Rn), in which case of course X(Q, Λ) = BM O. However, there is much more flexibility in this definition. For instance, we can ap- ply our results in settings where B is the collection of balls, dyadic cubes, rectangles, dyadic rectangles, or other collections of sets, hence including various BM O spaces in the literature associated to different geometries. The results are also applicable in the situation where F is the collection of locally integrable function, tempered distributions, or distributions modulo polynomials, and the functional Λ takes various appropriate forms leading to distribution spaces like Triebel-Lizorkin spaces. Moreover, the general approach we follow also works in a general σ-finite measure space and hence we obtain, not surprisingly, versions of results in the context of spaces of ho- mogenous type, but more interestingly also in non-doubling settings where many tools of harmonic analysis fail. The main properties for X(B, Λ) of interest to us are weight invariance and John-Nirenberg p-power integrability or p-invariance. Our abstract setup is reminiscent of the one in the work by Franchi, P´erez, and Wheeden in [9] and others. The authors in [9] considered inequalities of the form (2) 1 µ(B) B f (x) − fBdµ(x) . Γ(f, B) for appropriate functionals Γ taking values now in R, and which lead to self- improvements. In particular they investigated conditions on Γ that allow the left-hand side of (2) to be replaced by its p-power version for 1 < p < ∞. Several other authors (we shall give some references later on) have followed such abstract approaches too. Nonetheless, and although one of our appli- cations will overlap with results in [9], our results are geared more towards the establishment of several weighted norm equivalences that, incidentally, will also work for 0 < p < 1. We show that under remarkably minimal assumptions on the objects defining the spaces X(B, Λ), which apply to many situations, weight invariance and p-invariance are essentially equiv- alent properties. The precise statements are given in Theorems 3.2, 3.3, and 4.2. We then use known p-invariance results for several function spaces to prove their weight invariance as well. The following are a few examples of new invariance results obtained as corollaries of Theorems 3.2 and 3.3, which showcase the convenience and benefits of the very general framework 4 JAROD HART AND RODOLFO H. TORRES used. Here we state abridged versions of these results; see the corresponding theorems in Section 5 for their complete statements. Theorem 5.2. (The John-Nirenberg BMO space) Let 0 < p < ∞ and w, v be in A∞. Then kf kBM O ≈ sup Q⊂Rn(cid:18) 1 w(Q) Q f (x) − fv,Qpw(x)dx(cid:19) 1 p . Theorem 5.7. (BMO spaces with respect to non-doubling measures) Let µ be a non-negative measure in Rn such that µ(L) = 0 for any hyperplane L orthogonal to one of the coordinate axes. Also let 0 < p < ∞ and v, w ∈ A∞(µ). Then kf kBM Oµ(Rn) ≈ sup Q∈Q(cid:18) 1 µw(Q) Q f (x) − fµv,Qpdµw(x)(cid:19)1/p . Theorem 5.9. (Duals of weighted Hardy spaces) Let 0 < r < ∞, w be in A1, v be in A∞(w), and ρ = v · w. Then for all f in BM O∗,w kf kBM O∗,w : = sup Q∈Q f (x) − fQdx 1 w(Q) Q Q∈Q(cid:18) 1 ρ(Q) Q ≈ sup f (x) − fQrw(x)1−rv(x)dx(cid:19)1/r . Theorem 5.14. (Little bmo) Let 0 < p < ∞ and v and w be rectangular weights in the class A∞(Rn1 × Rn2). Then kf kbmo ≈ sup R∈R(cid:18) 1 w(R) R f (x) − fv,Rpw(x)dx(cid:19)1/p . Theorem 5.18. (End-point Triebel-Lizorkin spaces) Let α ∈ R, 0 < p, q < ∞, and w in A∞. Then for all f in F α,q ∞ kf k F α,q ∞ ≈ sup Q∈Q  1 w(Q) Q Xk∈Z:2−k≤ℓ(Q) (2αkϕk ∗ f (x))q 1 p . p q w(x)dx The rest of this article is organized as follows. We introduce the terminol- ogy employed above and most of the notation used throughout the article in Section 2. In Sections 3 and 4, we prove several results related to the equivalence between the p-invariance and weight invariance for any BM O type space X(B, Λ); in particular Theorems 3.3 and 4.2 alluded to before. Finally in Section 5 we present the proofs of the theorems stated above, as well as some extensions of them and other applications. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 5 2. Definitions and Preliminaries Let (Y, Σ, µ) be a measure space, with µ non-negative and σ-finite, and let B be a collection of sets in the σ-algebra Σ of µ-measurable sets such that 0 < µ(B) < ∞ for all B ∈ B. Let also MB be a sublinear operator that maps L1 loc(µ) denotes the collection of all µ-measurable functions f : Y → C such that loc(µ) into non-negative µ-measurable functions. Here L1 ´B f (x)dµ(x) < ∞ for all B ∈ B. Definition 2.1. Given 1 < p < ∞, a non-negative locally µ-integrable function w is in the Muckenhoupt class of weights Ap(B, µ) = Ap(B), or simply Ap, if [w]Ap(B) := sup B∈B(cid:18) 1 µ(B) B w(x)dµ(x)(cid:19)(cid:18) 1 µ(B) B w(x)−p′/pdµ(x)(cid:19)p/p′ < ∞, where, as usual, p′ denotes the Holder conjugate of 1 < p < ∞ given by p′ = p p−1. Define A∞(B) to be the union of all Ap(B) for 1 < p < ∞. We say that w ∈ A1(B) if MBw(x) ≤ [w]A1(B)w(x), where [w]A1(B) denotes the minimal constant satisfying this inequality. Define also for 1 < δ < ∞ the reverse Holder class RHδ(B) to be the collection of all w ∈ A∞(B) such that (cid:18) 1 µ(B) B w(x)δdµ(x)(cid:19)1/δ ≤ [w]RHδ 1 µ(B) B w(x)dµ(x), for all B ∈ B. Here again the reverse Holder constant [w]RHδ for w is the smallest constant such that the above inequality holds. It will be necessary sometimes to specify the reverse Holder class with respect to a measure µ and we will write RHδ(B, µ) or RHδ(µ) in such a case. Remark 2.2. In applications MB will be a maximal function associated to the base B. It is interesting, however, that we do not need MB to be so to prove our general results. Moreover even if MB is such a maximal function, we are not assuming that the base is a Muckenhoupt base, i.e. that the Ap classes characterize, or even suffice for, the boundedness of MB on weighted Lp spaces. On the other hand, we note that in the above general setting, we will not be able to use any self-improving property on the weights unless we impose a reverse Holder condition. Definition 2.3. Let F be a set (typically of functions or distributions), and let Λ be a mapping from F × B into the collection of µ-measurable functions on Y , such that Λ(f, B) is non-negative for all f ∈ F and B ∈ B. We consider the following spaces and properties. (i) For w ∈ A∞(B) and 0 < p < ∞, X p w(B, µ, F, Λ) is the collection of all f ∈ F such that kf kp X p w := sup B∈B 1 w(B) B Λ(f, B)(x)pw(x)dµ(x) < ∞, 6 JAROD HART AND RODOLFO H. TORRES w(B, Λ), or just X p where w(B) = ´B w(x)dµ(x). We will simply write X p w(B, µ, Λ), more typically X p w, when the other objects in the definition of X p w(B, µ, F, Λ) are clear from or do not play a significant role in the particular context being considered. We will also use the notation X p(B, Λ) = X p 1 (B, Λ) when w is the constant function 1, Xw(B, Λ) = X 1 1 (B, Λ) when both w = 1 and p = 1. We will call X p w(B, Λ) an "oscillation space" and Λ an "oscillation functional". w(B, Λ) when p = 1, and X(B, Λ) = X 1 (ii) The oscillation space X(B, Λ) satisfies the weight invariance property for a collection of weights W ⊂ A∞(B) if Xw(B, Λ) = Xv(B, Λ) for all w, v ∈ W, and there is a constant bw,v > 0 such that kf kXv ≤ bw,vkf kXw for all f ∈ Xw(B, Λ). Without loss of generality, we let bw,v be the smallest such constant. (iii) For an oscillation space X(B, Λ) that satisfies the weight invariance property for all weights in Ap(B), define for t ≥ 1 the function (3) Ψp(t) := sup{b1,v : v ∈ Ap(B), [v]Ap ≤ t}. (iv) The oscillation space Xw(B, Λ) satisfies the p-power John-Nirenberg property, or p-invariance property, for an interval I ⊂ (0, ∞) if for any p, q ∈ I with p < q there exists a constant C(w) > 0 (which may also depend on p and q) such that kf kX q w ≤ C(w)kf kX p w for all f ∈ Xw(B, Λ). We let cp,q(w) be the infimum of the constants such that the above inequality holds. When w = 1, we use the notation cp,q = cp,q(1). Throughout this article, we reserve the lower case subscripted letters bw,v and cp,q(w), to play the role they do in the definitions above. Remark 2.4. Note that since Λ(f, B) is a non-negative µ-measurable func- tion, it follows that Λ(f, B)p w is also a non-negative µ-measurable function for all f ∈ F, B ∈ B, 0 < p < ∞, and w ∈ A∞(B, µ). Hence the integral w is well-defined, though possibly infinite for some used to compute k · kX p elements f ∈ F. Note also that despite the notation k·kX p w and the name oscillating space, the collection X p w(B, Λ) may not be a Banach space, a normed space, or even a vector space. We use this notation because it is conducive to think of it in this context, even though we do not need linearity, completeness, or a normed space structure for our computations. If X and Z are either oscillation spaces or normed function spaces, the notation X ≈ Z will always mean that X = Z as sets and kf kX ≈ kf kZ for all of their elements. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 7 Remark 2.5. The reason for introducing the technically looking function Ψp in (3) will become apparent in Section 4, where it will be used to quantify some uniform estimates. Note that for each p, Ψp(t) is obviously a non- decreasing function of t that in principle could take the value ∞ for some or all t ≥ 1. Remark 2.6. It is important to remark that the inequality kf kX q w ≤ cp,q(w)kf kX p w w(B, Λ) or f ∈ w(B, Λ). Hence, there is a distinction between the p-power John-Nirenberg in (iv) is only required for f ∈ Xw(B, Λ), not for f ∈ X p X q property when p ≥ 1 and when p < 1. Indeed, suppose X(B, Λ) satisfies the p-power John-Nirenberg property for [1, ∞). This means that kf kX p ≤ c1,pkf kX for all 1 < p < ∞ and f ∈ X(B, Λ); hence X(B, Λ) ⊂ X p(B, Λ). It is immediate by Holder's (or Young's) inequality that kf kX ≤ kf kX p, and hence X p(B, Λ) ⊂ X(B, Λ). Therefore the John-Nirenberg property for X(B, Λ) on [1, ∞) implies that X(B, Λ) = X p(B, Λ) and k · kX ≈ k · kX p for all 1 ≤ p < ∞. On the other hand, when X(B, Λ) satisfies the p-power John-Nirenberg property for (0, 1], we cannot make such a strong conclusion. In this case, it follows that kf kX p ≤ kf kX ≤ cp,1kf kX p for all f ∈ X(B, Λ). The restriction to "functions" in X(B, Λ) is of substance here. We cannot conclude that f ∈ X p(B, Λ) implies f ∈ X(B, Λ) or that X(B, Λ) = X p(B, Λ) in this case. We can only conclude that k · kX ≈ k · kX p for all 0 < p < 1 when restricted to X(B, Λ). Remark 2.7. There is a plethora of BM O spaces that can be realized as X(B, µ, F, Λ) spaces with the appropriate choices of B, µ, F, and Λ. For example, every space mentioned in the introduction can be realized as such a space in an obvious way suggested by their standard definitions. More details are given in Section 5. We will now lay out several assumptions on µ, B, and MB. Different subsets of these assumptions will be used to prove different results. We will refer to them throughout as assumptions A1–A4, defined as follows. A1. A1(B) ⊂ Ap(B) and [w]Ap ≤ [w]A1 for all 1 < p < ∞. A2. A∞(B) =S1<δ<∞ RHδ(B). and lim supp→1+ kMBkLp,Lp = ∞. A3. For all 1 < p < ∞, the operator MB satisfies 0 < kMBkLp,Lp < ∞ A4. There exist functions ∆ : (1, ∞)2 → (1, ∞), non-increasing in each variable, and K : (1, ∞)2 → (1, ∞), non-decreasing in each variable, such that if u ∈ Ap(B) then u ∈ RH∆(p,[u]Ap)(B) with [u]RH∆(p,[u]Ap ) ≤ K(p, [u]Ap) for all 1 < p < ∞. 8 JAROD HART AND RODOLFO H. TORRES Remark 2.8. It is well-known that assumptions A1–A4 hold in many sit- uations when MB is the centered or uncentered Hardy-Littlewood maximal function associated to B. It is worth noting that in the non-doubling setting, it is necessary to choose MB to be the centered maximal operator to assure A3 holds, since the uncentered Hardy-Littlewood maximal operator associ- ated to a non-doubling measure µ may not be Lp(µ) bounded for 1 < p < ∞. Property A2 essentially represents the existence of a reverse Holder inequal- ity, while A4 is a quantified version of that. Actually A4 implies A2, but we list both since only A2 will be needed in some of our arguments. We direct the reader to [20, 19, 27] and the reference therein for more information on the recent interest in sharp quantitative versions of the reverse Holder inequality. In all the cases in which we are aware A4 holds, ∆ and K can be taken in the form ∆(p, t) = 1 + 1 τ (p)t and K(p, t) = C, for an appropriate function τ (p) and constant C > 0. For example, the following choices for ∆ and K are sufficient to satisfy A4 in the corresponding settings: 1 • ∆(p, t) = 1 + 2n+1t−1 and K(p, t) = 2 for the standard Euclidean setting using cubes and the Lebesgue measure (see [20, Theorem 2.3]). • ∆(p, t) = 1 + 1 2p+2t and K(p, t) = 2 for the Euclidean setting using rectangles and the Lebesgue measure (see [27, Theorem 1.2]). • ∆(p, t) = 1 + 1 τ t and K(p, t) = C for the space of homogeneous type setting, where τ and C are fixed constants depending on parameters of the underlying space (see [20, Theorem 1.1]). 1 • ∆(p, t) = 1+ 2p+1B(n)t and K(p, t) = 2 for the Euclidian setting using non-doubling measures, where B(n) is the Besicovitch constant for Rn (see [31, Lemma 2.3] and the remarks in the introduction of [27]). Actually, in some of these examples we can take K(p, t) = 21/∆(p,t). However this dependance on p and t is of no relevance since ∆(p, t) ≥ 1 and hence 1 ≤ K(p, t) ≤ 2. Therefore, for simplicity, we shall use K(p, t) = 2. 3. Sufficient Conditions for Weight Invariance In this section we provide sufficient conditions for weight invariance of X(B, Λ). The first theorem of this section holds solely based on the definition of the Ap(B) weights, while the second one only needs as an additional assumption A2. Proposition 3.4 needs the stronger assumption A4. Lemma 3.1. Fix w ∈ A∞(B). If there exist 0 < q0 < p0 < ∞, with q0 ≤ 1 such that Xw(B, Λ) satisfies the p-power John-Nirenberg property on the interval [q0, p0], then Xw(B, Λ) also satisfies the p-power John-Nirenberg property on the interval (0, p0]. Proof. Let q = inf{r > 0 : Xw(B, Λ) is p-invariant on [r, p0]}. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 9 Clearly 0 ≤ q ≤ q0. By way of contradiction, assume that q > 0, and let ǫ > 0 and r > q be selected so that q < r − ǫ < p0 but 0 < r − 2ǫ < q. Then, for any f ∈ Xw(B, Λ), kf kr−ǫ X r−ǫ w kΛ(f, B)r/2−ǫkL2 w(B)kΛ(f, B)r/2kL2 w(B) 1 ≤ sup B w(B) ≤ kf k(r−2ǫ)/2 X r−2ǫ w kf kr/2 X r w ≤ cr/2 r−ǫ,r(w)kf k(r−2ǫ)/2 X r−2ǫ w kf kr/2 X r−ǫ w . Because of the invariance on [q0, p0] and the fact that q0 ≤ 1, we also have kf kX r−ǫ w . kf kX q w ≤ kf kXw < ∞. Therefore r kf kX r−ǫ w ≤ c r−2ǫ r−ǫ,r(w)kf kX r−2ǫ w , contradicting the definition of q. (cid:3) The next theorem shows that p-invariance in a limited range of exponents implies appropriate weight invariance. Theorem 3.2. If X(B, Λ) satisfies the p-power John-Nirenberg property for [1, p0] for some p0 > 1 and X(B, Λ) = X p(B, Λ) for all 0 < p ≤ 1, then X(B, Λ) satisfies the weight invariant property for RHp′ (B). In particular, for each 1 < q < ∞ such that w ∈ Aq(B) ∩ RHp′ (B), it holds that 0 0 (c1/q,1[w]Aq )−1kf kX ≤ kf kXw ≤ c1,p0[w]RHp′ 0 kf kX for all f ∈ X(B, Λ). If in addition w ∈ RHσ for some σ > p′ p-power John-Nirenberg property for (0, p0/σ′ ] and 0, then Xw(B, Λ) satisfies the c1,p0/σ′(w) ≤ c1,p0c1/q,1 ([w]RHσ ) σ p0(σ−p′ 0 ) [w]Aq . Proof. Assume X(B, Λ) satisfies the p-power John-Nirenberg property for [1, p0] for some p0 > 1. Fix w ∈ RHp′ (B). Then for f ∈ X(B, Λ) 0 kf kXw = sup B∈B 1 w(B) B w(B)(cid:18) 1 µ(B) B µ(B) ≤ sup B∈B ≤ [w]RHp′ c1,p0kf kX . 0 Λ(f, B)(x)w(x)dµ(x) Λ(f, B)(x)p0dµ(x)(cid:19) 1 p0(cid:18) 1 µ(B) B w(x)p′ 0dµ(x)(cid:19) 1 p′ 0 Hence X(B, Λ) ⊂ Xw(B, Λ). Let w ∈ Aq(B) for some 1 < q < ∞. For f ∈ Xw(B, Λ), it follows that kf kX 1/q = sup B∈B(cid:18) 1 µ(B) B Λ(f, B)(x)1/qw(x)1/qw(x)−1/qdµ(x)(cid:19)q 10 JAROD HART AND RODOLFO H. TORRES ≤ sup B∈B(cid:18) 1 w(B) B ≤ [w]Aq kf kXw , Λ(f, B)(x)w(x)dµ(x)(cid:19) × µ(B)(cid:19)(cid:18) 1 (cid:18)w(B) µ(B) B w(x)−q′/qdµ(x)(cid:19)q/q′ and so f ∈ X 1/q(B, Λ). By assumption X 1/q(B, Λ) = X(B, Λ), and therefore X(B, Λ) = Xw(B, Λ) for all w ∈ RHp′ (B). It also follows that 0 kf kX ≤ c1/q,1kf kX 1/q ≤ c1/q,1[w]Aq kf kXw . By modifying some of the arguments we can obtain the p-invariance of Xw(B, Λ). Indeed, let 1 < p < p0 and w ∈ RH p0 p0−p (B). Note that p(cid:19)′ (cid:18) p0 = p0 p0 − p > p′ 0. Hence w ∈ RHp′ here. Then for f ∈ X(B, Λ) = Xw(B, Λ), 0 as well, and so the first part of this theorem is applicable kf kX p w = sup B∈B(cid:18) 1 w(B) B w(B)(cid:19) 1 B∈B(cid:18) µ(B) Λ(f, B)(x)pw(x)dµ(x)(cid:19) 1 p(cid:18) 1 µ(B) B p ≤ sup p0 Λ(f, B)(x)p0dµ(x)(cid:19) 1 (cid:18) 1 µ(B) B w(x) × p0 p0−p dµ(x)(cid:19) p0−p p0p ≤ [w] 1 p RH p0 p0−p c1,p0kf kX ≤ c1/q,1c1,p0[w] 1 p RH p0 p0−p [w]Aq kf kXw . Therefore Xw(B, Λ) satisfies the p-power John-Nirenberg property on [1, p] when w ∈ RH p0 (cid:3) p0−p , and by Lemma 3.1 it does so on (0, p] too. We now show that the full p-invariance for one weight, implies p-invariance and weight invariance for all weights in A∞(B). Theorem 3.3. Suppose µ and B satisfy the assumption A2. Assume that for some w0 ∈ A∞(B), Xw0(B, Λ) satisfies the p-power John-Nirenberg property for the interval [1, ∞), and that X p w0(B, Λ) = Xw0(B, Λ) for all 0 < p < 1. Then X(B, Λ) satisfies the weight invariance property for A∞(B), and Xw(B, Λ) satisfies the p-power John-Nirenberg property for (0, ∞) for every w ∈ A∞(B). Furthermore, if w0 ∈ Ap(B)∩RHσ(B) and w ∈ Aq(B)∩RHδ(B) for 1 < p, q, δ, σ < ∞, then bw0,w ≤ c1,pδ′(w0)[w0] 1 pδ′ Ap [w]RHδ , JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 11 bw,w0 ≤ c(qσ′)−1,1(w0)[w]Aq [w0]qσ′ , and c1,t(w) ≤ c1,tpδ′ (w0)bw,w0[w0] RHσ 1 pδ′t Ap [w] 1 t RHδ for all 1 < t < ∞. Proof. Assume that w0 ∈ A∞(B), that Xw0(B, Λ) satisfies the p-power John- Nirenberg inequality for the interval [1, ∞), and that X p w0(B, Λ) = Xw0(B, Λ) for all 0 < p < 1. In particular, by assumption A2, w0 ∈ Ap(B) ∩ RHσ(B) for some 1 < p, σ < ∞. Also let w ∈ A∞(B) with w ∈ Aq(B) ∩ RHδ(B) for some 1 < q, δ < ∞, where again such δ exists by assumption A2. Define r = pδ′ > 1 and s = δ′ + p′δ′ pδ > 1. These numbers are chosen so that sr′/r = p′/p and s′r′ = δ. Then for any f ∈ Xw0(B, Λ) we have Λ(f, B)(x)w0(x)1/rw0(x)−1/rw(x)dµ(x) Λ(f, B)(x)rw0(x)dµ(x)(cid:19) 1 w(B) (cid:18)B r(cid:18)B w0(x)−sr′/rdµ(x)(cid:19) 1 w0(B)1/r sr′(cid:18)B w0(x)−r′/rw(x)r′ w(x)s′r′ r′ dµ(x)(cid:19) 1 dµ(x)(cid:19) 1 s′r′ kf kXw = sup B∈B 1 w(B) B ≤ sup B∈B 1 w(B)(cid:18)B ≤ kf kX r w0 sup B∈B ≤ c1,r(w0)kf kXw0 µ(B) w(B) sup B∈B w0(B)1/r µ(B)1/r ×(cid:18) 1 µ(B) B [w]RHδ kf kXw0 1 r Ap w0(x)−p′/pdµ(x)(cid:19) p p′r(cid:18) 1 w(x)δdµ(x)(cid:19) 1 δ µ(B) B [w]RHδ kf kXw0 1 pδ′ Ap ≤ c1,r(w0)[w0] = c1,pδ′(w0)[w0] Therefore Xw0(B, Λ) ⊂ Xw(B, Λ) and Define bw0,w ≤ c1,pδ′(w0)[w0] 1 pδ′ Ap r = qσ′ > 1 and s = σ′ + [w]RHδ . q′σ′ qσ > 1, where now these numbers are chosen so that sr′/r = q′/q and s′r′ = σ. It follows that, for f ∈ Xw(B, Λ), kf k X 1/r w0 = sup B∈B(cid:18) 1 w0(B) B Λ(f, B)(x)1/rw0(x)w(x)1/rw(x)−1/rdµ(x)(cid:19)r 12 JAROD HART AND RODOLFO H. TORRES ≤ sup B∈B 1 w0(B)r(cid:18)B Λ(f, B)(x)w(x)dµ(x)(cid:19)(cid:18)B w0(x)r′ w(x)−r′/rdµ(x)(cid:19)r/r′ µ(B)r−1w(B) w0(B)r ×(cid:18) 1 µ(B) B µ(B)r−1w(B) w0(B)r ≤ kf kXw sup B∈B ≤ kf kXw sup B∈B ≤ [w]Aq [w0]r w0(x)σdµ(x)(cid:19)r/σ(cid:18) 1 (cid:18)[w0]RHσ µ(B) B µ(B) (cid:19)r(cid:18)[w]Aq w0(B) w(x)−q′/qdµ(x)(cid:19)q/q′ w(B)(cid:19) µ(B) RHσ kf kXw = [w]Aq [w0]qσ′ RHσ kf kXw . Therefore Xw(B, Λ) ⊂ X 1/r w0 (B, Λ). By assumption, Xw0(B, Λ) = X 1/r and hence it follows that Xw0(B, Λ) = Xw(B, Λ). It also follows that w0 (B, Λ), bw,w0 ≤ c(qσ′)−1,1[w]Aq [w0]qσ′ RHσ . It remains to be shown that for any w ∈ A∞(B), Xw(B, Λ) is also p- invariant. So let again w0 ∈ Ap(B) ∩ RHσ(B) and w ∈ Aq(B) ∩ RHδ(B) for 1 < p, q, δ, σ < ∞, and so that as before r = pδ′ and s = 1 + p′(δ′ − 1), sr′/r = p′/p and s′r′ = δ. Fix 1 < t < ∞ and f ∈ Xw(B, Λ). We already know that Xw(B, Λ) = Xw0(B, Λ) and that Xw0(B, Λ) is p-invariant for the interval [1, ∞), hence Λ(f, B)(x)tw(x)dµ(x) 1 w(B) B w(B)(cid:18)B ≤ 1 ≤ c1,tr(w0)tkf kt Xw0 ×(cid:18)B w0(B)1/r w(B) Λ(f, B)(x)trw0(x)dµ(x)(cid:19)1/r(cid:18)B w0(x)−r′/rw(x)r′ dµ(x)(cid:19)1/r′ w0(x)−sr′/rdµ(x)(cid:19)1/(sr′)(cid:18)B w(x)s′r′ dµ(x)(cid:19)1/(s′r′) ≤ c1,tr(w0)tbt w,w0kf kt Xw w0(B)1/r w(B) ×(cid:18)B w0(x)−p′/pdµ(x)(cid:19)p/(rp′)(cid:18)B w(x)δdµ(x)(cid:19)1/δ ≤ c1,tr(w0)tbt w,w0kf kt Xw [w0] [w]RHδ . 1 pδ′ Ap Therefore c1,t(w) ≤ c1,tpδ′ (w0)bw,w0[w0] 1 pδ′t Ap [w] 1 t RHδ . (cid:3) JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 13 In the next result we impose more structure on the behavior of µ and B through the assumption A4. In this way, we are able to ensure that Ψp, as defined in (3), is finite when X(B, Λ) satisfies the p-power John-Nirenberg property. The proposition below essentially says that if we can control in a quantitative way the reverse Holder constant for a weight by its Ap norm, then we can also quantify the b1,w constants in the weight variance estimates. Proposition 3.4. Suppose µ and B satisfy the assumption A4. If X(B, Λ) satisfies the p-power John-Nirenberg property for [1, ∞), then for all 1 < p < ∞. Ψp(t) ≤ K(p, t)c1,∆(p,t)′ Proof. Let 1 < p < ∞, w ∈ Ap(B) with [w]Ap ≤ t. By A4, we have ) ≤ K(p, [w]Ap). Since ∆ is non- that w ∈ RH∆(p,[w]Ap) and [w]RH∆(p,[w]Ap increasing in each variable, it follows that RH∆(p,[w]Ap) ⊂ RH∆(p,t) and ). However, since K is non-decreasing in each [w]RH∆(p,t) ≤ [w]RH∆(p,[w]Ap variable K(p, [w]Ap) ≤ K(p, t), and hence [w]RH∆(p,t) ≤ K(p, t). Therefore, Λ(f, B)(x)w(x)dµ(x) 1 w(B) B w(B)(cid:18) 1 µ(B) ≤ µ(B) B Λ(f, B)(x)∆(p,t)′ dµ(cid:19) 1 ∆(p,t)′(cid:18) 1 µ(B) B w(x)∆(p,t)dµ(cid:19) 1 ∆(p,t) ≤ c1,∆(p,t)′kf kX[w]RH∆(p,t) ≤ c1,∆(p,t)′K(p, t)kf kX . It follows that b1,w ≤ K(p, t)c1,∆(p,t)′ for all w as specified above and hence Ψp(t) ≤ c1,∆(p,t)′K(p, t). (cid:3) Remark 3.5. We compare Proposition 3.4 applied to the traditional John- Nirenberg BM O space to some estimates proved in [19, Theorem 1.19], which is one of the few articles we know of that track the constants for such inequalities. The authors in [19] proved that kf kBM Ow ≤ c[w]′ A∞ kf kBM O for some dimensional constant c > 0, where [w]′ A∞ = sup Q∈Q 1 w(Q) Q M(χQw)(x)dx and M is the standard Hardy-Littlewood maximal operator. They showed that this constant is sharp in terms of the power on the weight character, A∞ )ǫ in place of in the sense that one cannot obtain this estimate with ([w]′ [w]′ A∞ for any 0 < ǫ < 1 (in fact, they showed something slightly better). In terms of our notation, this says that b1,w ≤ c[w]′ A∞ . In this setting, we apply 2n+1t−1 and K(p, t) = 2 as in Remark Proposition 3.4 with ∆(p, t) = 1 + 2.8, which provides the estimate Ψp(t) . t for all 1 < p < ∞ and t ≥ 1, since it is known that c1,p . p for all 1 < p < ∞. Then for any w ∈ Ap, it 1 14 JAROD HART AND RODOLFO H. TORRES follows that b1,w ≤ Ψp([w]Ap) . [w]Ap. We fail to recover the A∞ constant of [19, Theorem 1.19], but we do recover the linear dependence on the power of the weight constant. Since we have not considered any A∞ constants in this work, aside from the current remark, this is the best possible result for Proposition 3.4. We will see in Section 5 other applications of Proposition 3.4 in other settings, where the results are new and obtain the same linear dependence on the Ap weight character. Remark 3.6. The somehow artificially imposed assumption X(B, Λ) = X p(B, Λ) for 0 < p < 1 in the hypotheses of the above results can be eliminated in some situations. One of them is when F = X(B, Λ) in The- orem 3.2 or F = Xw0(B, Λ) in Theorem 3.3. This immediately forces X(B, Λ) = X p(B, Λ), respectively Xw0(B, Λ) = X p w0(B, Λ), for 0 < p ≤ 1. Indeed, for any w, by definition X p w(B, Λ) always holds in the range 0 < p ≤ 1. w(B, Λ) ⊂ F while Xw(B, Λ) ⊂ X p Alternatively, Xw(B, Λ) = X p w(B, Λ) for 0 < p ≤ 1 holds true if more w is a quasi- structure on Xw is assumed. More precisely, suppose that k · kX p norm, X p w(B, Λ) endowed with it is a quasi-Banach space for all 0 < p ≤ 1, and Xw(B, Λ) is dense in X p w also for all 0 < p < 1. If this is the case and Xw(B, Λ) satisfies the p-invariance property for (0, 1], then clearly again Xw(B, Λ) = X p w(B, Λ) with respect to k · kX p w(B, Λ) for 0 < p ≤ 1. Finally, in the classical BM O context, very general conditions under which 1 Q Q sup Q h(f (x) − fQ) dx < ∞ for an appropriate function h implies f ∈ BM O were given by Stromberg [35], Lo and Ruilin [25], and Shi and Torchinsky [32]. See also the more recent work of Logunov et al in [26]. Hence, if X = BM O and Λ(f, Q)p = h((f (x)−fQ) for a certain such appropriate function, then one can conclude that X p = X for 0 < p < 1. We will adapt some of these works to several applications we present in Section 5. 4. Necessary Conditions for Weight Invariance In this section, we provide a partial converse to the results from the previous one. Although in applications we will obtain weight invariance from known p-invariance results, it is natural to ask whether the two concepts are actually equivalent. We indeed show that, essentially, if we have weight invariance estimates depending only on the norm of the weights, then p- power John-Nirenberg properties also hold. Assumptions A3 and A4 and the function Ψp defined in (3) play a pivotal role. They allow us to perform several computations to estimate kf kX p, and additional estimates for Ψp impose the right control in terms of the weights. In the end, we will be able to estimate c1,p using Ψp as stated in equation (4) of Theorem 4.2 below, which when combined with Proposition 3.4 provides a precise quantitative way to JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 15 associate the p-invariance and weight invariance through the constants c1,p and Ψp(t). The following lemma is likely known in many settings. We include the computations just to show that it does not depend on any particular prop- erty of the measure, the family of sets B, or its associated maximal function. Lemma 4.1. If u ∈ Ap(B) ∩ RHδ(B) for some 1 < p, δ < ∞, then uδ ∈ Aq(B) and [uδ]Aq ≤ [u]δ Ap where q = 1 + δ(p − 1). [u]δ RHδ Proof. Let u ∈ Ap(B) ∩ RHδ(B) for some 1 < δ, p < ∞, and define q = q = p′ 1 + δ(p − 1). Note that with this selection we have q′ = p′ δp . Then it follows that δp + 1, and q′ (cid:18) 1 µ(B) B RHδ(cid:18) 1 u(x)δdµ(x)(cid:19)(cid:18) 1 µ(B) B ≤ [u]δ u(x) dµ(x)(cid:19)δ(cid:18) 1 µ(B) B ≤ [u]δ RHδ [u]δ Ap u(x)−δq′/qdµ(x)(cid:19)q/q′ µ(B) B u(x)−p′/pdµ(x)(cid:19)δp/p′ for all B ∈ B. Therefore uδ ∈ Aq(B) with [uδ]Aq ≤ [u]δ RHδ [u]δ Ap. (cid:3) The next theorem shows that if X(B, Λ) is weight invariant for the class W = Ap with constants controlling the "norm" equivalences depending only on the characteristic of the weights, then X(B, Λ) is also p-invariant. More precisely, Theorem 4.2. Suppose µ, B, and MB satisfy the assumptions A1–A4. Assume there exists 1 < p0 < ∞ such that X(B, Λ) satisfies the weight invariance property for Ap0(B) and that Ψp(2kMBkLp,Lp) is finite for all 1 < p < p0. Then X(B, Λ) satisfies the p-power John-Nirenberg property for the interval (0, ∞). Moreover, it also holds that (4) c1,p ≤ 2Ψp′(2kMBkLp′ ,Lp′ ) < ∞ for all 1 < p < ∞. Proof. We will first use a bootstrapping argument to prove that c1,p < ∞ without proving the estimate asserted in (4), as the constants in these initial arguments are difficult to track. Once we know that c1,p is finite for all 1 < p < ∞, we can revisit and streamline the argument to obtain the estimate on (4). It should be noted that we cannot use the streamlined proof directly since it requires the a priori knowledge that c1,p < ∞. Assume X(B, Λ) satisfies the weight invariance property for Ap0(B) for some 1 < p0 < ∞. Choose 1 < s < p0 small enough so that 2kMBkLs,Ls ≥ (1 + K(p0, 2kMBkLp0 ,Lp0 ) 2kMBkLp0 ,Lp0 )p0 , which is possible by assumption A3. Also fix two more parameters 1 < r < s and 1 < δ < min(cid:16) s−1 r−1 , p0, ∆ (p0, 2kMBkLp0 ,Lp0 )(cid:17). 16 JAROD HART AND RODOLFO H. TORRES Define the Lp0-adapted Rubio de Francia algorithm R(p0)g(x) = Mk Bg(x) (2kMBkLp0 ,Lp0 )k . ∞Xk=0 Here M0 a function g. By assumption A3, we have kR(p0)kLp0 ,Lp0 ≤ 2. Bg(x) = g(x) and Mk Bg is the k-fold iterated application of MB to Fix f ∈ X(B, Λ). Let B ∈ B, and u(x) = R(p0)[Λ(f, B)1/p0 · χB](x). Note that Λ(f, B) is measurable and non-negative and B Λ(f, B)(x)dµ(x) ≤ µ(B)kf kX < ∞, and hence Λ(f, B)1/p0 ·χB is an Lp0(µ) function. Therefore R(p0)[Λ(f, B)1/p0 · χB] is well-defined as an Lp0(µ) function. Since MBu(x) ≤ 2kMBkLp0 ,Lp0 u(x), it follows by A1 that u ∈ A1(B) ⊂ Ar(B) ⊂ Ap0(B) with [u]Ap0 ≤ [u]Ar ≤ [u]A1 ≤ 2kMBkLp0 ,Lp0 . By assumption A4, [u]Ap0 ≤ 2kMBkLp0 ,Lp0 implies that u ∈ RH∆(p0,2kMBkLp0 ,Lp0 ) with [u]RH∆(p0,2kMB k Lp0 ,Lp0 ) ≤ K(p0, 2kMBkLp0 ,Lp0 ). With 1 < δ < min( s−1 r−1 , p0, ∆(p0, 2kMBkLp0 ,Lp0 )) as specified above, define v = uδ. Using Lemma 4.1 and the fact that our parameter selection implies s > 1 + δ(r − 1), it follows that v ∈ As(B) ⊂ Ap0(B) and [v]As ≤ ([u]RHδ [u]Ar )δ ≤ (K(p0, 2kMBkLp0 ,Lp0 ) 2kMBkLp0 ,Lp0 )δ ≤ 2kMBkLs,Ls. Then (cid:18) 1 µ(B) B 1+δ/p0 1+1/p0 Λ(f, B)(x)1+1/p0 dµ(x)(cid:19) 1 ≤(cid:18) 1 µ(B) B =(cid:18) 1 µ(B) B ≤(cid:18) 1 µ(B) B ≤(cid:18) v(B) Λ(f, B)(x)1+δ/p0 dµ(x)(cid:19) 1 Λ(f, B)(x)(cid:16)Λ(f, B)(x) p0(cid:17)δ Λ(f, B)(x)v(x)dµ(x)(cid:19) 1 kf kXv(Rn)(cid:19) 1 µ(B) 1+δ/p0 . 1+δ/p0 1 χB(x)dµ(x)(cid:19) 1 1+δ/p0 JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 17 Since δ < p0, it also follows that v(B) µ(B) 1 = µ(B) B(cid:16)R(p0)[Λ(f, B)1/p0 · χB](x)(cid:17)δ ≤(cid:18) 1 µ(B) B(cid:16)R(p0)[Λ(f, B)1/p0 · χB](x)(cid:17)p0 ≤ 2δ(cid:18) 1 µ(B) B ≤ 2δkf kδ/p0 X . Λ(f, B)(x)dµ(x)(cid:19)δ/p0 dµ(x) dµ(x)(cid:19)δ/p0 It follows that (cid:18) 1 µ(B) B Λ(f, B)(x)1+δ/p0 dµ(x)(cid:19) 1 1+δ/p0 ≤(cid:18) v(B) µ(B) kf kXv(cid:19) 1 1+δ/p0 δ/p0 1 kf k 1+δ/p0 Xv 1+δ/p0 1+δ/p0 kf k X 1 δ δ 1+δ/p0 b 1+δ/p0 1,v kf kX ≤ 2 ≤ 2 ≤ 2 δ 1+δ/p0 Ψs(2kMBkLs,Ls) 1 1+δ/p0 kf kX. Therefore X(B, Λ) satisfies the p-power John-Nirenberg property on the in- terval [1, 1 + 1/p0] and hence, by Lemma 3.1, it does so on (0, 1 + 1/p0] too. We will now bootstrap this argument to show that X(B, Λ) satisfies the p-power John-Nirenberg inequality for the interval (0, ∞). We will do so by induction. Write 1 + 1/p0 = (4p0 + 4)/4p0 and assume that X(B, Λ) satisfies the p- ] for some integer power John-Nirenberg property on the interval (0, 4p0+ℓ−1 ℓ ≥ 5, and define the Lℓ′ -adapted Rubio de Francia algorithm 4p0 R(ℓ′)g(x) = Mk Bg(x) (2kMBkLℓ′ ,Lℓ′ )k . ∞Xk=0 Here, ℓ′ = ℓ 1 < s < ℓ′ small enough so that ℓ−1 is the Holder conjugate of ℓ and now kR(ℓ′)kLℓ′ ,Lℓ′ ≤ 2. Fix 2kMBkLs,Ls ≥(cid:16)1 + K(ℓ′, 2kMBkLℓ′ ,Lℓ′ ) 2kMBkLℓ′ ,Lℓ′(cid:17)ℓ′ , and, like in the initial step, fix 1 < r < s and 1 < δℓ < min(cid:18) s − 1 r − 1 , ℓ′, ∆(ℓ′, 2kMBkLℓ′ ,Lℓ′ )(cid:19) . Fix f ∈ X(B, Λ) and B ∈ B. Let uℓ(x) = R(ℓ′)[Λ(f, B) 4p0+ℓ−1 4p0 ℓ′ · χB](x). 18 JAROD HART AND RODOLFO H. TORRES Note that Λ(f, B) 4p0+ℓ−1 4p0 ℓ′ · χB ∈ Lℓ′ (Rn) since, by the inductive hypothesis, kf k X 4p0 +ℓ−1 4p0 ≤ c 1, 4p0+ℓ−1 4p0 kf kX < ∞. It follows that with uℓ ∈ A1(B) ⊂ Ar(B) ⊂ Aℓ′(B), [uℓ]Aℓ′ ≤ [uℓ]Ar ≤ [uℓ]A1 ≤ 2kMBkLℓ′ ,Lℓ′ , and hence uℓ ∈ RH∆(ℓ′,2kMBk Lℓ′ [uℓ]RH∆(ℓ′,2kMB k ,Lℓ′ )(B), with ,Lℓ′ ) ≤ K(ℓ′, 2kMBkLℓ′ ,Lℓ′ ). Lℓ′ Define vℓ = uδℓ that vℓ ∈ As(B) with ℓ , and using the same arguments as before, Lemma 4.1 implies [vℓ]As ≤ ([uℓ]RHδℓ [uℓ]Ar )δℓ ≤(cid:16)2K(ℓ′, 2kMBkLℓ′ ,Lℓ′ ) kMBkLℓ′ ,Lℓ′(cid:17)δℓ ≤ 2kMBkLs,Ls. Using that ℓ2 ≤ (ℓ − 1)(4 + ℓ − 1) and δℓ, p0 > 1, it follows that 4p0 + ℓ 4p0 ≤ 1 + δℓ(4p0 + ℓ − 1) 4p0ℓ′ . Then it also follows that µ(B) B (cid:18) 1 µ(B) B ≤(cid:18) 1 = 1 ≤(cid:18) 1 µ(B) B ≤(cid:18) vℓ(B) µ(B) B µ(B) vℓ(B) µ(B) Λ(f, B)(x) 4p0 +ℓ 4p0 dµ(x)(cid:19) 4p0 4p0+ℓ Λ(f, B)(x) 1+ δℓ (4p0+ℓ−1) 4p0ℓ′ 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) Λ(f, B)(x)(cid:18)Λ(f, B)(x) Λ(f, B)(x)vℓ(x)dµ(x)(cid:19) kf kXvℓ(cid:19) µ(B) B ≤(cid:18) 1 µ(B) B R(ℓ′)[Λ(f, B) R(ℓ′)[Λ(f, B) ≤ 1 . 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) dµ(x)(cid:19) 4p0ℓ′ (cid:19)δℓ 4p0 +ℓ−1 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) dµ(x)! 4p0 +ℓ−1 4p0ℓ′ · χB](x)δℓ dµ(x) 4p0+ℓ−1 4p0ℓ′ · χB](x)ℓ′ ℓ′ dµ(x)(cid:19) δℓ Since 1 < δℓ < ℓ′, we also have that JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 19 Λ(f, B)(x) 4p0 +ℓ−1 4p0 dµ(x)(cid:19) δℓ ℓ′ δℓ(4p0+ℓ−1) 4p0 ℓ′ , . kf k X ≤ 2δℓ(cid:18) 1 µ(B) B ≤ 2δℓkf k δℓ (4p0+ℓ−1) 4p0ℓ′ 4p0+ℓ−1 4p0 X 4p0+ℓ−1 4p0 X where we use that kf k . kf kX by the inductive hypothesis. Com- bining the above computations we obtain (cid:18) 1 µ(B) B . kf k Λ(f, B)(x) δℓ(4p0+ℓ−1) 4p0ℓ′ X 4p0+ℓ 4p0 +ℓ 4p0 dµ(x)(cid:19) 4p0 kf kXvℓ! 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) 1,vℓ . b kf kX . Ψs(2kMBkLs,Ls) 4p0ℓ′ 4p0ℓ′+δℓ(4p0+ℓ−1) kf kX. ] for X(B, Λ) Therefore the p-power John-Nirenberg property on (0, 4p0+ℓ−1 4p0 implies the p-power John-Nirenberg property on (0, 4p0+ℓ ] for X(B, Λ). By 4p0 induction, X(B, Λ) satisfies the p-power John-Nirenberg property for (0, ∞). Now that we know c1,p < ∞ for all 1 < p < ∞, we can obtain the better estimate given in (4). In fact, by Proposition 3.4 the assumption on Ψp in the theorem hypotheses improves now to Ψp(2kMBkLp,Lp) < ∞, for all 1 < p < ∞. Suppose f ∈ X(B, Λ), and let B ∈ B. Fix p, 1 < p < ∞, and define u(x) = R(p′)[Λ(f, B)p−1 · χB](x), which is well defined since c1,p < ∞ implies Λ(f, B)p−1χB ∈ Lp′ -adapted Rubio de Francia algorithm, similar to before. It follows that u ∈ A1(B) ⊂ Ap′(B) with [u]Ap′ ≤ [u]A1 ≤ 2kMBkLp′ ,Lp′ . Then we have µ(B) B Λ(f, B)(x)pdµ(x) ≤ Λ(f, B)(x)u(x)dµ(x) µ(B) B (µ). Here Rp′ is the Lp′ 1 1 ≤ u(B) µ(B) kf kXu ≤ b1,u u(B) µ(B) kf kX ≤ Ψp′(2kMBkLp′ ,Lp′ ) u(B) µ(B) kf kX since [u]Ap′ ≤ 2kMBkLp′ ,Lp′ . We note that u(B) µ(B) Then we have ≤(cid:18) 1 µ(B) B ≤ 2(cid:18) 1 µ(B) B R(p′)[Λ(f, B)p−1 · χB](x)p′ Λ(f, B)pdx(cid:19) 1 p′ ≤ 2kf kp−1 X p . p′ dx(cid:19) 1 kf kp X p ≤ 2Ψp′(2kMBkLp′ ,Lp′ )kf kp−1 X p kf kX . 20 JAROD HART AND RODOLFO H. TORRES Rearranging terms, we obtain kf kX p ≤ 2Ψp′(2kMBkLp′ ,Lp′ )kf kX for all f ∈ X(B, Λ). Therefore c1,p ≤ 2 Ψp′(2kMBkLp′ ,Lp′ ). (cid:3) 5. Applications We present several applications including the theorems stated in the in- troduction. We repeat the statement of those theorems for the readers convenience and because we present some of them in a more general form. Since we will consider many different measures in several different settings, we find it convenient to change for this section some of the the notation involving measures of sets and the corresponding averages of functions that we have been using so far. For example, when the underlying measure µ used to define an oscillation space X(B, µ, F, Λ) is not the Lebesgue measure in Rn and w is a weight with respect to µ, we will now write µw(B) = B w(x)dµ(x), instead of w(B). However we will keep the latter notation in the Lebesgue setting. The precise meaning of other quantities will be consistent within each of the subsections and will be specified therein. In several of the results we will present, we will verify the condition X p = X for 0 < p < 1 that appears in Theorems 3.2 and 3.3. We were not able to find in the literature a proof of this property for each situation, but we can adapt the techniques from [25] for our applications (our arguments are similar to the proof of [25, Proposition 2] except that we use the p- power property in place of the traditional John-Nirenberg exponential one). Rather than rewriting the same argument for every application, we present the argument once in Lemma 5.1 in the terminology of our X(B, Λ) spaces. We should also note that Lemma 5.1 is a version of Lemma 3.1 with more structure imposed on Λ and X, and hence we are able to conclude something stronger that addresses the technicalities arising in the X p = X conditions. Lemma 5.1. Let X(B, µ, F, Λ) be so that F ⊂ L1 loc(µ) and Λ(f, B)(x) = f (x) − fB, where fB denotes the average of f over B with respect to µ. Then the following properties hold. a) For all for all f ∈ F (5) kf kX ≈ sup B∈B inf c∈C 1 µ(B) B f (x) − cdµ(x). b) If in addition X(B, Λ) satisfies the p-invariance property for [1, p0) for some 1 < p0 < ∞, then X p(B, Λ) = X(B, Λ) for all 0 < p < 1 and X satisfies the p-invariance property for (0, p0) Proof. The proof of part a) is well-known. Simply note that for any complex number c Λ(f, B)(x) = f (x) − fB ≤ f (x) − c + fB − c. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 21 Taken then the average over B, followed by the infimum in c, we easily obtain kf kX ≤ 2 sup B∈B inf c∈C 1 µ(B) B f (x) − cdµ(x). The reverse inequality is of course trivial. Our first step to prove part b) is to show that f ∈ X p implies Λ(f, B0)p ∈ X for all B0 ∈ B and 0 < p < 1. So fix 0 < p < 1, f ∈ X p, and B0 ∈ B. For every B ∈ B, and since 0 < p < 1, we have B Λ(f, B0)(x)p − fB − fB0p dµ(x) ≤ B f (x) − fBpdµ(x) ≤ µ(B)kf kp X p. Hence from (5) it follows that Λ(f, B0)p ∈ X and that kΛ(f, B0)pkX . kf kp X p. From here we use a bootstrapping argument to show X p ⊂ X for all 0 < p < 1, ranging all the way down to 0. More precisely, we prove (by induction) that X 1/pk B0 Λ(f, B0)(x)dµ(x) ≤ 2p0B0(cid:12)(cid:12)(cid:12)(cid:12)Λ(f, B0)(x)1/p0 −(cid:16)Λ(f, B0)1/p0(cid:17)B0(cid:12)(cid:12)(cid:12)(cid:12) 0 ⊂ X for all k ∈ N. For f ∈ X 1/p0 , we have dµ(x) p0 Λ(Λ(f, B0)(x)1/p0 , B0)(x)p0 dµ(x) B0 = 2p0 B0 + 2p0µ(B0)(cid:16)Λ(f, B0)1/p0(cid:17)p0 + 2p0µ(B0)(cid:18) 1 µ(B0) B0 ≤ 2p0µ(B0)kΛ(f, B0)1/p0kp0 . cp0 1,p0µ(B0)kf kX 1/p0 . Λ(f, B0)(x)1/p0 dµ(x)(cid:19)p0 X p0 + 2p0µ(B0)kf kX 1/p0 Note here that c1,p0 < ∞ by assumption, and we used that Λ(f, B0)1/p0 ∈ X implies kΛ(f, B0)1/p0 kX p0 ≤ c1,p0kΛ(f, B0)1/p0kX . c1,p0kf k1/p0 X 1/p0 . Therefore X 1/p0 ⊂ X. Now assume that X 1/pk Then for f ∈ X 1/pk+1 0 , by a similar argument to the k = 1 case, we have 0 ⊂ X holds for a given k ≥ 1. B0 Λ(f, B0)(x)1/pk 0 dµ(x) ≤ 2p0 B0(cid:12)(cid:12)(cid:12)(cid:12)Λ(f, B0)(x)1/pk+1 0 (cid:17)B0(cid:12)(cid:12)(cid:12)(cid:12) 0 −(cid:16)Λ(f, B0)1/pk+1 0 (cid:17)p0 + 2p0µ(B0)(cid:16)Λ(f, B0)1/pk+1 ≤ 2p0µ(B0)kΛ(f, B)1/pk+1 0 kp0 X p0 + 2p0µ(B0)kf k B0 1/pk 0 X 1/pk+1 0 p0 dµ(x) 22 JAROD HART AND RODOLFO H. TORRES 1,p0kf k . µ(B0)cp0 1/pk 0 X 1/pk+1 By induction, it follows that X 1/pk+1 p0 > 1 so that 1/pk for all 0 < p < 1. 0 . 0 ⊂ X 1/pk 0 ⊂ X for all k ∈ N. Since 0 → 0 as k → ∞, this is sufficient to prove that X p ⊂ X (cid:3) Lemma 5.1 will play a crucial role in the first three applications we shall In them the underlying metric spaces are, respectively, Rn with present. the Lebesgue measure, an abstract space of homogeneous type, and Rn with a non-necessarily doubling measure. In all three cases we will prove that the natural BM O space of each context can be characterized also by an appropriate oscillation space defined using a pair of A∞ weights. Of course, the case of Rn with the Lebesgue measure (and the Euclidean distance) is a particular case of the other two applications, but we chose to present a proof in this context to show how it relates to results in [29, 35, 25, 32] and for clarity in the exposition. The proofs for all of the first three applications are also almost identical. Therefore, for spaces of homogeneous type and for non- doubling measures on Rn, we provide details about some needed estimates, but leave the computations completely analogous to the classical case to the reader. The proof of Theorem 5.14 also uses very similar arguments, and hence we omit many of the details as well. 5.1. Weight invariance for the John-Nirenberg BMO space. The following extension of the weight invariant result in [29] holds. Theorem 5.2. Let 0 < p < ∞, w, v ∈ A∞, and f be a complex-valued function on Rn. Then f is in L1 loc(v) and satisfies sup w(Q) Q Q∈Q(cid:18) 1 v(Q) ´Q f (x)v(x)dx and w(Q) = ´Q w(x)dx, if and only if f is f (x) − fv,Qpw(x)dx(cid:19) 1 < ∞, p where fv,Q = 1 in BM O. In such a case (6) kf kBM O ≈ sup Q∈Q(cid:18) 1 w(Q) Q f (x) − fv,Qpw(x)dx(cid:19) 1 p . Proof. Set B = Q and µ to be the Lebesgue measure. For v ∈ A∞, we also set the notation Fv = L1 loc(v), Λv(f, Q)(x) = f (x)−fv,Q, and X(B, Fv, µ, Λv) = X(B, Λv). With this notation, X(B, F1, µ, Λ1) = X(B, Λ1) = BM O by definition and kf kX p w(B,Λv) is the term appearing on the right hand side of (6). So it is sufficient to show that X p w(B, Λv) ≈ X(B, Λ1) for all 0 < p < ∞ and w, v ∈ A∞. Muckenhoupt and Wheeden proved in [29] the John-Nirenberg inequality for Xv(B, Λv). More precisely, there exist constants c1, c2 > 0 (depending only on the dimension n) such that v({x ∈ Q : f (x) − fv,Q > λ}) ≤ c1v(Q)e − c2λ kf kXv(B,Λv ) , JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 23 for all f ∈ Xv(B, Λv), Q ∈ B, and λ > 0. It follows that Xv(B, Λv) satisfies the p-invariance property for [1, ∞) for all v ∈ A∞. We complete this proof in three steps for any w, v ∈ A∞: w(B, Λv) ≈ X q 1. X p 2. Xv(B, Λv) ≈ X(B, Λ1); 3. X p w(B, Λv) ≈ Xv(B, Λv) for all 0 < p < 1 v (B, Λv) for all 1 ≤ p < ∞ and 0 < q < ∞; Step 1. Since Xv(B, Λv) satisfies the p-power John-Nirenberg property for [1, ∞) and any v ∈ A∞, then it also satisfies by Lemma 5.1 the p-invariance property on (0, ∞) and Xv(B, Λv) ≈ X p v (B, Λv) for all such p. Applying Theorem 3.3 completes the proof statement 1. Note that Xw(B, Λv) satisfies the p-invariant property for all of (0, ∞), but we cannot yet conclude that the X p w(B, Λv) coincide for 0 < p < 1 (see Remark 2.6 for a discussion of this subtle point). Note also, that we cannot apply Lemma 5.1 to X p w(B, Λv) since the weight w does not match the weight on the oscillation functional Λv. Step 2. For any f ∈ Xv(B, Λv), it follows from what we proved in Step 1 that f ∈ L1 loc (since f ∈ X(B, Λv)). Furthermore, we have kf kX(B,Λ1) ≤ 2 sup Q inf c∈C 1 Q Q f (x) − cdx ≤ 2kf kX(B,Λv ) ≈ kf kXv(B,Λv), where the last estimate also follows from Step 1 with p = q = 1 and w = 1. Similarly, f ∈ X(B, Λ1) implies f ∈ L1 loc(v) and kf kXv(B,Λv) ≤ 2kf kXv(B,Λ1) ≈ kf kX(B,Λ1). So statement 2 holds as well. Step 3. By what we proved in Step 1 with p = q = 1, and using Holder's inequality, it follows that Xv(B, Λv) ≈ Xw(B, Λv) ⊂ X p w(B, Λv) loc(v). Con- holds for 0 < p < 1, with kf kX p versely, for f ∈ X p qδ′ where 1 < q, δ < ∞ are such that w ∈ Aq and v ∈ RHδ. Using what is by now a familiar argument (see the proof of Theorem 3.3), it follows that w(B, Λv) with 0 < p ≤ 1, set ǫ = p w(B,Λv) . kf kXv(B,Λv) for f ∈ L1 1 v(Q) Q f (x) − fv,Qǫv(x)dx ≤ [v]RHδ [w]1/(qδ′ ) Aq (cid:18) 1 w(Q) Q f (x) − fv,Qpw(x)dx(cid:19)1/(δ′q) . w(B, Λv) ⊂ X ǫ Hence X p L1 loc(v). Using Step 1 again, it also follows that X ǫ Therefore Step 3, and hence of the proof of Theorem 5.2, is complete. w(B,Λv) for all f ∈ v(B, Λv) ≈ Xv(B, Λv). (cid:3) v(B, Λv) with kf kX ǫ v (B,Λv) . kf kX p 24 JAROD HART AND RODOLFO H. TORRES Remark 5.3. A particular consequence of the estimate in Step 1 is that X p v (B, Λv) ≈ Xv(B, Λv) for 0 < p < 1, which we proved by applying Lemma 5.1, but this was already known in several situations; see for example [35, 25, 32]. In the classical John-Nirenberg BM O setting, the statement in Step 2 is exactly the weight invariance of BM O proved by Muckenhoupt and Wheeden in [29]. We are not aware of any result in the form of the estimate in Step 3 before this article. Notice that the proofs of Steps 1-3 in Theorem 5.2 depend only on the following facts: A2 holds, Xv(B, Λv) satisfies the p-power John-Nirenberg property for [1, ∞) for all v ∈ A∞, and that the oscillation functional is of the form f (x) − fµ,Q for some measure µ. These are the properties we will verify to apply the same scheme of proof in the applications in the next two subsections. Remark 5.4. A particular case of Theorem 5.2 is of course (7) kf kBM O ≈ sup Q∈Q(cid:18) 1 w(Q) Q f (x) − fQpw(x)dx(cid:19) 1 p . Using a notion of p-convexity for 1 ≤ p < ∞, Ho [16, Theorem 3.1] proved (7) when w ∈ Ap. Note that our theorem allows us to use f (x) − fv,Q in place of f (x) − fQ for any v ∈ A∞. Moreover, our methods allows us to prove (7) in the context of spaces of homogeneous type, recovering in the next subsection a particular case of a result of Franchi, P´erez and Wheeden [9, Theorem 3.1], but allowing again the use of two weights to characterize BM O. 5.2. Weighted BMO on spaces of homogeneous type. Let (S, d, µ) be a space of homogeneous type in the sense of Coifman and Weiss. Let B be the collection of all d-balls B in S, of the form B = Bd(x, r) = {y ∈ S : d(x, y) < r}. Define for w ∈ A∞, BM Ow(S, µ) := {f ∈ L1 loc(µw) : kf kBM Oµw (S) < ∞}, kf kBM Oµw (S) := sup B∈B f (x) − fµw,Bw(x)dµ(x) fµw,B = f (x)w(x)dµ(x). When w ≡ 1 we obtain the classical John-Nirenberg space on a space of homogenous type and we simply write BM Oµ(S) in stead of BM Oµ1(S). Likewise we will use the notation fµ,B = fµ1,B. where with and 1 µw(B) B µw(B) = B w(B) B 1 w(x)dµ(x), JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 25 Theorem 5.5. Let 0 < p < ∞, v, w ∈ A∞(B, µ), and f be a measurable complex-valued function on S. Then f is in L1 loc(µv) and satisfies f (x) − fµv,Bpdµw(x)(cid:19)1/p < ∞ B∈B(cid:18) 1 sup µw(B) B B∈B(cid:18) 1 if and only if f is in BM Oµ(S). In such a case, kf kBM Oµ(S) ≈ sup µw(B) B f (x) − fµv,Bpdµw(x)(cid:19)1/p . Proof. We first observe that A2 holds in this situation (even sharp forms of the reverse Holder inequality are known, as discussed in Remark 2.8). Note also that if (S, d, µ) is a space of homogeneous type and w ∈ A∞(B, µ), it follows that (S, d, µw) is also a space of homogeneous type. It is well-known that the John-Nirenberg inequality holds for spaces of homogeneous type. This fact has been reproved many times in the literature, but can actually be traced back to the the original work of Coifman and Weiss [6, p. 694, fn. 22], as observed in [25], where the details are also presented. Hence, there exist constants c1, c2 > 0 such that µw({x ∈ B : f (x) − fµw,B > λ}) ≤ c1µw(B)e−c2λ/kf kBM Oµw for all f ∈ BM Oµw (S). As explained before, this makes up all of the ingredients necessary to reproduce the proof of Theorem 5.2 in the setting of spaces of homogeneous type. The details are left to the reader. (cid:3) Remark 5.6. Recall from Remark 2.8 that if we take ∆(p, t) = 1 + 1 τ t and K(p, t) = C, then the classes Ap(B, µ), when B and m are as in Theorem 5.5, satisfy A4. Then applying Proposition 3.4, it follows that Ψp(t) . c1,1+τ t . t, and hence that b1,w ≤ Ψp([w]Ap ) . [w]Ap for any w ∈ Ap with 1 < p < ∞. In particular, sup B∈B 1 µw(B) B f (x) − fµ,Bw(x)dµ(x) . [w]Apkf kBM Oµ(S) for all f ∈ BM Oµ(S), 1 < p < ∞, and w ∈ Ap(B, µ), where the suppressed constant does not depend on f , p, or w. 5.3. Weighted BMO with respect to non-doubling measures. Let µ be a non-negative Radon measure on Rn (not necessarily doubling). Define BM Oµ(Rn) := {f ∈ L1 loc(Rn, µ) : kf kBM Oµ(Rn) < ∞}, where and kf kBM Oµ(Rn) := sup Q∈Q 1 µ(Q) Q f (x) − fµ,Qdµ(x) fµ,Q = 1 µ(Q) Q f (x)dµ(x). 26 JAROD HART AND RODOLFO H. TORRES Also, for any w ∈ A∞(Q, µ) let µw(Q) = Q w(x)dµ(x). Theorem 5.7. Let µ be a non-negative Radon measure on Rn such that µ(L) = 0 for any hyperplane L orthogonal to one of the coordinate axes. Let 0 < p < ∞, v, w ∈ A∞(Q, µ), and f be a measurable complex-valued function on Rn. Then f is in L1 loc(µv) and satisfies Q∈Q(cid:18) 1 sup µw(Q) Q f (x) − fµv,Qpdµw(x)(cid:19)1/p < ∞, if and only if f is in BM Oµ(Rn). In such a case, (8) kf kBM Oµ ≈ sup Q∈Q(cid:18) 1 µw(Q) Q f (x) − fµv,Qpdµw(x)(cid:19)1/p . Proof. We note that A2 holds in this situation as it was proved in [31, Lemma 2.3]. Next, it was shown in [28, Theorem 1] that there exist constants c1, c2 > 0 (depending only on the dimension n) such that (9) µ({x ∈ Q : f (x) − fµ,Q > λ}) ≤ c1µ(Q)e − c2λ kf kBM Oµ for all Q ∈ B and λ > 0. Finally, note also that for any w ∈ A∞(µ), µw is absolutely continuous with respect to µ. So in particular, µw(L) = 0 for any hyperplane L orthogonal to one of the coordinate axes. Hence it follows that (9) hold with µw in place of µ. Once again, this verifies everything needed to reproduce the proof of Theorem 5.2. (cid:3) Remark 5.8. Recall from Remark 2.8 that if we take ∆(p, t) = 1+ 2p+1B(n)t and K(p, t) = 2, then the classes Ap(µ), when µ is as in Theorem 5.7, satisfy A4. Then applying Proposition 3.4, it follows that 1 Ψp(t) ≤ 2c1,1+2p+1B(n)t . 2pB(n)t for 1 < p < ∞ and t ≥ 1. Note that the linear estimate c1,p . p holds for 1 < p < ∞ as a consequence of the John-Nirenberg inequality proved in [28]. So for a particular w ∈ Ap(µ), we obtain b1,w ≤ Ψp([w]Ap) . 2p[w]Ap. Here we suppress the dependence on B(n) since it is a dimensional constant. In particular, sup Q∈Q 1 w(Q) Q f (x) − fQw(x)dx . 2p[w]Apkf kBM Oµ for any f ∈ BM Oµ, 1 < p < ∞, and w ∈ Ap(µ), where the suppressed constant does not depend on f , p, or w. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 27 5.4. Duals of weighted Hardy spaces. Let w be an A∞ weight in Rn with respect to the Lebesgue measure. It is known that the dual of the Hardy space H 1(w) (we will not need the definition of H 1(w) here) is the space BM O∗,w := {f ∈ L1 loc(Rn) : kf kBM O∗,w < ∞}, taken modulo constants, where kf kBM O∗,w := sup Q∈Q 1 w(Q) Q f (x) − fQdx and fQ = 1 Q ´Q f (x)dx. See [11, Theorem II.4.4] for more information on H 1(w) and a proof that its dual is BM O∗,w. Our general setup easily gives the following invariance result in this context. Theorem 5.9. Let w ∈ Ap for some 1 < p < ∞. If v ∈ RHp(w), then (10) kf kBM O∗,w ≈ sup Q∈Q 1 ρ(Q) Q f (x) − fQv(x)dx, for all f ∈ BM O∗,w, where ρ(Q) = ´Q v(x)w(x)dx. Furthermore, if v ∈ RHσ(w) for some σ > p, then for all 0 < r ≤ p′/σ′, we have kf kBM O∗,w ≈ sup Q∈Q(cid:18) 1 ρ(Q) Q f (x) − fQrw(x)1−rv(x)dx(cid:19)1/r for all f ∈ BM O∗,w. If w ∈ A1, v ∈ A∞(w dx), and 0 < r < ∞, then kf kBM O∗,w ≈ sup Q∈Q(cid:18) 1 ρ(Q) Q f (x) − fQrw(x)1−rv(x)dx(cid:19)1/r for all f ∈ BM O∗,w. Proof. First fix w ∈ Ap for some 1 < p < ∞. Define F = BM O∗,w and Λ(f, Q)(x) = f (x) − fQw(x)−1. If we let dµ(x) = w(x)dx, then it follows that X(Q, µ, Λ) = BM O∗,w since kf kX = sup Q∈Q 1 µ(Q) Q Λ(f, Q)(x)dµ(x) = sup Q∈Q 1 w(Q) Q f (x) − fQdx. It was proved independently by Muckenhoupt and Wheeden [29, Theorem 4] and Garcia-Cuerva [11, Theorem II.4.4] that Q∈Q(cid:18) 1 sup w(Q) Q f (x) − fQrw(x)1−rdx(cid:19)1/r . kf kBM O∗,w for all f ∈ BM O∗,w and 1 ≤ r ≤ p′. In other words, X(Q, µ, Λ) satisfies the John-Nirenberg p-power inequality for [1, p′]. Note that since we defined F = BM O∗,w, it follows immediately that X p(Q, µ, Λ) = X(Q, µ, Λ) for all 28 JAROD HART AND RODOLFO H. TORRES 0 < p < 1, and Theorem 3.2 can be applied to X(Q, µ, Λ). The first part of Theorem 3.2 implies that for any v ∈ RHp(w), kf kBM O∗,w = kf kX ≈ kf kXv = sup Q∈Q 1 ρ(Q) Q f (x) − fQv(x)dx, for all f ∈ BM O∗,w, where ρ(Q) = ´Q v(x)w(x)dx. Furthermore, if v ∈ RHσ(w) for some σ > p, then the second part of Theorem 3.2 implies that Xv(Q, Λ) is p-invariant for the interval (0, p′/σ′]. That is, kf kBM O∗,w = kf kX ≈ kf kX r v = sup Q∈Q(cid:18) 1 ρ(Q) Q f (x) − fQrw(x)1−rv(x)dx(cid:19) 1 r for all f ∈ BM O∗,w, 0 < r ≤ p′/σ′, and v ∈ RHσ(w) when σ > p. This completes the proof for w ∈ Ap with 1 < p < ∞. Now let w ∈ A1. Since it is known that A1 ⊂ Tp>1 Ap and A∞(w) = Sp>1 RHp(w), it follows from the estimates proved above that Q∈Q(cid:18) 1 ρ(Q) Q f (x) − fQrw(x)1−rv(x)dx(cid:19) 1 r kf kBM O∗,w ≈ sup for all f ∈ BM O∗,w, v ∈ A∞(w), and 0 < r < ∞. (cid:3) Remark 5.10. Unlike the situation of the previous applications, we are not providing a full characterization of the space in question in Theorem 5.9. We only show that (10) holds if f ∈ BM O∗,w. Though we suspect that if the righthand of (10) is finite then f must be in BM O∗,w, our methods do not seem to be able to establish that. Remark 5.11. We note that Theorem 5.9 can be extended to weighted BM O∗,w(S) spaces defined in the context of a space of homogeneous type (S, d, µ). Indeed, the only additional information needed to reproduce the proof of Theorem 5.9 is that the p-power John-Nirenberg estimate for the space BM O∗,w, proved in [29, 11], can be extended to BM O∗,w(S). A recent paper by Trong and Tung [36] does exactly this, and hence Theorem 5.9 can also be extended to a space of homogeneous type setting as an application of Theorem 3.2. Remark 5.12. The space BM O∗,w was used by Bloom in [3] to characterize the boundedness of commutators of the classical Hilbert transform between weighted Lebesgue spaces with different weights. In a recent work Holmes, Lacey, and Wick [18], extended Bloom's result and characterized the two- weight boundedness of commutators of Calder´on-Zygmund operators in Rn. In particular, [18, Theorem 1.1] shows the following: Let 1 < p < ∞, µ, λ ∈ Ap, ν = (µ · λ−1)1/p, b ∈ BM O∗,ν, and T be a Calder´on-Zygmund operator (see [18] for the precise condition on T ). Then the commutator [b, T ]f = bT (f ) − T (bf ) JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 29 is bounded from Lp(µ) into Lp(λ) and k[b, T ]kLp(µ),Lp(λ) . kbkBM O∗,ν . It is not hard to see that ν ∈ A2, which was proved in [18] as well. It is also not hard to verify that ν = µ1/pλ−1/p ∈ A2 implies ν−1 = µ−1/pλ1/p ∈ RH2(ν). Then by Theorem 5.9, it follows that (11) k[b, T ]kLp(µ),Lp(λ) . sup Q∈Q 1 Q Q b(x) − bQν(x)−1dx. Conversely, if ν ∈ RHδ for some 1 < δ < ∞ and b ∈ L1 loc satisfies (12) sup Q∈Q 1 Q Q b(x) − bQδ′ ν(x)−δ′ dx < ∞, then b ∈ BM O∗,ν, and [b, T ] is bounded from Lp(µ) into Lp(λ), and the estimate in (11) holds. To verify this, it is enough to check that Q b(x) − bQdx ≤(cid:18)Q b(x) − bQδ′ ν(x)−δ′ ≤ [ν]RHδ ν(Q)(cid:18) 1 Q Q dx(cid:19)1/δ′(cid:18)Q b(x) − bQδ′ ν(x)−δ′ ν(x)δdx(cid:19)1/δ dx(cid:19)1/δ′ , and then apply [18, Theorem 1.1] and Theorem 5.9. If one imposes more on µ and/or λ, other estimates can be obtained by applying Theorem 5.9. For example, letting p, µ, λ, ν, b, and T be as above, it follows that • if λ1/p ∈ RH2(ν), then k[b, T ]kLp(µ),Lp(λ) . sup Q∈Q 1 µ1/p(Q) Q b(x) − bQλ(x)1/pdx. • if λ ∈ RH2(ν), then k[b, T ]kLp(µ),Lp(λ) . sup Q∈Q • if µ−1/p ∈ RH2(ν), then 1 µ1/pλ1/p′(Q) Q b(x) − bQλ(x)dx. k[b, T ]kLp(µ),Lp(λ) . sup Q∈Q 1 λ−1/p(Q) Q b(x) − bQµ(x)−1/pdx. Furthermore, [18, Theorem 1.2] shows that when 1 < p < ∞, µ, λ ∈ Ap, ν = (µ · λ−1)1/p, and b ∈ BM O∗,ν, then k[b, Rj ]kLp(µ),Lp(λ) ≈ kbkBM O∗,ν , where Rj is the jth Riesz transform. Hence, under the same assumptions, it follows that (13) k[b, Rj ]kLp(µ),Lp(λ) ≈ sup Q∈Q 1 Q Q b(x) − bQν(x)−1dx. Similarly, in each of the situations discussed above where k[b, T ]kLp(µ),Lp(λ) is bounded by some maximal oscillation expression of b, one can obtain the 30 JAROD HART AND RODOLFO H. TORRES same lower estimate for the commutator operator norm when T = Rj is a Riesz transform. Finally, if ν ∈ RHδ for some 1 < δ < ∞ and b ∈ L1 loc satisfies (12), then [b, Rj ] is bounded and (13) holds. Similar to the discussion above, this can be proved by noting that (12) implies b ∈ BM O∗,ν and then apply [18, Theorem 1.2], followed by Theorem 5.9. 5.5. BMO spaces associated to operators. There is a very extensive literature about BM O-type spaces defined by expression of the form sup Q∈Q 1 Q Q f − StQf dx, where tQ is a parameter associated to the cube Q and {St}t > 0 is a semi- group, an approximation to the identity, or other appropriate family of op- erators. Moreover, further generalizations based on the approach in [9], with more general oscillating functional than f − StQf , and in weighted or different geometric context have been considered too. We refer to the works of Duong-Yang [7, 8], Hofmann-Mayboroda [17], Jimenez-del-Toro [21], Jimenez-del-Toro-Martell [22], Bernicot-Zhao [2], Bernicot-Martell [1], and Bui-Dong [4], to name a few. Although we will not pursue the analysis of these spaces here, it is easy to also represent them as X(B, Λ) for appropriate Λ's. The interested reader could use exponential John-Nirembeg inequalities obtained in the mentioned works to obtain p-invariance properties and hence conclude using our meth- ods weight invariance as well. 5.6. Weighted little bmo. Let Ap(Rn1 × Rn2) be the Muckenhoupt classes associated to rectangles of the form R = Q1 × Q2 for cubes Q1 ⊂ Rn1 and Q2 ⊂ Rn2. We denote the class of all such rectangles by R = R(Rn1 × Rn2). Let w ∈ A∞(Rn1 × Rn2). For a function f ∈ L1 loc(w), define kf kbmow = sup R∈R 1 w(R) R f (x) − fw,Rw(x)dx, where fw,R = 1 w(R) ´R f (x)dx. Let bmow(Rn1 × Rn2) be the collection of all w-locally integrable functions f modulo constants such that kf kbmow < ∞. We include this notation of bmow(Rn1 × Rn2) to help with our presentation, but later, as a consequence of Theorem 5.14, we will see that the spaces bmow(Rn1 × Rn2) in fact coincide for all w ∈ A∞(Rn1 × Rn2). This repro- duces Muckenhoupt and Wheeden's result for BM O in [29]. To simplify the notation, we will write bmow = bmow(Rn1 × Rn2) with the convention that we are working with a fixed decomposition Rn1 × Rn2 of Rn. As usual, we also write bmo in place of bmo1 when w = 1. The proof of the next lemma is adapted from the proof of John-Nirenberg inequality in the lecture notes by Journ´e [24, pp. 31-32]. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 31 Lemma 5.13. Let w ∈ A∞(Rn1 × Rn2). There exists a constant η > 0 such that sup R∈R 1 w(R) R ef (x)−fw,R/(ηkf kbmow )w(x)dx < ∞ for all f ∈ bmow. Proof. Let w ∈ A∞(Rn1 × Rn2) and f ∈ bmow with norm 1. For N ∈ N and η > 0, define TN (η) = sup R∈R 1 w(R) R emin(f (x)−fw,R,N )/ηw(x)dx. Now fix R = Q1 × Q2 ∈ R, λ > 0 (to be specified later), and choose Rj = Q1,j × Q2,j to be disjoint dyadic sub rectangles of R such that ℓ(Q1,j)/ℓ(Q2,j) = ℓ(Q1)/ℓ(Q2), λ < 1 w(Rj) Rj f (x) − fw,Rw(x)dx ≤ D2 wλ for all j, and f (x) − fw,R ≤ λ almost everywhere on R\Sj Rj. Here i.e. Dw the constant so that Dw denotes the doubling constant for w; w(2R) ≤ Dww(R) for all R ∈ R. (We use D2 w here since for any dyadic rectangle R, 22R = 4R engulfs its dyadic father). Note that this selection implies that fw,R − fw,Rj ≤ D2 R wλ for all j. Then it follows that ef (x)−fw,R/ηw(x)dx emin(f (x)−fw,R,N )/ηw(x)dx ≤ R\ Sj Rj +Xj emin(f (x)−fw,R,N )/ηw(x)dx Rj efw,Rj −fw,R/η Rj λ Xj Rj ! w(R). wλ/ηTN (η) λ eD2 ≤ eλ/ηw(R) +Xj ≤ eλ/ηw(R) + ≤ eλ/η + eD2 wλ/ηTN (η) emin(f (x)−fw,Rj ,N )/ηw(x)dx f (x) − fw,Rw(x)dx Dividing both sides by w(R) and taking the supremum over all R ∈ R, it follows that TN (η) ≤ eλ/η + eD2 wλ/ηTN (η) λ . Finally fix λ and η so that eD2 w λ/η 2 (for example we can just set λ = λ wλ/η η = 2eD2 λ = 1 2 ). Then it follows that TN (η) ≤ 2eλ/η for all N ∈ N (note that truncating by N in TN (η) assures us that TN (η) < ∞). w to obtain eD2 ≤ 1 32 JAROD HART AND RODOLFO H. TORRES Then for every R ∈ R and with λ, η selected as above, it follows that by monotone convergence that 1 w(R) R ef (x)−fw,R/ηw(x)dx ≤ lim N→∞ TN (η) ≤ 2eλ/η. This completes the proof. (cid:3) Some standard consequences of Lemma 5.13 are that for any weight w in A∞(Rn1 × Rn2) and f in bmow, there exist constants c1, c2 > 0 such that w({x ∈ R : f (x) − fw,R > λ}) ≤ c1w(R)e−c2λ/kf kbmow for all R ∈ R and λ > 0, and that (14) kf kp bmop w := sup R∈R 1 w(R) R for all 1 ≤ p < ∞. f (x) − fw,Rpw(x)dx . kf kp bmow Theorem 5.14. Let 0 < p < ∞, w, v ∈ A∞(Rn1 × Rn2), and f be a complex-valued function on Rn. Then f is in L1 loc(v) and satisfies R∈R(cid:18) 1 sup w(R) R R∈R(cid:18) 1 f (x) − fv,Rpw(x)dx(cid:19) 1 p < ∞ w(R) R f (x) − fv,Rpw(x)dx(cid:19) 1 p . if and only if f is in bmo. In such a case, (15) kf kbmo ≈ sup Proof. Given that we have proved (14) and that A2 holds for A∞(R) as was verified for example in [12], this proof follows exactly as in previous cases using Theorem 3.3 and Lemma 5.1. (cid:3) Remark 5.15. Recall from Remark 2.8 that if we take ∆(p, t) = 1 + 1 2p+2t and K(p, t) = 2, and n1 = n2 = 1 as is the situation in [27], then the classes Ap(R × R) satisfies A4. Applying Proposition 3.4, it follows that Ψp(t) ≤ 2c1,1+2p+2t . 2pt for 1 < p < ∞ and t ≥ 1. Note that the linear estimate c1,p . p holds for 1 < p < ∞ as a consequence of the John-Nirenberg inequality proved in Lemma 5.13. So for w ∈ Ap(R × R), we set t = [w]Ap to obtain b1,w ≤ Ψp([w]Ap) . 2p[w]Ap. In particular, sup R∈R 1 w(R) R f (x) − fRw(x)dx . 2p[w]Apkf kbmo for any f ∈ bmo, 1 < p < ∞, and w ∈ Ap(R × R), where the suppressed constant does not depend on f , p, or w. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 33 5.7. Discrete Triebel-Lizorkin spaces. We consider now the invariance of certain discrete Triebel-Lizorkin spaces defined by Frazier and Jawerth [10]. Let Qd ⊂ Q be the collection of standard dyadic cubes in Rn and let w ∈ Ad ∞ = A∞(Qd, dx) be dyadic version of A∞ defined in terms of Qd. Theorem 5.16. Let α ∈ R, 0 < q < ∞, 0 < p < ∞. Then for any sequence {sQ}Q∈Qd indexed by dyadic cubes Qd, we have k{sQ}k f α,q ∞ ≈ sup P ∈Qd 1 q := sup P ∈Qd 1 P P XQ∈Qd:Q⊂P w(P ) P XQ∈Qd:Q⊂P (Q−1/2−α/nsQχQ(x))qdx w(x)dx (Q−1/2−α/nsQχQ(x))q 1 p q  1 p , where w(P ) = ´P w(x)dx. Proof. Let F be the collection of sequences indexed by the collection of dyadic cubes Qd in Rn. Define for P ∈ Qd, {sQ} ∈ F, and x ∈ Rn. By [10, Corollary 5.7], it follows that k{sQ}kp Λ({sQ}, P )(x) = XQ∈Qd:Q⊂P P P XQ∈Qd:Q⊂P P ∈Qd 1 P P XQ∈Qd:Q⊂P X p = sup P ∈Qd ≈ sup = k{sQ}kp X . 1 (Q−1/2−α/nsQχQ(x))q p (Q−1/2−α/nsQχQ(x))q (Q−1/2−α/nsQχQ(x))qdx dx p Therefore X(Qd, Λ) satisfies the John-Nirenberg p-power inequality for the interval (0, ∞). Since the characterization k{sQ}kX p ≈ k{sQ}kX is valid for all sequences {sQ} ∈ F, meaning in particular that if k{sQ}kX p is finite for some 0 < p < ∞ then it is finite for all 0 < p < ∞, it follows that X p(Qd, Λ) = X(Qd, Λ) for all 0 < p < 1. The proof is completed by applying Theorem 3.3. (cid:3) 5.8. Triebel-Lizorkin spaces. Fix a function ϕ ∈ S (the usual Schwartz space of smooth rapidly decreasing functions) such that bϕ(ξ) is supported in 1/2 < ξ < 2 and bϕ(ξ) ≥ c0 > 0 for 3/5 < ξ < 5/3. Also define ϕk(x) = 2knϕ(2kx). For α ∈ R, 0 < q < ∞, and w ∈ A∞, define the homogeneous p = ∞ type Trieble-Lizorkin space F α,q ∞,w to be the collection 34 JAROD HART AND RODOLFO H. TORRES of all f ∈ S ′/P (tempered distribution modulo polynomials) such that kf k F α,q ∞,w = sup Q∈Q 1 w(Q) Q Xk∈Z:2−k≤ℓ(Q) (2αkϕk ∗ f (x))qw(x)dx 1 q < ∞, where w(Q) = ´Q w(x)dx. When w ≡ 1, we simply write F α,q The following lemma is implicit in the work in [10] as it will be seen from ∞ = F α,q ∞,1. its proof. Lemma 5.17. Let α ∈ R and 0 < p, q < ∞. Then for f ∈ S ′/P, we have sup Q∈Q 1 Q Q Xk∈Z:2−k≤ℓ(Q) (2αkϕk ∗ f (x))q Proof. Define for f ∈ S ′/P p dx . kf kpq F α,q ∞ . sup(f )Q = Q1/2 sup y∈Q ϕk ∗ f (y) for Q ∈ Qd, where ℓ(Q) = 2−k. Let α ∈ R and 0 < p, q < ∞. Then for any f ∈ S ′/P, we have sup P ∈Q = sup P ∈Q p 1 1 dx P P P P Xk∈Z:2−k≤ℓ(P ) (2αkϕk ∗ f (x))q  Xk∈Z:2−k≤ℓ(P ) XQ∈Qd:Q⊂P, P P XQ∈Qd:Q⊂P P ∈Q 1 P P XQ∈Qd:Q⊂P (2αkϕk ∗ f (x)χQ(x))q (Q−1/2−α/n sup(f )QχQ(x))q (Q−1/2−α/n sup(f )QχQ(x))qdx ℓ(Q)=2−k dx 1 p p ≤ sup P ∈Q . sup = k sup(f )kpq . f α,q ∞ p dx The last inequality here holds by [10, Corollary 5.7] applied with {sQ} = {sup(f )Q}. Also by [10, Theorem 5.2 and equation (5.6)], it follows that k sup(f )k f α,q (cid:3) Theorem 5.18. Let α ∈ R, 0 < q, p < ∞, and w ∈ A∞. Then for any f ∈ F α,q , which completes the proof. ≈ kf k F α,q ∞ , we have ∞ ∞ kf k F α,q ∞ ≈ sup Q∈Q  1 w(Q) Q Xk∈Z:2−k≤ℓ(Q) (2αkϕk ∗ f (x))q 1 p . p q w(x)dx It follows that X(Q, Λ) = F α,q follows that . By Lemma 5.17, it (2αkϕk ∗ f (x))q. ∞ and k · kX = k · kq Λ(f, Q)(x) = Xk∈Z:2−k≤ℓ(Q) Q Q Xk∈Z:2−k≤ℓ(Q) Q∈Q 1 (2αkϕk ∗ f (x))q Q Q Xk∈Z:2−k≤ℓ(Q) . sup 1 F α,q ∞ p dx p (2αkϕk ∗ f (x))qdx sup Q∈Q JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 35 Proof. Let α ∈ R, 0 < q, p < ∞, w ∈ A∞, F = F α,q ∞ and for all 0 < p < ∞. By Remark 3.6, it follows that X(Q, Λ) = X p(Q, Λ) for all 0 < p < 1. Then by Theorem 3.3, it follows that kf k F α,q ∞ = kf k1/q X ≈ kf k1/q X p/q w = sup Q∈Q  1 w(Q) Q Xk∈Z:2−k≤ℓ(Q) (2αkϕk ∗ f (x))q p/q for any f ∈ F α,q ∞ , 0 < p < ∞, α ∈ R, 0 < q < ∞, and w ∈ A∞. w(x)dx 1/p (cid:3) References [1] F. Bernicot and J. M. Martell, Self-improving properties for abstract Poincar´e type inequalities, Trans. Amer. Math. Soc. 367 (2015), no. 7, 4793-4835. [2] F. Bernicot and J. Zhao, Abstract framework for John Nirenberg inequalities and applications to Hardy spaces, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 11 (2012), no. 3, 475-501. [3] S. Bloom, A commutator theorem and weighted BMO, Trans. Amer. Math. Soc. 292 (1985), no. 1, 103-122. [4] H.-Q. Bui and X.-T. Duong, Weighted BMO spaces associated to operators, preprint, arXiv:1201.5828v3 [math.FA] 26Mar 2013. [5] H.-Q. Bui and M. Taibleson, The characterization of the Triebel-Lizorkin spaces for p = ∞, J. Fourier Anal. Appl. 6 (2000), no. 5, 537-550. [6] R. R. Coifman and G. Weiss, Extensions of Hardy spaces and their use in analysis, Bull. Amer. Math. Soc. 83 (1977), no. 4, 569-645. [7] X. T. Duong and L. Yan, Duality of Hardy and BMO spaces associated with operators with heat kernel bounds, J. Amer. Math. Soc. 18 (2005), no.4, 943-973. [8] X. T. Duong and L. Yan, New function spaces of BMO type, the John-Nirenberg inequality, interpolation, and applications, Comm. Pure Appl. Math. 58 (2005), no. 10, 1375-1420. [9] B. Franchi, C. P´erez and R. Wheeden, Self-improving properties of John-Nirenberg and Poincar´e inequalities on spaces of homogeneous type, J. Funct. Anal. 153 (1998), no. 5, 108-146. [10] M. Frazier and B. Jawerth, A discrete transform and decompositions of distribution spaces, J. Funct. Anal. 93 (1990), no. 1, 34-170. 36 JAROD HART AND RODOLFO H. TORRES [11] J. Garc´ıa-Cuerva, Weighted H p spaces, Dissertationes Math. (Rozprawy Mat.) 162 (1979), 63 pp. [12] J. Garc´ıa-Cuerva and J.-L. Rubio de Francia, Weighted norm inequalities and related topics, North-Holland Mathematics Studies, 116, Notas de Matem´atica, 104. North- Holland Publishing Co., Amsterdam, 1985. [13] E. Harboure, O. Salinas, and B. Viviani, A look at BM Oϕ(ω) through Carleson measures, J. Fourier Anal. Appl. 13 (2007), no. 3, 267-284. [14] S. Hartzstein and O. Salinas, Weighted BMO and Carleson measures on spaces of homogeneous type, J. Math. Anal. Appl. 342 (2008), no. 2, 950–969. [15] J. Hart and L. Oliveira, Hardy space estimates for limited ranges of Muckenhoupt weights, Adv. in Math. 313 (2017), 803–838. [16] K.-P. Ho Characterizations of BM O by Ap weights and p-convexity, Hiroshima Math. J. 41 (2011), no. 2, 153-165. [17] S. Hofmann and S. Mayboroda, Hardy and BMO spaces associated to divergence form elliptic operators, Math. Ann. 344 (2009), no. 1, 37-116. [18] I. Holmes, M. T. Lacey, and W. D. Wick, Commutators in the two-weight setting, Math. Ann. 367 (2017), no. 1-2, 51-80. [19] T. Hytonen and C. P´erez, Sharp weighted bounds involving A∞, Anal. PDE 6 (2013), no. 4, 777-818. [20] T. Hytonen, C. P´erez, and E. Rela, Sharp reverse Holder property for A∞ weights on spaces of homogeneous types, J. Funct. Anal. 263 (2012), no. 12, 3883-3899. [21] A. Jim´enez-del-Toro, Exponential self-improvement of generalized Poincar´e inequal- ities associated with approximations of the identity and semigroups, Trans. Amer. Math. Soc. 364 (2012), no. 2, 637-660. [22] A. Jim´enez-del-Toro and J.M. Martell, Self-improvement of Poincar´e type inequalities associated with approximations of the identity and semigroups, Potential Anal. 38 (2013), no. 3, 805-841. [23] F. John and L. Nirenberg, On functions of bounded mean oscillation, Comm. Pure Appl. Math. 14 1961, 415-426. [24] J.-L. Journ´e, Calder´on-Zygmund operators, pseudodifferential operators and the Cauchy integral of Calder´on, Lecture Notes in Mathematics, 994, Springer-Verlag, Berlin, 1983. [25] Y. Lo and L. Ruilin, BM O functions in spaces of homogeneous type, Sci. Sinica Ser. A 27 (1984), no. 7, 695–708. [26] A. A. Logunov, L. Slavin, D. M. Stolyarov, V. Vasyunin, and P. B. Zatitskiy, Weak integral conditions for BMO, Proc. Amer. Math. Soc. 143 (2015), no. 7, 2913–2926. [27] T. Luque, C. P´erez, and E. Rela, Reverse Holder property for strong weights and general measures, J. Geom. Anal. 27 (2017), no. 1, 162-182. [28] J. Mateu, P. Mattila, A. Nicolau, and J. Orobitg, BMO for nondoubling measures, Duke Math. J. 102 (2000), no. 3, 533–565. [29] B. Muckenhoupt and R. Wheeden, Weighted bounded mean oscillation and the Hilbert transform, Studia Math., 54 (1975/76), no. 3, 221-237. [30] U. Neri, Fractional integration on the space H 1 and its dual, Studia Math. 53 (1975), no. 2, 175-189 [31] J. Orobitg and C. P´erez, Ap weights for nondoubling measures in Rn and applications, Trans. Amer. Math. Soc. 354 (2002), no. 5, 2013-2033. [32] X. L. Shi and A. Torchinsky, Local sharp maximal functions in spaces of homogeneous type, Sci. Sinica Ser. A 30 (1987), no. 5, 473-480. [33] R. Strichartz, Bounded mean oscillation and Sobolev spaces, Indiana Univ. Math. J. 29 (1980), no. 4, 539-558. [34] R. Strichartz, Traces of BMO-Sobolev spaces, Proc. Amer. Math. Soc. 83 (1981), no. 3, 509-513. JOHN-NIRENBERG INEQUALITIES AND WEIGHT INVARIANT BMO SPACES 37 [35] J.-O. Stromberg, Bounded mean oscillation with Orlicz norms and duality of Hardy spaces, Indiana Univ. Math, J. 28 (1979), no. 3, 511-544. [36] N. Trong and N. Tung, Weighted BM O type spaces associated to admissible functions and applications, Acta Math. Vietnam 41 (2016), no. 2, 209-241. [37] Y. Tsutsui, A∞ constants between BM O and weighted BM O, Proc. Japan Acad. Ser. A Math. Sci. 90 (2014), no. 1, 11-14. Higuchi Biosciences Center, University of Kansas, Lawrence, KS 66045 E-mail address: [email protected] Department of Mathematics, University of Kansas, Lawrence, KS 66045 E-mail address: [email protected]
1709.04194	3	1709	2017-09-15T08:13:56	An example of non-uniqueness for the weighted Radon transforms along hyperplanes in multidimensions	[ "math.FA", "math-ph", "math-ph" ]	We consider the weighted Radon transforms $R_W$ along hyperplanes in $R^d, \, d \geq 3$, with strictly positive weights $W = W (x, \theta), \, x \in R^d, \, \theta \in S^{d-1}$. We construct an example of such a transform with non-trivial kernel in the space of infinitely smooth compactly supported functions. In addition, the related weight $W$ is infinitely smooth almost everywhere and is bounded. Our construction is based on the famous example of non-uniqueness of J. Boman (1993) for the weighted Radon transforms in $R^2$ and on a recent result of F. Goncharov and R. Novikov (2016).	math.FA	math	An example of non-uniqueness for the weighted Radon transforms along hyperplanes in multidimensions F.O. Goncharov∗ R. G. Novikov∗† October 1, 2018 Abstract We consider the weighted Radon transforms RW along hyperplanes in Rd, d ≥ 3, with strictly positive weights W = W (x, θ), x ∈ Rd, θ ∈ Sd−1. We construct an example of such a transform with non-trivial kernel in the space of infinitely smooth compactly supported functions. In addition, the related weight W is infinitely smooth almost everywhere and is bounded. Our construction is based on the famous example of non-uniqueness of J. Boman (1993) for the weighted Radon transforms in R2 and on a recent result of F. Goncharov and R. Novikov (2016). Keywords: weighted Radon transforms, injectivity, non-injectivity AMS Mathematics Subject Classification: 44A12, 65R32 1 Introduction We consider the weighted Radon transforms RW , defined by the formulas: RW f (s, θ) = Z xθ=s W (x, θ)f (x) dx, (s, θ) ∈ R × Sd−1, x ∈ Rd, d ≥ 2, (1.1) where W = W (x, θ) is the weight, f = f (x) is a test function on Rd. We assume that W is real valued, bounded and strictly positive, i.e.: W = W ≥ c > 0, W ∈ L∞(Rd × Sd−1), (1.2) where W denotes the complex conjugate of W , c is a constant. If W ≡ 1, then RW is reduced to the classical Radon transform R along hyperplanes in Rd. This transform is invertible by the classical Radon inversion formulas; see [Rad17]. If W is strictly positive, W ∈ C∞(Rd × Sd−1) and f ∈ C∞ 0 (Rd), then in [Bey84] the inversion of RW is reduced to solving a Fredholm type linear integral equation. Besides, in [BQ87] it was proved that RW 0(Rd)) if W is real-analytic and strictly positive. In addition, an example of is injective (for example, in L2 RW in R2 with infinitely smooth strictly positive W and with non-trivial kernel KerRW in C∞ 0 (R2) was constructed in [Bom93]. Here C∞ 0 denote the spaces of functions from C∞, L2 with compact support, respectively. 0 , L2 In connection with the most recent progress in inversion methods for weighted Radon transforms RW , see [Gon17]. We recall also that inversion methods for RW in R3 admit applications in the framework of emission tomographies (see [GN16]). In the present work we construct an example of RW in Rd, d ≥ 3, with non-trivial kernel KerRW in 0 (Rd). The related W satisfies (1.2). In addition, our weight W is infinitely smooth almost everywhere C∞ on Rd × Sd−1. In our construction we proceed from results of [Bom93] and [GN16]. In Section 2, in particular, we recall the result of [GN16]. In Section 3 we recall the result of [Bom93]. In Section 4 we obtain the main result of the present work. ∗CMAP, Ecole Polytechnique, CNRS, Universit´e Paris-Saclay, 91128 Palaiseau, France; email: [email protected] †IEPT RAS, 117997 Moscow, Russia; email: [email protected] 1 2 Relations between the Radon and the ray transforms We consider also the weighted ray transforms Pw in Rd, defined by the formulas: Pwf (x, θ) = Z R w(x + tθ, θ)f (x + tθ) dt, (x, θ) ∈ T Sd−1, T Sd−1 = {(x, θ) ∈ Rd × Sd−1 : xθ = 0}, d ≥ 2, where w = w(x, θ) is the weight, f = f (x) is a test-function on Rd. We assume that w is real valued, bounded and strictly positive, i.e.: w = w ≥ c > 0, w ∈ L∞(Rd × Sd−1). (2.1) (2.2) (2.3) We recall that T Sd−1 can be interpreted as the set of all oriented rays in Rd. In particular, if γ = (x, θ) ∈ T Sd−1, then γ = {y ∈ Rd : y = x + tθ, t ∈ R}, (2.4) where θ gives the orientation of γ. We recall that for d = 2, transforms Pw and RW are equivalent up to the following change of variables: RW f (s, θ) = Pwf (sθ, θ⊥), s ∈ R, θ ∈ S1, W (x, θ) = w(x, θ⊥), x ∈ R2, θ ∈ S1, θ⊥ = (− sin φ, cos φ) for θ = (cos φ, sin φ), φ ∈ [0, 2π), where f is a test-function on R2. For d = 3, the transforms RW and Pw are related by the following formulas (see [GN16]): RW f (s, θ) = Z R Pwf (sθ + τ [θ, α(θ)], α(θ)) dτ, (s, θ) ∈ R × S2, W (x, θ) = w(x, α(θ)), x ∈ R3, θ ∈ S2, (2.5) (2.6) (2.7) (2.8) (2.9) (2.10) (2.11) (2.12) (2.13) (2.14) (2.15) (2.16) (2.17) , if θ 6= ±η, [η, θ] [η, θ] any vector e ∈ S2, such that e ⊥ θ, if θ = ±η, α(θ) =   where η is some fixed vector from S2, [·, ·] denotes the standard vector product in R3, ⊥ denotes the orthogonality of vectors. Actually, formula (2.7) gives an expression for RW f on R × S2 in terms of Pwf restricted to the rays γ = γ(x, θ), such that θ ⊥ η, where W and w are related by (2.8). Below we present analogs of (2.7)-(2.8) for d > 3. Let Σ(s, θ) = {x ∈ Rd : xθ = s}, s ∈ R, θ ∈ Sd−1, Ξ(v1, . . . , vk) = Span{v1, . . . , vk}, vi ∈ Rd, i = 1, k, 1 ≤ k ≤ d, Θ(v1, v2) = {θ ∈ Sd−1 : θ ⊥ v1, θ ⊥ v2} ≃ Sd−3, v1, v2 ∈ Rd, v1 ⊥ v2, (e1, e2, e3, . . . , ed) - be some fixed orthonormal, positively oriented basis in Rd. If (e1, . . . , ed) is not specified otherwise, it is assumed that (e1, . . . , ed) is the standard basis in Rd. For d ≥ 3, the transforms RW and Pw are related by the following formulas: RW f (s, θ) = Z Rd−2 Pwf (sθ + d−2 Xi=1 τiβi(θ), α(θ)) dτ1 . . . dτd−2, (s, θ) ∈ R × Sd−1, W (x, θ) = w(x, α(θ)), x ∈ Rd, θ ∈ Sd−1, where α(θ), βi(θ), i = 1, d − 2, are defined as follows: direction of one-dimensional intersection Σ(s, θ) ∩ Ξ(e1, e2), where the orientation of α(θ) is chosen such that det(α(θ), θ, e3, . . . , ed) > 0, if θ 6∈ Θ(e1, e2), α(θ) =   any vector e ∈ Sd−1 ∩ Ξ(e1, e2), if θ ∈ Θ(e1, e2), (α(θ), β1(θ), . . . , βd−2(θ)) is an orthonormal basis on Σ(s, θ), and Σ(s, θ), Θ(e1, e2) are given by (2.10), (2.12), respectively. Here, due to the condition θ 6∈ Θ(e1, e2): dim(Σ(s, θ) ∩ Ξ(e1, e2)) = 1. (2.18) 2 Formula (2.18) is proved in Section 5. Note that formulas (2.14)-(2.18) are also valid for d = 3. (2.7)-(2.9), where e3 = −η. In this case these formulas are reduced to Note that, formula (2.14) gives an expression for RW f on R × Sd−1 in terms of Pwf restricted to the rays γ = (x, α), such that α ∈ Sd−1 ∩ Ξ(e1, e2). Remark 1. In (2.16) one can also write: α(θ) = (−1)d−1 ⋆ (θ ∧ e3 ∧ · · · ∧ ed), if θ 6∈ Θ(e1, e2), (2.19) where ⋆-denotes the Hodge star, ∧ - is the exterior product in Λ∗Rd (exterior algebra on Rd); see, for example, Chapters 2.1.c, 4.1.c of [Mor01]. Note that the value of the integral in the right hand-side of (2.14) does not depend on the particular choice of (β1(θ), . . . , βd−2(θ)) of (2.17). Note also that, due to (2.8), (2.9), (2.15), (2.16), the weight W is defined everywhere on Rd × Sd−1, d ≥ 3. In addition, this W has the same smoothness as w in x on Rd and in θ on Sd−1\Θ(e1, e2), where Θ(e1, e2) is defined in (2.12) and has zero Lebesgue measure on Sd−1. 3 Boman's example For d = 2, in [Bom93] there were constructed a weight W and a function f , such that: RW f ≡ 0 on R × S1, 1/2 ≤ W ≤ 1, W ∈ C∞(R2 × S1), f ∈ C∞ 0 (R2), f 6≡ 0, supp f ⊂ B 2 = {x ∈ R2 : x ≤ 1}. In addition, as a corollary of (2.5), (2.6), (3.1)-(3.3), we have that Pw0 f0 ≡ 0 on T S1, 1/2 ≤ w0 ≤ 1, w0 ∈ C∞(R2 × S1), f0 ∈ C∞ 0 (R2), f0 6≡ 0, supp f ⊂ B 2 = {x ∈ R2 : x ≤ 1}, where 4 Main results Let w0(x, θ) = W (x, −θ⊥), x ∈ R2, θ ∈ S1, f0 ≡ f. Bd = {x ∈ Rd : x < 1}, Bd = {x ∈ Rd : x ≤ 1}, (e1, . . . , ed) - be the canonical basis in Rd. Theorem 1. There are W and f , such that RW f ≡ 0 on R × Sd−1, W satisfies (1.2), f ∈ C∞ 0 (Rd), f 6≡ 0, where RW is defined by (1.1), d ≥ 3. In addition, 1/2 ≤ W ≤ 1, W is C∞-smooth on Rd × (Sd−1\Θ(e1, e2)), (3.1) (3.2) (3.3) (3.4) (3.5) (3.6) (3.7) (3.8) (4.1) (4.2) (4.3) (4.4) (4.5) (4.6) where Θ(e1, e2) is defined by (2.12). Moreover, weight W and function f are given by formulas (2.15), (4.8)-(4.10) in terms of the J. Boman's weight w0 and function f0 of (3.7), (3.8). Remark 2. According to (2.15), (2.16), W (x, θ) for θ ∈ Θ(e1, e2) can be specified as follows: W (x, θ) = W (x1, . . . , xd, θ) def = w0(x1, x2, e1), θ ∈ Θ(e1, e2), x ∈ Rd. Proof of Theorem 1. We define w(x, α) = w(x1, . . . , xd, α) def = w0(x1, x2, α1, α2), f (x) = f (x1, . . . , xd) for x = (x1, . . . , xd) ∈ Rd, α = (α1, α2, 0, . . . , 0) ∈ Sd−1 ∩ Ξ(e1, e2) ≃ S1, def = ψ(x3, . . . , xd)f0(x1, x2), (4.7) (4.8) (4.9) 3 where ψ ∈ C∞ 0 (Rd−2), supp ψ = Bd−2 and ψ(x) > 0 for x ∈ Bd−2. (4.10) From (2.1), (3.4), (4.8)-(4.10) it follows that: Pwf (x, α) = Z R w(x1 + tα1, x2 + tα2, x3, . . . , xd, α)f (x1 + tα1, x2 + tα2, x3, . . . , xd) dt = ψ(x3, . . . , xd)Z R w0(x1 + tα1, x2 + tα2, α1, α2)f0(x1 + tα1, x2 + tα2) dt = ψ(x3, . . . , xd)Pw0 f0(x1, x2, α1, α2) = 0 for any α = (α1, α2, 0, . . . , 0) ∈ Ξ(e1, e2) ∩ Sd−1 ≃ S1. Properties (4.4)-(4.6) follow from (2.15)-(2.17), (2.19), (3.2), (3.3), (4.7), (4.8). Theorem 1 is proved. 5 Proof of formula (2.18) Note that dim(Ξ(e1, e2)) + dim(Σ(s, θ)) = d + 1 > d, which implies that the intersection Σ(s, θ) ∩ Ξ(e1, e2) is one of the following: 1. The intersection is the one dimensional line l = l(s, θ): l(s, θ) = {x ∈ Rd : x = x0(s, θ) + α(θ)t, t ∈ R}, α(θ) ∈ S2, (4.11) (5.1) (5.2) where x0(s, θ) is an arbitrary point of Σ(s, θ) ∩ Ξ(e1, e2), the orientation of α(θ) is chosen such that: det(α(θ), θ, e3, . . . , ed) > 0. (5.3) Condition (5.3) fixes uniquely the direction of α(θ) of (5.2). Formulas (2.10), (2.11), (2.12) imply that (5.3) can hold if and only if θ 6∈ Θ(e1, e2). 2. The intersection is the two-dimensional plane Ξ(e1, e2). Formulas (2.10), (2.11) imply that it is the case if and only if s = 0, θ ⊥ e1, θ ⊥ e2. 3. The intersection is an empty set. Formulas (2.10), (2.11) imply that it is the case if and only if Note that s 6= 0, θ ⊥ e1, θ ⊥ e2. (5.4) (5.5) cases 2 and 3 occur if and only if θ ⊥ e1, θ ⊥ e2, i.e., θ ∈ Θ(e1, e2). (5.6) This completes the proof of formula (2.18). 6 Acknowledgments This work is partially supported by the PRC n◦ 1545 CNRS/RFBR: ´Equations quasi-lin´eaires, probl`emes inverses et leurs applications. References [Bey84] G. Beylkin. The inversion problem and applications of the generalized Radon transform. Com- munications on pure and applied mathematics, 37(5):579-599, 1984. [BQ87] J. Boman, E. Quinto. Support theorems for real-analytic Radon transforms. Duke Mathematical Journal, 55(4):943-948, 1987. [Bom93] J. Boman. An example of non-uniqueness for a generalized Radon transform. Journal d'Analyse Mathematique, 61(1):395 -- 401, 1993. [GN16] F. O. Goncharov, R. G. Novikov. An analog of Chang inversion formula for weighted Radon transforms in multidimensions. Eurasian Journal of Mathematical and Computer Applications, 4(2):23-32, 2016. [Gon17] F. O. Goncharov. Iterative inversion of weighted Radon transforms in 3D. arXiv:1611.10209, 2017. 4 [MQ85] A. Markoe, E. Quinto. An elementary proof of local invertibility for generalized and attenuated Radon transforms. SIAM Journal on Mathematical Analysis, 16(5):1114 -- 1119, 1985. [Mor01] S. Morita. Geometry of differential forms. American Mathematical Society., 2001. [Rad17] J. Radon. Uber die Bestimmung von Funktionen durch ihre Integralwerte langs gewisser Mannig- faltigkeiten. Ber. Saechs Akad. Wiss. Leipzig, Math-Phys, 69:262 -- 267, 1917. 5
1009.6219	1	1009	2010-09-30T19:08:41	Universal commutative operator algebras and transfer function realizations of polynomials	[ "math.FA", "math.OA" ]	To each finite-dimensional operator space $E$ is associated a commutative operator algebra $UC(E)$, so that $E$ embeds completely isometrically in $UC(E)$ and any completely contractive map from $E$ to bounded operators on Hilbert space extends uniquely to a completely contractive homomorphism out of $UC(E)$. The unit ball of $UC(E)$ is characterized by a Nevanlinna factorization and transfer function realization. Examples related to multivariable von Neumann inequalities are discussed.	math.FA	math	UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS AND TRANSFER FUNCTION REALIZATIONS OF POLYNOMIALS MICHAEL T. JURY Abstract. To each finite-dimensional operator space E is asso- ciated a commutative operator algebra U C(E), so that E em- beds completely isometrically in U C(E) and any completely con- tractive map from E to bounded operators on Hilbert space ex- tends uniquely to a completely contractive homomorphism out of U C(E). The unit ball of U C(E) is characterized by a Nevanlinna factorization and transfer function realization. Examples related to multivariable von Neumann inequalities are discussed. 1. Introduction Consider the algebra Pn = C[z1, . . . zn] of polynomials in n variables with complex coefficients. If T = (T1, . . . Tn) is an n-tuple of bounded, commuting operators on a Hilbert space H, then we can define a semi- norm k · kT on Pn in the obvious way: kpkT := kp(T )k. If now T is a collection of such n-tuples which is separating for Pn, that is, p(T ) = 0 for all T ∈ T if and only if p = 0, then the supremum (1.1) kpk := sup T ∈T kp(T )k. defines a norm on Pn, and the closure of Pn with respect to this norm is a Banach algebra. Moreover, if p is an m× m matrix of polynomials (equivalently, a polynomial with m × m matrix coefficients), we can similarly define (1.2) kpkn := sup T ∈T kp(T )k. Date: May 22, 2018. Research supported by NSF grant DMS 0701268. 1 2 MICHAEL T. JURY Explicitly, if we write p in multi-index notation as p(z) = Pn Anzn with An ∈ Mm×m(C), then kp(T )k denotes the norm of the operator An ⊗ T n Xn acting on Cm⊗ H, where T n has the obvious meaning. The completion of Pn in the norm (1.1) (together with the system of matrix norms k · kn) thus becomes an operator algebra. While not every Banach algebra norm on Pn can be obtained in this way, a number of norms of this type arise naturally and have been extensively studied. It will help to consider some examples. 1) Fix a nice domain Ω ⊂ Cn (say, the unit ball) and let T range over the commuting normal operators with joint spectrum in ∂Ω. Then by the spectral theorem (and the maximum principle) the norm kpkT is equal to the supremum norm kpk∞ = supz∈Ω p(z). 2) Another important example is the universal or Agler norm k · ku. Here T ranges over all commuting n-tuples of contractive operators on Hilbert space. It is known that k · ku is equal to the supremum norm over the unit circle T when n = 1 (this follows from von Neumann's inequality), and equal to the supremum norm over the 2-torus T2 when n = 2 (Ando's inequality), but the analogous statements are false for all n ≥ 3; a counterexample was first given by Kaijser and Varopoulos [13]. 3) Yet another well-known example comes from the row contractions; these are the commuting n-tuples for which I − n Xj=1 TjT ∗ j ≥ 0. It is remarkable in this case that the supremum (1.1) is always attained on a single distinguished row contraction, namely the n-shift S1, . . . Sn where Sj is the operator of multiplication by the coordinate function zj on a certain Hilbert space of holomorphic functions on the unit ball of Cn. The resulting norm on a polynomial p is the norm p inherits by acting as a multiplication operator on this space, called the Drury- Arveson space [8, 3, 5], which is the reproducing kernel Hilbert space on the unit ball of Cn with kernel k(z, w) = (1 −Pj zjwj)−1. It is known that this norm is generically strictly greater than the supremum norm over the ball, and in fact the two are inequivalent [3]. (In the previous example, it is not known whether the universal norm is equivalent to the supremum norm over Tn when n ≥ 3.) UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 3 Following Ambrozie and Timotin [2] and Ball and Bolotnikov [4] (and the more general approach of Mittal and Paulsen [9]) the last two examples may be unified in the following way: consider the n×n-matrix valued function Then the operators T1, . . . Tn are all contractive if and only if Q(z1, . . . zn) = diag(z1, . . . zn). (1.3) Row contractions are similarly characterized by the positivity of I − Q(T )Q(T )∗ for the 1 × n-matrix valued function I − Q(T )Q(T )∗ ≥ 0. Q(z) = (z1, . . . zn). In general, then, fix an analytic N × M matrix-valued polynomial Q in n variables and consider the domain DQ = {z ∈ Cn : kQ(z)k < 1}. (Examples (2) and (3) above give the unit polydisk Dn and the unit ball Bn respectively.) Now consider the class of commuting operator n-tuples TQ = {T = (T1, . . . Tn) : I − Q(T )Q(T )∗ ≥ 0}. This class of operators may be used to define an operator algebra norm on the space of polynomials as in (1.1). Say a polynomial lies in the Schur-Agler class SAQ if kp(T )k ≤ 1 for all T such that I − Q(T )Q(T )∗ ≥ 0. (It is possible for different polynomials Q to deter- mine the same domain DQ but distinct Schur-Agler classes SAQ; one of the motivations of the present paper is to investigate these differences in the case of linear Q.) Among the main results of [2] and [4] is that the classes SAQ are characterized by a "Nevanlinna factorization," so named because it may be read as a kind of generalization of a 1919 theorem of R. Nevanlinna. This theorem says that a function f in the unit disk D ⊂ C is holomorphic and bounded by 1 if and only if the kernel (1 − f (z)f (w))(1 − zw)−1 is positive semidefinite. Equivalently, there exists a Hilbert space H and a holomorphic function F : D → H such that (1.4) = F (z)F (w)∗ 1 − f (z)f (w) 1 − zw which it will be helpful to rewrite as 1 − f (z)f (w) = F (z)(1 − zw)F (w)∗. (1.5) A version of this theorem in the bidisk D2 was obtained by Agler [1], who showed that f : D2 → C is holomorphic and bounded by 1 if 4 MICHAEL T. JURY and only if there exists Hilbert space H and holomorphic functions F1, F2 : D2 → H such that (1.6) 1− f (z)f (w) = F1(z)(1− z1w1)F1(w)∗ + F2(z)(1− z2w2)F2(w)∗ If we put F = [F1 F2] and define Q(z) =(cid:18)z1 0 0 z2(cid:19) (1.7) (1.8) then (1.6) takes a form more reminiscent of (1.5): 1 − f (z)f (w) = F (z) [IH ⊗ (I2 − Q(z)Q(w)∗)] F (w)∗ In general, we have the following, which is special case of [4, Theorem 1.5]. (We state the theorem only in the case of polynomials, since it is really the operator algebra norm induced by the operators T that is of interest in the present paper.) Theorem 1.1. Let Q and DQ be as above, and let p be a matrix-valued analytic polynomial. Then the following are equivalent: 1) Agler-Nevanlinna factorization. There exists a Hilbert space (1.9) K, and an analytic function F : DQ → B(K, CN ) such that 1 − p(z)p(w)∗ = F (z) [IK ⊗ (I − Q(z)Q(w)∗)] F (w)∗ 2) Transfer function realization. There exists a Hilbert space K ′, a unitary transformation U : K ′ ⊕ CN → K ′ ⊕ CN of the form (1.10) (1.11) such that K ′ CN K ′ CN (cid:18) A B C D (cid:19) p(z) = D + C(I − Q(z)A)−1Q(z)B. 3) von Neumann inequality. p lies in the (matrix-valued) Schur- Agler class SAQ, that is, for every commuting n-tuple T such that I − Q(T )Q(T )∗ ≥ 0, (1.12) kp(T )kMN ⊗B(K) ≤ 1. The inequality in statement (3) always implies that p is bounded by 1 in DQ. The converse holds in D (von Neumann's inequality) and in D2 (with Q given by (1.7)) by Ando's theorem. In all other cases the converse is either false or an open problem. The purpose of this paper is to prove an analog of the above result where the single matrix-valued polynomial Q is replaced by a family of linear maps σ : Cn → B(H). In this respect there is some overlap with UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 5 the more general results of [9], where general analytic σ are considered, though the point of view of the present paper is somewhat different. In particular the linear maps σ are exactly the maps that are completely contractive with respect to a given n-dimensional operator space E. (We assume the reader is familiar with the notions of operator spaces, completely contractive maps, etc. The books [10] and [11] are excellent references. The facts and definitions we require are briefly reviewed in Section 2.) The role of these operator spaces (and their duals) is a central theme. (In fact the reader who is familiar with the results of [2, 4, 9] may prefer to read sections 4 and 5 first.) To state our main theorem, we introduce one bit of notation: if S = (S1, . . . Sn) is an n-tuple of operators on a Hilbert space K, write σS for the map (1.13) σS(z) := zjSj n Xj=1 from Cn into B(K). (Evidently every linear map σ : Cn → B(K) has this form.) Our main theorem, proved in Section 3, is the following: Theorem 1.2. Let E be a finite dimensional operator space, with un- derlying Banach space V , and let Ω ⊂ Cn denote the open unit ball of V . For each analytic MN -valued polynomial p, the following are equivalent: 1) Agler-Nevanlinna factorization. There exists a Hilbert space K, a completely contractive map σ : E → B(K), and an ana- lytic function F : Ω → B(K, CN ) such that (1.14) 1 − p(z)p(w)∗ = F (z) [IK − σ(z)σ(w)∗] F (w)∗ for all z, w ∈ Ω. 2) Transfer function realization. There exists a Hilbert space K ′, a unitary transformation U : K ′ ⊕ CN → K ′ ⊕ CN of the form (1.15) (1.16) K ′ CN K ′ CN (cid:18) A B C D (cid:19) and a completely contractive map σ : E → B(K ′) so that p(z) = D + C(I − σ(z)A)−1σ(z)B. for all z ∈ Ω. 6 MICHAEL T. JURY 3) von Neumann inequality. If S is a commuting n-tuple in B(K) and σS is completely contractive for E ∗, then (1.17) A quick observation: kp(S)kMN ⊗B(K) ≤ 1. if the operators S = (S1, . . . Sn) in statement (3) are commuting matrices, then the condition that σS be completely contractive for E ∗ is just the condition that S belong to the unit ball of E. This duality is described more fully in the next section. We now let UC(E ∗) denote the completion of the polynomials in the norm (1.18) kpkU C(E ∗) := sup S kp(S)k where the supremum is taken over all S appearing in item 3 of Theo- rem 1.2, and let UC(E ∗) denote the resulting operator algebra (UC for "Universal Commutative"). It is easy to see that the norm (1.18) con- trols the supremum norm over Ω, and hence every element of UC(E ∗) is a continuous function on the closure of Ω and analytic in Ω. These algebras UC(E ∗) are the universal commutative operator al- gebras of the title. Indeed, it is evident that UC(E ∗) has the following if σ : E ∗ → B(H) is any completely contractive universal property: map with commutative range, then σ has a unique extension to a com- pletely contractive homomorphism σ : UC(E ∗) → B(H). (The map σ picks out a commuting n-tuple S, and σ just evaluates on S.) This is discussed further in Section 4. The results of this paper overlap with those of [2, 4] only for those operator spaces E which can be embedded completely isometrically in B(H) for some finite-dimensional Hilbert space H. However many finite-dimensional operator spaces of interest do not admit such an embedding. Indeed among the most interesting operator spaces in the present context are the so-called maximal operator spaces MAX(V ), which correspond to the minimal UC-norms discussed in Section 5. (In particular we show that the supremum norm on the tridisk D3 is not a UC-norm.) Outside the exceptional cases of the V = ℓ1 and V = ℓ∞ norms on C2, we know of no maximal space which embeds in a matrix algebra (or even a nuclear C-algebra; in fact for every n > 16 there exist V with dimV = n such that MAX(V ) cannot embed in a nuclear C-algebra. See [11, pp.340–341]). 2. Preliminaries 2.1. Operator spaces and duality. Let k·k be a norm on Cn; write V for the Banach space (Cn,k · k). Let Ω denote the open unit ball of UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 7 V : Let h·,·i denote the standard symmetric (not Hermitian) inner product on Cn: Ω = {z ∈ Cn : kzk < 1} hz, wi = zjwj n Xj=1 We consider the dual space V ∗ with respect to this pairing, and let Ω∗ denote the dual unit ball: Ω∗ = {z ∈ Cn : hz, wi < 1 for all w ∈ Ω} We will equip V with various (concrete) operator space structures; each of these is determined by an isometric mapping ϕ : V → B(H) where H is a Hilbert space (of arbitrary dimension). More explicitly, if we write vectors z ∈ V in coordinate form with respect to the standard basis e1, . . . en, we will write Tj = ϕ(ej) so that ϕ(z) = ϕ(X zjej) =X zjTj := hz, Ti The matrix norm structure on E = (V, ϕ) is determined explicitly as follows: if A1, . . . An are matrices (all of some fixed size k × l) then k(A1 . . . An)kϕ := kX Aj ⊗ Tjk where the latter norm is the standard one in Mkl ⊗ B(H), obtained by identifying Mkl ⊗ B(H) with B(H l, H k). It will be convenient to write hA, Ti :=X Aj ⊗ Tj. Given an operator space structure E = (V, ϕ) and a linear map ψ : V → B(K), we say ψ is completely contractive with respect to ϕ if (2.1) kX Aj ⊗ ψ(ej)k ≤ kX Aj ⊗ ϕ(ej)k for all n-tuples of matrices A = (A1, . . . An). The map ψ will be called completely isometric if equality holds in (2.1) for all A. An operator space structure E over V naturally determines a dual operator space structure E ∗ over V ∗, by declaring E ∗ := CB(E, C) At the first matrix level, M1(E ∗) is isometrically V ∗. The Mm(E ∗)- norm of an m × m matrix A with entries from V ∗ is then given by the CB norm of the map from V to Mm(C) induced by A. In the case that E is finite dimensional, the duality can be described much more 8 MICHAEL T. JURY concretely in terms of a pairing between completely contractive maps for E and E ∗. It will be helpful to work this out very explicitly: first we recall that, since V is identified with Cn as a vector space, by the "canonical shuffle" elements of Mm(E) may be presented in one of two ways: either as m × m matrices with entries from V , ij) ∈ V, A = [~aij]m ~aij = (a1 ij, . . . an i,j=1, or as n-tuples of m × m matrices A = [A1 . . . An] where the i, j entry of Ak is ak ij. In general we will prefer the latter form. In particular it will be desirable to describe the matrix norms on E ∗ using these expressions. Fix an element A ∈ Mm(E ∗). Then A induces a map from V to Mm(C) via A · ~v = [h~aij, ~vi]m i,j=1 In turn, A sends an element B ∈ Ml(E) to the ml × ml matrix A · B = [h~aij,~bpqi]m i,j=1 l p,q=1 Using the definition of the symmetric pairing h,·,·i this last matrix may be written as a sum pq#m Now, for fixed k, the ml × ml matrix pq]m " n Xk=1 ak ijbk ijbk [ak l i,j=1 p,q=1 l p,q=1 i,j=1 may be canonically identified with the Kronecker tensor product Ak ⊗ Bk. Thus, up to a canonical shuffle, the matrix A · B is equal to n Ak ⊗ Bk Xk=1 Finally, by the definition of the matrix norms on E ∗, we have that the norm of A in Mm(E ∗) is equal to (2.2) kAkMm(E ∗) := sup kA · BkMml(C) = sup(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) n Xk=1 Ak ⊗ Bk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mml(C) UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 9 where the supremum is taken over all l ≥ 1 and all B in the unit ball of Ml(E). Similarly, we have for all B ∈ Ml(E) (2.3) n kBkMl(E) = sup(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1 Ak ⊗ Bk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mml(C) where the supremum is taken over all m ≥ 1 and all A in the unit ball of Mm(E)∗. The above considerations extend naturally to the setting of com- pletely contractive maps. Given an n-tuple of operators T = (T1, . . . Tn) on a Hilbert space B(H), define a linear map σT : Cn → B(H) by Proposition 2.1. Let E be an n-dimensional operator space and let E ∗ denote its dual. Given an n-tuple of operators S = (S1, . . . Sn), the map σS is completely contractive for E ∗ if and only if (2.4) σT (z) = zjTj. n Xj=1 n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 Sj ⊗ Tj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)min ≤ 1 for all n-tuples T for which the map σT is completely contractive for E. Remark: Throughout this paper, the norm in expressions such as (2.4) is understood to be the minimal tensor norm, that is, if S and T act on Hilbert spaces H and K respectively, the norm of the sum Pn j=1 Sj ⊗ Tj is its norm as an operator on H ⊗ K. We will henceforth omit the min subscript. Proof. Suppose σS is completely contractive for E ∗. Then for all A = (A1, . . . An) in the unit ball of Ml(E ∗), n Xk=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Sk ⊗ Ak(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1 Now let T be an n-tuple of operators on a (separable) Hilbert space H, such that σT is completely contractive for E. Fix an orthonormal basis for H and let Pk be the projection onto the span of the first k basis vectors. Define an n-tuple of k × k matrices Ak j = PkTjPk. (The matrix of Ak claim that j is written with respect to the fixed basis of H.) We Ak = (Ak 1, . . . Ak n) 10 MICHAEL T. JURY belongs to the unit ball of Mk(E ∗). To see this, by (2.2) it suffices to prove that for all B = (B1, . . . Bn) in the unit ball of Ml(E), for all l. But since σT is completely contractive for E, the map σk := PkσT Pk is as well, and we have n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 Bj ⊗ Ak j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Bj ⊗ Tj! (I ⊗ Pk)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) since σT is completely contractive. This proves the claim. Now, by the hypothesis that σS is completely contractive for E ∗, ≤ 1 ≤ 1 n n Ak j ⊗ Bj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (I ⊗ Pk) n Xj=1 Bj ⊗ Tj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 Xj=1 Sj ⊗ Ak Sj ⊗ Ak ≤ 1 n n n j=1 Sj⊗Tjk ≤ 1, as desired. for all k, but since Xj=1 j → For the converse, suppose that Sj ⊗ Tj in the strong operator topology, we have kPn Sj ⊗ Tj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj=1 n for all T such that σT is completely contractive for E. By (2.2), the map σA is completely contractive for E whenever A is an n-tuple of matrices in the unit ball of Mm(E ∗). It is then immediate that σS is completely contractive for E ∗. (cid:3) 2.2. Factorization of positive semidefinite functions. Let Ω be a set and K a Hilbert space. Following [7], a function Γ : Ω×Ω → B(K)∗ is called positive semidefinite if (2.5) Γ(z, w)(f (z)f (w)∗) ≥ 0 Xz,w∈Λ UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 11 for every finite subset Λ ⊂ Ω and every function f : Λ → B(K). This definition may be naturally extended if we replace B(K)∗ with B(B(K), MN (C)): then Γ : Ω × Ω → B(B(K), MN ) is positive semi- definite if and only if (2.6) v(z)Γ(z, w)(f (z)f (w)∗)v(w)∗ ≥ 0 Xz,w∈Λ for all finite Λ ⊂ Ω, and all functions f : Λ → B(K), v : Λ → CN . In the scalar case, the following lemma reduces to [7, Proposition 4.1]. The proof of the matrix-valued version stated here is entirely analogous and is omitted. Lemma 2.2. A function Γ : Ω × Ω → B(B(K), MN ) is positive semi- definite if and only if there exists a Hilbert space H and a function G : Ω → B(B(K), B(H, CN )) such that (2.7) Γ(z, w)(ab∗) = (G(z)[a])(G(w)[b])∗ for all a, b ∈ B(K). Lemma 2.3. Suppose E is a finite dimensional operator space, ψ : E → B(K) is completely contractive map, and Γ : Ω×Ω → B(B(K), MN ) is a positive semidefinite function. Then there exists a Hilbert space H, a completely contractive map σ : E → B(H) and a function F : Ω → B(H, CN ) such that Γ(z, w)[IK − ψ(z)ψ(w)∗] = F (z)(IH − σ(z)σ(w)∗)F (w)∗. Proof. Given the completely isometric map ψ : E → B(K), let A denote the unital C-subalgebra of B(K) generated by the operators {ψ(z) : z ∈ Ω}. Choose G to factor Γ as in Lemma 2.2. Now, in the factorization (2.7), let H ′ denote the subspace of H spanned by vectors of the form (G(w)[a])∗v for w ∈ Ω, a ∈ A, v ∈ CN . We then obtain a "right regular representation" π : A → B(H ′) by defining (2.8) π(a)∗(G(w)[b])∗v = (G(w)[ba])∗v It is straightforward to check that π is a ∗-homomorphism: linearity is evident, and for all x, y ∈ A we have (2.9) π(xy)∗(G(w)[b])∗v = (G(w)[bxy])∗v (2.10) (2.11) = π(y)∗(G(w)[bx])∗v = π(y)∗π(x)∗(G(w)[b])∗v 12 MICHAEL T. JURY so π is multiplicative. Similarly (2.12) (2.13) (2.14) (2.15) u∗G(z)[a]π(x)∗(G(w)[b])∗v = u∗G(z)[a](G(w)[bx])∗v = u∗Γ(z, w)[ax∗b∗]v = u∗G(z)[ax∗](G(w)[b])∗v = u∗G(z)[a]π(x∗)(G(w)[b])∗v so π(x∗) = π(x)∗. Now define σ(z) := π(ψ(z)). It is evident that σ is a completely contractive map from E to B(H ′). It follows from (2.8) and the definition of A that (2.16) for all z ∈ Ω, b ∈ A and v ∈ CN . and Equation 2.16 that Now define H = K ′ and F (z) := G(z)[IK]. It follows from Lemma 2.2 σ(w)∗(G(w)[b])∗v = (G(w)[bψ(w)])∗v Γ(z, w)[IK] = G(z)[IK](G(w)[IK])∗ = F (z)F (w)∗ and Γ(z, w)[ψ(z)ψ(w)∗] = (G(z)[IKψ(z)])(G(w)[IKψ(w)])∗ = F (z)σ(z)σ(w)∗F (w)∗, and thus Γ(z, w)(IK − ψ(z)ψ(w)∗) = F (z)F (w)∗ − F (z)σ(z)σ(w)∗F (w)∗ as desired. (cid:3) 3. Main Theorem The proof of each implication in Theorem 1.2 is handled in a separate subsection. 3.1. 1 implies 2. Proof. This is a standard application of the "lurking isometry" tech- nique. Rearrange (1.14) to obtain 1 + F (z)σ(z)σ(w)∗F (w)∗ = p(z)p(w)∗ + F (z)F (w)∗ (3.1) Define subspaces M,N ⊂ H ′ ⊕ CN by M = span(cid:26)(cid:18)σ(w)∗F (w)∗x N = span(cid:26)(cid:18)F (w)∗x x (cid:19) : w ∈ Ω, x ∈ CN(cid:27) p(w)∗x(cid:19) : w ∈ Ω, x ∈ CN(cid:27) . UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 13 The equation (3.1) then implies the existence of an isometry U ∗ : M → N such that U ∗(cid:18)σ(w)∗F (w)∗x x (cid:19) =(cid:18)F (w)∗x p(w)∗x(cid:19) for all w ∈ Ω and x ∈ CN . Enlarging H ′ to a space H ′′ if necessary, we may extend U ∗ to a unitary (still denoted U ∗) taking H ′′⊕ CN to itself. We also regard σ as taking CN into B(H ′′), by declaring σ(w)x to be 0 for all w ∈ Cn and all x ∈ H ′′ ⊖ H ′. (Note this extended σ is still completely contractive.) We now write the action of U ∗ as a unitary colligation B∗ D∗(cid:19)(cid:18)σ(w)∗F (w)∗x (cid:18)A∗ C ∗ x (cid:19) =(cid:18)F (w)∗x p(w)∗x(cid:19) This corresponds to the linear system (3.2) (3.3) A∗σ(w)∗F (w)∗ + C ∗ = F (w)∗ B∗σ(w)∗F (w)∗ + D∗ = p(w)∗ This system may be solved to obtain p(z) = D + C(I − σ(z)A)−1σ(z)B for all z ∈ Ω. (Note that (I − σ(z)A) is invertible, since kAk ≤ 1 and kσ(z)k ≤ kzkV < 1 for all z ∈ Ω.) 3.2. 2 implies 3. (cid:3) Proof. Write n σ(z) = zjTj. Xj=1 Let S = (S1, . . . Sn) induce a completely contractive map σS of E ∗ on B(L). Then by Proposition 2.1, k n Xj=1 Sj ⊗ Tjk ≤ 1. Given the unitary colligation U, let A = IL ⊗ A, B = IL ⊗ B, etc. Fix 0 < r < 1; and observe that p(rS) = D + ChrS, Ti(I − AhrS, Ti)−1 B. (3.4) Since krSk < 1 and the Sj commute, the right-hand side of (3.4) may be expanded in a norm-convergent power series in the Sj. Using (1.16), we may also expand the left-hand side in the Sj, by first expanding 14 MICHAEL T. JURY p and then substituting rS. The equality then follows by matching coefficients. It is now easy to verify that I − p(rS)∗p(rS) ≥ 0 (3.5) for all r < 1 and hence I − p(S)∗p(S) ≥ 0 by letting r → 1. To prove (3.5), let A, B, C, D be any unitary colligation and X any operator with kXk < 1. Then if we define a well-known calculation shows that Q = D + CX(I − AX)−1B I − Q∗Q = B∗(I − AX)−1∗(I − X ∗X)(I − AX)−1B ≥ 0. Taking X = hrS, Ti and Q = p(rS) proves the claim. 3.3. 3 implies 1. (cid:3) Proof. This is the most involved part of the proof; the argument will be broken into several sub-arguments. We will first show that, given any finite set Λ ⊂ Ω, there exist F and ψ so that (1.14) holds for all z, w ∈ Λ. (This constitutes the bulk of the proof.) We then conclude that such a factorization is valid on all of Ω via a compactness argument (in particular, by appeal to Kurosh's theorem). So, fix a finite set Λ = {λ1, . . . λk} ⊂ Ω. Consider the cone C of kN × kN Hermitian matrices which can be written in the form (3.6) Aij = [F (λi) (1 − ψ(λi)ψ(λj)∗) F (λj)∗]ij where F is a function from Λ to a Hilbert space B(K, CN ) and ψ is a completely contractive map of E into B(K). It is easy to see that C contains all positive semidefinite matrices: if A is positive semidefinite we may factor it as Aij = F (λi)F (λj)∗ and take ψ = 0. Moreover, observe that for all A ∈ C, the Hilbert space K in the above map can be taken to be a fixed space of finite dimension at most 2kN. To see this, note that the factorization that appears in the right hand side of 3.6 takes place in the subspace of K given by span {F (λi)∗x, ψ(λi)∗F (λi)∗x : i = 1, . . . k, x ∈ CN} We now suppose that the kN × kN Hermitian matrix Pij = IN − p(λi)p(λj)∗ does not belong to C. Our first claim is the following: Claim 1: C is closed. It follows that there exists a real linear functional L : M sa kN (C) → R such that L(C) ≥ 0 but L(P ) < 0. We extend L to a complex linear functional on all of MkN (C) (still denoted L) in the standard way. UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 15 Using L we construct a pre-Hilbert space: B(K, CN ), define for functions F, G : Λ → hF, GiL := L([F (λi)G(λj)∗]) Since L is positive on C and C contains all positive matrices, it follows that h·,·iL is positive semidefinite. Denote by H the resulting pre- Hilbert space. We next construct an n-tuple of operators on H. First, if Q : Λ → B(K) is any function, we can define a "right multiplication operator" MQ on H via (MQF )(λ) = F (λ)Q(λ) (In fact, the only Q we need will be scalar multiples of the identity, but it will be helpful to think of this scalar multiplication as occurring on the right rather than the left.) Now, for each λi ∈ Λ, write its coordinates as λi = (λ1 and define operators Sk : H → H by i , . . . λn i ) (SkF )(λi) := MλkIF (λi) = F (λi)λk i We will construct from these operators a completely contractive map of the operator space E ∗: is any completely contractive map from E to B(E), then the operator ψ(z) = Claim 2: If E is any Hilbert space and Xk=1 zkTk Sk ⊗ Tk!∗ n Xk=1 I − n Xk=1 is nonnegative on H ⊗ E. n Sk ⊗ Tk! From this claim it follows easily that Claim 3: If hF, FiL = 0 then hSkF, SkFiL = 0 for all k = 1, . . . n. We may now construct a Hilbert space from H as usual, by passing to the quotient by the space of null vectors and completing; denote the resulting Hilbert space H′. Claims 2 and 3 show that the operators Sk pass to well-defined, bounded operators on H′, which will still denote Sk. It is also immediate from Claim 2 that k n Xk=1 Sk ⊗ Tkk ≤ 1 16 MICHAEL T. JURY whenever ψ(z) = P zkTk is completely contractive for E; thus by Proposition 2.1, the map n ϕ(z) = zkSk Xk=1 is completely contractive for E ∗. The proof that (1.14) is valid on finite sets will now be complete if we can show that 1 − p(S)∗p(S) is not positive on H′. Let J denote the kN × kN which has the N × N identity matrix IN in the i, j block for all i, j = 1, . . . k. The matrix J may be factored as G(λi)G(λj)∗ where G(λi) = IN for all i. Then h(I − p(S)∗p(S))G, GiH′ = h(I − p(S)∗p(S))G, GiL = L([G(λi)(IN − p(λi)p(λj)∗)G(λj)∗]) = L(IN − p(λi)∗p(λj)) < 0. We have now proved the existence of the factorization on finite sets, modulo the proofs of the claims, which are now provided. After these, the factorizations on finite sets will be pieced together, and the proof will be finished. Proof of Claim 1: To see that C is closed we appeal again to the lurking isometry technique. So, suppose X ∈ C. Since X is Hermitian, by the spectral theorem we may write X as a difference of two positive matrices X = P − N with kPk ≤ kXk,kNk ≤ kXk. Now factor P and N as Grammians: There exist F and ψ so that Pij = hpj, pii, Nij = hnj, nii (3.7) Xij = hpj, pii − hnj, nii = F (λj)∗(1 − ψ(λj)∗ψ(λi))F (λi) As before, the lurking isometry argument leads to the equation F (λi) = (I − Aψ(λi))−1Bpi where A, B belong to a unitary colligation. Now, kAψ(λi)k ≤ kλikV < 1 and kpik ≤ kXk for all i, and so (3.8) kF (λi)k ≤ (1 − kλikV )−1kXk for all i. Let Xn be a sequence in C and suppose Xn → X. For each n we obtain Fn, ψn so that (3.7) holds. By (3.8) the functions Fn are uni- formly bounded, and hence admit a subsequence converging to some UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 17 F : Λ → K. Since the ψn are also uniformly bounded, passing to a fur- ther subsequence if necessary, we may assume that ψn → ψ pointwise in norm for some completely contractive ψ. (The fact that ψ(z) acts on a finite-dimensional space is used here.) It follows that this F and ψ factor X as in (3.7), and hence X ∈ C. an orthonormal set in E. To prove Claim 2 it suffices to show that (3.9) I − n Xk=1 Proof of Claim 2: Let F1, . . . Fd : Λ → B(K, CN ) and let e1, . . . ed be Sk ⊗ Tk!! (X Fl ⊗ el), (X Fm ⊗ em)+H⊗E Sk ⊗ Tk!∗ n Xk=1 is positive; this will be the case because this is in fact may be written as the functional L applied to an kN × kN matrix lying in the cone C. To see this, let us write Tk for the operator IK ⊗ Tk on K ⊗ E, and F (λi) =P Fl(λi) ⊗ el. By the definition of S we have (3.10) (3.11) (3.12) (3.13) Xk Fl(λi)λk (Sk ⊗ Tk)(Fl ⊗ el)(λi) =Xk = Xk = hλi, Ti(Fl(λi) ⊗ el) = hλi, Ti F (λi) i ⊗ Tkel Tk! (Fl(λi) ⊗ el) λk i Now, using the fact that {el} is orthonormal, (3.14) DX Fl ⊗ el,X Fm ⊗ emEH⊗E (3.15) = L(cid:16)X Fl(λi)Fl(λj)∗(cid:17) = L( F (λi) F (λj)∗) Combining the above calculations, we find that (3.9) may be written as (3.16) L(cid:16) F (λi)[1 − hλi, Tihλj, Ti∗] F (λj)∗(cid:17) Since the map z → hz, Ti is completely contractive for E, the map obtained by replacing T with T is as well. It follows that the argument of L in (3.16) belongs to C, and hence (3.9) is positive, as desired. that, for each k = 1, . . . n, the map Proof of Claim 3: Trivially, there exists a real number α > 0 such σ(z) = αzk 18 MICHAEL T. JURY is a completely contractive map of E. Applying Claim 2 to this map (that is, taking Tk = α, Tj = 0 for j 6= k) we get I − α2S ∗ kSk ≥ 0 for each k. Thus the operators Sk are bounded on H, so in particular hSkF, SkFiL = 0 whenever hF, FiL = 0. We proved that a factorization (1.14) exists on every finite subset Λ ⊂ Ω. The extension to all of Ω is accomplished via a routine applica- tion of Kurosh's theorem. For each finite set Λ ⊂ Ω fix a factorization (1.14). Let HΛ be the Hilbert space on which ψ acts. Put H :=LΛ HΛ and ψ :=LΛ ψΛ. Now for each Λ let ΦΛ be the set of all positive semi- definite functions ΓΛ : Λ × Λ → B(B(H), MN ) such that 1 − p(z)p(w)∗ = ΓΛ(z, w)[IH − ψ(z)ψ(w)∗] (3.17) for all z, w ∈ Λ. Each ΦΛ is nonempty, since it contains (3.18) where PΛ : H → HΛ is the orthogonal projection. By identifying B(B(K), MN )) with MN (B(K)∗), the former space inherits the topol- ogy of entrywise weak-* convergence. The set of functions from Λ × Λ to B(B(K), MN ) may then be endowed with the topology of pointwise convergence in this topology on B(B(K), MN ) (in other words, the "pointwise entrywise weak-* topology"). The sets ΦΛ are then com- pact in this topology; this follows from the boundedness argument in the proof of Claim 1. It is evident that restriction induces a continuous map παβ : Φα → Φβ when β ⊂ α, so by Kurosh's theorem there exists a positive semidefinite Γ which satisfies (3.17) for all z, w ∈ Ω. Finally, applying Lemma 2.3 to this Γ and ψ finishes the proof. ΓΛ(z, w)[a] = F (z)PΛaPΛF (w)∗ (cid:3) 4. Universality of UC(E) In this section, to unclutter the notation a bit, we reverse the roles of E and E ∗ (which is harmless, since finite-dimensional operator spaces are reflexive), and consider the operator algebras UC(E). So (4.1) kpkU C(E) = sup S {kp(S)k} the supremum taken over commuting n-tuples S such that σS : E → B(K) is completely contractive. As noted earlier, if the Sj are matrices, then this is just the condition that S lies in the unit ball of E ∗. Pisier [11, Chapter 6] introduces the universal (unital) operator alge- bra associated to an operator space E; this algebra is denoted OAu(E). We will not require an explicit construction of OAu(E) here, only that OAu(E) has the following properties: UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 19 Proposition 4.1. The following properties characterize OAu(E): (1) There exists a canonical completely isometric embedding ι : E → OAu(E). (2) If σ : E → B(H) is completely contractive, there exists a unique completely contractive unital homomorphism σ : OAu(E) → B(H) extending σ, i.e. so that σ(ι(x)) = σ(x) for all x ∈ E. Similarly, the algebras UC(E) are "universal" among commutative operator algebras containing E completely contractively; in particular we have: Proposition 4.2. Let E be a finite-dimensional operator space. (1) There exists a canonical completely isometric embedding ι : E → UC(E). (2) If σ : E → B(H) is a completely contractive map with commu- tative range, then there exists a unique completely contractive unital homomorphism σ : UC(E) → B(H) extending σ. Proof. Everything is more or less immediate. For the embedding of E into UC(E), since the vector space underlying E is just Cn we let the map ι send the point a = (a1, . . . an) to the linear polynomial p(z) = from the definition of the UC(E) norms and the duality described in Section 2. For the extension property, the map σ has the form σ(a) = P ajzj. That this embedding is completely isometric is immediate P ajSj for commuting Sj's, and thus by definition σ(p) := p(S) works; uniqueness is clear since the linear polynomials generate C[z1, . . . zn] as a (unital) algebra, and σ extends uniquely to the completion UC(E). (cid:3) The observations in the proof of Proposition 4.2 may be organized slightly differently. Restricting the operator algebra structure of UC(E) to the linear polynomials, we get a completely isometric copy of E. Thus a homomorphism π from the polynomials into B(K) is com- pletely contractive for UC(E) if and only if its restriction to the linear polynomials is completely contractive for the induced operator space structure. This gives a way of detecting whether or not a given operator algebra structure on the polynomials agrees with some UC(E). This observation is exploited in the next section to show that the tridisk algebra A(D3) is not completely isometric to any UC(E). A routine categorical argument shows that the universal property of Proposition 4.2 characterizes UC(E) (up to complete isometry) among 20 MICHAEL T. JURY the commutative operator algebras which contain E completely iso- metrically. We then obtain: Proposition 4.3. Let E be a finite-dimensional operator space, OAu(E) the universal (unital) operator algebra over E, and C the commutator ideal of OAu(E). Then UC(E) ∼= OAu(E)/C, completely isometrically. Proof. We begin with the observation that the map of E into OAu(E)/C given by the composition E ֒→ OAu(E) → OAu(E)/C is completely isometric. (The first map is the canonical (completely isometric) embedding into OAu(E); the second is the quotient map.) To see this, it suffices to see that the restriction of the quotient map to E is completely isometric; this in turn follows from the existence of a completely isometric map σ : E → B(H) with commutative range. Such a map can be obtained by taking any complete isometry τ : E → B(K) and defining σ =(cid:18)0 τ 0 0(cid:19) . With this canonical embedding of E into the quotient in hand, it is straightforward to check that OAu(E)/C has the universal property of Proposition 4.2, and hence is completely isometrically isomorphic to UC(E). (cid:3) It is shown in [11, Chapter 6] that the operator algebra norm on OA(E) is realized by taking the supremum over just those completely contractive representations of OA(E) on finite-dimensional Hilbert spaces. It is not obvious that the same is true for UC(E)-the difficulty is that if σ : E → B(H) has commuting range and P is a projection in B(H), the map P σP need not have commuting range. However the proof of Theorem 1.2 shows that UC(E) is indeed determined by its finite- dimensional representations: Theorem 4.4. For every matrix-valued polynomial p, we have (4.2) kpkU C(E) = sup kp(S)k where the supremum is take over all n-tuples of commuting matrices for which σS is completely contractive for E; in other words, over all commuting matrices in the unit ball of E ∗. UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 21 Proof. This is really an immediate consequence of the fact that the operators Sk constructed in the proof of the "(3) implies (1)" implica- tion of Theorem 1.2 act on a finite-dimensional Hilbert space. More precisely, (recalling the terminology and notation used in the proof of Theorem 1.2), if p is given and does not admit a Nevanlinna factoriza- tion, then there exists a finite set Λ ⊂ Ω for which 1 − p(z)p(w)∗ does not belong to the cone C. In this setting, the proof of the (3) =⇒ (1) implication produces an n-tuple of operators S = (S1, . . . Sn) on the finite-dimensional Hilbert space H for which I − p(S)p(S)∗ is non- positive. The contrapositive of this statement is that if kp(S)k ≤ 1 for all admissible matrices S, then p admits a Nevanlinna factorization, and hence (4.2) holds. (cid:3) Another useful fact about OA(E) is that it "commutes" with Calderon interpolation [11, Section 2.7], that is, for any pair of compatible oper- ator spaces E0, E1, OA(Eθ) = [OA(E0), OA(E1)]θ completely isometrically. We do not know if the analogous statement is true for UC(E). Question 4.5. Is it the case that UC(Eθ) ∼= [UC(E0), UC(E1)]θ completely isometrically? 5. Examples Lacking a better name, in what follows we shall refer to the operator algebra norms described by Theorem 1.2 generically as UC-norms. One natural class of examples in the present context are those coming from the minimal and maximal operator space structures over the given Banach space V . We briefly recall the definitions. To define MIN(V ), we observe that the duality between V and V ∗ induces a natural map e from V into the space of continuous functions on the unit ball of V ∗ (denoted C(V ∗ 1 )), by sending z to the functional it induces on V ∗: z → h·, zi By the Hahn-Banach theorem, this map is isometric if C(V ∗ 1 ) is equipped with the supremum norm. Since this norm makes C(V ∗ 1 ) into a C*- algebra, the embedding thus defines an operator space structure on V , called the minimal operator space over V , and is denoted MIN(V ). The operator space MAX(V ) is defined by the matrix norms k(vij)kN := sup ϕ k(ϕ(vij)kB(H N ) 22 MICHAEL T. JURY where the supremum is taken over all contractive linear maps from V into B(H). In other words, every contractive map out of V is com- pletely contractive for MAX(V ). On the other hand, a map is com- pletely contractive for MIN(V ) if and only if it is completely contrac- tive for every operator space structure over V . It is well-known (and not too hard to prove) that MIN(V )∗ = MAX(V ∗) and MAX(V ∗) = MIN(V ). It follows that for each V , there is a unique minimal and maximal UC-norm associated to the domain Ω = ball(V ). We denote these norms kpkM IN (Ω) and kpkM AX(Ω) respectively. The largest norm has the smallest unit ball; and hence allows the fewest completely contractive maps to appear in the Nevanlinna factorization. This happens when we choose E = MIN(V ) in Theorem 1.2, so the maximal UC-norm over Ω = ball(V ) is obtained by taking the supremum over all commuting T such that the map σT is completely contractive for MIN(V )∗ = MAX(V ∗). For example, if Ω is the unit polydisk Dn, then V = ℓ∞ n and V ∗ = ℓ1 n) if and only if it is contractive, that is if and only if n. Now, σT is completely contractive for MAX(ℓ1 n n zjTjk ≤ k Xj=1 zj. Xj=1 for all z ∈ Cn. Clearly this happens if and only if each Tj is con- tractive, so by the von Neumann inequality of Theorem 1.2 we see that the maximal UC-norm over the polydisk is equal to the universal norm (the supremum over all commuting contractions) discussed in the introduction. 5.1. MIN(ℓ1 n). In fact, the above considerations allow us to observe a stronger consequence of the Kaiser-Varopoulos counterexample to the three-variable von Neumann inequality. The original example, inter- preted in the present setting, shows that kpkM AX(D3) > kpk∞ on the polydisk D3. In fact their example shows that kpkM IN (D3) > kpk∞. More precisely, the triple commuting contractions T of the Kaijser- Varopoulos example [13] are such that σT is completely contractive for MIN(ℓ1), and hence kpkM IN (D3) ≥ kp(T )k > kpk∞. It should be stressed that this is a particular feature of this example and not true generically of counterexamples to the three-variable von Neumann in- equality; in particular it is not true of the 8 × 8 example produced by Crabb and Davie [6]. Proposition 5.1. The Kaiser-Varopoulos contractions are completely contractive for MIN(ℓ1 n). UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 23 Proof. Let e1, . . . e5 denote the standard basis of C5. Consider the unit vectors v1 = v2 = v3 = 1 √3 1 √3 1 √3 (−e2 + e3 + e4) (e2 − e3 + e4) (e2 + e3 − e4) The Kaijser-Varopoulos contractions are the commuting 5× 5 matrices T1, T2, T3 defined by Tj = ej+1 ⊗ e1 + e5 ⊗ vj To prove the proposition we must show that if A1, A2, A3 are matrices which satisfy (5.1) kz1A1 + z2A2 + z3A3k ≤ 1 for all z ∈ Dn, then kP Aj ⊗ Tjk ≤ 1. Computing, we find (5.2) A1 ⊗ T1 + A2 ⊗ T2 + A3 ⊗ T3 = where   0 0 0 0 0 0 0 0 0 0 0 A1 0 A2 A3 0 0 B1 B2 B3 0 0 0 0 0   B1 = B2 = B3 = 1 √3 1 √3 1 √3 (−A1 + A2 + A3) (A1 − A2 + A3) (A1 + A2 − A3) The norm of the matrix (5.2) is equal to the maximum of the norms of the first column and the last row. By (5.1), we have k±A1±A2±A3k ≤ 1 for any choices of signs, so the last row of (5.2) has norm at most 1. To say that the first column has norm at most 1 amounts to saying that (5.3) I −Xj A∗ j Aj ≥ 0 24 MICHAEL T. JURY or, in fancier language, the identity map of Cn is completely contractive from MIN(ℓ1 n) to the column operator space Cn. This may be seen by averaging: by (5.1), the matrix valued function is positive semidefinite on Tn. Integrating against Lebesgue measure on Tn gives (5.3). (cid:3) I −X zizjA∗ j Ai n). We next consider the unit ball of Cn, n ≥ 2, with the 5.2. MIN(ℓ2 ℓ2 norm. Recall that the row operator space Rn and column operator space Cn are defined by embedding Cn into Mn(C) "along the first row" or "along the first column" respectively. We have R∗ n = Cn com- pletely isometrically, and thus a polynomial belongs to the unit ball of UC(Cn) if and only if it is contractive when evaluated on every row contraction, that is, if and only if it is a contractive multiplier of the Drury-Arveson space; this fact is Arveson's von Neumann inequality for row contractions [3]. 2) norm. It is known in general that kpkU C(Cn) > kpk∞ (here kpk∞ is the sup norm over Bn); probably the simplest example is p(z1, z2) = 2z1z2. The next example shows that the strict inequality persists if we replace the sup norm with the MIN(ℓ2 Proposition 5.2. Let p(z1, z2) = 2z1z2. Then kpkM IN (B2) = kpk∞ = 1 (so in particular kpkU C(C2) > kpkM IN (B2). Proof. By Theorem 1.2 and the discussion at the beginning of this section, it suffices to exhibit a contractive map σ : ℓ2 2 → B(H) and a holomorphic function F : B2 → H such that (5.4) 1 − 4z1z2w1w2 = F (z)(1 − σ(z)σ(w)∗)F ((w)∗. To do this , take H = ℓ2 σ(z1, z2) =  6; define 0 z2 z1 0 0 0  z1 z2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 z2 z1 0 0 0 0 0 0 z1 z2 0 0 0 0   (5.5) and (5.6) Clearly kσ(z1, z2)k = z12 + z22, and one may then check that (5.4) holds. F (z1, z2) =(cid:0)1 0 0 0 z1 z2(cid:1) (cid:3) UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 25 Actually, what is most interesting about this example is not that kpkR2 > kpkM IN (B2) but that kpkM IN (B2) = kpk∞. Similar factoriza- tions can be constructed in higher dimensions to show that kpkM IN (Bn) = 1 for the polynomials p(z1, . . . zn) = z2 1 + · · · + z2 n. p(z1, . . . zn) = nn/2z1z2 · · · zn, (5.7) This is interesting because these are polynomials satisfying kpk∞ = 1 that are in some sense "large"; in particular their Cayley transforms are extreme points of the space of holomorphic functions with positive real part in Bn [12, Section 19.2]. It is then natural to raise the following question, which seems quite difficult: Question 5.3. Is it the case that kpkM IN (Bn) = kpk∞ for all polyno- mials p? In the language of [10, Chapter 5], the algebra of polynomials in n variables with the operator algebra norm defined by taking the supre- mum over all commuting contractions is denoted (Pn,k · ku). Paulsen also considers the algebras A(Dn) and MAXA(A(Dn). The former is the operator algebra determined by the supremum norm, the latter is obtained by taking the supremum over all commuting contractions which satisfy von Neumann's inequality. In our notation (up to taking closures) the algebra (Pn,k · ku) is UC(MAX(ℓ1 n)). The above exam- ple shows that the norm on UC(MIN(ℓ1 n)) strictly dominates the sup norm when n ≥ 3; it follows that none of the algebras UC(MAX(ℓ1 n)), UC(MIN(ℓ1 n)), MAXA(A(Dn), A(Dn) are completely isometrically isomorphic to each other when n ≥ 3. However it is an open problem to determine if any of these are pairwise completely boundedly isomorphic. By definition chasing, the identity map on polynomials induces complete contractions UC(MAX(ℓ1 n)) → UC(MIN(ℓ1 n)) → A(Dn) and UC(MAX(ℓ1 n)) → MAXA(A(Dn) → A(Dn) The relationship between UC(MIN(ℓ1 n)) and MAXA(A(Dn) is less clear; each possesses completely contractive maps into B(H) which are not completely contractive for the other. 26 MICHAEL T. JURY 6. Further results One may view the presence of only polynomials in Theorem 1.2 as too restrictive, but the statement admits a simple modification to make it valid for arbitrary analytic functions on Ω. All that is required is to restrict the von Neumann inequality to strictly completely contractive tuples S; that is, S for which kσSkcb = r < 1. It is not hard to see that this condition implies that the Taylor spectrum of S lies in the closure of rΩ, and hence f (S) is a well-defined, bounded operator for any f holomorphic in Ω. We then have: Theorem 6.1. Let V be an n-dimensional Banach space, E an opera- tor space structure over V , and Ω = ball(V ) ⊂ Cn. For every function f holomorphic in Ω, the following are equivalent: 1) Agler-Nevanlinna factorization. There exists a Hilbert space K, a completely contractive map ψ : V → B(K), and an ana- lytic function F : Ω → B(K, CN ) such that (6.1) 1 − f (z)f (w)∗ = F (z) [IK − σ(z)σ(w)∗] F (w)∗ 2) Transfer function realization. There exists a Hilbert space K ′, a unitary transformation U : K ′ ⊕ CN → K ′ ⊕ CN of the form (6.2) (6.3) K ′ CN K ′ CN (cid:18) A B C D (cid:19) and a completely contractive map σ : V → B(K ′) so that f (z) = D + C(I − σ(z)A)−1σ(z)B. 3) von Neumann inequality. If S is a commuting n-tuple in B(K) and σS is strictly completely contractive for E ∗ (that is, kσSkcb < 1), then (6.4) kf (S)kMN ⊗B(K) ≤ 1. Proof (sketch). The "1 implies 2" and "2 implies 3" proofs are essen- tially unchanged. For "3 implies 1," fix the finite set Λ and the cone C as in the original proof. Since C is closed and IN − f (λi)f (λj)∗ is assumed to be outside of C, there exists 0 < r < 1 so that IN − fr(λi)fr(λj)∗ is still outside of C, where fr(z) := f (rz). Now continue the proof as before with fr in place of f . The GNS construction produces, as in the original proof, operators Si so that σS is completely contractive for E ∗. UNIVERSAL COMMUTATIVE OPERATOR ALGEBRAS 27 Finishing the proof shows that h(I − f (rS)f (rS)∗)G, GiH′ = h(I − fr(S)fr(S)∗)G, GiH′ = L(IN − fr(λi)fr(λj)) < 0. The operators rSi thus give a strictly completely contractive map for E ∗ and the desired contradiction. (cid:3) As is now well-understood, the equivalences in Theorem 6.1 also give rise to a Nevanlinna-Pick interpolation theorem for the Banach algebra of holomorphic functions on Ω with the norm whose unit ball is characterized by Theorem 6.1. (Extending the notation of the previous section, we will call this algebra UC ∞(E)). We state here only the most elementary scalar version; by well-known techniques the result may be extended to cover matrix-valued interpolation. Theorem 6.2. Given points λ1, . . . λN in Ω and scalars w1, . . . wN , there exists a function f ∈ UC ∞(E) satisfying f (λj) = wj for all j = 1, . . . N if and only if there exist matrices T1, . . . Tn such that σT is completely contractive for E and vectors v1, . . . vN such that (6.5) 1 − wiwj = vi[I − n Xk,l=1 λk i λl jTkT ∗ l ]v∗ j Proof. If f ∈ UC ∞(E) and f (λj) = wj, then (6.5) is simply the re- striction of (6.1) to the points λ1, . . . λN , with Tk = σ(ek). Conversely, if (6.5) holds, we set σ = σT and run the lurking isometry argument; this produces a unitary colligation such that (6.6) wj = D + C(I − σ(λj)A)−1σ(λj)B. But this transfer function realization extends to define a function f in all of Ω, and by Theorem 6.1 this f lies in UC ∞(E). (cid:3) References [1] Jim Agler and John E. McCarthy. Pick interpolation and Hilbert function spaces, volume 44 of Graduate Studies in Mathematics. American Mathemati- cal Society, Providence, RI, 2002. [2] C.-G. Ambrozie and D. Timotin. A von Neumann type inequality for certain domains in Cn. Proc. Amer. Math. Soc., 131(3):859–869 (electronic), 2003. [3] William Arveson. Subalgebras of C ∗-algebras. III. Multivariable operator the- ory. Acta Math., 181(2):159–228, 1998. [4] Joseph A. Ball and Vladimir Bolotnikov. Realization and interpolation for Schur-Agler-class functions on domains with matrix polynomial defining func- tion in Cn. J. Funct. Anal., 213(1):45–87, 2004. 28 MICHAEL T. JURY [5] Joseph A. Ball, Tavan T. Trent, and Victor Vinnikov. Interpolation and com- mutant lifting for multipliers on reproducing kernel Hilbert spaces. In Opera- tor theory and analysis (Amsterdam, 1997), volume 122 of Oper. Theory Adv. Appl., pages 89–138. Birkhauser, Basel, 2001. [6] M. J. Crabb and A. M. Davie. von Neumann's inequality for Hilbert space operators. Bull. London Math. Soc., 7:49–50, 1975. [7] Michael A. Dritschel, Stefania Marcantognini, and Scott McCullough. Inter- polation in semigroupoid algebras. J. Reine Angew. Math., 606:1–40, 2007. [8] S. W. Drury. A generalization of von Neumann's inequality to the complex ball. Proc. Amer. Math. Soc., 68(3):300–304, 1978. [9] Meghna Mittal and Vern I. Paulsen. Operator algebras of functions. J. Funct. Anal., 258(9):3195–3225, 2010. [10] Vern Paulsen. Completely bounded maps and operator algebras, volume 78 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2002. [11] Gilles Pisier. Introduction to operator space theory, volume 294 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cam- bridge, 2003. [12] Walter Rudin. Function theory in the unit ball of Cn, volume 241 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Science]. Springer-Verlag, New York, 1980. [13] 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, 16:83–100, 1974. Department of Mathematics, University of Florida, Box 118105, Gainesville, FL 32611-8105, USA E-mail address: [email protected]
1712.00278	2	1712	2017-12-13T16:55:04	Exterior convexity and classical calculus of variations	[ "math.FA", "math.AP" ]	We study the relation between various notions of exterior convexity introduced in Bandyopadhyay-Dacorogna-Sil \cite{BDS1} with the classical notions of rank one convexity, quasiconvexity and polyconvexity. To this end, we introduce a projection map, which generalizes the alternating projection for two-tensors in a new way and study the algebraic properties of this map. We conclude with a few simple consequences of this relation which yields new proofs for some of the results discussed in Bandyopadhyay-Dacorogna-Sil \cite{BDS1}.	math.FA	math	Exterior convexity and classical calculus of variations Saugata Bandyopadhyay Department of Mathematics & Statistics IISER Kolkata Mohanpur-741246, India [email protected] Swarnendu Sil Section de Math´ematiques Station 8, EPFL 1015 Lausanne, Switzerland [email protected] Abstract We study the relation between various notions of exterior convexity introduced in Bandyopadhyay-Dacorogna-Sil [1] with the classical notions of rank one convexity, quasiconvexity and polyconvexity. To this end, we introduce a projection map, which generalizes the alternating projection for two-tensors in a new way and study the algebraic properties of this map. We conclude with a few simple consequences of this relation which yields new proofs for some of the results discussed in Bandyopadhyay- Dacorogna-Sil [1]. Keywords: calculus of variations, rank one convexity, quasiconvexity, polycon- vexity, exterior convexity, exterior form, differential form. 2010 Mathematics Subject Classification: 49-XX. 1 Introduction The notion of exterior convexity introduced in Bandyopadhyay-Dacorogna-Sil [1] is of fundamental importance in calculus of variations on exterior spaces, playing a role similar to what is played by the usual notions of convexity in classical vectorial calculus of variations. However, the precise connection between these two sets of notions of convex- ity is a question of somewhat delicate balance. In this article, we explore this connection through the introduction of an appropriate projection map. While this projection map coincides with the canonical alternating projection of the 1 two-tensor fields onto the exterior two-forms, it is non-trivial in the context of higher order forms. Furthermore, the projection map has the crucial property that it projects the tensor product to the exterior product and the gradient to the exterior derivative. It also allows us to express the connection between the notions of exterior convexity and classical notions of convexity in a crisp and explicit way, which is the content of our main theorem stated as follows Theorem 1.1 Let 2 ≤ k ≤ n, f : Λk → R and π : R(cid:0) n projection map. Then the following equivalences hold k−1(cid:1)×n → Λk be the f ext. one convex ⇔ f ◦ π rank one convex. f ext. quasiconvex ⇔ f ◦ π quasiconvex. f ext. polyconvex ⇔ f ◦ π polyconvex. The aforementioned result essentially situates the circle of ideas discussed in Bandyopadhyay-Dacorogna-Sil [1] in its proper place with respect to classical calculus of variations, which is well-developed and by now, standard (cf. Da- corogna [3]). It allows us to do calculus of variations back and forth between exterior spaces and the space of matrices. In particular, some results which were directly proved in Bandyopadhya-Dacorogna-Sil [1] turn out be easy corollaries of the theorem aforementioned above, in conjunction with classical results of vectorial calculus of variations. Notable among them is the characterization theorem for ext. quasiaffine functions (compare the proof of Theorem 3.3 in Bandyopadhyay-Dacorogna-Sil [1] with that of Theorem 3.4). While in this process we do sacrifice the intrinsic character and the co-ordinate free advan- tage of a direct proof in exterior spaces, a simple proof is obtained nonetheless provided we are ready to assume the results of classical calculus of variations which are non-trivial and technical in their own right. In this article, our main goal is to prove the aforementioned theorem. While proving the first two equivalences in Theorem 1.1 is easy from the definition of the projection map, proving the third one turns out to be surprizingly difficult and is of our principal concern in this article. One of the obstacles to the proof is the burden of heavy notations. To clarify presentation and to facilitate bookkeeping, we employed a system of notations, which is explained in detail in Section 7 at the end of the article. However, once the cloud of heavy notations is cleared, the proof highlights many intricacies of the algebraic structure of alternating multilinear maps, namely the algebraic structure of determinants and minors and their interrelationship with the algebra of the wedge products which we believe should be of independent interest. The rest of the paper is organized as follows: In Section 2, we recall the definitions of exterior convexity and introduce the projection map. Section 3 states the main theorem and presents the consequences along with a charac- terization theorem and a weak lower semicontinuity result. Section 4 explores the algebraic structure of the projection map is greater detail and Section 5 is devoted to the proof of an instrumental lemma, which singles out the crux of 2 the matter. We conclude the proof of the main theorem in Section 6. Finally, the notations used throughout the article is explained in Section 7. 2 Preliminaries 2.1 Notions of exterior convexity We start by recalling the notions of exterior convexity as introduced in [1]. Definition 2.1 Let 1 ≤ k ≤ n and f : Λk → R. (i) We say that f is ext. one convex, if the function g : t → g (t) = f (ξ + t α ∧ β) is convex for every ξ ∈ Λk, α ∈ Λk−1 and β ∈ Λ1. If the function g is affine we say that f is ext. one affine. (ii) A Borel measurable and locally bounded function f is said to be ext. quasiconvex, if the inequality ZΩ f (ξ + dω) ≥ f (ξ) meas Ω holds for every bounded open set Ω ⊂ Rn, ξ ∈ Λk and ω ∈ W 1,∞ equality holds, we say that f is ext. quasiaffine. 0 (cid:0)Ω; Λk−1(cid:1). If (iii) We say that f is ext. polyconvex, if there exists a convex function F : Λk × · · · × Λ[n/k]k → R such that f (ξ) = F(cid:16)ξ, · · · , ξ[n/k](cid:17) , for all ξ ∈ Λk. If F is affine, we say that f is ext. polyaffine. There are analogous notions of interior convexity (cf. [1]). In what follows, we will discuss the case of exterior convexity only. The case of interior convexity can be derived from the case for exterior convexity by means of Hodge duality. 2.2 Projection maps To study the relationship between the notions introduced in [1] and the clas- sical notions of the vectorial calculus of variations namely rank one convexity, quasiconvexity and polyconvexity (see [3]), we will introduce a projection map. We first introduce some notations. As usual, by abuse of notations, we identify Λk (Rn) with R(cid:0)n k(cid:1). 3 Definition 2.2 (Projection map) Let 2 ≤ k ≤ n. We write a matrix Ξ ∈ R(cid:0) n k−1(cid:1)×n, the upper indices being ordered alphabetically, as Ξ = ... 1 ... Ξ1···(k−1) n Ξ1···(k−1) 1 Ξ(n−k+2)···n   =(cid:0)ΞI · · · . . . · · · Ξ(n−k+2)···n    i∈{1,··· ,n} = i(cid:1)I∈T k−1  = (Ξ1, · · · , Ξn) .  We define a linear map π : R(cid:0) n k−1(cid:1)×n → Λk (Rn) in the following way Ξ(n−k+2)···n Ξ1···(k−1) n ... π (Ξ) = nXi=1 Ξi ∧ ei, where Ξi = X1≤i1<···<ik−1≤n Ξi1···ik−1 i ei1 ∧ · · · ∧ eik−1 = XI∈T k−1 ΞI i eI. Remark 2.3 Observe that this projection map can also be written as, π(Ξ) = XI∈T k  Xj∈I sgn(j, Ij )ΞIj j   eI, see 3(vii) in Section 7 for the notations. Remark 2.4 1. Note that the map π : R(cid:0) n k−1(cid:1)×n → Λk (Rn) is onto. 2. It is easy to see that π : Rn×n → Λ2 (Rn) is given by π (ξ) = nXi=1 ξi ∧ ei = X1≤i<j≤n(cid:16)ξi j − ξj i(cid:17) ei ∧ ej, so that, with abuse of notation, 2 (cid:19) . π (ξ) = ξ − ξT = 2(cid:18) ξ − ξT So for k = 2, π is just twice the alternating projection for 2-tensors (or twice the skew-symmetric projection for square matrices). 4 3 Main theorem and consequences 3.1 Main theorem The main result of the article is the following: Theorem 3.1 Let 2 ≤ k ≤ n, f : Λk → R and π : R(cid:0) n projection map. Then the following equivalences hold k−1(cid:1)×n → Λk be the f ext. one convex ⇔ f ◦ π rank one convex. f ext. quasiconvex ⇔ f ◦ π quasiconvex. f ext. polyconvex ⇔ f ◦ π polyconvex. Remark 3.2 (i) Note that the theorem does not say that any quasiconvex or rank one convex function φ : R(cid:0) n k−1(cid:1)×n → R is of the form f ◦ π with f ext. quasiconvex or ext. one convex as the following example shows. We let n = k = 2 and φ (Ξ) = det Ξ which is clearly polyconvex (and thus quasiconvex and rank one convex). But there is no function f : Λk → R such that φ = f ◦ π. Indeed if such an f exists, we arrive at a contradiction, since setting X =(cid:18) 1 0 0 1 (cid:19) and Y =(cid:18) 0 0 0 0 (cid:19) , we have π (X) = π (Y ) = 0 and thus 1 = φ (X) = f (π (X)) = f (π (Y )) = φ (Y ) = 0. (ii) The following equivalence is, of course, trivially true f convex ⇔ f ◦ π convex. 3.2 Relations between notions of exterior convexity We now list a few simple consequences of the main theorem. Theorem 3.3 Let 1 ≤ k ≤ n and f : Λk → R. Then f convex ⇒ f ext. polyconvex ⇒ f ext. quasiconvex ⇒ f ext. one convex. Proof The result is immediate from theorem 3.1 and the classical results (cf. [3]). Another, more direct proof, without using the classical results, can be found in [1]. Theorem 3.4 Let 1 ≤ k ≤ n and f : Λk → R. The following statements are then equivalent. (i) f is ext. polyaffine. 5 (ii) f is ext. quasiaffine. (iii) f is ext. one affine. (iv) For every 0 ≤ s ≤ [n/k], there exist cs ∈ Λks such that, f (ξ) = [n/k]Xs=0 hcs; ξsi , for every ξ ∈ Λk. Proof From the definitions of ext. polyaffine functions, it is clear that The statements (i) ⇔ (iv). (i) ⇔ (ii) ⇔ (iii) follow at once from classical results (cf. Theorem 5.20 in [3]) by virtue of The- orem 3.1. For a more direct proof of the above, see [1]. See also [2] for yet another proof. Theorem 3.5 Let 1 ≤ k ≤ n, 1 < p < ∞, Ω ⊂ Rn be a bounded smooth open set and f : Λk → R be ext. quasiconvex verifying, for every ξ ∈ Λk, c1 (ξp − 1) ≤ f (ξ) ≤ c2 (ξp + 1) for some c1 , c2 > 0. if then αs ⇀ α in W 1,p(cid:0)Ω; Λk−1(cid:1) s→∞ ZΩ f (dαs) ≥ZΩ lim inf f (dα) . Proof According to Theorem 3.1, we have that f ◦ π is quasiconvex. Then classical results (see Theorem 8.4 in [3]) show that s→∞ ZΩ lim inf f (dαs) = lim inf s→∞ ZΩ f (π (∇αs)) ≥ZΩ f (π (∇α)) =ZΩ f (dα) as wished. 4 Algebraic properties of the projection We now start exploring the algebraic structure of the projection map in greater detail. The following properties are easily obtained. See [5] for a proof. Proposition 4.1 Let 2 ≤ k ≤ n and π : R(cid:0) n k−1(cid:1)×n → Λk (Rn) be the projection map. 6 (i) If α ∈ Λk−1 (Rn) ∼ R(cid:0) n k−1(cid:1) and β ∈ Λ1 (Rn) ∼ Rn, then, π (α ⊗ β) = α ∧ β. (ii) Let ω ∈ C1(cid:0)Ω; Λk−1(cid:1) , then, by abuse of notations, π (∇ω) = dω. The following result is crucial to establish the main theorem in the case of polyconvexity. See section 5.4 of [3] for the definition of adjugates and section 7 for the notations. Proposition 4.2 (Adjugate formula) If k is even, then for 2 ≤ s ≤ [n/k], [π(Ξ)]s = (s!) XI∈T sk(cid:18)XI s I sgn(J; I)(adjs Ξ) J(cid:19) eI , k−1(cid:1)o . for [n/k] < s ≤ minnn,(cid:0) n k−1(cid:1)o . for all s, 2 ≤ s ≤ minnn,(cid:0) n and If k is odd, [π(Ξ)]s = 0, [π(Ξ)]s = 0, Proof We prove only the first equality. Everything else follows by properties of the wedge power. So we prove the case when k is even and 2 ≤ s ≤ [n/k]. We prove it by induction. Step 1: To start the induction, we first prove the case when s = 2. We have, π(Ξ) = XI∈T k So,  Xj∈I sgn(j, Ij )ΞIj j   eI. [π(Ξ)]2 = π(Ξ) ∧ π(Ξ) sgn(cid:0)I 1, I 2(cid:1) Xj1∈I 1 XI 1, I 2 I 1∪I 2=I I 1∩I 2=∅ = XI∈T 2k   I 1 j1 j1 (cid:19) j1(cid:1) Ξ sgn(cid:0)j1, I 1  Xj2∈I 2 sgn(cid:0)j2, I 2 j2(cid:1) Ξ I 2 j2 j2     eI 7 Now, since k is even, we have, ( sgn([j1, I 1 j1 ], [j2, I 2 j2 ]) sgn(j1, I 1 j1 ) sgn(j2, I 2 j2 )Ξ I 1 j1 j1 Ξ I 2 j2 j2 + sgn([j1, I 2 j2 ], [j2, I 1 j1 ]) sgn(j1, I 2 j2 ) sgn(j2, I 1 j1 )Ξ (sgn(j1, I 1 j1 , j2, I 2 j2 )Ξ I 1 j1 j1 Ξ I 2 j2 + sgn(j1, I 2 j2 j2 , j2, I 1 j1 )Ξ I 1 j1 j2 Ξ I 1 j1 j2 Ξ I 2 j2 j1 )  eI j1 )(cid:19) eI I 2 j2 [π(Ξ)]2 = 2 XI∈T 2k  XI 2 2 (cid:18)XI = 2 XI∈T 2k = 2 XI∈T 2k(cid:18)XI = 2 XI∈T 2k(cid:18)XI 2 2 sgn(j1, I 1 j1 , j2, I 2 j2 )(Ξ I 1 j1 j1 Ξ I 2 j2 j2 − Ξ I 1 j1 j2 Ξ I 2 j2 j1 )(cid:19) eI sgn(j1, I 1 j1 , j2, I 2 j2 )(adj2 Ξ) I 2 j2 I 1 j1 j1j2 (cid:19) eI , which proves the case for s = 2. Step 2: We assume the result to be true for some s ≥ 2 and show that it holds for s + 1, thus completing the induction. Now we know, by Laplace expansion formula for the determinants, (adjs+1 Ξ)I 1...I s+1 j1...js+1 = = s+1Xm=1 1 s + 1 jl (−1)l+m(adjs Ξ)I 1...cIm...I s+1 ΞIm j1...bjl...js+1 s+1Xl=1 jl (−1)l+m(adjs Ξ)I 1...cIm...I s+1 ΞIm j1...bjl...js+1 s+1Xm=1 . , for any 1 ≤ l ≤ s + 1 Note that, for any 1 ≤ l, m ≤ s + 1, sgn(j1, I 1, . . . , js+1, I s+1) = (−1){(l−1)+(m−1)(k−1)} sgn(jl, I m, I l,m), where I l,m is a shorthand for the permutation (j1, I 1, . . . , js, I s) and • j1 < . . . < js and {j1, . . . , js} = {j1, . . . ,bjl, . . . , js+1}. • I 1 < . . . < I s and { I 1, . . . , I s} = {I 1, . . . ,cI m, . . . , I s+1}. Note that this means jr = jr for 1 ≤ r < l and jr = jr+1 for l ≤ r ≤ s. Similarly, I r = I r for 1 ≤ r < m and I r = I r+1 for m ≤ r ≤ s. Now since k is even, for any 1 ≤ l, m ≤ s + 1, sgn(j1, I 1, . . . , js+1, I s+1) = (−1)l+m sgn(jl, I m, I l,m) = (−1)l+m sgn(jl, I m) sgn( I lm) sgn([jl, I m], [ I l,m]). 8 Thus, sgn(j1, I 1, . . . , js+1, I s+1)(adjs+1 Ξ)I 1...I s+1 j1...js+1 = 1 (s + 1) s+1Xl,m=1 sgn([jl, I m], [ I l,m]) sgn(jl, I m)ΞIm jl sgn( I lm)(adjs Ξ)I 1...cIm...I s+1 j1...bjl...js+1 . (cid:18)XI     Hence, (s + 1)! XI∈T (s+1)k = (s + 1)! (s + 1) XI∈T (s+1)k XI s+1 s+1Xl,m=1 = (s!) XI∈T (s+1)k XI s+1 s+1Xl,m=1 = XI∈T (s+1)k   XI ′⊂I I ′∈T k   sgn(I ′, [I\I ′])(Xj∈I ′ (cid:18)s!(cid:18)X[I\I ′] s sgn(j1, I 1, . . . , js+1, I s+1)(adjs+1 Ξ)I 1...I s+1 s+1 j1...js+1(cid:19) eI  eI   eI   eI , )(cid:19)(cid:19) sgn([jl, I m], [ I l,m]) sgn(jl, I m)ΞIm jl sgn( I lm)(adjs Ξ)I 1...cIm...I s+1 j1...bjl...js+1  sgn([jl, I m], [ I l,m]) sgn(jl, I m)ΞIm jl sgn( I lm)(adjs Ξ)I 1...cIm...I s+1 j1...bjl...js+1 sgn(j, I ′ j )Ξ I ′ j j ) sgn(j1, I 1, . . . , js, I s)(adjs Ξ) I 1... I s j1...js where the last line is just a rewriting of the penultimate one. Indeed on So, we have, by induction hypothesis, bijection between the terms on the two sides of the last equality. expanding the sums the map, sending jl to j; I m to I ′ j; I 1, . . . ,cI m, . . . , I s+1 to I 1, . . . , I s respectively and j1, . . . ,bjl, . . . , js+1 to j1, . . . , js respectively is a (cid:18)XI (s + 1)! XI∈T (s+1)k sgn(I ′, [I\I ′]) in π(Ξ)  XI ′⊂I  ×  coefficient of e[I\I ′] in [π(Ξ)]s   j1...js+1(cid:19) eI  sgn(j1, I 1, . . . , js+1, I s+1)(adjs+1 Ξ)I 1...I s+1  coefficient of eI ′ = XI∈T (s+1)k   I ′∈T k eI s+1 9 = XI∈T (s+1)k(cid:16) coefficient of eI in [π(Ξ)]s+1(cid:17) eI = [π(Ξ)]s+1 . This completes the induction proving the desired result. Since we have seen that [π(Ξ)]s depends only on adjs Ξ, we are now in a position to define a linear projection for every value of s. These maps will be useful later. Definition 4.3 For every 2 ≤ s ≤ minnn,(cid:0) n tion maps πs : R(cid:0)(cid:0) n k−1(cid:1)s s(cid:1) → Λks(Rn) by the condition, πs(adjs(Ξ)) = [π(Ξ)]s for all Ξ ∈ R(cid:0) n k−1(cid:1)×n. (cid:1)×(cid:0)n k−1(cid:1)o, we define the linear projec- Remark 4.4 It is clear that this condition uniquely defines the projection maps. For the sake of consistency, we define, π1 = π and π0 is defined to be the identity map from R to R. 5 An important lemma Lemma 5.1 Let 2 ≤ k ≤ n and N =(cid:0) n g(X, d) = f (π(X)) − k−1(cid:1). Consider the function min{N,n}Xs=1 s(cid:1)×(cid:0)n hds, adjs Xi s(cid:1) for all 1 ≤ s ≤ min {N, n} and where d = (d1, . . . , dmin{N,n}), ds ∈ R(cid:0)N X ∈ RN ×n. If for a given vector d, the function X 7→ g(X, d) achieves a minimum over RN ×n, then for all 1 ≤ s ≤ min {N, n}, there exists Ds ∈ Λks such that, hds, adjs Y i = hDs, πs(adjs Y )i for all Y ∈ RN ×n. The lemma is technical and quite heavy in terms of notations. So before proceeding to prove the lemma as stated, it might be helpful to spell out the idea of the proof. The plan is always the same. In short, if the conclusion of the lemma does not hold, we can always choose a matrix X such that g(X, d) can be made to be smaller than any given real number, contradicting the hypothesis that the map X 7→ g(X, d) assumes a minimum. Proof Let us fix a vector d and assume that for this d, the function X 7→ g(X, d) achieves a minimum over RN ×n. We will first show that all adjugates with a common index between subscripts and superscripts must have zero coefficients. More precisely, we claim that, Claim 5.2 For any 2 ≤ k ≤ n and for every 1 ≤ s ≤ min {N, n}, for every J ∈ T s, I =(cid:8)I 1 . . . I s(cid:9) where I 1, . . . , I s ∈ T k−1,we have, (ds)I J = 0 whenever I ∩ J 6= ∅. 10 We prove claim 5.2, using induction over s. To start the induction, we first show the case s = 1. Let j ∈ I, where I ∈ T k−1. We choose X = λej ⊗ eI , then clearly π(X) = 0. Also, g(X, d) = f (0) − λ (d1)I j . By letting λ to +∞ and −∞ respectively, we deduce that (d1)I j = 0, since otherwise we obtain a contradiction to the fact that g achieves a finite minima. for s = p + 1. We consider (dp+1)I 1...I p+1 Now we assume that claim 5.2 holds for all 1 ≤ s ≤ p and prove the result with jl ∈ I m for some 1 ≤ l, m ≤ p + 1. Now we first order the rest of the indices (other than the common index) in subscripts and the rest of the multiindices (other than the one with the common index) in superscripts. Let I 1 < . . . < I p and j1 < . . . < jp represent the j1...jp+1 multiindices and indices in the sets(cid:8)I 1, . . . , I p+1(cid:9)\{I m} and {j1, . . . , jp+1}\{jl} respectively. Now we choose, X = λejl ⊗ eIm + pXr=1 jr ⊗ e e I r . Since jl ∈ I m, we get π(X) is independent of λ. Also, all lower order non- constant adjugates of X must contain the index jl both in subscript and in superscript and hence their coefficients are 0 by the induction hypothesis. Hence, the only non-constant adjugate of X appearing in the expression for g(X, d) is, (cid:0)adjp+1 X(cid:1)I 1...I p+1 j1...jp+1 where α is a fixed integer. Now, = (−1)α λ, g(X, d) = (−1)α+1 λ (dp+1)I 1...I p+1 j1...jp+1 + constants . Again as before, we let λ to +∞ and −∞ and we deduce, by the same argument, (dp+1)I 1...I p+1 j1...jp+1 = 0. This completes the induction and proves the claim. At this point we split the proof in two cases, the case when k is an even integer and the case when k is an odd integer. Case 1: k is even Note that, unless k = 2, it does not follow from above that ds = 0 for all s ≥ [ n k ]. The possibility that two different blocks of multiindices in the superscript have some index in common has not been ruled out. Now we will show that the coefficients of two different adjugates having the same set of indices are related in the following way: Claim 5.3 For every s ≥ 1, sgn(J; I) (ds)I J = sgn( J; I) (ds) I J , whenever J ∪ I = J ∪ I, with J, J ∈ T s , I = (cid:8)I 1 . . . I s(cid:9) = (cid:2)I 1, . . . , I s(cid:3), I = n I 1 . . . I so = [ I 1, . . . , I s], I 1, . . . , I s, I 1, . . . , I s ∈ T k−1 and J ∩ I = ∅. In 11 particular, given any U ∈ T ks, there exists a constant DU ∈ R such that, sgn(J; I) (ds)I J = DU , (5.1) for all J ∪I = U with J ∈ T s , I =(cid:8)I 1 . . . I s(cid:9) =(cid:2)I 1, . . . , I s(cid:3), I 1, . . . , I s ∈ T k−1. We will prove the claim again by induction over s. We first prove it for the case s = 1. For the case s = 1, we just need to prove, for any index j, any multindex I ∈ T k−1 such that j ∩ I = ∅, we have sgn(j, I) (d1)I I j = sgn(j, I) (d1) j , (5.2) where [j, I] = [j, I]. We choose X = λ sgn(j, I)ej ⊗eI −λ sgn(j, I)ej ⊗e I. Clearly, π(X) = 0 and this gives, g(X, d) = f (0) + λ(cid:16)sgn(j, I) (d1)I j − sgn(j, I) (d1) I j(cid:17) , where we have used claim 5.2 to deduce that (d2)[I I] [jj] −∞, we get (5.2). = 0. Letting λ to +∞ and Now we assume the result for all 1 ≤ s ≤ s0 and show it for s = s0 + 1. Let [I 1 . . . I s0+1j1 . . . js0+1] = [ I 1 . . . I s0+1j1 . . . js0+1]. Note that the sets (cid:8)I 1 . . . I s0+1j1 . . . js0+1(cid:9) andn I 1 . . . I s0+1j1 . . . js0+1o are permutations of each other, preserving an order relation given by j1 < . . . < js0+1, j1 < . . . < js0+1, I 1 < . . . < I s0+1 and I 1 < . . . < I s0+1. Thus the aforementioned sets can be related by any permutation (of k(s0 + 1) indices) that respects this order. Since any such permutation is a product of k-flips, it is enough to prove the claim in case of k-flips, cf. definition 7.2. We now assume (J, I) and ( J , I) are related by a k-flip interchanging the subscript jl with one index in the superscript block I m and keep all the other indices unchanged. Also, we assume that after the interchange, the position of the multiindex containing jl in the superscript is p and the new position of the index from the multiindex I m in the subscript is q, i.e, jl ∈ I p and jq ∈ I m. We also order the remaining indices and assume , I = [ I 1, . . . , I s0 ] = { I 1 . . . I s0} =nI 1 . . .cI m . . . I s0+1o , J = [j1 . . . js0 ] = {j1 . . . js0 } =nj1 . . .bjl . . . js0+1o + X1≤r≤s0 − λ sgn(jq, I p)e jq ⊗ e I p jr ⊗ e e Ir . and respectively. Now we choose, X = λ sgn(jl, I m)ejl ⊗ eIm 12 Note that π(X) is independent of λ. Also, all non-constant adjugates of X appearing with possibly non-zero coefficients in the expression for g(X, d) have, either jl in subscript and I m in superscript or has jq as a subscript and I p as a superscript, but never both as then they have zero coefficients by claim 5.2. Also, these adjugates occur in pairs. More precisely, for every non-constant adjugate of X appearing with possibly non-zero coefficients in the expression for g(X, d) having jl in subscript and I m in superscript, there is one having jq in subscript and I p in superscript. Let us show that, for any 1 ≤ s ≤ s0 + 1, any subset ¯Js−1 = {¯j1, . . . , ¯js−1} ⊂ J of s indices and any choice of of s − 1 multiindices ¯I 1, . . . , ¯I s−1 out of s0 multiindices I 1, . . . , I s0, we have, (adjs X)[Im, ¯I 1,..., ¯I s−1] [jl ¯Js−1] sgn([jl ¯Js−1]; [I m, ¯I 1, . . . , ¯I s−1]) = − (adjs X)[ I p, ¯I 1,..., ¯I s−1] [jq ¯Js−1] sgn([jq ¯Js−1]; [ I p, ¯I 1, . . . , ¯I s−1]) . (5.3) b1 be the position of I m in [I m, ¯I 1, . . . , ¯I s−1] and b2 be the position of I p in Let a1 be the position of jl in(cid:2)jl ¯Js−1(cid:3) , a2 be the position of jq in(cid:2)jq ¯Js−1(cid:3), h I p, ¯I 1, . . . , ¯I s−1i. Since k is even, sgn([jl ¯Js−1]; [I m, ¯I 1, . . . , ¯I s−1]) = (−1){(a1−1)+(b1−1)} sgn(jl, I m) sgn( ¯Js−1; { ¯I 1 . . . ¯I s−1}) sgn([jl, I m], [( ¯Js−1; { ¯I 1 . . . ¯I s−1})]), and sgn([jq ¯Js−1]; [ I p, ¯I 1, . . . , ¯I s−1]) = (−1){(a2−1)+(b2−1)} sgn(jq, I p) sgn( ¯Js−1; { ¯I 1 . . . ¯I s−1}) sgn([jq, I p], [( ¯Js−1; { ¯I 1 . . . ¯I s−1})]). We also have, (adjs X)[Im, ¯I 1,..., ¯I s−1] [jl ¯Js−1] = (−1)a1+b1 sgn(jl, I m)λ(cid:0)adjs−1 X(cid:1)[ ¯I 1,..., ¯Is−1] [ ¯Js−1] , and (adjs X)[ I p, ¯I 1,..., ¯I s−1] [jq ¯Js−1] = −(−1)a2+b2 sgn(jq, I p)λ(cid:0)adjs−1 X(cid:1)[ ¯I 1,..., ¯Is−1] [ ¯Js−1] . Combining the four equations above, the result follows. 13 We now finish the proof of claim 5.3. Using (5.3), we have, g(X, d) = λ I J − sgn( J; I) (ds0+1) J(cid:17)  (−1)α(cid:16)sgn(J; I) (ds0+1)I  ks,γ  + s0Xs=1Xs sgn([jl ¯Js−1]; [I m, ¯I 1, . . . , ¯I s−1]) (ds)[Im, ¯I 1,..., ¯I s−1] − sgn([jq ¯Js−1]; [ I p, ¯I 1, . . . , ¯I s−1]) (ds)[ I p, ¯I 1,..., ¯I s−1]   where Ps is a shorthand, for every 1 ≤ s ≤ s0, for the sum over all possible such choices of ¯Js−1, ¯I 1, ¯I 2, . . . , ¯I s−1 and ks,γ is a generic placeholder for the constants appearing before each term of the sum and α is an integer. + constants,   [jl ¯Js−1] [jq ¯Js−1] By the induction hypothesis, the sum on the right hand side of the above expression is 0. Hence, we obtain, g(X, d) = (−1)αλ(cid:16)sgn(J; I) (ds0+1)I J − sgn( J; I) (ds0+1) I J(cid:17) + constants . Letting λ to +∞ and −∞, the claim is proved by induction. Note that by virtue of claim 5.3, claim 5.2 now implies, that for every 1 ≤ s ≤ min {N, n} , for every J ∈ T s, I =(cid:8)I 1 . . . I s(cid:9) where I 1, . . . , I s ∈ T k−1, we have, (ds)I J = 0 whenever either I ∩ J 6= ∅ or I l ∩ I m 6= ∅ for some 1 ≤ l < m ≤ s. (5.4) Indeed, if I ∩ J 6= ∅, we are done, using claim 5.2. So let us assume I ∩ J = ∅ but I l ∩ I m 6= ∅ for some 1 ≤ l < m ≤ s. Then there exists an index i such that i ∈ I l and i ∈ I m, we consider the k-flip interchanging some index j from subscript with the index i in I l. More precisely, let J ∈ T s and I l ∈ T k−1 be such that i ∈ J, J \ {i} ⊂ J, I l \ {i} ⊂ I l and J ∪ I l = J ∪ I l, then by claim 5.3 we have, sgn(J; I) (ds)I J = sgn(cid:16) J;h I l, I 1, . . . , bI l, . . . , I si(cid:17) (ds) h I l,I 1,...,bI l,...,I si J . Since, i ∈ J and i ∈ I m, J ∩h I l, I 1, . . . ,bI l, . . . , I si 6= ∅, the right hand side of J = 0, which proves (5.4). So this now implies, above equation is 0 and so (ds)I ds = 0 for all s ≥ [ n k ]. Hence we have, using (5.1), (5.4) and proposition 4.2, hds, adjs Y i = XI∈T skXI = XI∈T skXI s s I I (ds) J (adjs Y ) J I I sgn(J; I)(ds) J sgn(J; I)(adjs Y ) J 14 = XI∈T sk 1 s! DIXI s I (s!) sgn(J; I)(adjs Y ) J = hDs, πs(adjs Y )i, DI eI , which finishes the proof when k is even. where Ds = 1 s! XI∈T sk Case 3: k is odd In this case, by proposition 4.2, it is enough to show that all coefficients of all terms, except the linear ones must be zero. As in the case above, the plan is to establish a relation between the coefficients of two different adjugates having the same set of indices. But when k is odd, the relationship is not as nice as in the even case and as such there is no general formula. However, we still have a weaker analogue of claim 5.3 for the case of k-flips. Claim 5.4 For s ≥ 1, if J, J ∈ T s, and I 1 . . . , I s, I 1, . . . , I s ∈ T k−1, where J = {j1 . . . js}, J = {j1 . . . js}, I =(cid:8)I 1 . . . I s(cid:9) =(cid:2)I 1, . . . , I s(cid:3) and I =n I 1 . . . I so = [ I 1, . . . , I s] be such that J ∩ I = ∅ and (J, I) and ( J, I) are related by a k-flip interchanging an index jl in the subscript with one from the multiindex I m in the superscript. Also, we assume that after the interchange, the position of the multiindex containing jl in the superscript is p and the new position of the index from the multiindex I m in the subscript is q , i.e , jl ∈ I p and jq ∈ I m. Then we have, sgn(J; I) (ds)I J = (−1)(m−p) sgn( J; I) (ds) I J . Since the proof of claim 5.4 is very similar to that of claim 5.3, we shall indicate only a brief sketch of the proof. Since k is odd, we deduce, sgn([jl ¯Js−1]; [I m, ¯I 1, . . . , ¯I s−1]) = (−1){(a1−1)} sgn(jl, I m) sgn( ¯Js−1; { ¯I 1 . . . ¯I s−1}) sgn([jl, I m], [( ¯Js−1; { ¯I 1 . . . ¯I s−1})]), sgn([jq ¯Js−1]; [ I p, ¯I 1, . . . , ¯I s−1]) = (−1){(a2−1)} sgn(jq, I p) sgn( ¯Js−1; { ¯I 1 . . . ¯I s−1}) sgn([jq, I p], [( ¯Js−1; { ¯I 1 . . . ¯I s−1})]), and hence, in a manner analogous to the proof of (5.3), we have, (adjs X)[Im, ¯I 1,..., ¯I s−1] [jl ¯Js−1] sgn([jl ¯Js−1]; [I m, ¯I 1, . . . , ¯I s−1]) = −(−1)(b1−b2) (adjs X)[ I p, ¯I 1,..., ¯I s−1] [jq ¯Js−1] sgn([jq ¯Js−1]; [ I p, ¯I 1, . . . , ¯I s−1]) , (5.5) 15 for any 1 ≤ s ≤ s0 + 1, any subset ¯Js−1 = {¯j1, . . . , ¯js−1} ⊂ J of s − 1 indices and any choice of of s multiindices ¯I 1, . . . , ¯I s−1 out of s0 +1 multiindices, where a1 is the position of jl in(cid:2)jl ¯Js−1(cid:3) , a2 is the position of jq in(cid:2)jq ¯Js−1(cid:3), b1 is the position of I m in [I m, ¯I 1, . . . , ¯I s−1] and b2 is the position of I p inh I p, ¯I 1, . . . , ¯I s−1i. Claim 5.4 follows from above. Note that claim 5.4 and claim 5.2 together now rule out the possibility that an adjugate with non-zero coefficient can have common indices between the blocks of multiindices in the superscript and proves ds = 0 for all s > [ n k ]. I Furthermore, by claim 5.4, the coefficients of any two adjugates (ds)I J , (ds) J such that I ∪ J = I ∪ J, can differ only by a sign. So clearly, all of them must be 0 if one of them is. So without loss of generality, we shall restrict our attention to the coefficient of a particularly ordered adjugates, one with all distinct indices in subscript and superscripts , for which j1 < . . . < js < i1 k−1 < . . . < 1 << . . . < is is k−1, henceforth referred to as the totally ordered adjugate, Hence for a given s, 2 ≤ s ≤ [ n 1 < . . . < i1 k ], and given I ∈ T ks, we shall show that, 1i1 {i1 2...i1 j1j2...js k−1}...{is k−1}{i2 = 0, 2...i2 2...is k−1} 1i2 1is (ds) (5.6) where j1 < . . . < js < i1 1 < . . . < i1 k−1 < . . . < is 1 < . . . < is k−1. To prove (5.6), we first need the following: Claim 5.5 For any 1 ≤ r ≤ k − 1, we have, (ds) 1i1 {i1 2...i1 j1j2...js r i2 r+1i2 r+2...i2 k−1}{i1 r+1i1 r+2...i1 k−1i2 1i2 2...i2 r }...{is 1is 2...is k−1} = − (ds) 1i1 2...i1 {i1 j1j2...js k−1}{i2 1i2 2...i2 k−1}...{is 1is 2...is k−1} . (5.7) We prove the claim by induction over r. The case for r = 1 follows from repeated applications of claim 5.4 as follows. Using claim 5.4 to the k-flip interchanging j1 and i1 interchanging i1 1 and finally to the k-flip interchanging j1 and i2 1, then to the k-flip 1, we get, = −(−1)s (ds) k−1}{i1 1i2 2...i2 k−1}...{is 1is 2...is k−1} = −(−1)s(−1)s−2 (ds) 1i2 {i1 2...i2 j1j2...js k−1}{i1 2i1 2...i1 k−1i2 1}...{is 1is 2...is k−1} . This proves the case for r = 1. We now assume that (5.7) is true for 1 ≤ r ≤ r0 − 1 and show the result for r = r0. To show this, it is enough to prove that for any 2 ≤ r0 ≤ k − 1, (ds) 1i1 {i1 2...i1 j1j2...js r0 −1i2 r0 i2 r0 +1...i2 k−1}{i1 r0 i1 r0 +1...i1 k−1i2 1i2 2...i2 r0 −1}...{is 1is 2...is k−1} = (ds) 1i1 {i1 2...i1 j1j2...js r0 −1i1 r0 i2 r0 +1...i2 k−1}{i1 r0 +1i1 r0 +2...i1 k−1i2 1i2 2...i2 r0 }...{is 1is 2...is k−1} . (5.8) 16 (ds) 1 and i2 {i1 2...i1 1i1 j1j2...js = (−1)s (ds) k−1}{i2 1i2 2...i2 k−1}...{is 1is 2...is k−1} k−1}{i2 1i2 2...i2 k−1}...{is 1is 2...is k−1} {j1i1 2...i1 j2...jsi1 1 {j1i1 2...i1 j2...jsi2 1 Indeed the result for r = r0 follows by combining the induction hypothesis and (5.8). The proof is similar to the case for r = 1. Indeed, by applying claim 5.4 to the k-flip interchanging j1 and i1 r0 and r0 and finally to the k-flip interchanging j1 and i2 i2 1is r0, then to the k-flip interchanging i1 r0, we deduce , k−1i2 1...i2 2...is k−1} ...i2 k−1}{i1 r0 r0 +1...i1 i1 r0 −1}...{is (ds) r0 −1i2 r0 {i1 1...i1 j1j2...js = (−1)s−1 (ds) = −(−1)s−1 (ds) r0 +1...i1 {j1i1 j2...jsi1 r0 r0+1...i1 {j1i1 j2...jsi2 r0 k−1i2 1...i2 r0 −1}{i1 1...i1 r0 −1i2 r0 ...i2 k−1}...{is 1is 2...is k−1} k−1i2 1...i2 r0−1}{i1 1...i1 r0 i2 r0+1...i2 k−1}...{is 1is 2...is k−1} = −(−1)s−1(−1)s−2 (ds) {i1 1...i1 r0 j1j2...js i2 r0+1...i2 k−1}{i1 r0 +1...i1 k−1i2 1...i2 r0 }...{is 1is 2...is k−1} . This proves (5.8)) and establishes claim 5.5. Now, using claim 5.5, in particular for r = k − 1, we obtain, (ds) 1i1 {i1 2...i1 j1j2...js k−1}{i2 1i2 2...i2 k−1}...{is 1is 2...is k−1} = − (ds) 1i1 {i1 2...i1 j1j2...js k−1}{i2 1i2 2...i2 k−1}...{is 1is 2...is k−1} . This proves (5.6) and finishes the proof of the lemma in the case when k is odd and thereby establishes lemma 5.1 in all cases. 6 Proof of the main theorem We start by recalling a result regarding ext. polyconvex functions which we will use later. See [1] (cf. Proposition 14(ii)) for the proof. Proposition 6.1 Let f : Λk → R. Then f is ext. polyconvex if and only if, for every ξ ∈ Λk and 1 ≤ s ≤ [n/k] , there exists cs = cs (ξ) ∈ Λks such that f (η) ≥ f (ξ) + [n/k]Xs=1 hcs (ξ) ; ηs − ξsi , for every η ∈ Λk. Now we are ready to prove the main theorem. Proof of Theorem 3.1 (i) Recall (cf. Proposition 4.1) that π (α ⊗ β) = α ∧ β. The rank one convexity of f ◦ π follows then at once from the ext. one convexity of f. We now prove the converse. Let ξ ∈ Λk, α ∈ Λk−1 and β ∈ Λ1; we have to show that g : t → g (t) = f (ξ + t α ∧ β) is convex. Since the map π is onto, we can find Ξ ∈ R(cid:0) n k−1(cid:1)×n so that π (Ξ) = ξ. Therefore, g (t) = f (π (Ξ) + t π (α ⊗ β)) = f (π (Ξ + t α ⊗ β)) , 17 and the convexity of g follows at once from the rank one convexity of f ◦ π. (ii) Similarly since (cf. Proposition 4.1) π (∇ω) = dω, we immediately infer the quasiconvexity of f ◦ π from the ext. quasiconvexity of f. The reverse implication follows also in the same manner. (iii) Since f is ext. polyconvex we can find, using proposition 6.1, for every α ∈ Λk and 1 ≤ s ≤ [ n k ], cs = cs (α) ∈ Λks, such that f (β) ≥ f (α) + [n/k]Xs=1 hcs (α) ; βs − αsi , for every β ∈ Λk. Appealing to the proposition 4.2 we get, for every ξ ∈ R(cid:0) n k−1(cid:1)×n, f (π (η)) ≥ f (π (ξ)) + = f (π (ξ)) + [n/k]Xs=1 [n/k]Xs=1 hcs (π (ξ)) ; [π (η)]s − [π (ξ)]si hecs (ξ) ; adjs η − adjs ξi , 5.6 in [3] . k−1(cid:1)×n, which shows that f ◦π is indeed polyconvex by theorem k−1(cid:1). Since f ◦ π is for every η ∈ R(cid:0) n We now prove the reverse implication. Take N = (cid:0) n exists ds = ds (ξ) ∈ R(cid:0)N polyconvex, we have, using theorem 5.6 in [3] again, for every ξ ∈ RN ×n, there s(cid:1)×(cid:0)n s(cid:1) for all 1 ≤ s ≤ min {N, n} such that f (π (η)) ≥ f (π (ξ)) + min{N,n}Xs=1 hds (ξ) ; adjs η − adjs ξi , (6.1) for every η ∈ RN ×n. But this means that there exists d, given by d = (d1, . . . , dmin{N,n}) such that the function X 7→ g(X, d), where g(X, d) is as defined in lemma 5.1, achieves a minima at X = ξ. Then lemma 5.1 implies, for every 1 ≤ s ≤ min {N, n}, there exists Ds ∈ Λks such that hds, adjs η − adjs ξi = hDs; πs(adjs η) − πs(adjs ξ)i , for every η ∈ RN ×n. Hence, we obtain from (6.1), for every ξ ∈ RN ×n, f (π (η)) ≥ f (π (ξ)) + [n/k]Xs=1 hDs (ξ) ; πs(adjs η) − πs(adjs ξ)i , (6.2) for every η ∈ RN ×n. Since π is onto, given any α, β ∈ Λk, we can find η, ξ ∈ RN ×n such that π(η) = β and π(ξ) = α. Now using (6.2) and the definition of 18 πs, we have, by defining cs(α) = Ds(ξ), for every α ∈ Λk, f (β) ≥ f (α) + [n/k]Xs=1 hcs (α) ; βs − αsi , for every β ∈ Λk. This proves f is ext. polyconvex and concludes the proof of the theorem. 7 Notations We gather here the notations which we will use throughout this article. 1. Let k be a nonnegative integer and n be a positive integer. • We write Λk (Rn) (or simply Λk) to denote the vector space of all → R. For k = 0, we set alternating k−linear maps f : Rn × · · · × Rn k−times {z } Λ0 (Rn) = R. Note that Λk (Rn) = {0} for k > n and, for k ≤ n, • ∧, y , h ; i and ∗ denote the exterior product, the interior product, dim(cid:0)Λk (Rn)(cid:1) =(cid:0)n k(cid:1). • If (cid:8)e1, · · · , en(cid:9) is a basis of Rn, then, identifying Λ1 with Rn, the scalar product and the Hodge star operator respectively. is a basis of Λk. An element ξ ∈ Λk (Rn) will therefore be written as (cid:8)ei1 ∧ · · · ∧ eik : 1 ≤ i1 < · · · < ik ≤ n(cid:9) ξi1i2···ik ei1 ∧ · · · ∧ eik = XI∈T k ξ = X1≤i1<···<ik≤n ξI eI where T k =(cid:8)I = (i1 , · · · , ik) ∈ Nk : 1 ≤ i1 < · · · < ik ≤ n(cid:9) . An element of T k will be referred to as a multiindex. We adopt the alphabetical order for comparing two multiindices and we do not re- serve a specific symbol for this ordering. The usual ordering symbols, when written in the context of multiindices will denote alphabetical ordering. • We write ei1 ∧ · · · ∧ceis ∧ · · · ∧ eik = ei1 ∧ · · · ∧ eis−1 ∧ eis+1 ∧ · · · ∧ eik . Similarly, c placed over a string of indices (or multiindices ) will signify the omission of the string under the c sign. 19 2. Let Ω ⊂ Rn be a bounded open set. are defined in the usual way. (cid:0)Ω; Λk(cid:1) , 1 ≤ p ≤ ∞ • The spaces C1(cid:0)Ω; Λk(cid:1) , W 1,p(cid:0)Ω; Λk(cid:1) and W 1,p • For any ω ∈ W 1,p(cid:0)Ω; Λk(cid:1) , the exterior derivative dω belongs to Lp(Ω; Λk+1) and is defined by, for all 1 ≤ i1 < · · · < ik+1 ≤ n, 0 (dω)i1···ik+1 = k+1Xj=1 (−1)j+1 ∂ωi1···ij−1ij+1···ik+1 ∂xij , 3. Notation for indices: The following system of notations will be employed throughout. (i) Single indices will be written as lower case english letters, multiindices will be written as upper case english letters. (ii) Multiindices will always be indexed by superscripts. The use of a subscript while writing a multiindex is reserved for a special purpose. See (vi) below. (iii) {i1 . . . ir} will represent the string of indices i1 . . . ir. In the same way, {I 1 . . . I r} will represent the string of multiindices obtained by writing out the multiindices in the indicated order. (iv) (i1 . . . ir) will stand for the permutation of the r indices that arranges the string {i1 . . . ir} of distinct indices in strictly increasing order. (v) [i1 . . . ir] will stand for the increasingly ordered string of indices con- sisting of the distinct indices i1, . . . , ir. However, [I 1, . . . , I r] will represent the corresponding string of distinct multiindices I 1, . . . , I r, arranged in the increasing alphabetical order, whereas [I 1 . . . I r] will represent the string of indices obtained by arranging all the distinct single indices contained in the multiindices I 1, . . . , I r in increasing order. (vi) For I ∈ T k and j ∈ I, Ij stands for the multiindex obtained by removing j from I. (vii) The symbol (J; I), where J = {j1 . . . js} is a string of s single indices and I = {I 1 . . . I s} is a string of s multiindices, I 1, . . . , I s ∈ T (k−1)s, will be reserved to denote the interlaced string (cid:8)j1I 1 . . . jsI s(cid:9). (viii) The abovementioned system of notations will be in force even when representing indices as subscripts of superscripts of different objects. 4. Flip: We shall be employing some particular permutations often. Definition 7.1 (1-flip) Let s ≥ 1, let J ∈ T s, I ∈ T l be written as, J = {j1 . . . js}, I = {i1 . . . il} with J ∩ I = ∅. Let J ∈ T s, I ∈ T l. We say that ( J, I) is obtained from (J, I) by a 1-flip interchanging jp with im, for some 1 ≤ p ≤ s, 1 ≤ m ≤ l, if J = [j1 . . . jp−1imip+1 . . . jl] and I = [i1 . . . im−1jpim+1 . . . ik] . 20 Definition 7.2 (k-flip) Let s ≥ 1, k ≥ 2. Let J ∈ T s, J = {j1 . . . js}, I = {I 1 . . . I s} = [I 1, . . . , I s], where I 1, . . . , I s ∈ T k, I r = {ir k} for all 1 ≤ r ≤ s and J ∩ I = ∅. We say that ( J , I) is obtained from (J, I) by a k-flip if there exist integers 1 ≤ m, p ≤ s and 1 ≤ q ≤ k such that, 1, . . . , ir J = [j1 . . . jp−1im q jp+1 . . . js], and I = [I 1, . . . I m−1, [ir 1 . . . ir q−1jpir q+1 . . . ir k], I m+1, . . . , I s]. Note that a k-flip can be seen as a permutation in an obvious way. 5. Notation for sum: For I ∈ T ks, where 1 ≤ k ≤ n and 1 ≤ s ≤ [ n k ], the shorthand PI s stands for the sum, XJ, I . J={j1...js}=[j1...js], I={I 1...I s}=[I 1,...,I s] J∪ I=I Acknowledgement. We have benefitted of interesting discussions with Pro- fessor Bernard Dacorogna. Part of this work was completed during visits of S. Bandyopadhyay to EPFL, whose hospitality and support is gratefully acknowl- edged. The research of S. Bandyopadhyay was partially supported by a SERB research project titled "Pullback Equation for Differential Forms". References [1] S. Bandyopadhyay, B. Dacorogna and S. Sil, Calculus of variations with differential forms, To appear in Journal of European Mathematical Society. [2] S. Bandyopadhyay and S. Sil, Characterization of functions affine in the direction of one-divisible forms, Preprint. [3] B. Dacorogna, Direct methods in the calculus of variations, volume 78 of Applied Mathematical Sciences. Springer, New York, second edition, 2008. [4] J. W. Robbin, R. C. Rogers, and B. Temple, On weak continuity and the Hodge decomposition. Trans. Amer. Math. Soc., 303(2):609 -- 618, 1987. [5] S. Sil, PhD Thesis. 21
1606.05568	1	1606	2016-06-17T15:51:37	Noncoherent uniform algebras in $\mathbb C^n$	[ "math.FA" ]	Let $\mathbf D=\bar{\mathbb D}$ be the closed unit disk in $\mathbb C$ and $\mathbf B_n=\bar{\mathbb B_n}$ the closed unit ball in $\mathbb C^n$. For a compact subset $K$ in $\mathbb C^n$ with nonempty interior, let $A(K)$ be the uniform algebra of all complex-valued continuous functions on $K$ that are holomorphic in the interior of $K$. We give short and non-technical proofs of the known facts that $A(\bar{\mathbb D}^n)$ and $A(\mathbf B_n)$ are noncoherent rings. Using, additionally, Earl's interpolation theorem in the unit disk and the existence of peak-functions, we also establish with the same method the new result that $A(K)$ is not coherent. As special cases we obtain Hickel's theorems on the noncoherence of $A(\bar\Omega)$, where $\Omega$ runs through a certain class of pseudoconvex domains in $\mathbb C^n$, results that were obtained with deep and complicated methods. Finally, using a refinement of the interpolation theorem we show that no uniformly closed subalgebra $A$ of $C(K)$ with $P(K)\subseteq A\subseteq C(K)$ is coherent provided the polynomial convex hull of $K$ has no isolated points.	math.FA	math	NONCOHERENT UNIFORM ALGEBRAS IN Cn RAYMOND MORTINI n Abstract. Let D = D be the closed unit disk in C and Bn = Bn the closed unit ball in Cn. For a compact subset K in Cn with nonempty interior, let A(K) be the uniform algebra of all complex-valued continuous functions on K that are holomorphic in the interior of K. We give short and non-technical proofs of the known facts that A(D ) and A(Bn) are noncoherent rings. Using, additionally, Earl's interpolation theorem in the unit disk and the existence of peak-functions, we also establish with the same method the new result that A(K) is not coherent. As special cases we obtain Hickel's theorems on the noncoherence of A(Ω), where Ω runs through a certain class of pseudoconvex domains in Cn, results that were obtained with deep and complicated methods. Finally, using a refinement of the interpolation theorem we show that no uniformly closed subalgebra A of C(K) with P (K) ⊆ A ⊆ C(K) is coherent provided the polynomial convex hull of K has no isolated points. 27.8.2018 1. Introduction In this paper we are interested in a certain algebraic property of some standard Banach algebras of holomorphic functions of several complex variables. By introducing new methods we are able to solve a fourty year old problem first considered by McVoy and Rubel in the realm of uniform algebras appearing in approximation theory and complex analysis of several variables. Let us start by recalling the notion of a coherent ring. Definition 1.1. A commutative unital ring A is said to be coherent if the intersection of any two finitely generated ideals in A is finitely generated. 1991 Mathematics Subject Classification. Primary 32A38; Secondary 46J15, 46J20, 30H05, 13J99. Key words and phrases. coherent ring, polydisk algebra, ball algebra, uniform algebras, peak-points, approximate identities. 1 2 RAYMOND MORTINI We refer the reader to the article [6] for the relevance of the property of coherence in commutative algebra. and Bn = {z = (z1, . . . , zn) ∈ Cn :Pn Definition 1.2. Let D := {z ∈ C : z < 1} be the open unit disk in C j=1 zj2 < 1} the open unit ball in Cn. Their Euclidean closures are denoted by D and Bn, respectively. For a bounded open set Ω in Cn, let H ∞(Ω) be the Banach algebra of all bounded and holomorphic functions f : Ω → C, with pointwise addition and multiplication, and the supremum norm: kf k∞ := sup z∈Ω f (z), f ∈ H ∞(Ω). For a compact set K ⊂ Cn, let A(K) be the uniform algebra of all complex-valued continuous functions on K that are holomorphic in the interior K ◦ of K. If K = Ω, then we view A(K) as a subalgebra of H ∞(Ω). If K = Dn, then A(K) is called the polydisk algebra; if K = Bn, then A(K) is the ball algebra. In the context of function algebras of holomorphic functions in the unit disk D in C, we mention [11], where it was shown that the Hardy algebra H ∞(D) is coherent, while the disk algebra A(D) isn't. For n ≥ 3, Amar [1] showed that the Hardy algebras H ∞(Dn), H ∞(Bn), the polydisk algebra A(Dn) and the ball algebra A(Bn) are not coherent. The missing n = 2 case for the bidisk algebra A(D2) (respectively the ball algebra A(B2)) follows as a special case of a general result due to Hickel [8] on the noncoherence of the algebra A(Ω) of continuous functions on Ω that are holomorphic in Ω, where Ω ⊂ Cn (n ≥ 2) is a bounded strictly pseudoconvex domain with a C ∞ boundary. But the proof in [8] is technical. To illustrate our subsequent methods, we first give a short, elegant proof of the noncoherence of A(Dn) and A(Bn). Let me mention that an entirely elementary proof, developed after this manuscript had been written in 2013, has been published in [15]. Using techniques from the theory of Banach algebras which are based on peak-functions, bounded approximate identities and Cohen's factor- ization theorem (compare with [13]), and, additionally, function theo- retic tools, as Earl's interpolation theorem for H ∞(D) in the unit disk (a refinement of Carleson's interpolation theorem) [5, p. 309], we suc- ceed to show the noncoherence of A(K) for every compact set K in Cn. Finally, by replacing Earl's theorem with a result on asymptotic in- terpolation, we can handle for compact sets K ⊆ Cn without isolated points the case of any uniformly closed algebra A with P (K) ⊆ A ⊆ NONCOHERENT UNIFORM ALGEBRAS 3 C(K), where P (K) is the smallest closed subalgebra of C(K) contain- ing the polynomials. To conclude, let me point out that the coherence of rings of sta- ble transfer functions of multidimensional systems, such as A(Bn) or A(Dn), plays a role in the stabilization problem in Control Theory via the factorization approach; see [17]. 2. Preliminaries In this section we collect some technical results which we will use in the proof of our main results. Lemma 2.1. Let A be a commutative unital ring and M an ideal in A such that M 6= A. Suppose that I is a finitely generated ideal of A which satisfies I = IM. Then there exists m ∈ M such that (1 + m)I = 0. If A has no zero divisors, then I = 0. Proof. This follows from Nakayama's lemma [10, Theorem 76]. (cid:3) Lemma 2.2. Let I be a non-finitely generated ideal in a commutative unital ring A. Suppose that a ∈ A is not a zero-divisor. Then aI is not finitely generated either. Proof. Suppose, on the contrary, that aI = (G1, . . . , Gm), for some elements G1, . . . , Gm in A. Then there exist elements F1, . . . , Fm ∈ I such that Gj = aFj, j = 1, . . . , m. We claim that I = (F1, . . . , Fm). Indeed, trivially (F1, . . . , Fm) ⊆ I. Also, for any f ∈ I, af ∈ aI = (G1, . . . , Gm) gives the existence of α1, . . . , αm ∈ A such that af = α1G1 + · · · + αmGm = α1aF1 + · · · + αmaFm. Since a is not a zero-divisor, it follows that f = α1F1 + · · · + αmFm ∈ (F1, . . . , Fm). This shows that the reverse inclusion I ⊆ (F1, . . . , Fm) is true, too. But this means that I, which coincides with (F1, . . . , Fm), is finitely generated, a contradiction. (cid:3) Here is an example that shows that the condition on a being a non- zero-divisor is necessary: Example 2.3. Let D1 and D2 be two disjoint copies of the unit disk, say D1 = {z − 0.5 < 0.5} and D2 = {z + 0.5 < 0.5}, and let A be the algebra of bounded analytic functions on D1 ∪ D2. Let S(z) = 4 RAYMOND MORTINI exp(−(1 + z)/(1 − z)) be the atomic inner function. Consider the associated elements fn of A given by if z ∈ D1 if z ∈ D2. S(z) fn(z) =(S 1/n(z) a(z) =(0 if z ∈ D1 1 if z ∈ D2. and let the function a ∈ A be defined as Then the ideal I = (f1, f2, . . . , ) generated by the functions fn in A is not finitely generated, although the ideal aI is finitely generated. Definition 2.4. Let X be a metrizable space. (1) Cb(X, C) denotes the space of bounded, complex-valued contin- uous functions on X. (2) A function algebra A on X is a uniformly closed, point separat- ing subalgebra of Cb(X, C), containing the constants. (3) A point x0 ∈ X is called a peak-point for A, if there is a function p ∈ A (called a peak-function) with p(x0) = 1 and (2.1) sup x∈X\U p(x) < 1 for every open neighborhood U of x. Note that in case X is compact, condition (2.1) is equivalent to p(x) < 1 for all x ∈ X, x 6= x0. Definition 2.5. Let A be a commutative Banach algebra (without an identity element), and M a closed ideal of A. Then a bounded sequence (en)n∈N in M is called a (strong) approximate identity for M if for all f ∈ M. lim n→∞ kenf − f k = 0 For compact spaces, the following Proposition is in [2, p. 74, Corol- lary 1.6.4]. Proposition 2.6. Let X be a metric space and x0 ∈ X a peak-point for the function algebra A on X. If p is an associated peak function, then the sequence (en) defined by en = 1 − pn is a bounded approximate identity for the maximal ideal M(x0) = {f ∈ A : f (x0) = 0}. NONCOHERENT UNIFORM ALGEBRAS 5 Proof. For the reader's convenience here is the outline: In fact, for f ∈ A, ekf − f = pkf . Let ǫ > 0. As f (x0) = 0, there is an open neighbourhood U of x0 such that f < ǫ on U. By assumption, m := sup X\U p < 1. Now choose k0 ∈ N large enough so that for k > k0, mkkf k∞ < ǫ. Thus for k > k0, ekf − f = pkf ≤(cid:26) mkkf k∞ on X \ U 1k · ǫ on U (cid:27) < ǫ. Hence kekf − f k∞ ≤ ε for k > k0. (cid:3) Our central Banach-algebraic tool will be Cohen's Factorization The- orem; see [2, p.74, Theorem 1.6.5]. Proposition 2.7. Let A be a commutative unital real or complex Ba- nach algebra, I a closed ideal of A, and suppose that I has an ap- proximate identity. Then every f ∈ I can be decomposed in a product f = gh of two functions g, h ∈ I. The main function-theoretic tool for the construction of our ideals in general uniform algebras in Cn will be the following result on asymp- totic interpolation given in [12, p. 515], with predecessors in [7] and [3]. Recall that ρ(z, w) = (z − w)/(1 − zw) is the pseudohyperbolic distance between z and w in D. Theorem 2.8. Let (an) be a thin sequence in D; that is a sequence such that the associated Blaschke product b satisfies (1 − an2)b′(an) = 1. lim n Then for any sequence (wn) ∈ ℓ∞ with supn wn ≤ 1 there exists a Blaschke product B and a sequence of positive numbers τn → 1 such that for any 0 ≤ τ ′ n → 1, the zeros of B can be chosen to be contained in the union of the pseudohyperbolic disks {z ∈ D : ρ(z, an) ≤ τ ′ n ≤ τn with τ ′ n} and such that B(an) − wn → 0. If the interpolating nodes (an) cluster only at the point 1, then the zeros of B can be chosen so that they cluster also only at 1. 6 RAYMOND MORTINI Proof. It remains to verify the assertion on the zeros of B whenever (an) clusters only at 1. Since the pseudoyperbolic disk Dρ(a, r) coincides with the Euclidean disk D(C, R) where and C = 1 − r2 1 − r2a2 a R = 1 − a2 1 − r2a2 r (see [5]) it suffices to choose τ ′ n := min{τn, rn}, where rn =s 1 −p1 − an2 an2 and to verify that in that case Rn → 0 and Cn → 1. (cid:3) 3. A sufficient criteria for noncoherence The following concept of multipliers is new and is the key for our short proofs of the noncoherence results. Definition 3.1. Let A be a function algebra on a metrizable space X and x0 ∈ X a non-isolated point 1. A function is called a multiplier for the maximal ideal S ∈ Cb(cid:0)X \ {x0}, C(cid:1) if the ideal M(x0) = {f ∈ A : f (x0) = 0}, L := LS := {f ∈ A : Sf ∈ A} coincides with M(x0) and if there exists p ∈ M(x0) such that pS is not a zero-divisor 2. As a canonical example we mention the atomic inner function S(z) = exp(cid:18)− 1 + z 1 − z(cid:19) , which is a multiplier for the maximal ideal M(1) of the disk algebra A(D). 1 This means that there is a sequence of distinct points in X converging to x0. 2 The notation Sf ∈ A is to be interpreted in the usual way that Sf : X \{x0} → C has a continuous extension F to X with F ∈ A NONCOHERENT UNIFORM ALGEBRAS 7 Theorem 3.2. Let A be a function algebra on a metrizable space X. Suppose that x0 ∈ X is a non-isolated peak-point for A and that the function S ∈ Cb(X \ {x0}, C) is a multiplier for the maximal ideal M(x0). Then A is not coherent. Proof. We shall unveil two principal ideals whose intersection is not finitely generated. By assumption, S is a multiplier for M(x0). In particular, there is a function p ∈ M(x0) so that pS ∈ A is not a zero-divisor. This implies that p is not a zero-divisor, either. Let := (p), I J := (pS), K := {pSf : f ∈ A and Sf ∈ A}, and L := {f ∈ A : Sf ∈ A}. We claim that K = I ∩ J. Trivially K ⊆ I ∩ J. On the other hand, if g ∈ I ∩ J, then there exist f, h ∈ A such that g = ph = pSf , and so Sf = h ∈ A. In other words, g ∈ K. Thus also I ∩ J ⊆ K. It remains to show that K is not finitely generated. Note that by definition, K = pSL. Moreover, since S is a multiplier for M, we have M = L. Let f ∈ L. Since M has an approximate identity (by Proposition 2.6), we may apply Cohen's factorization Theorem (Proposition 2.7) to conclude that there exists g, h ∈ M such that f = hg. Consequently, L = LM. Assuming that L is finitely generated, there exists, by Nakayama's Lemma 2.1, m ∈ M such that (1 + m)L = 0. Note that L = M. Since A is point separating, there exists for every x1 ∈ X \ {x0} a function f ∈ M = L such that f (x1) 6= 0. Hence (1 + m(x1))f (x1) = 0 implies that m(x1) = −1. Since, by assumption, x0 is not an isolated point in X, the continuity of m on X implies that m(x0) = −1; a contradiction to the fact that m ∈ M(x0). Thus we conclude that L cannot be finitely generated. Because S is a multiplier, pS ∈ M(x0). Moreover, pS is not a zero- divisor. Hence, by Lemma 2.2, K = pSL is not finitely generated either. (cid:3) In the next sections we apply Theorem 3.2 to concrete function al- gebras of several complex variables. 8 RAYMOND MORTINI 4. The noncoherence of the ball and polydisk algebra In view of Theorem 3.2, to prove the noncoherence, it suffices to unveil a peak-function and a multiplier for some distinguished maximal ideal. Theorem 4.1. The ball algebra A(Bn) is not coherent for any n = 1, 2, . . . . Proof. P (z1, . . . , zn) = 1 + z1 2 is a peak-function at (1, 0, . . . , 0) for A(Bn) (note that if 1 + z1 = 2, then z1 = 1 and the remaining coordinates z2, . . . , zn are automatically zero because (z1, . . . , zn) ∈ Bn), and S(z1, . . . , zn) = exp(cid:18)− 1 + z1 1 − z1(cid:19) is a multiplier for M(1, 0, . . . , 0). Theorem 4.2. The polydisk algebra A(Dn) is not coherent for any n = 1, 2, . . . . (cid:3) (cid:3) Proof. P1(z1, . . . , zn) =(cid:18)1 + z1 2 (cid:19) · · ·(cid:18) 1 + zn 2 (cid:19) 1 − zn(cid:19) 1 − z1(cid:19) · · · exp(cid:18)− 1 + zn 1 + z1 is a peak-function at a = (1, . . . , 1) for A(Dn) and S(z1, . . . , zn) = exp(cid:18)− is a multiplier for M(1, . . . , 1). Thus we have obtained a short proof of this result by Amar and Hickel [1, 8]. 5. The noncoherence of P (K) ⊆ A ⊆ C(K) For a compact set K ⊂ Cn, let C(K) denote the uniform algebra of complex-valued continuous functions on K, A(K) the uniform algebra of all functions continuous on K and holomorphic in K ◦ and let P (K) be the subalgebra of those functions in A(K) that can be uniformly approximated on K by holomorphic polynomials. Let us recall the following well-known result: Theorem 5.1. Let K ⊆ Cn be a compact set. Then the following assertions hold: NONCOHERENT UNIFORM ALGEBRAS 9 (1) Endowed with the usual pointwise operations 3 and the supre- mum norm kf k∞ = sup{f (z) : z ∈ K} A(K) = A(+, ·, • s , k · k∞) and P (K) = A(+, ·, • s , k · k∞) are uniformly closed point separating subalgebras of C(K). (2) Let A be A(K) or P (K). Standard maximal ideals in A are given by M(z0) := {f ∈ A : f (z0) = 0} for a uniquely determined z0 ∈ K.4 (3) The spectrum (or maximal ideal space) of P (K) coincides with (4) The Shilov-boundary, ∂A, of A is a non-void closed subset of the polynomial convex hull bK of K. ∂K. (5) The set Π(A) of peak-points for A is a non-void dense subset of ∂A. (6) For each z0 ∈ Π(A), the associated maximal ideal M(z0) has a bounded approximate identity. Proof. (1) is elementary; (2)-(5) are standard facts in the theory of uniform algebras (see for instance [2] and [4]); note that the Shilov- boundary is the closure of the set of weak-peak points and that for function algebras on metrizable spaces every weak-peak point actually (6) follows from is a peak-point ([2, p. 96]). Proposition 2.6. (cid:3) (3) is in [4, p. 67]. We note that if x0 ∈ ∂K is a peak-point for P (K), then it is a peak- point for any uniformly closed algebra A with P (K) ⊆ A ⊆ C(K). Lemma 5.2. Let K ⊆ Cn be compact with K ◦ 6= ∅. If z0 ∈ ∂(K ◦) is a peak-point for P (K ◦), then z0 is a peak-point for A(K). Proof. Let f ∈ P (K ◦) peak at z0. Then f (K ◦) ⊆ D ∪ {1} ⊆ D. Let F : Cn → C be a continuous extension of f to Cn. Since D is a retract for C, there is a retraction map r of C onto D with r(z) = z for z ∈ D. Hence the function r ◦ F is an extension of f with target space D. 3 addition +, multiplication · and multiplication • 4 Note that, in general, there are many more maximal ideals than those given by point-evaluation at points in K; even in the case where K = Ω, Ω a bounded pseudoconvex domain in Cn, n ≥ 2, every function f ∈ H ∞(Ω) (a fortiori f ∈ A(Ω)) may have a bounded holomorphic extension to a strictly larger domain Ω′ (see [9]). Hence, in that case, the spectrum of A(Ω) is strictly larger than Ω itself. by complex scalars s 10 RAYMOND MORTINI By Urysohn's Lemma in metric spaces, there is a continuous function u : Cn → [0, 1] such that {z ∈ Cn : u(z) = 1} = K ◦. Now consider φ(z) = (1 + z)/2 that maps D onto z − 1/2 ≤ 1/2. We claim that is a peak function at z0 that belongs to A(K). g := φ ◦(cid:0)u · (r ◦ F )(cid:1) : K → D ∪ {1} To see this, we note that u(z) · (r(F (z)) ∈ D for every z ∈ K. Moreover, for z ∈ K ◦, F (z) = f (z) ∈ D; hence r(F (z)) = f (z) and so u(z)r(F (z)) = f (z). Since φ and f are holomorphic, we deduce that g is holomorphic in K ◦. Thus g ∈ A(K). Now if for some z1 ∈ K, g(z1) = 1, then necessarily u(z1) r(F (z1)) = 1. Now r(F (z1)) ≤ 1; hence u(z1) = u(z1) = 1. We conclude that z1 ∈ K ◦. Therefore, as was shown previously, u(z1)r(F (z1)) = f (z1) = 1. Since f ∈ P (K ◦) peaks at z0, we finally obtain that z1 = z0. (cid:3) Definition 5.3. Let Ω ⊂ Cn be a bounded open set. For a ∈ ∂Ω and a function f ∈ H ∞(Ω), let Cl(f, a) denote the cluster set of f at a; that is Cl(f, a) is the set of all points w ∈ C such there exists a sequence (zn) in Ω such that (f (zn)) converges to w. It is obvious that Cl(f, a) is a compact, nonvoid subset of C. In fact Cl(f, a) = \0<r≤1 f (Ω ∩ B(a, r)). In the case of the polydisk or unit ball, Cl(f, a) is connected. The proof of the following fundamental Lemma was motivated by parts of the proof of [14, Theorem 3.1] concerning the pseudo-B´ezout property for P (K) and its siblings, where K ⊂ C is compact. It gives us the possibility to construct multipliers for maximal ideals. Lemma 5.4. For a compact set K ⊆ Cn, let A be a uniformly closed algebra with P (K) ⊆ A ⊆ C(K). Let x0 ∈ ∂K be a non-isolated peak- point for P (K) and p ∈ P (K) an associated peak-function. Then there exists a function S ∈ Cb(K \ {x0}) such that 0 ∈ Cl(S, x0) but Cl(S, x0) 6= {0}, and (1 − p)S ∈ A. Moreover, S is a multiplier for the maximal ideal M(x0) = {f ∈ A : f (x0) = 0}. NONCOHERENT UNIFORM ALGEBRAS 11 Proof. Case 1 We first deal with the case, where K is the closure of a domain D in Cn 5. Let (zn) ∈ K be a sequence of distinct points in K converging to x0. Then p(zn) → 1 and p(zn) ∈ D. By passing to a subsequence, if necessary, we may assume that (p(zn)) is a (thin) interpolating sequence for H ∞(D). Using Earl's interpolation theorem [5, p. 309], there is an interpolating Blaschke product B satisfying (5.1) B(p(z2n)) = 0 and B(p(z2n+1)) = δ for all n and some constant δ > 0 and such that the zeros of B cluster only at 1. Hence B ◦ p is discontinuous at x0. Now let S := B ◦ p. Since p < 1 everywhere on K \ {x0}, it follows from the fact that B is continuous on D \ {1} that S = B ◦ p is continuous on K \ {x0}. Moreover, since x0 is not an isolated point, 0 ∈ Cl(S, x0) and δ ∈ Cl(S, x0). It remains to show that (1 − p)S ∈ A and that S is the multiplier we are looking for. Let us point out that for any q ∈ C(Ω) with q(x0) = 0, the function qS = q · (B ◦ p) is continuous at x0. We claim that if q ∈ A and q(x0) = 0, then q(B ◦ p) ∈ A. To this end, consider the partial products Bn := Qn j=1 Lj of the Blaschke product B. Then Bn converges locally uniformly (in D) to B. Since Bn is analytic in a neighborhood of the P (K)-spectrum σ(p) of p, where σ(p) ⊆ D, we see that Bn ◦ p ∈ P (K) ⊆ A. Now q(Bn ◦ p) converges uniformly in K to q(B◦p). Hence q(B◦p) ∈ A. In particular, (1 − p)(B ◦ p) ∈ A. Thus we have shown that M(x0) ⊆ IS := {f ∈ A : Sf ∈ A}. To show the reverse inclusion, let f ∈ IS. Then the continuity of f and the discontinuity of S at x0 imply that f (x0) = 0. Hence f ∈ M(x0) and so IS ⊆ M(x0). Consequently IS = {f ∈ A : Sf ∈ A} = M(x0). To show that (1 − p)S is not a zero-divisor in A, we have to use the special structure of K, namely that K = D for a domain D in Cn. Note that for general K, S = B ◦ p may vanish identically on whole components of K ◦ (for example if B has a zero at p(a) and p ≡ p(a) on such a component). Now (1 − p)S is analytic on D; since its zeros are 5 This is only for the purpose of simplicity, because the tools applied in this case are more elementary than in the general case. 12 RAYMOND MORTINI isolated, we deduce that (1 − p)Sq ≡ 0 implies q ≡ 0 on D for every q ∈ A. Putting it all together, we have shown that S is a multiplier for M(x0). Case 2 Now let K ⊆ Cn be an arbitrary compact set. To avoid the phenomenon described in the last paragraph, we have to look for a multiplier S that has no zeros on K \ {x0}. It will have the form S = (1 + B) ◦ p = 1 + (B ◦ p) for some Blaschke product B whose zeros cluster only at 1. Note that B ◦ p does never take the value −1 on K \ {x0}, since B(ξ) = −1 only for ξ ∈ T \ {1} and the only unimodular value p takes, is 1. Here is now the construction of B. According to the asymptotic interpolation theorem 2.8, there is a Blaschke product B whose zeros cluster only at 1 such that (5.2) B(p(z2n)) → −1 and B(p(z2n−1)) → 1. Hence 0 ∈ Cl(S, x0) and 2 ∈ Cl(S, x0). The rest is now clear in view of the proof of Case 1, always having in mind that x0 is not an isolated point in K. (cid:3) Theorem 5.5. i) If Ω is a bounded domain in Cn, then A(Ω) is not coherent. ii) If K ⊂ Cn is compact with K ◦ 6= ∅, then A(K) is not coherent. Proof. i) By Theorem 5.1, there exists a peak-point z0 ∈ ∂Ω for A(Ω) and M(z0) has an approximate identity. Of course Ω is a compact set without isolated points. Hence, by Lemma 5.4, there is a multiplier S for M(z0). The noncoherence of A(Ω) now follows from Theorem 3.2. ii) Similar as i); just use Lemma 5.2 to get the non-isolated peak- (cid:3) point x0 for A(K). If K ◦ = ∅, then A(K) = C(K). In Section 6 we will give a charac- terization of those compacta in Cn for which C(K) is coherent. Let us also note that i) is not a special case of ii), because there are algebras of the form A(Ω) that do not belong to the class of algebras of type A(K): just take as Ω the unit disk deleted by a Cantor set (=compact and totally disconnected) of positive planar Lebesgue measure. Definition 5.6. A compact set K ⊆ Cn is called admissible if its polynomial convex hull bK does not contain any isolated points. Our final theorem contains (more or less) all the preceding ones as a special case. NONCOHERENT UNIFORM ALGEBRAS 13 Theorem 5.7. Let K ⊆ Cn be an admissible compact set and let A be a uniformly closed subalgebra of C(K) with P (K) ⊆ A ⊆ C(K). Then A is not coherent. Proof. Similar as the proof above; note that Theorem 5.1 (5) yields the desired peak-point for P (K) and Lemma 5.4 the associated multiplier. (cid:3) If we are considering algebras of a single complex variable, then we have the following refinement: Theorem 5.8. Let K ⊆ C be an infinite compact set and let A be a uniformly closed subalgebra of C(K) with P (K) ⊆ A ⊆ C(K). Then A is not coherent. now follows as in the preceding theorems. (cid:3) of K is an infinite compact set, too. This in turn implies that its Proof. The infinity of K implies that the polynomial convex hull bK topological boundary ∂bK is an infinite compact set. Hence, there exists a non-isolated point x0 ∈ ∂bK ⊆ K. Now by Mergelyan's Theorem P (bK) = R(bK) = A(bK). Using the fact that C \ bK is connected, shows that x0 is a peak-point for R(bK) = P (bK). The non-coherence Gonchar's peak-point criterium for R(K) (see [4, Corollary 4.4, p. 205]) We guess that this result can be extended to the case of several variables. 6. Noncoherence of C(K) Let K be a compact set in Cn. A general result in [16] tells us that for completely regular spaces X, C(X, R) is coherent if and only if X is basically disconnected 6. This result can be used to conclude that C(K, C) is coherent if and only if K is finite. Since our compacta K are metrizable, we would like to present, for the reader's convenience, the following independent easy proof. Theorem 6.1. If K ⊆ Cn is compact, then C(K) is coherent if and only if K is finite. Proof. Suppose that K is not finite. Then there is x0 ∈ K such that lim xn = x0 for some sequence (xn) of distinct points in K. Let E be a closed subset of K not containing x0. Then p(x) = d(x, E) d(x, E) + d(x, x0) 6 Recall that X is said to be basically disconnected if the closure of {x ∈ X : f (x) 6= 0} is open for every f ∈ C(X, R). 14 RAYMOND MORTINI is a peak-function for x0. By passing to a subsequence, if necessary, we may assume that p(xn) 6= p(xm) for n 6= m. Choose a continuous zero-free function B : [0, 1[→ ]0, 1] such that B(p(x2n)) = 1 and B(p(x2n−1)) = 1/n → 0, and let S := B ◦ p. Then S ∈ Cb(K \ {x0}). It is now straightforward to check that the continuous function (1 − p)S is not a zero-divisor and that S is a mul- tiplier for M(x0) (note that the cluster set of S at x0 is not a singleton and contains 0). Then we apply Theorem 3.2. If, on the other hand, X is finite, then C(X) is a principal ideal ring. In fact, if I ⊆ C(X) is an ideal, then we define a generator g of I by g(x) = 1 if x /∈ Z(I) and g(x) = 0 if x ∈ Z(I). Hence C(X) is trivially coherent. (cid:3) Acknowledgements. I thank Amol Sasane for many discussions on noncoherence and Peter Pflug for some valuable comments concerning extensions of bounded holomorphic functions in several variables, and for providing reference [9]. References [1] E. Amar, Non coh´erence de certains anneaux de fonctions holo- morphes, Illinois Journal of Math. 25 (1981), 68 -- 73. 2, 8 [2] A. Browder, Introduction to Function Algebras, W. A. Benjamin, New York-Amsterdam 1969. 4, 5, 9 [3] K. Dyakonov and A. Nicolau, Free interpolation by non-vanishing analytic functions. Trans. Amer. Math. Soc. 359 (2007), 4449 -- 4465. 5 [4] T.W. Gamelin, Uniform algebras, Chelsea Pub. Company, New York 1984. 9, 13 [5] J.B. Garnett, Bounded Analytic Functions, Academic Press, New York, 1981. 2, 6, 11 [6] S. Glaz. Commutative coherent rings: historical perspective and current developments, Nieuw Archief voor Wiskunde (4), 10 (1992), 37 -- 56. 2 [7] P. Gorkin and R. Mortini, Asymptotic interpolating sequences in uniform algebras, J. London Math. Soc. 67 (2003), 481 -- 498. 5 NONCOHERENT UNIFORM ALGEBRAS 15 [8] M. Hickel, Noncoh´erence de certains anneaux de fonctions holo- morphes, Illinois Journal of Mathematics, 34 (1990), 515 -- 525. 2, 8 [9] M. Jarnicki and P. Pflug, Extension of holomorphic functions, de Gruyter Expositions in Mathematics. 34. Berlin: de Gruyter. 487 p., (2000) 9, 14 [10] I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, Mass., 1970. 3 [11] W.S. McVoy and L.A. Rubel. Coherence of some rings of func- tions, Journal of Functional Analysis, 21 (1976), 76 -- 87. 2 [12] R. Mortini. Thin interpolating sequences in the disk, Archiv Math. 92 (2009), 504 -- 518. 5 [13] R. Mortini and M. von Renteln, Ideals in the Wiener algebra W +, Journal of the Australian Mathematical Society Series A, 46 (1989), 220 -- 228. 2 [14] R. Mortini and R. Rupp, The B´ezout properties for some classical function algebras, Indag. Mah. 24 (2013), 229 -- 253. 10 [15] R. Mortini and A. Sasane, Noncoherence of some rings of holo- morphic functions in several variables as an easy consequence of the one-variable case, Archiv Math. 101 (2013), 525 -- 529. 2 [16] C. Neville, When is C(X) a coherent ring? Proc. Amer. Math. Soc. 110 (1990), 505 -- 508. 13 [17] A. Quadrat, The fractional representation approach to synthesis problems: an algebraic analysis viewpoint. I. (Weakly) doubly co- prime factorizations. SIAM Journal on Control Optimization, 42 ( 2003), 266 -- 299. 3 Universit´e de Lorraine, D´epartement de Math´ematiques et In- stitut ´Elie Cartan de Lorraine, UMR 7502, Ile du Saulcy, F- 57045 Metz, France E-mail address: [email protected]
1209.0664	1	1209	2012-09-04T15:01:08	Spectra of measures and wandering vectors	[ "math.FA" ]	We present a characterization of the sets that appear as Fourier spectra of measures in terms of the existence of a strongly continuous representation of the ambient group that has a wandering vector for the given set.	math.FA	math	SPECTRA OF MEASURES AND WANDERING VECTORS DORIN ERVIN DUTKAY AND PALLE E.T. JORGENSEN Abstract. We present a characterization of the sets that appear as Fourier spectra of measures in terms of the existence of a strongly continuous representation of the ambient group that has a wandering vector for the given set. Contents 1. Introduction 2. Proof of Theorem 1.5 3. Examples References 1 3 4 7 Definition 1.1. For λ ∈ Rd, denote by 1. Introduction eλ(x) = e2πiλ·x, (x ∈ Rd) Let µ be a Borel probability measure on Rd. We say that the measure µ is spectral if there exists a set Λ in Rd, called a spectrum of µ, such that the set {eλ : λ ∈ Λ} is an orthonormal basis for L2(µ). A Lebesgue measurable subset Ω of Rd is called spectral if the renormalized Lebesgue measure on Ω is spectral. We say that Ω tiles Rd by translations if there exists a set T in Rd such that {Ω + t : t ∈ T } is a partition of Rd (up to Lebesgue measure zero). APa∈A δa is spectral, where δa is the Dirac A finite subset A of Rd is called spectral if the measure 1 measure at a. Fuglede's conjecture [Fug74] asserts that a Lebesgue measurable subset Ω of Rd is spectral if and only if it tiles Rd by translations. Tao [Tao04] found a union of cubes, in dimension 5 or higher, which is spectral but does not tile. Later, Tao's counterexample was improved by Matolcsi and his collaborators [KM06, FMM06], 2000 Mathematics Subject Classification. 42A32,05B45,43A25 . Key words and phrases. Fuglede conjecture, spectrum, unitary one-parameter groups, spectral pairs, locally compact Abelian groups, Fourier analysis, wandering vector. 1 2 DORIN ERVIN DUTKAY AND PALLE E.T. JORGENSEN to disprove Fuglede's conjecture in both directions, down to dimension 3. In dimension 1 and 2, the conjecture is still open in both directions. Lebesgue measure is not the only measure that provides examples of spectral sets. In [JP98], Jorgensen and Pedersen showed that the Hausdorff measure on a fractal Cantor set with scale 4 is also spectral and a spectrum has the form: Λ = { nXk=0 4klk : lk ∈ {0, 1}, n ∈ N}, but there are many more spectra for the same measure as shown in [DHS09]. Many more examples of fractal spectral measures have been constructed since [Str00, LW02, DJ07]. Finite spectral sets of integers are closely tied [ Lab02] to a conjecture of Coven and Meyerowitz [CM99] on translational tilings of Z. In this paper we focus on the following question: which sets appear as spectra of some measure? The main result is a characterization of spectra of measures in terms of the existence of a strongly continuous representation of the ambient group which has a wandering vector for the given set. Wandering vectors are vectors that generate orthonormal bases under the action of some system of unitary operators. They are ubiquitous throughout mathematics [HL00, HL01, Izu11, CPT11, Beg05]. Definition 1.2. Let U be a family of unitary operators acting on a Hilbert space H. We say that a vector v0 6= 0 in H is a wandering vector if {U v0 : U ∈ U} is an orthogonal family of vectors. We keep a higher level of generality and work with locally compact abelian groups. Definition 1.3. Let Γ be a locally compact abelian group and let G be its dual group (of all continuous characters); we will write G = bΓ and bG = bbΓ ≈ Γ where the isomorphism bbΓ ≈ Γ is the Pontryagin duality theorem; see [Rud90]. For a point γ ∈ Γ, write (1.1) hγ , gi = eγ(g), (g ∈ G). We say that a subset S of Γ is a spectrum for a Borel probability measure µ0 on G if the set {eγ : γ ∈ S} is an orthonormal basis for L2(µ0). Because of the interest in Fourier frames [OCS02] and since it does not affect the simplicity of the statement of our theorem, we formulate it not just for orthormal bases of exponential functions but also for frames. Definition 1.4. Let A, B > 0 and let H be a Hilbert space a family of vectors {ei : i ∈ I} in H is called a frame with bounds A, B if Akfk2 ≤Xi∈I hei , fi2 ≤ Bkfk2, (f ∈ H) A subset S of Γ is a frame spectrum with bounds A, B for a Borel probability measure µ0 on G if the set {eγ : γ ∈ S} is a frame with bounds A, B for L2(µ0). SPECTRA OF MEASURES AND WANDERING VECTORS 3 Theorem 1.5. Let S ⊂ Γ be an arbitrary subset. Then the subset S is a spectrum/frame spectrum with bounds A, B for a Borel probability measure µ0 on G if and only if there exists a triple (H, v0, U ) where H is a complex Hilbert space, v0 ∈ H, kv0k = 1 and U (·) is a strongly continuous representation of Γ on H such that {U (γ)v0 : γ ∈ S} is an orthonormal basis/frame with bounds A, B for H. Moreover, in this case µ0 can be chosen such that (1.2) hv0 , U (ξ)v0iH =ZG eξ(g) dµ0(g) for all ξ ∈ Γ and there is an isometric isomorphism W : L2(G, µ0) → H such that (1.3) W eγ = U (γ)v0 for all γ ∈ Γ. In section 3 we illustrate how this result can be used to determine spectral sets of a particular form. Our focus is on the techniques, more than on the examples themselves. As a corollary, we describe all spectral sets with 3 elements. 2. Proof of Theorem 1.5 For simplicity, we will prove Theorem 1.5 for orthonormal basis; for frames, the proof is identical, just replace the words "orthonormal basis" by the words "frame with bounds A, B". Proof. Suppose S is a spectrum for µ0. Set H = L2(G, µ0), v0 =the constant function 1 in L2(G, µ0) and take, for ξ ∈ Γ, U (ξ) on L2(G, µ0) to be the multiplication operator , i.e., (2.1) (U (ξ)f )(g) = eξ(g)f (g) (f ∈ L2(G, µ0), ξ ∈ Γ, g ∈ G) A simple check shows that all the requirements are satisfied and the isomorphism W is just the identity. Conversely, suppose (H, v0, U ) is a triple such that {U (γ)v0 : γ ∈ S} is orthonormal in H. Then by the Stone-Naimark-Ambrose-Godement theorem (the SNAG theorem [Mac92, Mac04]), there is an orthogonal projection valued measure PU defined on the Borel subsets of G, such that (2.2) Now set U (ξ) =ZG eξ(g) dPU (g) (ξ ∈ Γ) (2.3) dµ0(g) := kdPU (g)v0k2 and note that µ0 will then be a Borel probability measure on G. H We prove that (1.2) holds. Let ξ ∈ Γ. We have ZG (2.4) eξ(g) dµ0(g) =ZG =(cid:28)v0 , (cid:18)ZG H =ZG eξ(g)kdPU (g)v0k2 eξ(g) dPU (g)(cid:19) v0(cid:29) = hv0 , U (ξ)v0i . eξ(g) hv0 , dPU (g)v0i 4 DORIN ERVIN DUTKAY AND PALLE E.T. JORGENSEN We now show that there is an isometric isomorphism W : L2(G, µ0) → H that satisfies (1.3). The fact that {eγ : γ ∈ S} is an orthonormal basis will follow from this. Define W eγ = U (γ)v0 for γ ∈ Γ. We prove that the inner products are preserved by W and this shows that W can be extended to a well defined isometry from L2(G, µ0) onto H; it is onto because U (γ)v0 with γ ∈ S is an orthonormal basis for H, and it will be defined everywhere because the functions eγ, γ ∈ Γ are uniformly dense on any compact subset of G so they are dense in L2(G, µ0). But according to (2.4), we have for γ, γ′ ∈ Γ: eγ(g)eγ ′(g) dµ0(g). hU (γ)v0 , U (γ′)v0i =ZG (cid:3) Remark 2.1. In Theorem 1.5, for S to be the spectrum of some measure µ0, it is enough for the family {U (γ)v0 : γ ∈ S} to be an orthogonal basis for the closed linear span of Hv0 := {U (ξ)v0 : ξ ∈ Γ}, because, in this case Hv0 is a reducing subspace for U , and one can restrict the representation to it. Remark 2.2. For ψ continuous and compactly supported on G we have W ψ =ZΓ bψ(ξ)U (ξ)v0 dξ (2.5) bψ(ξ) =ZG Indeed, for γ ∈ Γ, we have (ξ ∈ Γ, g ∈ G) where dξ denotes the Haar measure on Γ and bψ denotes the Fourier transform, i.e., eγ(g)ψ(g) dg, and by the inversion formula, ψ(g) =ZΓ bψ(ξ)eξ(g) dξ, (cid:28)ZΓ bψ(ξ)U (ξ)v0 dξ , U (γ)v0(cid:29)H =ZΓ bψ(ξ)ZG =ZΓ bψ(ξ)hU (ξ)v0 , U (γ)v0iH dξ = (by (2.4)) eγ(g)ZΓ bψ(ξ)eξ(g) dξ dµ0(g) eξ(g)eγ(g) dµ0(g) dξ = (by Fubini) =ZG eγ(g)ψ(g) dµ0(g) = hψ , eγiL2(G,µ0) = hW ψ , W eγiH = hW ψ , U (γ)v0iH . = (by the inversion formula) =ZG But, since U (γ)v0, γ ∈ G span the entire space H, equation (2.5) follows. Remark 2.3. Let A be a finite spectral subset of Rd. Theorem 1.5 and its proof shows that one can chose a strongly continuous one parameter group (U (t))t∈Rd defined on l2(A), U (t) being the multiplication by the function et restricted to A. Thus the spectrum of U (t) is e2πit·a. 3. Examples In this section we show how our result can be used to determine spectral sets of a particular form. We urge the reader to focus more on the techniques than on the examples themselves, since we believe that these techniques can be applied to more general situations. For simplicity, we will introduce some notations and present a few techniques that we will use in this section. SPECTRA OF MEASURES AND WANDERING VECTORS 5 We will consider finite subsets A of R which we assume to be spectral. Note that, for finite sets, being spectral and being the spectrum of some measure are equivalent notions. If A has spectrum B then B has (e2πiab)a∈A,b∈B is unitary. Hence A is also a spectrum for B. to be finite, A = B and the matrix Conversely, if A is the spectrum of some measure µ0 then µ0 is atomic, supported on a finite set B, and the measures of the points in the support have to be equal (see e.g. [DL12]) and therefore A is a spectrum for 1√A B and vice-versa. By Theorem 1.5, there is a Hilbert space H, a strongly continuous one-parameter unitary group (U (t))t∈R on H and a vector v0 ∈ H such that {U (a)v0 : a ∈ A} is an orthonormal basis for H. We will identify a = va := U (a)v0, a ∈ A. We will denote by {a1, . . . , am} the linear subspace generated by the vectors a1, . . . , am. We also write a for the subspace {a}. We write {a1, . . . , am} t→ {b1, . . . , bn} if the unitary U (t) maps the subspace {a1, . . . , am} into the subspace {b1, . . . , bn}. Remark 3.1. We know that the vectors {a : a ∈ A} form an orthonormal basis for H. Suppose A = {a1, . . . , am} ∪ {a′ n}, both disjoint unions, distinct elements. Suppose {a1, . . . , am} t→ {b1, . . . , bm}. Then {a′ n} because U (t) is unitary so it maps n} and also A = {b1, . . . , bm} ∪ {b′ 1, . . . , b′ n} t→ {b′ 1, . . . , a′ 1, . . . , a′ 1, . . . , b′ orthogonal subspaces to orthogonal subspaces. 1, . . . , a′ Also, suppose we have {a1, . . . , ap} t→ {c1, . . . , ck, d1, . . . , dl} and in addition {a′ k} t→ {c1, . . . , ck}, k} are all distinct. Then {a1, . . . , ap} t→ {d1, . . . , dl}, because U (t) is unitary so where {a1, . . . , ap, a′ {a1, . . . , ap} must be mapped into the orthogonal complement of {c1, . . . , ck} in {c1, . . . , ck, d1, . . . , dl}. Definition 3.2. Let H = ⊕n i=1Vi be an orthogonal decomposition of the Hilbert space H. We say that a unitary U on H permutes the subspaces {Vi} if there exists a permutation σ of {1, . . . , n} such that U Vi = Vσ(i) for all i = 1, . . . , n. We say that U permutes non-trivially if the permutation σ is not the identity. 1, . . . , a′ Lemma 3.3. Let (U (t))t∈R be a strongly continuous one-parameter unitary group on H. Let a, b ∈ R, a, b 6= 0. Suppose U (a) and U (b) permute some subspaces ⊕n i=1Vi = H, one of them non-trivially. Then a/b is rational. Proof. Assume by contradiction that a/b is irrational. Then the set M := {ma + nb : m, n ∈ Z} is dense in R. Also, since U (ma + nb) = U (a)mU (b)n it follows that U (ma + nb) also permutes the subspaces Vi. Let σt be the permutation associated to U (t) for t ∈ M . We show that every U (t) permutes the subspaces Vi. Let t ∈ R. Approximate t by a sequence tn in M . Pick v ∈ V1 with kvk = 1. We have U (tn)v ∈ Vσtn (1) and U (tn)v → U (t)v. This implies that U (tn)v are close together for n large. But then σtn (1) and σtm (1) are equal for n, m large because, otherwise U (tn)v ∈ Vσtn (1) and U (tn)v → U (t)v would lie in orthogonal subspaces and therefore the distance between them would be √2 by Pythagora's theorem. Consequently, U (t)v must lie in the same subspace as U (tn)v for n large. Varying v, we see that U (t) permutes the subspaces {Vi}. This argument shows also that σt and σs are identical if t is close to s. But then the function t 7→ σt 6 DORIN ERVIN DUTKAY AND PALLE E.T. JORGENSEN is locally constant. Since R is connected, this means that σt is constant. But σ0 is the identity and one of σa or σb is not. The contradiction implies that a/b is rational. (cid:3) Lemma 3.4. Let S be a subset of R. Assume in addition that there exists α > 0 such that S ⊂ αZ. Suppose there exists a unitary U on a Hilbert space H and a vector v0 ∈ H such that {U sv0 : αs ∈ S} is an orthonormal basis for v0. Then there exists a measure µ0 on [0, 1 α ) such that S is a spectrum for µ0. Proof. The dual of the group αZ is T = [0, 1 α ; the duality pairing is ht , kαi = e2πitkα for t ∈ [0, 1 α ), k ∈ Z. U gives a representation of αZ on H by U (αk) = U k. According to Theorem 1.5, there is a measure µ0 on [0, 1 a ) are just restrictions of characters on R (which are the exponential functions et); therefore the result follows. (cid:3) a ) such that S is a spectrum for µ0, in this duality. But the characters on [0, 1 α ) with addition modulo 1 Proposition 3.5. Let A = {0, 1, . . . , n − 2, a} with n ∈ N, n ≥ 3 and a ∈ R, a 6= 0, 1, . . . , n − 2. Then A is spectral if and only if a is rational and in its reduced form a = p/q, with (p + q) ≡ 0 mod n. Proof. Assume that A is spectral and use the notation described above. We have 0 1→ 1, 1 1→ 2, . . . , n − 3 1→ n − 2. With the rules in Remark 3.1, we obtain that a 1→ {0, a}. Then 1 a−1→ a 1→ {0, a} so 1 a→ {0, a}. But since 0 a→ a we get with Remark 3.1 that 1 a→ 0. Then a 1−a→ 1 a→ 0 so a 1→ 0. Then only possibility for n− 2 is n − 2 1→ a. Hence, if we denote by vn−1 := va we have that vk m→ v(k+m) mod n for m ∈ Z and k ∈ {0, . . . , n − 1}. So U (m) permutes cyclically the one-dimensional subspaces generated by vk, k = 0, . . . , n − 1. 1→ v(k+1) mod n and vk Next, we compute how U (a) acts on these subspaces. We 0 a→ a. For k ∈ {0, . . . , n − 2} we have k a−k→ a k→ (k + n − 1) mod n = k − 1 so k a→ k − 1. The only remaining possibility for a is a a→ n − 1. Thus vk a→ v(k−1) mod n, for k ∈ {0, . . . , n − 1}. So U (a) permutes cyclically the subspaces vk. With Lemma 3.3 we see that a has to be rational, a = p/q, irreducible, for some p, q ∈ Z. Then qa = p so U (qa) = U (p). Then, apply this operator to the subspace of v0, since v0 → vp mod n we get that −q ≡ p mod n so p + q ≡ 0 mod n. For the converse, assume a has the given form. We will define a unitary operator U ( 1 q ) on l2({0, 1, . . . , n− 1}) as in Lemma 3.4. Since p and q are relatively prime, there exist k, l ∈ Z such that kp + lq = 1. Let δi be the canonical basis in l2({0, . . . , n − 1}). Define U ( 1 q )δi = δ(i+l−k) mod n for i = 0, . . . , n − 1. Then define U ( j → v−q mod n and v0 qa p q ) = U ( 1 We have p )j for j ∈ Z. Also U (1)δi = U ( U (a)δi = U ( 1 q 1 q )qδi = δ(i+ql−qk) mod n = δ(i+1−k(p+q)) mod n = δ(i+1) mod n. )pδi = δ(i+pl−pk) mod n = δ(i−1+l(p+q)) mod n = δ(i−1) mod n. Then, we see that {U (0)δ0, . . . , U (n− 2)δ0, U (a)δ0} = {δ0, . . . , δn−2, δn−1} is an orthonormal basis. With Lemma 3.4 we get that A is spectral. (cid:3) SPECTRA OF MEASURES AND WANDERING VECTORS 7 We consider now spectral sets A with 3 elements. The spectral property is invariant under translations and scaling. Therefore, using a translation we can assume that 0 is in A and then, rescaling we can assume that also 1 is in A. Corollary 3.6. Consider a set A with 3 elements A = {0, 1, a}. Then a is spectral if and only if a is rational and if a = p/q in its reduced form then p + q is divisible by 3. Acknowledgements. This work was done while the first named author (PJ) was visiting the University of Central Florida. We are grateful to the UCF-Math Department for hospitality and support. The authors are pleased to thank Professors Deguang Han and Qiyu Sun for helpful conversations. PJ was supported in part by the National Science Foundation, via a Univ of Iowa VIGRE grant. This work was partially supported by a grant from the Simons Foundation (#228539 to Dorin Dutkay). References [Beg05] Boris Begun. Weakly wandering vectors for abelian unitary actions. Ergodic Theory Dynam. Systems, 25(5):1471 -- 1483, 2005. [CM99] Ethan M. Coven and Aaron Meyerowitz. Tiling the integers with translates of one finite set. J. Algebra, 212(1):161 -- 174, 1999. [CPT11] Guizhen Cui, Wenjuan Peng, and Lei Tan. On the topology of wandering Julia components. Discrete Contin. Dyn. Syst., 29(3):929 -- 952, 2011. [DHS09] Dorin Ervin Dutkay, Deguang Han, and Qiyu Sun. On the spectra of a Cantor measure. Adv. Math., 221(1):251 -- 276, 2009. [DJ07] Dorin Ervin Dutkay and Palle E. T. Jorgensen. Fourier frequencies in affine iterated function systems. J. Funct. Anal., 247(1):110 -- 137, 2007. [DL12] Dorin Ervin Dutkay and Chun-Kit Lai. Uniformity of measures with fourier frames. http://arxiv.org/abs/1202.6028, 2012. [FMM06] B´alint Farkas, M´at´e Matolcsi, and P´eter M´ora. On Fuglede's conjecture and the existence of universal spectra. J. Fourier Anal. Appl., 12(5):483 -- 494, 2006. [Fug74] Bent Fuglede. Commuting self-adjoint partial differential operators and a group theoretic problem. J. Functional Analysis, 16:101 -- 121, 1974. [HL00] Deguang Han and David R. Larson. Frames, bases and group representations. Mem. Amer. Math. Soc., 147(697):x+94, 2000. [HL01] Deguang Han and D. Larson. Wandering vector multipliers for unitary groups. Trans. Amer. Math. Soc., 353(8):3347 -- 3370, 2001. [Izu11] Kou Hei Izuchi. Wandering subspaces and quasi-wandering subspaces in the Bergman space. New York J. Math., 17A:301 -- 305, 2011. Palle E. T. Jorgensen and Steen Pedersen. Dense analytic subspaces in fractal L2-spaces. J. Anal. Math., 75:185 -- 228, [JP98] 1998. [KM06] Mihail N. Kolountzakis and M´at´e Matolcsi. Complex Hadamard matrices and the spectral set conjecture. Collect. Math., (Vol. Extra):281 -- 291, 2006. [ Lab02] I. Laba. The spectral set conjecture and multiplicative properties of roots of polynomials. J. London Math. Soc. (2), 65(3):661 -- 671, 2002. [ LW02] Izabella Laba and Yang Wang. On spectral Cantor measures. J. Funct. Anal., 193(2):409 -- 420, 2002. 8 DORIN ERVIN DUTKAY AND PALLE E.T. JORGENSEN [Mac92] George W. Mackey. Harmonic analysis and unitary group representations: the development from 1927 to 1950. In L'´emergence de l'analyse harmonique abstraite (1930 -- 1950) (Paris, 1991), volume 2 of Cahiers S´em. Hist. Math. S´er. 2, pages 13 -- 42. Univ. Paris VI, Paris, 1992. [Mac04] George W. Mackey. Mathematical foundations of quantum mechanics. Dover Publications Inc., Mineola, NY, 2004. With a foreword by A. S. Wightman, Reprint of the 1963 original. [OCS02] Joaquim Ortega-Cerd`a and Kristian Seip. Fourier frames. Ann. of Math. (2), 155(3):789 -- 806, 2002. [Rud90] Walter Rudin. Fourier analysis on groups. Wiley Classics Library. John Wiley & Sons Inc., New York, 1990. Reprint of the 1962 original, A Wiley-Interscience Publication. [Str00] Robert S. Strichartz. Mock Fourier series and transforms associated with certain Cantor measures. J. Anal. Math., 81:209 -- 238, 2000. [Tao04] Terence Tao. Fuglede's conjecture is false in 5 and higher dimensions. Math. Res. Lett., 11(2-3):251 -- 258, 2004. [Dorin Ervin Dutkay] University of Central Florida, Department of Mathematics, 4000 Central Florida Blvd., P.O. Box 161364, Orlando, FL 32816-1364, U.S.A., E-mail address: [email protected] [Palle E.T. Jorgensen]University of Iowa, Department of Mathematics, 14 MacLean Hall, Iowa City, IA 52242- 1419, E-mail address: [email protected]
1702.08147	1	1702	2017-02-27T04:54:35	A Szeg\"o type theorem for truncated Toeplitz operators	[ "math.FA" ]	Truncated Toeplitz operators are compressions of multiplication operators on $L^2$ to model spaces (that is, subspaces of $H^2$ which are invariant with respect to the backward shift). For this class of operators we prove certain Szeg\"o type theorems concerning the asymptotics of their compressions to an increasing chain of finite dimensional model spaces.	math.FA	math	A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI Abstract. Truncated Toeplitz operators are compressions of multiplication operators on L2 to model spaces (that is, subspaces of H 2 which are invariant with respect to the backward shift). For this class of operators we prove certain Szego type theorems concerning the asymptotics of their compressions to an increasing chain of finite dimensional model spaces. The Toeplitz operators are compressions of multiplication operators on the space L2(T) to the Hardy space H 2; the multiplier is called the symbol of the operator. With respect to the standard exponential basis, their matrices are constant along diagonals; if we truncate such a matrix considering only its upper left finite corner, we obtain classical Toeplitz matrices. It does not come as a surprise that there are connections between the asymptotics of these Toeplitz matrices and the whole Toeplitz operator, or its symbol. A central result is Szego's strong limit theorem and its variants (see, for instance, [4] and the references within), which deal with the asymptotics of the eigenvalues of the Toeplitz matrix. On the other hand, certain generalizations of Toeplitz matrices have attracted a great deal of attention in the last decade, namely compressions of multiplication operators to subspaces of the Hardy space which are invariant under the backward shift. These "model spaces" are of the form H 2 ⊖uH 2 with u an inner function, and the compressions are called truncated Toeplitz operators. They have been formally introduced in [11]; see [8] for a more recent survey. Although classical Toeplitz matrices have often been a starting point for investigating truncated Toeplitz op- erators, the latter may exhibit surprising properties. It thus seems natural to see whether an analogue of Szego's strong limit theo- rem can be obtained in this more general context. Viewed as truncated Toeplitz operators, the Toeplitz matrices act on model spaces corresponding to the inner functions u(z) = zn, and Szego's theorem is about the asymptotical situation when n → ∞. The natural generalization is then to consider a sequence of zeros (λj) in D, and to let the truncations act on the model space corresponding to the finite Blaschke product associated to λj, 1 ≤ j ≤ n. 2010 Mathematics Subject Classification. 47B35, 30J10, 47A45. Key words and phrases. Model spaces, truncated Toeplitz operators, Szego Theorem. 1 2 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI Such a result has been obtained in [3]; it deals with the asymptotics of the de- terminant of a truncated Toeplitz operator. Let us note that in the case of classical Toeplitz operators and matrices one has different variants of Szego's Theorem, ei- ther in terms of the determinant of the truncation, or in terms of the trace, and one can pass from one to the other. However, this is no longer true in our generalized context, where the two different classes of results do not have a visible connection. The purpose of this paper is to find an analogue of the trace type Szego Theorem. We manage to obtain a complete result in the case when (λj) is not a Blaschke sequence. The Blaschke case seems to be less prone to an elegant solution, and we have only partial conclusions. The technique we use is inspired by one of the approaches to the Szego Theorem that is based on approximation by circulants (see for instance, [9]). In our case we use the analogue of circulants for truncated Toeplitz operators, namely elements in the so-called Sedlock algebras [12]. The plan of the paper is the following. After a rather extensive preliminary section that introduces the basic notions, we discuss the special case of finite di- mensional model spaces. We introduce then the Sedlock algebras in Section 3, and prove a result important in its own right, Theorem 3.2, which gives an alternate identification of these algebras. Sedlock algebras on finite dimensional model spaces are briefly discussed in Section 4, after which Section 5 develops the approximation technique based on them. The main result, Theorem 6.1, is proved in Section 6. The last section discusses through some examples the problems that appear when considering Blaschke sequences. 1. Preliminaries 1.1. Model spaces. Let H 2 be the Hardy space of square integrable functions on the circle with negative Fourier coefficients equal to 0. We recall that a model space is a subspace of H 2 which is invariant for the backwards shift, and that every such space is of the form KB = H 2 ⊖ BH 2 where B is an inner function, i.e., an element of H 2 of modulus 1 almost everywhere. We write PB for orthogonal projection from L2 onto KB. We will often make the supplementary assumption that B(0) = 0. In this case B is defined by B = z B. The reproducing kernels for KB at the points λ ∈ D are the functions kB λ (z) = 1 − B(λ)B(z) 1 − λz , z ∈ D. When B has an angular derivative in the sense of Caratheodory at a point in ζ ∈ T, that is, when the nontangential limit of B and B exist in ζ with the limit of B in ζ of module 1, then all functions in KB have a radial limit in ζ, and the corresponding ′ A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 3 reproducing kernel kB ζ belongs to KB. We have kB ζ (z) = 1 − B(ζ)B(z) 1 − ζz = ζB(ζ) B(ζ) − B(z) ζ − z , z ∈ D. We also use the fact that (1.1) kkB ζ k2 = kB ζ (ζ) = lim z→ζ 1 − B(ζ)B(z) 1 − ζz = ζB′(ζ) B(ζ) = B′(ζ). (The last equality follows from the fact that all of the above terms are positive real numbers.) 1.2. Truncated Toeplitz operators. If φ ∈ L2, then the map f 7→ PBφf is linear from the dense subspace H ∞ ∩ KB to KB. When this map is bounded, it can be extended to a bounded linear operator on the whole KB, that will be denoted by TB[φ] = PBMφKB and called a truncated Toeplitz operator. For φ(z) = z one obtains the compressed shift, which we will denote by SB. The closed linear space of all bounded truncated Toeplitz operators on KB will be denoted by TB. The function φ is called the symbol of the operator. It is known [11] that any operator in TB has a symbol φ in KB + KB, which is unique in case B(0) = 0. Such a symbol is called standard. Also, if B(0) = 0 we see that 1 ∈ KB, so that the orthogonal complement of the constants in KB equals zK B. Thus: KB + KB = KB ⊕ zK B and standard symbols can be uniquely written in the form φ+ + φ−, with φ+ ∈ KB and φ− ∈ zK B. Moreover, using the fact that for any inner function U the map h 7→ ¯z¯hU is an involution on KU one sees that {φ− : φ− ∈ zK B} = { ¯Bψ− : ψ− ∈ zK B} and so we will write our standard symbol (uniquely) as ψ+ + ¯Bψ− with ψ+ ∈ KB and ψ− ∈ zK B. For α ∈ T one defines Sα B = SB + α(1 ⊗ B). One can check easily that the B (called modified compressed shifts) are unitary. It is proved in [11] operators Sα that the modified compressed shifts belong to TB. Suppose now that a ∈ D, and define ba(z) = z−a 1−¯az ; ba(z) is an automorphism of the unit disk. We set Ba = B ◦ ba. The following result, that we will have the opportunity to use below, is Proposition 4.1 of [6]. Lemma 1.1. The formula (1.2) Ua(f ) =pb′ defines a unitary operator Ua from KB to KBa, and for any symbol φ we have a · f ◦ ba = p1 − a2 1 − ¯az f ◦ ba UaTB[φ]U ∗ a = TBa[φ ◦ ba]. 4 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI 1.3. Clark Measures. The following measures have been introduced by Clark [5]. For α ∈ T, the function α+B(z) α−B(z) has positive real part, and therefore, by Herglotz theorem, there exists a finite positive measure µB α on T such that (1.3) ℜ(cid:18) α + B(z) α − B(z)(cid:19) =ZT Pr,t(ζ)dµB α (ζ), where Pr,t(ζ) = 1−r2 case B(0) = 0 we have kµB continuous function f ζ−reit2 is the Poisson kernel. The measure is singular, and in α k = 1. A result of Alexandrov states that for every (1.4) ZT f dm =ZT(cid:18)ZT f dµα(cid:19) dm(α), where m is the normalized Lebesgue measure. We will often write µα instead of µB α when there is a single inner function B involved. We note that the Clark measures µb α are defined in the same way for any function b in the unit ball of H ∞. In particular for the zero function b = 0, for every α, µb α is the normalized Lebesgue measure. The next theorem combines results from [5] and [10]. Theorem 1.2. Suppose that B is an inner function, α ∈ T, and µα is defined by (1.3). Then any function f ∈ KB has a radial limit f ∗ almost everywhere with respect to µα, the operator V : Ku → L2(µα) defined by V f = f ∗ is unitary, and V Sα B = MzV . Suppose now that λj ∈ D, j ≥ 1. Define, for n ≥ 1, (1.5) Bn = n (− Yj=1 λj λj bλj ) (By convention, in case λj = 0, the corresponding factor will be −b0 = −z.) The following lemma is well known; we provide a proof for completeness. Lemma 1.3. Fix α ∈ T, and consider for each n the measure µBn α . α converges in the weak star topology to nor- malized Lebesgue measure. (i) If P(1 − λj) = ∞, then µBn (ii) If P(1 − λj ) < ∞, then µBn where B =Q∞ j=1(− λj α converges in the weak star topology to µB α , bλj ) is the infinite Blaschke product with zeros λj . λj Proof. The sequence (Bn)n converges uniformly on every compact set of D to the function B, where B is the zero function in case (i) and the infinite Blaschke product with zeros λj in case (ii). For 0 ≤ r < 1 and t ∈ R, the Poisson kernel is defined by Preit (ζ) = 1−r2 ζ−reit2 =Pn∈Z rne−intζn. We have α (ζ) = ℜ(cid:18) α + Bn(reit) α − Bn(reit)(cid:19) → ℜ(cid:18) α + B(reit) α − B(reit)(cid:19) =ZT Preit (ζ)dµBn ZT Preit (ζ)dµB α (ζ). A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 5 Since span{Preit; 0 ≤ r < 1, t ∈ R} is dense in the space of continuous functions, we get that (µBn α . Now the proof is finished since µB (cid:3) α )n converges in the weak star topology to µB α is the normalized Lebesgue measure when B ≡ 0. The behavior of the Clark measures with respect to change of variable by an automorphism of the disc is given by the next lemma, whose proof is a simple change of variable in (1.3). Lemma 1.4. With the above notations, we have µBa α = (b−a)∗(b′ −aµB α ). 1.4. The Cauchy transform. For the next facts about Cauchy transforms we refer to [7]. If µ is a Borel measure on T, its Cauchy transform Kµ is the analytic function on D defined by Kµ(z) =ZT 1 1 − ¯ζz dµ(ζ). The following lemma summarizes the properties of the Cauchy transform that we need. Lemma 1.5. Suppose µ is a Borel measure on T. Then: (1) Kµ ∈ H p for all 0 < p < 1. (2) Kµ ≡ 0 if and only if dµ = ¯φ dm for some φ ∈ H 1 0 , where H 1 0 = {f ∈ H 1, f (0) = 0}. Suppose Λ = (λj )j≥1, and each λj is repeated mj times, with mj a finite integer or infinite. We will denote by LΛ the linear span of zmkm+1 , with j ≥ 1 and 0 ≤ m < mj. The next lemma is probably known, but we have not found an appropriate reference. and ¯zm¯km+1 λj λj Lemma 1.6. Suppose P∞ Proof. Suppose µ is a Borel measure on T such that j=1(1 − λj ) = ∞. Then LΛ is dense in C(T). ZT zmkm+1 λj (z) dµ =ZT ¯zm¯km+1 λj (z) dµ = 0 for all j ≥ 1 and 0 ≤ m < mj. By replacing µ, if necessary, with 1 2 (µ + ¯µ) and 2i (µ − ¯µ), we may assume that µ is real. We have then K (m) 1 (λj ) = 0 for all j ≥ 1 and 0 ≤ m < mj. Since Kµ ∈ H p for 0 < p < 1, if it is not identically zero its zeros would have to satisfy the Blaschke condition. So Kµ ≡ 0, and then by Lemma 1.5 µ = ¯φ dm for some φ ∈ H 1 0 . Since µ is real, φ is real almost everywhere. Therefore φ ≡ 0 and µ ≡ 0, which finishes the proof. (cid:3) µ If Bn is defined by (1.5), then each of the functions zmkm+1 λj belongs to Bn for sufficiently large n. This remark yields the next corollary. Corollary 1.7. Suppose P∞ W∞ n=1(KBn + KBn ) is dense in C(T). j=1(1 − λj) = ∞, and Bn is defined by (1.5). Then 6 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI 2. Finite dimensional model spaces Let us suppose now that u = B is a finite Blaschke product of order N . In this case KB is of dimension N and contains only rational functions with poles outside the closed unit disc (more precisely, in the reciprocals of the zeros of B); we have dim KB = N and dim TB = 2N + 1. The measure µB α is concentrated on the roots ζα k (k = 1, . . . , N ) of the equation B(ζ) = α, and is given by the formula (2.1) µB α = N Xk=1 1 B′(ζα k ) δζ α k . The functions in KB as well as the standard symbols of TTOs have well defined values with respect to µB α . We will have the opportunity to use the next lemma, which completes (1.1). j=1 bλj . Then we have: (i) Lemma 2.1. Suppose B =Qn B′(eit) = eit B′(eit) B(eit) = d dt Arg B(eit). n 1 − λj 2 Xj=1 eit − λj 2 = 2Pn j=1(1 − λj ). (ii) For all ζ ∈ T we have B′(ζ) ≥ 1 Proof. (i) The second equality is just a computation, using the formula for B. Taking absolute values, we deduce the first equality. For the third equality, note that Arg B(eit) = ℑ log(B(eit)) = ℑ Xj=1  n Differentiating with respect to t, we obtain log bλj (eit)  . d dt n Arg B(eit) = ℑ Xj=1  i 1 − λj 2 eit − λj 2  = n Xj=1 1 − λj2 eit − λj2 as required. Part (ii) is a consequence of (i), once we note that for all ζ ∈ T and λ ∈ D we have 1 − λ ≤ ζ − λ ≤ 1 + λ, and therefore 1 − λ 1 + λ ≤ 1 − λ2 ζ − λ2 ≤ 1 + λ 1 − λ . (cid:3) For each α ∈ T the reproducing kernels kB ζ α k B has as eigenvalues ζα , k = 1, . . . , N , form an orthog- k (k = 1, . . . , N ), with onal base of KB. The operator Sα corresponding eigenvectors kB ζ α k . A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 7 k=1 νkδζ α k concentrated on the points ζα k and define νk kB ζ α k kkB ζ α k k ⊗ kB ζ α k kkB ζ α k k! . B, and therefore belongs to TB. In particular, we will the operator Dν by Consider a measure ν = PN Xk=1 Dν = (2.2) N Then Dν is a function of Sα denote ∆α ; thus B = DµB α (2.3) ∆α B := 1 k ) kB ζ α k kkB ζ α k kB ζ α k kkB ζ α k k! , ⊗ k B′(ζα N Xk=1 The next lemma gives more information about the unitary operators that have appeared in Lemma 1.1. Lemma 2.2. Suppose a ∈ D, α ∈ T, and the operator Ua : KB → KBa is defined by (1.2). (i) For each k = 1, . . . , N , Ua(kB ζ α k where ηα (ii) If ν =PN k = b−a(ζα k=1 νkδζ α k ). , then k where νa =PN (iii) We have k=1 νkδηα k ) =qb′ −a(ζα k )kBa ηα k . UaDνU ∗ a = Dνa, . Ua∆α BU ∗ a = Dνa α, where νa α = 1 b′ −a ◦ ba µBa α = b′ aµBa α . Proof. For g ∈ KBa we have hg, Ua(kB ζ α k )i = hU ∗ a g, kB ζ α k which proves (i). i = (U ∗ a g)(ζα −a(ζα k )g(ηα k ), k ) =qb′ To prove (ii), we use (i) to check the action of the left hand side operator on the reproducing kernels. Thus UaDνU ∗ a (kBa ηα k a(ηα k )UaDν(kB ζ α k ) =qb′ (iii) is a consequence of (ii): if ν = µB ) = νkqb′ α , then a(ηα k )qb′ −a(ζα k )kBa ηα k = νkkBa ηα k . 1 (Ba)′(ηα k ) · b′ −a(ζα k ) δηα k 1 B′(ζα k ) δηα k = N Xk=1 1 νa = = N N Xk=1 Xk=1 (Ba)′(ηα k ) · (b′ −a ◦ ba(ηα k ) δηα k = 1 b′ −a ◦ ba µBa α = b′ aµBa α . (cid:3) 8 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI 3. Sedlock algebras Sedlock algebras have been introduced in [12], where it was shown that they are the only algebras contained in TB. The family of Sedlock algebras is indexed by a parameter α ∈ C ∪ {∞}. We will be interested only in α ∈ T, in which case the Sedlock algebra BB α is defined to be the commutant of the unitary operator Sα B. In [12, Proposition 3.2] Sedlock characterizes the algebra BB α as the set of trun- cated Toeplitz operators with symbols of the form φ + αSB(Bzφ) = φ + αPB(Bφ), where φ ∈ KB. If B(0) = 0 we have PB(Bφ) = B(φ − φ(0)), and thus Sedlock's result yields the following Lemma. Lemma 3.1. Suppose B(0) = 0. A bounded operator A is in BB A = TB[φ + α ¯B(φ − φ(0))] for some φ ∈ KB. α if and only if Using Theorem 1.2, one can connect the two descriptions of the Sedlock algebras, as shown by the next theorem. Theorem 3.2. Let B be an arbitrary inner function satisfying B(0) = 0. Suppose T = TB[φ + α ¯B(φ − φ(0))] ∈ BB α . Then the function φ has radial limits almost everywhere with respect to µα. If we denote the limit function by φ∗, then φ∗ ∈ L∞(µα), and T = φ∗(Sα B). Proof. The first part of the statement follows from Theorem 1.2, since φ ∈ KB. The operator V : KB → L2(µα) defined by V f = f ∗ is unitary, and V Sα B = MzV . B}′, we have V T V ∗ ∈ {Mz}′, and thus V T V ∗ = Φ(Mz) = MΦ for Since BB some Φ ∈ L∞(µα). α = {Sα On the other hand, it follows from Proposition 3.2 of [12] that, when we write an operator T ∈ Bα as T = TB[φ + α ¯B(φ − φ(0))], we can identify φ as φ = T 1. So φ∗ = V φ = V T 1 = MΦV 1 = MΦ1 = Φ. Thus T = V ∗φ∗(Mz)V = φ∗(Sα B), which ends the proof of the theorem. (cid:3) Some consequences for operators in the Sedlock class BB α can be immediately deduced. Corollary 3.3. With the above notations, if we decompose µα = µc continuous and atomic parts, and the support of µa α is the sequence (ηn)n, then: α + µa α in its (i) TB[φ+α ¯B(φ−φ(0))] is compact if and only if φ∗ = 0 µc α-almost everywhere, while φ∗(ηj ) → 0. (ii) If p is a positive real number, TB[φ + α ¯B(φ − φ(0))] is in the class Cp if and only if φ∗ = 0 µc α-almost everywhere and Pn φ∗(ηn)p < ∞. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 9 (iii) If p ≥ 1 is an integer and T = TB[φ + ¯B(φ − φ(0))] ∈ Cp, then Tr T p =Xn φ∗(ηn)p. 4. Sedlock algebras and finite Blaschke products Let us suppose now that B is a finite Blaschke product such that B(0) = 0. α is given by (2.1). As noted above, for each k, k . Therefore, if φ ∈ KB, B associated to the eigenvalue ζα Then, for α ∈ T, the measure µB kB ζ α k Theorem 3.2 implies that is an eigenvector of Sα (4.1) TB[φ + α ¯B(φ − φ(0))] = Xk=1 If T = TB[φ + ¯B(φ − φ(0))] and ν =PN that N φ(ζα k ) kB ζ α k kkB ζ α k kB ζ α k kkB ζ α k k! , ⊗ k k=1 νkδζ α k , it follows from (4.1) and (2.2) (4.2) Tr(DνT p) = N Xk=1 νkφp(ζα k ) =ZT φp dν. In particular, (4.3) Tr(∆α BT p) =ZT φp dµB α and therefore, using also (1.3), (4.4) k∆α Bk1 = Tr(∆α B) =ZT dµB α = 1. We may integrate with respect to α to obtain some interesting consequences. Lemma 4.1. With the above notations, we have B′(cid:21) . k(cid:29) kB ζα k kkB ζα k ∆α ∆α ZT ZT kB ζα k kkB ζα k Bdm(α) = TB(cid:20) 1 k=1(cid:28)f, Proof. For f ∈ KB, we have f =PN B(f )dm(α) =ZT N Xk=1 =ZT ZTf, =ZT = TB(cid:20) 1 k )f, ζ k+ kB B′(cid:21) (f ). kB ζ kkB ζ dm(ζ) f (ζ)kB B′(ζ) B′(ζα 1 1 From (4.3) and Lemma 4.1 follows the next corollary. k . By (1.4), we get kB ζ α k kkB ζ α k k! dm(α) ζ α k kkB ζ α k k+ kB α (ζ)! dm(α) dµB ζ kkB ζ k (cid:3) 10 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI Corollary 4.2. With the above notations (including T = TB[φ + α ¯B(φ − φ(0))]), we have Tr(cid:18)TB(cid:20) 1 B′(cid:21) T p(cid:19) =Z φp dm. 5. Approximating by circulants In this section we will develop a procedure, in the case B(0) = 0, for deriving properties of a truncated Toeplitz operator from a closely associated element of a Sedlock algebra. The idea comes from the case of classical Toeplitz matrices (see, for instance, [9, Chapter 4]) and it is based on an enlargement of the model space. Fix α ∈ T, and let B be a finite Blaschke product with B(0) = 0 and ψ be a (general) standard symbol for a truncated Toeplitz on KB. Thus ψ ∈ KB + KB is of the form ψ = ψ+ + ¯Bψ− with ψ+ ∈ KB and ψ− ∈ zK B. Let u be an inner function that is a multiple of B; that is, u = Bv for some inner function v. Obviously, ψ is also a standard symbol for a truncated Toeplitz operator on Ku (since KB + KB ⊂ Ku + Ku). We intend to show that, if we set φ = ψ+ + αvψ− then we can have good estimates for the difference between Tu[ψ] and the operator Tu[φ + α¯u(φ − φ(0))] in Bu α. Notice that φ ∈ KB ⊂ Ku and that φ(0) = ψ+(0) (since ψ−(0) = 0). We see that: φ + α¯u(φ − φ(0)) = ψ+ + ¯αvψ− + α¯u(ψ+ + ¯αvψ− − ψ+(0)) = ψ+ + ¯Bψ− + ¯αvψ− + α¯u(ψ+ − ψ+(0)) = ψ + ¯αvψ− + α¯u(ψ+ − ψ+(0)) and thus (5.1) Tu[φ + α¯u(φ − φ(0))] − Tu[ψ] = ¯αTu[vψ−] + αTu[¯u(ψ+ − ψ+(0))]. We can now show that the norms of the two operators on the right hand side of (5.1) do not depend on which inner function u we choose. We will interpret these operators as operators on all of L2, rather than just on the subspace Ku. We use throughout the computations below the fact that ψ, ψ+, ψ− are all bounded functions. We use repeatedly that Ku = KB ⊕ BKv = Kv ⊕ vKB and that PvKB = MvPBM¯v. This gives us that: Tu[vψ−] = PuMvψ−Pu = PuMvψ−(PB + PBKv ). But, it is clear that Im(Mvψ−PBKv ) ⊂ uH 2 and so PuMvψ−PBKv = 0. Thus: Tu[vψ−] = PuMvψ−PB = (Pv + PvKB )Mvψ−PB. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 11 Now we use the fact that Im(Mvψ−PB) ⊆ vH 2 so that PvMvψ−PB = 0 to obtain that: (5.2) PuMvψ−Pu = PvKB Mvψ− PB = MvPBMψ−PB. So Tu[vψ−] is obtained by multiplying with the operator TB[ψ−] by a unitary operator. In particular, the norm (uniform or Schatten-von Neumann) of Tu[ψ−] depends only on B and ψ and is independent of the choice of v. In our analysis of the second operator, we simplify the notation by denoting ψ0 = ψ+ − ψ+(0) so that ψ0 ∈ zK B. We have PuM¯uψ0 Pu = PuM¯uψ0 (Pv + PvKB ). Now, using the fact that Kv ⊂ vH 2 Im(M¯uψ0Pv) ⊂ ¯uBvH 2 − = H 2 − and that ψ0 ∈ KB ⊂ BH 2 − we see that −. So it is clear that PuM¯uψ0Pv = 0. Thus Next we notice that: PuM¯uψ0 Pu = PuM¯uψ0PvKB . PuM¯uψ0PvKB = PuM¯uψ0 MvPBM¯v = PuM ¯Bψ0PBM¯v. Since ψ0 ∈ KB we know that ¯Bψ0 ⊂ H 2 − . But, an element of KB is a function h ∈ H 2 which is orthogonal to BH 2 and when we multiply such a function by a − we get an element h1 of L2 which is orthogonal to BH 2 (bounded) element of H 2 and is therefore orthogonal to BKv. Thus we can conclude that (5.3) PuM¯uψ0Pu = (PBKv + PB)M ¯Bψ0 PBM¯v = PBM ¯Bψ0PBM¯v. So, just as in the first case, we have obtained that any norm (uniform or Schatten- von Neumann) of Tu[¯u(ψ+ − ψ+(0))] is independent of the choice of v. We summa- rize the conclusion of our argument in the following lemma. Lemma 5.1. Suppose α ∈ T, ψ = ψ+ + ¯Bψ− ∈ KB + KB. If u = Bv for some inner function v, and φ = ψ+ + ¯αvψ−, then Tu[φ + α¯u(φ − φ(0))] ∈ Bu α and, for any 0 < p ≤ ∞, we have kTu[φ + α¯u(φ − φ(0))] − Tu[ψ]kp ≤ Cp, where Cp is a finite constant independent of v. 6. A Szego theorem for non-Blaschke sequences Suppose now that λj ∈ D, j ≥ 1. As in Section 1.3, define, for n ≥ 1, Bn = j=1(− λj The next theorem may be considered the central result of this paper; it gives a bλj ). λj Qn Szego theorem for truncated Toeplitz operators. (6.1) (6.2) Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) (TBn [ψ])p(cid:19) →ZT ψp dm. Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) g(TBn[ψ])(cid:19) →ZT g ◦ ψ dm. 12 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI Theorem 6.1. Suppose P(1 − λj ) = ∞, Bn is defined as above, ψ ∈ C(T), and p ∈ N. Then If ψ is real-valued, then for every continuous function g on [inf ψ, sup ψ] we have The proof of the theorem will use a series of intermediate results which have interest in themselves. The basic technique is introduced in the next lemma. Lemma 6.2. Fix N ∈ N, α ∈ T, assume BN (0) = 0, and denote by ζα n,k, k = 1, . . . , n, the roots of the equation Bn(z) = α. Suppose that for each n ≥ N we ; denote by Dνn the corresponding are given an atomic measure νn =Pn diagonal operator as defined by (2.2). Assume that k=1 νn k δζ α n,k (i) νn is weakly convergent to some measure ν. (ii) kDνn k → 0. Then for any ψ ∈ KBN + KBN and p ∈ N we have Tr(Dνn (TBn [ψ])p) →Z ψp dν. Proof. If ψ = ψ+ + ¯BN ψ− ∈ KBN + KBN , we define φn = ψ+ + ¯α Bn ψ− and BN φn = φn + α ¯Bn(φn − φn(0)) for n ≥ N . Applying Lemma 5.1 for B = BN and u = Bn we obtain that the trace norm kTBn[ψ]−TBn [ φn]k1 is bounded by a constant independent of n. Therefore, using condition (ii), Tr(Dνn (TBn [ψ])p) − Tr(Dνn (TBn[ φn])p) p−1 ≤ Tr(cid:16)Dνn(cid:16)TBn [ψ]k(cid:16)TBn [ψ] − TBn [ φn](cid:17) TBn [ φn]p−k−1(cid:17)(cid:17)(cid:12)(cid:12)(cid:12) ≤ M kDνnk · kTBn[ψ] − TBn[ φn]k1 → 0. Xk=0(cid:12)(cid:12)(cid:12) Now, according to (4.2), we have Tr(Dνn TBn[ φn]p) =Z φp n dνn. Since Bn(ζα n,k) = α, we have ψ(ζ(n) j ) = φn(ζ(n) j ) for all j = 1, . . . , n, and thus the last integral is R ψ dνn, which by (i) converges to R ψ dν. Next, with a suitable assumption on the symbol, we eliminate the restriction (cid:3) B(0) = 0. Lemma 6.3. Let P(1 − λj) = ∞, Bn be as above, α ∈ T fixed, and ∆α defined by (2.3). Suppose ψ ∈ (KB ) ◦ bλ1 and p ∈ N. Then + KB −λ1 N −λ1 N Bn as (6.3) Tr(∆α Bn (TBn [ψ])p) →Z ψp dm. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 13 Proof. Denote a = −λ1, so BN (−a) = 0, Then for any n ≥ N we have Ba n(0) = Bn ◦ ba(0) = Bn(−a) = 0. We have, by Lemma 1.1 and Lemma 2.2 (iii) Bn (TBn[ψ])p) = Tr((Ua∆α = Tr(Dνa,n Bn U ∗ α (TBa Tr(∆α a )(UaTBn[ψ]U ∗ n [ψ ◦ ba])p), a )p) where νa,n α = Consider the sequence of measures νa,n 1 µBa α . n b′ −a ◦ ba α . By Lemma 1.3 (i), νa,n α → 1 b′ −a ◦ ba m. On the other hand, kDνa,n α k = k∆α Bn k = sup 1≤j≤n 1 B′ n(ζ(n) j ≤ sup ζ∈T ) 1 n(ζ) B′ ≤ 2 j=1(1 − λj ) Pn , where the last inequality is a consequence of Lemma 2.1 (ii). + KBa The assumption on ψ implies that ψ ◦ ba ∈ KBa N N . Therefore, applying Lemma 6.2, Tr(∆α Bn (TBn [ψ])p) = Tr(Dνa,n α (TBa n[ψ ◦ ba])p) →Z 1 b′ −a ◦ ba (ψ ◦ ba)p dm. By Lemma 1.4, the last integral is precisely R ψp dm, which finishes the proof. (cid:3) The next step is to integrate with respect to α ∈ T in order to obtain a version of (6.1) for a restricted class of functions ψ. Lemma 6.4. Let P(1 − λj) = ∞, Bn be as above, ψ ∈ (KB and p ∈ N. Then −λ1 N + KB −λ1 N ) ◦ bλ1 , Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) (TBn [ψ])p(cid:19) →ZT ψp dm. Proof. By Lemma 6.3 we have for each α ∈ T (6.4) Tr(∆α Bn (TBn [ψ])p) →Z ψp dm. On the other hand, from Lemma 4.1 it follows that ZT Tr(∆α Bn (TBn [ψ])p)dm(α) = Tr(cid:18)(cid:18)ZT ∆α Bn dm(α)(cid:19) (TBn [ψ])p(cid:19) n(cid:21) (TBn [ψ])p(cid:19) . = Tr(cid:18)TBn(cid:20) 1 B′ Then (6.4) together with Lebesgue's dominated convergence theorem imply that Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) (TBn [ψ])p(cid:19) →ZT(cid:18)ZT ψp dm(cid:19) dm(α) =ZT ψp dm. (cid:3) 14 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI We may now go to the proof of the main theorem. Proof of Theorem 6.1. Suppose now ψ ∈ C(T) and p ∈ N. Let ǫ > 0 be given. ) ◦ bλ1 is also dense in C(T). Next Corollary 1.7 implies that Wn(KB −λ1 N −λ1 N + KB + KB we choose N ∈ N and ψN ∈ (KB where k.k∞ is the supremum norm; we assume also that kψk∞, kψN k∞ ≤ M . ) ◦ bλ1 such that kψ − ψN k∞ ≤ ǫ, −λ1 N −λ1 N Applying now Lemma 6.4 to ψN , we see that there exists n0 such that for n ≥ n0 we have Tr(TBn(cid:20) 1 B′ n(cid:21) ((TBn [ψN ])p) −Z ψp ≤ ǫ. N dm(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) Also, since kTBn[φ]k ≤ kφk∞ for any φ ∈ C(T), we have kTBn[ψN ])p − TBn [ψ])pk ≤ kTBn[ψN ]kkkTBn[ψ]p−k−1kkTBn[ψN ] − TBn [ψ]k p−1 Xk=0 ≤ pM p−1ǫ. Using (4.4), we have k∆α Bnk1 = Tr ∆α Bn = ℜ α + Bn(0) α − Bn(0) ≤ 1 + Bn(0) 1 − B(0) ≤ 1 + λ1 1 − λ1 . Therefore Bnk1dm(α) ≤ 1 + λ1 1 − λ1 , (6.5) and (cid:12)(cid:12)(cid:12)(cid:12) B′ B′ k∆α ≤ZT n(cid:21)(cid:13)(cid:13)(cid:13)(cid:13)1 (cid:13)(cid:13)(cid:13)(cid:13) TBn(cid:20) 1 n(cid:21) ((TBn [ψ])p) −Z ψp dm(cid:12)(cid:12)(cid:12)(cid:12) Tr(TBn(cid:20) 1 ≤ Tr(cid:18)TBn(cid:20) 1 n(cid:21) (TBn [ψ])p − TBn(cid:20) 1 n(cid:21) (TBn [ψN ])p −Z ψp Tr(TBn(cid:20) 1 +(cid:12)(cid:12)(cid:12) ≤ pM p−1 1 + λ1 1 − λ1 ǫ + ǫ + pM p−1ǫ. B′ B′ B′ n(cid:21) (TBn [ψN ])p(cid:19) +Z ψp N dm(cid:12)(cid:12)(cid:12) N − ψp dm The above estimate proves formula (6.1). Suppose now that ψ is real valued and g is continuous on [inf ψ, sup ψ]. Note that TBn [ψ] is then a bounded self-adjoint operator with spectrum contained in [inf ψ, sup ψ]. Let (Pk)k∈N be a sequence of polynomials that converges to g uni- formly on [inf ψ, sup ψ]. By (6.1), for every k we have Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) Pk(TBn [ψ])(cid:19) →ZT Pk ◦ ψ dm. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 15 But (cid:12)(cid:12)(cid:12)(cid:12) B′ n(cid:21) Pk(TBn[ψ])(cid:19) − Tr(cid:18)TBn(cid:20) 1 Tr(cid:18)TBn(cid:20) 1 n(cid:21)(cid:13)(cid:13)(cid:13)(cid:13)1 ≤(cid:13)(cid:13)(cid:13)(cid:13) TBn(cid:20) 1 k(Pk − g)(TBn [ψ])k kPk − gk∞ B′ ≤ 1 + λ1 1 − λ1 B′ n(cid:21) g(TBn[ψ])(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) where on the last line we have used (6.5). Since obviously letting k → ∞ proves (6.2). (cid:12)(cid:12)(cid:12)(cid:12) ZT Pk ◦ ψ − g ◦ ψ dm(cid:12)(cid:12)(cid:12)(cid:12) ≤ kPk − gk∞, (cid:3) In the classical case of Toeplitz matrices, we have Bn = zn, B′ n ≡ n, and n I. We obtain therefore a version of a classical Szego limit theorem: B ′ if Tn are truncated Toeplitz matrices corresponding to the symbol a, then TBnh 1 ni = 1 1 n Tr T p n →ZT an dm. It is interesting to compare Theorem 6.1 with the main result in [3], where a formula is obtained for the asymptotics of det TBn[ψ]. In the classical case the determinant and the trace versions of the Szego limit theorem are closely connected, and one can deduce one version from the other. This seems not to be the case for general truncated Toeplitz operators. 7. Blaschke sequences Theorem 6.1 concerns sequences that do not satisfy the Blaschke condition. It is possible to obtain Szego results for Blaschke sequences, but these are less elegant and require supplementary hypotheses. First, Lemma 6.3 is true for a Blaschke sequence, provided we add the condition k∆α Bn k → 0. Similarly, Lemma 6.4 is true if we assume that k∆α Bn k → 0 for almost all α ∈ T. As for the main Theorem 6.1, a new complication arises, since Lemma 1.6 is no longer true.. We do, however, have the following precise result; recall that the space LΛ is defined in section 1, before Lemma 1.6. Theorem 7.1. Suppose that (λj) is a Blaschke sequence (i.e. thatP∞ j=1(1−λj) < ∞) and each λj is repeated mj times, with mj a finite integer. Let Bn be given by (1.5) and p ∈ N. Suppose that, for some fixed i ≥ 1, ψ ◦ b−λi ∈ LΛ. (1) If α ∈ T and if k∆α Bn k → 0, then (7.1) Tr(∆α Bn (TBn[ψ])p) →Z ψp dµB α . 16 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI (2) If k∆α Bn k → 0 for almost all α ∈ T, then Tr(cid:18)TBn(cid:20) 1 B′ n(cid:21) (TBn [ψ])p(cid:19) →ZT ψp dm. Unfortunately, it does not seem possible to give simple conditions that would characterize those sequences for which (7.2) k∆α Bnk → 0. The next two examples suggest the intricacy of the problem. First we give an example of a Blaschke sequence for which (7.2) is satisfied. Example 7.2. Suppose (λj ) is a sequence of points in D obtained by choosing m2 evenly distributed points on each circle of radius rm = 1 − 1 m4 . We have then X(1 − λj) =Xm and therefore the sequence is Blaschke. m2 · 1 m4 < ∞, Fix m ∈ N. For any eit ∈ T there exists a j ∈ N such that λj = rm and eit − λj ≤ 10 m2 ; therefore, using Lemma 2.1 Pλj (eit) = 1 − λj2 eit − λj2 ≥ 1/m4 100/m4 = 1 100 . and so, for any eit ∈ T, Pλj (eit) ≥ 1 100 . Xλj =rm If we fix N , and choose n ≥ max{j ∈ N : λj = rm, 0 ≤ m ≤ N }, then inf{B′ n(ζ) : ζ ∈ T} ≥ N/100. So in this case inf{B′ for any α ∈ T. n(ζ) : ζ ∈ T} → ∞ when n → ∞, and thence k∆α Bn k → 0, Example 7.3. The second example exhibits the opposite behavior. We show that, for certain Blaschke sequences (λj) the hypothesis (7.2) is not satisfied, and, in the case where α = 1, λ1 = 0 and ψ = ¯z, the conclusion (7.1) is also false; that is: Tr(∆1 Bn TBn [ψ]) 6→Z ψ dµB 1 . So, let (λj ) be a real valued Blaschke sequence with λ1 = 0. We will add other hypotheses as necessary. Following the approximation procedure of Section 5 for the finite Blaschke product B2(z) = −z(−bλ2(z)) (where Bn is defined by (1.5) j=3(−bλj ). Then u = Bn = B2v, ψ+ = 0, ψ− = bλ2(z), and φ = Bn/z. Formula (5.1) then becomes and the remark after it), we set N = 2, ψ(z) = ¯z and v = Qn TBn [Bn/z + ¯Bn(Bn/z − Bn/z(0))] − TBn[¯z] = TBn [Bn/z]. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 17 From (4.3) and the fact that Bn ≡ 1 on the support of µBn 1 , we obtain Tr(cid:0)∆1 Thus, for n → ∞, Bn TBn[Bn/z + ¯Bn(Bn/z − (Bn/z)(0))](cid:1) =ZT Tr(cid:0)∆1 Bn TBn [Bn/z + ¯Bn(Bn/z − (Bn/z)(0))](cid:1) →ZT So, in order to achieve our purpose of showing that ¯z dµB 1 . Bn/z dµBn 1 =ZT ¯z dµBn 1 . we have to prove that Tr(∆1 Bn TBn [¯z]) 6→ZT ¯z dµB 1 , (7.3) Tr(∆1 Bn TBn [Bn/z]) 6→ 0. We next use [11, Section 5], where it is shown that TBn[Bn/z] = kBn 0 ⊗ kBn 0 . Denote by ζn k the solutions to Bn(ζ) = 1. Notice that, since −1 is such a solution, ζn 1 = −1. We have ∆1 Bn = n Xk=1 1 n(ζn k ) kBn ζ n k kkBn ζ n k B′ kBn ζ n k kkBn ζ n k k! = ⊗ k n Xk=1 (since kkBn ζ n k k2 = kBn ζ n k (ζn k ) = B′ n(ζn k ) by (1.1)). 1 n(ζn k )2 (kBn ζ n k B′ ⊗ kBn ζ n k ) Using the usual formulas (x ⊗ y)(z ⊗ t) = hz, yi(x ⊗ t), Tr(x ⊗ y) = hx, yi, we obtain ∆1 Bn TBn[Bn/z] = and n Xk=1 1 n(ζn k )2 hkBn 0 , kBn ζ n k B′ i(kBn ζ n k ⊗kBn 0 ) = kBn 0 (ζn k ) k )2 (kBn n(ζn B′ ζ n k ⊗kBn 0 ), n Xk=1 Tr(cid:0)∆1 Bn TBn[Bn/z](cid:1) = We have kBn 1. Thus we have arrived at the formula 0 = Bn/z, so kBn 0 (ζn n Xk=1 kBn 0 (ζn k ) k )2 hkBn n(ζn B′ k )¯ζn k ) = Bn(ζn ζ n k , kBn 0 i = kBn 0 (ζn B′ k )kBn 0 (ζn k ) n(ζn k )2 . n Xk=1 k = ¯ζn k , while kBn 0 ≡ 1, so kBn 0 (ζn k ) = (7.4) Tr(cid:0)∆1 Bn TBn[Bn/z](cid:1) = n Xk=1 ¯ζn k k )2 = − n(ζn B′ 1 n(−1)2 + B′ ¯ζn k k )2 . n(ζn B′ n Xk=2 Now we begin putting hypotheses on the rest of the sequence (λj ). We suppose that, λj ր 1 and that λj > 1/2 for j ≥ 2, which implies that B′ n(−1) ≤ 1 + 4 9 n Xj=2 (1 − λ2 j ), 18 ELIZABETH STROUSE, DAN TIMOTIN, AND MOHAMED ZARRABI so, B′ if we assume that the λj tend sufficiently fast to 1, we may assume that n(−1) ≤ 3/2. To estimate the remaining sum, remember that by Lemma 2.1 (i) we have B′ n(ζ) = n Xj=1 1 − λ2 j ζ − λj2 = 1 + 1 − λ2 j ζ − λj2 . n Xj=2 It follows that for fixed ζ, B′ on both half circles of T as we are getting closer to 1. n(ζ) increases to B′(ζ), while for fixed n it increases Finally, we add the assumption that the sequence λj has been chosen so that B′(ζ) ≤ 11/10 for all ζ belonging to the arc J that connects ei/10 and e−i/10 and contains −1. Then, for ζ ∈ J, B′ n(ζ) ≤ B′(ζ) ≤ 11/10, and so, by Lemma 2.1, d dt Arg Bn(eit) ≤ 11/10. Since d dt Arg Bn(eit) ≤ 11/10 for t ∈ [1/10, π], we have Arg Bn(−1) − Arg Bn(ei/10) ≤ 11/10(π − θ) ≤ 11π/10. Suppose ζn 2 = eiη2 is the next zero of Bn(ζ) = 1 going from −1 towards 1 on the upper semicircle. Then η2 ∈ (0, 1/10), and, since Arg Bn(−1)−Arg Bn(eiη2 ) = 2π, we must have Arg Bn(ei/10) − Arg Bn(eiη2 ) ≥ 9π/10. But, using Lagrange's formula, there is some point θ0 ∈ (η2, 1/10) such that Arg Bn(ei/10) − Arg Bn(eiη2 = ( − η2) d dt Arg Bn(eiθ0 ) ≤ B′ n(eiθ0) ≤ 1 10 B′ n(eiη2 ), 1 10 1 10 the last inequality being a consequence of 0 < η2 < θ0. Therefore B′ A similar argument can be used on the lower semicircle. Finally, since B′ n(eiη2 ) ≥ 9π. n(ζ) k ) ≥ 9π for any increases as we are getting closer to 1, we may conclude that B′ k ≥ 2. n(ζn Remember now that n Xk=1 1 n(ζn k ) B′ = kµBn 1 k = 1. We have then, starting with (7.4), Tr(cid:0)∆1 Bn TBn[Bn/z](cid:1) ≥ ≥ ≥ n Xk=2 1 n(ζn k )2 B′ 1 n(ζn k ) B′ B′ 1 n(−1)2 − Xk=2 1 9π n − − > 1 9π 1 3 . 4 9 4 9 Therefore the inequality (7.3) is proved, which concludes the example. A SZEG O TYPE THEOREM FOR TRUNCATED TOEPLITZ OPERATORS 19 References [1] A. B. Aleksandrov, Isometric embeddings of co-invariant subspaces of the shift operator, Zap. Nauchn. Sem. S.-Peterburg. Odtel. Mat. Inst. Steklov (POMI) 232 (1996), Issled. po Linein. Oper. i Teor. Funktsii. 24, 5 -- 15, 213; translation in J. Math. Sci. (New York) 92 (1998), no. 1, 3543 -- 3549. [2] A. Baranov, I. Chalendar, E. Fricain, J. Mashreghi, and D. Timotin, Bounded symbols and reproducing kernels thesis for truncated Toeplitz operators, Journal of Functional Analysis 259 (2010), 2673 -- 2701. [3] A. Bottcher, Borodin -- Okounkov and Szego for Toeplitz operators on model spaces, Integral Equations Operator Theory 78 (2014), 407 -- 414. [4] A. Bottcher, B. Silbermann, An Introduction to Large Truncated Toeplitz Matrices, Springer, New York, 1999. [5] D.N. Clark, One dimensional perturbations of restricted shifts, J. Analyse Math. 25 (1972), 169 -- 191. [6] J.A. Cima, S.R. Garcia, W.T. Ross, W.R Wogen, Truncated Toeplitz operators: spatial iso- morphism, unitary equivalence, and similarity, Indiana University Mathematical Journal 59 (2010), 595 -- 620. [7] J.A. Cima, A. Matheson, W.T. Ross, The Cauchy Transform, Mathematical Surveys and Monographs, 125, American Mathematical Society, Providence, RI, 2006. [8] S.R. Garcia and W.T. Ross, Recent progress on truncated Toeplitz operators, Blaschke prod- ucts and their applications, Fields Inst. Commun., 65, Springer, New York, 2013, 275 -- 319. [9] R.M. Gray, Toeplitz and Circulant Matrices: A Review, Foundations and Trends in Commu- nications and Information Theory, Vol. 2 (2006): No. 3, 155 -- 239. [10] A. Poltoratskii, Boundary behavior of pseudocontinuable functions, Algebra i Analiz 5 (1993), no. 2, 189 -- 210; translation in St. Petersburg Math. J. 5 (1994), no. 2, 389 -- 406. [11] D. Sarason, Algebraic properties of truncated Toeplitz operators, Oper. Matrices 1 (2007), no. 4, 491 -- 526. [12] N.A. Sedlock, Algebras of truncated Toeplitz operators, Oper. Matrices 5 (2011), no.2, 309- 326. Universit´e de Bordeaux, Institut de Math´ematiques de Bordeaux UMR 5251, 351, cours de la Lib´eration, F-33405 Talence cedex, France E-mail address: [email protected] Institute of Mathematics of the Romanian Academy, P.O. Box 1-764, Bucharest 014700, Romania E-mail address: [email protected] Universit´e de Bordeaux, Institut de Math´ematiques de Bordeaux UMR 5251, 351, cours de la Lib´eration, F-33405 Talence cedex, France E-mail address: [email protected]
1602.04173	1	1602	2016-02-12T19:17:02	Limited operators and differentiability	[ "math.FA" ]	We characterize the limited operators by differentiability of convex continuous functions. Given Banach spaces $Y$ and $X$ and a linear continuous operator $T: Y \longrightarrow X$, we prove that $T$ is a limited operator if and only if, for every convex continuous function $f: X \longrightarrow \R$ and every point $y\in Y$, $f\circ T$ is Fr\'echet differentiable at $y\in Y$ whenever $f$ is G\^ateaux differentiable at $T(y)\in X$.	math.FA	math	Limited operators and differentiability. Mohammed Bachir February 15, 2016 Laboratoire SAMM 4543, Université Paris 1 Panthéon-Sorbonne, Centre P.M.F. 90 rue Tolbiac 75634 Paris cedex 13 Email : [email protected] Abstract. We characterize the limited operators by differentiability of convex continuous functions. Given Banach spaces Y and X and a linear continuous operator T : Y −→ X, we prove that T is a limited operator if and only if, for every convex continuous function f : X −→ R and every point y ∈ Y , f ◦ T is Fréchet differentiable at y ∈ Y whenever f is Gâteaux differentiable at T (y) ∈ X. Keyword, phrase: Limited operators, Gâteaux differentiability, Fréchet differentiability, con- vex functions, extreme points. 2010 Mathematics Subject: Primary 46G05, 49J50, 58C20, 46B20, secondary 47B07. 1 Introduction. A subset A of a Banach space X is called limited, if every weak∗ null sequence (pn)n in X ∗ converges uniformly on A, that is, lim n→+∞ sup x∈A hpn, xi = 0. We know that every relatively compact subset of X is limited, but the converse is false in general. A bounded linear operator T : Y −→ X between Banach spaces Y and X is called limited, if T takes the closed unit ball BY of Y to a limited subset of X. It is easy to see that T : Y −→ X is limited if and only if, the adjoint operator T ∗ : X ∗ −→ Y ∗ takes weak∗ null sequence to norm null sequence. For useful properties of limited sets and limited operators we refer to [11], [4], [6] and [1]. We know that in a finite dimensional Banach space, the notions of Gâteaux and Fréchet differentiability coincide for convex continuous functions. In [5], Borwein and Fabian proved that a Banach space Y is infinite dimensional if and only if, there exists on Y functions f having points at which f is Gâteaux but not Fréchet differentiable. They also pointed in the introduction of [5] the observation that if the sup-norm k.k∞ on c0 is Gâteaux differentiable at some point, then it is Fréchet differentiable there. In this article we observe that this phenomenon is not just related to the sup-norm but more generally, for each convex lower semicontinuous function g : l∞ → R ∪ {+∞}, if g is Gâteaux differentiable at some point a ∈ c0 which is in the interior of its domain, then the restriction of g to c0 is Fréchet differentiable at a. This hold in particular when g = (f ∗)∗ is the Fenchel biconjugate of a convex continuous function f : c0 → R. In fact, this phenomenon is due, (see Corollary 1 in the Appendix and the comment just before), to the fact that the canonical embedding i : c0 −→ l∞ is a limited operator (see the reference [6]). 1 The goal of this paper, is to prove the following characterization of limited operators in terms of the coincidence of Gâteaux and Fréchet differentiability of convex continuous functions. Theorem 1. Let Y and X be two Banach spaces and T : Y −→ X be a continuous linear operator. Then, T is a limited operator if and only if, for every convex continuous function f : X −→ R and every y ∈ Y , the function f ◦ T is Fréchet differentiable at y ∈ Y whenever f is Gâteaux differentiable at T (y) ∈ X. As consequence we give, in Theorem 2 below, new characterizations of infinite dimensional Banach spaces, complementing a result of Borwein and Fabian in [[5], Theorem 1.]. A real valued function f on a Banach space will be called a PGNF-function (see [5]) if there exists a point at which f is Gâteaux but not Fréchet differentiable. A JN-sequence (due to Josefson-Nissenzweig theorem, see [[7], Chapter XII]) is a sequence (pn)n in a dual space Y ∗ that is weak∗ null and inf n kpnk > 0. We say that a function g on X ∗ has a norm-strong minimum (resp. weak∗-strong minimum) at p ∈ X ∗ if g(p) = inf q∈X ∗ g(q) and (pn)n norm converges (resp. weak∗ converges) to p whenever g(pn) −→ g(p). A norm-strong minimum and weak∗-strong minimum are in particular unique. Theorem 2. Let Y be a Banach space. Then the following assertions are equivalent. (1) Y is infinite dimensional. (2) There exists a JN-sequence in Y ∗. (3) There exists a convex norm separable and weak∗ compact metrizable subset K of Y ∗ contaning 0 and a continuous seminorm h on X ∗ which is weak∗ lower semicontinuous and weak∗ sequentially continuous, such that the restriction hK has a weak∗-strong minimum but not norm-strong minimum at 0. (4) There exists a Banach space X and a linear continuous non-limited operator T : Y −→ X. (5) There exists on Y a convex continuous PGNF-function. In Section 2 we give some preliminary results, specially the key Lemma 2. In Section 3, we give the proof of Theorem 1 (divided into two part, Theorem 3 and Theorem 4) and the proof of Theorem 2. In Section 4 we give some complementary remarks. 2 Preliminaries. We recall the following classical result. Lemma 1. Let T1 and T2 be topologies on a set K such that (1) K is Hausdorff with respect T1, (2) K is compact with respect to T2, (3) T1 ⊂ T2. Then T1 = T2. Proof. Let F ⊂ K be a T2-closed set. It follows that F is T2-compact, since K is T2-compact. Let {Oi : i ∈ I} be any cover of F by T1-open sets. Since T1 ⊂ T2, then each of these sets is also T2-open. Hence, there exist a finite subcollection that covers F . It follows that F is T1-compact and therefore is T1-closed since T1 is Hausdorff. This implies that T2 ⊂ T1. Consequently, T1 = T2. Now, we establish the following useful lemma. If B is a subset of a dual Banach space X ∗, we denote by cow∗ (B) the weak∗ closed convex hull of B. 2 Lemma 2. Let X be a Banach space and K be a subset of X ∗. (1) Suppose that K is norm separable, then there exists a sequence (xn)n in the unit sphere SX of X which separate the points of K i.e. for all p, p′ ∈ K, if hp, xni = hp′, xni for all n ∈ N, then p = p′. Consequently, if K is a weak∗ compact and norm separable set of X ∗, then the weak∗ topology of X ∗ restricted to K is metrizable. (2) Let (pn)n be a weak∗ null sequence in X ∗. Then, the set cow∗ {pn : n ∈ N} is convex weak∗ compact and norm separable. Proof. (1) Since K is norm separable, then K − K := {a − b/(a, b) ∈ K × K} is also norm separable and so there exists a sequence (qn)n of K − K which is dense in K − K. According to the Bishop-Phelps theorem [3], the set D = {r ∈ X ∗ r attains its supremum on the sphere SX } is norm-dense in the dual X ∗. Thus, for each n ∈ N, there exists rn ∈ D such that kqn − rnk < 1+n . For each n ∈ N, let xn ∈ SX be such that krnk = hrn, xni. We claim that the sequence (xn)n separate the points of K. Indeed, let q ∈ K − K and suppose that hq, xni = 0, for all k for all k ∈ N∗ and n ∈ N. There exists a subsequence (qnk )k ⊂ K − K such that kqnk − qk < 1 so we have krnk − qk < 1 + 1 k . It follows that 1 1+nk krnk k = hrnk , xnk i = hrnk , xnk i − hq, xnk i ≤ krnk − qk < 1 1 + nk + 1 k . Hence, for all k ∈ N∗, kqk ≤ kq − rnk k + krnk k < 2( k ), which implies that q = 0, and so that (xn)n separate the points of K. Now, suppose that K is weak∗ compact subset of X ∗. We show that the weak∗ topology of X ∗ restricted to K is metrizable. Indeed, each x ∈ X determines a seminorm νx on X ∗ given by + 1 1+nk 1 νx(p) = hp, xi, p ∈ X ∗. The family of seminorms (νx)x∈X induces the weak∗ topology σ(X ∗, X) on X ∗. The subfamily (νxn)n also induces a topology on X ∗, which we will call T . Since this is a smaller family of seminorms, we have T ⊆ σ(X ∗, X). Suppose that p, p′ ∈ K and νxn (p − p′) = 0 for all n ∈ N. Then we have hp, xni = hp′, xni for all n ∈ N and so we have that p = p′ since (xn)n separates the points of K. Consequently, K is Hausdorff with respect to the topology TK (the restriction of T to K). Thus TK is a Hausdorff topology on K induced from a countable family of seminorms, so this topology is metrizable. More precisely, TK is induced from the metric d(p, p′) := +∞ X n=0 2−n νxn(p − p′) 1 + νxn (p − p′) . Then we have that K is Hausdorff with respect to TK, and is compact with respect to σ(X ∗, X)K. Lemma 1 implies that TK = σ(X ∗, X)K. Hence σ(X ∗, X)K is metrizable. (2) Let (pn)n be a weak∗ null sequence in X ∗ and set K = cow∗ {pn : n ∈ N}. Clearly K is a convex and weak∗ compact subset of X ∗. According to Haydon's theorem [[8], Theorem 3.3] the weak∗ compact convex set K is the norm closed convex hull of its extreme points whenever ex(K) (the set of extreme points of K) is norm separable. By the Milman theorem [[10], p.9] ex(K) ⊂ {pn : n ∈ N} Haydon's theorem, K itself is weak∗ compact, convex, and norm separable. = {pn : n ∈ N} ∪ {0} so that ex(K) is norm separable and, hence, by w∗ The following proposition will be used in the proof of Theorem 3. 3 Proposition 1. Let X be a Banach space and K be a weak∗ compact and norm separable subset of X ∗ containing 0. Then, there exists a continuous seminorm h on X ∗ satisfying (1) h is weak∗ lower semicontinuous and sequentially weak∗ continuous, (2) the restriction hK of h to K has a weak∗-strong minimum at 0. Proof. Using Lemma 2, there exists a sequence (xk)k ⊂ SX which separate the points of K. Define the function h : X ∗ −→ R as follows: h(x∗) = (X k≥0 2−k(hx∗, xki)2) 1 2 , ∀x∗ ∈ X ∗. It is clear that h is a seminorm, and since h(x∗) ≤ kx∗k for all x∗ ∈ X ∗, it is also continu- ous. Since h is the supremum of a sequence of weak∗ continuous functions, it is weak∗ lower semicontinuous. On the other hand, since the series Pk≥0 2−k(hx∗, xki)2 uniformly converges on bounded sets of X ∗ and since the maps xk : x∗ 7→ hx∗, xki are weak∗ continuous for all k ∈ N, then h is sequentially weak∗ continuous. If p ∈ K and h(p) = 0, then hp, xki = 0 for all k ∈ N which implies that p = 0, since the sequence (xk)k separate the points of K. Hence, the restriction of h to K has a unique minimum at 0. This minimum is necessarily a weak∗-strong minimum since K is weak∗ metrizable by Lemma 2, this follows from a general fact which say that for every lower semicontinuous function on a compact metric space (K, d), a unique minimum is necessarily a strong minimum for the metric d in question. 3 Limited operators and differentiability. Recall that the domain of a function f : X −→ R ∪ {+∞}, is the set dom(f ) := {x ∈ X/f (x) < +∞}. For a function f with dom(f ) 6= ∅, the Fenchel transform of f is defined on the dual space for all p ∈ X ∗ by f ∗(p) := sup x∈X {hp, xi − f (x)}. The second transform (f ∗)∗ is defined on the bidual X ∗∗ by the same formula. We denote by f ∗∗, the restriction of (f ∗)∗ to X, where X is identified to a subspace of X ∗∗ by the canonical embedding. Recall that the Fenchel theorem state that f = f ∗∗ if and only if f is convex lower semicontinuous on X. The "if" part of Theorem 1 is given by the following theorem. Theorem 3. Let Y and X be Banach spaces and let T : Y −→ X be a linear continuous operator. Suppose that f ◦ T is Fréchet differentiable at y ∈ Y whenever f : X −→ R is convex continuous and Gâteaux differentiable at T (y) ∈ Y . Then T is a limited operator. Proof. Let (pn)n be a weak∗ null sequence in X ∗. We want to prove that kT ∗(pn)kY ∗ → 0. Set K = cow∗ {pn : n ∈ N}. According to Lemma 2, K is convex weak∗ compact and norm separable. Using Proposition 1, there exists a continuous seminorm which is weak∗ lower semicontinuous and sequentially weak∗ continuous h : X ∗ −→ R such that the restriction hK of h to K has a weak∗-strong minimum at 0 and in particular minK h = h(0) = 0. Since the sequence (pn)n weak∗ converges to 0, it follows that limn h(pn) = h(0) = minK h. Thus, (pn)n is a minimizing sequence for hK. Set g = h + δK, where δK denotes the indicator function, which is equal to 0 on K and equal to +∞ otherwise. Since K is convex, weak∗-closed and norm bounded, then g is a convex and 4 weak∗ lower semicontinuous function with a norm bounded domain dom(g) = K. Moreover we have, (1) g(p) > 0 = g(0) = minX ∗ (g) for all p ∈ X ∗ \ {0}. (2) limn→+∞ g(pn) = minX ∗ (g). Hence, there exists a convex and Lipschitz continuous function f : X −→ R such that g = f ∗ (we can take f = g∗ X ). The function f is Gâteaux differentiable at 0 with Gâteaux derivative ∇f (0) = 0, this is due to the fact that f ∗ = g has a weak∗-strong minimum at 0 (we can see [Corollary 1. [2]]). Thus, from our hypothesis, f ◦ T is Fréchet differentiable at 0 with Fréchet derivative equal to 0. It follows that (f ◦ T )∗ has a norm-strong minimum at 0 (see [Corollary 2. [2]]). Now, we prove that (T ∗(pn))n is a minimizing sequence for (f ◦ T )∗, which will implies that kT ∗(pn)kY ∗ → 0. Indeed, on one hand, we have 0 = minX ∗ (g) = −g∗(0) = −f (0). On the other hand we have 0 = −f (0) ≤ sup y∈Y {−f ◦ T (y)} := (f ◦ T )∗(0) ≤ sup x∈X {−f (x)} = f ∗(0) = g(0) = 0. It follows that (f ◦ T )∗(0) = 0. Hence, since (f ◦ T )∗ has a minimum at 0, we obtain 0 = (f ◦ T )∗(0) ≤ (f ◦ T )∗(T ∗(pn)) {hT ∗(pn), yi − f ◦ T (y)} := sup y∈Y = sup y∈Y ≤ sup x∈X {hpn, T (y)i − f (T (y))} {hpn, xi − f (x)} = f ∗(pn) = g(pn). Since, g(pn) → 0, it follows that (f ◦ T )∗(T ∗(pn)) → 0 = (f ◦ T )∗(0). In other words, (T ∗(pn))n is a minimizing sequence for (f ◦ T )∗. Since (f ◦ T )∗ has a norm-strong minimum at 0, we obtain that kT ∗(pn)kY ∗ → 0, which implies that T is a limited operator. The "only if" part of Theorem 1 is given by the following theorem. Theorem 4. Let Y and X be two Banach spaces and T : Y −→ X be a limited operator. Let f : X −→ R ∪ {+∞}, be a convex lower semicontinuous function and let a ∈ Y such that T (a) belongs to the interior of dom(f ). Then, f ◦ T is Fréchet differentiable at a ∈ Y with Fréchet-derivative T ∗(Q) ∈ Y ∗, whenever f is Gâteaux differentiable at T (a) ∈ X with Gâteaux-derivative Q ∈ X ∗. Proof. Since f is convexe lower semicontinuous and T (a) is in the interior of dom(f ), there exists ra > 0 and La > 0 such that f is La-Lipschitz continuous on the closed ball BX (T (a), ry). It is well known that there exists a convex La-Lipschitz continuous function fa on X such that fa = f on BX (T (a), ra) (See for instance Lemma 2.31 [9]). It follows that fa ◦ T = f ◦ T on BY (a, ra kT k )) is a subset of BX (T (a), ra) (we can assume that T 6= 0). fa, we can assume without loss of generality that f is convexe 1-Lipschitz Replacing f by 1 La continuous on X. It follows that dom(f ∗) ⊂ BX ∗ (the closed unit ball of X ∗). kT k ), since T (BX(a, ra 5 Claim. Suppose that f is Gâteaux differentiable at T (a) ∈ X with Gâteaux-derivative Q ∈ X ∗, then the function q 7→ f ∗(q) − hq, T (a)i has a weak∗-strong minimum on BX ∗ at Q. Proof of the claim. See [Corollary 1. [2]]. Now, suppose by contradiction that T ∗(Q) is not the the Fréchet derivative of f ◦ T at a. There exist ε > 0, tn −→ 0+ and hn ∈ Y , khnkY = 1 such that for all n ∈ N∗, f ◦ T (a + tnhn) − f ◦ T (a) − hT ∗(Q), tnhni > εtn. Let rn = tn/n for all n ∈ N∗ and choose pn ∈ BX ∗ such that f ∗(pn) − hpn, T (a + tnhn)i < inf p∈BX∗ {f ∗(p) − hp, T (a + tnhn)i} + rn. (1) (2) From (2) we get f ∗(pn) − hpn, T (a)i < inf p∈BX∗ {f ∗(p) − hp, T (a)i} + 2tnkT k + rn. This implies that the sequence (pn)n minimize the function q 7→ f ∗(q) − hq, T (a)i on BX ∗ . Using the claim, the function q 7→ f ∗(q) − hq, T (a)i has a weak∗-strong minimum on BX ∗ at Q, it follows that (pn)n weak∗ converges to Q and so (since T is limited) we have kT ∗(pn − Q)kY ∗ −→ 0. (3) On the other hand, since f (T (a + tnhn)) = f ∗∗(T (a + tnhn)) = − inf p∈BX∗ {f ∗(p) − hp, T (a + tnhn)i}, using (2) we obtain for all y ∈ Y f ◦ T (a + tnhn) − hpn, T (a + tnhn)i < −f ∗(pn) + rn ≤ f ◦ T (y) − hpn, T (y)i + rn. Replacing y by a in the above inequality we obtain f ◦ T (a + tnhn) − hpn, T (tnhn)i ≤ f ◦ T (a) + rn. (4) Combining (1) and (4) we get ε < hpn, T (hn)i − hT ∗(Q), hni + rn/tn = hT ∗(pn), hni − hT ∗(Q), hni + 1 n ≤ kT ∗(pn − Q)kY ∗ + 1 n which is a contradiction with (3). Thus f ◦T is Fréchet differentiable at a with Fréchet derivative T ∗(Q). Now, we give the proof of Theorem 2. Proof of Theorem 2. (1) =⇒ (2) is the deeper Josefson-Nissenzweig theorem [[7], Chapter XII]. (2) =⇒ (1) is well known. (2) =⇒ (3) Let (pn)n be a weak∗ null sequence in Y ∗ such that inf n kpnk > 0 and set {pn : n ∈ N}. By Lemma 2, the set K is convex norm separable and weak∗ compact K = cow∗ 6 metrizable. On the other hand, from Proposition 1, there exists a continuous seminorm h which is weak∗ lower semicontinuous and weak∗ sequentially continuous on Y ∗ such that the restriction of h to K has a weak∗-strong minimum at 0. It remains to show that 0 is not a norm-strong minimum for hK. Indeed, since (pn)n is weak∗ null and h is weak∗ sequentially continuous, then limn h(pn) = h(0) = minK h. So (pn)n is a minimizing sequence for hK which not converges to 0 since inf n kpnk > 0. Hence, 0 is not a norm-strong minimum for hK. (3) =⇒ (2) Since 0 is not a norm-strong minimum for the restriction hK, there exists a sequence (pn)n that minimize h on K but kpnk 9 0. Since hK has a weak∗-strong minimum at 0, it follows that (pn)n weak∗ converges to 0. Hence, (pn)n weak∗ converges to 0 but kpnk 9 0. Thus, there exists a JN-sequence in Y ∗. (2) =⇒ (4) This part is given by taking X = Y and T = I the identity map. Indeed, there exists a sequence (pn)n which weak∗ converges to 0 but inf n kI ∗(pn)k = inf n kpnk > 0. So I cannot be a limited operator. (4) =⇒ (5). Indeed, if there exists a Banach space X and a non-limited operator T : Y −→ X, by using Theorem 1, there exists a convex continuous function f : X −→ R and a point y ∈ Y such that f is Gâteaux differentiable at T (y) ∈ X but f ◦ T is not Fréchet differentiable at y. So f ◦ T is Gâteaux but not Fréchet differentiable at y. Hence, f ◦ T is a convex continuous PGNF-function on Y . (5) =⇒ (2) Let f be a PGNF-function on Y . We can assume without loss of generality that f is Gâteaux differentiable at 0 with Gâteaux-derivative equal to 0, but f is not Fréchet differentiable at 0. It follows from classical duality result (see Corollary 1. and Corollary 2. in [2]) that f ∗ has a weak∗-strong minimum but not norm-strong minimum at 0. Since 0 is not a norm-strong minimum for f ∗, there exists a sequence (pn)n ∈ X ∗ minimizing f ∗ such that kpnk 9 0. On the other hand, since f ∗ has a weak∗-strong minimum at 0, and (pn)n minimize f ∗, we have that (pn)n weak∗ converges to 0. Thus, (pn)n weak∗ converges to 0 but kpnk 9 0. Hence, there exists a JN-sequence. Canonical construction of PGNF-function. There exist different way to build a PGNF- function in infinite dimentional Banach spaces. We can find examples of such constructions in [5]. We present below a different method for constructing a PGNF-function on a Banach space X canonically from a JN-sequence. Given a JN-sequence (pn)n ⊂ X ∗, we set K = cow∗ {pn : n ∈ N}. Using Lemma 2, there exists a sequence (xn)n ∈ SX which separates the points of K, and as in the proof of Proposition 1, there exist a continuous seminorm h which is weak∗ lower semicontinuous and weak∗ sequentially continuous such that hK has a weak∗-strong minimum at 0. The function h is given explicitly as follows h(x∗) = (X n≥0 2−n(hx∗, xni)2) 1 2 , ∀x∗ ∈ X ∗. Since (pn)n weak∗ converges to 0, it follows that (pn) is a minimizing sequence for hK. Since (pn)n is a JN-sequence, it follows that 0 is not a norm-strong minimum for hK. Define the function f by f (x) = (h + δK)∗(x), ∀x ∈ X, where δK denotes the indicator function, which is equal to 0 on K and equal to +∞ otherwise and where for each x ∈ X, we denote by x ∈ X ∗∗ the linear map x∗ 7→ hx∗, xi for all x∗ ∈ X ∗. Then f is convex Lipschitz continuous, Gâteaux differentiable at 0 (since h + δK has a weak∗- strong minimum) but not Fréchet differentiable at 0 (because 0 is not a norm-strong minimum for h + δK). 4 Appendix. There exists a class of Banach spaces (E, k.kE) such that the canonical embedding i : E −→ E∗∗ is a limited operator. This class contains in particular the space c0 and any closed subspace 7 F of c0 (This class is also stable by product and quotient. For more information see [6]). In this setting, Theorem 4 gives immediately the following corollary. Corollary 1. Suppose that the canonical embedding i : E −→ E∗∗ is a limited operator. Let g : E∗∗ −→ R ∪ {+∞} be a convex lower semicontinuous function. Suppose that x ∈ E belongs to the interior of dom(g) and that g is Gâteaux differentiable at x ∈ E (we use the identification i(x) = x), then the restriction of g to E is Fréchet differentiable at x. In particular, if f : E −→ R ∪ {+∞} is convex lower semicontinuous function, x ∈ E belongs to the interior of dom((f ∗)∗) and (f ∗)∗ is Gâteaux differentiable at x, then f is Fréchet differentiable at x. We obtain the following corollary by combining Proposition 2 and a delicate result due to Zajicek (see [Theorem 2; [12]]), which say that in a separable Banach space, the set of the points where a convex continuous function is not Gâteaux differentiable, can be covered by countably many d.c (that is, delta-convex) hypersurf ace. Recall that in a separable Banach space Y , each set A which can be covered by countably many d.c hypersurf ace is σ-lower porous, also σ-directionally porous; in particular it is both Aronszajn (equivalent to Gauss) null and Γ-null. For details about this notions of small sets we refer to [13] and references therein. Note that a limited set in a separable Banach space is relatively compact [4]. Proposition 2. Let Y and X be Banach spaces and T : Y −→ X be a limited operator with a dense range. Let f : X −→ R be a convex continuous function. Then f ◦ T is Gâteaux differentiable at a ∈ Y if and only if, f ◦ T is Fréchet differentiable at a ∈ Y . Proof. Suppose that f ◦ T is Gâteaux differentiable at a ∈ Y . It follows that f is Gâteaux differentiable at T (a) with respect to the direction T (Y ) which is dense in X. It follows (from a classical fact on locally Lipschitz continuous functions) that f is Gâteaux differentiable at T (a) on X. So by Theorem 4, f ◦ T is Fréchet differentiable at a ∈ Y . The converse is always true. Corollary 2. Let Y be a separable Banach space, X be a Banach spaces and T : Y −→ X be a compact operator with a dense range. Let f : X −→ R, be a convex and continuous function. Then, the set of all points at which f ◦ T is not Fréchet differentiable can be covered by countably many d.c hypersurf ace. References [1] K. T. Andrews, Dunford-Pettis sets in the space of Bochner integrable functions, Math. Ann. 241, (1979), 35-41. [2] E. Asplund and R. T. Rockafellar, Gradients of convex functions. Trans. Amer. Math. Soc. 139, (1969), 443-467. [3] E. Bishop and R. R. Phelps, A proof that every Banach space is subreflexive, Bull. Amer. Math. Soc. 67, (1961), 97-98 . [4] J. Bourgain and J. Diestel, Limited operators and strict cosingularity, Math. Nachr. 119, (1984), 55-58. [5] J. M. Borwein and M. Fabian, On convex functions having points of Gâteaux differen- tiability which are not points of Féchet-differentiability. Can. J. Math.Vol.45 (6), 1993, 1121-1134. [6] H. Carrión, P. Galindo, and M.L Lourenco, Banach spaces whose bounded sets are bound- ing in the bidual Annales Academiae Scientiarum Fennicae Mathematica Volumen 31, (2006), 61-70. 8 [7] J. Diestel, Sequences and series in Banach spaces, Graduate texts in Mathematics, Springer Verlag, N.Y., Berlin, Tokyo, 1984. [8] R. Haydon, An extreme point criterion for separability of a dual Banach space, and a new proof of a theorem of Corson, Quart. J. Math. Oxford Ser. 27 (1976), 379-385. [9] R. R. Phelps, Convex Functions, Monotone Operators and Differentiability. Lecture Notes in Mathematics 1364, Springer-Verlag, Berlin (1993). [10] R. R. Phelps, Lectures on Chaquet's theorem, Van Nostrand, Princeton, N. J., 1966. [11] Th. Schlumprecht, Limited sets in Banach spaces, Dissertation, München, 1987. [12] L. Zajicek, On the differentiation of convex functions in finite and infinite dimensional spaces, Czechoslovak Math. J. 29, (1979), no. 3, 340-348. [13] L. Zajicek, On sigma-porous sets in abstract spaces, Abstract and Applied Analysis, vol. (2005), issue 5, 509-534, 9
1912.00910	1	1912	2019-12-02T16:39:57	Shorting, parallel addition and form sums of nonnegative selfadjoint linear relations	[ "math.FA" ]	We extend the operations of shorting and parallel addition from the cone of bounded nonnegative selfadjoint operators in a Hilbert space to the set of all nonnegative selfadjoint linear relations. New properties of these operations and connections with the Cayley transforms are established. It is shown that for a pair of nonnegative selfadjoint linear relations there exist, in general, two arithmetic--harmonic means. Applications of the arithmetic, harmonic, and arithmetic--harmonic means to the theory of nonnegative selfadjoint extensions of nonnegative symmetric linear relations are given.	math.FA	math	SHORTING, PARALLEL ADDITION AND FORM SUMS OF NONNEGATIVE SELFADJOINT LINEAR RELATIONS YU.M. ARLINSKII Abstract. We extend the operations of shorting and parallel addition from the cone of bounded nonnegative selfadjoint operators in a Hilbert space to the set of all nonnegative selfadjoint linear relations. New properties of these operations and connections with the Cayley transforms are established. It is shown that for a pair of nonnegative selfadjoint linear relations there exist, in general, two arithmetic -- harmonic means. Applications of the arithmetic, harmonic, and arithmetic -- harmonic means to the theory of nonnegative selfadjoint extensions of nonnegative symmetric linear relations are given. 9 1 0 2 c e D 2 ] . A F h t a m [ 1 v 0 1 9 0 0 . 2 1 9 1 : v i X r a 1. Introduction We will use the following notations: dom A, ran A, and ker A are the domain, the range, and the kernel of a linear operator/linear relation A, ran A and clos L denote the closure of ran A and of the set L, respectively. By s − lim we denote the strong limit of operators and s − R − lim means the strong resolvent limit of operators/linear relations [17, 33]. The Banach space of all bounded operators acting between Hilbert spaces H1 and H2 is denoted by B(H1,H2) and B(H) := B(H,H). The cone of all bounded self-adjoint nonnegative operators in a complex Hilbert space H is denoted by B+(H). If A1, A2 are two bounded selfadjoint operators in H and the difference A1 − A2 belongs to B+(H), then we write A1 ≥ A2. If L is a subspace (closed linear manifold) in H, then PL is the orthogonal projection in H onto L, and L⊥ def = H ⊖ L. N is the set of natural numbers, N0 := N ∪ {0}, C is the field of complex numbers, R+ := [0,∞). Let H be a complex Hilbert space and let S ∈ B+(H). It was discovered by M.G. Kreın [34] that for an arbitrary subspace L of H the set (1.1) Ξ(S,L) :=neS ∈ B+(H) : eS ≤ S, raneS ⊆ Lo has a maximal element SL. The operation hS,Li 7→ SL was applied in [34] to the problem of a description contractive selfadjoint extensions of a non-densely defined Hermitian contraction. The following representations of SL and associated quadratic form were established in [34]: (1.2) 2 PMS 1 1 2 , (SLf, f ) = inf ϕ∈L⊥ {(S(f + ϕ), f + ϕ)} , f ∈ H, SL = S 2010 Mathematics Subject Classification. 47A06, 47A64, 47A20. Key words and phrases. Nonnegative selfadjoint linear relation, shorting, parallel addition, harmonic mean, arithmetic -- harmonic mean, selfadjoint extension. 1 2 YU.M. ARLINSKII where PM is the orthogonal projection in H onto the subspace M = nf ∈ H : S Besides [34] 1 2 f ∈ Lo . (1.3) ran S = ran S 1 2 L 1 2 ∩ L, SL = 0 ⇐⇒ ran S 2 ∩ L = {0}. 1 A bounded selfadjoint operator S admits the block operator matrix representation w.r.t. the orthogonal decomposition H = L⊥ ⊕ L: S =(cid:20)S11 S12 S∗12 S22(cid:21) : L⊥ ⊕ L → , L⊥ ⊕ L where S∗11 = S11, S∗22 = S22. It is well known (see e.g. [47]) that (1.4) (1.5) and the operator SL is given by the block matrix 0 , 1 2 [− 1 2 ] S11 ≥ 0, ran S12 ⊂ ran S 11, [− 1 2 ] S22 −(cid:16)S 0 S22 −(cid:16)S S ≥ 0 ⇐⇒ 11 S12(cid:17)∗(cid:16)S SL ="0 11 S12(cid:17)∗(cid:16)S 11 S12(cid:17) takes the form S22 − S∗12S−1 11 S12(cid:17) ≥ 0 11 S12(cid:17)# , If S−1 [− 1 2 ] [− 1 2 ] [− 1 2 ] is the Moore-Penrose pseudo-inverse. where S S22 −(cid:16)S [− 1 2 ] 11 [− 1 2 ] 11 S12(cid:17)∗(cid:16)S 11 ∈ B(L⊥), then the operator 11 S12 and is said to be the Schur complement of S w.r.t. L [53]. From (1.5) it follows that SL = 0 ⇐⇒ ran S12 ⊂ ran S 1 2 11 and S22 = S∗12S−1 11 S12. The function B+(H) ∋ S 7→ SL has the following property [47]: in the strong convergence sense (1.6) B+(H) ⊃ {Sn} ց S∞ =⇒ (Sn)L ց (S∞)L For the case of a nonnegative S acting on a finite-dimensional Hilbert space H, the operator SL in (1.5) was called by W.N. Anderson in [1] the shorted operator due to the its connection with electrical network theory. It was shown [1, Theorem 1] that if the shorted operator is defined by such a way, then it is the maximal element of the set Ξ(S,L) defined in (1.1). Basic properties of SL were studied in [1, 2, 3, 26, 34, 35, 39, 40, 42, 43, 47]. The parallel sum B : G for B, G ∈ B+(H) for the case of finite-dimensional H was defined in [2] as follows B : G = B(B + G)[−1]G, where (B + G)[−1] is the Moore-Penrose pseudo-inverse. In the general case of dim H ≤ ∞ the operator B : G is of the form (1.7) B : G = (B−1 + G−1)−1, if both B and G have bounded inverses and if this is not the case, then B : G = s − lim ε↓0 (B + εI) : (G + εI) = s − lim ε↓0(cid:0)(B + εI)−1 + (G + εI)−1(cid:1)−1 = s − lim ε↓0 B (B + G + εI)−1 G. SHORTING, PARALLEL ADDITION AND FORM SUMS 3 In [26] the following important property of the parallel sum is established: (1.8) 2 ∩ ran G The quadratic form ((B : G)h, h) admits the representation ran (B : G) 2 = ran B 2 . 1 1 1 ((B : G)ϕ, ϕ) = inf f,g∈H{ (Bf, f ) + (Gg, g) : ϕ = f + g } , see [3, 5, 26, 44]. It follows that B : G ≤ B and B : G ≤ G. As is known [44], B : G can be calculated as follows i.e., the parallel addition can be expressed [3] by means of the shorting: B : G = B −(cid:0)(B + G)[−1/2]B(cid:1)∗(cid:0)(B + G)[−1/2]B(cid:1) , B : G =(cid:18)(cid:20)B + G B B(cid:21)(cid:19){0}⊕H B ↾ {0} ⊕ H. The latter representation leads to the following statement, see [44, Theorem 2.5]: (1.9) B+(H) ⊃ {Bn},{Gn}, Bn ց B∞, Gn ց G∞ in the strong convergence sense =⇒ s − lim n→∞ (Bn : Gn) = B∞ : G∞. Another connections between shorting and parallel addition were established in [3] and are given by the following equalities (1.10) (1.11) BL = s − lim t↑+∞ (B : tPL), (B : G)L = BL : G = B : GL = BL : GL. Various applications of shorting and parallel addition in complex analysis, operator theory, probability and statistics, numerical analysis are scattered in the literature. For matrices some of applications can be found in the books [38, 53]. Generalizations of the shorting and parallel additions to the cases of bounded operators, acting between two Hilbert spaces, on linear spaces, and on Kreın spaces, are given in [8, 19, 27, 37]. The operation of parallel addition has been extended to quadratic forms in [31, 32, 51, 52]. In this article the shorting and parallel addition are defined and studied for nonnegative selfadjoint linear relations (l.r. for short) and, in particular, for unbounded nonnegative selfadjoint operators. We define the shorting AL similarly to Kreın's definition (1.1) for a bounded nonnegative selfadjoint operator. The parallel addition of two nonnegative selfad- joint l.r. we define as in (1.7), replacing the sum by the form sum. Recall that the form sum A +B of nonnegative selfadjoint l.r. A and B is the nonnegative selfadjoint l.r. associated 2 ϕ2 +B with the closed quadratic form g[ϕ] = A 2 [25, 29, 33]. 2 holds. In the We show, see Theorem 5.4, that the inclusion ran A sequel the nonnegative selfadjoint l.r. (A +B)/2 will be called the arithmetic mean of A and B. The harmonic mean h(A, B) of A and B (in [36] the notation A!B was used) is defined as follows 2 ∩dom B 2 ⊆ ran (A +B) 2 ϕ2, ϕ ∈ dom A 2 + ran B 1 1 1 1 1 1 1 h(A, B) = 2(A : B) =(cid:18)1 2 (A−1 +B−1)(cid:19)−1 . For A, B ∈ B+(H) the inequality h(A, B) ≤ (A + B)/2 is valid [6]. For nonnegative selfadjoint l.r. we show that equalities (1.3), (1.8), (1.11) remain valid and some new properties of shorting and parallel additions we establish in Theorems 3.1, 3.3, 4.2. 4 YU.M. ARLINSKII There is a one-to-one correspondence between nonnegative selfadjoint l.r. L and selfadjoint contractions T given by the Cayley transform: L 7→ C(L) = T := −I + 2(I + L)−1 ∈ [−I, I], [−I, I] ∋ T 7→ C(T ) = L := −I + 2(I + T )−1 = {{(I + T )f, (I − T )f} : f ∈ H} , (1.12) C(C(L)) = L, C(L−1) = −T = −C(L). In Theorem 3.1 and Theorem 5.2 we establish the following equalities, connecting the Cayley transforms: C(AL) = I − (I − C(A))L, C(cid:0) 1 2(A +B)(cid:1) = h(I + C(A), I + C(B)) − I, C (h(A, B)) = I − h(I − C(A), I − C(B)), where A and B are nonnegative selfadjoint l.r. and L is a subspace. Using the above equali- ties and replacing the strong convergence by the strong resolvent convergence, in Proposition 3.2 and Corollary 5.3 we extend properties (1.6), (1.9), (1.10) to the set of all nonnegative selfadjoint l.r.. The mean c0(a, b) := a + b + 2ab 2 + a + b of positive numbers a and b satisfies the inequalities 2ab a + b ≤ a + b + 2ab 2 + a + b ≤ a + b 2 with " = " if and only if a = b. Using the Cayley transforms, the mean c0 can be extended to arbitrary nonnegative selfadjoint l.r. A and B: We prove in Theorem 5.4 analogues of the above inequalities: c0(A, B) := C(cid:18)C(A) + C(B) 2 (cid:19) . h(A, B) ≤ c0(A, B) ≤ A +B 2 with " = " if and only if A = B. The geometric mean A#B for A, B ∈ B+(H) can be defined as follows [6, 36, 45]: A#B := maxnX ∈ B+(H) : (Xϕ, ψ) ≤ A 2(cid:16)A− 1 2(cid:17) 1 2 BA− 1 A A 1 2 1 1 1 2 ϕ B 1 2 ψ ϕ, ψ ∈ Ho , . 2 , if A−1 and B−1 are bounded 2 . The following definition [4] of geometric For commuting A and B one has A#B = (AB) mean as the arithmetic -- harmonic mean is a straightforward generalization of this notion for positive numbers. Define A0 := A, B0 := B, An := 1 2 (An−1 + Bn−1), Bn := h(An−1, Bn−1), n ∈ N. Then {An}, {Bn} are non-increasing and non-decreasing sequences, respectively, and they have the common strong limit g(A, B), which coincides with A#B and thus h(A, B) ≤ A#B ≤ 1 2 (A + B). SHORTING, PARALLEL ADDITION AND FORM SUMS 5 In Section 6 we define similar sequences for nonnegative selfadjoint l.r. A and B and show that, in general, there is no a common strong resolvent limit, i.e., there are two arithmetic- harmonic means ah(A, B). According to Kreın's results [34] and their generalizations in [7, 11, 18, 28, 29, 30] (see Section 7), the Cayley transform of the set of all nonnegative selfadjoint extensions of a nonnegative symmetric l.r. forms an operator interval. In Theorem 7.1, in particular, we find parameters corresponding to the arithmetic and harmonic means of given two nonnegative selfadjoint extensions, in Proposition 7.2 we show that the pair of arithmetic -- harmonic symmetric operator or linear relation consists of 1 we show for differential operators in L2(R+) means ah(eS1,eS2) for two extremal nonnegative selfadjoint extensions eS1, eS2 of a nonnegative 2(eS1 +eS2) and h(eS1,eS2). In Proposition 7.3  2 (R+), f′(0) = cf (0)} , c ∈ [0, +∞], W 2 Lcf = − dom Lc = {f ∈ W 2 2 (R+) is the Sobolev space the validity of the equalities d2f dx2 1 2(Lc +Ld) = L 1 , ah(Lc, Ld) = L√cd, 2(L0 +L∞) = L∞, h(L0, L∞) = L0, ah(L∞, L0) = hL∞, L0i . 2 (c+d), h(Lc, Ld) = L 2cd c+d 1 In the case of a nonnegative symmetric l.r. having one-dimensional resolvent difference of its Friedrichs and Kreın extensions, we calculate in Theorem 7.4 parameters correspond- ing to the resolvents of the arithmetic -- harmonic means of any two nonnegative selfadjoint extensions. 2. Nonnegative selfadjoint linear relations Let H be a Hilbert space and let JH : {f, f′} 7→ {if′,−if}, f, f′ ∈ H. Then JH is selfadjoint and unitary operator in the Hilbert space H2 = H ⊕ H. Let A be a l.r. in H [9], i.e., A is a closed linear manifold (a subspace) in the Hilbert space H2. Recall that the l.r. A∗ := H2 ⊖ JHA is called the adjoint to A. If A = {{ϕ, ϕ′}}, then, clearly, A∗ = {{g, g′} : (ϕ′, g) = (ϕ, g′) ∀ {ϕ, ϕ′} ∈ A} . A l.r. A is called selfadjoint if A∗ = A ⇐⇒ A = H2 ⊖ JHA. The domain, range, kernel and multi-valued part of a l.r. A are defined as follows: dom A = {f ∈ H : {f, f′} ∈ A} , ran A = {f′ ∈ H : {f, f′} ∈ A} , ker A = {f ∈ H : {f, 0} ∈ A} , mul A = {f′ ∈ H : {0, f′} ∈ A} . The l.r. A−1 := {{f′, f} : {f, f′} ∈ A} is called the inverse to A. Clearly dom A−1 = ran A, ran A−1 = dom A, ker A−1 = mul A, mul A−1 = ker A. Recall that a symmetric (selfadjoint) l.r. A admits the orthogonal decomposition A = Graph(Ao)L{{0} ⊕ mul A}, where Ao is the symmetric (selfadjoint) operator part acting The resolvent (A − λI)−1 of a selfadjoint l.r. A is defined for all regular points ρ(Ao) of on dom A (dom A = H ⊖ mul A when A is selfadjoint). the operator part Ao and takes the form where PA is the orthogonal projection onto dom A. We will identify ρ(A) with ρ(Ao). (A − λI)−1f = (Ao − λI)−1PAf, f ∈ H, 6 YU.M. ARLINSKII If L is a subspace of H, then PL(A − λI)−1↾L is called the compressed resolvent of A. The holomorphic family MA,L of l.r. (2.1) is a Nevanlinna function or a Nevanlinna family [21, 22]. Note that (2.2) If a bounded selfadjoint operator A in the Hilbert space H = L⊥ ⊕ L is given by the block operator matrix MA,L(λ) := −(cid:0)PL(A − λI)−1↾ L(cid:1)−1 − λIL, λ ∈ ρ(Ao) MA−1,L(λ) =(cid:0)MA,L(λ−1)(cid:1)−1 , λ ∈ ρ(A) \ {0}. A =(cid:20)A11 A12 A∗12 A22(cid:21) , A11 = A∗11, A22 = A∗22, then B(L)-valued function (2.3) is a Nevanlinna function and MA,L(λ) := −A22 + A∗12 (A11 − λIL⊥)−1 A12, λ ∈ ρ(A11) PL (A − λIL)−1 ↾L = − (MA,L(λ) + λIL)−1 , λ ∈ ρ(A) ∩ ρ(A11). In the sequel the notation A↾L is used for the l.r. A↾L := {{f, f′} : f ∈ dom A ∩ L, {f, f′} ∈ A} . It is established in [50] that if A is a selfadjoint unbounded linear operator and L is a subspace with finite codimension, then the operator PLA↾L is selfadjoint in the space L. In [41] has been proved that the compression PLA↾ L of a maximal dissipative operator A is maximal dissipative in L. Further in [16] similar assertions were obtained for selfadjoint and maximal dissipative l.r.. Let A = {{f, f′}} be a symmetric l.r. in the Hilbert space H. Then for any {f, f′} ∈ A one has the equality [46] (f′, f ) = (Aof, f ). A l.r. A is called nonnegative (we will write A ≥ 0) if (f′, f ) ≥ 0 for all {f, f′} ∈ A. 2 is defined as follows If A is a nonnegative selfadjoint l.r. then the square root A 1 A Hence 1 2 = Graph(A 1 2 o )M{{0} ⊕ mul A}}, ran A 1 2 = ran A 1 2 o ⊕ mul A. A− 1 2 = {{A o f, f}, f ∈ dom A 1 2 1 2 o }M{{mul A ⊕ {0}}. Let a = a[·,·] be a nonnegative sesquilinear form in the Hilbert space H with domain dom a and let a[ϕ] := a[ϕ, ϕ], ϕ ∈ dom a. The form a is closed [33] if ( lim n→∞ {ϕn} ⊂ dom a The form a is closable [33] if ϕn = ϕ, lim n,m→∞ a[ϕn − ϕm] = 0, =⇒ ϕ ∈ dom a, lim n→∞ a[ϕn − ϕ] = 0. ( lim n→∞ ϕn = 0, {ϕn} ⊂ dom a lim n,m→∞ a[ϕn − ϕm] = 0, =⇒ lim n→∞ a[ϕn] = 0. The form a is closable if and only if it has a closed extension, and in this case the closure of the form is the smallest closed extension of a. The inequality a1 ≥ a2 for semi-bounded forms a1 and a2 is defined by (2.4) dom a1 ⊆ dom a2, a1[ϕ] ≥ a2[ϕ], ϕ ∈ dom a1. SHORTING, PARALLEL ADDITION AND FORM SUMS 7 In particular, a1 ⊂ a2 implies a1 ≥ a2. If the forms a1 and a2 are closable, the inequality a1 ≥ a2 is preserved by their closures. If A = {{f, f′}} is a nonnegative symmetric l.r., then the form a[f ] = (f′, f ), {f, f′} ∈ A, f ∈ dom A = dom Ao is closable [33, 46]. There is a one-to-one correspondence between all closed nonnegative forms a and all nonnegative selfadjoint l.r. A in H (the first representation theorem [33]), see [33, 46], via dom A ⊂ dom a and (2.5) In what follows the closed form associated with A is denoted by A[·,·] and its domain by D[A]. By the second representation theorem (2.6) f ∈ dom A, ψ ∈ dom a. a[f, ψ] = (Aof, ψ), A[ϕ, ψ] = (A o ϕ, A 1 2 1 2 1 2 o ψ), ϕ, ψ ∈ D[A] = dom A o . The formulas (2.5), (2.6) are analogs of Kato's representation theorems for, in general, nondensely defined closed semi-bounded forms in [33, Section VI]; see e.g. [46, 14, 30]. If A is a nonnegative selfadjoint l.r. and if T = C(A) is its Cayley transform, then [12] (2.7)  1 2 , D[A] = ran (I + T ) A[u, v] = −(u, v) + 2(cid:16)I + T )[− 1 2 ]u, (I + T )[− 1 2 ]v(cid:17) , u, v ∈ D[A], 2 ] is the Moore-Penrose pseudo-inverse. Replacing A by A−1 in (2.7), we where (I + T )[− 1 conclude that ran A 1 2 = ran (I − T ) 1 2 . Given a sequence {An} of nonnegative selfadjoint l.r., we say that {An} converges in the An = A) if the sequence of the resolvent {(An + I)−1} converges in the strong strong resolvent sense [33] to a nonnegative selfadjoint l.r. A (s − R − lim n→∞ sense to the resolvent (A + I)−1 (s − lim n→∞ inequality A1 ≥ A2 if dom A 1 2 1 2 1 2 1 2 1,o ⊂ dom A 2,o and kA 1,oϕk ≥ kA (An + I)−1 = (A + I)−1). Let A1 and A2 be nonnegative selfadjoint l.r. in H, then A1 and A2 are said to satisfy the 2,oϕk, ϕ ∈ dom A 1 2 1,o. This means that the closed nonnegative forms A1[·,·] and A2[·,·] associated with A1 and A2 satisfy the inequality A1[ϕ] ≥ A2[ϕ] for all ϕ ∈ D[A1]; see (2.4), (2.6). Finally we note that the inequalities A1 ≥ A2 ≥ 0 are equivalent to the inequality A−1 2 , see [29]. Besides, (2.8) 1 ≤ A−1 A1 ≥ A2 ≥ 0 ⇐⇒ −I ≤ C(A1) ≤ C(A2) ≤ I. in H. Then the resolvent (A + xI)−1 is a bounded Let A be a nonnegative selfadjoint l.r. nonnegative selfadjoint operator for an arbitrary positive number x. If 0 ≤ x1 ≤ x2, then, clearly, A−1 ≥ (A + x1I)−1 ≥ (A + x2I)−1 and (2.9) ((A + xI)−1f, f ) = lim x↓0(cid:13)(cid:13)(cid:13)(Ao + xI)− 1 2 2 PAf(cid:13)(cid:13)(cid:13) =(cid:26) A−1[f ], +∞, f ∈ mul A 0, [− 1 2 ] o PAf2, PAf ∈ ran A A +∞, f ∈ H \ ran A 1 2 o 1 2 lim x↓0 = f ∈ ran A 1 2 f ∈ H \ ran A . 1 2 8 YU.M. ARLINSKII Let A and B be two nonnegative selfadjoint l.r.. If dom A∩dom B 6= 0, then it is naturally defined the sum A + B := {{f, f′ + g′} : {f, f′} ∈ A, {f, g′} ∈ B} which is nonnegative symmetric l.r. and it, in general, is not selfadjoint. On the other side, one can define a quadratic form g[ϕ] = A[ϕ] + B[ϕ], ϕ ∈ D[A] ∩ D[B], which is nonnegative and closed [33, 25, 29]. Hence, by the first representation theorem, [33, 46] there is a nonnegative selfadjoint l.r. associated with g. This l.r. is called the form sum of A and B [24, 25, 29, 33] and is denoted by A +B. Observe that A +B = {0} ⊕ H ⇐⇒ D[A] ∩ D[B] = {0}. 3. Shorted operators for non-negative selfadjoint linear relations Theorem 3.1. Let A be a nonnegative selfadjoint l.r. in the Hilbert space H and let L be a subspace of H. Then the set (3.1) Ξ(A,L) =neA is a nonnegative selfadjoint l.r., eA ≤ A, ran eA ⊆ Lo has a unique maximal element AL and ker AL ⊇ L⊥. Moreover, (1) the l.r. AL↾L is selfadjoint in L and (AL↾L)−1 is associated with the closed sesquilin- ear form cL defined as follows (3.2) cL[f, g] := A−1[f, g], f, g ∈ dom cL = ran A 1 2 ∩ L; (2) the equality ran (AL) 1 ran A 2 ∩ L = {0}; (3) if C(A) is the Cayley transform (1.12) of A, then 1 2 = ran A 1 2 ∩ L holds and AL = 0 (dom AL = H) if and only if (3.3) AL = C (I − (I − C(A))L) ; (4) AL is an operator (mul AL = {0}) if and only if mul A ∩ L = {0}; (5) AL ∈ B+(H) \ {0} if and only if the linear manifold (3.4) M :=nf ∈ D[A] : A 1 2 f ∈ Lo is a subspace and mul A ∩ L = {0}. Proof. The operator T := C(A) admits the following block operator matrix representation (3.5) T =(cid:20) D C C∗ F(cid:21) : L⊥ ⊕ L → L⊥ ⊕ L . If eA ∈ Ξ(A,L) and if eT = C(eA) = −I + 2(I + eA)−1 is the Cayley transform of eA, then because eA + I ≤ A + I and 2(eA + I)−1 = I + eT , 2(A + I)−1 = I + T , we conclude that I + eT ≥ I + T. On the other hand, the representation eA =n{(I + eT )g, (I − eT )g}, g ∈ Ho SHORTING, PARALLEL ADDITION AND FORM SUMS 9 0 and the inclusion ran eA ⊆ L yield that ran (I − eT ) ⊆ L. The latter is equivalent to the equality eT ↾L⊥ = IL⊥. It follows that eT is of the form eT =(cid:20)I eF − F(cid:21) : L⊥ 2 ]C(cid:17)∗ (I − D)[− 1 and (I + eT ) − (I + T ) ≥ 0, from (1.4) we get the following inequality 2 ]C ≥ 0. (I + eT ) − (I + T ) =(cid:20)I − D −C eF − F −(cid:16)(I − D)[− 1 [15, 20, 48]), that the block operator-matrix (3.5) is a selfadjoint contraction if 0 eF(cid:21) : L⊥ Recall (cf. and only if the following properties of the entries are valid ⊕ L L⊥ ⊕ L L⊥ ⊕ L −C∗ . Since ⊕ L → → From the structure (3.6) of entries of T we obtain that . is a contraction, is a selfadjoint contraction D ∈ [−IL⊥, IL⊥], C∗ = NDD, F = −NDN∗ + DN ∗XDN ∗, N ∈ B(DD,L) X ∈ B(DN ∗)  F ′ := F +(cid:16)(I − D)[− 1 Υ(T,L) := C (Ξ(A,L)) =neT ∈ [−I, I] : eT ↾L⊥ = IL⊥,eT ≥ To 2 ]C(cid:17)∗ (I − D)[− 1 2 ]C = NN∗ + DN ∗XDN ∗. the set Thus, −IL ≤ F ′ ≤ eF ≤ IL. Since the set Ξ(A,L) coincides with the Cayley transforms of (3.6) and we obtain that (I − T )L ="0 Therefore , L⊥ ⊕ L min eT∈Υ(BL)eT = T ′ =(cid:20)I 0 → max ⊕ L 0 F ′(cid:21) : L⊥ eA∈Ξ(A,L) eA = C(T ′) =: AL. 2 ]C(cid:17)∗ (I − D)[− 1 0 0 I − F −(cid:16)(I − D)[− 1 T ′ = I − (I − T )L, AL = C (I − (I − T )L) . Let (I − T )L be the Kreın shorted operator. Then (3.5) and (1.5) yield 2 ]C# =(cid:20)0 0 0 I − F ′(cid:21) . Due to the properties of the shorted operator we have It follows that L ∩ ran (I − T ) 1 2 = ran ((I − T )L) 1 2 = ran (I − F ′) 1 2 . AL = C(T ′) = {{(I + T ′)g, (I − T ′)g}, g ∈ H} =(cid:8){h, 0} : h ∈ L⊥(cid:9)L{{(IL + F ′)f, (IL − F ′)f} : f ∈ L} =(cid:0)L⊥ ⊕ {0}(cid:1)L C(F ′). 10 YU.M. ARLINSKII Hence AL↾L = C(F ′) ⊂ L ⊕ L. Set Because (see (1.2)), and we get that L′ := {f ∈ H : (I − T ) 1 2 f ∈ L}. (I − T )L = (I − T ) 1 2 PL′(I − T ) 1 2 (I − T )L↾ L = I − F ′ 2 ↾L′ = (I − F ′) 1 1 (I − T ) 2 V ′, where V ′ is an isometry from L′ onto ran (I − F ′). It follows that (I − T )[− 1 Now (2.7) yields that 2 ]f = (I − F ′)[− 1 2 ]f, f ∈ ran (I − T ) 2 ∩ L = ran (I − F ′) 1 1 2 . A−1[f, g] = (C(F ′))−1 [f, g] = (AL↾ L)−1 [f, g], f, g ∈ D[A−1] ∩ L = ran A 1 2 ∩ L. Using definition (3.2) of the form cL, we get mul AL = {0} ⇐⇒ ker(AL)−1 = {0} ⇐⇒ ker cL = {0} ⇐⇒ mul A ∩ L = {0}, AL ∈ B+(H) \ {0} ⇐⇒( mul A ∩ L = {0}, cL[f ] ≥ mf2 ∀f ∈ L ∩ ran A o , m > 0 1 2 . The latter is equivalent to the conditions: mul A ∩ L = {0} and the linear manifold defined in (3.4) is closed. (cid:3) Observe that since AL ≤ A and D[AL] = L⊥ ⊕ D[AL↾ L], we have D[A] ∩ L ⊆ D[AL↾L], AL[f ] ≤ A[f ] ∀f ∈ D[A] ∩ L. Let A1 and A2 be nonnegative selfadjoint l.r.. Suppose A1 ≤ A2. Then Ξ(A1,L) ⊂ Ξ(A2,L) for any subspace L in H. Hence The next proposition is an extension of the property (1.6) to the set of all nonnegative selfadjoint l.r.. Proposition 3.2. Let {An} be a sequences of nonnegative selfadjoint l.r. and let L be a subspace in H. Then in the strong resolvent convergence sense An ց A∞ =⇒ (An)L ց (A∞)L. (A1)L ≤ (A2)L. Proof. Set Tn := C(An), n ∈ N, T∞ := C(A∞). Then s − R − lim n→∞ An = A∞ =⇒ s − lim n→∞ Tn = T∞. Moreover, {An} is non-increasing ⇐⇒ {Tn} is non-decreasing ⇐⇒ {I−Tn} is non-increasing =⇒ {(I − Tn)L} is non-increasing for each subspace L in H. Because {I − Tn} ⊂ B+(H) we get [47] that in the strong convergence sense (I − Tn)L ց (I − T∞)L. Hence from (3.3) s − R − lim n→∞ The proof is complete. (An)L = s − R − lim n→∞ C (I − (I − Tn)L) = C (I − (I − T∞)L) = (A∞)L. (cid:3) SHORTING, PARALLEL ADDITION AND FORM SUMS 11 Replacing A by A−1 in Theorem 3.1 we get that (A−1)L = max{eB : eB ∈ Ξ(A−1,L)} = minneA : eA = eA∗, eA ≥ A, dom eA ⊆ Lo and ((A−1)L↾L)−1 is associated with the closed form (3.7) Besides, ran ((A−1)L) Theorem 3.3. Let A be a nonnegative selfadjoint l.r. in the Hilbert space H and let L be a subspace of H. Then Let MA,L(λ) be the Nevanlinna family defined in (2.1). Then 2 = D[A] ∩ L and (A−1)L = 0 if and only if D[A] ∩ L = {0}. dL[ϕ, ψ] = A[ϕ, ψ], ϕ, ψ ∈ D[A] ∩ L. AL ≤ ((A−1)L)−1. 1 Proof. If C(A) = T , then C(A−1) = −T . Hence from (3.3) we get (3.8) Then (3.8) implies λ↓−∞ MA,L(λ)(cid:19) . AL↾L = −(cid:18)s − R − lim λ↑0 MA,L(λ)(cid:19) , ((A−1)L)−1↾L = −(cid:18)s − R − lim (A−1)L = C (I − (I + T )L) . (cid:0)(A−1)L(cid:1)−1 (3.9) The equality (I − T ) + (I + T ) = 2I and the definition of the shorted operator produce the inequality = C ((I + T )L − I) . Hence From (2.8) it follows that (I − T )L + (I + T )L ≤ 2I. I − (I − T )L ≥ (I + T )L − I. The operator T takes the block operator matrix form (3.5). The entries C and F of T admit the representations (3.6), i.e., AL = C (I − (I − T )L) ≤ C ((I + T )L − I) =(cid:0)(A−1)L(cid:1)−1 . C∗ = NDD, C = DDN∗, F = −NDN∗ + DN ∗XDN ∗. Φ(z) = F + zC∗(IL⊥ − zD)−1C, z ∈ C \ {(−∞,−1] ∪ [1, +∞)}. Define the function (3.10) The function Φ belongs to the class RS(L) [13], i.e., Φ is a Nevanlinna function and the Schur class function in the unit disk D. Then there exist strong limit values Φ(±1) and 2 ]C = −NN∗ + DN ∗XDN ∗, Φ(−1) = s − lim x↓−1 Φ(1) = s − lim x↑1 Observe that 2 ]C(cid:17)∗ (I + D)[− 1 Φ(x) = F −(cid:16)(I + D)[− 1 Φ(x) = F +(cid:16)(I − D)[− 1 2 ]C(cid:17)∗ (I − D)[− 1 0 I ± Φ(∓1)(cid:21) : L⊥ (I ± T )L =(cid:20)0 → 0 ⊕ L . L⊥ ⊕ L 2 ]C = NN∗ + DN ∗XDN ∗. 12 YU.M. ARLINSKII Therefore (3.3) and (3.9) yield AL↾L = C(Φ(1)), ((A−1)L)−1↾L = C(Φ(−1)). But then due to Φ(1) = C(AL↾L) and Φ(−1) = C(((A−1)L)−1↾L) we have −Φ(1) = C ((AL↾L)−1) , IL − Φ(1) = 2(IL + (AL↾L)−1↾ L)−1, IL + Φ(−1) = 2(cid:0)IL + ((A−1)L)−1↾L(cid:1)−1 Note that according to the Schur-Frobenius formula for the resolvent of block-operator matrix the relation (3.11) holds. PL(IH − zT )−1↾L = (IL − zΦ(z))−1, z ∈ C \ {(−∞,−1] ∪ [1, +∞)} Further we will use the following notations NA,L(λ) := PL(A − λI)−1↾L, λ ∈ C \ R+, NT,L(ξ) := PL(T − ξI)−1↾L, ξ ∈ C \ [−1, 1]. The resolvents of A and T are connected by the relations (A − λI)−1 = − 1 (A + I)−1 = 1 2 ⇐⇒ (A − λI)−1 =  (cid:0)N −1 A,L Hence (3.12) (λ) + I(cid:1)−1 = −(cid:18)I + 1 (T + I) , 1 − λ , λ ∈ ρ(A). 1+λ(cid:16)I + 2 1+λ I(cid:1)−1(cid:17) , λ ∈ ρ(A) \ {−1} 1−λ T(cid:1)−1 1+λ(cid:0)T − 1−λ (T + I)(cid:0)I − 1+λ = NA,L(λ) (NA,L(λ) + I)−1 1 + λ(cid:19)(cid:19)(cid:18)λ − 1 + λ(cid:19)(cid:19)−1 1 + λNT,L(cid:18)1 − λ 1 + λNT,L(cid:18) 1 − λ 2 T,L(cid:18)1 − λ = −(cid:18)N −1 T,L(cid:18)1 − λ 1 + λ(cid:19) + 1 + λ(cid:19)(cid:18)λN −1 1 + λ(cid:19) − 1 + λ(cid:19) 1 + λ(cid:19) = −N −1 T,L(cid:18)− −T,L(cid:18)1 − λ N −1 1 − λ 2 2 , 2 1 + λ(cid:19)−1 . From the equality C(A−1) = −T and the equality one gets (3.13) (cid:16)N −1 A−1,L Using (3.11) in the form (λ) + I(cid:17)−1 = −(cid:18)−N −1 T,L(cid:18)− 1 + λ(cid:19)(cid:18)−λN −1 T,L(cid:18)− NT,L(ξ) = (Φ(ξ−1) − ξI)−1, ξ ∈ C \ [−1, 1], 1 + λ(cid:19) + 1 − λ 2 where Φ(·) is defined by (3.10), we get the equalities (ξ) = Φ(1) − IL, s − lim s − lim ξ↓1 N −1 T,L ξ↑−1N −1 T,L (ξ) = IL + Φ(−1). 1 − λ 1 + λ(cid:19) − 2 1 + λ(cid:19)−1 . we have From (2.1) Hence Similarly = −MA,L(λ) − λIL, λ ∈ ρ(A). (cid:0)PL(A − λ)−1↾ L(cid:1)−1 AL↾L = −(cid:18)s − R − lim (A−1)L↾L = −(cid:18)s − R − lim (cid:0)((A−1)L)↾L(cid:1)−1 = −(cid:18)s − R − lim = −(cid:18)s − R − lim µ↓−∞MA,L(µ)(cid:19) λ↑0 λ↑0 MA,L(λ)(cid:19) . λ↑0 MA−1,L(λ)(cid:19) . (MA−1,L(λ))−1(cid:19) . Consequently Now (2.2) yields that (cid:0)((A−1)L↾L(cid:1)−1 SHORTING, PARALLEL ADDITION AND FORM SUMS 13 Using (3.12), (3.13) and since 2(I + (AL)−1↾L)−1 = I − Φ(1), 2(I + ((A−1)L)−1↾L)−1 = I + Φ(−1), (PL(A − λ)−1↾L)−1 , AL↾L = s − R − lim λ↑0 (PL(A−1 − λ)−1↾L)−1 . (A−1)L = s − R − lim λ↑0 = s − R − lim µ↓−∞(cid:16)µIL +(cid:0)PL(A − µI)−1↾L(cid:1)−1(cid:17) . (cid:3) Let A =(cid:20)A11 A12 A∗12 A22(cid:21) : L⊥ ⊕ L L⊥ ⊕ L Due to (1.4) the entry A12 admits the representation → be a bounded nonnegative selfadjoint operator in H. A12 = A 11Y A 1 2 1 2 22, Y ∈ B(ran A22, ran A11) is a contraction. 22(I−Y ∗Y )A From (1.5) we get the equality AL = A is associated with the form A[ϕ, ψ], ϕ, ψ ∈ L. Hence 1 2 1 2 22. In general A−1 is a l.r.. But ((A−1)L)−1 ((A−1)L↾L)−1 = A22. So, if A22 has bounded inverse, then (A−1)L is bounded and vice versa. If this is a case, then (A−1)L = A−1 22 . On the other side the function MA,L defined in (2.3) takes the form MA,L(λ) = −A22 + A 22Y ∗(A11 − λIL⊥)−1A11Y A 22, λ ∈ ρ(A11). 1 2 1 2 Hence −(cid:18)s − lim −(cid:18)s − lim λ↑0 M(λ)(cid:19) = A λ↓−∞M(λ)(cid:19) = A22 = ((A−1)L↾L)−1. 22(I − Y ∗Y )A 1 2 1 2 22 = AL↾L, 14 YU.M. ARLINSKII Remark 3.4. Suppose that codimL < ∞. Let A be a nonnegative selfadjoint l.r. in H. Then due to [50, 16] the l.r. PLA−1↾L and PLA↾L are selfadjoint and, therefore, they are associated with the closed forms cL and dL defined in (3.2) and in (3.7), respectively. Hence, due to Theorem 3.1 one gets AL↾L =(cid:0)PLA−1↾L(cid:1)−1 , ((A−1)L)−1 = PLA↾ L. 4. Parallel addition of nonnegative selfadjoint linear relations Definition 4.1. Let A and B be nonnegative selfadjoint l.r. in H. Then the nonnegative selfadjoint l.r. A : B defined as follows where + is the form sum, is said to be the parallel sum of A and B. The l.r. A : B =(cid:0)A−1 +B−1(cid:1)−1 , h(A, B) := 2(A : B) =(cid:18) A−1 +B−1 2 (cid:19)−1 is said to be the harmonic mean of A and B. Clearly, the parallel sum of two nonnegative selfadjoint l.r. is the nonnegative selfadjoint l.r. as well. From Definition 4.1 it follows that (4.1) and 1 A−1 +B−1 = (A : B)−1, A−1 : B−1 = (A +B)−1, 2 (cid:19)−1 2(A−1 +B−1) = (h(A, B))−1, h(A−1, B−1) =(cid:18)A +B mul (A : B) = ker(A−1 +B−1) = mul A ∩ mul B. , Hence, A : B is a selfadjoint operator if and only if mul A ∩ mul B = {0}. quadratic form The parallel sum A : B is a bounded operator if and only if mul A ∩ mul B = {0} and the q[g] := A−1[g] + B−1[g], g ∈ ran A 2 ∩ ran B 1 1 2 is positive definite. In particular, if A or B is a bounded nonnegative selfadjoint operator, then the nonnegative selfadjoint l.r. A−1 +B−1 has a bounded inverse, hence in this case the l.r. A : B is a bounded nonnegative selfadjoint operator. The theorem below shows that such way defined parallel addition for nonnegative selfad- joint l.r. preserves basic properties of parallel addition for bounded nonnegative selfadjoint linear operators. Theorem 4.2. The parallel sum of nonnegative selfadjoint l.r. possesses the properties: (1) λA : λB = λ(A : B), λ > 0, (2) A : B = B : A, (3) (A : B) : C = A : (B : C) =(cid:0)A−1 +B−1 +C−1(cid:1)−1 , (4) A : B ≤ A, A : B ≤ B, (5) A1 ≤ A2 =⇒ A1 : B ≤ A2 : B, 2 ∩ ran B (6) ran (A : B) (7) A : B = 0 (i.e., dom (A : B) = H and (A : B)f = 0 ∀f ∈ H) if and only if (8) if the quadratic form A−1[·] is the closed restriction of the quadratic form B−1[·], then 2 ∩ ran B 2 = {0}, 2 = ran A ran A 1 2 , 1 1 1 1 h(A : B) = A. SHORTING, PARALLEL ADDITION AND FORM SUMS 15 Besides, the following relation (4.2) A : B = s − R − lim n→∞(cid:18)(cid:18)A + 1 n I(cid:19) :(cid:18)B + 1 n I(cid:19)(cid:19) n→∞ (cid:18)A + = s − R − lim 1 n I(cid:19)−1 +(cid:18)B + I(cid:19)−1!−1 1 n is valid. Proof. The first three properties follows directly from Definition 4.1. Because A−1 +B−1 ≥ A−1, we get A : B =(cid:0)A−1 +B−1(cid:1)−1 ≤ A. Similarly, A : B ≤ B. A : B = 0 ⇐⇒ mul (A−1 +B−1) = H ⇐⇒ ran A If A1 ≤ A2, then A−1 +B−1. Hence 2 ∩ ran B 2 = {0}. 1 1 2 +B−1 ≥ A−1 +B−1(cid:1)−1 ≤(cid:0)A−1 2 2 = dom A− 1 +B−1(cid:1)−1 = A2 : B. 2 ∩ dom B− 1 2 = ran A 1 2 ∩ ran B 1 2 . Further, ran (A : B) If the quadratic form A−1[·] is the closed restriction of the quadratic form B−1[·], then for the form sum A−1 +B−1 one has that A−1 +B−1 = 2A−1. Hence A : B = 1 2A and h(A, B) = A. For each n ∈ N define the nonnegative selfadjoint l.r. Hn as follows 1 1 2 and A−1 1 1 ≥ A−1 A1 : B =(cid:0)A−1 2 = dom(cid:0)A−1 +B−1(cid:1) 1 Hn = (cid:18)A + Then {Hn} is a non-increasing sequence. By [17, Theorem 3.7] there exists 1 n Hn = s − R − lim H∞ := s − R − lim n→∞ 1 n I(cid:19)−1!−1 and . 1 n 1 n I(cid:19)−1!−1 I(cid:19)−1 +(cid:18)B + n→∞ (cid:18)A + I(cid:19)−1 n→∞(cid:13)(cid:13)(cid:13)(H− 1 n )og(cid:13)(cid:13)(cid:13) ∞Tn=1 n→∞(cid:13)(cid:13)(cid:13)(H− 1 n )og(cid:13)(cid:13)(cid:13) n : lim ran H 1 2 . 2 2 2 +(cid:18)B + < ∞(cid:27) , 2 H∞ ≤ A + I, H∞ ≤ B + I ∀n ∈ N. 1 n 1 n 1 2 ran H ∞ =(cid:26)g ∈ ∞ )og(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(H− 1 2 2 = lim Since H∞ ≤ Hn ∀n, we get Hence H∞ ≤ A and H∞ ≤ B. Note that is a bounded nonnegative seldfadjoint operator in H for each n ∈ N and H−1 n =(cid:18)A + I(cid:19)−1 1 n 1 n I(cid:19)−1 +(cid:18)B + g, g! + (cid:18)B + 1 n 1 n I(cid:19)−1 I(cid:19)−1 = (cid:18)A + g, g! , g ∈ H. 2 2 n g(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)H− 1 16 Besides YU.M. ARLINSKII ker H−1 n = mul A ∩ mul B ∀n ∈ N. For each vector h ∈ H the sequence of numbers (cid:26)(cid:13)(cid:13)(cid:13)H− 1 2(cid:27)∞ n g(cid:13)(cid:13)(cid:13) (2.9) it follows that 2 n=1 < ∞ ⇐⇒ g ∈ ran A 1 2 ∩ ran B 1 2 . 2 2 lim n→∞(cid:13)(cid:13)(cid:13)H− 1 n g(cid:13)(cid:13)(cid:13) 1 1 Thus 1 2 is non-decreasing. From 2 2 ran H 2 = D[A−1] ∩ D[B−1], ∞ = ran A = H−1 ∞ ∞ g(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)H− 1 2 ∩ ran B [g] = A−1[g] + B−1[g] = (A−1 +B−1)[g], g ∈ D[A−1] ∩ D[B−1]. = A−1 +B−1, H∞ =(cid:0)A−1 +B−1(cid:1)−1 = A : B, i.e., (4.2) holds true. Therefore, H−1 ∞ Theorem 4.3. Let A be a nonnegative selfadjoint l.r. in H and let L be a subspace of H. Then the following equality (cid:3) AL = s − R − lim t↑+∞ (A : tPL) holds. Proof. Set Then A(t) ≤ A for all t > 0, A(t) is a non-decreasing sequences of nonnegative selfadjoint l.r., A(t) := A : tPL, t > 0. ran (A(t)) 1 2 = ran A 1 2 ∩ L ∀t > 0. Hence ran A(t) ⊆ L for all t > 0 and, therefore, see (3.1), A(t) ∈ Ξ(A,L) for all t > 0. From [17, Theorem 3.1] it follows that there exists s − R − lim t↑+∞ A(t) =: A0. (4.3) Since ker A(t) ⊇ L⊥, we get ker A0 ⊇ L⊥ ⇐⇒ ran A0 ⊆ L. Hence A0 ∈ Ξ(A,L). , n ∈ N, t > 0. Let eA ∈ Ξ(A,L). Set 1 n I(cid:19)−1 I(cid:19)−1!−1 Because ker eA ⊇ L⊥ and ker PL = L⊥, PL↾L = IL, we have that IL(cid:19)−1 +(cid:18)tP + IL⊥, (eHn(t))−1↾L =(cid:18)eA + eHn(t) := (cid:18)eA + eHn(t)↾L⊥ = for all n ∈ N and for all t > 0. Hence 1 2n 1 n 1 n ↾L +(cid:18)t + (eHn(t))−1↾L = eA−1↾ L + t−1IL ∀t > 0. eA : tPL = s − R − lim n→∞ eHn(t) = eA. s − R − lim n→∞ Therefore, 1 n(cid:19)−1 ↾L SHORTING, PARALLEL ADDITION AND FORM SUMS 17 Now the inequality A ≥ eA for all eA ∈ Ξ(A,L) implies A : tPL ≥ eA : tPL for all t > 0 and letting t to +∞ from (4.3) we get A0 ≥ eA ∀eA ∈ Ξ(A,L). Due to Theorem 3.1, the shorted l.r. AL is the maximal element of Ξ(A,L). Therefore, A0 = AL. The proof is complete. (cid:3) Corollary 4.4. Let A and B be nonnegative selfadjoint l.r. and let L be a subspace. Then (A : B)L = AL : B = A : BL = AL : BL. Proof. By definition Since AL : B ≤ A : B and ran (AL : B) 1 (A : B)L = max{C = C∗, C ≥ 0 : C ≤ A : B, ran C ⊂ L} 1 1 1 2 = ran (AL) 2 ∩ ran B 2 = L ∩ ran A 2 ∩ ran B 1 2 ⊂ L, we get that (A : B)L ≥ AL : B and similarly (A : B)L ≥ A : BL. BL, one has On the other side because B : tPL is a non-decreasing function w.r.t. t > 0 and B : tPL ≤ (A : B)L = s − R − lim t↑+∞ ((A : B) : tPL) = s − R − lim t↑+∞ (A : (B : tPL)) ≤ A : BL Hence Hence and similarly (A : B)L = AL : B. A : BL ≥ (A : B)L ≥ A : BL ⇐⇒ (A : B)L = A : BL Since AL ≤ A, we get AL : BL ≤ A : BL = (A : B)L. On the other side AL ≥ A : tPL =⇒ AL : BL ≥ (A : tPL) : BL = (A : BL) : tPL ∀t > 0 AL : BL ≥ s − R − lim t↑+∞ ((A : BL) : tPL) = (A : BL)L = A : BL = (A : B)L. (cid:3) 5. The inequality between the harmonic and arithmetic means of nonnegative selfadjoint linear relations If a and b are positive numbers, then the harmonic mean and the arithmetic mean of a and b are connected by the inequality h(a, b) =(cid:18)a−1 + b−1 2 (cid:19)−1 a + b , 2 ≤ and the equality holds if and only if a = b. Let A, B ∈ B+(H) and let L := ran (A + B), then since A + B ≥ A, B, there exists a nonnegative selfadjoint contraction K in L such that A = (A + B) 1 2 K(A + B) 1 2 , B = (A + B) 1 2 (IL − K)(A + B) 1 2 . 1 1 Actually, since (A + B) [23] there exists a contraction X ∈ B(H, L) such that A 2 XX∗(A + B) (A + B) 2 f2 ≥ A 2 . The following equality is established in [10]: 2 f2 for all f ∈ H, by the Douglas factorization theorem 2 X. Then A = 2 = (A + B) 1 1 1 1 (5.1) h(A, B) = 2(A + B) 1 2 (IL − K)K(A + B) 1 2 . 18 YU.M. ARLINSKII It follows that h(A, B) ≤ A + B 2 and the equality holds if and only if A = B. The goal of this section is to prove a similar inequality for the case of nonnegative selfad- joint l.r.. First, we find the Cayley transforms of the arithmetic and harmonic means. Lemma 5.1. Let M be a nonnegative selfadjoint contraction in the Hilbert space L. Then (M − M 2)−1[f ] = M−1[f ] + (I − M)−1[f ] ∀f ∈ ran (M − M 2) 1 2 . Proof. It is sufficient to consider the case ker M = ker(I − M) = {0}. Then for f = (M − M 2) 2 u we get 1 M− 1 2 f = (I − M) 1 2 u, (I − M)− 1 2 f = M 1 2 u, (M − M 2)− 1 2 f = u. Hence (M − M 2)− 1 (I − M)− 1 2 f2 = u2, M− 1 2 f2 = M 2 u2. 1 2 f2 = (I − M) This completes the proof. 1 2 u2 = u2 − M 1 2 u2, (cid:3) Theorem 5.2. Let A and B be nonnegative selfadjoint l.r. and let TA := C(A), TB := C(B) be the Cayley transforms of A and B, respectively. Then (5.2) (5.3) Proof. Set Since C(cid:18)1 2 (A +B)(cid:19) = h(I + TA, I + TB) − I, C (h(A, B)) = I − h(I − TA, I − TB). 2(A +B), LA = I + TA, LB = I + TB, eS := 1 eT = h(I + TA, I + TB) − I, eL = I + eT . eT ≥ −I, eT ≤ h(I + TA, I + TB) = I +eL = 2(LA + LB) (I + TA + I + TB) − I = 1 2 1 2 the operator eT is a selfadjoint contraction. Using (5.1) we get the representation 2 (M − M 2)(LA + LB) (5.4) 2 , 1 1 where M is a nonnegative selfadjoint contraction acting in the subspace ran (LA + LB) and such that (TA + TB) ≤ I, (5.5) LA = (LA + LB) 1 2 M(LA + LB) 1 2 , LB = (LA + LB) From (2.7) we have 1 2 (I − M)(LA + LB) 1 2 . D[A] = ran L D[B] = ran L A ∩ ran L 1 2 1 2 1 2 A, A[v] + v2 = 2L−1 B, B[u] + u2 = 2L−1 B, eS[f ] + f2 = L−1 1 2 A [v], v ∈ D[A], B [u], u ∈ D[B], A [f ] + L−1 B [f ], f ∈ D[eS]. D[eS] = ran L SHORTING, PARALLEL ADDITION AND FORM SUMS 19 From (5.5) one gets , 1 2 A 1 1 2 2 1 1 1 1 2 1 2 Hence ran L L−1 ran L L−1 A = (LA + LB) B = (LA + LB) 2 ran M 2 , 2 ](LA + LB)[− 1 , v ∈ ran L 2 ran (I − M) 2 , 2 ](LA + LB)[− 1 A [v] =(cid:13)(cid:13)(cid:13)M [− 1 B [u] =(cid:13)(cid:13)(cid:13)(I − M)[− 1 2(cid:16)ran M 2 ]f(cid:13)(cid:13)(cid:13)    D[eS] = (LA + LB) eS[f ] + f2 =(cid:13)(cid:13)(cid:13)M [− 1 (cid:26) D[eS] = ran (I +eL) eS[f ] + f2 = (M − M 2)−1[(LA + LB)[− 1 Taking into account (2.7) we obtain that C(eS) =eL. 2 ]v(cid:13)(cid:13)(cid:13) 2 ]u(cid:13)(cid:13)(cid:13) 2(cid:17) = (LA + LB) +(cid:13)(cid:13)(cid:13)(I − M)[− 1 2 ]f ] = 2(I +eL)[− 1 Now Lemma 5.1 and (5.4) yield that 2 ∩ ran (I − M) 2 ](LA + LB)[− 1 Since 1 2 , 1 1 2 C(A−1) = −TA, C(B−1) = −TB, (A−1 +B−1)(cid:19) = h(I − TA, I − TB) − I. C(cid:18)1 2 we get Then , u ∈ ran L . 1 2 B 2 ](LA + LB)[− 1 1 1 2 2(cid:17) , , f ∈ D[eS] 2(cid:16)ran (M − M 2) 2 ]f(cid:13)(cid:13)(cid:13) 2 ]f2, f ∈ D[eS] . . (cid:3) C (h(A, B)) = C (cid:18)A−1 +B−1 2 (cid:19)−1! = I − h(I − TA, I − TB). Now from (1.12), (2.7) and (5.3) we get the equalities h(A, B) = {{(2I − h(I − C(A), I − C(B))) f, (h(I − C(A), I − C(B))) f} , f ∈ H} (5.6) D[A : B] = D[h(A, B)] = ran (2I − h(I − C(A), I − C(B))) 1 2 . The next statement is an analogue of [44, Theorem 2.5]. Corollary 5.3. Let {An} and {Bn} be two sequences of nonnegative selfadjoint l.r.. Then in the strong resolvent convergence sense An ր A∞, Bn ր B∞ =⇒ An +Bn ր A∞ +B∞, An ց A∞, Bn ց B∞ =⇒ (An : Bn) ց (A∞ : B∞). Proof. Set TAn := C(An), TBn := C(Bn), n ∈ N, TA∞ := C(A∞), TB∞ := C(B∞). in the strong sense. It follows from (5.2) and [44, Theorem 2.5.] that If An ր A∞ and Bn ր B∞ in the strong resolvent sense, then TAn ց TA∞ and TBn ց TB∞ C(cid:18)1 (An +Bn)(cid:19) = h(I + TAn, I + TBn) − I 2 = 2(I + TAn : I + TBn) − I ց 2(I + TA∞ : I + TB∞) − I = h(I + TA∞, I + TB∞) − I. 20 Now 1 2 (A∞ YU.M. ARLINSKII +B∞) = C (h(I + TA∞, I + TB∞) − I) = s − R − lim n→∞ 1 2 (An +Bn). Similarly, if An ց A∞ and Bn ց B∞, then, taking into account equality 5.3 and [44, Theorem 2.5.] once again, we get that (An : Bn) ց (A∞ : B∞). (cid:3) For nonnegative selfadjoint l.r. A and B let us define a nonnegative selfadjoint l.r. c0(A, B) as follows (5.7) c0(A, B) := C(cid:18) C(A) + C(B) =(cid:26)(cid:26)(cid:18)I + 2 If a and b are positive numbers, then (cid:19) C(A) + C(B) 2 (cid:19) f,(cid:18)I − C(A) + C(B) 2 (cid:19) f(cid:27) : f ∈ H(cid:27) . (5.8) c0(a, b) = a + b + 2ab 2 + a + b . Observe that the following inequalities are valid 2ab a + b ≤ a + b + 2ab 2 + a + b ≤ a + b , 2 and the sign " = " holds if and only if a = b. Theorem 5.4. Let A and B be a nonnegative selfadjoint l.r.. Then (1) D[c0(A, B)] = D[A] + D[B]; (2) ran (c0(A, B)) (3) the inclusion 2 = ran A 1 1 2 + ran B 1 2 ; (5.9) holds and the inequality D[A : B] ⊇ D[A] + D[B] h(A, B) ≤ c0(A, B) ≤ A +B 2 is valid; moreover, the sign " = " holds if and only if A = B. In addidion (5.10) Proof. Set ran (A +B) 1 2 ⊇ ran A 1 2 + ran B 1 2 . TA := C(A), TB := C(B). Because I ± TA and I ± TB are bounded nonnegative operators, we have the inequalities h(I ± TA, I ± TB) ≤ Using (5.2), (5.3), and (5.7) we get 1 2 (I ± TA + I ± TB) = I ± (TA + TB). 1 2 1 2 h(I + TA, I + TB) − I ≤ (TA + TB) ≤ I − h(I − TA, I − TB) ⇐⇒ C(cid:18)1 2 (A +B)(cid:19) ≤ C (c0(A, B)) ≤ C(h(A, B)) ⇐⇒ h(A, B) ≤ c0(A, B) ≤ 1 2 (A +B). SHORTING, PARALLEL ADDITION AND FORM SUMS 21 If A = B, then, clearly, c0(A, B) = A, 1 2 (A +B) = A, 1 2 (A−1 +B−1) = A−1 =⇒ h(A, B) = A. Thus, h(A, B) = c0(A, B) = 1 Suppose that h(A, B) = 1 2(A +B). 2(A +B) for some nonnegative selfadjoint l.r. A and B. Then I − h(I − TA, I − TB) = h(I + TA, I + TB) − I. (TA + TB) ≤ I − h(I − TA, I − TB) = h(I + TA, I + TB) − I ≤ 1 2 (TA + TB), Since we get 1 2 h(I + TA, I + TB) − I = 1 2 (TA + TB) ⇐⇒ h(I + TA, I + TB) = 1 2 (I + TA + I + TB). Because I + TA and I + TB are bounded, the latter equality yields the equality TA = TB and hence A = C(TA) = C(TB) = B. Similarly, one can prove that h(A, B) = c0(A, B) =⇒ A = B, and c0(A, B) = 1 A = B. 2(A +B) =⇒ From (2.7) and (5.3) we have D[A : B] = D[h(A, B)] = ran (2I − h(I − TA, I − TB)) 1 2 , D[c0(A, B)] = ran(cid:18)I + TA 2 + I + TB 2 2 (cid:19) 1 , ran (c0(A, B)) 1 2 = ran(cid:18)I − TA 2 + 2 I − TB 2 (cid:19) 1 . Since h(I ± TA, I ± TB) ≤ we obtain the inequality (I ± TA) + (I ± TB) 2 = I ± TA + TB 2 , 2I − h(I ± TA, I ± TB) ≥ Now the Douglas theorem [23] yields the inclusions I ∓ TA 2 + I ∓ TB 2 . ran (2I − h(I ± TA, I ± TB)) But, see [26, Theorem 2.2], 1 2 ⊇ ran(cid:18)I ∓ TA 2 + 2 I ∓ TB 2 (cid:19) 1 . ran(cid:18)I ± TA 2 Using (5.6) and equalities + 2 I ± TB 2 (cid:19) 1 = ran (I ± TA) 1 2 + ran (I ± TB) 1 2 . D[A] = ran (I + TA) we get the equalities 1 2 , D[B] = ran (I + TB) ran A 1 2 , 2 = ran (I − TA) 1 1 2 , ran B 1 2 = ran (I − TA) 1 2 D[c0(A, B)] = D[A] + D[B], ran (c0(A, B)) 1 2 = ran A 1 2 + ran B 1 2 , 22 YU.M. ARLINSKII and finally arrive at (5.9). Since ran (2I − h(I + TA, I + TB)) 1 2 ⊇ ran ((I − TA) + (I − TB)) 1 2 , we get (5.10). The proof is complete. (cid:3) As a consequence of (5.9) we note, that if one of two nonnegative selfadjoint l.r. is a linear operator, then their parallel sum is a nonnegative selfadjoint linear operator as well, moreover, if one of two is a bounded operator, then the parallel sum is a bounded operator. 6. Arithmetic -- harmonic means We consider analogs of the arithmetic-harmonic means for the case of nonnegative selfad- joint l.r.. Theorem 6.1. Let A and B be nonnegative selfadjoint l.r.. Define (6.1) A0 := A, B0 := B, An := 1 2 (An−1 +Bn−1), Bn := h(An−1, Bn−1), n ∈ N. Then the sequence {An} is a non-increasing, while the sequence {Bn} is a non-decreasing and An ≥ Bn for all n ∈ N. Moreover, the nonnegative selfadjoint l.r. Bn (6.2) A∞ := s − R − lim n→∞ An, B∞ := s − R − lim n→∞ posses the following properties: 1 2 2(A +B), h(A, B) ≤ B∞ ≤ A∞ ≤ 1 D[B∞] ⊇ D[A∞] ⊇ D[A] ∩ D[B], ran A A∞[ϕ] = B∞[ϕ] ∀ϕ ∈ D[A∞], ∞ [ψ] = B−1 A−1 ∞ [ψ] ∀ψ ∈ ran B ∞. ∞ ⊇ ran B ∞ ⊇ ran A 1 2 1 2 1 2 ∩ ran B 1 2 , If A−1 and B−1 are bounded (i.e. they are graphs of bounded nonnegative selfadjoint opera- tors), then Proof. By induction from the definitions of form sums, harmonic means and from Theorem 5.4 one can show that A∞ = B∞ = (A−1#B−1)−1. D[An] = D[A] ∩ D[B], D[Bn] ⊇ D[A] + D[B], h(A, B) = B1 ≤ Bn ≤ Bn+1 ≤ An+1 ≤ An ≤ A1 = 1 2(A +B) ∀n ∈ N. Since the sequence {Bn} is a non-decreasing and the sequence {An} is a non-increasing, from [33, Theorem VII.3.11], [49, Theorem 3.1, Theorem 3.2], [17, Theorem 4.2, Theorem 4.3] it follows that the strong resolvent limits of {An} and {Bn} exist and h(A, B) ≤ B∞ ≤ A∞ ≤ Since D[B∞] =(h ∈ ∞\n=1 1 2 (A +B), D[B∞] ⊇ D[A∞]. Bn[h] < ∞) D[Bn] : lim n→∞ SHORTING, PARALLEL ADDITION AND FORM SUMS 23 and Bn ≤ A1 for all n ∈ N, ∞Tn=1D[Bn] ⊇ D[A] ∩ D[B] = D[A1], we get that D[B∞] ⊇ D[A] ∩ D[B]. Fix n ∈ N. Then for any f ∈ D[A] ∩ D[B] by definition and from Bn+1 ≥ Bn we have An+1[f ] − Bn+1[f ] = Bn[f ] − Bn+1[f ] An[f ] + 1 2 1 2 Hence It follows that (6.3) 0 ≤ An+1[f ] − Bn+1[f ] ≤ lim n→∞ An[f ] = lim n→∞ Define the sesquilinear form t as follows 1 2 ≤ An[f ] + 1 2 Bn[f ] − Bn[f ] ≤ An[f ] − Bn[f ] . 2 A1[f ] − B1[f ] 2n ∀f ∈ D[A] ∩ D[B], ∀n ∈ N. Bn[f ] = B∞[f ] ∀f ∈ D[A] ∩ D[B]. dom t = D[An], t[f, g] = lim n→∞ An[f, g], f, g ∈ dom t. ∞[n=1 Since D[An] = D[A] ∩ D[B], we have the equality dom t = D[A] ∩ D[B] and from (6.3): t[f, g] = B∞[f, g] ∀f, g,∈ D[A] ∩ D[B]. Since the form B∞[·,·] restricted on D[A] ∩ D[B] is closable, using [49, Theorem 3.2], [17, Theorem 4.3], we obtain that the closure of the form t[·,·] coincides with A∞[·,·]. Hence A∞[ϕ] = B∞[ϕ] for all ϕ ∈ D[A∞]. Since n = 2(cid:0)An−1 +Bn−1(cid:1)−1 = h(A−1 A−1 n = (h(An−1, Bn−1))−1 = 1 B−1 2(A−1 n−1, B−1 n−1), n−1 +B−1 n−1), n ∈ N, the sequences {A−1 tively, and n },{B−1 n } ⊂ B+(H) are the non-decreasing and non-increasing, respec- Arguing as above we see that s − R − lim n→∞ A−1 n = A−1 ∞ , s − R − lim n→∞ B−1 n = B−1 ∞ . 1 2 , ran A B−1 ∞ ≥ A−1 ∞ A−1 ∞ [ψ] = B−1 ∞ ⊇ ran B ∞ ⊇ ran A 1 2 ∩ ran B 1 2 , 1 2 ∞ [ψ] ∀ψ ∈ ran B 1 2 ∞. 0 = A−1 and B−1 0 = B−1 are bounded. Then from (4.1) it follows that n } ⊂ B+(H), and there is a common limit B−1 s − lim n→∞ A−1 n = s − lim n→∞ n = A−1#B−1. and Suppose that A−1 n },{B−1 {A−1 Due to the equalities A∞ =(cid:16)s − lim n→∞ A−1 n (cid:17)−1 , B∞ =(cid:16)s − lim n→∞ B−1 n (cid:17)−1 , (cid:3) we get A∞ = B∞ = (A−1#B−1)−1 . The proof is complete. 24 YU.M. ARLINSKII In the sequel we will denote by ah(A, B) the arithmetic -- harmonic means of nonnegative selfadjoint l.r. A and B. By Theorem 6.1 we have where A∞ and B∞ are defined in (6.1) and (6.2). Then from (4.1) ah(A, B) = hA∞, B∞i , ah(A−1, B−1) =(cid:10)B−1 ∞ , A−1 ∞(cid:11) . The next corollaries, show that in general A∞ 6= B∞. Corollary 6.2. Let B be a nonnegative selfadjoint l.r.. Define B′ := {dom B ⊕ {0}}L{{0} ⊕ mul B} =(cid:8){ϕ, ψ} : ϕ ∈ dom B, ψ ∈ mul B(cid:9) , B′′ := {ker B ⊕ {0}}L{{0} ⊕ ran B} = {{f, g} : f ∈ ker B, g ∈ ran B} . Then (1) for A = 0 (dom A = H) one has ah(A, B) = hB′, Ai , (2) for A = {0} ⊕ H one has ah(A, B) = hA,B′′i . Proof. (1) If A is zero operator, defined on H, then for sequences {An} and {Bn}, defined in (6.1), we have A1 := 1 2B, B1 := h(A, B) = A, 2 (A +B) = 1 An = 1 2nB, Bn = A, n ∈ N. Thus A∞ = B′, B∞ = A. and the arguments above. (2) If A = {0} ⊕ H, then A−1 is the zero operator defined on H. Hence we can use (4.1) (cid:3) Corollary 6.3. Let A be a nonnegative selfadjoint l.r.. Set (6.4) Then ah(A, A−1) =(cid:10)(PM⊥)−1 , PM⊥(cid:11). In particular ah(A, A−1) = I ⇐⇒ M = {0}. Proof. Set T := C(A), then C(A−1) = −T, M := mul A ⊕ ker A. (A +A−1), B1 := h(A, A−1) =(cid:18) 1 2 (A−1 +A)(cid:19)−1 = A−1 1 . A1 := 1 2 By induction An := 1 2 (An−1 +Bn−1) = 1 2 Using (5.2) and (5.3) we get (An−1 +A−1 n−1), Bn := h(An−1, Bn−1) = A−1 n , n ≥ 2. C(A1) = h(I + T, I − T ) − I = 2(I + T )(I − T )(I + T + I − T )−1 − I = −T 2, C(B1) = C(A−1 , C(Bn) = T 2n 1 ) = −C(A1) = T 2, C(An) = −T 2n , n ≥ 2. Since T is a selfadjoint contraction and M = ker(I − T 2), where M is defined by (6.4), we get the equality It follows that s − lim n→∞ T 2n = PM. B∞ = C(PM) = (I − PM)(I + PM)−1 = PM⊥, A∞ = C(−PM) = (PM⊥)−1 . The proof is complete. (cid:3) SHORTING, PARALLEL ADDITION AND FORM SUMS 25 7. Arithmetic, harmonic, arithmetic -- harmonic means and nonnegative selfadjoint extensions of nonnegative symmetric linear relations Basic results of the Kreın theory [34] can be extended to the case of a non-densely defined symmetric operator or a symmetric l.r. [7, 11, 18, 28, 29, 30]. Let S be a nonnegative in H. Then the set of all its nonnegative selfadjoint extensions consists symmetric l.r. of maximal and minimal elements. The maximal nonnegative selfadjoint extension is the Friedrichs extension SF [46], which is associated with the closure of the sesquilinear form The minimal nonnegative selfadjoint extension SK is called the Kreın or the Kreın-von Neu- mann extension and can be defined as follows [18]: s[f, g] = (f′, g), {f, f′},{g, g′} ∈ S. SK = ((S−1)F)−1. S is called extremal [11] if The closed form SF[·,·] is a closed restriction of the closed form eS[·,·] associated with any nonnegative selfadjoint extension eS of S [34, 11]. A nonnegative selfadjoint extension eS of It is proved in [11, Proposition 3] that if eS is an extremal extension of S, then the closed form eS[·,·] is a closed restriction of the closed form SK[·,·] associated with the Kreın extension of Let Q = C(S) be the Cayley transform of S. Then Q is a non-densely defined Hermitian for all {f, f′} ∈ eS. inf {(f′ − g′, f − g) : {g, g′} ∈ S} = 0 contraction with dom Q = ran (S + I). There is a one-to-one correspondence [34, 18] S. between the set of all nonnegative selfadjoint extensions and the set of all selfadjoint con- tractive extensions of Q. Let eQ = C(eS), eS = C(eQ) Qµ = C(SF), QM = C(SK). The operators Qµ and QM posses the following properties [34]: if N := H ⊖ dom Q, then (7.1) N is the deficiency subspace of S corresponding to λ = −1. Set N0 := ran (QM − Qµ). The set of all selfadjoint contractive extensions of Q forms the operator interval [Qµ, QM ] [34], which admits the parameterizations (I + Qµ)N = (I − QM )N = 0. 1 2 1 1 2 (QM − Qµ) 1 2 1 = Qµ + (Qµ + QM ) + (QM − Qµ) 2eZ(QM − Qµ) (7.2) eQ = 2 (IN0 −eZ)(QM − Qµ) A nonnegative selfadjoint extension eS is extremal if and only if the parameter eZ for eQ = C(eS) in the subspace N0 [12]. This is equivalent to the equality (I − eQ2)N = 0, see [12]. in the right hand side in (7.2) is a selfadjoint and unitary operator (a fundamental symmetry) If 2 (IN0 +eZ)(QM − Qµ) dim N0 = 1, then extremal extensions of S are only SF and SK. From (7.2) and (7.1) it follows the equalities (QM − Qµ) 2 = QM − 1 2 1 2 1 1 1 2 . (7.3) 1 2 (QM − Qµ) (I ± eQ)N = 1 2 (IN0 ± eZ)(QM − Qµ) 1 2 . 26 YU.M. ARLINSKII Note that from (7.1) and (3.3), (3.8) one gets the equalities for shorted operators (SK)N = 0, ((SF)−1)N = 0. Moreover, the Kreın uniqueness criteria [34] is equivalent to the equality (SF)N = 0. The next theorem shows connections of the arithmetic and harmonic means with the theory of nonnegative selfadjoint extensions of nonnegative symmetric l.r.. Theorem 7.1. Let S be a nonnegative symmetric l.r.. Then the operators mean 1 given by (5.7), are nonnegative selfadjoint extensions of S; (1) for an arbitrary nonnegative selfadjoint extensions eS1 and eS2 of S their arithmetic 2(eS1 +S2), harmonic mean h(eS1,eS2), and nonnegative selfadjoint l.r. c0(eS1,eS2), (2) if eZ1 and eZ2 are parameters of C(eS1) and C(eS2) in formulae (7.2), respectively, then the corresponding parameters for C(cid:16) 1 2(eS1 +S2)(cid:17), C(cid:16)h(eS1, S2)(cid:17) and C(c0(eS1,eS2)) are h(IN0 + eZ1, IN0 + eZ2) − IN0, IN0 − h(IN0 − eZ1, IN0 − eZ2), (3) for an arbitrary nonnegative selfadjoint extension eS of S the equalities (SF +eS) = SF, h(SK,eS) = SK (4) if eS1 and eS2 are extremal extensions of S, then 1 2(eS1 +S2) and h(eS1,eS2) are extremal extensions of S as well, moreover, for an arbitrary extremal extension eS one has (eZ1 + eZ2), respectively; hold; and 1 2 1 2 1 2 (SK +eS) = eS, h(SF,eS) = eS. (7.4) (7.5) and Proof. If eS is a nonnegative selfadjoint extension of S, then (see [11, 18, 29, 30]) K , the form S−1 K ] = D[S−1 K ]. • SK ≤ eS ≤ SF, the form SF[·,·] is the closed restriction of the form eS[·,·], and D[eS] ∩ D[SF] = D[SF]. • eS−1 is a nonnegative selfadjoint extension of S−1, S−1 F ≤ eS−1 ≤ S−1 is the closed restriction of the form eS−1[·,·], and D[eS−1] ∩ D[S−1 Let eS1 and eS2 be two nonnegative selfadjoint extensions of S. Then 1 ] ∩ D[eS−1 2 ] ⊇ D[S−1 K ], 2(cid:16)eS1[f ] +eS2[f ](cid:17) ≥ SK[f ], f ∈ D[eS1] ∩ D[eS2], 2(eS1 +eS2)[f ] = 1 2(eS1 +eS2)[ϕ] = SF[ϕ], ϕ ∈ D[SF], 2(cid:16)eS−1 1 [g] +eS−1 F [g], g ∈ D[eS−1 2(eS−1 2(eS−1 K [ψ], ψ ∈ D[S−1 K ]. 2(eS−1 2(eS1 +eS2) is a nonnegative selfadjoint extension of S and 1 D[eS1] ∩ D[eS2] ⊇ D[SF], D[eS−1 Hence 1 selfadjoint extension of S−1. It follows that 1 ] ∩ D[eS−1 +eS−1 2 [g](cid:17) ≥ S−1 2 )[g] = 1 2 )[ψ] = S−1 +eS−1 +eS−1 2 ) is a nonnegative K [·,·] 2 ], 1 1 1 1 1 1 1 h(eS1,eS2) = 2(eS−1 1 +eS−1 2 )−1 SHORTING, PARALLEL ADDITION AND FORM SUMS 27 is a nonnegative selfadjoint extension of S. Besides, we get equalities in (7.4). 1 1 1 2 2 , 1 2 1 2 1 2 1 2 1 2 (Qµ + QM ) + (Qµ + QM ) + (QM − Qµ) (QM − Qµ) selfadjoint contraction and by means of expression (7.2): the operator 1 C( 1 Set Q = C(S), Qµ = C(SF), QM = C(SK), N = H ⊖ dom Q, N0 = ran (QM − Qµ). From (5.2) and (5.3) we get the equalities Let eQ1 := C(eS1) and eQ2 := C(eS2). The operators eQk, k = 1, 2 admit the representations 2eZk(QM − Qµ) eQk = where eZk, k = 1, 2 are selfadjoint contractions in N0. Since the operator 1 2(eZ1 + eZ2) is a (eZ1 + eZ2)(cid:19) (QM − Qµ) 2(cid:18)1 (eQ1 + eQ2) = 2(eQ1 + eQ2) is a selfadjoint contractive extension of Q. Hence c0(eS1,eS2) = 2(eQ1 + eQ2)) is a nonnegative selfadjoint extension of S. (eS1 +eS2)(cid:19) = h(I + eQ1, I + eQ2), I − C(cid:16)h(eS1,eS2)(cid:17) = h(I − eQ1, I − eQ2). I + C(cid:18)1 (cid:18)I + C(cid:18)1 (eS1 +eS2)(cid:19)(cid:19)N = h(cid:16)(I + eQ1)N, (I + eQ2)N(cid:17) . h(cid:16)(I + eQ1)N, (I + eQ2)N(cid:17) = 2(cid:18)1 =(cid:16)h(I + eQ1, I + eQ2)(cid:17)N 2(cid:19) :(cid:18) 1 From (7.3) and [10, Proposition 1] it follows the equality Further, using (1.11), one obtains (QM − Qµ) 2 , 2 2 2 2 1 1 1 (QM − Qµ) 2(cid:19) 2 (IN0 + eZ2)(QM − Qµ) 2 h(IN0 + eZ1, IN0 + eZ2)(QM − Qµ) 2 . 1 1 1 (QM − Qµ) Now from (7.2), (7.3) and the uniqueness of the representation we get that (Qµ + QM ) + 1 2 (QM − Qµ) 1 2 , 1 2eZ(QM − Qµ) 2fW (QM − Qµ) 1 1 2 (Qµ + QM ) + (QM − Qµ) where eZ = h(IN0 + eZ1, IN0 + eZ2) − IN0. Similarly wherefW = IN0 − h(IN0 − eZ1, IN0 − eZ2). Let eS be an extremal nonnegative selfadjoint extension of S and let eQ = C(eS) be its Cayley transform. Then eQ admits the representation (7.2) with a fundamental symmetry eZ in N0. The operator eS−1 is a nonnegative selfadjoint extension of the operator S−1 and C(S−1) = −Q, C((S−1)F) = −QM , C((S−1)K) = −Qµ, C(eS−1) = −eQ. Hence 2 , 1 1 1 (Qµ + QM ) + 1 2 (QM − Qµ) 2 (−eZ)(QM − Qµ) 2 . 1 = 1 2 2 (IN0 + eZ1)(QM − Qµ) (eS1 +eS2)(cid:19) = C(cid:18)1 C(cid:16)h(eS1,eS2)(cid:17) = 1 2 1 2 2 1 2 −eQ = − 28 YU.M. ARLINSKII sesquilinear form 1 that 1 Because eZ is a selfadjoint and unitary (in N0), the operator −eZ is selfadjoint and unitary as well. Therefore, the operator eS−1 is an extremal extension of S−1. Let eS1 and eS2 be two extremal nonnegative selfadjoint extensions of S. Then the closed form eS1[·,·] and eS2[·,·] are closed restrictions of the closed form SK[·,·]. Therefore the 2(eS1 +eS2)[·,·] is a closed restriction of the closed form SK[·,·]. It follows 2 (eS1 +eS2) is an extremal extension of S. 2 (cid:17) +eS−1 Since eS−1 h(eS1,eS2) =(cid:18)1 2 (cid:17)(cid:19)−1 +eS−1 For an arbitrary extremal extension eS of S we have equalities 2 are extremal nonnegative selfadjoint extensions of S−1, 1 is an extremal extension of S−1 and this implies that 2(cid:16)eS−1 and eS−1 is an extremal extension of S. 2(cid:16)eS−1 1 1 1 and 1 2 (S−1 F 1 2 (SK +eS)[f ] = eS[f ] ∀f ∈ D[(SK +eS)/2] = D[eS], +eS−1)[h] = eS−1[h] ∀h ∈ D[(S−1 +eS−1)/2] = D[eS−1]. F It follows that equalities in (7.5) are valid. The proof is complete. extensions too. Moreover, (cid:3) Proposition 7.2. Let S be a nonnegative symmetric l.r. in H, having a non-unique nonneg- ative selfadjoint extension. Then the arithmetic-harmonic means ah(eS1,eS2) of an arbitrary two extremal nonnegative selfadjoint extensions eS1, eS2 are extremal nonnegative selfadjoint where eS(1) 2 = h(eS1,eS2). In particular, ah(SF, SK) = hSF, SKi. ah(eS1,eS2) =DeS(1) 2 E , 1 ,eS(1) Proof. Set A = SF, B = SK. From Theorem 7.1 it follows that 2(eS1 +eS2), eS(1) 1 = 1 h(A, B) = h(SF, SK) = SK = B, 1 2 Then from (6.1) and (6.2) we get A∞ = SF and B∞ = SK. (A +B) = 1 2 (SF +SK) = SF = A. Recall [26, Theorem 4.3] that if L1, L2 are two subspaces in the Hilbert space H, then (7.6) 2(PL1 : PL2) = PL1∩L2. 1 Let eSk, k = 1, 2 be two extremal extensions of S. Then the corresponding operators eZk, k = 1, 2 in (7.2) for C(eSk) are fundamental symmetries in N0. Therefore, the operators 2(IN0 ± eZk) are orthogonal projections in N0. Hence, the operators 2 2 (IN0 ± eZ1)(cid:19) :(cid:18)1 2(cid:18)1 (IN0 ± eZ2)(cid:19) = 1 = h(IN0 + eZ1, IN0 + eZ2) − IN0, eZ (1) eZ (1) 1 2 h(IN0 ± eZ1, IN0 ± eZ2) 2 = IN0 − h(IN0 − eZ1, IN0 − eZ2) are orthogonal projections in N0. It follows that the operators SHORTING, PARALLEL ADDITION AND FORM SUMS 29 are fundamental symmetries in N0. Define two sequences 1 ), 1 2 1 eS(0) 1 = 1 2 1 2 1 2 1 1 . Hence ), n ≥ 2. ), n ∈ N. 2 n ≥ 2. 1 2 1 Let us show that ) − IN0, 2 Since the operators 1 2 ,eS(n−1) 2 := eS1, eS(0) 1 = C(eS(n) Define also two sequences of operators in N0: 2 ) in their representation in (7.2), respectively. ), eS(n) 2 = h(eS(n−1) := eS2, eS(n) (eS(n−1) +eS(n−1) eS(n) 1 = eS(1) 1 , eS(n) 2 = eS(1) , IN0 + eZ (n−1) 1 = h(IN0 + eZ (n−1) eZ (n) , IN0 − eZ (n−1) 2 = IN0 − h(IN0 − eZ (n−1) eZ (n) correspond for all n ∈ N to eQ(n) and eZ (n) By Theorem 7.1 the operators eZ (n) eQ(n) 2 = C(eS(n) 2 ≥ eZ (1) 1 , we get eZ (1) 2 ≤ eS(1) Because eS(1) 1 ≥ IN0 − eZ (1) 1 , IN0 − eZ (1) 2 ≥ IN0 + eZ (1) IN0 + eZ (1) 2 (IN0±eZ (1) 2 (IN0±eZ (1) 2 = h(IN0 − eZ (1) 2 ), IN0 − eZ (2) 1 , IN0 + eZ (1) 1 = h(IN0 + eZ (1) IN0 + eZ (2) 2 = IN0 − eZ (1) 1 , IN0 − eZ (2) 1 = IN0 + eZ (1) IN0 + eZ (2) Therefore eZ (2) 1 = eZ (1) 1 , eZ (2) 2 = eZ (1) eZ (n) 1 , eZ (n) 1 = eZ (1) 2 = eS(1) 1 , eS(n) 1 = eS(1) 2 , eS(n) 2 = eZ (1) 2 E . By Theorem 7.1 the operators eS(1) Thus, ah(eS1,eS2) =DeS(1) 1 , eS(1) 1 ,eS(1) dx2,  2 (R+), f′(0) = cf (0)} , c > 0, c 6= ∞, 2 (R+), f′(0) = 0} , 2 (R+), f (0) = 0} , dom Lc = {f ∈ W 2 dom L0 = {f ∈ W 2 dom L∞ = {f ∈ W 2 are extremal non- negative selfadjoint extensions of S. (cid:3) Proposition 7.3. Consider in the Hilbert space L2(R+) the following differential operators 2 ) are orthogonal projections in N0, the equalities 1 , IN0 − eZ (1) 2 (R+) is the Sobolev space. Then and equality (7.6) imply that 2 , and by induction 2 ∀n ≥ 2. d2f Lcf = − 2 . 2 . 1 ), 2 ) where W 2 1 1 1 2 (c+d), c0(Lc, Ld) = Lc0(c,d), h(Lc, Ld) = Lh(c,d) = L 2cd 2(Lc +Ld) = L 1 2(Lc +L∞) = L∞, c0(Lc, L∞) = L2c+1, h(Lc, L∞) = L2c, ah(Lc, L∞) = L∞, c > 0 2(Lc +L0) = Lc/2, c0(Lc, L0) = L c 2+c , h(Lc, L0) = L0, ah(Lc, L0) = L0, c > 0 2(L0 +L∞) = L∞, c0(L0, L∞) = L1, h(L0, L∞) = L0, ah(L∞, L0) = hL∞, L0i . Proof. The operators Lc are nonnegative selfadjoint extensions of the nonnegative symmetric operator , ah(Lc, Ld) = L√cd, c+d 1 The operator S has one-dimensional deficiency subspaces. In particular, Sf = − d2f dx2, dom S =(cid:8)f ∈ W 2 2 (R+), f (0) = f′(0) = 0(cid:9) . N = {ξ exp(−x), ξ ∈ C} 30 YU.M. ARLINSKII is the deficiency subspace corresponding to λ = −1. One can show that (Lc + I)−1 exp(−x) = 1 2 exp(−x)(cid:18)x + 1 c + 1(cid:19) , x ∈ R+. ((Lc + I)−1 − (L∞ + I)−1) exp(−x) = ((L0 + I)−1 − (L∞ + I)−1) exp(−x) = 1 exp(−x) , 2(c + 1) 2 exp(−x). Then (7.7) Set Hence The operators L∞ and L0 are the Friedrichs and the Kreın extension of S, respectively. Hence as it is shown above 1 2 (L0 +L∞) = L∞, h(L0, L∞) = L0, ah(L∞, L0) = hL∞, L0i . eQc = C(Lc), Qµ = C(L∞), QM = C(L0). From (1.12) we have eQc − Qµ = 2(cid:0)(Lc + I)−1 − (L∞ + I)−1(cid:1) , QM − Qµ = 2(cid:0)(L0 + I)−1 − (L∞ + I)−1(cid:1) and ran (eQc − Qµ) = ran (QM − Qµ) = N. Now from (7.2) it follows that 2(cid:0)(Lc + I)−1 − (L∞ + I)−1(cid:1) = (1 +ezc)(cid:0)(L0 + I)−1 − (L∞ + I)−1(cid:1) . Using (7.7) we obtain 1 − c 1 + c , z0 = 1, z∞ = −1. ezc = h(1 +ezd, 1 +ezd) − 1 = 1 − 1 1 + 1 2(c + d) 2 (c + d) , 1 − h(1 −ezd, 1 −ezd) = 1 − 1 + 2cd c + d 2cd c + d Applying Theorem 7.1 and (5.7), (5.8), we get 1 1 2 (c+d), c0(Lc, Ld) = Lc0(c,d), h(Lc, Ld) = Lh(c,d) = L 2cd 2(Lc +Ld) = L 1 2(Lc +L∞) = L∞, c0(Lc, L∞) = L2c+1, h(Lc, L∞) = L2c, 2(Lc +L0) = Lc/2, c0(LcL0) = L c 2+c , h(Lc, L0) = L0. 1 c+d , Set A1 = 1 2 Then (Lc +Ld), B1 = h(Lc, Ld), . . . , An = c1 = 1 2(c + d), d1 = h(c, d) = 1 (An−1 +Bn−1), Bn = h(An−1, Bn−1), n ≥ 2, 2 2cd c + d , . . . , cn = 1 2(cn−1 + dn−1), dn = h(cn−1, dn−1) = 1 − dn 1 + dn 1 − cn 1 + cn , n ∈ N. ezcn = , ezdn = An = Lcn, Bn = Ldn ∀n ∈ N. 2cn−1dn−1 cn−1 + dn−1 , n ≥ 2. SHORTING, PARALLEL ADDITION AND FORM SUMS 31 Due to Theorem 7.1 we have the equalities for the Cayley transforms C(An) = 1 C(Bn) = 1 2(Qµ + QM ) + 1 2(Qµ + QM ) + 1 From it follows that lim n→∞ cn = lim n→∞ s − lim n→∞ C(An) = s − lim n→∞ C(Bn) = Therefore 2ezcn(QM − Qµ), 2ezdn(QM − Qµ), n ∈ N. dn = √cd, 1 − √cd 1 + √cd (Qµ + QM ) + 1 2 QM − Qµ 2 . (cid:3) s − R − lim n→∞ Lcn = s − R − lim n→∞ Ldn = L√cd. Thus, ah(Lc, Ld) = L√cd for the case {c > 0, d ∈ [0, +∞]}. Observe that the equality takes place [34]. Besides if dim ran(cid:0)(SK + I)−1 − (SF + I)−1(cid:1) = dim (D[SK]/D[SF]) (7.8) that (7.9) 1 = Theorem 7.4. Let S be a nonnegative symmetric l.r. such that (7.8) holds. Suppose that 2(cid:0)(SK + I)−1 − (SF + I)−1(cid:1) , dim ran(cid:0)(SK + I)−1 − (SF + I)−1(cid:1) = 1, 2(cid:0)(SK + I)−1 + (SF + I)−1(cid:1) +ez then for an arbitrary nonnegative selfadjoint extension eS of S from (1.12) and (7.2) it follows (cid:16)eS + I(cid:17)−1 for someez ∈ [−1, 1]. eS1 and eS2 are two different non-extremal nonnegative selfadjoint extensions of S. Then the arithmetic-harmonic mean ah(eS1,eS2) is singleton. Moreover, if the numbers ez1,ez2 ∈ [−1, 1] correspond to eS1 and eS2 in the resolvent formula (7.9), then (7.10) (cid:16)ah(eS1,eS2) + I(cid:17)−1 2(cid:0)(SK + I)−1 + (SF + I)−1(cid:1) 2(cid:0)(SK + I)−1 − (SF + I)−1(cid:1) , + ew ew = p(1 +ez1)(1 +ez2) −p(1 −ez1)(1 −ez2) p(1 +ez1)(1 +ez2) +p(1 −ez1)(1 −ez2) In particular, ah(SF,eS) = SF, when eS 6= SK and ah(SK,eS) = SK, when eS 6= SF. Proof. Clearly,DeS1,eS2E 6=DeSF,eSKE . Define the iterations ), eS(n) 2 = h(eS(n−1) := eS2, eS(n) := eS1, eS(0) +eS(n−1) (eS(n−1) ), n ∈ N. 1 eS(0) , S(n−1) 2 (7.11) 1 = 1 2 1 = where 1 2 . 1 2 32 YU.M. ARLINSKII Then from (7.9) and Theorem 7.1 j 2 (cid:0)(SK + I)−1 − (SF + I)−1(cid:1) , j = 1, 2, n ∈ N, 2 } are non-decreasing and non-increasing sequences, respectively, and 1 1 1 ) = ) = 1 = ) , ) . 1 1 1 (7.12)  2 2 )(1 +ez(n−1) +ez(n−1) )(1 −ez(n−1) −ez(n−1) 2 2 (cid:16)eS(n) j + I(cid:17)−1 whereez(0) 2(1 +ez(n−1) 2 +ez(n−1) 2(1 −ez(n−1) 2 −ez(n−1) 2(cid:0)(SK + I)−1 + (SF + I)−1(cid:1)+ez(n) 1 =ez1,ez(0) 2 =ez2, , 1 +ez(n−1) 1 = h(1 +ez(n−1) 1 +ez(n) , 1 −ez(n−1) 2 = h(1 −ez(n−1) 1 −ez(n) 1 } and {ez(n) Besides, {ez(n) 2 ∀n ∈ N. Hence the sequences {ez(n) 1 } and {ez(n) ez(n) 1 ≤ez(n) n→∞ez(n) ewj = lim Note that becauseDeS1,eS2E 6=DeSF,eSKE we haveez1ez2 6= −1. If ez1 = −1, ez2 6= 1, then from (7.12) one has ez(n) ew1 = ew2 = −1. Ifez2 = 1,ez1 6= −1, then ew1 = ew2 = 1. Supposeez1 6= −1,ez2 6= −1. Then (7.12) yields ew1 = ew2 := ew. Let us find ew. Set 1 −ez(n) , bn := C(ez(n) 1 +ez(n) 1 Then, using (5.2) and (5.3), we obtain an := C(ez(n) 1 −ez(n) 1 +ez(n) 2 } converge. Set , n ∈ N0. , j = 1, 2. 2 ) = 1 ) = 2 2 j 1 2 n 2an−1bn−1 an−1 + bn−1 1 = −1 for all n ∈ N and therefore an = Hence, 1 2 (an−1 + bn−1), bn = h(an−1, bn−1) = Thus, we arrive at (7.10) -- (7.11). lim n→∞ an = lim n→∞ , n ∈ N0. (cid:3) bn =pa0b0 =s 1 −ez1 1 +ez1 References . 1 −ez2 1 +ez2 [1] W.N. Anderson, Shorted operators, SIAM J. Appl. Math. 20 (1971), 520 -- 525. [2] W.N. Anderson and R.J. Duffin, Series and parallel addition of matrices, J. Math. Anal. Appl. 26 (1969), 576 -- 594. [3] W.N. Anderson and G.E. Trapp, Shorted operators, II, SIAM J. Appl. Math. 28 (1975), 60 -- 71. [4] W.N. Anderson, T.D. Morlye, and G.E. Trapp, Characterizations of parallel subtraction, Proc. Nat. Acad. Sci. USA, 76 (1975), No.8, 3599 -- 3601. [5] T. Ando, Lebesgue type decomposition of positive operators. Acta Sci.Math. (Szeged) 38 (1976), 253 -- 260. [6] T. Ando, Topics on operator inequalities, Lecture Notes, Hokkaido University, Sapporo, Japan, 1978. [7] T. Ando and K. Nishio, Positive self-adjoint extensions of positive symmetric Operators. Toh´oku Math. J. 22 (1970), 65 -- 75. [8] J. Antezana, G. Corach, D. Stojanoff, Bilateral shorted operators and parallel sums, Linear Algebra and Application 414 (2006), 570 -- 588. [9] R. Arens, Operational calculus of linear relations, Pacific J. Math. 11 (1961), 9 -- 23. SHORTING, PARALLEL ADDITION AND FORM SUMS 33 [10] Yu.M. Arlinskiı, On the theory of operator means. Ukr. Mat. Zh. 42 (1990), No.6, 723 -- 730 (in Russian). English translation in Ukr. Math. Journ. 42 (1990), No.6, 639 -- 645. [11] Yu.M. Arlinskiı, Extremal extensions of sectorial linear relations. Matematychnii Studii 7 (1997), No.1, 81 -- 96. [12] Yu. Arlinskiı, S. Belyi, and E. Tsekanovskiı, Conservative realizations of Herglotz-Nevanlinna func- tions, Operator Theory: Advances and Applications 217. Basel: Birkhauser, xviii, 528 p.(2011). [13] Yu. Arlinskiı and S. Hassi, Holomorphic operator valued functions generated by passive selfadjoint systems, in Interpolation and Realization Theory with Applications to Control Theory, Operator Theory: Advances and Appl. 272 (2019), 35 -- 76. [14] Yu.M. Arlinskiı, S. Hassi, Z. Sebesty´en, and H.S.V. de Snoo, On the class of extremal extensions of a nonnegative operator, Oper. Theory: Adv. Appl. 127, 41 -- 81 (2001). [15] Gr. Arsene and A. Gheondea, Completing matrix contractions, J. Operator Theory 7 (1982), 179 -- 189. [16] T.Ya. Azizov, A. Dijksma, and G. Wanjala, Compressions of maximal dissipative and self-adjoint linear relations and of dilations. Linear Algebra Appl. 439 (2013), 771 -- 792. [17] Ju. Behrndt, S. Hassi, H. de Snoo, and R. Wietsma, Monotone convergence theorems for semi-bounded operators and forms with applications, Proc. Royal Soc. of Edinburgh 140 A (2010), 927 -- 951. [18] E.A. Coddington and H.S.V. de Snoo, Positive selfadjoint extensions of positive symmetric subspaces, Math. Z. 159 (1978), 203 -- 214. [19] M. Contino, A. Maestripieri, and S. Marcantognini, Schur complements of selfadjoint Kreın space operators, Linear Algebra and its Applications, 581 (2019), 214 -- 246. [20] Ch. Davis, W.M. Kahan, and H.F. Weinberger, Norm preserving dilations and their applications to optimal error bounds, SIAM J. Numer. Anal. 19 (1982), 445 -- 469. [21] V.A. Derkach and M.M. Malamud, Generalized resolvents and the boundary value problems for Hermitian operators with gaps, J. Funct. Anal. 95 (1991), No. 1, 1 -- 95. [22] V. Derkach, S. Hassi, M.M. Malamud, and H.S.V. de Snoo, Boundary relations and their Weyl families, Trans. Amer. Math. Soc. 358 (2006), No. 12, 5351 -- 5400 [23] R.G. Douglas, On majorization, factorization and range inclusion of operators in Hilbert space, Proc. Amer. Math. Soc. 17 (1966), 413 -- 416. [24] W.G. Faris, Selfadjoint operators, Lecture Notes in Mathematics, 433 (1975). [25] B. Farkas and M. Matolsci, Commutation properties of the form sum of positive, symmetric operators, Acta Sci.Math. (Szeged) 67 (2001), 777 -- 790. [26] P.A. Fillmore and J.P. Williams, On operator ranges. Advances in Math. 7 (1971), 254 -- 281. [27] J. Fridrich, M. Guntner, and L. Klotz, A generalized Schur complement for nonnegative operators on linear spaces, Banach J. Math. Anal 12 (2018), No.3, 617 -- 633. [28] S. Hassi, M. Malamud, and H.S.V. de Snoo, On Kreın's extension theory of nonnegative operators, Math. Nach. 274-275 (2004), 40 -- 73. [29] S. Hassi, A. Sandovici, H.S.V. de Snoo, and H. Winkler, Form sums of nonnegative selfadjoint oper- ators, Acta Math. Hungar. 111 (1-2) (2006), 81 -- 105. [30] S. Hassi, A. Sandovici, H.S.V. de Snoo, and H. Winkler, A general factorization approach to the extension theory of nonnegative operators and relations, J. Operator Theory 58 (2007), 351 -- 386. [31] S. Hassi, Z. Sebestyen, and H.S.V. de Snoo, Lebesgue type decompositions for nonnegative forms, J. Funct. Anal. 257 (2009), No.12, 3858 -- 3894. [32] S. Hassi, Z. Sebestyen, and H.S.V. de Snoo, Domain and range descriptions for adjoint relations, and parallel sums and differences of forms. Recent advances in operator theory in Hilbert and Krein spaces, 211 -- 227, Oper. Theory Adv. Appl. 198, Birkhauser Verlag, Basel, 2010. [33] T. Kato, Perturbation theory for linear operators, Springer-Verlag, Berlin, Heidelberg, 1995. [34] M.G. Kreın, Theory of selfadjoint extensions of semibounded operators and its applications. I, Mat. Sb. 20 (1947), No.3, 431 -- 498. (in Russian) [35] M.G. Kreın and I.E. Ovcharenko, On Q-functions and sc-extensions of nondensely defined Hermit- ian contractions, Sibirsk. Mat. Zh. 18 (1977), No.5, 1032 -- 1056 (in Russian). English translation in Siberian Math. Journ. 18 (1977), No.5, 728 -- 746. [36] F. Kubo and T. Ando, Means of positive linear operators. Math. Ann. 246 (1980), No.3, 205 -- 224. 34 YU.M. ARLINSKII [37] A. Maestripieri and F.M. Peria, Schur complements in Kreın spaces, Integral Equations Operator Theorey, 59 (2007), No.2, 207 -- 221. [38] S.K. Mitra, P. Bhimasankaram, S.B. Malik, Matrix partial orders, shorted operators and applications, Series in Algebra 10. Hackensack, NJ: World Scientific, 2010. [39] T.D. Morley, Shorts of block operators and infinite networks -- A note of the shorted operator: II, Circuits Systems Signal Process 9 (1990), No.2, 161 -- 170 [40] K. Nishio and T. Ando, Characterizations of operators derived from network connections, J. Math. Anal. Appl. 53 (1976), 539 -- 549. [41] M.A. Nudelman, A generalization of Stenger's lemma to maximal dissipative operators, Integral Equations Operator Theory 70 (2011), 301 -- 305. [42] E.L. Pekarev, Shorts of operators and some extremal problems, Acta Sci. Math. (Szeged) 56 (1992), 147 -- 163. [43] E.L. Pekarev, A note on characterizations of the shorted operators, Electronic Journal of Linear Algebra 27 (2014), 155 -- 161. [44] E.L. Pekarev and Yu.L. Shmulyan, Parallel addition and parallel subtraction of operators. Izv. AN SSSR 40 (1976), 366 -- 387 (1976) (in Russian). English translation in Math. USSR Izv. 10, 289 -- 337. [45] W. Pusz and S.L. Woronowics, Functional calculus for sesquilinear forms and purification map, Rep. Math. Phys., 8 (1975), No.2, 159 -- 170. [46] F.S. Rofe-Beketov, Numerical range of linear relation and maximal linear relations, Functions Theory, Functional Anal. and their Appl. 44 (1985), 103 -- 111 (in Russian). English translation in Journal of Math. Sci. 48 (1990), No.3, 329 -- 336. [47] Yu.L. Shmul'yan, Hellinger's operator integral, Mat. Sb. 49 (1959), No.4, 381 -- 430 (in Russian). [48] Yu.L. Shmul'yan and R.N. Yanovskaya, Blocks of a contractive operator matrix, Izv. Vuzov, Mat. 7 (1981), 72-75 (in Russian). English translation in Sov. Math. 25 (1981), No. 7, 82 -- 86. [49] B. Simon, A canonical decomposition for quadratic forms with applications to monotone convergence theorems, J. Funct. Analysis 28 (1978), 377 -- 385. [50] W. Stenger, On the projection of a selfadjoint operator, Bull. Amer.Math.Soc. 74 (1968), 369 -- 372. [51] Z. Tarcsay, Lebesgue type decomposition of positive operators. Positivity 17 (2013), No.3, 803 -- 817. [52] Z. Tarcsay, On the parallel sum of positive operators, forms and functionals, Acta Math. Hungar. 147 (2015), No.2, 408 -- 426. [53] F. Zhang, The Schur complement and its applications. Springer, 2005. Volodymyr Dahl East Ukrainian National University, pr. Central 59-A, Severodonetsk, 93400, Ukraine E-mail address: [email protected]
1103.4682	1	1103	2011-03-24T05:48:02	Generalized Fourier representation of the absolutely continuous part of a selfadjoint operator	[ "math.FA", "math-ph", "math-ph", "math.SP" ]	We formulate and prove the existence and uniqueness of the generalized Fourier transform associated with the absolutely continuous part of an arbitrary selfadjoint operator on a separable Hilbert space. To this end we develop a novel method to decompose an absolutely continuous operator into a variable fiber direct integral of selfadjoint operators.	math.FA	math	Generalized Fourier representation of the absolutely continuous part of a selfadjoint operator Take-Yuki Nagaoa aAdvanced Institute of Industrial Technology, 1-10-40 Higashi-Ohi, Shinagawa-City, Tokyo, JAPAN 140-0011 Abstract We formulate and prove the existence and uniqueness of the generalized Fourier transform associated with the absolutely continuous part of an ar- bitrary selfadjoint operator on a separable Hilbert space. To this end we develop a novel method to decompose an absolutely continuous operator into a variable fiber direct integral of selfadjoint operators. spectral decomposition, stationary method, eigenfunction Keywords: expansion, scattering theory 2000 MSC: 35P10, 35P25, 47A40 1. Introduction The most fundamental result of the spectral theory of selfadjoint opera- tors is the spectral theorem which states H = ZR λdEλ where H is a selfadjoint operator on some Hilbert space H and {Eλ}λ∈R is a spectral family associated with H. In this paper we establish a new spectral representation theorem that connects the spectral family and the generalized eigenfunctions. By generalized eigenfunction we mean a solution u to the formal eigenequa- tion (H − λ)u = 0, u ∈ X (1.1) Email address: [email protected] (Take-Yuki Nagao) Preprint submitted to Elsevier November 21, 2018 where λ ∈ R is a spectral parameter. The function space X must be chosen suitably, depending on the operator H of interest. Basically we choose X in such a way that the solutions to the problem (1.1) carry a complete set of information about the operator H at the energy λ. We can take X = H when λ /∈ σ(H) or λ ∈ σdisc(H). But we need to have X larger than H if λ ∈ σess(H) in order to capture the generalized eigenfunctions corresponding to the continuum. The goal of this paper is show the existence of a family of Hilbert spaces {X(λ)}λ∈R such that there is a selfadjoint operator Hλ on each X(λ) with the same spectrum as H and the collection of the solutions to the problem (Hλ − λ)u = 0, u ∈ D(Hλ) (1.2) for λ from a set of full Lebesgue measure determines the absolutely continuous part of H. We shall show that the spectral density E′ λ is a pullback of the orthogonal projection Q(λ) onto the null space of Hλ − λ in X(λ) in the sense that hE′ λf f i = kQ(λ)Y (λ)f k2, f ∈ D(Y (λ)), (1.3) where Y (λ) : H −→ X(λ) is a possibly non-closable injective linear map. This means that the generalized eigenvalue problem (1.2) determines the absolutely continuous part of H. We now briefly compare the idea of this paper with the known results. In the case of partial differential opeartors, it is common to use weighted L2 spaces [1] and Besov spaces [2, 3, 4] for the function space X in (1.1). Well known technique is the following. Given an absolutely continuous operator H on some function space, we look for a function space Y such that we have h(H − z)−1f gi ≤ Ckf kYkgkY, f, g ∈ Y with a uniform constant C > 0. The choice of Y depends on the problem in general. The point is that we have Imh(H − z)−1f f i ≤ Ckf k2 Y, f ∈ Y, Im z > 0 so that there is a bounded operator δ(H − λ) : Y −→ Y∗ such that Imh(H − λ − iε)−1f f i = πhδ(H − λ)f f i, f ∈ Y (1.4) lim ε↓0 2 for all λ ∈ R. The operator δ(H − λ) is identical to the spectral density E′ λ realized as an opeartor. One usually fix a function space D and argues that any solution u ∈ Y∗ of the problem hu(H − λ)f i = 0, f ∈ D can be written as u = δ(H − λ)ϕ for some ϕ ∈ Y and that we have hHf f i = ZR λhδ(H − λ)f f idλ. (1.5) (1.6) The identity (1.6) guarantees that the collection of solutions to the problem (1.5) for all λ ∈ R is complete in the sense that we can recover the original operator H from this collection. The main idea of this paper is to realize the left hand side of (1.3), which roughly corresponds to (1.4), as a quadratic from and choose the completion of its graph as Y. Although the quadratic form is nonnegative, it is not always closable. We shall establish a new representation theorem of nonnegative quadratic forms to overcome this difficulty. Another problem is how to give a rigorous meaning to the problem (1.5). To solve this, we propose the notion of continuation of a selfadjoint operator to some other function spaces. The author expects that the results of this paper is useful in scattering theory, since one can prove the existence of the generalized Fourier trans- form, which is a key tool for the theory, without depending on the limiting absorption principle nor on the Mourre inequality. In this paper we focus on the absolutely continuous part of H and the results on the case where H has a non-trivial singular continuous part will appear elsewhere. We close this section by collecting the notations used throughout the paper. A quadratic form q is linear in the first argument and anti-linear in the second. We adopt the same convention about inner products. We write q(f ) = q(f, f ). By ρ(T ), D(T ), R(T ) and N(T ) we mean the resolvent set, domain, range and null space of a linear opeartor T . If T is closable we write T for its closure unless otherwise noted. By Pac(H) we denote the spectral projection onto the absolutely continuous subspace of a selfadjoint operator H. The set of infinitely differentiable complex valued functions f on R such that f and all its derivatives tend to zero at infinity is denoted by C ∞ 0 (R). We sometimes write 0 for a zero space. The identity operators of various function spaces are denoted by 1. T ↾ D is the operator T with domain restricted to a subspace D of D(T ). 3 2. Continuation In this section we develop a method to continue a selfadjoint operator H initially defined on a Hilbert space H to another Banach space X that is close to H in a certain sense. We will use the results of this section to formulate the generalized eigenspace of a selfadjoint operator. Throughout this section we assume that X is a Banach space obtained by the completion of a dense subspace D of H with respect to some norm on D. The goal of this section is to formulate a condition that guarantees the existence of a closed operator HX : X −→ X, which we will call a continuation of H to X. The operator HX will be chosen so that it coincides with H on a common core of HX and H. We start with the observation that X is isometrically isomorphic to the completion of D by some norm if and only if there is a densely defined injective linear map Y : H −→ X with dense range such that D(Y ) = D. In this paper we call such Y a completion operator. Note that a completion operator need not be closable. Just like in the case of closed operators, we say that a subspace D of X is a core of Y if for any f ∈ D(Y ) there is a sequence fj ∈ D such that fj → f and Y fj → Y f . In this case the restriction Y ↾ D is a completion operator from H to X. Given a completion operator Y , we someteimes identify D(Y ) with R(Y ) through the correspondence D(Y ) ∋ f 7→ Y f ∈ R(Y ) in order to simplify notations. We also remark that since X is a completion, it cannot always be identified with a subspace of H. As we will see later, however, it is important to consider such a case when dealing with a non-closable form. Our strategy is that we first continue the resolvent (H − z)−1 to X and then define the continuation HX using the inverse of the continued resolvent. We introduce the following Definition 2.1. Let Y : H −→ X be a completion operator and z ∈ ρ(H). We say that H is compatible with Y at z if there is a subspace D ⊂ D(Y ) with the following properties: i. (H − z)−1D is a core of Y ; ii. there exists a constant Cz > 0 depending of z such that kY (H − z)−1f kX ≤ CzkY f kX for all f ∈ D; 4 iii. u ∈ X, fj ∈ D, Y fj → u in X, and Y (H − z)−1fj → 0 in X imply u = 0. Given a subset Ω ⊂ ρ(H), we say that H is compatible with Y in Ω if H is compatible with X through Y for all z ∈ Ω. Theorem 2.1. Let Ω be a non-empty subset of ρ(H). Suppose that H is compatible with a completion operator Y : H −→ X in Ω. Let D be as in Definition 2.1. Then there exists a densely defined closed operator HX : X −→ X such that Ω ⊂ ρ(HX) and that (HX − z)−1Y f = Y (H − z)−1f (2.1) for all f ∈ D and z ∈ Ω. The operator HX is independent of the choice of D and uniquely determined by H and Y . Proof. There exists a unique bounded operator B(z) on X such that B(z)Y f = Y (H − z)−1f for all f ∈ D. The first resolvent identity of (H − z)−1 yields B(z) − B(w) = (z − w)B(z)B(w) (2.2) (2.3) for all z, w ∈ Ω and it follows that R(B(z)) is constant for all z ∈ Ω. Remark that B(z) is injective and has a dense range. We define the operator HX : X −→ X with domain D(HX) = R(B(z)) by HXB(z)f = f + zB(z)f, f ∈ X. (2.4) Since B(z) is bounded and injective, HX is closed. Moreover, HX is inde- pendent of the choice of z ∈ Ω due to (2.3). The identity (2.4) implies that HX − z is bijective and satisfies (HX − z)−1 = B(z) for all z ∈ Ω. The unique- ness of HX follows from the uniqueness of B(z) in (2.2). This completes the proof. (cid:3) When we identify f ∈ D(Y ) and Y f ∈ R(Y ), (2.1) means that H and HX share D0 = (H − z)−1D = (HX − z)−1D as a common core and that we have HXu = Hu for all u ∈ D0. It is thus reasonable to regard HX as an extension of H to the Banach space X. In order to avoid confusion with the usual notion of extension of operators, let us refer to the operator HX as a continuation of H to X induced by Y or simply a continuation of H. In general, HX depends on the choice of Y and not uniquely determined by H and X. 5 Corollary 2.2. Suppose that X is a Hilbert space and that H is compatible with a completion operator Y : H −→ X in {i, −i}. Suppose further that HX is symmetric. Then HX is selfadjoint. 3. Pullback representation of a nonnegative form In this section we shall develop our main tool to deal with nonnegative quadratic forms that might not be closable. Our goal here is to formulate and prove a representation theorem of nonnegative quadratic forms. Recall that a nonnegative quadratic form q on a Hilbert space H is closable if and only if D(q) is complete under the norm k · kq where kf k2 q = kf k2 + q(f ). So it is reasonable to embed the domain D(q) into its completion by the above norm. We thus begin with the following Definition 3.1. Suppose that H, X and E are Hilbert spaces. A pullback triple τ = (J, F, Y ) is a triplet of operators J : X −→ H, F : X −→ E, Y : H −→ X (3.1) with the following properties: i. J, F are bounded with dense range; ii. Y is a completion operator; iii. for any u ∈ X we have kuk2 iv. JY ⊂ 1. X = kJuk2 + kF uk2 E; We often say that τ is a pullback triple on H in order to specify the function space H explicitly. We call D(Y ) the domain of τ and denote it by D(τ ). Note that the condition (iv) is equivalent to the commutativity of the diagram D(Y ) ι yyyyyyyy J Y X H where ι is the natural injection. As is done in depicting the operator Y of the above diagram, we put the domain of a linear operator in front of the tail of the arrow indicating the operator. We prefer this rule in order to simplify notations. 6 o o Lemma 3.1. Let (J, F, Y ) be a pullback triple. Then i. Y −1 is closable and J = Y −1; ii. Y is closable if and only if J is injective and in this case we have Y = J −1; iii. Y is bounded if and only if J is an isomorphism. E ≥ kf k2 for any f ∈ D(Y ). X = kf k2 + Proof. (i). Since JY f = f for any f ∈ D(Y ), we have kY f k2 kF Y f k2 It follows that Y −1 is closable and J = Y −1. (ii). If fj ∈ D(Y ) and Y fj → u in X, then we have fj → Ju in H. It follows that Y is closable if and only if J is injective. It is easy to see that D(Y ) = R(J) and that Y Ju = u for all u ∈ X. (iii). Note that kY f k ≤ Ckf k for all f ∈ D(Y ) is equivalent to kuk ≤ CkJuk for all u ∈ X. Since J is bounded and has a dense range, the statement follows. (cid:3) Lemma 3.2. Let Jj : Xj −→ Hj, Fj : Xj −→ Ej, Yj : Hj −→ Xj, j = 1, 2 be pullback triples. Then the following conditions are equivalent: i. J1 and J2 are unitarily equivalent; ii. F1 and F2 are unitarily equivalent; iii. there exist unitary isomorphisms WH, WX, and WE such that the following diagram commutes. H1 J1 F1 X1 E1 WH WX WE H2 J2 X2 F2 / E2 We introduce the following equilvalence relation. Definition 3.2. We say that two pullback triples (Jj, Fj, Yj), j = 1, 2 are equivalent if any of the conditions of Lemma 3.2 is satisfied. Lemma 3.3. Two pullback triples (Jj, Fj, Yj), j = 1, 2 on a Hilbert space H are equivalent if and only if there is a unitary isomorphism W such that (J1, F1, W Y2) is a pullback triple. 7 o o / / o o / Lemma 3.4. Let (J, F, Y ) be a pullback triple and Y0 ⊂ Y . Then (J, F, Y0) is a pullback triple if and only if D(Y0) is a core of Y . In this case (J, F, Y ) and (J, F, Y0) are equivalent. We say that a linear space D is a core of a pullback triple (J, F, Y ) if D is a core of Y . Theorem 3.5. Let q be a densely defined nonegative quadratic form on a Hilbert space H. Then there exists a pullback triple (J, F, Y ) with D(Y ) = D(q) such that the following diagram commutes. D(q) q(·) / R k·k2 E ι zzzzzzzz J Y X H !CCCCCCCC F / E The pullback triple is unique up to equivalence. Moreover, q is closable if and only if so is Y . In this case (J, F, Y ) is a pullback triple of q and we have D(q) = R(J), q(Ju) = kF uk2 E, u ∈ X. (3.2) Remark 3.1. The commutativity of the above diagram implies q(f ) = kF Y f k2 E, f ∈ D(Y ). This means that any nonnegative form q is a pullback of a bounded nonneg- ative form by a completion operator Y . Proof. It is easy to see that D(q) × D(q) ∋ (f, g) 7→ hf gi + q(f, g) ∈ C is a positive definite inner product on D(q). Let X be the completion of D(q) by this inner product and Y : H −→ X be the associated completion operator. Since q is nonnegative, we have q(f, g)2 ≤ q(f )q(g) ≤ kY f k2 XkY gk2 X for any f, g ∈ D(q). By the Riesz representation theorem there exists a unique bounded selfadjoint operator δ : X −→ X such that hδY f Y giX = q(f, g), f, g ∈ D(q). 8 ! / o o / O O It is easy to see that the mapping R(δ) × R(δ) ∋ (δu, δv) 7→ hδuviX ∈ C is well-defined and is a positive definite inner product on R(δ). Let E be the completion of R(δ) with respect to this inner product. We define the operator F by F : X −→ E, F u = δu, u ∈ X. By our choice of E, R(F ) is dense in E and we have kF uk2 E = hδuuiX for any u ∈ X. Note that Y −1 is closable and has a bounded extension by Lemma 3.1. We set J = Y −1. Clearly, JY f = f, q(f ) = kF Y f k2 E, f ∈ D(Y ). (3.3) Therefore (J, F, Y ) has the desired properties. The uniqueness follows from (3.3) and Lemma 3.2. The statement about closability follows from Lemma 3.1. The relation (3.2) is obvious. This completes the proof. (cid:3) We refer to the triple (J, F, Y ) as a pullback triple of q and we say that two nonnegative forms q1 and q2 are equivalent if the corresponding pullback triples are equivalent. There are cases where the domains of q1 and q2 have no inclusion relation but q1 and q2 are equivalent. Similarly any linear operator is a pullback of a bounded operator. We only state the result and the proof is left to the reader. Theorem 3.6. Let T : H −→ K be a linear operator between Hilbert spaces. Then there exists a pullback triple (J, F, Y ) with D(Y ) = D(T ) such that the following diagram commutes. D(T ) ι }{{{{{{{{{ J Y X H T #GGGGGGGG F / R(T ) The pullback triple is unique up to equivalence. Moreover, T is closable if and only if so is Y . In this case (J, F, Y ) is a pullback triple of T and we have D(T ) = R(J), T Ju = F u, u ∈ X. 9 } # o o / Fix a pullback triple (3.1). Then N(J) and N(F ) are orthogonal in X. Moreover, N(J) = {f ∈ X : f = Qf }, N(F ) = {f ∈ X : f = Qf }, (3.4) where Q and Q are bounded operators in X defined by Q = F ∗F, Q = 1 − Q = J ∗J. Lemma 3.7. Let (J, F, Y ) be a pullback triple. Then the following conditions are equivalent: i. F is a partial isometry; ii. X = N(J) ⊕ N(F ); iii. JF ∗ = 0; iv. J N(F ) is dense in H. A pullback triple (J, F, Y ) is said to be splitting if any of the conditions of Lemma 3.7 is satisfied. We close this section by summarizing the properties of a splitting pullback triple for later use. Proposition 3.8. Suppose that (J, F, Y ) is a splitting pullback triple. Then J and F are partial isometries with N(F ) = R(J ∗), R(J) = H, N(J) = R(F ∗), R(F ) = E. (3.5) (3.6) Remark 3.2. Proposition 3.8 implies that under the condition that the pull- back triple (J, F, Y ) is splitting, the form q associated with the triple is clos- able if and only if F and thus q vanishes identically. In this case we have E = 0. Note also that X is unitarily isomorphic to the splitting direct sum H ⊕ E. This means that one can measure by E how far is q from being closable. 10 4. Pullback representation of spectral densities Suppose that H is a selfadjoint operator on a separable Hilbert space H. Let {Eλ}λ be a spectral family of H. We shall derive a representation of the absolutely continuous part of H in terms of the spectral density E′ λ. The main diffculty is that the spectral density is never closable unless it is zero identically. The pullback triples we have developed in the previous section facilitate the understanding of the spectral density as a quadratic form. Since kEλf k2 is a non-decreasing function of λ for any f ∈ H, it is differentiable almost everywhere in the sense of Dini, and so is hEλf gi for any f, g ∈ H. It is easy to see that dhEλf gi dλ = dhEλPac(H)f gi dλ (4.1) almost everywhere. By Radon-Nikodym theorem we see that (4.1) is inte- grable and that hPac(H)f gi = ZR dhEλf gi dλ dλ (4.2) for any f, g ∈ H. We shall show that (4.1) is well-defined as a densely defined nonnegative form. Lemma 4.1. Let f, g ∈ H. Suppose that hEλf gi is differentiable at λ. Then hEλχ(H)f gi is differentiable at λ for any χ ∈ C ∞ 0 (R) and we have dhEλχ(H)f gi dλ = χ(λ) dhEλf gi dλ . (4.3) Proof. The statement follows from the fact that [χ(H)−χ(λ)]EH (I)/I → 0 strongly as I → 0 for any interval I containing λ as an internal point. (cid:3) Note that the spectral density has the following representation by the boundary value of the resolvent. Lemma 4.2. Let f ∈ H. Then lim ε↓0 ε π k(H − λ − iε)−1Pac(H)f k2 = dkEλf k2 dλ (4.4) for almost every λ with respect to Lebesgue measure. 11 Suppose that D0 is a subset of a Hilbert space H and A is a family of bounded linear operators on H. Let D be a linear space of all finite linear combinations of the vectors of the form Af, f ∈ D0, where A is the product of zero or finitely many elements of A. It is easy to see that D0 ⊂ D and AD ⊂ D for all A ∈ A. We call D the span of D0 by A. We are interested in the case where D0 is an orthonormal basis of H and A = {(H − z)−1}z∈ρ(H). In this case we say that D is the span of an orthonormal basis by the resolvent of H. Lemma 4.3. Let D be the span of an orthonormal basis by the resolvent of H. Then there exists a subset Λ ⊂ R of full Lebesgue measure such that hEλf gi is differentiable at any λ ∈ Λ whenever f, g ∈ D. Remark 4.1. We remark that Λ does not contain an eigenvalue of H. Proof. Let {ϕj} be the orthonormal basis generating D, which is countable by assumption. There is a subset Λ ⊂ R of full Lebesgue measure such that hEλϕjϕki is differentiable at any λ ∈ Λ for all j, k. By Lemma 4.1 hEλAϕjBϕki is differentiable at any λ ∈ Λ for all j, k, where A and B are zero or finitely many products of operators in {(H − z)−1}z∈ρ(H). The statement follows by linearity. (cid:3) Lemma 4.4. Suppose that each of D1 and D2 is the span of an orthonormal basis by the resolvent of H. Then there exists a set Λ0 of full Lebesgue measure with the following properties: i. hEλf gi is differentiable at λ whenever λ ∈ Λ0 and f, g ∈ D1 + D2; ii. for any f ∈ D2 there exists a sequence fj ∈ D1 such that kf − fjk → 0, dkEλ(f − fj)k2 dλ → 0 (4.5) as j → ∞ for any λ ∈ Λ0; iii. the same statement as (ii) with D1 and D2 interchanged. Proof. (i) is obvious from Lemma 4.3. We prove (ii) and (iii). By (4.2) one can see that (4.5) holds almost everywhere for fixed f ∈ D2. Let A be a finite product of operators from {(H −z)−1}z∈ρ(H). By Lemma 4.1 we see that (4.5) remains valid when we replace f and fj by Af and Afj respectively. It follows that one can make Λ0 independent of the choice of f ∈ D2. We can make Λ0 smaller so that (iii) holds as well. This completes the proof. (cid:3) 12 Definition 4.1. Suppose that H is a selfadjoint operator on a Hilbert space H and {τ (λ)}λ∈R is a family of pullback triples, where τ (λ) = (J(λ), F (λ), Y (λ)) and the function spaces are as given by the commutative diagram below. D(Y (λ)) H X(λ) / E(λ) ι {vvvvvvvvvv J(λ) Y (λ) F (λ)Y (λ) $JJJJJJJJJ F (λ) We call {τ (λ)}λ∈R a pullback representation of the absolutely continuous part of H if there exist a subset Λ ⊂ R of full Lebesgue measure and a span D of an orthonormal basis by the resolvent of H such that i. D is a core of τ (λ) for all λ; ii. for any f ∈ D and λ ∈ Λ we have dkEλf k2 dλ = kF (λ)Y (λ)f k2 E(λ); iii. for λ ∈ R \ Λ we have X(λ) = H, Y (λ) = J(λ) = 1, and F (λ) = 0. We say that two representations {τj(λ)}λ∈R, j = 1, 2 are equivalent if τ1(λ) and τ2(λ) are equivalent almost everywhere. We refer to D as a core of the representation. Theorem 4.5. Any selfadjoint operator on a separable Hilbert space admits a pullback representation of the absolutely continuous part. The representa- tion is unique up to equivalence. Proof. Let D and Λ be as given by Lemma 4.3. We see that the function qλ : H × H −→ C defined for each λ ∈ Λ by qλ(f, g) = dhEλf gi dλ , f, g ∈ D(q) = D (4.6) is well-defined as a densely defined nonnegative quadratic form on H. We set qλ = 0, D(qλ) = D for λ ∈ R \ Λ. For each λ ∈ R let J(λ) : X(λ) −→ H, F (λ) : X(λ) −→ E(λ), Y (λ) : H −→ X(λ) be a pullback triple of qλ. By Theorem 3.5 we have qλ(f, g) = hF (λ)Y (λ)f F (λ)Y (λ)giE(λ), f, g ∈ X(λ). 13 { $ o o / We have thus proven that any selfadjoint operator admits a pullback repre- sentation. We now show the uniqueness. Suppose that τj = {τj(λ)}, j = 1, 2 are pullback representations of H. Write τj(λ) = (Jj(λ), Fj(λ), Yj(λ)). In view of Lemma 3.4 we may assume that D(Yj(λ)) = Dj, j = 1, 2, where each Dj is a span of an orthonormal basis by the resolvent of H. Let Λ0 be as in Lemma 4.4. Then for each λ ∈ Λ0 there is a well-defined linear operator Y3(λ) : H −→ X1(λ) with D(Y3(λ)) = D2 such that J1(λ)Y3(λ)f = f, dkEλf k2 dλ = kF1(λ)Y3(λ)f k2 E1(λ), (4.7) dhEλf gi dλ = hF1(λ)Y3(λ)f F1(λ)Y1(λ)gi, g ∈ D1 (4.8) for all f ∈ D(Y3(λ)). In view of Lemma 3.3 it suffices to show that τ3(λ) = (J1(λ), F1(λ), Y3(λ)) is a pullback triple for all λ ∈ Λ0. Clearly Y3(λ) is densely defined and injective. We now show that Y3(λ) has a dense range. Let ϕ ∈ D1. Then there exists a sequence ϕj ∈ D2 so that kϕ − ϕjk → 0, dkEλ(ϕ − ϕj)k2 dλ → 0 (4.9) for all λ ∈ Λ0. By (4.7) we see that Y3(λ)ϕj → u in X1(λ) for some u ∈ X1(λ) and (4.8) yields hF1(λ)[Y3(λ)ϕj − Y1(λ)ϕ]F1(λ)Y1(λ)gi → 0, g ∈ D1. Therefore J1(λ)u = ϕ and F1(λ)u = F1(λ)Y1(λ)ϕ. This means u = Y1(λ)ϕ and hence Y3(λ) has a dense range. We have thus shown that Y3(λ) is a completion operator. This and (4.7) means that τ3(λ) is a pullback triple. (cid:3) 5. Generalized eigenfunction expansion The purpose of this section is to show that the absolutely continuous part of an arbitrary selfadjoint operator H admits a generalized eigenfunction ex- pansion. We shall first show that a pullback representation τ = {τ (λ)}λ∈R of H induces a family of continuations {Hλ}λ∈R of H. Then we show that the spectral structue of Hλ and H are identical except at energy λ and the spectral density E′ λ of H at energy λ is identical to a pullback of the orthog- onal projection onto the nullspace of Hλ − λ for almost every λ. This means 14 that the structure of the absolutely continuous part of H is encoded into the family {N(Hλ − λ)}λ∈R of nullspaces and we can recover the absolutely continous part of the original operator from this family. Throughout this section we fix a pullback representation τ of the abso- lutely continuous part of H. We use the same operators and function spaces as in the Definition 4.1. Lemma 5.1. Let λ ∈ R. Then i. f ∈ D and Im z 6= 0 imply F (λ)Y (λ)(H − z)−1f = (λ − z)−1F (λ)Y (λ)f ; (5.1) ii. (H − z)−1D is a core of τ (λ) and we have kY (λ)(H − z)−1f kX(λ) ≤ Im z−1kY (λ)f kX(λ) (5.2) for f ∈ D and Im z 6= 0; iii. H is compatible with Y (λ) in ρ(H). Proof. (i) is immediate from Lemma 4.1. (ii). From (i) we see that kY (λ)(H − z)−1f k2 X(λ) = k(H − z)−1f k2 + (λ − z)−12kF (λ)Y (λ)f k2 E(λ). This implies (5.2). Suppose that u ∈ X(λ) and that huY (λ)(H −z)−1giX(λ) = 0 whenever g ∈ D. It follows from (i) that hJ(λ)uJ(λ)Y (λ)(H − z)−1gi + (λ − z)−1hF (λ)uF (λ)Y (λ)gi = 0 for g ∈ D. Let fj be a sequence in D such that Y (λ)fj → u in X(λ). Note that fj → J(λ)u in H. The inequality (5.2) implies hfj(H − z)−1fji + (λ − z)−1kF (λ)Y (λ)fjk2 → 0 and by taking the imaginary parts we see that kfjk → 0, kF (λ)Y (λ)fjkE(λ) → 0. Hence u = 0. By the Hahn-Banach theorem we conclude that Y (λ)(H − z)−1D is dense in X(λ). (iii). If λ ∈ R ∩ ρ(H), then F (λ) = 0 and thus Y (λ) is isometric. This combined with (i) and (ii) imply (i) through (iii) of Definition 2.1. To verify the condition (iv) suppose Y (λ)fj → u in X(λ) with fj ∈ D and Y (λ)(H − z)−1fj → 0 in X(λ). Then (i) yields k(H − z)−1fjk → 0, kF (λ)Y (λ)fjkE(λ) → 0 so we must have u = 0. This completes the proof. (cid:3) 15 Lemma 5.2. Let λ ∈ R. Then H has a selfadjoint continuation Hλ to X(λ) induced by Y (λ) and it satisfies F (λ)χ(Hλ) = χ(λ)F (λ), J(λ)χ(Hλ) = χ(H)J(λ), (5.3) for χ ∈ C ∞ 0 (R). Proof. Lemma 5.1 implies that H has a continuation Hλ to X(λ) and that In view of Lemma 5.1 the first identity of (5.3) holds for ρ(H) ⊂ ρ(Hλ). χ(H) = (H − z)−1 with z ∈ ρ(H) and thus for all χ ∈ C ∞ 0 (R). Take fj ∈ D such that Y (λ)fj → u in X(λ). Then we have J(λ)(Hλ − z)−1Y (λ)fj = (H − z)−1fj → (H − z)−1Ju for any z ∈ ρ(H). This implies the second identity of (5.3). For any z ∈ ρ(H) we have Imh(Hλ − z)−1uuiX(λ) = Im zk(Hλ − z)−1uk2 X(λ) for all u ∈ Y (λ)D and thus for all u ∈ X(λ). Therefore Hλ is symmetric and hence selfadjoint by Corollary 2.2. (cid:3) By Lemma 5.2 J(λ)N(F (λ)) is dense in H for all λ ∈ R, so we have the following Corollary 5.3. The pullback triple τ (λ) is splitting for all λ ∈ R. Theorem 5.4. For any λ ∈ R the spectrum of H and Hλ are identical. We have J(λ)Hλ = HJ(λ), F (λ)Hλ ⊂ λF (λ). J(λ) and F (λ) are partial isometries. For λ ∈ Λ we have R(Hλ − λ) = N(F (λ)) = R(J(λ)∗), R(J(λ)) = H, N(Hλ − λ) = N(J(λ)) = R(F (λ)∗), R(F (λ)) = E(λ). Proof. Lemma 5.2 implies (5.4). We now prove N(Hλ − λ) = N(J(λ)). (5.4) (5.5) (5.6) (5.7) As Hλ is selfadjoint and F (λ)(Hλ−λ)v = 0 for any v ∈ D(Hλ), u ∈ N(Hλ−λ) if and only if hJ(λ)uJ(λ)(Hλ − λ)vi = 0 for all v ∈ D(Hλ). This condition 16 is equivalent to J(λ)u = 0 because J(λ)(Hλ − λ) has a dense range. Hence we obtain (5.7). By Proposition 3.8, (5.5) and (5.6) follow. We now prove σ(H) = σ(Hλ). This is obvious if λ ∈ R \ Λ, so we may assume λ ∈ Λ. Then J(λ) is a partial isometry with initial space N(Q(λ)) and final space H, where Q(λ) = F (λ)∗F (λ). Note that Q(λ) is the orthogonal projection onto the eigenspace of Hλ with eigenvalue λ. This and (5.4) imply that EH (I) is unitarily equivalent to EHλ(I \ {λ}) for any interval I. If λ ∈ ρ(H), then EHλ({λ}) = Q(λ) = 0 and thus λ ∈ ρ(Hλ). Conversely, λ ∈ ρ(Hλ) implies Q(λ) = 0 and so λ ∈ ρ(H). This completes the proof. (cid:3) Theorem 5.5. There exists a partial isometry F : H −→ Z ⊕ R E(λ)dλ (5.8) such that We have (F f )(λ) = F (λ)Y (λ)f, f ∈ D, λ ∈ R. F ∗F = Pac(H), F F ∗ = 1. (5.9) Remark 5.1. We remark that dhEλf gi dλ = hQ(λ)Y (λ)f Y (λ)giX(λ), f, g ∈ D almost everywhere, where Q(λ) = F (λ)∗F (λ). It follows that the spectral density E′ λ is a pullback of the eigenprojection Q(λ) of Hλ by the completion operator Y (λ) and Theorem 5.5 implies hHacf gi = ZR λhQ(λ)Y (λ)f Y (λ)giX(λ)dλ. The absolutely continuous part of a selfadjoint operator thus admits a gener- alized eigenfunction expansion. It should be noted that Q(λ) = 0 if and only if the spectral density is closable at λ and that the size of the generalized eigenspace N(Hλ − λ) = R(Q(λ)) = N(J(λ)) is determined by how far is the spectral density from being closable at energy λ. 17 Proof. It is easy to see that kPac(H)f k2 = ZR kF (λ)Y (λ)f k2 E(λ)dλ f ∈ D. It follows that the partial isometry F exists and we have F ∗F = Pac(H). We now show that F is surjective. By the Hahn-Banach theorem it suffices to prove that ZR hg(λ)F (λ)Y (λ)f iE(λ)dλ = 0, f ∈ D, ZR kg(λ)k2 E(λ)dλ < ∞ implies g(λ) = 0 for almost every λ. The identity (5.3) yields ZR ϕ(λ)hg(λ)F (λ)Y (λ)f iE(λ)dλ = 0 for any ϕ ∈ C ∞ 0 (R) and f ∈ D. Since D is the span of an orthonormal basis by the resolvents of H, there exists a set Λ0 ⊂ R of full Lebesgue measure which is independent of f so that hg(λ)F (λ)Y (λ)f iE(λ) = 0 for λ ∈ Λ0 and f ∈ D. Since F (λ) is bounded and surjective, we must have g(λ) = 0 for almost every λ. This completes the proof. (cid:3) 6. Acknowledgements The author is grateful to Professor Shu Nakamura for valuable discussions and constructive remarks. References [1] S. Agmon, Spectral properties of Schrodinger operators and scattering theory, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 2 (1975) 151 -- 218. [2] S. Agmon, L. Hormander, Asymptotic properties of solutions of differ- ential equations with simple characteristics, Journal d'Analyse Mathma- tique 30 (1976) 1 -- 38. [3] L. Hormander, The Analysis of Linear Partial Differential Operators II, Springer-Verlag, 1990. [4] H. Isozaki, Asymptotic properties of solutions to 3-particle schrodinger equations, Communications in Mathematical Physics 222 (2001) 371 -- 413. 18
1205.0318	1	1205	2012-05-02T04:27:42	Restricted normal cones and the method of alternating projections	[ "math.FA" ]	The method of alternating projections (MAP) is a common method for solving feasibility problems. While employed traditionally to subspaces or to convex sets, little was known about the behavior of the MAP in the nonconvex case until 2009, when Lewis, Luke, and Malick derived local linear convergence results provided that a condition involving normal cones holds and at least one of the sets is superregular (a property less restrictive than convexity). However, their results failed to capture very simple classical convex instances such as two lines in three-dimensional space. In this paper, we extend and develop the Lewis-Luke-Malick framework so that not only any two linear subspaces but also any two closed convex sets whose relative interiors meet are covered. We also allow for sets that are more structured such as unions of convex sets. The key tool required is the restricted normal cone, which is a generalization of the classical Mordukhovich normal cone. We thoroughly study restricted normal cones from the viewpoint of constraint qualifications and regularity. Numerous examples are provided to illustrate the theory.	math.FA	math	Restricted normal cones and the method of alternating projections Heinz H. Bauschke∗, D. Russell Luke†, Hung M. Phan‡, and Xianfu Wang§ May 2, 2012 Abstract The method of alternating projections (MAP) is a common method for solving feasibility prob- lems. While employed traditionally to subspaces or to convex sets, little was known about the behavior of the MAP in the nonconvex case until 2009, when Lewis, Luke, and Malick de- rived local linear convergence results provided that a condition involving normal cones holds and at least one of the sets is superregular (a property less restrictive than convexity). How- ever, their results failed to capture very simple classical convex instances such as two lines in three-dimensional space. In this paper, we extend and develop the Lewis-Luke-Malick framework so that not only any two linear subspaces but also any two closed convex sets whose relative interiors meet are covered. We also allow for sets that are more structured such as unions of convex sets. The key tool required is the restricted normal cone, which is a generalization of the classical Mordukhovich normal cone. We thoroughly study restricted normal cones from the viewpoint of constraint qualifications and regularity. Numerous examples are provided to illustrate the theory. 2010 Mathematics Subject Classification: Primary 49J52, 49M20; Secondary 47H09, 65K05, 65K10, 90C26. Keywords: Constraint qualification, convex set, Friedrichs angle, linear convergence, method of alternating projections, normal cone, projection operator, restricted normal cone, superregularity. 1 Introduction Throughout this paper, we assume that ∗Mathematics, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. †Institut f ur Numerische und Angewandte Mathematik, Universitat G ottingen, Lotzestr. 16 -- 18, 37083 G ottingen, Germany. E-mail: [email protected]. ‡Mathematics, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. §Mathematics, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. 1 (1) X is a Euclidean space (i.e., finite-dimensional real Hilbert space) with inner product h·, ·i, induced norm k · k, and in- duced metric d. Let A and B be nonempty closed subsets of X. We assume first that A and B are additionally convex and that A ∩ B 6= ∅. In this case, the projection operators PA and PB (a.k.a. projectors or nearest point mappings) corresponding to A and B, respectively, are single-valued with full do- main. In order to find a point in the intersection A and B, it is very natural to simply alternate the operator PA and PB resulting in the famous method of alternating projections (MAP). Thus, given a starting point b−1 ∈ X, sequences (an)n∈N and (bn)n∈N are generated as follows: (2) bn := PBan. (∀n ∈ N) an := PAbn−1, In the present consistent convex setting, both sequences have a common limit in A ∩ B. Not surprisingly, because of its elegance and usefulness, the MAP has attracted many famous math- ematicians, including John von Neumann and Norbert Wiener and it has been independently rediscovered repeatedly. It is out of scope of this article to review the history of the MAP, its many extensions, and its rich and convergence theory; the interested reader is referred to, e.g., [4], [7], [11], and the references therein. Since X is finite-dimensional and A and B are closed, the convexity of A and B is actually not needed in order to guarantee existence of nearest points. This gives rise to set-valued projection op- erators which for convenience we also denote by PA and PB. Dropping the convexity assumption, the MAP now generates sequences via (3) (∀n ∈ N) an ∈ PAbn−1, bn ∈ PBan. This iteration is much less understood than its much older convex cousin. For instance, global convergence to a point in A ∩ B cannot be guaranteed anymore [9]. Nonetheless, the MAP is widely applied to applications in engineering and the physical sciences for finding a point in A ∩ B (see, e.g., [25]). Lewis, Luke, and Malick achieved a break-through result in 2009, when there are no normal vectors that are opposite and at least one of the sets is superregular (a property less restrictive than convexity). Their proof techniques were quite different from the well known convex approaches; in fact, the Mordukhovich normal cone was a central tool in their analysis. However, their results were not strong enough to handle well known convex and linear scenarios. For instance, the linear convergence of the MAP for two lines in R3 cannot be obtained in their framework. The goal of this paper is to extend the results by Lewis, Luke and Malick to make them applicable in more general settings. We unify their theory with classical convex convergence results. Our principal tool is a new normal cone which we term the restricted normal cone. A careful study of restricted normal cones and their applications is carried out. We also allow for constraint sets that are unions of superregular 2 (or even convex) sets. We shall recover the known optimal convergence rate for the MAP when studying two linear subspaces. In a parallel paper [5] we apply the tools developed here to the important problem of sparsity optimization with affine constraints. The remainder of the paper is organized as follows. In Section 2, we collect various auxiliary results that are useful later and to make the later analysis less cluttered. The restricted normal cones are introduced in Section 3. Section 4 focuses on normal cones that are restricted by affine subspaces; the results achieved are critical in the inclusion of convex settings to the linear conver- gence framework. Further examples and results are provided in Section 5 and Section 6, where we illustrate that the restricted normal cone cannot be obtained by intersections with various natural conical supersets. Section 7 and Section 8 are devoted to constraint qualifications which describe how well the sets A and B relate to each other. In Section 9, we discuss regularity and superreg- ularity, notions that extend the idea of convexity, for sets and collections of sets. We are then in a position to provide in Section 10 our main results dealing with the local linear convergence of the MAP. Notation The notation employed in this article is quite standard and follows largely [6], [22], [23], and [24]; these books also provide exhaustive information on variational analysis. The real num- of X. Then the closure of S is S, the interior of S is int(S), the boundary of S is bdry(S), and the smallest affine and linear subspaces containing S are aff S and span S, respectively. The lin- ear subspace parallel to aff S is par S := (aff S) − S = (aff S) − s, for every s ∈ S. The rela- tive interior of S, ri(S), is the interior of S relative to aff(S). The negative polar cone of S is bers are R, the integers are Z, and N := (cid:8)z ∈ Z (cid:12)(cid:12) z ≥ 0(cid:9). Further, R+ := (cid:8)x ∈ R (cid:12)(cid:12) x ≥ 0(cid:9), R++ := (cid:8)x ∈ R (cid:12)(cid:12) x > 0(cid:9) and R− and R−− are defined analogously. Let R and S be subsets S⊖ =(cid:8)u ∈ X(cid:12)(cid:12) sup hu, Si ≤ 0(cid:9). We also set S⊕ := −S⊖ and S⊥ := S⊕ ∩ S⊖. We also write R ⊕ S for R + S := (cid:8)r + s(cid:12)(cid:12) (r, s) ∈ R × S(cid:9) provided that R ⊥ S, i.e., (∀(r, s) ∈ R × S) hr, si = 0. We cone if (∀λ ∈ R+) λK :=(cid:8)λk(cid:12)(cid:12) k ∈ K(cid:9) ⊆ K. The smallest cone containing S is denoted cone(S); thus, cone(S) := R+ · S :=(cid:8)ρs(cid:12)(cid:12) ρ ∈ R+, s ∈ S(cid:9) if S 6= ∅ and cone(∅) := {0}. The smallest con- and ρ ∈ R++, then ball(z; ρ) :=(cid:8)x ∈ X(cid:12)(cid:12) d(z, x) ≤ ρ(cid:9) is the closed ball centered at z with radius ρ while sphere(z; ρ) :=(cid:8)x ∈ X(cid:12)(cid:12) d(z, x) = ρ(cid:9) is the (closed) sphere centered at z with radius ρ. If u and v are in X, then [u, v] :=(cid:8)(1 − λ)u + λv(cid:12)(cid:12) λ ∈ [0, 1](cid:9) is the line segment connecting u and v. write F : X ⇒ X, if F is a mapping from X to its power set, i.e., gr F, the graph of F, lies in X × X. Abusing notation slightly, we will write F(x) = y if F(x) = {y}. A nonempty subset K of X is a vex and closed and convex subset containing S are conv(S) and conv (S), respectively. If z ∈ X 2 Auxiliary results In this section, we fix some basic notation used throughout this article. We also collect several auxiliary results that will be useful in the sequel. 3 Projections Definition 2.1 (distance and projection) Let A be a nonempty subset of X. Then (4) dA : X → R : x 7→ inf a∈A d(x, a) is the distance function of the set A and is the corresponding projection. PA : X ⇒ X : x 7→(cid:8)a ∈ A(cid:12)(cid:12) dA(x) = d(x, a)(cid:9) (5) (6) Proposition 2.2 (existence) Let A be a nonempty closed subset of X. Then (∀x ∈ X) PA(x) 6= ∅. Proof. Let z ∈ X. The function f : X → R : x 7→ kx − zk2 is continuous and limkxk→+∞ f (x) = +∞. Let (xn)n∈N be a sequence in A such that f (xn) → inf f (A). Then (xn)n∈N is bounded. Since A is closed and f is continuous, every cluster point of (xn)n∈N is a minimizer of f over the set A, i.e., an element in PAz. (cid:4) Example 2.3 (sphere) Let z ∈ X and ρ ∈ R++. Set S := sphere(z; ρ). Then (∀x ∈ X) PS(x) =(z + ρ x−z kx−zk , if x 6= z; otherwise. S, Proof. Let x ∈ X. The formula is clear when x = z, so we assume x 6= z. Set (7) c := z + ρ x − z kx − zk ∈ S, and let s = z + ρb ∈ S r {c}, i.e., kbk = 1 and b 6= (x − z)/kx − zk. Hence, using that kuk − kvk < ku − vk ⇔ hu, vi < kukkvk and because of Cauchy-Schwarz, we obtain (8a) kx − ck =(cid:12)(cid:12)kx − zk − ρ(cid:12)(cid:12) =(cid:12)(cid:12)kx − zk − kρbk(cid:12)(cid:12) =(cid:12)(cid:12)kx − zk − ks − zk(cid:12)(cid:12) < kx − sk. (8b) We have thus established (6). (cid:4) In view of Proposition 2.2, the next result is in particular applicable to the union of finitely many nonempty closed subsets of X. Lemma 2.4 (union) Let (Ai)i∈I be a collection of nonempty subsets of X, set A := Si∈I Ai, let x ∈ X, and suppose that a ∈ PA(x). Then there exists i ∈ I such that a ∈ PAi(x). Proof. Indeed, since a ∈ A, there exists i ∈ I such that a ∈ Ai. Then d(x, a) = dA(x) ≤ dAi (x) ≤ d(x, a). Hence d(x, a) = dAi (x), as claimed. (cid:4) The following result is well known. 4 Fact 2.5 (projection onto closed convex set) Let C be a nonempty closed convex subset of X, and let x, y and p be in X. Then the following hold: (i) PC(x) is a singleton. (ii) PC(x) = p if and only if p ∈ C and suphC − p, x − pi ≤ 0. (iii) kPC(x) − PC(y)k2 + k(Id −PC)(x) − (Id −PC)(y)k2 ≤ kx − yk2. (iv) kPC(x) − PC(y)k ≤ kx − yk. Proof. (i)&(ii): [4, Theorem 3.14]. (iii): [4, Proposition 4.8]. (iv): Clear from (iii). (cid:4) Miscellany Lemma 2.6 Let A and B be subsets of X, and let K be a cone in X. Then the following hold: (i) cone(A ∩ B) ⊆ cone A ∩ cone B. (ii) cone(K ∩ B) = K ∩ cone B. Proof. (i): Clear. (ii): By (i), cone(K ∩ B) ⊆ (cone K) ∩ (cone B) = K ∩ cone B. Now assume that x ∈ (K ∩ cone B) r {0}. Then there exists β > 0 such that x/β ∈ B. Since K is a cone, x/β ∈ K. Thus x/β ∈ K ∩ B and therefore x ∈ cone(K ∩ B). (cid:4) Note that the inclusion in Lemma 2.6(i) may be strict: indeed, consider the case when X = R, A := {1}, and B = {2}. Lemma 2.7 (a characterization of convexity) Let A be a nonempty closed subset of X. Then the follow- ing are equivalent: (i) A is convex. (ii) P−1 A (a) − a is a cone, for every a ∈ A. (iii) PA(x) is a singleton, for every x ∈ X. Proof. "(i)⇒(ii)": Indeed, it is well known in convex analysis (see, e.g., [24, Proposition 6.17]) that for every a ∈ A, P−1 A (a) − a is equal to the normal cone (in the sense of convex analysis) of A at a. "(ii)⇒(iii)": Let x ∈ X. By Proposition 2.2, PAx 6= ∅. Take a1 and a2 in PAx. Then kx − a1k = A a1 − a1. Hence A a − a is a cone, we have 2(x − a1) ∈ P−1 kx − a2k and x − a1 ∈ P−1 y := 2x − a1 ∈ P−1 (9a) A a1 − a1. Since P−1 A a1 and y − x = x − a1. Thus, hy − a2, a1 − a2i = h(y − x) + (x − a2), (a1 − x) + (x − a2)i 5 (9b) (9c) (9d) = hy − x, a1 − xi + hy − x, x − a2i + hx − a2, a1 − xi + kx − a2k2 = hx − a1, a1 − xi + hx − a1, x − a2i + hx − a2, a1 − xi + kx − a2k2 = −kx − a1k2 + kx − a2k2 = 0. (9e) Since a1 ∈ PAy, it follows that (10a) ky − a1k2 = ky − a2k2 + 2hy − a2, a2 − a1i + ka1 − a2k2 (10b) (10c) (10d) = ky − a2k2 + ka1 − a2k2 ≥ ky − a2k2 ≥ ky − a1k2. Hence equality holds throughout (10). Therefore, a1 = a2. "(iii)⇒(i)": This classical result due to Bunt and to Motzkin on the convexity of Chebyshev sets is well known; for proofs, see, e.g., [11, Chapter 12] or [4, Corollary 21.13]. (cid:4) Proposition 2.8 Let S be a convex set. Then the following are equivalent. (i) 0 ∈ ri S. (ii) cone S = span S. (iii) cone S = span S. Proof. Set Y = span S. Then (i) ⇔ 0 belongs to the interior of S relative to Y. "(i)⇒(ii)": There exists δ > 0 such that for every y ∈ Y r {0}, δy/kyk ∈ S. Hence y ∈ cone S. "(ii)⇒(i)": For every y ∈ Y, there exists δ > 0 such that δy ∈ S. Now [23, Corollary 6.4.1] applies in Y. "(ii)⇔(iii)": Set K = cone S, which is convex. By [23, Corollary 6.3.1], we have ri K = ri Y ⇔ K = Y ⇔ ri Y ⊆ K ⊆ Y. Since ri Y = Y = Y, we obtain the equivalences: ri K = Y ⇔ K = Y ⇔ K = Y. (cid:4) 3 Restricted normal cones: basic properties Normal cones are fundamental objects in variational analysis; they are used to construct subd- ifferential operators, and they have found many applications in optimization, optimal control, nonlinear analysis, convex analysis, etc.; see, e.g., [4], [6], [8], [19], [22], [23], [24]. One of the key building blocks is the Mordukhovich (or limiting) normal cone NA, which is obtained by limits of proximal normal vectors. In this section, we propose a new, very flexible, normal cone of A, denoted by NB A, by constraining the proximal normal vectors to a set B. 6 Definition 3.1 (normal cones) Let A and B be nonempty subsets of X, and let a and u be in X. If a ∈ A, then various normal cones of A at a are defined as follows: (i) The B-restricted proximal normal cone of A at a is (ii) The (classical) proximal normal cone of A at a is (12) (11) A(a) := cone(cid:16)(cid:0)B ∩ P−1 A a(cid:1) − a(cid:17) = cone(cid:16)(cid:0)B − a(cid:1) ∩(cid:0)P−1 bNB A a − a(cid:1). (a) := bNX A(a) is implicitly defined by u ∈ NB sequences (an)n∈N in A and (un)n∈N in bNB A (a) = cone(cid:0)P−1 A(an) such that an → a and un → u. Nprox A A a − a(cid:1)(cid:17). (iv) The Fr´echet normal cone NFr´e A (a) is implicitly defined by u ∈ NFr´e A (a) if and only if (∀ε > 0) (iii) The B-restricted normal cone NB A(a) if and only if there exist (∃ δ > 0) (∀x ∈ A ∩ ball(a; δ)) hu, x − ai ≤ εkx − ak. (v) The normal convex from convex analysis Nconv only if suphu, A − ai ≤ 0. A (a) is implicitly defined by u ∈ Nconv A (a) if and (vi) The Mordukhovich normal cone NA(a) of A at a is implicitly defined by u ∈ NA(a) if and only if there exist sequences (an)n∈N in A and (un)n∈N in Nprox A (an) such that an → a and un → u. If a /∈ A, then all normal cones are defined to be empty. The proximal normal cone A a Nprox A (a) A a Remark 3.2 Some comments regarding Definition 3.1 are in order. The restricted proximal normal cone P−1 A (a) ∩ B B A(a) bNB (i) Clearly, the restricted proximal normal cone generalizes the notion of the classical proximal A a is restricted normal cone. The name "restricted" stems from the fact that the pre-image P−1 to the set B. 7 (ii) See [24, Example 6.16] and [22, Subsection 2.5.2.D on page 240] for further information re- garding the classical proximal normal cone, including the fact that u ∈ Nprox (13) This also implies that: Nprox A A (a) + (A − a)⊖ ⊆ Nprox A (a). (a) ⇔ a ∈ A and (∃ δ > 0)(∀x ∈ A) hu, x − ai ≤ δkx − ak2. A. Put differently, NB A(a) is the outer (or upper Kura- (iii) Note that gr NB towski) limit of bNB (14) A = (A × X) ∩ gr bNB A(x) as x → a in A, written See also [24, Chapter 4]. NB A(a) = lim x→a A(x). x∈A bNB (iv) See [22, Definition 1.1] or [24, Definition 6.3] (where this is called the regular normal cone) for further information regarding NFr´e A (a). that NA = NX (v) The Mordukhovich normal cone is also known as the basic or limiting normal cone. Note and once again NA(a) (x) as x → a in A. See also [22, A and gr NA = (A × X) ∩ grbNX is the outer (or upper Kuratowski) limit of bNX A = (A × X) ∩ gr Nprox A (x) or Nprox page 141] for historical notes. A A The next result presents useful characterizations of the Mordukhovich normal cone. Proposition 3.3 (characterizations of the Mordukhovich normal cone) Let A be a nonempty closed subset of X, let a ∈ A, and let u ∈ X. Then the following are equivalent: (i) u ∈ NA(a). (ii) There exist sequences (λn)n∈N in R+, (bn)n∈N in X, (an)n∈N in A such that an → a, λn(bn − an) → u, and (∀n ∈ N) an ∈ PAbn. (iii) There exist sequences (λn)n∈N in R+, (xn)n∈N in X, (an)n∈N in A such that xn → a, λn(xn − an) → u, and (∀n ∈ N) an ∈ PAxn. (This also implies an → a.) (iv) There exist sequences (an)n∈N in A and (un)n∈N in X such that an → a, un → u, and (∀n ∈ N) un ∈ NFr´e A (an). Proof. "(i)⇔(ii)": Clear from Definition 3.1(vi). "(iii)⇔(iv)": Noting that the definition of NA(a) in [22] is the one given in (iv), we see that this equivalence follows from [22, Theorem 1.6]. "(ii)⇒(iii)": Let (λn)n∈N, (an)n∈N, and (bn)n∈N be as in (ii). For every n ∈ N, since an ∈ PAbn, [24, Example 6.16] implies that an ∈ PA[an, bn]. Now let (εn)n∈N be a sequence in ]0, 1[ such that εnan → 0 and εnbn → 0. Set (∀n ∈ N) (15) xn = (1 − εn)an + εnbn = an + εn(bn − an) ∈ [an, bn]. 8 Then xn → a and (∀n ∈ N) an ∈ PAxn. Furthermore, (λn/εn)n∈N lies in R+ and (16) (λn/εn)(xn − an) = λn(bn − an) → u. "(iii)⇒(ii)": Let (λn)n∈N, (xn)n∈N, and (an)n∈N be as in (iii). Since xn → a and a ∈ A, we deduce that 0 ≤ kxn − ank = dA(xn) ≤ kxn − ak → 0. Hence xn − an → 0 which implies that an − a = an − xn + xn − a → 0 + 0 = 0. Therefore, (ii) holds with (bn)n∈N = (xn)n∈N. (cid:4) Here are some basic properties of the restricted normal cone and its relation to various classical cones. Lemma 3.4 (basic inclusions among the normal cones) Let A and B be nonempty subsets of X, and let a ∈ A. Then the following hold: (i) Nconv A (a) ⊆ Nprox A (a). A a − a)) ⊆ (cone(B − a)) ∩ Nprox A (a). (a) and NB A(a) ⊆ NA(a). A(a) = cone((B − a) ∩ (P−1 A(a) ⊆ bNX A(a) ⊆ NB A (a) = Nprox A(a). A (ii) bNB (iii) bNB (iv) bNB (v) If A is closed, then Nprox (vi) If A is closed, then NFr´e A (a). A (a) ⊆ NFr´e A (a) ⊆ NA(a). A (a) = Nprox A (a) = NFr´e A (a) = Nconv A (a) = NA(a). A (vii) If A is closed and convex, then bNX (viii) If a ∈ ri(A), then bNaff(A) (ix) (aff(A) − a)⊥ ⊆ (A − a)⊖. (x) (A − a)⊖ ∩ cone(B − a) ⊆ bNB Proof. (i): Take u ∈ Nconv In view of (13), u ∈ Nprox A A (a) = Naff(A) A (a) = {0}. A(a) ⊆ cone(B − a). (a) and fix an arbitrary δ > 0. Then (∀x ∈ A) hu, x − ai ≤ 0 ≤ δkx − ak2. (a). (ii): In view of Lemma 2.6, the definitions yield (17a) (17b) (17c) A(a) = cone(cid:0)(B ∩ P−1 bNB ⊆ cone(cid:0)(B − a) ∩ cone(P−1 = cone(B − a) ∩ Nprox A a) − a(cid:1) = cone(cid:0)(B − a) ∩ (P−1 A a − a)(cid:1) A a − a)(cid:1) = cone(cid:0)(B − a) ∩ Nprox (a). A A (a)(cid:1) (iii), (iv) and (ix): This is obvious. 9 (v): Assume that A is closed and take u ∈ Nprox (a). By (13), there exists ρ > 0 such that (∀x ∈ A) hu, x − ai ≤ ρkx − ak2. Now let ε > 0 and set δ = ε/ρ. If x ∈ A ∩ ball(a; δ), then hu, x − ai ≤ ρkx − ak2 ≤ ρδkx − ak = εkx − ak. Thus, u ∈ NFr´e A (a). A (vi): This follows from Proposition 3.3. (vii): Since A is closed, it follows from (i), (v), and (vi) that (18) Nconv A (a) ⊆ Nprox A (a) ⊆ NFr´e A (a) ⊆ NA(a). On the other hand, by [22, Proposition 1.5], NA(a) ⊆ Nconv A (a) because A is convex. A Therefore, Naff(A) (a) = {0}. Since a ∈ ri(A), it follows that (∀x ∈ ball(a; δ/2) ∩ aff(A)) bNaff(A) (viii): By assumption, (∃ δ > 0) ball(a; δ) ∩ aff(A) ⊆ A. Hence aff(A) ∩ P−1 bNaff(A) (x): Take u ∈ ((A − a)⊖ ∩ cone(B − a)) r {0}, say u = λ(b − a), where b ∈ B and λ > 0. Then 0 ≥ suphA − a, ui = λ suphA − a, b − ai = sup λ hconv A − a, b − ai. By Fact 2.5(ii), a = Pconv Ab and hence a = PAb. It follows that u ∈ cone((B ∩ P−1 A a) − a). The left inclusion thus holds. The right inclusion is clear. A a = {a} and thus (x) = {0}. (a) = {0}. (cid:4) A A A(a), NA(a), and Nconv Remark 3.5 (on closedness of normal cones) Let A be a nonempty subset of X, let a ∈ A, and let B be a subset of X. Then NB (a) are obviously closed -- this is also true for NFr´e A (a) but requires some work (see [24, Proposition 6.5]). On the other hand, the classical proximal normal cone Nprox A (a) is not necessarily closed (see, e.g., [24, page 213]), and A(a). For a concrete example, suppose that X = R2, that A = {(0, 0)}, that A(a) = (cid:0)R × R++(cid:1) ∪ {(0, 0)}, which is not closed; hence neither is bNB B = R × {1} and that a = (0, 0). Then bNB however, the classical proximal normal cone Nprox (a) = bNX (a) = R2 is closed. A A A The sphere is a nonconvex set for which all classical normal cones coincide: Example 3.6 (classical normal cones of the sphere) Let z ∈ X and ρ ∈ R++. Set S := sphere(z; ρ) and let s ∈ S. Then Nprox Proof. By Example 2.3, we have P−1 Hence, using Lemma 3.4(v)&(vi), we have S (s) = z + R+(s − z) and so P−1 S (s) − s = [−1, +∞[ · (s − z). S (s) = NS(s) = R(s − z). (s) = bNX S (s) = NFr´e S Nprox S (19a) (19b) (19c) as announced. (s) = bNX S (s) = R(s − z) ⊆ NFr´e = lim s′→S s′∈S (s′) = lim s′→S s′∈S Nprox S = Nprox S S (s) ⊆ NS(s) R(s′ − z) = R(s − z) (s), (cid:4) Here are some elementary yet useful calculus rules. 10 Proposition 3.7 Let A, A1, A2, B, B1, and B2 be nonempty subsets of X, let c ∈ X, and suppose that a ∈ A ∩ A1 ∩ A2. Then the following hold: A A (a). A(a) = NB A2 (a). A1 (a) = NB1 A A(a) = N−B A (a) ∪ NB2 −A(−a), and −NA(a) = N−A(−a). A(a) is convex. A (a) and NB1∪B2 A(a) = {0}. (i) If A and B are convex, then bNB (ii) bNB1∪B2 A (a) ∪ bNB2 (a) = bNB1 (iii) If B ⊆ A, then bNB (a) ⊆ bNB (iv) If A1 ⊆ A2, then bNB A(a) = bN−B (v) −bNB −A(−a), −NB A(a) = bNB−c (vi) bNB A−c(a − c) and NB Proof. It suffices to establish the conclusions for the restricted proximal normal cones since the restricted normal cone results follows by taking closures (or outer limits). (i): We assume that B ∩ P−1 A a 6= ∅, for otherwise the conclusion is clear. Then P−1 a = (Id +NA)a is convex (as the image of the maximally monotone operator Id +NA at a). Hence (B ∩ P−1 A a) − a is convex A(a). (ii): Since ((B1 ∪ B2) ∩ P−1 as well, and so is its conical hull, which is bNB A a) − a = ((B1 ∩ P−1 A a) − a) ∪ ((B2 ∩ P−1 A a) − a), the result follows by taking the conical hull. (iii): Clear, because (B ∩ P−1 A a) − a is either empty or equal to {0}. (iv): Suppose λ(b − a) ∈ bNB (a), where λ ≥ 0, b ∈ B, and a ∈ PA2b. Since a ∈ A1 ⊆ A2, we have a ∈ PA1b. Hence λ(b − a) ∈ bNB (a). (v): This follows by using elementary manipulations and the fact that P−A = (− Id) ◦ PA ◦ (− Id). (vi): This follows readily from the fact that P−1 A−c(a − c). A (a) = P−1 A A(a) = NB−c A2 A1 (cid:4) A−c(a − c) = P−1 A (a) − c. Remark 3.8 The restricted normal cone counterparts of items (i) and (iv) are false in general; see Example 5.1 (and also Example 5.4(iv)) below. The Mordukhovich normal cone (and hence also the Clarke normal cone which contains the Mordukhovich normal cone) strictly contains {0} at boundary points (see [22, Corollary 2.24] or [24, Exercise 6.19]); however, the restricted normal cone can be {0} at boundary points as we illustrate next. Example 3.9 (restricted normal cone at boundary points) Suppose that X = R2, set A := ball(0; 1) =(cid:8)x ∈ R2(cid:12)(cid:12) kxk ≤ 1(cid:9) and B := R × {2}, and let a = (a1, a2) ∈ A. Then (20) Consequently, (21) A(a) =(R+a, bNB A(a) =(R+a, NB if kak = 1 and a2 > 0; {(0, 0)}, otherwise. if kak = 1 and a2 ≥ 0; {(0, 0)}, otherwise. 11 Thus the restricted normal cone is {(0, 0)} for all boundary points in the lower half disk that do not "face" the set B. Remark 3.10 In contrast to Example 3.9, we shall see in Corollary 4.11(ii) below that if A is closed, B is the affine hull of A, and a belongs to the relative boundary of A, then the restricted normal cone NB A(a) strictly contains {0}. 4 Restricted normal cones and affine subspaces In this section, we consider the case when the restricting set is a suitable affine subspace. This results in further calculus rules and a characterization of interiority notions. The following four lemmas are useful in the derivation of the main results in this section. Lemma 4.1 Let A and B be nonempty subsets of X, and suppose that c ∈ A ∩ B. Then (22) aff(A ∪ B) − c = span(B − A). Proof. Since c ∈ A ∩ B ⊆ A ∪ B, it is clear that the aff(A ∪ B) − c is a subspace. On the one hand, if a ∈ A and b ∈ B, then b − a = 1 · b + (−1) · a + 1 · c − c ∈ aff(A ∪ B) − c. Hence B − A ⊆ aff(A ∪ B) − c and thus span(B − A) ⊆ aff(A ∪ B) − c. On the other hand, if x ∈ aff(A ∪ B), say x = ∑i∈I λiai + ∑j∈J µjbj, where each ai belongs to A, each bj belongs to B, and ∑i∈I λi + ∑j∈J µj = 1, then x − c = ∑i∈I(−λi)(c − ai) + ∑j∈I µj(bj − c) ∈ span(B − A). Thus aff(A ∪ B) − c ⊆ span(B − A). Lemma 4.2 Let A be a nonempty subset of X, let a ∈ A, and let u ∈ (aff(A) − a)⊥. Then (23) (cid:4) (∀x ∈ X) PA(x + u) = PA(x). Proof. Let x ∈ X. For every b ∈ A, we have (24a) ku + x − bk2 = kuk2 + 2hu, x − bi + kx − bk2 (24b) = kuk2 + 2hu, x − ai + 2hu, a − bi + kx − bk2 = kuk2 + 2hu, x − ai + kx − bk2. (24c) Hence PA(x + u) = argminb∈A ku + x − bk2 = argminb∈A kx − bk2 = PAx, as announced. Lemma 4.3 Let A be a nonempty subset of X, and let L be an affine subspace of X containing A. Then (cid:4) (25) PA = PA ◦ PL. Proof. Let a ∈ A and x ∈ X, and set b = PLx. Using [4, Corollary 3.20(i)], we have x − b ∈ (L − a)⊥ ⊂ (aff(A) − a)⊥. In view of Lemma 4.2, we deduce that (PA ◦ PL)x = PA(b) = PA(b + (x − b)) = PAx. (cid:4) 12 Lemma 4.4 Let A be a nonempty subset of X, let a ∈ A, and let L be an affine subspace of X containing A. Then the following hold: (i) bNL (ii) NL A(a)⊥(L − a)⊥. A(a)⊥(L − a)⊥. Proof. Observe that L − a = par(A) does not depend on the concrete choice of a ∈ A. (i): Using A(a) ⊆ cone(L − a) ⊆ span(L − a) ⊥ (span(L − a))⊥ = (L − a)⊥ = A ⊆ par A. Since ran NL A ⊆ par A = A, it follows that ran NL Lemma 3.4(x), we see that bNL (par A)⊥. (ii): By (i), ran bNL L − a. A ⊆ ran bNL (cid:4) For a normal cone restricted to certain affine subspaces, it is possible to derive precise relation- ships to the Mordukhovich normal cone. Theorem 4.5 (restricted vs Mordukhovich normal cone) Let A and B be nonempty subsets of X, suppose that a ∈ A, and let L be an affine subspace of X containing A. Then the following hold: (26a) (26b) (26c) (26d) A (a) = bNL bNX A(a) = bNX bNL A (a) + (L − a)⊥, A(a) ⊕ (L − a)⊥ = bNX A (a) ∩ (L − a), A(a) ⊕ (L − a)⊥ = NA(a) + (L − a)⊥, NA(a) = NL NL A(a) = NA(a) ∩ (L − a). Consequently, the following hold as well: (27a) (27b) (27c) (27d) (27e) A A (a) = bNaff(A) bNX bNaff(A) (a) = bNX (a) = NA(a) ∩(cid:0) aff(A) − a(cid:1), Naff(A) a ∈ A ∩ B ⇒ Naff(A∪B) (a) = NA(a) ∩ span(A − B). A (a) ∩ (aff(A) − a), NA(a) = Naff(A) (a) ⊕ (aff(A) − a)⊥ = bNX A (a) + (aff(A) − a)⊥, (a) ⊕(cid:0) aff(A) − a(cid:1)⊥ = NA(a) +(cid:0) aff(A) − a(cid:1)⊥, A A A A (28) Proof. (26a): Take u ∈ bNX A (a). Then there exist λ ≥ 0, x ∈ X, and a ∈ PAx such that λ(x − a) = u. Set b = PLx. By Lemma 4.3, we have a ∈ PAx = (PA ◦ PL)x = PAb. Using [4, Corollary 3.20(i)], we thus deduce that λ(b − a) ∈ bNL A(a) and λ(x − b) ∈ (L − b)⊥ = (L − a)⊥. Hence u = λ(b − a) + A(a) + (L − a)⊥ = bNL λ(x − b) ∈ bNL A(a) ⊕ (L − a)⊥ by Lemma 4.4(i). We have thus shown that A (a) ⊆ bNL bNX On the other hand, Lemma 3.4(iii) implies that bNL A(a) + (L − a)⊥ ⊆ bNX bNL A(a) ⊕ (L − a)⊥. A(a) ⊆ bNX A (a) + (L − a)⊥. A (a) and thus (29) 13 Altogether, (30) bNX A (a) ⊆ bNL A(a) ⊕ (L − a)⊥ ⊆ bNX A (a) + (L − a)⊥. A (a), as required. A (a) + (L − a)⊥ ⊆ bNX To complete the proof of (26a), it thus suffices to show that bNX A (a). To this end, let u ∈ bNX A (a) and v ∈ (L − a)⊥ ⊆ (aff(A) − a)⊥. Then there exist λ ≥ 0, b ∈ X, and a ∈ PAb such that u = λ(b − a). If λ = 0, then u = 0 and u + v = v ∈ (aff(A) − a)⊥ ⊆ (A − a)⊖ = (A − a)⊖ ∩ X = (A − a)⊖ ∩ cone(X − a) ⊆ bNX A (a) by Lemma 3.4(ix)&(x). Thus, we assume that λ > 0. By Lemma 4.2, we have a ∈ PAb = PA(b + λ−1v). Hence b + λ−1v − a ∈ bNX A (a) and therefore λ(b + λ−1v − a) = λ(b − a) + v = u + v ∈ bNX A (a) ∩ (L − a). Now let u ∈ bNX A(a) ⊆ bNX (26b): By Lemma 3.4(iii)&(x), bNL A (a) ∩ (L − a). By (26a), we have u = v + w, where v ∈ bNL A(a) ⊆ L − a and w ∈ (L − a)⊥. On the other hand, w = u − v ∈ (L − a) − (L − a) = L − a. Altogether w ∈ (L − a) ∩ (L − a)⊥ = {0}. Hence u = v ∈ bNL (26c): Let u ∈ NA(a). By definition, there exist sequences (an)n∈N in A and (un)n∈N in X such that an → a, un → u, and (∀n ∈ N) un ∈ bNX A (an). By (26a), there exists a sequence (vn, wn)n∈N such that (an, vn)n∈N lies in grbNL A, (wn)n∈N lies in (L − a)⊥, and (∀n ∈ N) un = vn + wn and vn ⊥ wn. Since kuk2 ← kunk2 = kvnk2 + kwnk2, the sequences (vn)n∈N and (wn)n∈N are bounded. After passing to subsequences and relabeling if necessary, we assume (vn)n∈N and (wn)n∈N are A(a) and w ∈ (L − a)⊥; convergent, with limits v and w, respectively. A(a) ⊕ (L − a)⊥. consequently, u = v + w ∈ NL On the other hand, by Lemma 3.4(iii), NL A(a) ⊕ (L − a)⊥ by Lemma 4.4(ii). Thus NA(a) ⊆ NL It follows that v ∈ NL A(a) ⊕ (L − a)⊥ ⊆ NA(a) + (L − a)⊥. Altogether, A(a). (31) NA(a) ⊆ NL A(a) ⊕ (L − a)⊥ ⊆ NA(a) + (L − a)⊥. It thus suffices to prove that NA(a) + (L − a)⊥ ⊆ NA(a). To this end, take u ∈ NA(a) and v ∈ (L − a)⊥. Then there exist sequences (an)n∈N in A and (un)n∈N in X such that an → a, un → u, and (∀n ∈ N) un ∈ bNX A (an). For every n ∈ N, we have L − a = L − an and hence un + v ∈ A (an) + (L − an)⊥ = bNX bNX A (an) by (26a). Passing to the limit, we conclude that u + v ∈ NA(a). (26d): First, take u ∈ NL A(a). On the one hand, by Lemma 3.4(iii), u ∈ NA(a). On the other hand, by Lemma 4.4(ii), u ∈ (L − a)⊥⊥ = L − a. Altogether, we have shown that (32) NL A(a) ⊆ NA(a) ∩ (L − a). A(a) and w ∈ (L − a)⊥ Conversely, take u ∈ NA(a) ∩ (L − a) ⊆ NA(a). By (26c), there exist v ∈ NL such that u = v + w and v ⊥ w. By (32), v ∈ L − a. Hence w = u − v ∈ (L − a) − (L − a) = L − a. Since w ∈ (L − a)⊥, we deduce that w = 0. This implies u = v ∈ NL A(a). Therefore, NA(a) ∩ (L − a) ⊆ NL A(a). "Consequently" part: Consider (26) when L = aff(A) or L = aff(A ∪ B), and recall Lemma 4.1 (cid:4) in the latter case. An immediate consequence of Theorem 4.5 (or of the definitions) is the following result. 14 Corollary 4.6 (the X-restricted and the Mordukhovich normal cone coincide) Let A be a nonempty subset of X, and let a ∈ A. Then (33) NX A (a) = NA(a). The next two results provide some useful calculus rules. Corollary 4.7 (restricted normal cone of a sum) Let C1 and C2 be nonempty closed convex subsets of X, let a1 ∈ C1, let a2 ∈ C2, and let L be an affine subspace of X containing C1 + C2. Then (34) (a1 + a2) = NL−a2 (a2). NL C1+C2 C1 (a1) ∩ NL−a1 C2 Proof. Set C = C1 + C2 and a = a1 + a2. Then (26d) and [24, Exercise 6.44] yield (35a) (35b) NL C(a) = NC(a) ∩ (L − a) = NC1 (a1) ∩ NC2 (a2) ∩ (L − a) =(cid:0)NC1 (a1) ∩ (L − a)(cid:1) ∩(cid:0)NC2 (a2) ∩ (L − a)(cid:1). Note that L − a is a linear subspace of X containing C1 − a1 and C2 − a2. Thus, L − a2 = L − a + a1 is an affine subspace of X containing C1, and L − a1 = L − a + a2 is an affine subspace of X containing C2. By (26d), (36) NL−a2 C1 (a1) = NC1 (a1) ∩ (L − a) and NL−a1 C2 (a2) = NC2 (a2) ∩ (L − a). The conclusion follows by combining (35) and (36). (cid:4) Corollary 4.8 (an intersection formula) Let A and B be nonempty closed convex subsets of X, and sup- pose that a ∈ A ∩ B. Let L be an affine subspace of X containing A ∪ B. Then (37) NL Proof. Using (26d), Proposition 3.7(v), [24, Exercise 6.44], and again (26d), we obtain Let us now work towards relating the restricted normal cone to the (relative and classical) inte- rior and to the boundary of a given set. 15 (cid:4) NL A(a) ∩(cid:0) − NL A−B(0). B (a)(cid:1) = NL−a A(a) ∩(cid:0) − NL B (a)(cid:1) = NA(a) ∩ (L − a) ∩(cid:0) − NB(a)(cid:1) ∩ (L − a) =(cid:16)NA(a) ∩(cid:0) − NB(a)(cid:1)(cid:17) ∩ (L − a) =(cid:0)NA(a) ∩ N−B(−a)(cid:1) ∩ (L − a) = NA−B(0) ∩ (L − a) = NL−a A−B(0), (38a) (38b) (38c) (38d) (38e) as required. Proposition 4.9 Let A be a nonempty subset of X, let a ∈ A, let L be an affine subspace containing A, and suppose that NL A(a) = {0}. Then L = aff(A). Proof. Using 0 ∈ Naff(A) A (a) ⊆ NL A(a) = {0} and applying (26c) and (27c), we have (39) NA(a) = 0 + (L − a)⊥ = 0 + (aff(A) − a)⊥. So L − a = aff(A) − a, i.e., L = aff(A). Theorem 4.10 Let A and B be nonempty subsets of X, and let a ∈ A. Then (40) A(a) = {0} ⇔ (∃ δ > 0)(cid:0)∀x ∈ A ∩ ball(a; δ)(cid:1) P−1 NB A (x) ∩ B ⊆ {x}. Furthermore, if A is closed and B is an affine subspace of X containing A, then the following are equivalent: (cid:4) (i) NB A(a) = {0}. (ii) (∃ ρ > 0) ball(a; ρ) ∩ B ⊆ A. (iii) B = aff(A) and a ∈ ri(A). Proof. Note that NB from the definition of bNB A(a) = {0} ⇔ (∃ δ > 0) (∀x ∈ A ∩ ball(a; δ)) bNB A(x). A(x) = {0}. Hence (40) follows Now suppose that A is closed and B is an affine subspace of X containing A. "(i)⇒(ii)": Let δ > 0 be as in (40) and set ρ := δ/2. Let b ∈ B(a; ρ) ∩ B, and take x ∈ PAb, which is possible since A is closed. Then kb − xk = dA(b) ≤ kb − ak ≤ ρ and hence (41) Using (40), we deduce that b ∈ P−1 kx − ak ≤ kx − bk + kb − ak ≤ ρ + ρ = 2ρ = δ. A (x) ∩ B ⊆ {x} ⊆ A. "(ii)⇒(iii)": It follows that B = aff(B) ⊆ aff(A) ⊆ B; hence, B = aff(A). Thus ball(a; ρ) ∩ aff(A) ⊆ A, which means that a ∈ ri(A). "(iii)⇒(i)": Lemma 3.4(viii). (cid:4) Corollary 4.11 (interior and boundary characterizations) Let A be a nonempty closed subset of X, and let a ∈ A. Then the following hold: (i) Naff(A) A (ii) Naff(A) A (a) = {0} ⇔ a ∈ ri(A). (a) 6= {0} ⇔ a ∈ A r ri(A). (iii) NA(a) = {0} ⇔ a ∈ int(A). 16 (iv) NA(a) 6= {0} ⇔ a ∈ A r int(A). Proof. (i): Apply Theorem 4.10 with B = aff(A). (ii): Clear from (i). (iii): Apply Theorem 4.10 with B = X, and recall Corollary 4.6. (iv): Clear from (iii). (cid:4) A second look at the proof of (i)⇒(ii) in Theorem 4.10 reveals that this implication does actually not require the assumption that B be an affine subspace of X containing A. The following example illustrates that the converse implication fails even when B is a superset of aff(A). Example 4.12 Suppose that X = R2, and set A := R × {0}, a = (0, 0), and B = R × {0, 2}. Then A = aff(A) ⊆ B and ball(a; 1) ∩ B ⊆ A; however, (∀x ∈ A) bNB A(x) = {0} × R+ and therefore NB A(a) = {0} × R+ 6= {(0, 0)}. Two convex sets It is instructive to interpret the previous results for two convex sets: Theorem 4.13 (two convex sets: restricted normal cones and relative interiors) Let A and B be nonempty convex subsets of X. Then the following are equivalent: (i) ri A ∩ ri B 6= ∅. (ii) 0 ∈ ri(B − A). (iii) cone(B − A) = span(B − A). (iv) NA(c) ∩ (−NB(c)) ∩ cone(B − A) = {0} for some c ∈ A ∩ B. (v) NA(c) ∩ (−NB(c)) ∩ cone(B − A) = {0} for every c ∈ A ∩ B. (vi) NA(c) ∩ (−NB(c)) ∩ span(B − A) = {0} for some c ∈ A ∩ B. (vii) NA(c) ∩ (−NB(c)) ∩ span(B − A) = {0} for every c ∈ A ∩ B. (viii) Naff(A∪B) A (c)) = {0} for some c ∈ A ∩ B. (c)) = {0} for every c ∈ A ∩ B. B (c) ∩ (−Naff(A∪B) (c) ∩ (−Naff(A∪B) (0) = {0}. B (ix) Naff(A∪B) A (x) Nspan(B−A) A−B Proof. By [23, Corollary 6.6.2], (ii) ⇔ ri A ∩ ri B 6= ∅ ⇔ 0 ∈ ri A − ri B ⇔ (ii). Applying Proposition 2.8 to B − A, and [3, Proposition 3.1.3] to cone (B − A), we obtain (42a) (ii) ⇔ (iii) ⇔ cone (B − A) = span(B − A) 17 (42b) ⇔ cone (B − A) ∩(cid:0)cone (B − A)(cid:1)⊕ = {0}. Let c ∈ A ∩ B. Then Corollary 4.8 (with L = X) yields NA(c) ∩(cid:0) − NB(c)(cid:1) = NA−B(0) = (A − B)⊖ = (B − A)⊕ = (cone(B − A))⊕. Hence (43) (∀c ∈ C) NA(c) ∩(cid:0) − NB(c)(cid:1) ∩ cone (B − A) =(cid:0)cone (B − A)(cid:1)⊕ ∩ cone (B − A) (∀c ∈ C) NA(c) ∩(cid:0) − NB(c)(cid:1) ∩ span(B − A) =(cid:0)cone (B − A)(cid:1)⊕ ∩ span(B − A). Combining (42), (43), and (44), we see that (ii) -- (vii) are equivalent. and (44) Next, Lemma 4.1 and Corollary 4.8 yield the equivalence of (viii) -- (x). Finally, (x)⇔(ii) by Corollary 4.11(i). (cid:4) Corollary 4.14 (two convex sets: normal cones and interiors) Let A and B be nonempty convex sub- sets of X. Then the following are equivalent: (i) 0 ∈ int(B − A). (ii) cone(B − A) = X. (iii) NA(c) ∩ (−NB(c)) = {0} for some c ∈ A ∩ B. (iv) NA(c) ∩ (−NB(c)) = {0} for every c ∈ A ∩ B. (v) NA−B(0) = {0}. Proof. We start by notating that if C is a convex subset of X, then 0 ∈ int C ⇔ 0 ∈ ri C and span C = X. Consequently, (45) (i) ⇔ 0 ∈ ri(B − A) and span(B − A) = X. Assume that (i) holds. Then (45) and Theorem 4.13 imply that cone(B − A) = cone (B − A) = span(B − A) = X. Hence (ii) holds, and from Theorem 4.13 we obtain that (ii)⇒(iii)⇔(iv)⇔(v). Finally, Corollary 4.11(iii) yields the implication (v)⇒(i). (cid:4) 5 Further examples In this section, we provide further examples that illustrate particularities of restricted normal cones. As announced in Remark 3.8, when a ∈ A2 $ A1, it is possible that the nonconvex restricted (a) even when A1 and A2 are both convex. This lack of inclusion normal cones satisfy NB A1 A2 (a) 6⊆ NB 18 (46a) (46b) Consequently, (47a) (47b) (cid:0)∀x = (x1, x2) ∈ A1(cid:1) (cid:0)∀x = (x1, x2) ∈ A2(cid:1) A2 A1 R+(1, −1), R+(−1, −1), {(0, 0)}, R+(2, −1), R+(−2, −1), {(0, 0)}, (x) = bNB (x) = bNB (a) = cone(cid:8)(1, −1), (−1, −1)(cid:9), (a) = cone(cid:8)(2, −1), (−2, −1)(cid:9). (a) 6⊆ NB (a); in fact, NB A1 A1 NB A1 NB A2 is also known for the Mordukhovich normal cone (see [22, page 5], where however one of the sets is not convex). Furthermore, the following example also shows that the restricted normal cone cannot be derived from the Mordukhovich normal cone by the simple relativization procedure of intersecting with naturally associated cones and subspaces. Example 5.1 (lack of convexity, inclusion, and relativization) Suppose that X = R2, and define two nonempty closed convex sets by A := A1 := epi( · ) and A2 := epi(2 · ). Then a := (0, 0) ∈ A2 $ A1. Furthermore, set B := R × {0}. Then if x2 = x1 > 0; if x2 = −x1 > 0; otherwise, if x2 = 2x1 > 0; if x2 = −2x1 > 0; otherwise. A2 (a) nor NB A2 (a) and NB A2 (a) 6⊆ NB Note that NB (a) = {(0, 0)}. Fur- A1 thermore, neither NB (a) is convex even though A1, A2, and B are. Finally, observe A1 that cone(B − a) = span(B − a) = B, that cone(B − A) = R × R−, that span(B − A) = X, A(a). Consequently, cone(B − a) ∩ NA(a) = and that NA(a) = cone[(1, −1), (−1, −1)] span(B − a) ∩ NA(a) = {(0, 0)}, cone(B − A) ∩ NA(a) = NA(a) = span(B − A) ∩ NA(a). There- fore, NB A(a) cannot be obtained by intersecting the Mordukhovich normal cone with one of the sets cone(B − a), span(B − a), cone(B − A), and span(B − A). (a) ∩ NB 6= NB A2 We shall present some further examples. The proof of the following result is straight-forward and hence omitted. Proposition 5.2 Let K be a closed cone in X, and let B be a nonempty cone of X. Then (48) NB K (0) = [x∈K bNB K (x) = [x∈bdry K bNB K (x) = [x∈K NB K (x) = [x∈bdry K NB K (x). Example 5.3 Let K be a closed convex cone in X, suppose that u0 ∈ int(K) and that K ⊆ {u0}⊕, and set B := {u0}⊥. Then: (i) (∀x ∈ K ∩ B) bNB (ii) (∀x ∈ K r B) bNB K (x) = NK(x) = K⊖ ∩ {x}⊥. K (x) = {0}. K (x) = NB 19 (iii) NB K (x) =Sx∈KrB(K⊖ ∩ {x}⊥) = K⊖ ∩Sx∈KrB{x}⊥. If one of these unions is closed, then all closures may be omitted. K (0) =Sx∈K bNB Proof. (i): Let x ∈ K ∩ B. It suffices to show that B∩ P−1 By definition of B, we have hu0, xi = 0 and hu0, yi = 0. Hence (49) hu0, y − xi = 0. K (x) = {x}. To this end, take y ∈ B∩ P−1 K (x). Furthermore, x = PKy and hence, using e.g. [4, Proposition 6.27], we have y − x ∈ K⊖. Since u0 ∈ int K, there exists δ > 0 such that ball(u0; δ) ⊆ K. Thus y − x ∈ (ball(u0; δ))⊖. In view of (49), δky − xk ≤ 0. Therefore, y = x. (ii): Let x ∈ K r B. Using Lemma 3.4(iii)&(iv), Corollary 4.6, Lemma 3.4(vii), and [4, Exam- ple 6.39], we have bNB K (x) ⊆ bNX (50) K (x) ⊆ NX K (x) = NK(x) = Nconv K (x) = K⊖ ∩ {x}⊥. Since x ∈ K ⊆ {u0}⊕ and x /∈ B, we have hu0, xi > 0. Now take u ∈ (K⊖ ∩ {x}⊥) r {0}. Since u ∈ K⊖ and u0 ∈ int(K), we have hu, u0i < 0. Now set b := x − hu0, xi hu0, ui (51) u. Then b ∈ B and b − x = − hu0, xi hu0, ui−1 u ∈ R++u ⊆ K⊖ ∩ {x}⊥ = Nconv tion 6.46], x = PKb. Hence b − x ∈ bNB view of (50), and since bNB K (x) ⊆ NB Example 5.4 (ice cream cone) Suppose that X = Rm = Rm−1 × R, where m ∈ {2, 3, 4, . . .}, and let β > 0. Define the corresponding closed convex ice cream cone by (x). By [4, Proposi- K (x). In K (x) ⊆ NK(x) by Lemma 3.4(iii)&(iv), we have established (ii). K (x). Therefore, K⊖ ∩ {x}⊥ ⊆ bNB K (x) and thus u ∈ bNB (iii): Combine (i), (ii), and Proposition 5.2. (cid:4) K (52) K :=(cid:26)x ∈ Rm(cid:12)(cid:12)(cid:12) βqx2 1 + · · · + x2 m−1 ≤ xm(cid:27) , K (0, 0) = {(0, 0)}. and set B := Rm−1 × {0}. Then the following hold: (i) bNB (ii) NK(0, 0) =(cid:8)y ∈ Rm(cid:12)(cid:12) β−1qy2 (iii) (∀z ∈ Rm−1 r {0}) bNB K (0, 0) =S z∈Rm−1 1 + · · · + y2 (iv) NB kzk=1 m−1 ≤ −ym(cid:9) =S z∈Rm−1 kzk≤1 R+(βz, −1). K (z, βkzk) = NB R+(βz, −1), which is a closed cone that is not convex. K (z, βkzk) = NK(z, βkzk) = R+(βz, −kzk). 20 Proof. Clearly, K is closed and convex. Note that K is the lower level set of height 0 of the continu- ous convex function (53) f : Rm = Rm−1 × R → R : x = (z, xm) 7→ βkzk − xm; hence , by [26, Exercise 2.5(b) and its solution on page 205], Lemma 3.4(iii)&(iv), Corollary 4.6, and Corollary 4.11(iii) imply that int(K) =(cid:8)x = (z, xm) ∈ Rm−1 × R(cid:12)(cid:12) βkzk < xm(cid:9). (cid:0)∀x ∈ int(K)(cid:1) K (x) ⊆ NX bNB K (x) ⊆ bNX K (x) = NK(x) = {0}. Write x = (z, xm) ∈ Rm−1 × R = X, and assume that x ∈ K. We thus assume that x ∈ bdry(K), i.e., βkzk = xm by (54), i.e., x = (z, βkzk). Combining [4, Proposition 16.8] with [26, Corollary 2.9.5] (or [4, Lemma 26.17]) applied to f , we obtain (54) (55) (56) (57) NK(cid:0)z, βkzk(cid:1) = cone(cid:0)β∂k · k(z) × {−1}(cid:1), where ∂k · k denotes the subdifferential operator from convex analysis applied to the Euclidean norm in Rm−1. In view of [4, Example 16.25] we thus have NK(cid:0)z, βkzk(cid:1) =(cone(cid:0)βkzk−1z × {−1}(cid:1), cone(cid:0) ball(0; β) × {−1}(cid:1), if z 6= 0; if z = 0. The case z = 0 in (57) readily leads to (ii). Now set u0 := (0, 1) ∈ Rm−1 × R. Then {u0}⊥ = B and {u0}⊕ = Rm−1 × R+ ⊇ K. Note that (0, 0) ∈ K ∩ B and thus bNB Now assume that z 6= 0. Then NK(z, βkzk) = R+(βz, −kzk). Note that βz 6= 0 and so K (0, 0) = {(0, 0)} by Example 5.3(i). We have thus established (i). (z, βkzk) /∈ B. The formulas announced in (iii) therefore follow from Example 5.3(ii). Next, combining (54), (55), and Example 5.3(iii) as well as utilizing the compactness of the unit sphere in Rm−1, we see that (58) NB K (0, 0) = [z∈Rm−1r{0} R+(βz, −kzk) = [z∈Rm−1 kzk=1 R+(βz, −1) = [z∈Rm−1 kzk=1 R+(βz, −1). This establishes (iv). (cid:4) Remark 5.5 Consider Example 5.4. Note that NB thermore, since NK(0, 0) = Nconv which is therefore an ice cream cone as well. K K (0, 0) is actually the boundary of NK(0, 0). Fur- (0, 0) by Lemma 3.4(vii), the formulas in (ii) also describe K⊖, 21 6 Cones containing restricted normal cones In this section, we provide various examples illustrating that the restricted (proximal) normal cone does not naturally arise by considering various natural cones containing it. Let A and B be nonempty subsets of X, and let a ∈ A. We saw in Lemma 3.4(ii) that A(a) = cone(cid:0)(B − a) ∩ (P−1 bNB A a − a)(cid:1) ⊆ cone(B − a) ∩ Nprox A (a). This raises the question whether or not the inclusion in (59) is strict. It turns out and as we shall now illustrate, both conceivable alternatives (equality and strict inclusion) do occur. Therefore, A(a) is a new construction. We start with a condition sufficient for equality in (59), (59) bNB Proposition 6.1 Let A and B be nonempty subsets of X. Let A be closed and a ∈ A. Assume that one of the following holds: (i) P−1 A (a) − a is a cone. (ii) A is convex. A(a) = cone(B − a) ∩ Nprox A (a). Then bNB Proof. (i): Lemma 2.6(ii). (ii): Combine (i) with Lemma 2.7. (cid:4) The next examples illustrates that equality in (59) can occur even though P−1 A (a) − a is not a A (a) − a be a cone in Proposition 6.1 is sufficient -- but cone. Consequently, the assumption that P−1 not necessary -- for equality in (59). Example 6.2 Suppose that X = R2, and let A := X r R2 one verifies that ++, B := R+(1, 1), and a := (0, 1). Then P−1 A (a) − a = [0, 1] × {0}, Nprox (a) = cone(P−1 A A a − a) = R+ × {0}, cone(B − a) =(cid:8)(t1, t2) ∈ R2(cid:12)(cid:12) t1 ≥ 0, t2 < t1(cid:9) ∪ {(0, 0)}, A(a) = R+ × {0}. bNB A(a) = R+ × {0} = cone(B − a) ∩ Nprox A (a). (60a) (60b) (60c) (60d) Hence bNB We now provide an example where the inclusion in (59) is strict. 22 +, B := R+(2, 1), and Example 6.3 Suppose that X = R2, let A := cone{(1, 0), (0, 1)} = bdry R2 a := (0, 1) ∈ A. Then one verifies that (61a) P−1 A (a) − a = ]−∞, 1] × {0}, Nprox (a) = cone(P−1 A A a − a) = R × {0}, cone(B − a) =(cid:8)(x1, x2) ∈ R2(cid:12)(cid:12) x1 ≥ 0, 2x2 < x1(cid:9) ∪ {(0, 0)}, A(a) = {(0, 0)}. (61b) (61c) (61d) bNB A(a) = {(0, 0)} $ R+ × {0} = cone(B − a) ∩ Nprox A is strict. In accordance with Proposition 6.1, neither is P−1 Hence bNB (a), and therefore the inclusion in (59) A (a) − a a cone nor is A convex. Let us now turn to the restricted normal cone NB A(a). Taking the outer limit in (59) and recalling (14), we obtain (62a) (62b) (62c) NB A(a) = lim x→a ⊆ lim x→a A(x) x∈A bNB x∈A(cid:0) cone(B − x) ∩ Nprox (x)(cid:1) cone(B − x)(cid:1) ∩ NA(a). ⊆(cid:0) lim x→a x∈A A The inclusions in (62) are optimal in the sense that all possible combinations (strict inclusion and equality) can occur: • For results and examples illustrating equality in (62b) and equality in (62c), see Proposi- tion 6.5 and Example 6.6 below. • For an example illustrating equality in (62b) and strict inequality in (62c), see Example 6.7 below. • For an example illustrating strict inequality in (62b) and equality in (62c), see Example 6.10 below. • For examples illustrating strict inequality in (62b) and strict inequality in (62c), see Exam- ple 6.8 and Example 6.9 below. The remainder of this section is devoted to providing these examples. Proposition 6.4 Let A and B be nonempty subsets of X. Let A be closed a ∈ A. Assume that one of the following holds: (i) P−1 A (x) − x is a cone for every x ∈ A sufficiently close to a. (ii) A is convex. 23 Then (62b) holds with equality, i.e., NB A(a) = lim x→a x∈A(cid:0) cone(B − x) ∩ Nprox A (x)(cid:1) Proof. Indeed, if x ∈ A is sufficiently close to a, then Proposition 6.1 implies that bNB x) ∩ Nprox (x). Now take the outer limit as x → a in A. A A(x) = cone(B− (cid:4) Proposition 6.5 Let A be a nonempty closed convex subset of X, let B be a nonempty subset of X, and let a ∈ A. Assume that x 7→ cone(B − x) is outer semicontinuous at a relative to A, i.e., (63) cone(B − x) = cone(B − a), lim x→a x∈A Then (62) holds with equalities, i.e., (64) NB A(a) = lim x→a x∈A(cid:0) cone(B − x) ∩ Nprox A (x)(cid:1) =(cid:0) lim x→a x∈A cone(B − x)(cid:1) ∩ NA(a). Proof. The convexity of A and Lemma 3.4(vii) yield (65) cone(B − a) ∩ NA(a) = cone(B − a) ∩ Nprox A (a). On the other hand, Proposition 6.1(ii) and Lemma 3.4(iv) imply (66) cone(B − a) ∩ Nprox A Altogether, cone(B − a) ∩ NA(a) ⊆ NB (cid:0) lim (67) x→a x∈A (a) = bNB A(a). In view of (63), A(a) ⊆ NB A(a). cone(B − x)(cid:1) ∩ NA(a) ⊆ NB A(a). Recalling (62), we therefore obtain (64). (cid:4) Example 6.6 Let A be a linear subspace of X, set B := A, and a := (0, 0). Then NB (26d), NA(a) = A⊥, and cone(B− x) = A, for every x ∈ A. Hence (lim x→a x∈A {0} and (62) holds with equalities. A(a) = {0} by cone(B− x))∩ NA(a) = In Proposition 6.5, the convexity and the outer semicontinuity assumptions are both essential in the sense that absence of either assumption may make the inclusion (62c) strict; we shall illustrate this in the next three examples. Example 6.7 Suppose that X = R2, and let A := epi( · ), B := R × {0}, and a := (0, 0). If x = (x1, x2) ∈ A r {a}, then x2 > 0, B − x = R × {−x2}, and so cone(B − x) = R × R−− ∪ {(0, 0)}. Hence (68) cone(B − x) = R × R− 6= R × {0} = cone(B − a), lim x→a x∈A 24 i.e., (63) fails. Since A is closed and convex, Lemma 3.4(vii) implies that NA(a) = Nconv Thus A (a) = −A. (69) (70) Proposition 6.4(ii) yields equality in (62b), i.e., (cid:0) lim x→a x∈A cone(B − x)(cid:1) ∩ NA(a) = −A. (x)(cid:1). x∈A(cid:0) cone(B − x) ∩ Nprox A NB A(a) = lim x→a Already in Example 5.1 did we observe that (71) Therefore we have NB A(a) = cone{(1, −1), (−1, −1)}. (72) NB A(a) = lim x→a x∈A(cid:0) cone(B − x) ∩ Nprox A (x)(cid:1) $(cid:0) lim x→a x∈A cone(B − x)(cid:1) ∩ NA(a), i.e., the inclusion (62c) is strict. Example 6.8 Suppose that X = R2, and let A := cone{(1, 0), (0, 1)} = bdry R2 +, B := R × {1} ∪ {(1, 0), (−1, 0)}, and a := (0, 0). Clearly, A is not convex. If x = (x1, x2) ∈ A is sufficiently close to a, we have (73) This yields (74) cone(B − x) =(R × R+, R × R++ ∪ cone{(1, −x2), (−1, −x2)}, if x1 ≥ 0; if x2 > 0. cone(B − x) = R × R+ = cone(B − a), lim x→a x∈A i.e., (63) holds. Next, if x = (x1, x2) ∈ A, then P−1 A (x) = (x) = cone(cid:0)P−1 (75) and so (76) It follows that (77) Nprox A {x1} × ]−∞, x1] , ]−∞, x2] × {x2}, R2 −, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0, A (x) − x(cid:1) = {0} × R, R × {0}, R2 −, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0. NA(a) = lim x→a x∈A Nprox A (x) = R2 − ∪(cid:0){0} × R(cid:1) ∪(cid:0)R × {0}(cid:1). 25 If x ∈ A is sufficiently close a, then (78) It follows that (79) A(x) =({(0, 0)}, bNB R− × {0}, if x 6= a; if x = a. NB A(a) = R− × {0}. Combining (73) and (76), we obtain for every x = (x1, x2) ∈ A sufficiently close to a that A cone(B − x) ∩ Nprox {0} × R+, {(0, 0)}, R− × {0}, (x) = (x)(cid:1) =(cid:0){0} × R+(cid:1) ∪(cid:0)R− × {0}(cid:1). x∈A(cid:0) cone(B − x) ∩ Nprox if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0. lim x→a A (80) Thus (81) (83) (84) and (85) Using (79), (81), (74), and (77), we conclude that (82a) (82b) (82c) NB A(a) = R− × {0} $(cid:0){0} × R+(cid:1) ∪(cid:0)R− × {0}(cid:1) = lim $(cid:0){0} × R+(cid:1) ∪(cid:0)R × {0}(cid:1) =(cid:16) lim a′∈A(cid:0) cone(B − x) ∩ Nprox (x)(cid:1) cone(B − x)(cid:17) ∩ NA(a). x→a x∈A a′→a A Therefore, both inclusions in (62) are strict; however, A is not convex while (63) does hold. Example 6.9 Suppose that X = R2, let A := cone{(1, 0), (0, 1)} = bdry R2 a := (0, 0). Let x = (x1, x2) ∈ A. Then (see Example 6.8) +, B := R+(2, 1) and P−1 A (x) − x = (x) = Nprox A {0} × ]−∞, x1] , ]−∞, x2] × {0}, R2 −, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0, {0} × R, R × {0}, R2 −, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0, NA(a) = lim x→a x∈A Nprox A (x) = R2 − ∪(cid:0){0} × R(cid:1) ∪(cid:0)R × {0}(cid:1). 26 A(x) = cone(cid:0)(P−1 bNB A (x) − x) ∩ (B − x)(cid:1) =({0} × R+, {(0, 0)}, if x1 > 0 and x2 = 0; if x1 = 0 and x2 ≥ 0. Thus (86) Hence (87) (88) NB A(a) = lim x→a A(x) = {0} × R+. x∈A bNB (cid:8)(y1, y2)(cid:12)(cid:12) y2 ≥ 0, y1 < 2y2(cid:9) ∪ {(0, 0)}, (cid:8)(y1, y2)(cid:12)(cid:12) y1 ≥ 0, 2y2 < y1(cid:9) ∪ {(0, 0)}, B, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0. On the other hand, cone(B − x) = Combining (84) and (88), we deduce that (89) cone(B − x) ∩ Nprox A Using (88) and (89), we compute (x) = {0} × R+, R+ × {0}, {(0, 0)}, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0. (90) and (91) lim x→a x∈A cone(B − x) =(cid:8)(y1, y2)(cid:12)(cid:12) y1 ≥ 0 or y2 ≥ 0(cid:9) = X r R2 x∈A(cid:0) cone(B − x) ∩ Nprox (x)(cid:1) =(cid:0){0} × R+(cid:1) ∪(cid:0)R+ × {0}(cid:1) = cone{(0, 1), (1, 0)}. −− 6= B = cone(B − a) lim x→a A Using (87), (91), (90), and (85), we conclude that (92a) (92b) (92c) NB A(a) = {0} × R+ $(cid:0){0} × R+(cid:1) ∪(cid:0)R+ × {0}(cid:1) = lim $(cid:0){0} × R(cid:1) ∪(cid:0)R × {0}(cid:1) =(cid:0) lim x∈A(cid:0) cone(B − x) ∩ Nprox cone(B − x)(cid:1) ∩ NA(a). x→a x∈A x→a A (x)(cid:1) Therefore, both inclusions in (62) are strict; however, A is not convex and (63) does not hold (see (90)). Finally, we provide an example where the inclusion (62b) is strict while the inclusion (62c) is an equality. 27 Example 6.10 Suppose that X = R2, let A := cone{(1, 0), (0, 1)}, B := (cid:8)(y1, y2)(cid:12)(cid:12) y1 + y2 = 1(cid:9), and a := (0, 0). Let x = (x1, x2) ∈ A be sufficiently close to a. We compute (93a) cone(B − x) =(cid:8)(y1, y2)(cid:12)(cid:12) y1 + y2 > 0(cid:9) ∪ {(0, 0)}, if x1 > 0 and x2 = 0; if x1 = 0 and x2 > 0; if x1 = x2 = 0, (x) = {0} × R, R × {0}, R2 −, A(x) = {(0, 0)}. Nprox A bNB (93b) (93c) (94a) (94b) (94c) Furthermore, Example 6.8 (see (77)) implies that NA(a) = R2 deduce that − ∪(cid:0){0} × R(cid:1) ∪(cid:0)R × {0}(cid:1). We thus NB A(a) = {(0, 0)} $(cid:0){0} × R+(cid:1) ∪(cid:0)R+ × {0}(cid:1) = lim =(cid:0){0} × R+(cid:1) ∪(cid:0)R+ × {0}(cid:1) =(cid:0) lim x∈A(cid:0) cone(B − x) ∩ Nprox (x)(cid:1) cone(B − x)(cid:1) ∩ NA(a). x→a x∈A x→a A Therefore, the inclusion (62b) is strict while the inclusion (62c) is an equality. 7 Constraint qualification conditions and numbers Utilizing restricted normal cones, we introduce in this section the notions of CQ-number, joint-CQ- number, CQ condition, and joint-CQ condition, where CQ stands for "constraint qualification". CQ and joint-CQ numbers (95) θδ := θδ(cid:0)A, eA, B,eB(cid:1) := sup(cid:26)hu, vi(cid:12)(cid:12)(cid:12)(cid:12) Definition 7.1 (CQ-number) Let A, eA, B, eB, be nonempty subsets of X, let c ∈ X, and let δ ∈ R++. The CQ-number at c associated with (A, eA, B,eB) and δ is u ∈ bNeB The limiting CQ-number at c associated with (A, eA, B,eB) is θ := θ(cid:0)A, eA, B,eB(cid:1) := lim A(a), v ∈ −bNeA θδ(cid:0)A, eA, B,eB(cid:1). ka − ck ≤ δ, kb − ck ≤ δ. B (b), kuk ≤ 1, kvk ≤ 1, (cid:27). (96) δ↓0 Clearly, (97) θδ(cid:0)A, eA, B,eB(cid:1) = θδ(cid:0)B,eB, A, eA(cid:1) and θ(cid:0)A, eA, B,eB(cid:1) = θ(cid:0)B,eB, A, eA(cid:1). 28 Note that, δ 7→ θδ is increasing; this makes θ well defined. Furthermore, since 0 belongs to nonempty B-restricted proximal normal cones and because of the Cauchy-Schwarz inequality, we have (98) c ∈ A ∩ B and 0 < δ1 < δ2 ⇒ 0 ≤ θ ≤ θδ1 ≤ θδ2 ≤ 1, while θδ, and hence θ, is equal to −∞ if c /∈ A ∩ B and δ is sufficiently small (using the fact that sup ∅ = −∞). Using Proposition 3.7(ii)&(vi), we see that (99) and, for every x ∈ X, (100) eA ⊆ A′ and eB ⊆ B′ ⇒ θδ(A, eA, B,eB) ≤ θδ(A, A′, B, B′) θδ(cid:0)A, eA, B,eB(cid:1) at c = θδ(cid:0)A − x, eA − x, B − x,eB − x(cid:1) at c − x. To deal with unions, it is convenient to extend this notion as follows. Definition 7.2 (joint-CQ-number) Let A := (Ai)i∈I, eA := (eAi)i∈I, B := (Bj)j∈J, eB := (eBj)j∈J be nontrivial collections1 of nonempty subsets of X, let c ∈ X, and let δ ∈ R++. The joint-CQ-number at c associated with (A, eA, B, eB) and δ is (101) θδ(cid:0)Ai, eAi, Bj,eBj(cid:1), and the limiting joint-CQ-number at c associated with (A, eA, B, eB) is θδ(cid:0)A, eA, B, eB(cid:1). θδ = θδ(cid:0)A, eA, B, eB(cid:1) := sup θ = θ(cid:0)A, eA, B, eB(cid:1) := lim (i,j)∈I×J (102) δ↓0 For convenience, we will simply write θδ, θ and omit the possible arguments (A, eA, B,eB) and (A, eA, B, eB) when there is no cause for confusion. If I and J are singletons, then the notions of CQ-number and joint-CQ-number coincide. Also observe that (103) Bj ⇒ (∀δ ∈ R++) 0 ≤ θ ≤ θδ ≤ 1 c ∈[i∈I Ai ∩[j∈J while θ = θδ = −∞ when δ > 0 is sufficiently small and c does not belong to both Si∈I Ai andSj∈J Bj. Furthermore, the joint-CQ-number (and hence the limiting joint-CQ-number as well) really depends only on those sets Ai and Bj for which c ∈ Ai ∩ Bj. To illustrate this notion, let us compute the CQ-number of two lines. The formula provided is the cosine of the angle between the two lines -- as we shall see in Theorem 8.12 below, this happens actually for all linear subspaces although then the angle must be defined appropriately and the proof is more involved. 1The collection (Ai)i∈I is said to be nontrivial if I 6= ∅. 29 Proposition 7.3 (CQ-number of two distinct lines through the origin) Suppose that wa and wb are two vectors in X such that kwak = kwbk = 1. Let A := Rwa, B := Rwb, and δ ∈ R++. Assume that A ∩ B = {0}. Then the CQ-number at 0 is (104) θδ(A, A, B, B) = hwa, wbi . Proof. Set s := hwa, wbi. Assume first that s 6= 0. Let a = αwa ∈ A and b = βwb ∈ B. Then P−1 A (a) − a = NA(a) = {wa}⊥; A (a) − a) ∩ (B − a) = βwb − αwa and Similarly, considering (B − a) ∩ {wa}⊥ leads to βs = α. Hence (P−1 (105) A(a) = cone(cid:0)αs−1wb − αwa). bNB B (b) = cone(cid:0)βwb − βs−1wa). −bN A A(a) and v := βwb − βs−1wa ∈ −bN A Now set u := αs−1wb − αwa ∈ bNB , kvk = β√1 − s2 kuk = α√1 − s2 , and hu, vi = (106) (107) s s B (b). One computes αβ(1 − s2) s . Hence (108) hu, vi kuk · kvk = sgn(α) sgn(β)s. Choosing α and β in {−1, 1} appropriately, we arrange for hu, vi /(kuk · kvk) = s, as claimed. Now assume that s = 0. Arguing similarly, we see that (109) (∀a ∈ A) if a 6= 0; if a = 0, and (∀b ∈ B) A(a) =({0}, bNB B, B (b) =({0}, bN A A, if b 6= 0; if b = 0. (cid:4) This leads to θδ(A, A, B, B) = 0 = s, again as claimed. Let A := (Ai)i∈I, eA := (eAi)i∈I, B := (Bj)j∈J and eB := (eBj)j∈J be nontrivial collections of nonempty closed subsets of X and let δ ∈ R++. Set A := Si∈I Ai, eA := Si∈I eAi, B := Sj∈J Bj, eB := Sj∈JeBj, and suppose that c ∈ A ∩ B. It is interesting to compare the joint-CQ-number of collections, i.e., θδ(cid:0)A, eA, B, eB(cid:1), to the CQ-number of the unions, i.e., θδ(cid:0)A, eA, B,eB(cid:1). We shall see in the following two examples that neither of them is smaller than the other; in fact, one of them can be equal to 1 while the other is strictly less than 1. Example 7.4 (joint-CQ-number < CQ-number of the unions) Suppose that X = R3, let I := J := {1, 2}, A1 := R(0, 1, 0), A2 := R(2, 0, −1), B1 := R(0, 1, 1), B2 := R(1, 0, 0), c := (0, 0, 0), and let δ > 0. Furthermore, set A := (Ai)i∈I, B := (Bj)j∈J, A := A1 ∪ A2, and B := B1 ∪ B2. Then (110) θδ(cid:0)A, A, B, B(cid:1) = 2√5 < 1 = θδ(cid:0)A, A, B, B(cid:1). 30 Proof. Using Proposition 7.3, we compute, for the reference point c, θδ(A1, A1, B2, B2) = h(0, 1, 0), (1, 0, 0)i = 0, θδ(A1, A1, B1, B1) =(cid:12)(cid:12)(cid:10)(0, 1, 0), 1√2 θδ(A2, A2, B1, B1) =(cid:12)(cid:12)(cid:10) 1√5 θδ(A2, A2, B2, B2) =(cid:12)(cid:12)(cid:10) 1√5 (0, 1, 1)(cid:11)(cid:12)(cid:12) = 1√2 (0, 1, 1)(cid:11)(cid:12)(cid:12) = 1√10 (2, 0, −1), (1, 0, 0(cid:11)(cid:12)(cid:12) = 2√5 (2, 0, −1), 1√2 , . , Hence θδ(A, A, B, B) = max(i,j)∈I×J θδ(Ai, Ai, Bj, Bj) = 2√5 < 1. To estimate the CQ-number of the union, set (112) a := (0, δ, 0) ∈ A1 ⊆ A and b := (δ, 0, 0) ∈ B2 ⊆ B. Note that ka − ck = kak = δ and kb − ck = kbk = δ. Now define (113) ea := (δ, 0, −δ/2) ∈ A2 ⊆ A and eb := (0, δ, δ) ∈ B1 ⊆ B. Since kea − PB2eak < kea − PB1eak and PB2ea = b, we have b = PBea. Since keb − PA1ebk < keb − PA2ebk and PA1eb = a, we have a = PAeb. Therefore,eb ∈ B ∩ P−1 A (a) andea ∈ A ∩ P−1 B (b). It follows that A(a), (114a) u := 1 v := 2 δ (eb − a) = (0, 0, 1) ∈ bNB δ (b −ea) = (0, 0, 1) ∈ −bN A B (b). (111a) (111b) (111c) (111d) (114b) Since kuk = kvk = 1, we obtain 1 = hu, vi ≤ θδ(A, A, B, B) ≤ 1. Example 7.5 (CQ-number of the unions < joint-CQ-number) Suppose that X = R, let I := J := {1, 2}, A1 := B1 := R−, A2 := B2 := R+, c := 0, and δ > 0. Furthermore, set A := (Ai)i∈I, B := (Bj)j∈I, A := A1 ∪ A2 = R, and B := B1 ∪ B2 = R. Then (115) (cid:4) θδ(cid:0)A, A, B, B(cid:1) = 0 < 1 = θδ(cid:0)A, A, B, B(cid:1). (0) = R− and bN R+ R− Proof. Lemma 3.4(viii) implies that (∀x ∈ R) bNR On the other hand, bN therefore θδ(cid:0)A, A, B, B(cid:1) = 1 as well. R−R+ R (x) = {0}. Hence θδ(R, R, R, R) = 0 as claimed. (0) = R+. Hence θδ(R−, R−, R+, R+) = 1 and (cid:4) The two preceding examples illustrated the independence of the two types of CQ-numbers (for the collection and for the union). In some cases, such as Example 7.4, it is beneficial to work with a suitable partition to obtain a CQ-number that is less than one, which in turn is very desirable in applications (see Section 10). CQ and joint-CQ conditions Definition 7.6 (CQ and joint-CQ conditions) Let c ∈ X. 31 (116) (i) Let A, eA, B andeB be nonempty subsets of X. Then the (A, eA, B,eB)-CQ condition holds at c if (ii) Let A := (Ai)i∈I, eA := (eAi)i∈I, B := (Bj)j∈J and eB := (eBj)j∈J be nontrivial collections of nonempty subsets of X. Then the (A, eA, B, eB)-joint-CQ condition holds at c if for every (i, j) ∈ I × J, the (Ai, eAi, Bj,eBj)-CQ condition holds at c, i.e., A(c) ∩(cid:0) − NeA NeB B (c)(cid:1) ⊆ {0}. (117) (cid:0)∀(i, j) ∈ I × J(cid:1) NeBj Ai (c) ∩(cid:0) − NeAi Bj (c)(cid:1) ⊆ {0}. In view of the definitions, the key case to consider is when c ∈ A ∩ B (or when c ∈ Ai ∩ Bj in the joint-CQ case). The CQ-number is based on the behavior of the restricted proximal normal cone in a neighborhood of the point under consideration -- a related notion is that of the exact CQ-number, where we consider the restricted normal cone at the point instead of nearby restricted proximal normal cones. Definition 7.7 (exact CQ-number and exact joint-CQ-number) Let c ∈ X. (118) (i) Let A, eA, B and eB be nonempty subsets of X. The exact CQ-number at c associated with (A, eA, B,eB) is 2 B (c), kuk ≤ 1, kvk ≤ 1(cid:27). (ii) Let A := (Ai)i∈I, eA := (eAi)i∈I, B := (Bj)j∈J and eB := (eBj)j∈J be nontrivial collections of nonempty subsets of X. The exact joint-CQ-number at c associated with (A, B, eA, eB) is α := α(cid:0)A, eA, B,eB(cid:1) := sup(cid:26)hu, vi(cid:12)(cid:12)(cid:12)(cid:12) u ∈ NeB A(c), v ∈ −NeA (119) α := α(A, eA, B, eB) := sup (i,j)∈I×J α(Ai, eAi, Bj,eBj). The next result relates the various condition numbers defined above. Theorem 7.8 Let A := (Ai)i∈I, eA := (eAi)i∈I, B := (Bj)j∈J and eB := (eBj)j∈J be nontrivial collections of nonempty subsets of X. Set A := Si∈I Ai and B := Sj∈J Bj, and suppose that c ∈ A ∩ B. Denote the exact joint-CQ-number at c associated with (A, eA, B, eB) by α (see (119)), the joint-CQ-number at c associated with (A, eA, B, eB) and δ > 0 by θδ (see (101)), and the limiting joint-CQ-number at c associated with (A, eA, B, eB) by θ (see (102)). Then the following hold: (i) If α < 1, then the (A, eA, B, eB)-CQ condition holds at c. (ii) α ≤ θδ. 2Note that if c /∈ A ∩ B, then α = sup ∅ = −∞. 32 (iii) α ≤ θ. Now assume in addition that I and J are finite. Then the following hold: (iv) α = θ. Ai Proof. (c) ∩ (−NeAi (i): Suppose that α < 1. The condition for equality in the Cauchy-Schwarz inequality (c)) is either empty or {0}. In (v) The (A, eA, B, eB)-joint-CQ condition holds at c if and only if α = θ < 1. implies that for all (i, j) ∈ I × J, the intersection NeBj view of Definition 7.6, we see that the (A, eA, B, eB)-joint-CQ holds at c. (ii): Let (i, j) ∈ I × J. Take u ∈ NeBj (c) such that kuk ≤ 1 and kvk ≤ 1. Then, by definition of the restricted normal cone, there exist sequences (an)n∈N in Ai, (bn)n∈N in Bj, (un)n∈N and (vn)n∈N in X such that an → c, bn → c, un → u, vn → v, and (∀n ∈ N) un ∈ bNeBj (an) and vn ∈ −bNeAi hun, vni ≤ θδ(Ai, eAi, Bj,eBj). Taking the limit as n → +∞, we obtain hu, vi ≤ θδ(Ai, eAi, Bj,eBj) ≤ θδ. Now taking the supremum over suitable u and v, followed by taking the supremum over (i, j), we conclude that α ≤ θδ. (bn). Note that since δ > 0, eventually an and bn lie in ball(c; δ); consequently, (c) and v ∈ −NeAi Ai Ai Bj Bj Bj (iii): This is clear from (ii) and (102). (iv): Let (δn)n∈N be a sequence in R++ such that δn → 0. Then for every n ∈ N, there exist (120) such that (121) in ∈ I, jn ∈ J, an ∈ Ain, bn ∈ Bjn , un ∈ bNeBjn Ain (an), vn ∈ −bNeAin Bjn (bn) kan − ck ≤ δn, kbn − ck ≤ δn, kunk ≤ 1, kvnk ≤ 1, and hun, vni > θδn − δn. Since I and J are finite, and after passing to a subsequence and relabeling if necessary, we can and do assume that there exists (i, j) ∈ I × J such that un → u ∈ NeBj (c). Hence θ ← θδn − δn < hun, vni → hu, vi ≤ α. Hence θ ≤ α. On the other hand, α ≤ θ by (iii). Altogether, α = θ. (c) and vn → v ∈ −NeAi Ai Bj (v): "⇒": Let (i, j) ∈ I × J. c ∈ Ai ∩ Bj. Since the (A, eA, B, eB)-joint-CQ condition holds, we have NeBj By Cauchy-Schwarz, If c 6∈ Ai ∩ Bj, then α(Ai, eAi, Bj,eBj) = −∞. Now assume that (c) = {0}. (c) ∩ −NeAi Ai Bj (122) α(Ai, eAi, Bj,eBj) = sup(cid:26)hu, vi(cid:12)(cid:12)(cid:12)(cid:12) u ∈ NeBj Ai Since I and J are finite and because of (iv), we deduce that θ = α < 1. "⇐": Combine (i) with (iv). (c), v ∈ −NeAi Bj (c), kuk ≤ 1, kvk ≤ 1(cid:27) < 1. 33 (cid:4) 8 CQ conditions and CQ numbers: examples In this section, we provide further results and examples illustrating CQ conditions and CQ num- bers. First, let us note that the assumption that the sets of indices be finite in Theorem 7.8(iv) is essential: Example 8.1 (α < θ) Suppose that X = R2, let Γ ⊆ R++ be such that sup Γ = +∞, set (∀γ ∈ Γ) 2 γ · 2), B := R+ × R, A := (Aγ)γ∈Γ, eA := (X)γ∈Γ, B := (B), eB := (X), and Aγ := epi( 1 c := (0, 0). Denote the exact joint-CQ-number at c associated with (A, eA, B, eB) by α (see (119)), the joint-CQ-number at c associated with (A, eA, B, eB) and δ > 0 by θδ (see (101)), and the limiting joint-CQ-number at c associated with (A, eA, B, eB) by θ (see (102)). Then Proof. Let γ ∈ Γ and pick x > 0 such that a := (x, 1 x > 0 and 2 γx2) ∈ Aγ satisfies kak = ka − ck = δ, i.e., α = 0 < 1 = θδ = θ. (123) (124) Hence (125) γ2x2 = 2(cid:16)q1 + γ2δ2 − 1(cid:17) → +∞ as γ → +∞ in Γ. γx → +∞, as γ → +∞ in Γ. Since Aγ is closed and convex, it follows from Lemma 3.4(vii) that (γx, −1) Aγ (a) = NX Aγ (a) = NAγ (a). u := (126) pγ2x2 + 1 ∈ R+(γx, −1) = Nconv Furthermore, v := (1, 0) ∈ −(R− × {0}) = −bNX in view of (125), Aγ (a) = bNX B (c) = −NX (127a) (127b) 1 ≥ θδ ≥ θδ(Aγ, X, B, X) ≥ hu, vi = → 1 as γ → +∞ in Γ. γx pγ2x2 + 1 B (c) = −NB(c), kuk = kvk = 1, and, Thus θδ = 1, which implies that θ = 1. Finally, NAγ (c) = ({0} × R−) ⊥ (R+ × {0}) = −NB(c), which shows that α = 0. (cid:4) For the eventual application of these results to the method of alternating projections, the condi- tion α = θ < 1 is critical to ensure linear convergence. The following example illustrates that the CQ-number can be interpreted as a quantification of the CQ condition. 34 Example 8.2 (CQ-number quantifies CQ condition) Let A and B be subsets of X, and suppose that c ∈ A ∩ B. Let L be an affine subspace of X containing A ∪ B. Then the following are equiva- lent: (i) NL A(c) ∩ (−NL B (c)) = {0}, i.e., the (A, L, B, L)-CQ condition holds at c (see (116)). (ii) NA(c) ∩ (−NB(c)) ∩ (L − c) = {0}. (iii) θ < 1, where θ is the limiting CQ-number at c associated with (A, L, B, L) (see (96)). Proof. The identity (26d) of Theorem 4.5 yields NL (L − c). Hence (128) NL A(c) ∩(cid:0) − NL B (c)(cid:1) = NA(c) ∩(cid:0) − NB(c)(cid:1) ∩ (L − c), A(c) = NA(c) ∩ (L − c) and NL B (c) = NB(c) ∩ and the equivalence of (i) and (ii) is now clear. Finally, Theorem 7.8(iv)&(v) yields the equivalence of (i) and (iii). (cid:4) either hold or fail: Depending on the choice of the restricting sets eA and eB, the (A, eA, B,eB)-CQ condition may Example 8.3 (CQ condition depends on restricting sets) Suppose that X = R2, and set A := epi( · ), B := R × {0}, and c := (0, 0). Then we readily verify that NA(c) = NX A (c) = −A, A(c) = − bdry A, NB(c) = NX NB B (c) = {0} × R+. Consequently, B (c)(cid:1) = {(0, 0)}. A (c) ∩(cid:0) − NX (129) B (c) = {0} × R, and N A B (c)(cid:1) = {0} × R− while NB Therefore, the (A, A, B, B)-CQ condition holds, yet the (A, X, B, X)-CQ condition fails. A(c) ∩(cid:0) − N A NX For two spheres, it is possible to quantify the convergence of θδ to δ = α: Proposition 8.4 (CQ-numbers of two spheres) Let z1 and z2 be in X, let ρ1 and ρ2 be in R++, set S1 := sphere(z1; ρ1) and S2 := sphere(z2; ρ2) and assume that c ∈ S1 ∩ S2. Denote the limiting CQ-number at c associated with (S1, X, S2, X) by θ (see Definition 7.1), and the exact CQ-number at c associated with (S1, X, S2, X) by α (see Definition 7.7). Then the following hold: (i) θ = α = hz1 − c, z2 − ci ρ1ρ2 . (ii) α < 1 unless the spheres are identical or intersect only at c. Now assume that α < 1, let ε ∈ R++, and set δ := (p(ρ1 + ρ2)2 + 4ρ1ρ2ε − (ρ1 + ρ2))/2 > 0. Then (130) α ≤ θδ ≤ α + ε, where θδ is the CQ-number at c associated with (S1, X, S2, X) (see Definition 7.1). 35 Proof. (i): This follows from Theorem 7.8(iv) and Example 3.6. (ii): Combine (i) with the characterization of equality in the Cauchy-Schwarz inequality. Let us now establish (130). By Theorem 7.8(ii), we have α ≤ θδ. Let s1 ∈ S1 be such that (s1) be such that ku1k = 1, let s2 ∈ S2 be such that ks2 − ck ≤ δ, and let S1 (s2) be such that ku2k = 1. By Example 3.6, ks1 − ck ≤ δ, let u1 ∈ bNX u2 ∈ bNX S2 (131) u1 = ± s1 − z1 ks1 − z1k = ± s1 − z1 ρ1 and u2 = ± s2 − z2 ks2 − z2k = ± s2 − z2 ρ2 . Hence (132a) (132b) (132c) (132d) (132e) ρ1ρ2 hu1, u2i ≤ hs1 − z1, s2 − z2i = h(s1 − c) + (c − z1), (s2 − c) + (c − z2)i ≤ hs1 − c, s2 − ci + hs1 − c, c − z2i + hc − z1, s2 − ci + hc − z1, c − z2i ≤ δ2 + δ(ρ1 + ρ2) + ρ1ρ2α and thus, using the definition of δ, (133) hu1, u2i ≤ α + δ2 + δ(ρ1 + ρ2) ρ1ρ2 = α + ε. Therefore, by the definition of θδ, we have θδ ≤ α + ε. (cid:4) Two convex sets Let us turn to the classical convex setting. We start by noting that well known constraint qualifi- cations are conveniently characterized using our CQ conditions. Proposition 8.5 Let A and B be nonempty convex subsets of X such that A ∩ B 6= ∅, and set L = aff(A ∪ B). Then the following are equivalent: (i) ri A ∩ ri B 6= ∅. (ii) The (A, L, B, L)-CQ condition holds at some point in A ∩ B. (iii) The (A, L, B, L)-CQ condition holds at every point in A ∩ B. Proof. This is clear from Theorem 4.13. (cid:4) Proposition 8.6 Let A and B be nonempty convex subsets of X such that A ∩ B 6= ∅. Then the following are equivalent: 36 (i) 0 ∈ int(B − A). (ii) The (A, X, B, X)-CQ condition holds at some point in A ∩ B. (iii) The (A, X, B, X)-CQ condition holds at every point in A ∩ B. Proof. This is clear from Corollary 4.14. (cid:4) In stark contrast to Proposition 8.5 and 8.6, if the restricting sets are not both equal to L or to X, then the CQ-condition may actually depend on the reference point as we shall illustrate now: Example 8.7 (CQ condition depends on the reference point) Suppose that X = R2, and let f : R → R : x 7→ (max{0, x})2, which is a continuous convex function. Set A := epi f and B := R × {0}, which are closed convex subsets of X. Consider first the point c := (−1, 0) ∈ A ∩ B. Then NB A(c) = {(0, 0)} and N A B (c) = {0} × R+; hence, (134) i.e., the (A, A, B, B)-CQ condition holds at c. On the other hand, consider now d := (0, 0) ∈ A ∩ B. Then NB A(d) = {0} × R− and N A B (d) = {0} × R+; thus, NB (135) i.e., the (A, A, B, B)-CQ condition fails at d. NB A(c) ∩(cid:0) − N A A(d) ∩(cid:0) − N A B (c)(cid:1) = {(0, 0)}, B (d)(cid:1) = {0} × R−, Two linear (or intersecting affine) subspaces We specialize further to two linear subspaces of X. A pleasing connection between CQ-number and the angle between two linear subspaces will be revealed. But first we provide some auxiliary results. Proposition 8.8 Let A and B be linear subspaces of X, and let δ ∈ R++. Then (136) A(a) = [ a∈A∩(B+A⊥)∩ball(0;δ) bNB [a∈A∩ball(0;δ)bNB A(a) = [a∈A bNB A(a) = A⊥ ∩ (A + B). A (a) = a + A⊥ and hence P−1 Proof. Let a ∈ A. Then P−1 A (a) − a = A⊥. If B ∩ (a + A⊥) = ∅, then bNB A(a) = {0}. Thus we assume that B ∩ (a + A⊥) 6= ∅, which is equivalent to a ∈ A ∩ (B + A⊥). Next, by Lemma 3.4(ii), bNB A(a) = A⊥ ∩ cone(B − a). This implies (∀λ ∈ R++) cone(B − λa) = cone(λ(B − a)) = cone(B − a). Thus, A(λa) = A⊥ ∩ cone(B − λa) = A⊥ ∩ cone(B − a) = bNB bNB (137) This establishes not only the first two equalities in (136) but also the third because (∀λ ∈ R++) A(a). (138a) [a∈A bNB A(a) = [a∈A(cid:0)A⊥ ∩ cone(B − a)(cid:1) = A⊥ ∩ [a∈A cone(B − a) 37 (138b) (138c) = A⊥ ∩ cone(cid:16) [a∈A = A⊥ ∩ (B + A). (B − a)(cid:17) = A⊥ ∩ cone(B − A) = A⊥ ∩ (B − A) The proof is complete. (cid:4) We now introduce two notions of angles between subspaces; for further information, we highly recommend [10] and [11]. Definition 8.9 Let A and B be linear subspaces of X. (i) (Dixmier angle) [15] The Dixmier angle between A and B is the number in [0, π 2 ] whose cosine is (ii) (Friedrichs angle) [16] The Friedrichs angle (or simply the angle) between A and B is the number c0(A, B) := sup(cid:8) ha, bi (cid:12)(cid:12) a ∈ A, b ∈ B, kak ≤ 1, kbk ≤ 1(cid:9). in [0, π 2 ] whose cosine is given by given by (139) (140a) (140b) c(A, B) := c0(A ∩ (A ∩ B)⊥, B ∩ (A ∩ B)⊥) = sup(cid:26) ha, bi (cid:12)(cid:12)(cid:12)(cid:12) a ∈ A ∩ (A ∩ B)⊥, kak ≤ 1, b ∈ B ∩ (A ∩ B)⊥, kbk ≤ 1 (cid:27). Let us gather some properties of angles. Fact 8.10 Let A and B be linear subspaces of X. Then the following hold: (i) If A ∩ B = {0}, then c(A, B) = c0(A, B). (ii) If A ∩ B 6= {0}, then c0(A, B) = 1. (iii) c(A, B) < 1. (iv) c(A, B) = c0(A, B ∩ (A ∩ B)⊥) = c0(A ∩ (A ∩ B)⊥, B). (v) (Solmon) c(A, B) = c(A⊥, B⊥). Proof. (i) -- (iii): Clear from the definitions. (iv): See, e.g., [10, Lemma 2.10(1)] or [11, Lemma 9.5]. (v): See, e.g., [10, Theorem 2.16]. (cid:4) Proposition 8.11 (CQ-number of two linear subspaces and Dixmier angle) Let A and B be linear subspaces of X, and let δ > 0. Then (141a) (141b) (141c) θδ(A, A, B, B) = c0(cid:0)A⊥ ∩ (A + B), B⊥ ∩ (A + B)(cid:1), θδ(A, X, B, B) = c0(cid:0)A⊥ ∩ (A + B), B⊥(cid:1), θδ(A, A, B, X) = c0(cid:0)A⊥, B⊥ ∩ (A + B)(cid:1), where the CQ-numbers at 0 are defined as in (95). 38 Proof. This follows from Proposition 8.8. (cid:4) We are now in a position to derive a striking connection between the CQ-number and the Friedrichs angle, which underlines a possible interpretation of the CQ-number as a generalized Friedrichs angle between two sets. Theorem 8.12 (CQ-number of two linear subspaces and Friedrichs angle) Let A and B be linear subspaces of X, and let δ > 0. Then (142) θδ(A, A, B, B) = θδ(A, X, B, B) = θδ(A, A, B, X) = c(A, B) < 1, where the CQ-number at 0 is defined as in (95). Proof. On the one hand, using Fact 8.10(v), we have On the other hand, Fact 8.10(iv) yields (143a) (143b) (143c) (144a) (144b) (144c) (144d) c(A, B) = c(A⊥, B⊥) = c0(cid:0)A⊥ ∩ (A⊥ ∩ B⊥)⊥, B⊥ ∩ (A⊥ ∩ B⊥)⊥(cid:1) = c0(cid:0)A⊥ ∩ (A + B), B⊥ ∩ (A + B)(cid:1). c0(cid:0)A⊥ ∩ (A + B), B⊥(cid:1) = c0(cid:0)A⊥ ∩ (A⊥ ∩ B⊥)⊥, B⊥(cid:1) = c0(cid:0)A⊥, B⊥ ∩ (A⊥ ∩ B⊥)⊥(cid:1) = c0(cid:0)A⊥, B⊥ ∩ (A + B)(cid:1). = c(A⊥, B⊥) Altogether, recalling Proposition 8.11, we obtain the result. (cid:4) The results in this subsection have a simple generalization to intersecting affine subspaces. In- deed, if A and B are intersecting affine subspaces, then the corresponding Friedrichs angle is (145) c(A, B) := c(par A, par B). Combining (100) with Theorem 8.12, we immediately obtain the following result. Corollary 8.13 (CQ-number of two intersecting affine subspaces and Friedrichs angle) Let A and B be affine subspaces of X, suppose that c ∈ A ∩ B, and let δ > 0. Then (146) θδ(A, A, B, B) = θδ(A, X, B, B) = θδ(A, A, B, X) = c(A, B) < 1, where the CQ-number at c is defined as in (95). 39 9 Regularities In this section, we study a notion of set regularity that is based on restricted normal cones. Definition 9.1 (regularity and superregularity) Let A and B be nonempty subsets of X, and let c ∈ X. (i) We say that B is (A, ε, δ)-regular at c ∈ X if ε ≥ 0, δ > 0, and ky − ck ≤ δ, kb − ck ≤ δ, (y, b) ∈ B × B, B (b)  ⇒ hu, y − bi ≤ εkuk · ky − bk. (147) u ∈ bN A If B is (X, ε, δ)-regular at c, then we also simply speak of (ε, δ)-regularity. (ii) The set B is called A-superregular at c ∈ X if for every ε > 0 there exists δ > 0 such that B is (A, ε, δ)-regular at c. Again, if B is X-superregular at c, then we also say that B is superregular at c. Remark 9.2 Several comments on Definition 9.1 are in order. (i) Superregularity with A = X was introduced by Lewis, Luke and Malick in [17, Section 4]. Among other things, they point out that amenability and prox regularity are sufficient con- ditions for superregularity, while Clarke regularity is a necessary condition. (ii) The reference point c does not have to belong to B. If c 6∈ B, then for every δ ∈ ]0, dB(c)[, B is (0, δ)-regular at c; consequently, B is superregular at c. (iii) If ε1 > ε2 and B is (A, ε2, δ)-regular at c then B is also (A, ε1, δ)-regular at c. (iv) If ε ∈ [1, +∞[, then Cauchy-Schwarz implies that B is (ε, +∞)-regular at every point in X. (v) It follows from Proposition 3.7(ii) that B is (A1 ∪ A2, ε, δ)-regular at c if and only if B is both (A1, ε, δ)-regular and (A2, ε, δ)-regular at c. (vi) If B is convex, then it follows with Lemma 3.4(vii) that B is (A, 0, +∞)-regular at c; conse- quently, B is superregular. (vii) Similarly, if B is locally convex at c, i.e., there exists ρ ∈ R++ such that B ∩ ball(c; ρ) is convex, then B is superregular at c. (viii) If B is (A, 0, δ)-regular at c, then B is A-superregular at c; the converse, however, is not true in general (see Example 9.3 below). As a first example, let us consider the sphere. Example 9.3 (sphere) Let z ∈ X and ρ ∈ R++. Set S := sphere(z; ρ), suppose that s ∈ S, let ε ∈ R++, and let δ ∈ R++. Then S is (ε, ρε)-regular at s; consequently, S is superregular at s (see Definition 9.1). However, S is not (0, δ)-regular at s. 40 Proof. Let b ∈ S and y ∈ S. Then ρ2 = kz − yk2 = kz − bk2 + ky − bk2 − 2hz − b, y − bi = ρ2 + ky − bk2 − 2hz − b, y − bi, which implies (148) 2hz − b, y − bi = ky − bk2. On the other hand, by Example 3.6, we have S (b) ∩ sphere(0; 1). Combining (148) and (149), we obtain (149) S (b) ∩ sphere(0; 1) =(cid:26) ± bNX S (b) ∩ sphere(0; 1), y − bE =(cid:26) ± DbNX Suppose that u ∈ bNX Thus if ky − sk ≤ ρε, kb − sk ≤ ρε, and u ∈ bNX (151) (150) 1 hu, y − bi ≤ z − b kz − bk(cid:27) =(cid:26) ± z − b ρ (cid:27). 2ρky − bk2(cid:27). 1 S (b) ∩ sphere(0; 1), then 2ρ(cid:0)ky − sk + ks − bk(cid:1)ky − bk ≤ ρε + ρε 2ρ ky − bk 1 2ρky − bk2 ≤ = εkuk · ky − bk, (152) (153) which verifies the (ε, ρε)-regularity of S at s. Finally, by (150), max(cid:8)hbNX S (b) ∩ sphere(0; 1), y − bi(cid:9) = 1 2ρky − bk2 > 0 (cid:4) and therefore S is not (0, δ)-regular at s. We now characterizes A-superregularity using restricted normal cones. Theorem 9.4 (characterization of A-superregularity) Let A and B be nonempty subsets of X, and let c ∈ X. Then B is A-superregular at c if and only if for every ε ∈ R++, there exists δ ∈ R++ such that (154) ky − ck ≤ δ, kb − ck ≤ δ (y, b) ∈ B × B u ∈ N A B (b)  ⇒ hu, y − bi ≤ εkuk · ky − bk. Proof. "⇐": Clear from Lemma 3.4(iv). "⇒": We argue by contradiction; thus, we assume there exists ε ∈ R++ and sequences (yn, bn, un)n∈N in B × B × X such that (yn, bn) → (c, c) and for every n ∈ N, (155) By the definition of the restricted normal cone, for every n ∈ N, there exists a sequence B (bn,k). Hence there exists a subsequence (kn)n∈N of (n)n∈N such that bn,kn → c and ε 2kun,knk · kyn − bn,knk. (156) (bn,k, un,k)k∈N in B × X such that limk∈N bn,k = bn, limk∈N un,k = un, and (∀k ∈ N) un,k ∈ bN A and hun, yn − bni > εkunk · kyn − bnk. hun,kn, yn − bn,kni > (∀n ∈ N) un ∈ N A B (bn) However, this contradicts the A-superregularity of B at c. (cid:4) When B = X, then Theorem 9.4 turns into [17, Proposition 4.4]: 41 Corollary 9.5 (Lewis-Luke-Malick) Let B be a nonempty subset of X and let c ∈ B. Then B is super- regular at c if and only if for every ε ∈ R++ there exists δ ∈ R++ such that (157) ky − ck ≤ δ, kb − ck ≤ δ (y, b) ∈ B × B u ∈ NB(b)  ⇒ hu, y − bi ≤ εkuk · ky − bk. We now introduce the notion of joint-regularity, which is tailored for collections of sets and which turns into Definition 9.1 when the index set is a singleton. Definition 9.6 (joint-regularity) Let A be a nonempty subset of X, let B := (Bj)j∈J be a nontrivial collection of nonempty subsets of X, and let c ∈ X. (i) We say that B is (A, ε, δ)-joint-regular at c if ε ≥ 0, δ > 0, and for every j ∈ J, Bj is (A, ε, δ)-regular at c. (ii) The collection B is A-joint-superregular at c if for every j ∈ J, Bj is A-superregular at c. As in Definition 9.1, we may omit the prefix A if A = X. Here are some verifiable conditions that guarantee joint-(super)regularity. Proposition 9.7 Let A := (Aj)j∈J and B := (Bj)j∈J be nontrivial collections of nonempty subsets of X, let c ∈ X, let (ε j)j∈J be a collection in R+, and let (δj)j∈J be a collection in ]0, +∞]. Set A := Tj∈J Aj, ε := supj∈J ε j, and δ := infj∈J δj. Then the following hold: (i) If δ > 0 and (∀j ∈ J) Bj is (Aj, ε j, δj)-regular at c, then B is (A, ε, δ)-joint-regular at c. (ii) If J is finite and (∀j ∈ J) Bj is (Aj, ε j, δj)-regular at c, then B is (A, ε, δ)-joint-regular at c. (iii) If J is finite and (∀j ∈ J) Bj is Aj-superregular at c, then B is A-joint-superregular at c. Proof. (i): Indeed, by Remark 9.2(v), Bj is (A, ε, δ)-regular at c for every j ∈ J. (ii): Since J is finite, we have δ > 0 and so the conclusion follows from (i). (iii): This follows from (ii) and the definitions. (cid:4) Corollary 9.8 (convexity and regularity) Let B := (Bj)j∈J be a nontrivial collection of nonempty con- vex subsets of X, let A ⊆ X, and let c ∈ X. Then B is (0, +∞)-joint-regular, (A, 0, +∞)-joint-regular, joint-superregular, and A-joint-superregular at c. Proof. By Remark 9.2(vi), Bj is (0, +∞)-regular, superregular, and A-superregular at c, for every j ∈ J. Now apply Proposition 9.7(i)&(iii). (cid:4) The following example illustrates the flexibility gained through the notion of joint-regularity. 42 Example 9.9 (two lines: joint-superregularity 6⇒ superregularity of the union) Suppose that d1 and d2 are in sphere(0; 1). Set B1 := Rd1, B2 := Rd2, and B := B1 ∪ B2, and assume that B1 ∩ B2 = {0}. By Corollary 9.8, (B1, B2) is joint-superregular at 0. Let δ ∈ R++, and set b := δd1 and y := δd2. Then ky − 0k = δ, kb − 0k = δ, and 0 < ky − bk = δkd2 − d1k. Using Proposition 3.3(iii), we see that NB(b) = {d1}⊥. Note that there exists v ∈ {d1}⊥ such that hv, d2i 6= 0 (for otherwise {d1}⊥ ⊆ {d2}⊥ ⇒ B2 ⊆ B1, which is absurd). Hence there exists u ∈ {d1}⊥ = {b}⊥ = NB(b) such that kuk = 1 and hu, d2i > 0. It follows that hu, y − bi = hu, yi = δ hu, d2i = hu, d2i kukky − bk/kd2 − d1k. Therefore, B is not superregular at 0. Let us provide an example of an A-superregular set that is not superregular. To do so, we require the following elementary result. Lemma 9.10 Consider in R2 the sets C := [(0, 1), (m, 1 + m2)] = (cid:8)(x, 1 + mx) (cid:12)(cid:12) x ∈ [0, m](cid:9) and D := [(m, 1), (m, 1 + m2)], where m ∈ R++. Let z ∈ R. Then PC∪D(z, 0) = (0, 1), {(0, 1), (m, 1)}, (m, 1), if z < m/2; if z = m/2; if z > m/2. (158) (159) (160) Proof. It is clear that PD(z, 0) = (m, 1). We assume that 0 < z < m for otherwise (158) is clearly true. We claim that PC(z, 0) = (0, 1). Indeed, f : x 7→ k(x, 1 + mx) − (z, 0)k2 is a convex quadratic with minimizer xz := (z − m)/(1 + m2). The requirement xz ≥ 0 from the definition of C forces z ≥ m, which is a contradiction. Hence PC(z, 0) is a subset of the relative boundary of C, i.e., of {(0, 1), (m, 1 + m2)}. Clearly, (0, 1) is the closer to (z, 0) than (m, 1 + m2). This verifies the claim. Since PC∪D(z, 0) is the subset of points in PC(z, 0) ∪ PD(z, 0) closest to (z, 0), the result follows. (cid:4) Example 9.11 (A-superregularity 6⇒ superregularity) Suppose that X = R2. As in [17, Exam- ple 4.6], we consider c := (0, 0) ∈ X and B := epi f , where f : R → ]−∞, +∞] : x 7→ 2k(x − 2k), 0, +∞, if 2k ≤ x < 2k+1 and k ∈ Z; if x = 0; if x < 0. Then B is not superregular at c; however, B is A-superregular at c, where A := R × {−1}. Proof. It is stated in [17, Example 4.6] that B is not superregular at c (and that B is Clarke regular at c). To tackle A-superregularity, let us determine PB(A). Let us consider the point a = (α, −1), where α ∈(cid:2)2−1, 1(cid:2). Then Lemma 9.10 (see also the picture below) implies that PB(α, −1) = ( 1 2 , 0), (cid:8)( 1 2 , 0), (1, 0)(cid:9), (1, 0), 2 ≤ α < 3 4 ; if 1 if α = 3 4 ; if 3 4 < α < 1; 43 ✻ B = epi f ✘✘✘✘✘✘ ✭✭✭ ✆✆ ❊❊ ❈❈ 1 4 1 2 ✭✭ 1 1 8 16 0 ✟✟✟✟✟✟✟✟✟✟✟ ✄✄ 1 ❈❈ ✆✆ ❈❈ ❊❊ ❈❈ ❈❈ ❊❊ ✆✆ ❊❊ ✆✆ −1 A = R × {−1} ❊❊ ✆✆ 3 8 s 1 4 s 1 2 and more generally, ✄✄ ✄✄ ✄✄ ✄✄ ❈❈ ❈❈ ✄✄ ✄✄ ❈❈ ✄✄ ❈❈ ✄✄ 3 4 s 1 (2k, 0), (cid:8)(2k, 0), (2k+1, 0)(cid:9), (2k+1, 0), Clearly, if a ∈ R− × {−1}, then PB(a) = (0, 0). Let b ∈ B. Then ✲ if 2k ≤ α < 2k + 2k−1; if α = 2k + 2k−1; if 2k + 2k−1 < α < 2k+1. (161) (162) Thus (163) A ∩ P−1 2k ≤ α < 2k+1 ⇒ PB(α, −1) = B (b) = B (b) = bN A R− × {−1}, ∅, {(0, 0)}, (cid:2)2k−2 + 2k−1, 2k−1 + 2k(cid:3) × {−1}, if b = (2k, 0) and k ∈ Z; if b = (0, 0); otherwise. cone(cid:16)(cid:2) − 2k−2, 2k−1(cid:3) × {−1}(cid:17), {(0, 0)} ∪(cid:0)R− × R−−(cid:1), if b = (2k, 0) and k ∈ Z; if b = (0, 0); otherwise. Let ε ∈ R++. Let K ∈ Z be such that 2K−1 ≤ ε, and let δ ∈(cid:3)0, 2K(cid:3). Furthermore, let y = (y1, y2) ∈ B, let b = (b1, b2) ∈ B, let u ∈ bN A B (b), and assume that ky − ck ≤ δ and that kb − ck ≤ δ. We consider three cases. 44 +; consequently, hu, y − bi = hu, yi ≤ 0 ≤ εkuk · ky − − and y ∈ R2 Case 1: b = (0, 0). Then u ∈ R2 Case 2: b /∈ ({0} ∪ 2Z) × {0}. Then bN A Case 3: b ∈ 2Z × {0}, say b = (2k, 0), where k ∈ Z. Since 2k = kb − 0k = kb − ck ≤ δ ≤ 2K, we have k ≤ K. Furthermore, y2 ≥ 0, max{y1 − b1, y2 − b2} ≤ ky − bk, and u = λ(t, −1) = (λt, −λ) where t ∈ [−2k−2, 2k−1] and λ ≥ 0. Hence λ ≤ kuk and (164a) B (b) = {(0, 0}; hence u = 0 and so hu, y − bi = 0 ≤ hu, y − bi = λt(y1 − b1) − λ(y2 − b2) = λt(y1 − b1) − λ(y2 − 0) ≤ λt(y1 − b1) ≤ λt · y1 − b ≤ kuk · 2k−1 · ky − bk ≤ 2K−1kuk · ky − bk ≤ ε · kuk · ky − bk. bk. εkuk · ky − bk. (164b) (164c) Therefore, in all three cases, we have shown that hu, y − bi ≤ εkuk · ky − bk. We now use Example 9.11 to construct an example complementary to Example 9.9. (cid:4) Example 9.12 (superregularity of the union 6⇒ joint-superregularity) Suppose that X = R2, set B1 := epi f , where f is as in Example 9.11, B2 := X r B1, and c := (0, 0). Since B1 ∪ B2 = X is convex, it is clear from Remark 9.2(vi) that B1 ∪ B2 is superregular at c. On the other hand, since B1 is not superregular at c (see Example 9.11), it is obvious that (B1, B2) is not joint-superregular at c. 10 The method of alternating projections (MAP) We now apply the machinery of restricted normal cones and associated results to derive linear convergence results. On the composition of two projection operators The method of alternating projections iterates projection operators. Thus, in the next few results, we focus on the outcome of a single iteration of the composition. Lemma 10.1 Let A and B be nonempty closed subsets of X. Then the following hold3: (i) PA(B r A) ⊆ bdryaff A∪B A ⊆ bdry A. (ii) PB(A r B) ⊆ bdryaff A∪B(B) ⊆ bdry B. 3We denote by bdryaff A∪B(S) the boundary of S ⊆ X with respect to aff(A ∪ B). 45 (iii) If b ∈ B and a ∈ PAb, then: (165) a ∈ (bdry A) r B ⇔ a ∈ A r B ⇒ b ∈ B r A ⇒ a ∈ bdry A. (iv) If a ∈ A and b ∈ PBa, then: (166) b ∈ (bdry B) r A ⇔ b ∈ B r A ⇒ a ∈ A r B ⇒ b ∈ bdry B. Proof. (i): Take b ∈ B r A and a ∈ PAb. Assume to the contrary that there exists δ ∈ R++ such d(a, b) = dA(b), which is absurd. that aff(A ∪ B) ∩ ball(a; δ) ⊆ A. Henceea := a + δ(b − a)/kb − ak ∈ A and thus dA(b) ≤ d(ea, b) < (ii): Interchange the roles of A and B in (i). (iii): If a ∈ (bdry A) r B, then clearly a ∈ A r B. Now assume that a ∈ A r B. If b ∈ A, then a ∈ PAb = {b} ⊆ B, which is absurd. Hence b ∈ B r A and thus (i) implies that a ∈ PA(B r A) ⊆ bdry A. (iv): Interchange the roles of A and B in (iii). (cid:4) Lemma 10.2 Let A and B be nonempty closed subsets of X, let c ∈ X, let y ∈ B, let a ∈ PAy, let b ∈ PBa, and let δ ∈ R+. Assume that dA(y) ≤ δ and that d(y, c) ≤ δ. Then the following hold: (i) d(a, c) ≤ 2δ. (ii) d(b, y) ≤ 2d(a, y) ≤ 2δ. (iii) d(b, c) ≤ 3δ. Proof. Since y ∈ B, we have (167) Thus, (168) d(a, b) = dB(a) ≤ d(a, y) = dA(y) ≤ δ. d(a, c) ≤ d(a, y) + d(y, c) ≤ δ + δ = 2δ, which establishes (i). Using (167), we also conclude that d(b, y) ≤ d(b, a) + d(a, y) ≤ 2d(a, y) ≤ 2δ; hence, (ii) holds. Finally, combining (167) and (168), we obtain (iii) via d(b, c) ≤ d(b, a) + d(a, c) ≤ δ + 2δ = 3δ. (cid:4) Corollary 10.3 Let A and B be nonempty closed subsets of X, let ρ ∈ R++, and suppose that c ∈ A ∩ B. Then (169) PAPBPA ball(c; ρ) ⊆ ball(c; 6ρ). 46 Proof. Let b−1 ∈ ball(c; ρ), a0 ∈ PAb−1, b0 ∈ PBa0, and a1 ∈ PAb0. We have d(a0, b−1) = dA(b−1) ≤ d(b−1, c) ≤ ρ, so dB(a0) ≤ d(a0, c) ≤ d(a0, b−1) + d(b−1, c) ≤ 2ρ. Applying Lemma 10.2(iii) to the sets B and A, the points a0, b0, a1, and δ = 2ρ, we deduce that d(a1, c) ≤ 3(2ρ) = 6ρ. (cid:4) The next two results are essential to guarantee a local contractive property of the composition. Proposition 10.4 (regularity and contractivity) Let A and B be nonempty closed subsets of X, let eA and eB be nonempty subsets of X, let c ∈ X, let ε ≥ 0, and let δ > 0. Assume that B is (eA, ε, 3δ)-regular at c (see Definition 9.1). Furthermore, assume that y ∈ B ∩eB, that a ∈ PA(y) ∩ eA, that b ∈ PB(a), that ky − ck ≤ δ, and that dA(y) ≤ δ. Then (170) ka − bk ≤ (θ3δ + 2ε)ka − yk, where θ3δ the CQ-number at c associated with (A, eA, B,eB) (see (95)). Proof. Lemma 10.2(i)&(iii) yields ka − ck ≤ 2δ and kb − ck ≤ 3δ. On the other hand, y − a ∈ bNeB and b − a ∈ −bNeA Since a − b ∈ bNeA B (b), ky − ck ≤ δ, and kb − ck ≤ 3δ, we obtain, using the (eA, ε, 3δ)-regularity of B, that ha − b, y − bi ≤ εka − bk · ky − bk. Moreover, Lemma 10.2(ii) states that ky − bk ≤ 2ka − yk. It follows that hb − a, y − ai ≤ θ3δkb − ak · ky − ak. B (b). Therefore, A(a) (171) (172) ha − b, y − bi ≤ 2εka − bk · ka − yk. Adding (171) and (172) yields ka − bk2 ≤ (θ3δ + 2ε)ka − bk · ka − yk. The result follows. (cid:4) We now provide a result for collections of sets similar to -- and relying upon -- Proposition 10.4. Proposition 10.5 (joint-regularity and contractivity) Let A := (Ai)i∈I and B := (Bj)j∈J be nontriv- ial collections of closed subsets of X, Assume that A := Si∈I Ai and B := Sj∈J Bj are closed, and that c ∈ A ∩ B. Let eA := (eAi)i∈I and eB := (eBj)j∈J be nontrivial collections of nonempty subsets of X such that (∀i ∈ I) PAi((bdry B) r A) ⊆ eAi and (∀j ∈ J) PBj((bdry A) r B) ⊆ eBj. Set eA := Si∈I eAi and eB :=Sj∈JeBj, let ε ≥ 0 and let δ > 0. (i) If b ∈ (bdry B) r A and a ∈ PA(b), then (∃ i ∈ I) a ∈ PAi(b) ⊆ Ai ∩ eAi. (ii) If a ∈ (bdry A) r B and b ∈ PB(a), then (∃ j ∈ J) b ∈ PBj(a) ⊆ Bj ∩eBj. (iii) If y ∈ B, a ∈ PA(y) and b ∈ PB(a), then: (173) b ∈(cid:0)(bdry B) r A(cid:1) ∩[j∈J (Bj ∩eBj) ⇔ b ∈ B r A ⇒ a ∈ A r B. 47 (iv) If x ∈ A, b ∈ PB(x), and a ∈ PA(b), then: a ∈(cid:0)(bdry A) r B(cid:1) ∩[i∈I (Ai ∩ eAi) ⇔ a ∈ A r B ⇒ b ∈ B r A. (174) (175) kb − ak ≤ (θ3δ + 2ε)ka − yk, (v) Suppose that B is (eA, ε, 3δ)-joint-regular at c (see Definition 9.6), that y ∈ ((bdry B) r A) ∩ Sj∈J(Bj ∩eBj), that a ∈ PA(y), that b ∈ PB(a), and that ky − ck ≤ δ. Then where θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) (see (101)). (vi) Suppose that A is (eB, ε, 3δ)-joint-regular at c (see Definition 9.6), that x ∈ ((bdry A) r B) ∩ Si∈I(Ai ∩ eAi), that b ∈ PB(x), that a ∈ PA(b), and that kx − ck ≤ δ. Then where θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) (see (101)). ka − bk ≤ (θ3δ + 2ε)kb − xk, (176) Proof. (i)&(ii): Clear from Lemma 2.4 and the assumptions. (iii): Note that Lemma 10.1(iv)&(iii) and (ii) yield the implications (177) b ∈ B r A ⇔ b ∈ (bdry B) r A ⇒ a ∈ A r B ⇔ a ∈ (bdry A) r B ⇒ b ∈ [j∈J which give the conclusion. (iv): Interchange the roles of A and B in (iii). (Bj ∩eBj), (178) ka − bk = dB(a) ≤ dBj (a) = ka − b′k. (v): There exists j ∈ J such that y ∈ Bj ∩eBj ∩ ((bdry B) r A). Let b′ ∈ PBj a. Then Since B is (eA, ε, 3δ)-joint-regular at c, it is clear that Bj is (eA, ε, 3δ)-regular at c. Since y ∈ (bdry B) r A and because of (i), there exists i ∈ I such that a ∈ PAiy ⊆ eAi. Since eAi ⊆ eA, it follows that (see also Remark 9.2(v)) Bj is (eAi, ε, 3δ)-regular at c. Since y ∈ Bj ∩eBj, a ∈ PAiy ∩ eAi, b′ ∈ PBja, and dAi (y) = dA(y) = ky − ak ≤ ky − ck ≤ δ, we obtain from Proposition 10.4 that (179) Combining with (178), we deduce that ka − bk ≤ ka − b′k ≤ (θ3δ + 2ε)ka − yk. ka − b′k ≤(cid:0)θ3δ(Ai, eAi, Bj,eBj) + 2ε(cid:1)ka − yk. (vi): This follows from (v) and (97). (cid:4) 48 An abstract linear convergence result Let us now focus on algorithmic results (which are actually true even in complete metric spaces). Definition 10.6 (linear convergence) Let (xn)n∈N be a sequence in X, let ¯x ∈ X, and let γ ∈ [0, 1[. Then (xn)n∈N converges linearly to ¯x with rate γ if there exists µ ∈ R+ such that (180) (∀n ∈ N) d(xn, ¯x) ≤ µγn. Remark 10.7 (rate of convergence depends only on the tail of the sequence) Let (xn)n∈N be a sequence in X, let ¯x ∈ X, and let γ ∈ ]0, 1[. Assume that there exists n0 ∈ N and µ0 ∈ R+ such that (cid:0)∀n ∈ {n0, n0 + 1, . . .}(cid:1) d(xn, ¯x) ≤ µ0γn. (181) Set µ1 := max(cid:8)d(xm, ¯x)/γm(cid:12)(cid:12) m ∈ {0, 1, . . . , n0 − 1}(cid:9). Then (182) (∀n ∈ N) d(xn, ¯x) ≤ max{µ0, µ1}γn, and therefore (xn)n∈N converges linearly to ¯x with rate γ. Proposition 10.8 (abstract linear convergence) Let A and B be nonempty closed subsets of X, let (an)n∈N be a sequence in A, and let (bn)n∈N be a sequence in B. Assume that there exist constants α ∈ R+ and β ∈ R+ such that (183a) γ := αβ < 1 and (183b) (∀n ∈ N) d(an+1, bn) ≤ αd(an, bn) and d(an+1, bn+1) ≤ βd(an+1, bn). Then (∀n ∈ N) d(an+1, bn+1) ≤ γd(an, bn) and there exists c ∈ A ∩ B such that d(a0, b0) · γn; (184) 1 + α 1 − γ consequently, (an)n∈N and (bn)n∈N converge linearly to c with rate γ. (∀n ∈ N) max(cid:8)d(an, c), d(bn, c)(cid:9) ≤ Proof. Set δ := d(a0, b0). Then for every n ∈ N, (185) d(an, bn) ≤ βd(an, bn−1) ≤ αβd(an−1, bn−1) = γd(an−1, bn−1) ≤ · · · ≤ γnδ; hence, (186a) (186b) d(bn, bn+1) ≤ d(bn, an+1) + d(an+1, bn+1) ≤ αd(bn, an) + γd(an, bn) = (α + γ)d(an, bn) ≤ (α + γ)δγn. 49 Thus (bn)n∈N is a Cauchy sequence, so there exists c ∈ B such that bn → c. On the other hand, by (185), d(an, bn) → 0 and (an)n∈N lies in A. Hence, an → c and c ∈ A. Thus, c ∈ A ∩ B. Fix n ∈ N and let m ≥ n. Using (186), m−1 ∑ k=n d(bn, bm) ≤ (α + γ)δγk = d(bk, bk+1) ≤ ∑ k≥n d(bk, bk+1) ≤ ∑ k≥n 1 − γ (187) (α + γ)δγn . Hence, using (185) and (187), we estimate that (188) d(an, bm) ≤ d(an, bn) + d(bn, bm) ≤ δγn + Letting m → +∞ in (187) and (188), we obtain (184). (α + γ)δγn 1 − γ = (1 + α)δγn 1 − γ . (cid:4) The sequence generated by the MAP We start with the following definition, which is well defined by Proposition 2.2. Definition 10.9 (MAP) Let A and B be nonempty closed subsets of X, let b−1 ∈ X, and let (189) (∀n ∈ N) an ∈ PA(bn−1) and bn ∈ PB(an). Then we say that the sequences (an)n∈N and (bn)n∈N are generated by the method of alternating projections (with respect to the pair (A, B)) with starting point b−1. The MAP between A = A1 ∪ A2 and B, A ∩ B = {c1, c2} a0 b−1 a1 A1 A2 c2 B c1 b0 b1 Our aim is to provide sufficient conditions for linear convergence of the sequences generated by the method of alternating projections. The following two results are simple yet useful. Proposition 10.10 Let A and B be nonempty closed subsets of X, and let (an) and (bn) be sequences generated by the method of alternating projections. Then the following hold: 50 b c b c b b b b b (i) The sequences (an)n∈N and (bn)n∈N lie in A and B, respectively. (ii) (∀n ∈ N) kan+1 − bn+1k ≤ kan+1 − bnk ≤ kan − bnk. (iii) If {an}n∈N ∩ B 6= ∅, or {bn}n∈N ∩ A 6= ∅, then there exists c ∈ A ∩ B such that for all n sufficiently large, an = bn = c. Proof. (i): This is clear from the definition. (ii): Indeed, for every n ∈ N, kan+1 − bn+1k = dB(an+1) ≤ kan+1 − bnk = dA(bn) ≤ kbn − ank using (i). (iii): Suppose, say that an ∈ B. Then bn = PBan = an =: c ∈ A ∩ B and all subsequent terms of the sequences are equal to c as well. (cid:4) New convergence results for the MAP We are now in a position to state and derive new linear convergence results. In this section, we shall often assume the following: (190)  of nonempty closed subsets of X; Bj are closed; Ai and B := [j∈J A := (Ai)i∈I and B := (Bj)j∈J are nontrivial collections A :=[i∈I c ∈ A ∩ B; eA := (eAi)i∈I and eB := (eBj)j∈J are collections (∀i ∈ I) PAi(cid:0)(bdry B) r A(cid:1) ⊆ eAi, (∀j ∈ J) PBj(cid:0)(bdry A) r B(cid:1) ⊆ eBj; eA :=[i∈I eAi andeB := [j∈JeBj. of nonempty subsets of X such that Lemma 10.11 (backtracking MAP) Assume that (190) holds. Let (an)n∈N and (bn)n∈N be generated by the MAP with starting point b−1. Let n ∈ {1, 2, 3, . . .}. Then the following hold: (i) If bn /∈ A, then an ∈ ((bdry A) r B) ∩Si∈I(Ai ∩ eAi) and bn ∈ ((bdry B) r A) ∩Sj∈J(Bj ∩eBj). (ii) If an /∈ B, then an ∈ ((bdry A) r B) ∩Si∈I(Ai ∩ eAi). 51 (iii) If an /∈ B and n ≥ 2, then bn−1 ∈ ((bdry B) r A) ∩Sj∈J(Bj ∩eBj). Proof. (i): Applying Proposition 10.5(iii) to bn−1 ∈ B, an ∈ PAbn−1, bn ∈ PBan, we obtain (191) bn ∈ B r A ⇔ bn ∈(cid:0)(bdry B) r A(cid:1) ∩[j∈J (Bj ∩eBj) ⇒ an ∈ A r B. On the other hand, applying Proposition 10.5(iv) to an−1 ∈ A, bn−1 ∈ PBan−1, an ∈ PAbn−1, we see that (192) an ∈ A r B ⇔ an ∈(cid:0)(bdry A) r B(cid:1) ∩[i∈I Altogether, (i) is established. (Ai ∩ eAi). (ii)&(iii): The proofs are analogous to that of (i). (cid:4) Let us now state and prove a key technical result. Proposition 10.12 Assume that (190) holds. Suppose that there exist ε ≥ 0 and δ > 0 such that the following hold: (i) A is (eB, ε, 3δ)-joint-regular at c (see Definition 9.6) and set (193) σ :=(1, 2, if B is not known to be (eA, ε, 3δ)-joint-regular at c; if B is also (eA, ε, 3δ)-joint-regular at c. (ii) θ3δ < 1− 2ε, where θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.2). Set θ := θ3δ + 2ε ∈ ]0, 1[. Let (an)n∈N and (bn)n∈N be sequences generated by the MAP with starting point b−1 satisfying (194) kb−1 − ck ≤ (1 − θσ)δ 6(2 + θ − θσ) . Then (an)n∈N and (bn)n∈N converge linearly to some point ¯c ∈ A ∩ B with rate θσ; in fact, 2 + θ − θσ θσ(n−1). (∀n ≥ 1) max(cid:8)kan − ¯ck, kbn − ¯ck(cid:9) ≤ k ¯c − ck ≤ δ δ(1 + θ) (195) and Proof. In view of a1 ∈ PAPBPAb−1 and (194), Corollary 10.3 yields (1 − θσ)δ δ (2 + θ − θσ) ≤ 2 β := ka1 − ck ≤ (196) . 52 Since c ∈ A ∩ B, we have θ3δ ≥ 0 by (98) and hence θ > 0. Using (196), we estimate (197a) (197b) (197c) (197d) (197e) (∀n ≥ 1) βθσ(n−1) + β + β(1 + θ) n−2 ∑ k=0 θσk ≤ β + β(1 + θ) = β + β(1 + θ) n−1 ∑ k=0 θσk 1 − θσn 1 − θσ 1 + θ 1 − θσ ≤ β + β = β(cid:16) 2 + θ − θσ 1 − θσ (cid:17) ≤ δ. We now claim that if (198) then (199a) (199b) n ≥ 1, kan − bnk ≤ βθσ(n−1) and kan − ck ≤ β + β(1 + θ) n−2 ∑ k=0 θσk, kan+1 − bn+1k ≤ θσ−1kan+1 − bnk ≤ θσkan − bnk ≤ βθσn, kan+1 − ck ≤ β + β(1 + θ) n−1 ∑ k=0 θσk. To prove this claim, assume that (198) holds. Using (198) and (197), we first observe that (200a) (200b) max(cid:8)kan − ck, kbn − ck(cid:9) ≤ kbn − ank + kan − ck ≤ βθσ(n−1) + β + β(1 + θ) n−2 ∑ k=0 θσk ≤ δ. We now consider two cases: Case 1: bn ∈ A ∩ B. Then bn = an+1 = bn+1 and thus (199a) holds. Moreover, kan+1 − ck = kbn − ck and (199b) follows from (200a). Case 2: bn 6∈ A ∩ B. Then bn ∈ B r A. Lemma 10.11(i) implies an ∈ ((bdry A) r B) ∩Si∈I(Ai ∩ eAi) and bn ∈ ((bdry B) r A) ∩Sj∈J(Bj ∩eBj). Note that kan − ck ≤ δ by (200a), and recall that A is (eB, ε, 3δ)-joint-regular at c by (i). It thus follows from Proposition 10.5(vi) (applied to an, bn, an+1) that (201) kan+1 − bnk ≤ θkan − bnk. On the one hand, if σ = 1, then Proposition 10.10(ii) yields kan+1 − bn+1k ≤ kan+1 − bnk = θσ−1kan+1 − bnk. On the other hand, if σ = 2, then B is (eA, ε, 3δ)-joint-regular at c by (i); hence, 53 Proposition 10.5(v) (applied to bn, an+1, bn+1) yields kan+1 − bn+1k ≤ θkan+1 − bnk = θσ−1kan+1 − bnk. Altogether, in either case, (202) kan+1 − bn+1k ≤ θσ−1kan+1 − bnk. Combining (202) with (201) and (198) gives (203) kan+1 − bn+1k ≤ θσ−1kan+1 − bnk ≤ θσkan − bnk ≤ βθσn, which is (199a). Furthermore, (201), (198) and (200a) yield (204a) (204b) (204c) (204d) kan+1 − ck ≤ kan+1 − bnk + kbn − ck ≤ θkan − bnk + kbn − ck ≤ θβθσ(n−1) + βθσ(n−1) + β + β(1 + θ) n−2 ∑ k=0 θσk = β + β(1 + θ) n−1 ∑ k=0 θσk, which establishes (199b). Therefore, in all cases, (199) holds. Since ka1 − b1k = dB(a1) ≤ ka1 − ck = β, we see that (198) holds for n = 1. Thus, the above claim and the principle of mathematical induction principle imply that (199) holds for every n ≥ 1. Next, (199a) implies (205) (∀n ≥ 1) kan+1 − bnk ≤ θkan − bnk and kan+1 − bn+1k ≤ θσ−1kan+1 − bnk. (∀n ≥ 1) max(cid:8)kan − ¯ck, kbn − ¯ck(cid:9) ≤ In view of (205) and ka1 − b1k ≤ β, Proposition 10.8 yields ¯c ∈ A ∩ B such that 1 − θσka1 − b1k · θσ(n−1) (206) 1 − θσ β · θσ(n−1) 2 + θ − θσ θσ(n−1). δ(1 + θ) ≤ ≤ (207) 1 + θ 1 + θ (208) On the other hand, (199b) and (197) imply (∀n ≥ 1) kan+1 − ck ≤ δ; thus, letting n → +∞, we obtain k ¯c − ck ≤ δ. This completes the proof of (195). Remark 10.13 In view of Lemma 10.1(i)&(ii), an aggressive choice for use in (190) is (∀i ∈ I) (cid:4) eAi = bdry Ai and (∀j ∈ J) eBj = bdry Bj. Our main convergence result on the linear convergence of the MAP is the following: Theorem 10.14 (linear convergence of the MAP and superregularity) Assume that (190) holds and that A is eB-joint-superregular at c (see Definition 9.6). Denote the limiting joint-CQ-number at c associ- ated with (A, eA, B, eB) (see Definition 7.2) by θ, and the the exact joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.7) by α. Assume further that one of the following holds: 54 (i) θ < 1. (ii) I and J are finite, and α < 1. Let θ ∈(cid:3)θ, 1(cid:2) and set ε := (θ − θ)/3 > 0. Then there exists δ > 0 such that the following hold: (iii) A is (eB, ε, 3δ)-joint-regular at c (see Definition 9.6). (iv) θ3δ ≤ θ + ε < 1 − 2ε, where θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.2). Consequently, suppose the starting point of the MAP b−1 satisfies kb−1 − ck ≤ (1 − θ)δ/12. Then (an)n∈N and (bn)n∈N converge linearly to some point in ¯c ∈ A ∩ B with k ¯c − ck ≤ δ and rate θ: (209) δ(1 + θ) θn−1. (∀n ≥ 1) max{kan − ¯ck, kbn − ¯ck} ≤ 2 Proof. Observe that (ii) implies (i) by Theorem 7.8(iv). The definitions of eB-joint-superregularity and of θ allow us to find δ > 0 sufficiently small such that both (iii) and (iv) hold. The result thus follows from Proposition 10.12 with σ = 1. (cid:4) Corollary 10.15 Assume that (190) holds and that, for every i ∈ I, Ai is convex. Denote the limiting joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.2) by θ, and assume that θ < 1. Let θ ∈(cid:3)θ, 1(cid:2), and let b−1, the starting point of the MAP, be sufficiently close to c. Then (an)n∈N and (bn)n∈N converge linearly to some point in A ∩ B with rate θ. Proof. Combine Theorem 10.14 with Corollary 9.8. (cid:4) Example 10.16 (working with collections and joint notions is useful) Consider the setting of Example 7.4, and suppose that eA = A and eB = B. Note that Ai is convex, for every i ∈ I. Then θδ(A, eA, B, eB) < 1 = θδ(A, A, B, B) = θ(A, X, B, X). Hence Corollary 10.15 guarantees lin- ear convergence of the MAP while it is not possible to work directly with the unions A and B due to their condition number being equal to 1 and because neither A nor B is superregular by Exam- ple 9.9! This illustrates that the main result of Lewis-Luke-Malick (see Corollary 10.24 below) is not applicable because two of its hypotheses fail. The following result features an improved rate of convergence θ2 due to the additional presence of superregularity. Theorem 10.17 (linear convergence of the MAP and double superregularity) Assume that (190) holds, that A is eB-joint-superregular at c and that B is eA-joint-superregular at c (see Definition 9.6). Denote the limiting joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.2) by θ, and the the exact joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.7) by α. Assume further θ ∈(cid:3)θ, 1(cid:2) and ε := θ−θ that (a) θ < 1, or (more restrictively) that (b) I and J are finite, and α < 1 (and hence θ = α < 1). Let 3 . Then there exists δ > 0 such that 55 (see Definition 7.2). (i) A is (eB, ε, 3δ)-joint-regular at c; (ii) B is (eA, ε, 3δ)-joint-regular at c; and (iii) θ3δ < θ + ε = θ − 2ε < 1 − 2ε, where θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) Consequently, suppose the starting point of MAP b−1 satisfies kb−1 − ck ≤ (1−θ)δ 6(2−θ) . Then (an)n∈N and (bn)n∈N converge linearly to some point in ¯c ∈ A ∩ B with k ¯c − ck ≤ δ and rate θ2; in fact, 2 − θ(cid:0)θ2(cid:1)n−1 (210) (∀n ≥ 1) max(cid:8)kan − ¯ck, kbn − ¯ck(cid:9) ≤ δ . Proof. The existence of δ > 0 such that (i) -- (iii) hold is clear. Then apply Proposition 10.12 with σ = 2. (cid:4) In passing, let us point out a sharper rate of convergence under sufficient conditions stronger than superregularity. Corollary 10.18 (refined convergence rate) Assume that (190) holds and that there exists δ > 0 such that (i) A is (eB, 0, 3δ)-joint-regular at c; (ii) B is (eA, 0, 3δ)-joint-regular at c; and (iii) θ < 1, where θ := θ3δ is the joint-CQ-number at c associated with (A, eA, B, eB) (see Definition 7.2). Suppose also that the starting point of the MAP b−1 satisfies kb−1 − ck ≤ (1−θ)δ 6(2−θ) . Then (an)n∈N and (bn)n∈N converge linearly to some point in ¯c ∈ A ∩ B with k ¯c − ck ≤ δ and rate θ2; in fact, 2 − θ(cid:0)θ2(cid:1)n−1 (211) (∀n ≥ 1) max(cid:8)kan − ¯ck, kbn − ¯ck(cid:9) ≤ δ . Proof. Apply Proposition 10.12 with σ = 2. (cid:4) Let us illustrate a situation where it is possible to make δ in Theorem 10.17 precise. Example 10.19 (the MAP for two spheres) Let z1 and z2 be in X, let ρ1 and ρ2 be in R, set A := sphere(z1; ρ1) and B := sphere(z2; ρ2), and assume that {c} $ A ∩ B $ A ∪ B. Then α := hz1 − c, z2 − ci /(ρ1ρ2) < 1. Let θ ∈ ]α, 1[. Then the conclusion of Theorem 10.17 holds with (212) δ := min(p(ρ1 + ρ2)2 + ρ1ρ2(θ − α) − (ρ1 + ρ2) 6 , ερ1 3 , ερ2 3 ) 56 Proof. Combine Example 9.3 (applied with ε = (θ − α)/4 there), Proposition 8.4, and Theo- rem 10.17. (cid:4) Here is a useful special case of Theorem 10.17: Theorem 10.20 Assume that A and B are L-superregular, and that (213) where L := aff(A ∪ B). Then the sequences generated by the MAP converge linearly to a point in A ∩ B provided that the starting point is sufficiently close to c. NA(c) ∩(cid:0) − NB(c)(cid:1) ∩(cid:0)L − c(cid:1) = {0}, Proof. Combine Example 8.2 with Theorem 10.17 (applied with I and J being singletons, and with (cid:4) eA = eB = L). We now obtain a well known global linear convergence result for the convex case, which does not require the starting point to be sufficiently close to A ∩ B: Theorem 10.21 (two convex sets) Assume that A and B are convex, and A ∩ B 6= ∅. Then for every starting point b−1 ∈ X, the sequences (an)n∈N and (bn)n∈N generated by the MAP converge to some point in A ∩ B. The convergence of these sequences is linear provided that ri A ∩ ri B 6= ∅. Proof. By Fact 2.5(iv), we have (∀c ∈ A ∩ B) ka0 − ck ≥ kb0 − ck ≥ ka1 − ck ≥ kb1 − ck ≥ · · · (214) After passing to subsequences if needed, we assume that akn → a ∈ A and bkn → b ∈ B. We show that a = b by contradiction, so we assume that ε := ka − bk/3 > 0. We have eventually max{kakn − ak, kbkn − bk} < ε; hence kakn − bknk ≥ ε eventually. By Fact 2.5(iii), we have kakn − ck2 ≥ kakn − bknk2 + kbkn − ck2 ≥ ε2 + kakn +1 − ck2 ≥ ε2 + kakn+1 − ck2 (215) eventually. But this would imply that for all n sufficiently large, and for every m ∈ N, we have kakn − ck2 ≥ mε2 + kakn+m − ck2 ≥ mε2, which is absurd. Hence ¯c := a = b ∈ A ∩ B and now (214) (with c = ¯c) implies that an → ¯c and bn → ¯c. Next, assume that ri A ∩ ri B 6= ∅, and set L := aff(A ∪ B). By Proposition 8.5, the (A, L, B, L)- CQ conditions holds at ¯c. Thus, by Example 8.2, NA( ¯c) ∩ (−NB( ¯c)) ∩ (L − ¯c) = {0}. Furthermore, Corollary 9.8 and Remark 9.2(vi)&(viii) imply that A and B are L-superregular at ¯c. The conclusion now follows from Theorem 10.20, applied to suitably chosen tails of the sequences (an)n∈N and (bn∈N). Example 10.22 (the MAP for two linear subspaces) Assume that A and B are linear subspaces of X. Since 0 ∈ A ∩ B = ri A ∩ ri B, Theorem 10.21 guarantees the linear convergence of the MAP to some point in A ∩ B, where b−1 ∈ X is the arbitrary starting point. On the other hand, A and B are (0, +∞)-regular (see Remark 9.2(vi)). Since (∀δ ∈ R++) θδ(A, A, B, B) = c(A, B) < 1, where c(A, B) is the cosine of the Friedrichs angle between A and B (see Theorem 8.12), we obtain from Corollary 10.18 that the rate of convergence is c2(A, B). In fact, it is well known that this is the optimal rate, and also that limn an = limn bn = PA∩B(b−1); see [10, Section 3] and [11, Chapter 9]. (cid:4) 57 Remark 10.23 For further linear convergence results for the MAP in the convex setting we refer the reader to [1], [2], [3], [12], [13], [14], and the references therein. See also [20] and [21] for recent related work for the nonconvex case. Comparison to Lewis-Luke-Malick results and further examples The main result of Lewis, Luke, and Malick arises as a special case of Theorem 10.14: Corollary 10.24 (Lewis-Luke-Malick) (See [17, Theorem 5.16].) Suppose that NA(c) ∩ (−NB(c)) = {0} and that A is superregular at c ∈ A ∩ B. If the starting point of MAP is sufficiently close to c, then the sequences generated by the MAP converge linearly to a point in A ∩ B. Proof. Since NA(c) ∩ (−NB(c)) = {0}, we have θ < 1. Now apply Theorem 10.14(i) with eA := eB := (X), A := (A) and B := (B). However, even in simple situations, Corollary 10.24 is not powerful enough to recover known (cid:4) convergence results. Example 10.25 (Lewis-Luke-Malick CQ may fail even for two subspaces) Suppose that A and B are two linear subspaces of X, and set L := aff(A ∪ B) = A + B. For c ∈ A ∩ B, we have (216) NA(c) ∩ (−NB(c)) = A⊥ ∩ B⊥ = (A + B)⊥ = L⊥. Therefore, the Lewis-Luke-Malick CQ (see [17, Theorem 5.16] and also Corollary 10.24) holds for (A, B) at c if and only if (217) NA(c) ∩ (−NB(c)) = {0} ⇔ A + B = X. On the other hand, the CQ provided in Theorem 10.20 (see also Example 10.22) always holds and we obtain linear convergence of the MAP. However, even for two lines in R3, the Lewis-Luke-Malick CQ (see Corollary 10.24) is unable to achieve this. (It was this example that originally motivated us to pursue the present work.) Example 10.26 (Lewis-Luke-Malick CQ is too strong even for convex sets) Assume that A and B are convex (and hence superregular). Then the Lewis-Luke-Malick CQ condition is 0 ∈ int(B − A) (see Corollary (i)) while the (A, aff(A ∪ B), B, aff(A ∪ B))-CQ is equivalent to the much less restrictive condition ri A ∩ ri B 6= ∅ (see Theorem 4.13). The flexibility of choosing (eA,eB) Theorem 10.20. However, there are situations when this choice for eA and eB is not helpful but Often, L = aff(A ∪ B) is a convenient choice which yields linear convergence of the MAP as in when a different, more aggressive, choice does guarantee linear convergence: 58 Example 10.27 ((eA,eB) = (A, B)) Let A, B, and c be as in Example 8.3, and let L := aff(A ∪ B). Since A and B are convex and hence superregular, the (A, L, B, L)-CQ condition is equivalent to ri A ∩ ri B 6= ∅ (see Proposition 8.5), which fails in this case. However, the (A, A, B, B)-CQ con- dition does hold; hence, the corresponding limiting CQ-number is less than 1 by Theorem 7.8(v). Thus linear convergence of the MAP is guaranteed by Theorem 10.17. The next example illustrates a situation where the choice (eA,eB) = (A, B) fails while the even tighter choice (eA,eB) = (bdry A, bdry B) results in success: Example 10.28 ((eA,eB) = (bdry A, bdry B)) Suppose that X = R2, that A = epi( · /2), that B = − epi( · /3), and that c = (0, 0). Note that aff(A ∪ B) = X and ri A ∩ ri B = ∅. Then (218a) and so the (A, A, B, B)-CQ condition fails because NB N A A(c) = NX B (c) = NX A(c) ∩ (−N A NB A (c) = NA(c) =(cid:8)(u1, u2) ∈ R2(cid:12)(cid:12) u2 + 2u1 ≤ 0(cid:9), B (c) = NB(c) =(cid:8)(u1, u2) ∈ R2(cid:12)(cid:12) −u2 + 3u1 ≤ 0(cid:9), B (c)) =(cid:8)(u1, u2) ∈ R2(cid:12)(cid:12) u2 + 3u1 ≤ 0(cid:9) 6= {0}. (218b) (219) (220a) (220b) because α = θ = 1: indeed, u = (0, −1) ∈ NA(c) and v = (0, −1) ∈ −NB(c), so 1 = hu, vi ≤ ¯α ≤ 1. Consequently, for either (eA,eB) = (A, B) or (eA,eB) = (X, X), Theorem 10.17 is not applicable On the other hand, let us now choose (eA,eB) = (bdry A, bdry B), which is justified by Re- mark 10.13. Then A(c) =(cid:8)(u1, u2) ∈ R2(cid:12)(cid:12) u2 + 2u1 = 0(cid:9), NeB B (c) =(cid:8)(u1, u2) ∈ R2(cid:12)(cid:12) −u2 + 3u1 = 0(cid:9), NeA rem 7.8(v), Theorem 10.21 and Theorem 10.17, we deduce linear convergence of the MAP. linear convergence as the following shows. In this example, we employ the tightest possibility A(c) ∩ (−NeA NeB B (c)) = {0} and the (A, eA, B,eB)-CQ condition holds. Hence, using also Theo- However, even the choice (eA,eB) = (bdry A, bdry B) may not be applicable to yield the desired allowed by our framework, namely (eA,eB) = (PA((bdry B) r A), PB((bdry A) r B)). Example 10.29 ((eA,eB) = (PA((bdry B) r A), PB((bdry A) r B)) ) Suppose that X = R2, that A = epi( · ), that B = −A, and that c = (0, 0). Then Nbdry B (c) = bdry B = − bdry A and Nbdry A bdry A; hence, the (A, bdry A, B, bdry B)-CQ condition fails because Nbdry B (c) ∩ (−Nbdry A bdry B 6= {0}. On the other hand, if (eA,eB) = (PA((bdry B) r A), PB((bdry A) r B)), then NeB {0} = NeA B = {0} because eA = {c} = eB. Thus, the (A, eA, B,eB)-CQ conditions holds. (Note that the MAP converges in finitely many steps.) (c) = (c)) = A = B B A A 59 Conclusion We have introduced restricted normal cones which generalize classical normal cones. We have presented some of their basic properties and shown their usefulness in describing interiority con- ditions, constraint qualifications, and regularities. The corresponding results were employed to yield new powerful sufficient conditions for linear convergence of the sequences generated by the method of alternating projections applied to two sets A and B. A key ingredient were suitable basic settings from convex analysis into our framework. Thus, the framework provided here unifies the recent nonconvex results by Lewis, Luke, and Malick with classical convex-analytical restricting sets (eA and eB). The least aggressive choice, (eA,eB) = (X, X), recovers the framework by Lewis, Luke, and Malick. The choice (eA,eB) = (aff(A ∪ B), aff(A ∪ B)) allows us to include settings. When the choice (eA,eB) = (aff(A ∪ B), aff(A ∪ B)) fails, one may also try more aggres- sive choices such as (eA,eB) = (A, B) or (eA,eB) = (bdry A, bdry B) to guarantee linear convergence. In a follow-up work [5] we demonstrate the power of these tools with the important problem of sparsity optimization with affine constraints. Without any assumptions on the regularity of the sets or the intersection we achieve local convergence results, with rates and radii of convergence, where all other sufficient conditions, particularly those of [18] and [17], fail. Acknowledgments HHB was partially supported by the Natural Sciences and Engineering Research Council of Canada and by the Canada Research Chair Program. This research was initiated when HHB vis- ited the Institut f ur Numerische und Angewandte Mathematik, Universitat G ottingen because of his study leave in Summer 2011. HHB thanks DRL and the Institut for their hospitality. DRL was supported in part by the German Research Foundation grant SFB755-A4. HMP was partially supported by the Pacific Institute for the Mathematical Sciences and and by a University of British Columbia research grant. XW was partially supported by the Natural Sciences and Engineering Research Council of Canada. References [1] H.H. Bauschke and J.M. Borwein, On the convergence of von Neumann's alternating pro- jection algorithm for two sets, Set-Valued Analysis 2 (1993), 185 -- 212. [2] H.H. Bauschke and J.M. Borwein, On projection algorithms for solving convex feasibility problems, SIAM Review 38 (1996), 367 -- 426. [3] H.H. Bauschke, J.M. Borwein, and A.S. Lewis, The method of cyclic projections for closed convex sets in Hilbert space, in Recent Developments in Optimization Theory and Nonlin- ear Analysis (Jerusalem 1995), Y. Censor and S. Reich (editors), Contemporary Mathemat- ics vol. 204, American Mathematical Society, pp. 1 -- 38, 1997. 60 [4] H.H. Bauschke and P.L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, Springer, 2011. [5] H.H. Bauschke, D.R. Luke, H.M. Phan, and X. Wang, Restricted normal cones and sparsity optimization with affine constraints, arXiv preprint, May 2012, http://arxiv.org [6] J.M. Borwein and Q.J. Zhu, Techniques of Variational Analysis, Springer-Verlag, 2005. [7] Y. Censor and S.A. Zenios, Parallel Optimization, Oxford University Press, 1997. [8] F.H. Clarke, Y.S. Ledyaev, R.J. Stern and P.R. Wolenski, Nonsmooth Analysis and Control The- ory, Springer-Verlag, 1998. [9] P.L. Combettes and H.J. Trussell, Method of successive projections for finding a common point of sets in metric spaces, Journal of Optimization Theory and Applications 67 (1990), 487 -- 507. [10] F. Deutsch, The angle between subspaces of a Hilbert space, in Approximation theory, wavelets and applications (Maratea, 1994), S.P. Singh, A. Carbone, and B. Watson (editors), NATO Ad- vanced Science Institutes Series C: Mathematical and Physical Sciences vol. 454, Kluwer, pp. 107 -- 130, 1995. [11] F. Deutsch, Best Approximation in Inner Product Spaces, Springer, 2001. [12] F. Deutsch and H. Hundal, The rate of convergence for the cyclic projections algorithm I: angles between convex sets, Journal of Approximation Theory 142 (2006), 36 -- 55. [13] F. Deutsch and H. Hundal, The rate of convergence for the cyclic projections algorithm II: norms of nonlinear operators, Journal of Approximation Theory 142 (2006), 56 -- 82. [14] F. Deutsch and H. Hundal, The rate of convergence for the cyclic projections algorithm III: regularity of convex sets, Journal of Approximation Theory 155 (2008), 155 -- 184. [15] J. Dixmier, ´Etude sur les vari´et´es et les op´erateurs de Julia, avec quelques applications, Bul- letin de la Soci´et´e Math´ematique de France 77 (1949), 11 -- 101. [16] K. Friedrichs, On certain inequalities and characteristic value problems for analytic func- tions and for functions of two variables, Transactions of the AMS 41 (1937), 321 -- 364. [17] A.S. Lewis, D.R. Luke, and J. Malick, Local linear convergence for alternating and averaged nonconvex projections, Foundations of Computational Mathematics 9 (2009), 485 -- 513. [18] A.S. Lewis and J. Malick, Alternating projection on manifolds, Mathematics of Operations Research 33 (2008), 216 -- 234. [19] P.D. Loewen, Optimal Control via Nonsmooth Analysis, CRM Proceedings & Lecture Notes, AMS, Providence, RI, 1993. [20] D.R. Luke, Finding best approximation pairs relative to a convex and a prox-regular set in a Hilbert space, SIAM Journal on Optimization 19(2) (2008), 714 -- 739. 61 [21] D.R. Luke, Local linear convergence and approximate projections onto regularized sets, Nonlinear Analysis 75 (2012), 1531 -- 1546. [22] B.S. Mordukhovich, Variational Analysis and Generalized Differentiation I, Springer-Verlag, 2006. [23] R.T. Rockafellar, Convex Analysis, Princeton University Press, Princeton, 1970. [24] R.T. Rockafellar and R.J-B Wets, Variational Analysis, Springer, corrected 3rd printing, 2009. [25] H. Stark and Y. Yang, Vector Space Projections, Wiley, 1998. [26] C. Zalinescu, Convex Analysis in General Vector Spaces, World Scientific Publishing, 2002. 62
1905.10543	1	1905	2019-05-25T07:28:02	A generalization of $b$-weakly compact operators	[ "math.FA" ]	A. Bahramnezhad and K. Haghnejad Azar introduced the classes of $KB$-operators and $WKB$-operators, and they studied some of theirs properties. In the present paper, we give answer for an open problem from that paper, which two classifications of operators, $b$-weakly compact operators and $KB$-operators are different. A continuous operator $T$ from a normed vectoe lattice $E$ into a normed space $X$ is said to be $KB$-operator (respectively, $WKB$-operator) if $\{Tx_n\}_n$ has a norm (respectively, weak) convergent subsequence in $X$ for every positive increasing sequence $\{x_n\}_n$ in the closed unit ball $B_E$ of $E$. We investigate some other properties of $KB$-operators and its relationships with $b-$weakly compact operators.	math.FA	math	A GENERALIZATION OF b-WEAKLY COMPACT OPERATORS KAZEM HAGHNEJAD AZAR Abstract. A. Bahramnezhad and K. Haghnejad Azar introduced the classes of K B-operators and W K B-operators, and they studied some of theirs proper- ties. In the present paper, we give answer for an open problem from that paper, which two classifications of operators, b-weakly compact operators and K B- operators are different. A continuous operator T from a normed vectoe lattice E into a normed space X is said to be K B-operator (respectively, W K B- operator) if {T xn}n has a norm (respectively, weak) convergent subsequence in X for every positive increasing sequence {xn}n in the closed unit ball BE of E. We investigate some other properties of K B-operators and its relationships with b−weakly compact operators. 1. Introduction and preliminaries A subset A of a vector lattice E is called b-order bounded in E if it is order bounded in E∼∼. An operator T : E → X, mapping each b-order bounded subset of E into a relatively weakly compact subset of X is called a b-weakly compact operator. Alpay, Altin and Tonyali introduced the class of b-weakly compact operators for vector lattices having separating order duals [3]. In [4], Alpay and Altin proved that a continuous operator T from a Banach lattice E into a Ba- nach space X is b-weakly compact if and only if {T xn}n is norm convergent for each b-order bounded increasing sequence {xn}n in E+ if and only if {T xn}n is norm convergent to zero for each b-order bounded disjoint sequence {xn}n in E+. Authors in [6] proved that an operator T from a Banach lattice E into a Banach space X is b-weakly compact if and only if {T xn}n is norm convergent for every positive increasing sequence {xn}n of the closed unit ball BE of E. The aim of this paper is studied the classes of operators on Banach lattices are called KB-operators and W KB-operators. We will investigate on theirs properties. A continuous operator T from a Banach lattice E into a normed space X is said to be KB-operator ( W KB-operato), if {T xn}n has a norm ( weak) convergent subsequence in X for every positive increasing sequence {xn}n in the closed unit ball BE of E, see [9]. To state our results, we need to fix some notation and recall some definitions. A Banach lattice E is is said to be an AM-space if for each x, y ∈ E such that x ∧ y = 0, we have kx + yk = max{kxk, kyk}. A Banach lattice E is an AL- space if its topological dual E′ is an AM-space. A Banach lattice E is said to be Copyright 2018 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. Banach lattice, K B-operator, W K B-operator, b-weakly compact operator. 1 2 KAZEM HAGHNEJAD AZAR KB-space whenever each increasing norm bounded sequence of E+ is norm con- vergent. An operator T : E → F between two Riesz spaces is positive if T (x) ≥ 0 in F whenever x ≥ 0 in E. Note that each positive linear mapping on a Banach lattice is continuous. An operator T from a Banach space X into a Banach space Y is compact (resp. weakly compact) if T (BX) is compact (resp. weakly compact) where BX is the closed unit ball of X. For terminology concerning Banach lat- tice theory and positive operators, we refer the reader to the excellent book of [1]. As b-weakly compact operators [7], the class of KB-operators and W KB-operators does not satisfy duality property. In fact the identity operator of the Banach lat- tice ℓ1 is a KB-operator (respectively, W KB-operator); but its adjoint which is the identity operator of the Banach lattice ℓ∞, is not a KB-operator (respectively, W KB-operator). Conversely, the identity operator of the Banach lattice c0 is not a KB-operator (respectively, not W KB-operator); but its adjoint, which is the identity operator of the Banach lattice ℓ1, is a KB-operator (respectively, W KB- operator). Recall that a Banach space is said to have Schur property whenever w−→ 0 im- every weak convergent sequence is norm convergent, i.e., whenever xn plies kxnk → 0. Let E, F be Banach lattices. If either E or F has the Schur property then L(E, F ) = Wb(E, F ) [4]. It is also clear that for a Banach lattice E and a Banach space X with Schur property, every W KB-operator T : E → X is a KB-operator. Let E be a vector lattice. A sequence {xn}∞ 1 ⊂ E is called order convergent to x as n → ∞ if there exists a sequence {yn}∞ 1 such that yn ↓ 0 o1−→ x when {xn} is order as n → ∞ and xn − x ≤ yn for all n. We will write xn convergent to x. A sequence {xn} in a vector lattice E is strongly order conver- o2−→ x whenever there exists a net {yβ}β∈B in E gent to x ∈ E, denoted by xn such that yβ ↓ 0 and that for every β ∈ B, there exists n0 such that xn − x ≤ yβ for all n ≥ n0. It is clear that every order convergent sequence is strongly order convergent, but two convergence are different, for information see, [2]. A net (xα)α in Banach lattice E is unbounded norm convergent (or, un-convergent for k.k −→ 0 for all u ∈ E+. We denote this convergence short) to x ∈ E if xα − x ∧ u by xα un−→ x. This convergence has been introduced and studied in [10, 13]. 2. Main Results In each parts of this manuscript, E is a normed vector lattice and X normed space. The definition of b−weakly compact operator holds whenever E is a normed vector lattice and X normed space. The collections of KB-operators, W KB-operators, b-weakly compact operators, order weakly compact operators, weakly compact operators and compact operators will be denoted by LKB(E, X), WKB(E, X), Wb(E, X), Wo(E, X), W (E, X) and K(E, X), respectively, when- ever there is not confused. By notice an example from [12], page 95, the classifi- cation of order compact or order weakly compact operators from Banach lattice E into Banach space X, in general, are not subspace of LKB(E, X), but by The- orem 3.4.4, from [12], every interval-bounded and order weakly compact operator T : E → X is b−weakly compact operator, and so KB−operator. On the other hand, if e ∈ E is an atom, then LKB(Ee, F ) = Wo(Ee, F ). We have the following A GENERALIZATION OF b-WEAKLY COMPACT OPERATORS 3 relationships between these spaces: K(E, X) ⊆ W (E, X) ⊆ Wb(E, X) ⊆ LKB(E, X) ⊆ WKB(E, X). Let E be normed vector lattices, X normed space and T : E → X be a positive operator. By using Prpostion 2.1 from [9], T ∈ Wb(E, X) iff T ∈ LKB(E, X) iff T ∈ WKB(E, X) as follows. Proposition 2.1. [9] Let E and F be Banach lattices and T : E → F be a positive operator. Then the following statements are equivalent. (1) T is b-weakly compact. (2) T is KB-operator. (3) T is W KB-operator. In the following example, we give answer for an open problem from [9], which two classifications of operators Wb(E, X) and LKB(E, X) are different. In the other words, there is a KB-operator T from normed vector lattice E into Banach space X that is not b−weakly compact operator. Example 2.2. In the following, we show that the inclusion Wb(E, F ) ⊆ LKB(E, F ) may be proper whenever E is normed vector lattice and F is a Banach space. Proof. Authors in Theorem 2.10 of [6] proved that an operator T from Banach lattice E into Banach lattice F is b−weakly compact if and only if (T (xn))+∞ n=1 is norm convergent for every positive increasing sequence (xn) in the closed unit ball BE of E. Now this theorem holds when E is normed vectoe lattice. : r ∈ R+}i where χ[0,r] is a characteristic function on Suppose that E = h{χ[0,r] [0, r]. It is clear that E is a subspace of L∞([0, +∞)) and with k.k∞ is normed vector lattice. We define an operator T : E → l∞ with T (f ) = (Z +∞ 0 f (x) cos nπxdx)+∞ n=1 for all f ∈ E. T is a bounded linear operator of course Z +∞ 0 f (x) cos nπxdx ≤ Z +∞ 0 f (x)dx < +∞. 2 ], then T f = ( 1 nπ sin nπ 2 )+∞ Now the proof will be based upon three claims as follows: claim 1: T is not positive operator. If we set f = χ[0, 1 operator. claim 2: T is not b−weakly compact operator. Let fm = χ[0,m+ 1 (T fm)+∞ claim 3: T is a KB−operator. Let {fm} ⊆ E, 0 ≤ fm ↑ and supm kfmk∞ < +∞. We write T fm = am,n where 2 ] where m ∈ N. Then 0 ≤ fm ↑ and supm kfmk∞ < +∞, but n=1 which shows that T is not positive m=1 is not convergent. 0 fm(x) cos nxdx. am,n = R +∞ For fix n ∈ N, the sequence (am,n)+∞ sequence (akm,n)+∞ m=1 is norm bounded in R, and so it has a sub- m=1 which is convergent. It is clear 0 ≤ fkm ↑ and supm kfkmk∞ < 4 KAZEM HAGHNEJAD AZAR +∞. It follows that (akm,n+1)+∞ quence (a(k+1)m,n+1)+∞ of (am,n)+∞ is subsequence of (T fm)+∞ m=1 is norm bounded in R, and so it has also subse- m=1 is a subsequence m=1 which is convergent for each n ∈ N. On the other hand, (T fmm)+∞ m=1 m=1 which is convergent. Now, T is a KB−operator. (cid:3) m=1 which is convergent. Thus (amm,n)+∞ Proposition 2.1, shows that LKB(E, F ) = Wb(E, F ) whenever LKB(E, F ) is vector lattice. Consequently, the above example implies that in general LKB(E, F ) is not vector lattice. On the other hand, in general, LKB(E, F ) is not order dense in B(E, F ). Note that if we set E = F = c0 and Tn ∈ LKB(c0, c0) defined by Tn(x1, x2, ...) = (x1, x2, ..., xn, 0, 0, ...), then 0 ≤ Tn ↑ Ic0. But Ic0 is not KB- operator. Then by Proposition 2.1, Wb(E, X), in general, is also not order dense in B(E, F ). In the following we show that LKB(E, F ) is a norm closed subspace of B(E, F ) whenever F is a Banach lattice. Theorem 2.3. The collection of all KB-operators from E into Banach lattice F is norm closed vector subspace of B(E, F ). Proof. Let {Tn} ⊆ B(E, F ) be a sequence of KB-operators such that kTn − T k → 0 holds in B(E, F ) and let {xm} be a positive increasing sequence in E with supmkxmk < ∞. Since T1 ∈ LKB(E, F ), there is a subsequence {xm,1} of {xm} which {T1xm,1} is norm convergent in F . Choose subsequence {xm,n} and {xm,n+1} of {xm} as follows (1) {xm,n+1} is a subsequence of {xm,n} for each n ∈ N.. (2) {T xm,n}m is norm convergent in F for each n ∈ N. It follows that {Tnxm,m}m is norm convergent for each n ∈ N. From following inequalities kT xm,m − T xk,kk ≤ kT xm,m − Tnxm,mk + kTnxm,m − Tkxm,mk + kTkxm,m − T xm,mk, the sequence {T xm,m}m is a norm Cauchy, and so norm convergent in F . (cid:3) Theorem 2.4. Let E and F be Banach lattices where E has order continuous norm. Let G be a sublattice order dense in E and T be a positive operator from E into F . If T ∈ LKB(G, F ), then T ∈ LKB(E, F ). Proof. By using Propostion 2.1 from [9], it is enough to give a proof for b−weakly compact operators. Let {xn} be a positive increasing sequence in E with supnkxnk < ∞. Since G is order dense in E, by Theorem 1.34 from [1], we have {y ∈ G : 0 ≤ y ≤ xn} ↑ xn, m=1 ⊂ G with 0 ≤ ymn ↑ xn for each n. Put zmn = ∨n for each n. Let {ymn}∞ i=1ymi and 0 ≤ T ∈ L(E, F ). Its follows that zmn ↑m xn and supm,nkzm,nk ≤ supnkxnk < ∞. Now, if T ∈ Wb(G, F ), then {T zmn} is norm convergent to some point y ∈ F . Now, we have the following inequalities kT xn − T zmnk ≤ kT kkxn − zmnk ≤ kT kkxn − ymnk → 0. A GENERALIZATION OF b-WEAKLY COMPACT OPERATORS 5 Thus by the following inequality proof holds kT xn − yk ≤ kT xn − T zmnk + kT zmn − yk. (cid:3) By notice to Remark 2.26 from [9], LKB(E, F ) is not vector lattice whenever E and F are Banach lattices. On the other hand, if T ∈ LKB(E, F ), in general, the modulas of T need not be a KB−operators. Now in the following theorem, we show that T ∈ LKB(E, F ) whenever T ∈ LKB(E, F ). Theorem 2.5. Let E and F be normed vector lattices. We have the following assertions. (1) If T : E → F is an order bounded operator and F is KB−space, then T and T are KB−operators. (2) If T is KB−operator or b−weakly compact operator, then T ∈ LKB(E, F ) ∩ Wb(E, F ) Proof. (1) Let xn ↑ and supn kxnk < ∞. Since T + is positive, by Theorem 4.3 [1], T + is norm bounded. It follows that {T +xn} is norm bounded. Thus T + is KB−operator. In similar way, T − is also KB-operator. It follows that T and T are KB−operators. (2) Since 0 ≤ T −, T + ≤ T , then T − and T + are KB−operators. By Proposition 2.9, from [9], T is KB−operators. Similar argument holds for b−weakly compact operators, and so by using proposition 2.1, proof holds. (cid:3) Theorem 2.5 shows that the modulus of the operator T in example of page 3 is not KB−operator or b−weakly compact operator. Note that each weakly compact operator is a KB-operator but the converse may be false in general. For example, the identity operator I : L1[0, 1] → L1[0, 1] is a KB-operator but is not weakly compact. Let E and F be two Banach lattices such that the norm of E′ is order continuous. Then, by using Proposition 2.1 and [8, Theorem 2.3], it is clear that each positive KB-operator T : E → F is weakly compact. Now in the following, we show that B(E, F ) = LKB(E, F ) whenever E has order unit and order continuous norm. Theorem 2.6. Let E be a normed vectoe lattice and F Dedekind Banach Lattice. By one of the following conditions B(E, F ) = LKB(E, F ) (1) E has order unit and order continuous norm. (2) E and E′ have order continuous norm and F has Schur property Proof. (1) Let T ∈ B(E, F ) and xn ↑ and supn kxnk < ∞. Let e ∈ E+ be an order unit for E. For each x ∈ E, the norm kxk∞ = inf{λ > 0 : x ≤ λe} 6 KAZEM HAGHNEJAD AZAR is equivalent to the original norm k.k of E. It follows that sup n kxnk∞ = sup n kxnk. For each n ∈ N, there exists λn > 0 such that λn ≤ kxnk∞ + 1 where xn ≤ λne. It follows that 0 < λ = supλn ≤ sup n kxnk∞ + 1 < sup n kxnk + 1 < ∞. Then {xn} ⊆ [−λe, λe]. Since E has order continuous norm, by using theorem 4.9 from [1], it follows that [−λe, λe] is weakly compact. On the other hand, since E has order unit, T is order bounded, and so T + exists. Thus T +([−λe, λe]) is weakly compact, since T + is weak to weak continuous. As {T +xn} ⊆ T +([−λe, λe]), there is subsequence {T +xnj } from {T +xn} which is weak convergent to some point z ∈ F . Since (T +xn)n is an increasing sequence, then by Theorem 1.4.1 from [12], T +xn is norm convergent to z in F . Thus T + ∈ LKB(E, F ). By using equality T = T + − T −, we conclude that T ∈ LKB(E, F ). (2) Let T ∈ B(E, F ) and (xn) ⊆ E be increasing positive sequence where sup kxnk < +∞. n By Theorem 4.25 from [1], there is a weakly Cauchy subsequence {xnj } in E. It follows that {T xnj } is weakly Cauchy in F . Since F has Schur property, {T xnj } is norm Cauchy in F , and so convergence in F . Thus T ∈ LKB(E, F ). (cid:3) Theorem 2.7. Assume that an order bounded operator T : E → F between two Banach lattices has preserves disjointness. Then T ∈ Wb(E, F ) if and only if T ∈ LKB(E, F ). Proof. By using Theorem 2.40 from [1], T is KB−operator if and only if T is KB−operator, and by proposition 2.1 from [9] it is equivalent to that T is b−weakly compact operator. Another using Theorem 2.40 [1] shows that T is b−weakly compact and proof follows. (cid:3) Let E and F be Banach lattices and let T ∈ LKB(E, F ). By Proposition 3.4 from [10], {T xn}n has an order convergence subsequence in F for every positive increasing sequence {xn}n in the closed unit ball BE of E. Now, in the following we introduce two classifications of operators which are generalization of order and strongly order continuous operators and we establish the relationships between them and positive KB−operators. Definition 2.8. Let E and F be vector lattice. L(1) is the collection of operators T ∈ Lb(E, F ), which xn T xnk o1−→ 0 (resp. T xnk o2−→ 0) whenever {xnk } is a subsequence of {xn}. c (E, F ) (resp. L(2) o1−→ 0 (xn c (E, F )) o2−→ 0) implies In [2], there are some examples which shows that two classifications of operators c (E, F ) and L(2) c (E, F ) are different. L(1) A GENERALIZATION OF b-WEAKLY COMPACT OPERATORS 7 Theorem 2.9. Let E, F be a Banach lattices and E has order continuous norm. Then (1) Wb(E, F )+ = LKB(E, F )+ ⊆ L(2) (2) If LKB(E, F ) is vector lattice and F Dedekind complete, then LKB(E, F ) c (E, F ). is an ideal in L(1) c (E, F ) = L(2) c (E, F ). Proof. with P+∞ k=1 kxnj k < +∞. Define ym = Pm (1) Let T be a positive KB−operator and {xn} ⊂ E strongly or- o2−→ 0, der convergence in E. Without lose generality, we set 0 ≤ xn which follows {xn} is norm convergent to zero. Set {xnj } as subsequence j=1 xnj . Then 0 ≤ ym ↑ and supmkymk < ∞. Since T is KB−operator, {T ym} is norm convergent to some point z ∈ F . Now by [11], page 7, it has a subsequence as {T ymk} which is strongly order convergent to z ∈ F . Thus there is {zβ} ⊂ F + and that for each β there exists n0 T ymk − z ≤ zβ ↓ 0 whenever k ≥ n0. If we set n0 ≤ k′ ≤ k, then we the following inequalities 0 ≤ T xnmk ≤ T ymk − T ymk′ ≤ T ymk − z + T ymk′ − z ≤ zβ + zβ ↓ 0, shows that T ∈ L(2) c (E, F ) and proof immediately follows. (2) Since LKB(E, F ) is a vector lattice, by equality T = T +−T − and Theorem c (E, F ) and by part (1), LKB(E, F ) c (E, F ). By Proposition 2.9, from [9], it is also obvious 1.7 from [2], we have L(2) is a subspace of L(1) that LKB(E, F ) is an ideal in L(1) c (E, F ) = L(1) c (E, F ). (cid:3) Question. By conditions Theorem 2.9, Is LKB(E, F ) a band in L(1) L(2) c (E, F )? c (E, F ) = Definition 2.10. (1) An operator T : E → F between two normed vector lattice is unbounded b-weakly compact if {T xn} is un−convergent for every positive increasing sequence {xn}n in the closed unit ball BE of E. (2) An operator T : E → F between two normed vectoe lattice is unbounded KB-operator if {T xn} is un−convergent for every positive increasing se- quence {xn}n in the closed unit ball BE of E. For normed vector lattices E and F , the collection of unbounded KB−operators (resp. b−weakly compact operators) will be denoted by LU KB(E, F ) (resp. U Wb(E, F )). As example of page 3, two classifications of the above operators are different. If a Banach lattice F has strong unit, by using Theorem 2.3 [13], we have LKB(E, F ) = LU KB(E, F ) and Wb(E, F ) = U Wb(E, F ). It is clear that every KB−operators (or b−weakly compact operators) are unbounded KB−operators (or b−weakly compact operators), but the following example shows that the converse in general, not holds. 8 KAZEM HAGHNEJAD AZAR Example 2.11. Let Ic0 be an identity mapping from c0 into itself. Then Ic0 is an unbounded KB-operator and unbounded b-weakly compact operator. But Ic0 is not KB-operator or b-weakly compact operator. Let X be a Banach space, F a Banach lattice, and T ∈ L(X, F ). We say that T is (sequentially) un−compact, if for every bounded net {xα} (resp. {xn} its image has a subnet (resp. subsequence) which is un−convergent. The collections of un−compact (resp. sequentially un−compact) will be denoted by KU(X, F ) (resp. K σU(X, F )), that is KU(X, F ) = {T : X → F T is un − convergent}, K σU(X, F ) = {T : X → F T is sequentially un − convergent}. It is clear that K σU(E, F ) ⊆ LU KB(E, F ) where E and F are Banach lattices. As example of page 3, this inclusion may be proper. Theorem 2.12. Let I be an ideal in Banach lattice E and T ∈ B(E, F ), where F is Banach lattice. If T ∈ LKB(I, F ) is surjective homomorphism, then T ∈ LU KB(E, F ). Proof. Let {xn} be a positive increasing sequence in E with supnkxnk < ∞ and let x ∈ I. Then xn ∧ x ∈ I, 0 ≤ xn ∧ x ↑ and supnkxn ∧ xk ≤ kxnk < ∞. Since T ∈ LKB(I, F ), there is a subsequence {xnj } which {T (xnj ∧ x)} is convergent for each x ∈ I. As T is homomorphism and surjective, {T (xnj ) ∧ y} is convergent for all y ∈ F and proof follows. (cid:3) Theorem 2.13. Let E be a KB−space and T ∈ B(E, F ) be an surjective homo- morphism, where F is Banach lattice. Then T ∈ LU KB(E, F ) Proof. Let {xn} ⊂ E+ be increasing sequence. Since for each x ∈ E+ supkxn ∧ xk < ∞, it follows {xn ∧ x} is norm convergent, and so {T (xn ∧ x)} is norm convergent for each x ∈ E+. As T is surjective homomorphism, proof follows. (cid:3) Theorem 2.14. Let E has order continuous norm and atomic and F normed lattice. Then WKB(E, F ) ⊆ LU KB(E, F ). Proof. Let {xn} ⊂ E+ be increasing sequence and supkxnk < ∞. If T ∈ WKB(E, F ), there is a subsequence {xnj } which {T xnj } is weak convergent to some point y ∈ F . By Proposition 6.2, from [10], {T xnj } is unbounded norm convergent to y. Thus T ∈ LU KB(E, F ). (cid:3) Theorem 2.15. Assume that E and F are Banach lattices and F has order continuous norm. Then (1) WKB(E, F )+ ⊆ LU KB(E, F ). (2) If T : E → F is surjective lattice homomorphism, then T ∈ LU KB(E, F ). Proof. Let {xn} ⊂ E+ be increasing sequence and supkxnk < ∞. Then (1) if T ∈ WKB(E, F )+, {T xn} is weakly convergence in F . By Proposition 6.3 from [10], {T xn} is unbounded norm convergence in F , and so T ∈ LU KB(E, F ). A GENERALIZATION OF b-WEAKLY COMPACT OPERATORS 9 (2) let x ∈ E+. Set yn = xn ∧ x, which follows that yn ↑≤ x and supnkynk ≤ kxk. Since T is lattice homomorphism, T is positive, which follows T yn ↑≤ T x. By using Theorem 4.11 [1], {T yn} norm Cauchy, and so norm conver- gence in F . On the other hand, T (xn ∧ x) = T xn ∧ T x is norm convergent and proof follows. (cid:3) In the preceeding theorem, if WKB(E, F ) is a vector lattice, we conclude that WKB(E, F ) is a subspace of LU KB(E, F ) whenever E and F are Banach lattices. On the other hand, if F ′ has order continuous norm, then by using Theorem 6.4 [10], LU KB(E, F )+ ⊆ WKB(E, F ) whenever E and F are Banach lattices. References 1. C. D. Aliprantis and O. Burkinshaw, Positive Operators, Springer, Berlin 2006. 2. Y. Abramovich and G. Sirotkin, On order convergence of nets, Positivity 9 (2005), 287-292. 3. S. Alpay and B. Altin, C. Tonyali, On property (b) of vector lattices, Positivity 7 (2003), 135-139. 4. S. Alpay and B. Altin, A note on b-weakly compact operators, Positivity 11 (2007), 575-582. 5. B. Altin, On b-weakly compact operators on Banach lattices, Taiwanese J. Math. 11, (2007), 143 -- 150. 6. B. Aqzzouz, M. Moussa and J. Hmichane, Some Characterizations of b-weakly compact operators on Banach lattices, Math. Reports, 62 (2010), 315-324. 7. B. Aqzzouz, A. Elbour and J. Hmichane, The duality problem for the class of b-weakly compact operators, Positivity 13 (2009), 683-692. 8. B. Aqzzouz and A. Elbour, On the weak compactness of b-weakly compact operators, Posi- tivity, (2010), 75-81. 9. A. Bahramnezhad and K. Haghnejad Azar, K B-operators on Banach lattices and their relationships with Dunford-Pettis and order weakly compact operators, U. P. F. Sci. Bull. 80 (2018), 91-99. 10. Y. Deng, M. O,Brien and V. G. Troitsky, Unbounded norm convergence in Banach lattices, Positivity 21 (2017), 963-974. 11. N. Gao and F. Xanthos, Unbounded order convergence and application to martingales with- out probability, J. Math. Anal. Appl. 415 (2014), 931-947. 12. P. Meyer-Nieberg, Banach lattices, Universitex. Springer, Berlin. MR1128093, 1991. 13. M. Kandic, M. A. A. Marabeh and V. G. Troitsky, Unbounded norm topologies in Banach lattices, J. Math. Anal. Appl. 451 (2017), 259-279. Department of Mathematics, University of Mohaghegh Ardabili, Ardabil, Iran. E-mail address: [email protected];
1109.5300	1	1109	2011-09-24T20:07:01	Strongly non embeddable metric spaces	[ "math.FA", "math.GN" ]	Enflo constructed a countable metric space that may not be uniformly embedded into any metric space of positive generalized roundness. Dranishnikov, Gong, Lafforgue and Yu modified Enflo's example to construct a locally finite metric space that may not be coarsely embedded into any Hilbert space. In this paper we meld these two examples into one simpler construction. The outcome is a locally finite metric space $(\mathfrak{Z}, \zeta)$ which is strongly non embeddable in the sense that it may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness. Moreover, we show that both types of embedding may be obstructed by a common recursive principle. It follows from our construction that any metric space which is Lipschitz universal for all locally finite metric spaces may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness. Our construction is then adapted to show that the group $\mathbb{Z}_\omega=\bigoplus_{\aleph_0}\mathbb{Z}$ admits a Cayley graph which may not be coarsely embedded into any metric space of non zero generalized roundness. Finally, for each $p \geq 0$ and each locally finite metric space $(Z,d)$, we prove the existence of a Lipschitz injection $f : Z \to \ell_{p}$.	math.FA	math	STRONGLY NON EMBEDDABLE METRIC SPACES CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON Abstract. Enflo [3] constructed a countable metric space that may not be uniformly embedded into any metric space of positive generalized roundness. Dranishnikov, Gong, Lafforgue and Yu [2] modified Enflo's example to con- struct a locally finite metric space that may not be coarsely embedded into any Hilbert space. In this paper we meld these two examples into one sim- pler construction. The outcome is a locally finite metric space (Z, ζ) which is strongly non embeddable in the sense that it may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness. Moreover, we show that both types of embedding may be obstructed by a common re- cursive principle. It follows from our construction that any metric space which is Lipschitz universal for all locally finite metric spaces may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness. Z admits Our construction is then adapted to show that the group Z a Cayley graph which may not be coarsely embedded into any metric space of non zero generalized roundness. Finally, for each p ≥ 0 and each locally finite metric space (Z, d), we prove the existence of a Lipschitz injection f : Z → ℓp. ω = Lℵ0 1. Introduction The notion of the generalized roundness of a metric space was introduced by Enflo [3] as a means to show that not every separable metric space may be uniformly embedded into a Hilbert space. The crux of Enflo's proof was the construction of a countable metric space that may not be uniformly embedded into any metric space of positive generalized roundness. Enflo then noted that Hilbert space has generalized roundness 2 in order to draw the desired conclusion. As in [3], the terms uniform embedding and generalized roundness are defined in the following way. Definition 1.1. A uniform embedding of a metric space (X, d) into a metric space (Y, ς) is an injection f : X → Y such that both f and f −1 are uniformly continuous, where the domain of definition of f −1 is its natural domain as a subspace of (Y, ς). Definition 1.2. The generalized roundness q(X) of a metric space (X, d) is the supremum of the set of all p ≥ 0 that satisfy the following property: For all integers n ≥ 2 and all choices of (not necessarily distinct) points x1, . . . , xn, y1, . . . , yn ∈ X, (1) d(xi, yj)p. X1≤i<j≤n(cid:8)d(xi, xj )p + d(yi, yj)p(cid:9) ≤ X1≤i,j≤n The configuration of points (X , Y ) that underlies (1), where X denotes x1, . . . , xn and Y denotes y, . . . , yn, is called a double-n-simplex. A pair of the form (xi, xj) or (yi, yj), where i 6= j, will be called an edge in (X , Y ). While a pair of the form (xi, yj), with no restriction on i and j, will be called a connecting line in (X , Y ). 2000 Mathematics Subject Classification. 46C05, 46T99. Key words and phrases. Locally finite metric space, coarse embedding, uniform embedding. 1 2 CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON Metric spaces of generalized roundness 0 may have very rigid uniform structures. For example, the metric space constructed in the proof of [3, Theorem 2.1] may not be uniformly embedded into any metric space of positive generalized roundness. Dranishnikov, Gong, Lafforgue and Yu [2, Proposition 6.3] modified Enflo's example to construct a metric space that may not be coarsely embedded into any Hilbert space. A simple check shows that this modified metric space also has generalized roundness 0. It follows from these examples that the quality of a metric space having generalized roundness 0 may induce unique properties on the given space, not only on the fine scale of uniform structure but also on the global scale of coarse geometry. The notion of a coarse embedding was introduced by Gromov [4]. Definition 1.3. A map f from a metric space (X, d) into a metric space (Y, ς) is a coarse embedding if there exist two non decreasing functions ρ1, ρ2 : [0,∞) → [0,∞) such that (1) ρ1(d(x, y)) ≤ ς(f (x), f (y)) ≤ ρ2(d(x, y)) for all x, y ∈ X, and (2) limt→∞ ρ1(t) = ∞. Although the focus of this paper is squarely on uniform and coarse embeddings we will have the occasion to refer to Lipschitz embeddings. A Lipschitz embedding of a metric space (X, d) into a metric space (Y, ς) is an injection f : X → Y such that both f and f −1 are Lipschitz maps of order 1, where the domain of definition of f −1 is its natural domain as a subspace of (Y, ς). Our interest in Lipschitz embeddings stems from their quality of being both uniform and coarse embeddings. We further recall that a metric space (Y, ς) is Lipschitz universal for a class U of metric spaces if every (X, d) ∈ U admits a Lipschitz embedding into Y . In this paper we meld the examples of Enflo [3, Theorem 2.1] and Dranishnikov et al. [2, Proposition 6.3] into one simpler construction. The outcome is a locally finite metric space (Z, ζ) which is strongly non embeddable in the sense that it may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness (Theorems 3.1 and 4.1). Our approach shows that both types of embedding may be obstructed by a common recursive principle (Theorem 2.5). It follows that any metric space which is Lipschitz universal for all locally finite metric spaces is neither uniformly nor coarsely embeddable into any metric space of non zero generalized roundness (Corollary 4.7). The main construction is then adapted to show that the group Zω = Lℵ0 Z admits a Cayley graph which may not be coarsely embedded into any metric space of non zero generalized roundness. Remark 4.2 shows that it is not possible to construct an analogue of the metric space (Z, ζ) that has positive generalized roundness. Finally, for each p ≥ 0 and each locally finite metric space (Z, d), we prove the existence of a Lipschitz injection Z → ℓp (Theorem 4.5). This indicates that the two types of embeddings under consideration cannot be weakened if one wants to maintain an obstruction to embedding into metric spaces of non zero generalized roundness. To close out this section, we note that examples of metric spaces of non zero generalized roundness include all real normed spaces that are linearly isometric to a subspace of some Lp-space with 0 < p ≤ 2, all metric trees, all ultrametric spaces, all Hadamard manifolds, hyperbolic spaces such as Hk C, and many CAT(0)- spaces. By way of comparison, provided p > 2, Lp-spaces of dimension at least 3 are examples of metric spaces of generalized roundness 0. The same is true of certain CAT(0)-spaces as well as the Schatten p-classes Cp provided 0 < p < 2. The bulk of these examples are discussed more expansively in the survey paper [10]. STRONGLY NON EMBEDDABLE METRIC SPACES 3 2. The Main Construction Let Ln be the set {kn−n : 1 ≤ k ≤ n2n} endowed with the cyclic metric χ(x, y) = Ln endowed with min{x − y, nn − x − y}. Let Mn be the cartesian product Πnn the supremum metric: i=1 d∞(x, y) = max 1≤i≤nn χ(xi, yi), where in this case x = (x1,··· , xnn ) and y = (y1,··· , ynn ). Definition 2.1. Let 0 ≤ t < n and 0 < m < n. If x = (xi) and y = (yi), then we call {x, y} ⊂ Mn a (t, m)-pair if the following two conditions hold: (1) χ(xi, yi) = 2t whenever xi 6= yi, and (2) the cardinality of the set {i : xi 6= yi} is nm. We denote the set of (t, m)-pairs by Θm t . Definition 2.2. If (X , Y ) is a double-n-simplex in Mn and 0 < m < n, 0 ≤ t < n, then we call (X , Y ) a (t, m)-simplex if each connecting line is a (t, m + 1)-pair, and each edge is a (t + 1, m)-pair. For the rest of the paper we will only consider the case where n is even. Also, note that the (t, m)-pairs {x, y} and {y, x} will be regarded as identical, as will be the double-n-simplices (X , Y ) and (Y , X ). Lemma 2.3. If 0 < m < n, t < n, there exists a (t, m)-simplex in Mn. Proof. Divide the nn coordinates into 3 groups, the first two of size nm+1/2 and the third of size nn− nm+1. All coordinates in the third group are 0 in both the xi's and the yi's. For the xi's, all coordinates in the second group are 2t, and the first group is divided up into n subgroups of size nm/2. As i varies, the ith subgroup is filled with 2t+1, and the other n − 1 subgroups are filled with 0. For the yi's, just switch the first and second groups. The xi's and yi's may be visualized in the following way: nm/2 xi = z 0,··· , 0 {z } yi = nm+1/2 { } z 2t,··· , 2t, nm+1/2 nm+1/2 , nm/2 nm/2 { } z 2t,··· , 2t, } { ,··· , 2t+1,··· , 2t+1 ,··· , 0,··· , 0 {z } {z } { ,··· , 2t+1,··· , 2t+1 ,··· , 0,··· , 0 , {z } } z 0,··· , 0 {z } } {z nm+1/2 nm/2 nm/2 nm/2 nn−nm+1 0,··· , 0!, z } { 0,··· , 0!. z } { nn−nm+1 It is easy to verify that each xi and yj all differ in exactly nm+1 coordinates by 2t, while each xi and xj differ in exactly nm coordinates by 2t+1. (cid:3) Lemma 2.4. The isometry group of Mn acts transitively on Θm t whenever m ≥ 1. Proof. First note that any rearrangement of coordinates is an isometry of Mn. Given (t, m)-pairs {x, y} and {u, v}, first rearrange the coordinates of x and y so that {i : xi = yi} = {i : ui = vi}. This can be done because of condition (2) in the definition of a (t, m)-pair. One can then rotate the coordinates of x and y indexed by {i : xi = yi} so that xi = yi ⇒ xi = ui. Given i such that xi 6= yi, flip xi and yi as necessary to ensure that (xi, yi) and (ui, vi) have the same orientation in Ln, and then rotate the ith coordinate in xi and yi so that xi = ui, yi = vi. This 4 CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON reflection can be carried out because nn is even. Clearly the operations performed above constitute an isometry of Mn. (cid:3) Let N(t,m) be the number of (t, m)-pairs in Mn. It is clear that there is a natural number L(t,m) such that any (t, m)-pair in Mn is a connecting line in L(t,m) different ways in (t, m − 1)-simplices. Similarly, there exists a natural number K(t,m) such that any (t, m)-pair is an edge in K(t,m) different ways in (t − 1, m)-simplices. Let S(t,m) be the number of distinct (t, m)-simplices. Any (t + 1, m)-pair may be embedded as an edge within a fixed (t, m)-simplex in n(n− 1) different ways, so S(t,m)n(n− 1) = N(t+1,m)K(t+1,m). Similarly, any (t, m + 1)-pair may be embedded as a connecting line within a fixed (m, t)-simplex in n2 different ways, so S(t,m)n2 = N(t,m+1)L(t,m+1). This yields the following expression: (2) L(t,m+1) K(t+1,m) =(cid:18) n n − 1(cid:19) N(t+1,m) N(t,m+1) , which holds whenever 0 < m < n − 1 and −n < t < n − 1. It is worth noting that N(t,m) depends on n as well as t and m. The context will make the value n clear. To simplify later expressions we will denote f (x) by xf and d(x, y)p by dp(x, y) for the rest of this paper. 1 Theorem 2.5. Let (B, d) be a metric space of finite generalized roundness q(B) and let f : Mn+2 → B be a function. Provided 0 ≤ p ≤ q(B) and −n ≤ t ≤ 0, the following inequality holds: N(t,n+1) X{x,y}∈Θn+1 In the event that q(B) = ∞, the above inequality holds for all p ≥ 0 and −n ≤ t ≤ 0. Proof. Consider any non negative real number p that does not exceed q(B). We first note that if 0 < m < n + 1 and if (X , Y ) is any (t, m)-simplex in Mn+2, then dp(xf , yf ) ≥ (cid:18)1 − N(t+n,1) X{u,v}∈Θ1 n + 2(cid:19)n dp(uf , vf ). t+n 1 1 t Xi,j dp(xf i , yf j ) ≥Xi<j dp(xf i , xf j ) +Xi<j dp(yf i , yf j ) because (1) holds for any exponent that does not exceed q(B) by [6, Corollary 2.5]. By adding up all such inequalities for a fixed m, 0 < m < n + 1, we obtain L(t,m+1) X{x,y}∈Θm+1 t dp(xf , yf ) ≥ K(t+1,m) X{u,v}∈Θm t+1 dp(uf , vf ). Applying (2), we see that t 1 1 1 n + 2(cid:19) dp(xf , yf ) ≥(cid:18)1 − N(t+1,m) X{u,v}∈Θm N(t,m+1) X{x,y}∈Θm+1 Start with m = n, and iterate this inequality n times to arrive at the desired result. The case where q(B) = ∞ follows from [6, Corollary 2.5] in the same way. (cid:3) We conclude this section by describing the locally finite metric space (Z, ζ) that will be the focus of the subsequent sections. Theorems 3.1 and 4.1 determine that this metric space may not be embedded uniformly or coarsely into any metric space of non zero generalized roundness. dp(uf , vf ). t+1 STRONGLY NON EMBEDDABLE METRIC SPACES 5 Definition 2.6. Let Z = `n∈2N Mn, with metric ζ such that ζ restricted to any Mn gives the original metric on Mn, and such that ζ(x, y) = 4m+n if x ∈ Mm and y ∈ Mn. The metric space (Z, ζ) is locally finite by the nature of its definition. 3. The Obstruction to Coarse Embeddings Theorem 3.1. The locally finite metric space (Z, ζ) does not coarsely embed into any metric space of non zero generalized roundness. Proof. We first suppose that there exists a metric space (B, d) with q(B) = p > 0 and a coarse embedding f : Z → B. By the definition of a coarse embedding, there exist non-decreasing functions ρ1, ρ2 : [0,∞) → [0,∞) such that ρ1(ζ(x, y)) ≤ d(xf , yf ) ≤ ρ2(ζ(x, y)) for all x, y ∈ Z, and such that limr→∞ ρ1(r) = ∞. Then select a positive even integer n together with a real number α such that ρ1(2n) ≥ αρ2(1), and αpe−1 > 1. Now consider Mn+2 ⊂ Z, and set t = 0. By Theorem 2.5, 1 1 N(n,1) X{u,v}∈Θ1 n dp(uf , vf ) N(0,n+1) X{x,y}∈Θn+1 0 1 n + 2(cid:19)n dp(xf , yf ) ≥(cid:18)1 − ≥ e−1ρp 1(2n) ≥ αpe−1ρp 2(1). On the other hand, we have that 1 N(0,n+1) X{x,y}∈Θn+1 0 dp(xf , yf ) ≤ ρp 2(1), which clearly contradicts αpe−1 > 1. In the case where q(B) = ∞, we may argue as above by using any positive real number p in place of q(B). This is due to the second assertion of Theorem 2.5. (cid:3) It is worth noting that if one is only interested in locally finite metric spaces that do not embed coarsely into Hilbert space, then there is a simpler construction than that of (Z, ζ). Let (Xn) be an infinite sequence of finite, connected, k-regular graphs which is an expanding family, and let X be the disjoint union of the Xn's endowed with a metric that induces the original metric on each Xn, then X does not coarsely embed into any Hilbert space. This example is due to Gromov [5]. In view of the emphasis that is often placed on locally finite metric spaces in coarse geometry it makes sense to isolate the following natural definition. Definition 3.2. A metric space (X, d) is said to be a universal coarse embedding space if every locally finite metric space coarsely embeds into X. Examples of universal coarse embedding spaces include the Banach space C[0, 1] and the Urysohn space U. We note that background information on the Urysohn space may be found in Pestov [8, Chapter 5]. Theorem 3.1 automatically implies that no universal coarse embedding space can have non zero generalized roundness. This mirrors the corresponding result for uniform embeddings [3, Theorem 2.1]. Corollary 3.3. No universal coarse embedding space has non zero generalized roundness. In particular, metric subspaces of Lp-spaces, where 0 < p ≤ 2, and ultrametric spaces are not universal coarse embedding spaces. 6 CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON We conclude this section by noting that there is an interesting way to adapt Theo- rem 3.1 to Cayley graphs of certain abelian groups. We provide one such example. Recall that if G is an abelian group, then for each generating set Σ, one can consider the Cayley graph Γ = Cay (G, Σ) of G with respect to Σ, endowed with the usual graph metric. In most cases, q(Γ) ≤ 1. Simply select g, h ∈ Σ such that g + h, g − h /∈ Σ, and consider the double-2-simplex (X , Y ) with X = {0, g}, Y = {h, g + h}. It is readily checked that the generalized roundness inequality for (X , Y ) holds only if p ≤ 1. Non zero distances in Cayley graphs are all at least one, so it is not possible for Mn to be isometric to a subset of any Cayley graph. However, if we define M∗ n to be the subset of Mn consisting of points with integer coordinates, then it is apparent that the proofs of Theorems 2.5 and 3.1 can be adapted to any space that contains isometric copies of M∗ n for arbitrarily large n. This leads to the following theorem. Theorem 3.4. There exists a generating set Σ for Zω = Lℵ0 Z such that the Cayley graph of Zω with respect to Σ contains isometric copies of M∗ n for arbitrarily large n. In particular, the Cayley graph Cay (Zω, Σ) does not coarsely embed into any metric space of non zero generalized roundness. n=1 An, where An = Znn Proof. We may rewrite Zω asL∞ . For each n, let Σn be the set of vectors contained in An whose coordinates are either 0 or ±(nn − 1). Let Σ beS∞ n=1 Σn together with all maps ω → {0,±1} of finite support. It is easy to see that the restriction of the graph metric on Cay (Zω, Σ) to Cay (An, An ∩ Σ) yields the usual metric on M∗ n. The proof of the second statement is a straightforward application of Theorem 3.1. (cid:3) Remark 3.5. In the case of obstructing coarse embeddings into Hilbert space, it is worth noting that the second statement of Theorem 3.4 is similar in spirit to an example of Nowak [7, Remark 5 (A)]. A feature of our approach is that it is more general and it uses a less sophisticated machinery. 4. The Obstruction to Uniform Embeddings The metric space (Z, ζ) constructed in Definition 2.6 has enough fine structure that it also obstructs uniform embeddings into metric spaces of non zero generalized roundness. Once again, the obstruction is provided by Theorem 2.5. Theorem 4.1. The locally finite metric space (Z, ζ) does not uniformly embed into any metric space of non zero generalized roundness. Proof. We first suppose that there exists a metric space (B, d) with q(B) = p > 0 and a uniform embedding f of Z into B. If we set t = −n and consider Mn+2 ⊂ Z, then it follows from Theorem 2.5 that and so (3) sup {x,y}∈Θn+1 −n dp(xf , yf ) ≥ e−1 inf {u,v}∈Θ1 0 dp(uf , vf ), sup {x,y}∈Θn+1 −n d(xf , yf ) ≥ p√e−1 inf {u,v}∈Θ1 0 d(uf , vf ). However, ζ(u, v) ≥ 1 for all {u, v} ∈ Θ1 0 and f has a uniformly continuous inverse, d(uf , vf ) ≥ ε. The net effect is so there must exist an ε > 0 such that inf {u,v}∈Θ1 that the right side of (3) must be bounded away from zero, which is impossible. 0 STRONGLY NON EMBEDDABLE METRIC SPACES 7 This is because ζ(x, y) = 2−n for all {x, y} ∈ Θn+1 −n and f is uniformly continuous, providing a contradiction in the limit as n → ∞. In the case where q(B) = ∞, we may argue as above by using any positive real number p in place of q(B). This is due to the second assertion of Theorem 2.5. (cid:3) Remark 4.2. The metric space (Z, ζ) whose properties are examined in Theorems 3.1 and 4.1 has generalized roundness p = 0 together with ample circular symmetry. It is not possible to adapt this construction to positive values of p in any way whatsoever. No form of Theorem 3.1 or Theorem 4.1 exists for metric spaces of generalized roundness p > 0. In other words, for any metric space (X, d) of generalized roundness p > 0, there exists a metric space of generalized roundness q > p that is both uniformly and coarsely equivalent to (X, d). To see this we simply consider the metric transform ρ(x, y) = d(x, y)p/q. It is easy to see that (X, ρ) has generalized roundness q, and that the identity map on X is both a uniform and a coarse equivalence between (X, d) and (X, ρ). It is also worth noting that the uniform continuity of the inverse of f in the proof of Theorem 4.1 is absolutely necessary. We illustrate this by showing that every locally finite metric space admits a Lipschitz injection into ℓ0 and a uniformly continuous injection into ℓp for each p > 0. Theorem 4.3. Every locally finite metric space may be Lipschitz injected into ℓ0. Proof. Let (Z, d) be a locally finite metric space. We assume that Z > 1 to avoid a triviality. For each x ∈ Z, let xg = min{d(x, y) : x 6= y}. We may then define, for each n ∈ N, the (possibly empty) set An = {x ∈ Z : 21−n > xg}. Provided Z \ An 6= ∅, we may choose an injection hn : Z \ An → (0,∞) because Z \ An is at most countable. This allows us to define a map f : Z → ℓ0 : x 7→ (xf1 , xf2 , . . .) by xfn =(cid:26) 0 xhn if x ∈ An, otherwise. Notice that if x 6= z, then we may choose n ∈ N so that both x, z /∈ An. Then xhn 6= zhn because hn : Z \ An → (0,∞) is an injection. Thus f (x) 6= f (y), and so it follows that f : Z → ℓ0 is an injection. For x, y ∈ Z, we now define κ(x, y) = min{n ∈ N : 21−n ≤ max{xg, yg}}. Notice that, κ(x, y) ≤ min{κ(x, x), κ(y, y)} for all x, y ∈ Z. Moreover, if k = κ(x, x) and if n < k, then xfn = 0. Thus, given any x, y ∈ Z, we see that kxf − yfk0 = ∞ xfn − yfn · 2−n 1 + xfn − yfn xfn − yfn · 2−n 1 + xfn − yfn ∞ ∞ = Xn=1 Xn=κ(x,y) Xn=κ(x,y) = 21−κ(x,y) ≤ max{xg, yg} ≤ d(x, y), 2−n ≤ thereby establishing that f is a Lipschitz injection of order 1. (cid:3) 8 CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON Theorem 4.4. Let p > 0. Every locally finite metric space admits a uniformly continuous injection into ℓp. Proof. Let p > 0 and let (Z, d), g, hn, f and κ be as in the proof of Theorem 4.3, with the exception that xhn ∈ (0, 2−n/p) when 21−n ≤ xg. We then evidently have: kxf − yfkp p = xfn − yfnp ≤ ∞ Xn=κ(x,y) ∞ Xn=κ(x,y) 2−n/pp. (cid:3) The latter sum is 21−κ(x,y) as before, so we see that kxf − yfkp ≤ d(x, y)1/p. It turns out that Theorems 4.3 and 4.4 are special instances of a more general phenomenon concerning the existence of Lipschitz injections. Recall that a metric space (Z, d) is said to be discrete if, for each x ∈ Z, there exists a δ > 0 such that d(x, y) > δ for all y 6= x. For example, all locally finite metric spaces are discrete. Theorem 4.5. Let (Z, d) be a given discrete metric space and suppose that (Y, ρ) is any metric space that contains a nested sequence of open balls B1 ⊃ B2 ⊃ B3 ⊃ . . . such that (1) the diameter tn of Bn tends to 0 as n → ∞, and (2) Z ≤ Bn for all n ∈ N. Then, Z may be Lipschitz injected into Y . Proof. Suppose (Z, d) and (Y, ρ) satisfy the hypotheses of the theorem with Z > 1. For each x ∈ Z and t > 0, we define the functions xg = inf{d(x, y) : y 6= x} and th = min{n ∈ N : tn ≤ t}. By conditions (1) and (2) there must exist a function f : Z → Y that maps each x ∈ Z uniquely into B(xg )h. For any x, y ∈ Z, we then have xf , yf ∈ B(max{xg ,yg})h , implying that ρ(xf , yf ) ≤ max{xg, yg} ≤ d(x, y). Hence the function f is a Lipschitz injection of order 1. (cid:3) Remark 4.6. Theorem 4.5 is easily seen to imply Theorems 4.3 and 4.4. In fact, Theorem 4.5 implies a considerably stronger but less explicit version of Theorem 4.4: For each locally finite metric space (Z, d) and each p > 0, there exists a Lipschitz injection f : Z → ℓp. This is significant in the context of Theorem 3.1. For coarse embeddings the behavior of f −1 is governed by the condition on the function ρ1 that is described in Definition 1.3. The existence of the Lipschitz injections implied by Theorem 4.5, combined with the fact that ℓp has positive generalized roundness whenever 0 < p ≤ 2, makes it clear that we cannot dispense with the function ρ1 and expect to maintain the conclusion of Theorem 3.1 even if ρ2(t) = t for all t > 0. We further deduce from Theorem 4.5 that any locally finite metric space may be Lipschitz injected into any metric space that contains a Cauchy sequence of distinct elements. Indeed, any countable discrete metric space (Z, d) may be Lipschitz injected into any metric space that contains a Cauchy sequence of distinct elements. For if {xn} is such a Cauchy sequence in (Z, d) and m ∈ N, we may let Bm be any open ball of diameter ≤ 1/m that contains all but finitely many of the xn's. In conjunction with existing theory, Theorems 3.1 and 4.1 say something interesting about the Banach space c0. Recall that c0 consists of all real sequences that converge It was shown by Aharoni [1] that to zero endowed with the supremum norm. every separable metric space admits a Lipschitz embedding into c0. Thus, there exists a Lipschitz embedding (Z, ζ) → c0. Now let B be a metric space of non zero generalized roundness. Lipschitz embeddings are both uniform and coarse embeddings, so if there existed an embedding c0 → B that was either a uniform or STRONGLY NON EMBEDDABLE METRIC SPACES 9 coarse embedding, then there would exist such an embedding (Z, ζ) → B. However, this is not possible according to Theorems 3.1 and 4.1. More generally, these remarks apply to any metric space which is Lipschitz universal for all locally finite metric spaces and not just c0. Interesting metric spaces of this nature include the widely studied Urysohn space U. We therefore obtain the final result of this paper. We should note, however, that in the case of uniform embeddings, this result follows from [3, Theorem 2.1] by a minor modification. Corollary 4.7. The Banach space c0 is neither uniformly nor coarsely embeddable into any metric space of non zero generalized roundness. More generally, any met- ric space which is Lipschitz universal for all locally finite metric spaces is neither uniformly nor coarsely embeddable into any metric space of non zero generalized roundness. As already noted, the preceding corollary implies that the Urysohn space U is neither uniformly nor coarsely embeddable into any metric space of non zero gener- alized roundness. It is very interesting to compare this result to a theorem of Pestov [9]: The Urysohn space U is neither uniformly nor coarsely embeddable into any uniformly convex Banach space. Notice, for example, that the Banach space ℓ1 has non zero generalized roundness but it is not uniformly convex. On the other hand, the Banach space ℓ3 is uniformly convex but it does not have non zero generalized roundness. Acknowledgments The research presented in this paper was undertaken and completed at the 2011 Cornell University Summer Mathematics Institute (SMI). The authors would like to thank the Department of Mathematics and the Center for Applied Mathematics at Cornell University for supporting this project, and the National Science Foundation for its financial support of the SMI through NSF grant DMS-0739338. The last named author thanks Canisius College for a faculty summer research fellowship. References [1] I. Aharoni, Every separable metric space is Lipschitz equivalent to a subset of c+ 0 , Israel J. Math 19 (1974), 284 -- 196. 8 [2] A. N. Dranishnikov, G. Gong, V. Lafforgue and G. Yu, Uniform embeddings into Hilbert space and a question of Gromov, Can. Math. Bull. 45 (2002), 60 -- 70. 1, 2 [3] P. Enflo, On a problem of Smirnov, Ark. Mat. 8 (1969), 107 -- 109. 1, 2, 5, 9 [4] M. Gromov, Asymptotic invariants of infinite groups, Cambridge University Press (Cam- bridge), Proc. Symp. Sussex, 1991: II, London Math. Soc. Lecture Notes 182 (1993), vii+1 -- 295. 2 [5] M. Gromov, Spaces and questions, Geom. Funct. Anal. Special Volume (2000), 118 -- 161. 5 [6] C. J. Lennard, A. M. Tonge and A. Weston, Generalized roundness and negative type, Michigan Math. J. 44 (1997), 37 -- 45. 4 [7] P. Nowak, Coarse embeddings of metric spaces into Banach spaces, Proc. Amer. Math. Soc. 133 (2005), 2589 -- 2596. 6 [8] V. Pestov, Dynamics of Infinite-Dimensional Groups: The Ramsey-Dvoretzky-Milman Phe- nomenon, Univ. Lecture Ser., vol. 40, American Mathematical Society, 2006. 5 [9] V. Pestov, A theorem of Hrushovski -- Solecki -- Vershik applied to uniform and coarse embed- dings of the Urysohn metric space, Topology Appl. 155 (2008), 1561 -- 1575. 9 [10] E. Prassidis and A. Weston, Manifestations of non linear roundness in analysis, discrete geometry and topology. In: G. Arzhantseva and A. Valette (editors), Limits of Graphs in Group Theory and Computer Science, Research Proceedings of the ´Ecole Polytechnique F´ed´erale de Lausanne, CRC Press, Boca Raton (2009). 2 10 CASEY KELLEHER, DANIEL MILLER, TRENTON OSBORN, AND ANTHONY WESTON Department of Mathematics, California Polytechnic State University, San Luis Obispo, CA 93407 E-mail address: [email protected] Department of Mathematics, University of Nebraska at Omaha, Omaha, NE 68182 E-mail address: [email protected] Department of Mathematics, Baylor University, Waco, TX 76798 E-mail address: trenton [email protected] Department of Mathematics and Statistics, Canisius College, Buffalo, NY 14208 E-mail address: [email protected]
1603.00506	1	1603	2016-03-01T22:08:18	Asymptotic formulas for determinants of a special class of Toeplitz + Hankel matrices	[ "math.FA" ]	We compute the asymptotics of the determinants of certain $n\times n$ Toeplitz + Hankel matrices $T_n(a)+H_n(b)$ as $n\to\infty$ with symbols of Fisher-Hartwig type. More specifically we consider the case where $a$ has zeros and poles and where $b$ is related to $a$ in specific ways. Previous results of Deift, Its and Krasovsky dealt with the case where $a$ is even. We are generalizing this in a mild way to certain non-even symbols.	math.FA	math	Asymptotic formulas for determinants of a special class of Toeplitz + Hankel matrices ∗ Estelle L. Basor† Torsten Ehrhardt‡ American Institute of Mathematics Department of Mathematics Palo Alto, CA 94306, USA University of California Santa Cruz, CA 95064, USA Dedicated to Albrecht Bottcher on the occasion of his sixteith birthday Abstract We compute the asymptotics of the determinants of certain n× n Toeplitz + Hankel matrices Tn(a) + Hn(b) as n → ∞ with symbols of Fisher-Hartwig type. More specif- ically we consider the case where a has zeros and poles and where b is related to a in specific ways. Previous results of Deift, Its and Krasovsky dealt with the case where a is even. We are generalizing this in a mild way to certain non-even symbols. 1 Introduction For many recent applications, an asymptotic formula for determinants of the sum of finite Toeplitz and Hankel matrices has been of interest. For example, if we let a be in L1(T) and denote the kth Fourier coefficients of a by ak then understanding the behavior of det (aj−k + aj+k+1)j,k=0,...,n−1 as n → ∞ is important in random matrix theory. It has been shown in [5] that the above determinant behaves asymptotically like GnE with certain explicitly given constants G and ∗2010 MSC: 47B35; keywords: Toeplitz operator, Hankel operator, Toeplitz plus Hankel operator †[email protected] ‡[email protected] 1 E if a is a sufficiently well-behaved function. Such a result is an analogue of the classical Szego-Widom limit theorem [17] for Toeplitz determinants. The above determinant is a special case of more general determinants, det (aj−k + bj+k+1)j,k=0,...,n−1, where both a and b are in L1(T). We refer to the functions a and b as symbols. The goal is to find the asymptotics in the case of well-behaved a and b and also for the singular Fisher-Hartwig type symbols (symbols, with say, jump discontinuities or zeros). While an asymptotic formula in such a general case (with explicit description of the constants) is probably not doable, much recent progress has been made in some special cases. To be more precise a Fisher-Hartwig symbol is one of the form a(eiθ) = c(eiθ) RYr=1 vτr,αr (eiθ)uτr,βr(eiθ) (1) where c is a sufficiently well-behaved function (i.e., sufficiently smooth, nonvanishing, and with winding number zero) and for τ = eiφ, uτ,β(eiθ) = exp(iβ(θ − φ − π)), vτ,α(eiθ) = (2 − 2 cos(θ − φ))α. 0 < θ − φ < 2π, The symbol uτ,β has a jump at the point τ on the unit circle and the function vτ.α can be singular (say if α has negative real part,) or be zero at τ . We will generally refer to this last factor as having a singularity of "zero" type. Furthermore, αr and βr are complex parameters (where we assume Re αr > −1/2) and τ1, . . . , τR are distinct points on the unit circle T. In the case of smooth symbols, we cite the results in [7] where the case of det (aj−k + bj+k+1)j,k=0,...,n−1 with b(eiθ) = ±eiℓθa(eiθ) and ℓ fixed is considered. It is worth mentioning that among those cases, there are four special cases of particular interest, (i) b(eiθ) = a(eiθ), (ii) b(eiθ) = −a(eiθ), (iii) b(eiθ) = eiθa(eiθ), (iv) b(eiθ) = −e−iθa(eiθ), 2 in which the asymptotics have the form GnE with non-zero E. In the case of even symbols, i.e., a(eiθ) = a(e−iθ), these four cases are also related to the random matrices taken from the classical groups [2, 14]. Furthermore, these Toeplitz+Hankel determinants are expressable as Hankel determinants as well. In the case of b = a, two earlier papers of the authors consider the case of jump discontinuities [5, 6]. Furthermore, in the above four cases where in addition a is even, the results of Deift, Its, and Krasovsky [12] are quite complete and impressive. They allow quite general Fisher-Hartwig symbols with both zeros and jumps. In [12] the asymptotics are of the form Gn np E where G, p, and E are explicitly given constants. However, none of the earlier mentioned papers cover the case where the symbol is allowed to be non-even and with singularities of the zero type. So this is the focus of this paper, a non-even symbol with certain specified types of Fisher-Hartwig symbols. We prove an asymptotic formula of the same form as above. This is a step in the ultimate goal of asymptotics for non-even symbols with general Fisher-Hartwig singularitites. In order to briefly sketch the main ideas of the paper, let T (a) and H(b) stand for the Toeplitz and Hankel operators with symbols a and b acting on ℓ2, and let Pn stand for the finite section projection on ℓ2. The precise definition of these operators will be given in the next section. The above determinants can be understood as the determinant of Pn(T (a) + H(b))Pn for certain symbols a, b of Fisher-Hartwig type. Since we will allow not only for jumps, but also for zeros and poles the underlying operator (or its inverse) is generally not bounded. Hence the first step is to reformulate the problem as one which involves only bounded operators. This will be done by establishing an identity of the kind det Pn(T (a) + H(b))Pn = det(cid:16)PnT −1(ψ)(T (c) + H(cdφ))T −1(ψ−1)Pn(cid:17) (2) where the functions c and d are smooth and φ and ψ have only jump discontinuities. The next major step is a separation theorem, which allows to "remove" the smooth functions c and d, i.e., det(cid:16)PnT −1(ψ)(T (c) + H(cdφ))T −1(ψ−1)Pn(cid:17) det(cid:16)PnT −1(ψ)(I + H(φ))T −1(ψ−1)Pn(cid:17) ∼ E · Gn, n → ∞, with explicit constant E and G. Then by using the first identity again we relate the last determinant back to a Toeplitz+Hankel determinant, det Pn(T (a0) + H(b0))Pn = det(cid:16)PnT −1(ψ)(I + H(φ))T −1(ψ−1)Pn(cid:17) , 3 where, as it turns out, a0 and b0 are Fisher-Hartwig symbols which in their product do not have the smooth part. Of course, the whole procedure is only as useful as far as we are able to obtain the asymptotics of det Pn(T (a0) + H(b0))Pn. Here we apply the results of Deift, Its, and Krasovsky mentioned above [12] to identify this asymptotics in four special cases. The relation between our symbols a and b, and the symbols a0 and b0 to which we apply [12], is in fact given by a = c a0, b = c d b0, where a0 is even and of Fisher-Hartwig type and b0 relates to a0 as in (i) -- (iv). The functions c and d, while required to be sufficiently well-behaved, do not have to be even. The function d is required to satisfy d(eiθ)d(e−iθ) = 1, d(±1) = 1. In this sense, our results generalize some of the results of [12]. Note that in the most general case (for which we are able to do the separation theorem), the asymptotics of the corresponding det Pn(T (a0) + H(b0))Pn is not known. The idea for establishing an identity of the kind (2) and proving a separation theorem is due to Bottcher and Silbermann. In 1985 they proved the Fisher-Hartwig conjecture for symbols with small parameters (i.e, symbols (1) with Re αr < 1/2 and Re βr < 1/2), which was considered a major breakthrough at the time (see [10] and [11, Sect. 10.10]). Although they considered bounded invertible operators acting between different weighted L2-spaces, the essential point of their analysis can be expressed as identity of the kind det Tn(a) = det PnT −1(ψ)T (φ)T −1(ψ−1)Pn , where a is a Fisher-Hartwig symbol with jumps and zeros/poles, while φ and ψ are Fisher- Hartwig symbols with jumps only (the corresponding Toeplitz operators being bounded and invertible under certain conditions). Here is an outline of the paper. We begin with the operator theoretic preliminaries. This is done is section 2. In section 3, we reformulate the problem so that we can consider determinants of bounded operators only. In section 4, some additional operator theoretic results are given that are particularly useful for our situation. We then prove, in section 5, a "separation" theorem, that is, a theorem that allows us to compute the asymptotics from a combination of the smooth symbols and some specific cases of singular symbols where the results can be computed by other means. This is done in more generality than is needed for our final results, but it may prove to be useful in the future if other specific cases of singular symbols are obtained. Section 6 is devoted to infinite determinant computations that are required to describe constants explicitly and the next section contains the known results for the specific known singular symbols. Everything is collected in section 8 where the final asymptotics are com- puted. 4 Finally, the last section contains some additional results. In the course of the compu- tations for the main results of this paper, we discovered that the inverse of certain Identity plus Hankel operators had inverses that could be described using Toeplitz operators, their inverses and Hankel operators. So the inverse expressions may be of independent interest and are also included. 2 Preliminaries We denote by ℓ2 the space of all complex-valued sequences {xn}∞ n=0 quipped with usual 2- norm. The set L(ℓ2) is the set of bounded operators on ℓ2 and C1(ℓ2) is the set of trace class operators on ℓ2. The Toeplitz operator T (a) and Hankel operator H(a) with symbol a ∈ L∞(T) are the bounded linear operators defined on ℓ2 with matrix representations and T (a) = (aj−k), 0 ≤ j, k < ∞, H(a) = (aj+k+1), 0 ≤ j, k < ∞. It is well-known and not difficult to prove that Toeplitz and Hankel operators satisfy the fundamental identities (3) (4) T (ab) = T (a)T (b) + H(a)H(b) and H(ab) = T (a)H(b) + H(a)T (b). In the last two identities and throughout the paper we are using the notation It is worthwhile to point out that these identities imply that b(eiθ) := b(e−iθ). T (abc) = T (a)T (b)T (c), H(abc) = T (a)H(b)T (c) (5) for a, b, c, ∈ L∞(T) if an = c−n = 0 for all n > 0. We define the (finite section) projection Pn by Pn : {xk}∞ k=0 ∈ ℓ2 7→ {yk}∞ k=0 ∈ ℓ2, yk =(cid:26) xk 0 if k < n if k ≥ n . Using Pn we can view our determinants of interest as determinants of truncations of infinite matrices, det(Tn(a) + Hn(b)) = Pn(T (a) + H(b))Pn. 5 For bounded a and b this is the truncation of a sum of bounded operators, but even more generally for a, b ∈ L1(T) providing we view the operators as being defined on the space of sequences with only a finite number of non-zero terms. In the next sections we will be mostly concerned with functions a that are products of continuous functions times those with certain specific types of singularities. It will be convenient for the continuous function factors to satisfy certain properties. To describe this, we consider the Banach algebra called the Besov class B1 1. This is the algebra of all functions a defined on the unit circle for which kakB1 1 :=Z π −π 1 y2Z π −π a(eix+iy) + a(eix−iy) − 2a(eix) dx dy < ∞. A function a is in B1 Moreover, the Riesz projection is bounded on B1 1 if and only if the Hankel operators H(a) and H(a) are both trace class. 1, and an equivalent norm is given by where kAkC1 is the trace norm of the operator A. a0 + kH(a)kC1 + kH(a)kC1, Let us also recall the notion of Wiener-Hopf factorization. There are several versions of it. We say that c ∈ L∞(T) has a bounded (canonical) factorization if we can write c = c−c+ with c+, c−1 + (T) and c−, c−1 − (T), where + ∈ H ∞ − ∈ H ∞ H ∞ ± (T) = {f ∈ L∞(T) : fn = 0 for all ± n < 0 }. + ∈ H ∞ + (T) ∩ B1 1 has a canonical factorization in B1 We say that c ∈ B1 1 if we can write c = c−c+ with c+, c−1 − (T) ∩ B1 It is well known (see, e.g., [11, Sect. 1. 10.24]) that c admits a canonical factorization in B1 1 if and only if the function c does not vanish on T and has winding number zero. In this case, the logarithm exists, log c ∈ B1 1, and one can define normalized factors, 1 and c−, c−1 − ∈ H ∞ c±(t) = exp ∞Xk=1 t±k[log c]±k! , which yield a factorization c = c−G[c]c+ with the constant G[c] := exp([log c]0) representing the geometric mean. For our purposes it is also important to consider a factorization of the kind d = d−1 + d+ with d+, d−1 + ∈ B1 1 ∩ H ∞ + (T), 6 (6) (7) (8) in which the "minus" factor d−1 + is given by the "plus" factor d+. It is not too difficult to show (using the above result and the uniqueness of factorization up to multiplicative constants) that d ∈ B1 1 possesses a factorization of the above kind if and only if d does not vanish on T, has winding number zero and satisfies the conditions d d = 1 and d(±1) = 1. Notice that in this case log d ∈ B1 1 is an odd function and thus G[d] = 1. 3 Reformulating the problem As described in the introduction, we are interested in determinants of Toeplitz plus Hankel matrices with singular symbols. Let us denote the corresponding (infinite) operator by M(a, b) := T (a) + H(b). Notice that when the symbols involve zeros or poles, then either M(a, b) or its inverse are in general not bounded operators anymore. The purpose of this section is to reformulate the problem about the asymptotics of det(PnM(a, b)Pn) as one for det(PnAPn) where A is a bounded (and invertible) operator on ℓ2. More precisely, we are going to prove a formula det PnM(a, b)Pn = det PnT −1(ψ)M(c, c d φ)T −1(ψ−1)Pn, (9) where a and b are certain functions of Fisher-Hartwig type (allowing in particular for zeros and jumps) while on the right hand side ψ and φ are functions with jump discontiniuties only and with ranges of parameters such that T (ψ) and T (ψ−1) are invertible Toeplitz operators. The functions c and d are smooth and nonvanishing functions with winding number zero. Since the above formula involves inverses of Toeplitz operators, let us first recall a well-known sufficient invertibility criterion (see, e.g., [16] or [11]). Theorem 3.1 Let c be a continuous and nonvanishing function on T with winding number zero, let τ1, . . . , τR ∈ T be distinct, and ψ(eiθ) = c(eiθ) uτr,βr(eiθ) . RYr=1 If Re βr < 1/2 for all 1 ≤ r ≤ R, then T (ψ) is invertible on ℓ2. Let us now introduce the functions for which identity (9) will be proved. For these functions the separation theorem will be proved later on as indicated in the introduction. 7 We consider a = c v1,α+ v−1,α− RYr=1 vτr,α+ r v¯τr,α− r , b = c d v1,α+ u1,β+ v−1,α− u−1,β− vτr,αruτr,βr v¯τr,αru¯τr,βr . RYr=1 (10) (11) The functions c and d are smooth nonvanishing functions with winding number zero. In addition, we will require that d d = 1 and d(±1) = 1. We also assume that τ1, . . . , τR ∈ T+ are distinct, where and that T+ := { t ∈ T : Im (t) > 0 } , α±, β±, α± r , βr are complex parameters satisfying the conditions (16) and (17) stated below, whereas αr := r + α− α+ r 2 for 1 ≤ r ≤ R. The functions ψ and φ that will appear in the identity are ψ = u1,α+ u−1,α− φ = u1,γ+ u−1,γ− r u¯τr,α− r , uτr,γr u¯τr,γr , uτr,α+ RYr=1 RYr=1 where γ± := α± + β± , γr := αr + βr . The restrictions which we are going to impose on the parameters are the following: which guarantee the invertibility of T (ψ) and T (ψ−1), and Re α± < 1/2, Re α± r < 1/2, (12) (13) (14) (15) (16) −3/2 < Re γ+ < 1/2, −1/2 < Re γ− < 3/2, Re γr < 1/2. (17) The last conditions are needed later on. Theorem 3.2 Let a, b, c, d, φ, ψ be as above with (16) being assumed. Then det PnM(a, b)Pn = det(cid:16)PnT −1(ψ)M(c, c d φ)T −1(ψ−1)Pn(cid:17). 8 Proof. We first notice that a, b ∈ L1(T). Hence the PnM(a, b)Pn is a well-defined matrix, although M(a, b) may be an unbounded operator. The proof of the identity is based on In order to avoid the Wiener-Hopf factorization of the underlying generating functions. unbounded factors, let us assume for the time being that all the parameters α±, β±, α± r , βr are purely imaginary. The general case follows by observing that both sides of the identity are analytic in each of these parameters. In order to obtain the factorization introduce the functions ητ,γ(t) = (1 − t/τ )γ, ξτ,δ(t) = (1 − τ /t)δ, where the branches of η (analytic inside the unit circle) and ξ (analytic outside the unit circle) are chosen so that ητ,γ(0) = ξτ,δ(∞) = 1. Using the above definitions we can produce the well-known Wiener-Hopf factorizations for uτ,β = ξτ,−β ητ,β , vτ,α = ξτ,α ητ,α . Now put Then, indeed, ψ = ψ−ψ+. Furthermore, ψ−1 − ψ+ = v1,α+ v−1,α− ψ−1 − ψ+ = ξ1,2α+ ξ−1,2α− ψ− = ξ1,−α+ ξ−1,−α− ψ+ = η1,α+ η−1,α− r . r r , ητr,α+ η¯τr,α− ξτr,−α+ r ξ¯τr,−α− RYr=1 RYr=1 RYr=1 RYr=1 u1,−α+ u−1,−α− Y uτr,−αr u¯τr,−αr v¯τr,α− vτr ,α+ r r +α− r ξτr,α+ r +α− r ξ¯τr,α+ , r . Here notice that ητ,α = ξ¯τ ,α. The latter can be written as the product of v1,α+ v−1,α− Y vτr ,αr v¯τr,αr r )/2. Thus we see that r + α− as αr = (α+ and ψ = ψ−ψ+ , a = c ψ−1 − ψ+ , b = c d φ ψ−1 − ψ+ . It follows that det PnM(a, b)Pn = det PnM(ψ−1 − c ψ+, ψ−1 − c d φ ψ+)Pn 9 which equals det PnT (ψ−1 − )M(c, c d φ)T (ψ+)Pn by using (5). Also, notice that the determinants of PnT (ψ±)Pn and PnT (ψ−1 ± )Pn are one since they are either upper or lower triangular matrices with ones on the diagonal. Using this and the observation that PnT (ψ−1 + )Pn = PnT (ψ−1 + ), PnT (ψ−)Pn = T (ψ−)Pn, the above equals det PnT (ψ−1 + )T (ψ−1 − )M(c, c d ψ)T (ψ+)T (ψ−)Pn. Applying the following formulas for the inverses, T −1(ψ) = T (ψ−1 + )T (ψ−1 − ), T −1(ψ−1) = T (ψ+)T (ψ+), concludes the proof of the identity. ✷ It is interesting to consider certain special cases. What we have in mind is the case where the Fisher-Hartwig part of a (i.e., the product without the function c) is even. This happens if If in addition, we put βr = 0, then α+ r = α− r = αr . b = u1,β+ u−1,β− d a . There are four specific choices of parameters β± where the factor φ0 := u1,β+ uβ− is actually continuous: (1) β+ = β− = 0, φ0(t) = 1 ; (2) β+ = −1, β− = 1, φ0(t) = −1 ; (3) β+ = 0, β− = 1, φ0(t) = t ; (4) β+ = −1, β− = 0, φ0(t) = −1/t . Notice that the conditions (16) and (17) on the parameters α± and αr amount to the following To summarize, in these special cases we have Re α± < 1/2, Re αr < 1/2. a = c v1,α+ v−1,α− RYr=1 vτr,αr v¯τr,αr , b = φ0 d a. 10 (18) (19) Notice that the cases (1)-(4) correspond to the cases (i)-(iv) considered in the introduction, but are slightly more general due to the factor d. The reason why we single out these four special cases, is because for the computations that are made later in this paper, it is in these cases that we can actually determine the asymptotics, whereas in the more general case we can only reduce the asymptotics to a simplified determinant problem for which an answer is unknown. 4 Additional operator theoretic results We need some results about Toeplitz operators and Hankel operators (see [5] and [11] for the general theory). First of all, in addition to the projections Pn, and Qn = I − Pn we define Wn(f0, f1, . . . ) = (fn−1, fn−2, . . . , f1, f0, 0, 0, . . . ), Vn(f0, f1, . . . ) = (0, 0, . . . , 0, 0, f0, f1, f2, . . . ), V−n(f0, f1, . . . ) = (fn, fn+1, fn+2, . . . ). It is easily seen that W 2 also that n = Pn, Wn = WnPn = PnWn, VnV−n = Qn and V−nVn = I. Note PnT (a)Vn = WnH(a), V−nT (a)Pn = H(a)Wn. (20) Moreover, we have V−nT (a)Vn = T (a), V−nH(a) = H(a)Vn, WnT (a)Wn = PnT (a)Pn. (21) In the proofs that follow we will need the notions of stability and strong convergence and we describe those now. Let An be a sequence of operators. This sequence is said to be stable if there exists an n0 such that the operators An are invertible for each n ≥ n0 and supn≥n0kA−1 n k < ∞. Moreover, we say that An converges strongly on ℓ2 to an operator A as n → ∞ if Anx → Ax in the norm of ℓ2 for each x ∈ ℓ2. When dealing with finite matrices An, we identify the matrices and their inverses with operators acting on the image of Pn. It is well known (see [11, Th. 4.15] and worthy to note that stability is related to strong convergence of the inverses (and their adjoints) in the following sense. Lemma 4.1 Suppose that An is a stable sequence such that An → A and A∗ Then A is invertible, and A−1 n )∗ → (A−1)∗ strongly. n → A−1 and (A−1 n → A∗ strongly. Recall that for trace class operators, the trace "trace A" and the operator determinant "det(I + A)" are well defined and continuous with respect to A in the trace class norm. The following well known result shows the connection with strong convergence. 11 Lemma 4.2 Let B be a trace class operator and suppose that An and Cn are sequences such that An → A and C ∗ n → C ∗ strongly. Then AnBCn → ABC in the trace class norm. We can use the first lemma to obtain information about the strong convergence of the inverses of Toeplitz matrices. Proposition 4.3 Let φ ∈ L∞(T). If Tn(φ) is stable, then T (φ) is invertible and T −1(φ) = s-lim T −1 n (φ), T −1(eφ) = s-lim WnT −1 n (φ)Wn. Since P ∗ Proof. n = Pn → I strongly, it follows that Tn(φ) → T (φ) strongly and the same holds for the adjoints. Lemma 4.1 implies the first statement. For the second one, observe that WnTn(φ)Wn = Tn(eφ) and proceed similarly. Also note that T (eφ) is the transpose of T (φ), thus also invertible. ✷ We will need a new definition and additional results about strong convergence in what follows (see also [11, Thm. 7.13]). Let A equal the set of all bounded operators A defined on ℓ2 such that the operator (cid:18) Wn V−n (cid:19) A(cid:0) Wn Vn (cid:1) =(cid:18) WnAWn WnAVn V−nAWn V−nAVn (cid:19) along with its adjoint (which replaces A with A∗) converge strongly to operators defined on ℓ2 ⊕ ℓ2. In other words π(A) := s-lim(cid:18)(cid:18) Wn V−n (cid:19) A(cid:0) Wn Vn (cid:1)(cid:19) exists. Lemma 4.4 The set A is a (closed in the operator topology) C ∗-subalgebra of L(ℓ2), and the map π : A → L(ℓ2 ⊕ ℓ2) is a -homomorphism. Proof. It is easy to see that the sum, the product and the involution are closed operations in A and at the same time that π is a -homomorphism. Using that the norms of Wn and V±n are one, one can conclude the map π is bounded. The fact that A is closed can be shown straightforwardly using a Cauchy sequence argument (see also [11, Thm. 7.13]). ✷ We now relate Toeplitz and Hankel operators to A and π. 12 Lemma 4.5 For φ in L∞(T) the operators T (φ) and H(φ) belong to A. Moreover, H(φ) T (φ) (cid:19) , π(T (φ)) =(cid:18) T ( φ) H( φ) 0 0 (cid:19) . π(H(φ)) =(cid:18) 0 0 Proof. We consider first the Toeplitz operator. We use the identities WnT (φ)Wn = PnT ( φ)Pn, V−nT (φ)Wn = H(φ)Pn, WnT (φ)Vn = PnH( φ), V−nT (φ)Vn = T (φ), which we stated at the beginning of the section, to show the strong convergence. For the Hankel operator we consider WnH(φ)Wn = WnV−nT (φ)Pn, V−nH(φ)Wn = V−2nT (φ)Pn, WnH(φ)Vn = WnV−nH(φ), V−nH(φ)Vn = V−2nH(φ), and the strong convergence follows because V−n → 0 strongly. For the adjoints the argu- mentation is analogous. ✷ If we abbreviate then (3) and (4) imply that H(φ) T (φ) (cid:19) , L(φ) :=(cid:18) T ( φ) H( φ) L(φ1φ2) = L(φ1)L(φ2) . (22) This is not surprising since by an appropriate identification of ℓ2 ⊕ ℓ2 with ℓ2(Z) it is easily seen that L(φ) is the Laurent operator with symbol φ. The following result is what we will need in the next section. Notice that in the case of c = 1 we have that π(A) is the identity operator on ℓ2 ⊕ ℓ2. Proposition 4.6 Let A = T −1(ψ)(T (c) + H(φ))T −1(ψ−1) where c, φ, ψ ∈ L∞(T) are such that T (ψ±1) are invertible. Then A ∈ A and π(A) =(cid:18) T (c) H(c) H(c) T (c) (cid:19) . Proof. Since A is a C-algebra (hence inverse closed) and π is a -homomorphism, it follows that T −1(ψ) ∈ A and π(T −1(ψ)) = (π(T (ψ)))−1 =(cid:18) T ( ψ) H( ψ) H(ψ) T (ψ) (cid:19)−1 =(cid:18) T ( ψ−1) H( ψ−1) H(ψ−1) T (ψ−1) (cid:19) = L(ψ−1) . 13 The inversion of the operator matrix follows from (3) and (4). Similarly, we obtain π(T −1(ψ−1)) =(cid:18) T ( ψ) H( ψ) H(ψ) T (ψ) (cid:19) = L(ψ) and π(T (c) + H(φ)) = L(c). Using (22) we obtain π(A) = L(ψ−1)L(c)L(ψ) = L(c), which is the formula for π(A). ✷ 5 Separation theorems We now establish a separation theorem, which we are formulation in a quite general setting. Theorem 5.1 Let ψ, φ ∈ L∞(T) with φ φ = 1 be such that T (ψ) is invertible on ℓ2 and such that the sequence An = PnT −1(ψ)M(1, φ)T −1(ψ−1)Pn (23) is stable. Moreover, assume that c ∈ B1 that d ∈ B1 1 has a Wiener-Hopf factorization d = d+ d−1 1 is nonvanishing and has winding number zero and + in B1 1. Then lim n→∞ det(cid:16)PnT −1(ψ)M(c, cdφ)T −1(ψ−1)Pn(cid:17) G[c]n det(cid:16)Pn T −1(ψ)M(1, φ)T −1(ψ−1)Pn(cid:17) = E, + )(cid:17) × E = det(cid:16)T −1(c d+)T (c)T ( d+)(cid:17) × det(cid:16)T (cd+)T (c−1d−1 where G[c] = exp([log c]0) and det(cid:16)T −1(cd+)T −1(ψ)M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1)M −1(1, φ)T (ψ)(cid:17). (24) Proof. We note that the conditions on c and d+ imply the invertibility of T (cd+) and T (c d+) and the stability of Tn(cd+). Because T (ψ) is invertible (and hence ψ−1 ∈ L∞(T)) and one can conclude that T ( ψ−1) is invertible. Indeed, the formula T −1( ψ−1) = T ( ψ) − H( ψ)T −1(ψ)H(ψ). can be verified straightforwardly using (4) and (5). Note that T ( ψ−1) is the transpose of T (ψ−1), which thus is also invertible. Furthermore the stability of An implies the invertibil- ity of T −1(ψ)M(1, φ)T −1(ψ−1). Hence M(1, φ) is invertible. From the proof below it will follow that the operator determinants in (24) are well-defined, by which we mean that the underlying operator is identity plus a trace class operator. 14 We start by looking at M(c, cdφ) modulo trace class operators. It equals T (c) + H(cd+ d−1 + φ) = (T (cd+) + H(cd+φ))T (d−1 + ) =(cid:16)T (cd+) + T (cd+)H(φ) + H(cd+)T ( φ)(cid:17) T (d−1 = T (cd+)M(1, φ)T (d−1 + ) + trace class. + ) Hence modulo trace class, T −1(ψ)M(c, cdφ)T −1(ψ−1) equals T −1(ψ)T (cd+)M(1, φ)T (d−1 + )T −1(ψ−1). Since the commuators [T −1(ψ), T (cd+)] and [T (d−1 + ), T −1(ψ−1)] are trace class, it follows that T −1(ψ)M(c, cdφ)T −1(ψ−1) = T (cd+)T −1(ψ)M(1, φ)T −1(ψ−1)T (d−1 + ) + K1 with a certain trace class operator K1. Now multiply with Pn from the left and the right hand side and write PnT (cd+)AT (d−1 + )Pn = PnT (cd+)PnAPnT (d−1 + )Pn + PnT (cd+)QnAQnT (d−1 + )Pn+ PnT (cd+)QnAPnT (d−1 + )Pn + PnT (cd+)PnAQnT (d−1 + )Pn with A := T −1(ψ)M(1, φ)T −1(ψ−1). We analyse the last three terms. First, using Pn = W 2 n , Qn = VnV−n, we see that PnAQnT (d−1 + )Pn = Wn(cid:16)WnAVn(cid:17)H(d−1 + )Wn tends to zero in trace norm because WnAVn → 0 strongly (see Proposition 4.6). Secondly, PnT (cd+)QnAPn = WnH(c d+)(cid:16)V−nAWn(cid:17)Wn tends also to zero in trace norm because (V−nAWn)∗ → 0 strongly (again by Prop. 4.6). This implies that the last two terms of the above expressions tend to zero in trace norm. Finally, PnT (cd+)QnAQnT (d−1 + )Pn = WnH(c d+)(cid:16)V−nAVn(cid:17)H(d−1 + )Wn Here V−nAVn → I strongly (by Prop. 4.6). Hence the latter is WnH(c d+)H(d−1 sequence tending to zero in trace norm. + )Wn plus a Summarizing, we have so far Bn := PnT −1(ψ)M(c, cdφ)T −1(ψ−1)Pn = Tn(cd+)PnAPnTn(d−1 + ) + PnK1Pn + WnL1Wn + D(1) n 15 with K1 and L1 = H(c d+)H(d−1 n being a sequence tending to zero in trace norm. Now we take the inverses of Tn(cd+), PnAPn =: An, and Tn(d−1 + ) being trace class operators and D(1) + ). Thus, T −1 n (cd+)BnTn(d−1 + )A−1 n = Pn + T −1 n (cd+)PnK1PnTn(d−1 + )A−1 n + T −1 n (cd+)WnL1WnTn(d−1 + )A−1 n + D(2) n with D(2) the above sequences and their adjoints it follows that n → 0 in trace norm due to stability. Using stability and the strong convergence of T −1 n (cd+)PnK1PnTn(d−1 + )A−1 n = PnKPn + D(3) n and T −1 n (cd+)WnL1WnTn(d−1 + )A−1 n = WnLWn + D(4) n with D(j) limits of n → 0 in trace norm and K, L being trace class. In the latter we use that the strong WnTn(cd+)Wn, WnAWn = WnAnWn, WnTn(d−1 + )Wn (and their adjoints) exist. Indeed, apply Proposition 4.6. Also, due to stability their inverses have a strong limit. Thus T −1 n (cd+)BnT −1 n (d−1 + )A−1 n = Pn + PnKPn + WnLWn + Dn (25) with Dn → 0 in trace norm. Write Pn + PnKPn + WnLWn = (Pn + PnKPn)(Pn + WnLWn) − PnKWnLWn with the last term tending to zero in trace norm as Wn → 0 weakly. Now take determinants and it follows that lim n→∞ det Bn det Tn(cd+) · det An · det Tn(d−1 + ) = det(I + K) · det(I + L). From the standard Szego Limit theorem we get det Tn(cd+) ∼ G[cd+]n · det T (cd+)T (c−1d−1 + ), n → ∞, while det Tn(d−1 + ) = G[d−1 + ]n. Together we get the exponential factor G[c] = G[cd+] · G[d−1 + ]. It remains to identify trace class operators K and L. This is most conveniently done by passing to strong limits (and the strong limits after applying Wn from both sides) in (25). We obtain T −1(cd+)BT −1(d−1 + )A−1 = I + K 16 with B = T −1(ψ)M(c, cdφ)T −1(ψ−1) and A = T −1(ψ)M(1, φ)T −1(ψ−1), i.e., I + K = T −1(cd+)T −1(ψ)M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1)M(1, φ)−1T (ψ). This gives one of the operator determinant in (24). As for the Wn-limits we obtain T −1(c d+)T (c)T −1( d−1 + ) = I + L. Here notice that WnBnWn → T (c) and WnAnWn → I (again by Proposition 4.6). Thus, det(I + L) = det T −1(c d+)T (c)T ( d+), which is the remaining term in (24) along with the constant term from the Szego-Limit theorem above. ✷ In order to use the previous theorem we have to know the stability of the sequence An defined in (23). This is a non-trivial issue and is addressed in [9], where the following two theorems are proved. These results include certain "local" operators, which we are not going to define here, but instead refer to [9]. Theorem 5.2 Let φ and ψ be of the form (13) and (14). Assume that conditions (16) are satisfied. Then the sequence is stable if and only if the following conditions are satisfied An = Pn T −1(ψ)M(1, φ)T −1(ψ−1)Pn (i) the operator M(1, φ) is invertible on ℓ2, (ii) Re γ+ /∈ 2Z + 1/2 and Re γ− /∈ 2Z − 1/2, (iii) for each 1 ≤ r ≤ R, a certain the "local" operator B(α+ r , α− r , γr) is invertible. This theorem is proved in [9] by using general stability results of [15]. These gen- eral stability results imply that An is stable if and only if a certain collection of oper- ators is invertible. Among these operators is the strong limit of An, i.e., the operator A = T −1(ψ)M(1, φ)T −1(ψ−1). Thus it is necessary for stability that M(1, φ) is invertible. In addition, there occur "local" operators (associated to each point where ψ or φ have jump discontinuities). Invertibility of the local operators at t = ±1 lead to conditions (ii). For the jumps at t = τr and t = ¯τr the local operators are Mellin convolution operators B(α+ r , γr) with 2 × 2 matrix valued symbol defined in terms of the three parameters α+ r , γr ∈ C. As well known, the invertibility of such operators is equivalent to a Wiener-Hopf factorization of the matrix symbol, which in general is an unaccessible problem. Therefore, we have only the following results availabe [9]. On the positive side, part (c) covers the special cases we are particularly interested in. r , α− r , α− 17 Theorem 5.3 Let Re α± r /∈ Z + 1/2. (a) If B(α+ r , α− r , γr) is invertible, then Re γr /∈ Z + 1/2; (b) If Re γr /∈ Z + 1/2, then B(α+ r , α− r , γr) is Fredholm with index zero; (c) If αr = α+ r = α− r and Re γr /∈ Z + 1/2, then B(αr, αr, γr) is invertible. Let us return to the invertibility of the operator M(1, φ). For general Toeplitz+Hankel operators (with jump discontinuities) invertibility is a delicate issue. In [8] necessary and sufficient condtions for invertibility conditions were established for Toeplitz+Hankel operator T (a) + H(b) with piecewise continuous a, b satisfying the additional condition aa = bb. Since φ φ = 1 our operator M(1, φ) = I + H(φ) falls into this class and we cite the corresponding result (Corollary 5.5 of [8]). For sake of simple presentation we only state it in the form that provides us a sufficient condition. Note that guτ,γ = u¯τ ,−γ , which ensures φ φ = 1. Theorem 5.4 Let φ be of the form φ = u1,γ+ u−1,γ− uτr,γr u¯τr,γr , RYr=1 with distinct τ1, . . . , τR ∈ T+ and assume −3/2 < Re γ+ < 1/2, −1/2 < Re γ− < 3/2, −1/2 < Re γr < 1/2. Then M(1, φ) = I + H(φ) is invertible on ℓ2. We will make further remarks on the invertibility of M(1, φ) in Section 9. As a conclusion of the previous three results we can give some sufficient condition for stability. Notice that if (b) in Theorem 5.3 would imply invertibility (as is the case in (c)), we would not need the extra condition α+ r = α− r . Corollary 5.5 Let φ and ψ be of the form (13) and (14). Assume that conditions (16) and (17) are satisfied and that in addition αr = α+ r . Then An is stable. r = α− 6 Determinant computations In view of our separation theorem, we need to do two things. One is to evaluate the constant (24) and the other is to compute the asymptotics of the determinant of An. The goal of 18 this section is to do the first, that is, evaluate the constant (24) which is given in terms of operator determinants. Some of the factors have been computed before, and the complicated one can be reduced to simpler ones, which also have been computed before. We start with the following definitions and observations. For A, B ∈ L∞(T) for which H(A)H( B) is trace class and for which T (A) and T (B) are invertible, let E[A, B] = det(cid:16)T −1(A)T (AB)T −1(B)(cid:17). Note that H(A)H( B) is trace class if one of the functions is in B1 smooth (and have a continuous logarithm), then it has been shown that 1. If both functions are E[A, B] = exp(cid:16)trace H(log A)H(log B)(cid:17) = exp(cid:16)Xk≥1 k[log A]k[log B]−k(cid:17). From this formula it follows that E[A, B] = E[ B, A] = E[A+, B−], where A = A−A+ and B = B−B+ are Wiener-Hopf factorization of functions in B1 1. This constant is related to the constant E[C] = det T (C)T (C −1) = exp(cid:16)Xk≥1 k[log C]k[log C]−k(cid:17). appearing in the Szego-Widom limit theorem. In fact, we have E[C] = E[C, C] = E[C+, C−]. Finally, for a nonvanishing function C ∈ B1 1 with winding number zero let It was computed in [7] that F [C] = det(I + T −1(C)H(C)) F [C] = exp(cid:16) − = exp(cid:16) − 1 2 1 trace H(log C)2 + trace H(log C)(cid:17) [log C]2k−1(cid:17). 2Xk≥1 k +Xk≥1 k[log C]2 (26) The previous determinant relates to a slighty more complicated determinant. Lemma 6.1 Let C ∈ B1 has a bounded Wiener-Hopf factorization φ = φ− φ−1 − . Then 1 be nonvanishing and have winding number zero. Assume that φ det(I + H(C)T −1(Cφ)) = det(I + H(C)T −1(C)) = F [C]. 19 Proof. We have T (Cφ) = T (φ−)T (C)T ( φ−1 − ) with the factors being bounded invertible operators. Using H(C)T ( φ−) = H(Cφ−) = T (φ−)H(C) we obtain H(C)T −1(Cφ) = H(C)T ( φ−)T −1(C)T (φ−1 − ) = T (φ−)H(C)T −1(C)T (φ−1 − ), which gives the assertion. ✷ This next lemma illustrates some properties of the constant E[A, B] which will be used later to simplify determinants. Lemma 6.2 Let A, B, C ∈ L∞(T) be such that T (A) is invertible and B and C admit bounded factorizations. (a) If H(A)H( B) and H(A)H( C) are trace class, then E[A, BC] = E[A, B] · E[A, C]. (b) If H(B)H( A) and H(C)H( A) are trace class, then E[BC, A] = E[B, A] · E[C, A]. Proof. Let B = B−B+ and C = C−C+ be the bounded factorizations. Using (3) -- (5) in what follows, we remark that H(A)H( B−) = H(A)H( B)T (B−1 + ) and H(A)H( C−) = H(A)H( C)T (C −1 + ) are trace class. Analogously, H(A)H( B C) = H(A)H( B− C−)T (B+C+). Observe that H(A)H( B− C−) = H(A)H( B−)T (C−) + H(A)T ( B−)H( C−) = H(A)H( B−)T (C−) + T (B−)H(A)H( C−) where we used T (B−)H(A) = H(B−A) = H(AB−) = H(A)T ( B−). Therefore we conclude that H(A)H( B C) is trace class, too The Toeplitz operators T (B), T (C), and T (BC) are in- vertible due to the bounded factorizations. This implies that the three operator determinants are well-defined. A straightforward compuation using the factorizations yields that E[A, B] = det T −1(A)T (AB)T −1(B) = det T −1(A)T (AB−)T (B−1 − ) = E[A, B−] 20 and similar statements for the other two determinants. In fact, we can write E[A, B] = det T (B−1 − )T −1(A)T (AB−), E[A, C] = det T −1(A)T (AC−)T (C −1 − ) and multiplication yields det T (B−1 − )T −1(A)T (B−)T (A)T −1(A)T (AC−)T (C −1 − ) = det T (B−1 − )T −1(A)T (B−)T (AC−)T (C −1 − ) = det T −1(A)T (AB−C−)T (C −1 − B−1 − ), which is E[A, BC]. This proof (a). The proof of (b) is analogous. ✷ The next theorem states that the operator determinant occuring in (24) is well defined under certain conditions (invertibility of M(1, φ)), and that it can be expressed in terms of above constants in case of a slightly stronger condition (invertibility of T (φ)). Afterwards, when we specialize to the functions ψ and φ with jump discontinuities we will see that the stronger condition is redundant for the evaluation of the constant. Notice that we have already shown in Theorem 5.1 that operator determinant is well-defined, but under the (perhaps stronger) assumption of the stability of a certain sequence (which in fact implies invertibility of M(1, φ)). Theorem 6.3 Let ψ, φ ∈ L∞(T) with φ φ = 1 such that T (ψ) is invertible on ℓ2. Assume that c ∈ B1 1 has a factorization d = d+ d−1 1 is nonvanishing and has winding number zero and that d ∈ B1 + in B1 1. (a) If M(1, φ) is invertible on ℓ2, then the following operator determinant is well-defined: E1 = det(cid:16)T −1(cd+)T −1(ψ)M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1)M −1(1, φ)T (ψ)(cid:17). (b) If T (φ) is invertible on ℓ2, then M(1, φ) is invertible, and E1 = E[ψ, cd+] E[cd+, ψ] × E[d−1 + , ψ−1] × E[cd+, φ] × det(cid:16)(T (cd+φ) + H(cd+))T −1(cd+φ)(cid:17) . (c) If φ and ψ have a bounded factorization, then E1 = E[ψ, c] E[c, ψ] × E[cd+, φ] × F [cd+] . 21 Proof. (a): Abbreviate e = cd+ and write M(e, eφ) = T (e) + H(eφ) = T (e) + T (e)H(φ) + H(e)T ( φ) = T (e)M(1, φ) + K1 with K1 being trace class. This implies det M(cd+, cd+φ)M −1(1, φ)T −1(cd+) = det M −1(1, φ)T −1(cd+)M(cd+, cd+φ) is well-defined. Multiplying this determinant with the well-defined determinant E[d−1 + , ψ−1] = det T −1(d−1 = det T (d−1 + )T (ψ−1d−1 + )T −1(ψ−1) + )T −1(ψ−1)T (d+)T (ψ−1) and observing M(cd+, cd+φ)T (d−1 + ) = M(c, cdφ) yields the well-defined determinant det M −1(1, φ)T −1(cd+)M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1), which we can also write as det M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1)M −1(1, φ)T −1(cd+). Next observe that E[ψ, cd+] E[cd+, ψ] = det T (cd+)T (ψ)T −1(ψcd+) · det T (ψcd+)T −1(cd+)T −1(ψ). = det T (cd+)T (ψ)T −1(cd+)T −1(ψ). This well-defined determinant we multiply to the above one to obtain det M(c, cdφ)T −1(ψ−1)T (d+)T (ψ−1)M −1(1, φ)T (ψ)T −1(cd+)T −1(ψ), which is E1. Summarizing, besides the issue of E1 being well-defined we have shown that E1 = E[ψ, cd+] E[cd+, ψ] × E[d−1 + , ψ−1] × det M(cd+, cd+φ)M −1(1, φ)T −1(cd+). (b) Now we are going to show that M(1, φ) is invertible if so is T (φ) and express the inverse. Then we will compute the remaining determinant. The identity (I + H(φ)) (I − H(φ)) = I − H(φ)H( φ−1) = T (φ) T (φ−1) implies that M(1, φ) = I + H(φ) is invertible. Moreover, M −1(1, φ) = (I + H(φ))−1 = (I − H(φ)) T −1(φ−1) T −1(φ). 22 Next observe that M(cd+, cd+φ)(I − H(φ)) = T (cd+) + H(cd+φ) − T (cd+)H(φ) − H(cd+φ)H(φ) = T (cd+) + H(cd+φ) − H(cd+φ) + H(cd+)T ( φ) − T (cd+φ φ) + T (cd+φ)T ( φ) = H(cd+)T ( φ) + T (cd+φ)T ( φ) = (T (cd+φ) + H(cd+)) T (φ−1). Therefore, and thus M(cd+, cd+φ)M −1(1, φ) = M(cd+, cd+φ)(I − H(φ))T −1(φ−1) T −1(φ) = (T (cd+φ) + H(cd+))T −1(φ), det(cid:16)M(cd+, cd+φ)M −1(1, φ)T −1(cd+)(cid:17) = det(cid:16)(T (cd+φ) + H(cd+))T −1(φ)T −1(cd+)(cid:17). We split this into det(cid:16)(T (cd+φ) + H(cd+))T −1(cd+φ)(cid:17) × det(cid:16)T (cd+φ)T −1(φ)T −1(cd+)(cid:17) with the last determinant equal to E[cd+, φ]. For part (c), we apply Lemma 6 and Lemma 6.2. ✷ Now we specialize to the functions we are interested in. Corollary 6.4 Let ψ and φ be given by (13) and (14) and assume that the condition (16) and (17) hold. Moreover, let c = c−G[c]c+ and d = d−1 1 (see (6) -- (8)). Then the operator determinant E1 of Theorem 6.3 is well-defined and given by + d+ be factorizations in B1 E1 = E1 F [e+] = c0(1)−α+ c0(−1)−α− RYr=1 c0(τr)−α+ r c0(¯τr)−α− r e+(τr)α+ r +βre+(¯τr)α− r +βr RYr=1 × e+(1)α++β+ e+(−1)α−+β− with e+ = c+d+ and c0 = c+c− = c/G[c]. 23 Proof. The assumptions imply that T (ψ) and M(1, φ) are invertible. Hence by the previous theorem, part (a), the operator determinant E1 is well-defined. It also depends analytically on the parameters α±, β±, α± r , βr. Therefore it is sufficient to prove the identity under the assumption that the real parts of all parameters vanish. This implies that T (φ) is invertible and hence part (c) of the previous theorem can be applied. Moreover, F [cd+] = F [c+d+] as can be easily seen from (26). To carry out the computation of the various E[·, ·] terms, we observe that for functions 1 (see (6)) the following general formulas were established C admitting a factorization in B1 in [11, Sect. 10.62]), E[uτ,β, C] = C−(τ )−β, E[C, uτ,β] = C+(τ )β. Recalling the definition of ψ and φ, ψ = u1,α+ u−1,α− RYr=1 uτr,α+ r u¯τr,α− r , φ = u1,α++β+ u−1,α−+β− we get E[ψ, c−] = c−(1)−α+ c−(−1)−α− E[c+, ψ]−1 = c+(1)−α+ c+(−1)−α− uτr,αr+βr u¯τr,αr+βr , c−(τr)−α+ r c−(¯τ )−α− r , c+(τr)−α+ r c+(¯τ )−α− r , RYr=1 RYr=1 RYr=1 E[c+d+, φ] = e+(1)α++β+ e+(−1)α−+β− from which the formula follows. e+(τr)α+ r +βre+(¯τ )α− r +βr, RYr=1 ✷ We remark that in the special case of αr = α± r and βr = 0, which we are going to consider later, the constant E1 = c+(1)β+ c+(−1)β−c−(1)−α+ c−(−1)−α− × d+(1)α++β+ d+(−1)α−+β− RYr=1 c−(τr)−αr c−(¯τr)−αr RYr=1 d+(τr)αr d+(¯τr)αr . (27) Let us now turn to the constant E appearing in Theorem 5.1. 24 Corollary 6.5 Let ψ and φ be given by (13) and (14) and assume that the condition (16) and (17) hold. Moreover, let c = c−G[c]c+ and d = d−1 1 (see (6) -- (8)). Then the constant E in (24) is well-defined and given by + d+ be factorizations in B1 E = E1 ×(cid:18) c+(1)d+(1) c+(−1)d+(−1)(cid:19)1/2 × exp(cid:16)Xk≥1 k[log c]k[log c]−k − 1 2Xk≥1 k([log c]k + [log d]k)2(cid:17). where E1 is the expression in the previous corollary. Proof. We have to identify the additional constants, which are det(cid:16)T −1(c d+)T (c)T ( d+)(cid:17) × det(cid:16)T (cd+)T (c−1d−1 + )(cid:17) E−1[c, d+] · E[cd+] = E[cd+, cd+] E[d+, c] = E[c+, c−]. This we combine with F [cd+] = F [e+] and the previous corollary. ✷ Putting all this together, we have the following. Corollary 6.6 Let ψ and φ be given by (13) and (14) and assume that the condition (16) and (17) hold. Moreover, let c = c−G[c]c+ and d = d−1 1 (see (6) -- (8)). Finally, suppose that (iii) of Theorem 5.2 holds. Then + d+ be factorizations in B1 lim n→∞ det(cid:16)PnT −1(ψ)M(c, cdφ)T −1(ψ−1)Pn(cid:17) G[c]n × det(cid:16)Pn T −1(ψ)M(1, φ)T −1(ψ−1)Pn(cid:17) = E, where E is as in the previous corollary. 7 Known asymptotics In the previous separation theorem and the constant computation we have reduced the asymptotics of to the asymptotics of det(cid:16)PnT −1(ψ)M(c, cdφ)T −1(ψ−1)Pn(cid:17) det(cid:16)Pn T −1(ψ)M(1, φ)T −1(ψ−1)Pn(cid:17). 25 The separation theorem is of course only as useful as far as we are able to obtain this last asymptotics. We can reverse the considerations of Section 3 and Theorem 3.2 and obtain det(cid:16)Pn T −1(ψ)M(1, φ)T −1(ψ−1)Pn(cid:17) = det Mn(a0, b0) where a0 = v1,α+ v−1,α− RYr=1 vτr,α+ r v¯τr,α− r , b0 = v1,α+ u1,β+ v−1,α− u−1,β− vτr,αr uτr,βr v¯τr,αr u¯τr,βr . RYr=1 (28) (29) These are the original functions a and b without the c and d terms. The asymptotics of det Mn(a0, b0) are known in the cases of βr = 0 and αr = α+ r = α− r and β+ ∈ {0, −1}, β− ∈ {0, 1}. We remark that in these cases a0 is an even function which is a product of pure Fisher-Hartwig type functions with zeros/poles only (no jumps). Furthermore, depending on the values of β±, we have four cases, b0(t) = ±a0(t), b0(t) = ta0(t), b0(t) = −t−1a0(t), which are precisely the cases (i) -- (iv) described in the introduction. We list them here and note that G(1+z) is the Barnes G-function, an entire function satisfying G(1+z) = Γ(z)G(z) (see [3]). The asymptotics of det Mn(a0, b0) are given by (1) b0(t) = a0(t), βr = 0 and αr = α+ r = α− r , and β+ = 0, β− = 0. n{ 1 ×Yr 2 ((α+)2+(α−)2−α++α−)+P α2 2 (α++α−)2+ 1 2 (α++α−)+P α2 r r} 2− 1 1 − τ 2 r −α2 r 1 − τr−2αr(α+−1/2)1 + τr−2αr(α−+1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(1/2 + α+)G(3/2 + α−)Yr 2 (α++α−)G(1/2)G(3/2) π 1 × G(1 + αr)2 G(1 + 2αr) 26 (2) b0(t) = −a0(t), βr = 0 and αr = α+ r = α− r , and β+ = −1, β− = 1. n{ 1 ×Yr 2 ((α+)2+(α−)2+α+−α−)+P α2 2 (α++α−)2+ 1 2 (α++α−)+P α2 r r} 2− 1 1 − τ 2 r −α2 r 1 − τr−2αr(α++1/2)1 + τr−2αr(α−−1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(3/2 + α+)G(1/2 + α−)Yr 2 (α++α−)G(3/2)G(1/2) π 1 × G(1 + αr)2 G(1 + 2αr) (3) b0(t) = ta0(t), βr = 0 and αr = α+ r = α− r , and β+ = 0, β− = 1. n{ 1 2 ((α+)2+(α−)2−α+−α−)+P α2 2 (α++α−−1)2+ 1 2 (α++α−−1)+P α2 r} r} 2{2− 1 1 − τ 2 r −α2 r 1 − τr−2αr(α+−1/2)1 + τr−2αr(α−−1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(1/2 + α+)G(1/2 + α−)Yr 2 (α++α−)G(1/2)2 π 1 × G(1 + αr)2 G(1 + 2αr) ×Yr ×Yr (4) b0(t) = −t−1a0(t), βr = 0 and αr = α+ r = α− r , and β+ = −1, β− = 0. n{ 1 2 ((α+)2+(α−)2+α++α−)+P α2 2 (α++α−+1)2+ 1 2 (α++α−+1)+P α2 r} r} 2{− 1 1 − τ 2 r −α2 r 1 − τr−2αr(α++1/2)1 + τr−2αr(α−+1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(3/2 + α+)G(3/2 + α−)Yr 2 (α++α−)G(3/2)2 π 1 × G(1 + αr)2 G(1 + 2αr) The above asymptotics have been proved in [12, Theorem 1.25]. In the special case of α+ = α− = 0 and case (1), see also [6]. On the other hand, if all αr = 0, then the determinants can be evaluated explicitely because they are related to Hankel determinants constructed from the moments of classical Jacobi orthogonal polynomials. In this later case the Hankel determinant asymptotics can also be described as Toeplitz determinants [6]. 27 8 Final computations for the four important cases Using the previous section and then our computations for the constant E we have the final four answers. We assume in what follows that we have the factorizations c = c−G[c]c+ and d = d+ d−1 + (see (6) -- (8)) and that condition (18) holds. Then the asymptotics of det Mn(a, b) are given by the following products. (1) b(t) = a(t), βr = 0 and αr = α+ r = α− r , and β+ = 0, β− = 0. G[c]n n{ 1 2 ((α+)2+(α−)2−α++α−)+P α2 2 (α++α−)2+ 1 2 (α++α−)+P α2 r r} 2− 1 × exp Xk≥1(cid:16)k[log c]k[log c]−k − 1 2 k([log c]k + [log d]k)2(cid:17)! × c+(1)1/2c+(−1)−1/2c−(1)−α+ c−(τr)−αr c−(¯τr)−αr c−(−1)−α− Yr × d+(1)α++1/2d+(−1)α−−1/2 Yr d+(τr)αr d+(¯τr)αr τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k ×Yr 1 − τ 2 r −α2 r 1 − τr−2αr(α+−1/2)1 + τr−2αr(α−+1/2) × 1 π 2 (α++α−)G(1/2)G(3/2) G(1/2 + α+)G(3/2 + α−)Yr G(1 + αr)2 G(1 + 2αr) (2) b(t) = −a(t), βr = 0 and αr = α+ r = α− r , and β+ = −1, β− = 1. G[c]n n{ 1 2 ((α+)2+(α−)2+α+−α−)+P α2 2 (α++α−)2+ 1 2 (α++α−)+P α2 r r} 2− 1 × exp Xk≥1(cid:16)k[log c]k[log c]−k − 1 2 k([log c]k + [log d]k)2(cid:17)! × c+(1)−1/2c+(−1)1/2c−(1)−α+ c−(τr)−αr c−(¯τr)−αr c−(−1)−α− Yr × d+(1)α+−1/2d+(−1)α−+1/2 Yr d+(τr)αr d+(¯τr)αr 28 ×Yr 1 − τ 2 r −α2 r 1 − τr−2αr(α++1/2)1 + τr−2αr(α−−1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(3/2 + α+)G(1/2 + α−)Yr 2 (α++α−)G(3/2)G(1/2) π 1 × G(1 + αr)2 G(1 + 2αr) (3) b(t) = ta(t), βr = 0 and αr = α+ r = α− r , and β+ = 0, β− = 1. G[c]n n{ 1 2 ((α+)2+(α−)2−α+−α−)+P α2 2 (α++α−−1)2+ 1 2 (α++α−−1)+P α2 r} r} 2{2− 1 × exp Xk≥1(cid:16)k[log c]k[log c]−k − 1 2 k([log c]k + [log d]k)2(cid:17)! ×c+(1)1/2c+(−1)1/2 c−(1)−α+ c−(τr)−αrc−(¯τr)−αr c−(−1)−α− Yr d+(τr)αr d+(¯τr)αr × d+(1)α++1/2d+(−1)α−+1/2 Yr ×Yr 1 − τ 2 r −α2 r 1 − τr−2αr(α+−1/2)1 + τr−2αr(α−−1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(1/2 + α+)G(1/2 + α−)Yr 2 (α++α−)G(1/2)2 π 1 × G(1 + αr)2 G(1 + 2αr) (4) b(t) = −t−1a(t), βr = 0 and αr = α+ r = α− r , and β+ = −1, β− = 0. G[c]n n{ 1 2 ((α+)2+(α−)2+α++α−)+P α2 2 (α++α−+1)2+ 1 2 (α++α−+1)+P α2 r} r} 2{− 1 × exp Xk≥1(cid:16)k[log c]k[log c]−k − 1 2 k([log c]k + [log d]k)2(cid:17)! ×c+(1)−1/2c+(−1)−1/2 c−(1)−α+ c−(τr)−αr c−(¯τr)−αr c−(−1)−α− Yr × d+(1)α+−1/2d+(−1)α−−1/2 Yr d+(τr)αr d+(¯τr)αr 29 ×Yr 1 − τ 2 r −α2 r 1 − τr−2αr(α++1/2)1 + τr−2αr(α−+1/2) τk − τj−2αkαj τk − 1/τj−2αkαj ×Yj<k G(3/2 + α+)G(3/2 + α−)Yr 2 (α++α−)G(3/2)2 π 1 × G(1 + αr)2 G(1 + 2αr) These formulas follow in a straightforward manner from simply using the known results quoted in the previous section, along with our separation theorem formulas for the constant and the evaluation of those for the specific values of β+ and β−. More specifically, use (27) and Corollary 6.5. 9 The inverse of I + H(ψ) In the course of the computations in the previous sections, we discovered that when ψ−1 = ψ, explicit forms of the inverse of I +H(φ0ψ) can be found in the four cases φ0 = ±1, φ0(z) = z, φ0(z) = −z−1 considered in the previously. The inverse will be expressed in terms of the inverse of T (ψ) and T (ψ−1). The first two cases are rather simple. Proposition 9.1 Suppose that ψ−1 = ψ and that the operator T (ψ) is invertible. Then I ± H(ψ) is invertible and (I ± H(ψ))−1 = T (ψ−1)−1(I ± H(ψ−1))T (ψ)−1 . Proof. From ψ−1 = ψ and thus T (ψ)T (ψ−1) = I − H(ψ)2, we know that (T (ψ)T (ψ−1))−1 = T (ψ−1)−1T (ψ)−1 = (I − H(ψ)2)−1. This means that T (ψ−1)−1(I ± H(ψ−1))T (ψ)−1 = (I − H(ψ)2)−1T (ψ) (I ± H(ψ−1)) T (ψ)−1. Now from (4) we know that T (ψ)H(ψ−1) = −H(ψ)T (ψ). So the term T (ψ) (I ± H(ψ−1) T (ψ)−1 = I ∓ H(ψ) 30 and this yields the result. ✷ In order to continue with the other two case, we need to make a connection with factorization. If T (ψ) is invertible then there exists a Wiener-Hopf factorziation of ψ such If in addition ψ−1 = ψ, then using the that ψ = ψ−ψ+ and ψ± and ψ−1 uniqueness of the Wiener-Hopf factorization (up to multiplicative constants) it is not difficult to see that the factors are related to each other either by ± are all in L2. ψ+ = ψ−1 − or ψ+ = − ψ−1 − . This means that we have either ψ = ψ−1 + ψ+ or ψ = − ψ−1 + ψ+ . We can equivalently express this dichotomy as P1T −1(ψ)P1 = 1 or P1T −1(ψ)P1 = −1 by using that T −1(ψ) = T (ψ−1 + )T (ψ−1 − ). We need one basic result before we proceed. Lemma 9.2 If ψ−1 = ψ, if the Toeplitz operator T (ψ) is invertible and if P1T −1(ψ)P1 = 1, then (H( ψ) + T (ψ)−1H(ψ) − H( ψ/z)T (ψ)−1H(ψ))H(z) = 0. and ψ± and ψ−1 If T (ψ) is invertible then there exists a factorization of ψ such that ψ = ψ−ψ+ where Proof. −1 ψ− = ψ+ ± are all in L2. The operator H(z) is the rank one projection P1, so the above will be zero if the operator above applied to the constant function f (z) = 1 is zero. The operator T (ψ)−1H(ψ) is the same as the operator T (ψ−1 + )H(ψ+) since it agrees with this operator on all trigonometric polynomials. Hence T (ψ)−1H(ψ) f = T (ψ−1 + )H(ψ+) f = (ψ−1 + (ψ+ − (ψ+)0)/z) = (1 − ψ−1 + (ψ+)0)/z. We also have H( ψ) f = P ( ψ/z). Finally H( ψ/z)((1 − ψ−1 + (ψ+)0)/z) = P ( ψ/z) − (1 − ψ−1 + (ψ+)−1 0 )/z. Putting the three terms together yields the desired result. ✷ Theorem 9.3 Suppose that ψ−1 = ψ, that the operator T (ψ) is invertible, and that P1T −1(ψ)P1 = 1. Then I − H(z−1ψ) is also invertible and its inverse is given by T (ψ−1)−1(I − H(z−1ψ−1))T (ψ)−1 . 31 Proof. The inveribility of T (ψ) implies the invertibility of T (ψ−1) (see the beginning of the proof of Theorem 5.1). This also means that the operator T (ψ)T (ψ−1) = I − H(ψ)2 is invertible. We do the same steps as in the first two cases, except that this time we are dealing with the term T (ψ) (I − H(ψ−1/z) T (ψ)−1 = T (ψ) (I − H( ψ/z)) T (ψ)−1 in a slightly different manner. First T (ψ) (I − H( ψ/z)) T (ψ)−1 = I + H(ψ)T (z) + H(ψ)H(z)H( ψ)T (ψ)−1. Multiplying on the right by I − H(ψ/z) we have that T (ψ) (I − H( ψ/z)) T (ψ)−1(I − H(ψ/z)) is the same as I + H(ψ)H(z)H( ψ)T (ψ)−1 − H(ψ)T (z) H(ψ/z) − H(ψ)H(z)H( ψ)T (ψ)−1H(ψ/z). The above uses the identity H(ψ/z) = H(ψ) T (z) = T (1/z) H(ψ). Now H(ψ) T (z) H(ψ/z) = H(ψ) T (z) T (1/z)H(ψ) = H(ψ) Q1 H(ψ). This means the above is the same as I −H(ψ)H(ψ) + H(ψ)H(z)H(ψ) + H(ψ)H(z)H( ψ)T (ψ)−1 −H(ψ)H(z)H( ψ)T (ψ)−1H(ψ/z). Thus we will have our result if we can show that the last three terms of this operator sum to zero. To do this we consider the transpose of H(ψ) + H( ψ)T (ψ)−1 − H( ψ)T (ψ)−1H(ψ/z), that is, H(ψ) + T ( ψ)−1H( ψ) − H(ψ/z)T ( ψ)−1H( ψ). Using Lemma 9.2, with ψ replaced by eψ the statement is proved. Theorem 9.4 Suppose that ψ−1 = ψ, that the operator T (ψ) is invertible, and that P1T −1(ψ)P1 = 1. Then I + H(zψ) is also invertible and the inverse is given by ✷ (T (ψ−1) + H(z))−1(I + H(zψ−1))(T (ψ) + H(z))−1. 32 Proof. The inverse of (T (ψ) + H(z))−1 is the same as T (ψ)−1(I − 1 be verified using basic linear algebra. Now consider the product 2H(z)T (ψ)−1) which can T (ψ)−1(I − 1 2 H(z)T (ψ)−1) (I + H(zψ)). Using the factorization of ψ we see this is the same as the operator Note that T (ψ−1 + )T (fψ+) = T (ψ)−1 and that + )H(zψ+) = T (ψ−1 T (ψ−1 T (ψ−1 + )[(I − 1 2 H(z))(T (fψ+) + H(zψ+)]. + )T (fψ+)H(zψ+fψ+ −1 ) = T (ψ)−1H(zψ) and thus is also a bounded operator. Now multiplying by I + H(zeψ) and doing a similar computation yields (I + H(zeψ))T (ψ)−1(I − H(z)T (ψ)−1) (I + H(zψ)) = T (eψ) + H(z). which gives the desired result. 1 2 ✷ A concise way to say both of the previous theorems is: The inverse of M(1, φ0ψ) is given by References (T (ψ−1) + H(φ0))−1M(1, φ0ψ−1)(T (ψ) + H(φ0))−1. [1] J.C. Andrade, Random Matrix Theory and L-functions in function fields, PhD Thesis (2012), University of Bristol. [2] J. Baik, E.M. Rains, Algebraic aspects of increasing subsequences, Duke Math. J. 109, no.1(2001), 1 -- 65. [3] E.W. Barnes, The theory of the G-function, Quart. J. Pure Appl. Math. 31 (1900), 264 -- 313. [4] E.L. Basor, Y. Chen, T. Ehrhardt, Painlev`e V and time-dependent Jacobi polynomials, J. Phys. A: Math. Theor. 43 (2010), 01524 (25 pages). [5] E. L. Basor, T. Ehrhardt, Asymptotic formulas for determinants of a sum of finite Toeplitz and Hankel matrices, Math. Nachr. 122 (2001), 5 -- 45. 33 [6] E. L. Basor, T. Ehrhardt, Asymptotic formulas for the determinants of symmetric Toeplitz plus Hankel matrices, in: Toeplitz matrices and singular integral equations (Pobershau, 2001), 61 -- 90, Oper. Theory Adv. Appl., 135, Birkhauser, Basel, 2002. [7] E.L. Basor, T. Ehrhardt, Determinant computations for some classes of Toeplitz-Hankel matrices, Oper. Matrices 3, no. 2 (2009), 167 -- 186. [8] E.L. Basor, T. Ehrhardt, Fredholm and invertibility theory for a special class of Toeplitz + Hankel operators, J. Spectral Theory 3 (2013), 171 -- 214. [9] E.L. Basor, T. Ehrhardt, Stability of finite sections of Toeplitz-Hankel type operators, in preparation. [10] A. Bottcher, B. Silbermann, Toeplitz matrices and determinants with Fisher-Hartwig symbols. J. Funct. Anal. 63 (1985), no. 2, 178 -- 214. [11] A. Bottcher, B. Silbermann, Analysis of Toeplitz operators, Springer, Berlin, 1990; also: 2nd ed., Springer, 2006. [12] P. Deift, A. Its, I. Krasovsky, Asymptotics of Toeplitz, Hankel, and Toeplitz+Hankel determinants with Fisher-Hartwig singularities, Annals of Mathematics, 174 (2011), 143 -- 1299. [13] T. Ehrhardt, A status report on the asymptotic behavior of Toeplitz determinants with Fisher-Hartwig singularities, in: Recent advances in operator theory (Groningen, 1998), 217 -- 241, Oper. Theory Adv. Appl., 124, Birkhauser, Basel, 2001. [14] P.J. Forrester, N.E. Frankel, Applications and generalizations of Fisher-Hartwig asymp- totics, J. Math. Phys. 45, no. 5 (2004), 2003 -- 2028. [15] S. Roch, Finite Sections of operators generated by singular integrals with Carleman shift, Preprint Nr. 52, November 1987, Sektion Mathematik, Technische Universitat Karl-Marx-Stadt. [16] H. Widom, Inversion of Toeplitz Matrices. II, Illinois J. Math. 4 (1960), 88 -- 99. [17] H. Widom, Asymptotic behavior of block Toeplitz determinants. II, Adv. Math. 21 (1976), 1 -- 29. 34
1907.06878	2	1907	2019-12-11T09:47:17	Toeplitz operators with singular symbols in polyanalytic Bergman spaces on the half-plane	[ "math.FA", "math.CV" ]	Using the approach based on sesquilinear forms, we introduce Toeplitz operator in the analytic Bergman space on the upper half-plane with strongly singular symbols, derivatives of measures. Conditions for boundedness and compactness of such operators are found. A procedure of reduction of Toeplitz operators in Bergman spaces of polyanalytic functions to operators with singular symbols in the analytic Bergman space by means of the creation-annihilation structure is elaborated, which leads to the description of the properties of the former operators	math.FA	math	TOEPLITZ OPERATORS WITH SINGULAR SYMBOLS IN POLYANALYTIC BERGMAN SPACES ON THE HALF-PLANE GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI From the first named author, with best wishes to the second named one, on the occasion of his Jubilee Abstract. Using the approach based on sesquilinear forms, we introduce Toeplitz operator in the analytic Bergman space on the upper half-plane with strongly singular symbols, derivatives of measures. Conditions for bounded- ness and compactness of such operators are found. A procedure of reduction of Toeplitz operators in Bergman spaces of polyanalytic functions to opera- tors with singular symbols in the analytic Bergman space by means of the creation-annihilation structure is elaborated, which leads to the description of the properties of the former operators. 9 1 0 2 c e D 1 1 ] . A F h t a m [ 2 v 8 7 8 6 0 . 7 0 9 1 : v i X r a 1. Introduction In a series of papers [7, 8, 9] the authors developed the approach to defining the Toeplitz operators in Bergman type spaces by means of bounded sesquilinear forms, which permitted them to study Toeplitz operators with strongly singular symbols. The cases of the classical Bergman space of analytic functions on the unit disk D and the Fock space on the whole complex plane were considered. The present paper is devoted to the study of such Toeplitz operators in one more classical Bergman space, the one of analytic functions on the upper half-plane Π, square integrable with respect to the Lebesgue measure. It is well known that for sufficiently regular, say, bounded symbols, the theories of Toeplitz operators in the Bergman spaces on the disk and on the upper half-plane are equivalent, in the sense that the Mobius transform z 7→ ζ = M(z) := z − i 1 − iz : Π → D, generates the mapping U : f 7→ g = Uf, (Uf )(z) = f (M(z)) 1 (1 − iz)2 , z ∈ Π, 2010 Mathematics Subject Classification. 47A75 (primary), 58J50 (secondary). Key words and phrases. Bergman space, Toeplitz operators. The first-named author is supported by grant RFBR No 17-01-00668. 1 (1.1) (1.2) 2 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI which is an isometry of the Bergman spaces A2(D) and A1 := A2(Π). How- ever, when passing to more singular symbols that generate Toeplitz operators via sesquilinear forms, the boundedness conditions carry over not in such simple way. In the present paper, we introduce Carleson measures for derivatives of order k (k-C measures) for the Bergman space on the half-plane and find conditions for a given measure to be a k-C measure. As usual, these conditions are sufficient for complex measures but are also necessary for positive ones. An estimate for the properly defined norm of k-C measures, with explicit dependence on the derivative order k, is obtained. This makes it possible to consider strongly singular symbols, containing derivatives of unbounded order. The above results are applied for studying Toeplitz operators in polyanalytic Bergman spaces on the upper half-plane, using the creation-annihilation structure discovered in [4, 14]. The first-named author is grateful to the Mittag-Leffler institute for hospitality and support while a considerable part of the paper was written. 2. The Bergman space on the half-plane and related structures. Singular symbols Similar to the hyperbolic metric on the disk, the pseudo-hyperbolic metric on the half-plane is useful. As known, for z, w ∈ Π, the pseudo-hyperbolic distance between z, w is defined by Thus, the area of the disk D(z, R) equals 4πR2y2 (1−R2)2 . Conversely, the Euclidean disk B(x + iη, r) ∈ Π is the pseudo-hyperbolic disk D(x + iy, R) with R = η r −r η2 r2 − 1 and y = 1 − R2 1 + R2 η. Note that, while the pseudo-hyperbolic radius is fixed, the y-coordinate of the center of the pseudo-hyperbolic disk is proportional to the corresponding coordi- nate of the Euclidean disk. Recall that the Bergman space A1 = A2(Π) is a subspace in L2(Π) which It is a reproducing kernel space, with the consists of analytic in Π functions. reproducing kernel κ(z, w) = −(π(z−w)2)−1. Thus, the integral operator P = PΠ with kernel κ(z, w) is the orthogonal (Bergman) projection of L2(Π) onto A2(Π). d(z, w) =(cid:12)(cid:12)(cid:12)(cid:12) z − w z − w(cid:12)(cid:12)(cid:12)(cid:12) . We will denote by D(z, R), z = x + iy, R < 1, the pseudo-hyperbolic disk in Π, centered at z with 'radius' R. It is easy to check that the disk D(z, R) coincides with the Euclidean disk B(w, r), D(x + iy, R) = B(w, r), where w = x + i 1 + R2 1 − R2 y, r = 2Ry 1 − R2 . (2.1) POLYANALYTIC BERGMAN SPACES 3 This projection is connected with the Bergman projection PD for the Bergman space A2(D) by means of the operator U in (1.2): PΠ = U∗PDU. (2.2) Given a bounded function a(z), z ∈ Π, the Toeplitz operator in A1 is defined in the usual way as (Taf )(z) = (Paf )(z) =ZΠ κ(z, w)a(w)f (w)dA(w), (2.3) dA being the Lebesgue measure. This operator is bounded in A1, as a composition of two bounded operators. The function a(z) is called the symbol of the Toeplitz operator. It was the object of many studies to extend this definition to symbols being objects, more singular than bounded functions, still producing bounded operators. This program was implemented, in particular, in [7, 8] for the Bergman space on the disk and for the Fock space. Now we follow the pattern of these papers for the upper half-pane case. The first stage here is considering (complex) measures as symbols. Let µ be an absolutely continuous measure on Π, with a bounded density a(z) with respect to the Lebesgue measure dA. Then the action of operator (2.3) can be written as κ(z, w)f (w)dµ(w). (Taf )(z) = (Paf )(z) =ZΠ κ(z, w)a(w)f (w)dA(w) =ZΠ (2.4) Such a definition of the Toeplitz operator by means of the expression on the right- hand side of (2.4) can be, at least formally, extended to measures µ which are not necessarily absolutely continuous with respect to A with bounded density, and even to those that are not absolutely continuous with respect to A at all; what is, actually, needed is just the boundedness of the operator defined by the right-hand side in (2.4). To find some effective analytical conditions for this boundedness is rather a quite hard task. At the same time the approach based upon sesquilinear forms turns out to be very efficient here. In fact, in case of dµ = a(z)dA(z), we consider the sesquilinear form Fµ[f, g] = hTaf, gi, where f, g ∈ A1. By the above definition of the operator Ta, Fµ[f, g] = hPaf, gi = haf, Pgi = haf, gi = (2.5) ZΠ a(z)f (z)g(z)dA(z) =ZΠ f (z)g(z)dµ(z), f, g ∈ A1. The left-hand side in (2.5) is defined for a being a (sufficiently nice) function, however, the right-hand side makes sense for a measure µ and can be thus used for a definition of a Toeplitz operator. The sesquilinear form (2.5) defines a bounded operator in A1 in case it is bounded, i.e., Fµ[f, g] ≤ Ckfkkgk. This 4 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI boundedness follows, as soon as the inequality (cid:12)(cid:12)(cid:12)(cid:12)ZΠ f (z)2dµ(z)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Ckfk2 ZΠ f (z)2dµ(z) ≤ Ckfk2, is satisfied for all f ∈ B. This estimate is, surely, satisfied provided (2.6) where µ denotes the variation of the measure µ. Note that the considerations involving sesquilinear forms are essentially more convenient in analysis since they evade using reproducing kernels and deal with inequalities containing only the functions f, g and the measure µ. Measures µ subject to the estimate (2.6) are called Carleson measures for the space A1. A description for such Carleson measures was given for the case of the Bergman space on the disk, for the Fock space and some other Bergman and Hardy type spaces, see, e.g., [16, 17, 18]. The criterion for a measure to be a Carleson measure for A1 follows from the known one for the space A(D). Proposition 2.1. Let µ be a complex Borel measure on Π. For a fixed R ∈ (0, 1), if (2.7) with constant Cµ not depending on z, then the sesquilinear form (2.5) is bounded in A1. That is, the inequality (2.6) is satisfied for all f ∈ A1, with constant C = C(R)Cµ, where C(R) depends only on R. If the measure µ is positive, the condition (2.7) is also necessary for the the sesquilinear form (2.5) to be bounded. µ(D(z, R)) ≤ CµA(D(z, R)), z ∈ Π, R < 1, Proof. The result follows from a similar statement concerning Carleson measures for the Bergman space on the disk, see, e.g., [16], by means of the unitary equiv- alence (2.2). A reasoning establishing directly this property can be found, for example, in [3, Proposition 2.6]. (cid:3) We note here that the condition (2.7) can be (although just formally !) relaxed, when replaced by µ(D(z, R)) ≤ CµA(D(z, R1)), z ∈ Π, R < 1, (2.8) with any fixed R1 > R. This remark will be used when considering Carleson measures for derivatives later on. A measure µ on Π, having compact support, can be considered as a distribution in E ′(Π). At the same time, the function f (z)g(z) is infinitely differentiable in Π, f (z)g(z) ∈ E(Π), moreover, it is real- analytic in Π. Thus, this function can be represented as f (z)g(z) = Diag ∗(f ⊗ ¯g) = Diag ∗((f ⊗ 1)(1 ⊗ ¯g)), where Diag is the diagonal embedding of Π into Π × Π, Diag (z) = (z, z), and Diag ∗ is the induced mapping A1 ⊗ A1 to A(Π) (the space of real-analytic func- tions), Diag ∗(f ⊗ g) = f (z)g(z). We denote by M the image of A1 ⊗A1 in A(Π) POLYANALYTIC BERGMAN SPACES 5 under the mapping Diag ∗. Therefore the expression (2.5) can be understood as Fµ[f, g] = (µ, f (z)g(z)) = (µ, h), with h = Diag ∗(f ⊗ ¯g) ∈ A(Π), (2.9) where parentheses denote the intrinsic paring of E ′(Π) and E(Π). Proposition 2.1 can be now understood in the sense that as soon as the condition (2.7) is satisfied, the sesquilinear form (2.9) can be extended to M for the measure µ not neces- sarily having compact support. Moreover, the estimate (µ, h) ≤ Cµkfkkgk, h ∈ M, holds. We recall here how such extension of the distribution is stan- dardly performed; this will make our further considerations for distributions of more general nature more clear. Let µj be a sequence of measures with compact support, each one in a (closed) quasi-hyperbolic disk Dj Π of radius r, and let these disks form a covering of Π with finite multiplicity m: each point of Π is covered by not more than m disks Dj. Suppose that for each measure µj, the estimate Fµj [f, g] ≤ Ckfkkgk is satisfied, with the same constant C. Then, we can sum up such estimates over j and, using the finite multiplicity property, arrive at the same estimate for the measure µ = P µj, automatically locally finite in Π (just with a controllably larger constant.) This line of reasoning, considering first the distributions in E ′(Π), with compact support, obtaining proper estimates, and extending then these estimates to certain distributions without compact support condition, so that the estimates hold on A1, will be further implemented for more general distributions. A few words to explain our philosophy. Let Ω be an open subset in C (Ω = Π If F is a distribution in E ′(Ω) i.e., a distribution with compact in our case). support in Ω, its derivative, say, ∂F is standardly defined as the distribution ∂F acting on functions φ ∈ E(Ω) by the rule (∂F, φ) = −(F, ∂φ). If, however, F is a distribution in a wider space, F ∈ D′(Ω), the action (∂F, φ) is not necessarily defined for all φ ∈ E(Ω). In particular, if φ is a nontrivial function in the Bergman space or a function in M, it can never belong to D(Ω), so the action of F on such functions and, further on, the definition of derivatives of F needs to be specified anew, however being consistent with the usual definition. In what follows, we consider a class of distributions for which such construction works, preserving the usual properties of distributions. The natural compensation for this frivolity is the narrowing of the set of functions on which such 'distributions' act. 3. Carleson measures for derivatives Following the pattern in [7, 8], we introduce now a class of sesquilinear forms involving derivatives of functions f, g, corresponding thus to (formal) distribu- tional derivatives of Carleson measures. 6 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI Definition 3.1. Let µ be a regular fix sign measure on Π and α, β be two non- negative integers. We denote by Fα,β,µ the sesquilinear form Fα,β,µ[f, g] = (−1)α+βZΠ ∂αf (z)∂βg(z)dµ(z), f, g ∈ A1, (3.1) (3.2) which we denote as well by Fα,β,µ[f, g] = (∂α ¯∂βµ, f ¯g). This definition is consistent with our explained above approach. In fact (3.1), (3.2) act as the definition of the action of the 'distribution' ∂α ¯∂βµ on elements in A1. Note that this is consistent with the standard distributional definition of the derivative of a measure in the case when µ is a compactly supported measure in Π. The first set of properties of such forms and corresponding operators, similar to the ones for other Bergman spaces, is the following. Theorem 3.2. Let µ be a measure with compact support in Π and let α, β be nonnegative integers. Then (1) For any f, g ∈ A1, the integral in (3.1) converges, moreover, F[f, g] = Fα,β,µ[f, g] = (∂α ¯∂βµ, f ¯g), where the derivatives are understood in the sense of distributions in (2) The sesquilinear form (3.1) is bounded, considered on A1 ×A1 and there- E ′(Π) and the parentheses mean the intrinsic paring in (E ′(Π),E(Π)). fore defines a bounded Toeplitz operator by hTFf, gi = F[f, g], or (TFf )(z) = F[f, κz(·)], where, as usual, κz(w) = k(z, w) = k(w, z). in A1. lowing estimate holds (4) If sn(TF) denote the singular numbers of the operator TF, then the fol- (3) The sesquilinear form (3.1) determines a compact Toeplitz operator TF sn(T) ≤ C exp(−nσ), n ∈ N, (3.3) where σ > 0 is a constant determined by the measure µ and integers α, β. Proof. The property (4) absorbs the other ones, so we will prove only it. Due to Ky Fan's inequalities for singular numbers of compact operators, it is sufficient to establish (3.3) for a positive measure µ. Consider a closed Euclidean disk B ⊂ Π with radius R such that for some r > 0, the support of µ lies strictly inside B, thus dist (z, ∂B) > r > 0 for all z ∈ supp µ. The Cauchy integral formula implies that for any z ∈ supp µ and any α ∈ Z+, ∂αf (z)2 ≤ CαZ∂B f (ζ)2dl(ζ) POLYANALYTIC BERGMAN SPACES 7 for each function f ∈ A1, with constant Cα depending only on α, but not de- pending on f and z. By the same reason, for any z ∈ supp µ, the estimate ∂αf (z)2 ≤ CαZζ−z∈(R,R+r/2) f (ζ)2dA(ζ) holds. By the Cauchy-Schwartz inequality, Fα,β,µ[f, g] ≤ Zsupp µ ∂αf (z)2dµ(z)! 1 2 Zsupp µ ∂βg(z)2dµ(z)! 1 2 (3.4) , for all f, g ∈ A1(Π), and then, due to (3.4), Fα,β,µ[f, g] ≤ C ′ αC ′ βµ(B)kfkL2(B′)kgkL2(B′). The last relation means that the sesquilinear form Fα,β,µ is bounded not only in A1(Π), but in A1(B′) as well, where B′ is the Euclidean disk B′ = B(z, R + r/2). Now we represent the Toeplitz operator TF as the composition A1(B′) TF = TF IB⇛B′ IB′⇛Π, (3.5) where IB⇛B′ : A1(B′) → A1(B), IB′⇛Π : A1(Π) → A1(B′) are operators gen- erated by restrictions of functions defined on a larger set to the corresponding smaller set. The equality (3.5) can be easily checked by writing the sesquilinear forms of operators on the left-hand and on the right-hand side. Finally, the first and the third operators on the right-hand side are bounded, while the middle one, the operator generated by the embedding of the disk B to B′, is known to have the exponentially fast decaying sequence of singular numbers see, e.g., [5]. (cid:3) Remark 3.3. Using the results in [5], one can give an upper estimate for the constant σ in (3.3). In fact, let Q ⊂ D be the image of supp µ in the unit disk, under the mapping (1.1). We denote by C(Q) the set of connected closed sets V ⊂ D containing Q. For each V ∈ C(Q), let cap(V ) denote the logarithmic capacity of V (see the definition, e.g., in [5]) and we set c(µ) as c(µ) = inf V ∈C(Q) cap(V ). Then estimate (3.3) holds for any σ < c(µ). In fact, for a positive measure µ, α, β = 0, and for a connected set Q, the asymptotics of the singular numbers of the Toeplitz operator was found in [5], and our estimate for the exponent in (3.3) follows from this result and the natural monotonicity of singular numbers under the extension of the set where the measure is supported. Our considerations give only the upper estimate for these singular numbers. It is remarkable that for a non-connected support of the measure, even in the setting of [5], the terms in which the singular numbers asymptotics or even sharp order estimates can be expressed are so far unknown. Next, we present our main definition. 8 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI Definition 3.4. A measure on Π is called a Carleson measure for derivatives of order k (k-C measure) if for some constant Ck(µ) > 0, Fk,µ[f, f ] ≡(cid:12)(cid:12)(cid:12)(cid:12)ZΠ ∂kf (z)2dµ(z)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Ck(µ)kfk2 for all f ∈ A1. (3.6) Now, we find a sufficient condition for a measure to be a k-C measure. Here it is important to control the dependence of the constant Ck(µ) in (3.6) on the number k. Theorem 3.5. Let γ ∈ (0, 1) be fixed, and let the measure µ on Π satisfy the condition (3.7) with some κ(µ) for all z0 ∈ Π. Then inequality (3.6) is satisfied, with Ck(µ) = (k!)2κ(µ)γ−2k, that is, µ is a k-C measure. µ(D(z0, γ)) ≤ κ(µ) Im z02kA(D(z0, γ)) Proof. Given a point z0 ∈ Π, the pseudo-hyperbolic disk D(z0, γ) coincides with the Euclidean disk B(w0, s), where their centra and radii are connected by (2.1). We write then the standard representation of the derivative at a point w, w0 − w < s1 < s of an analytic function f (w): f (k)(w) = k!(2πı)−1Zw0−ζ=σ (ζ − w)−k−1f (ζ)dζ, s1 < σ < s. (3.8) Now we fix s2 ∈ (s1, s) and integrate (3.8) in σ variable from s2 to s, which gives us the estimate f (k)(w) ≤ (s − s2)−1k!(2π)−1Zs2≤w0−ζ≤s ζ − w−k−1f (ζ)dA(ζ). We choose then s1 = 1 Cauchy-Schwartz, one yields 4 s, s2 = 3 4s. The inequality ζ − w ≥ 1 2s, together with the f (k)(w) ≤ (s/2)−k−2k!π−1Zw0−ζ≤s f (ζ)dA(ζ) (s/2)−(k+1)k!(cid:18)Zw0−ζ≤s f (ζ)2dA(ζ)(cid:19) 1 2 √π ≤ 2 , f (k)(w)2 ≤ (4/π)(s/2)−2(k+1)(k!)2Zw0−ζ≤s f (ζ)2dA(ζ). (3.9) or The estimate (3.9) holds for all w, w − w0 < s1 = s/4. Therefore we can integrate it over the Euclidean disk B(w0, s1) with respect to the measure µ, POLYANALYTIC BERGMAN SPACES 9 which gives (3.10) (3.11) We substitute now the estimate (3.7) into (3.10) and use (2.1) to arrive at (cid:12)(cid:12)(cid:12)(cid:12)ZB(w0,s1) f (k)(w)2dµ(w)(cid:12)(cid:12)(cid:12)(cid:12) ≤ µ(B(w0, s1))(4/π)(s/2)−2(k+1)(k!)2 Zw0−ζ≤s f (ζ)2dA(ζ). (cid:12)(cid:12)(cid:12)(cid:12)ZB(w0,s1) f (k)(w)2dµ(w)(cid:12)(cid:12)(cid:12)(cid:12) ≤ k(µ) y2k ≤ k(µ) γ−2k(k!)2Zw0−ζ≤s f (ζ)2dA(ζ), (cid:12)(cid:12)(cid:12)(cid:12)ZD(z1,γ1) f (k)(w)2dµ(w)(cid:12)(cid:12)(cid:12)(cid:12) ≤ k(µ) γ−2k(k!)2ZD(z0,γ) f (ζ)2dA(ζ), 0 π(s/2)2(4/π)(s/2)−2(k+1)(k!)2 Zw0−ζ≤s f (ζ)2dA(ζ) or, returning to the pseudo-hyperbolic disks, where D(z1, γ1) = B(w0, s1) ⊂ B(w0, s) = D(z0, γ). Now we follow the reasoning in [16, Theorem 7.4], where estimates of the type (3.11) were summed to obtain the required k-Carleson property. One just should replace the hyperbolic disks with pseudo-hyperbolic ones. Like in [16], it is possible to find a locally finite covering Ξ of the half-plane Π by disks of the Π of finite, moreover, controlled multiplicity. The latter means that the number type D(z1, γ1), with z1 ∈ Π, so that the larger disks D(z0, γ) form a coveringeΞ of m(eΞ) = maxz∈Π #{D ∈ eΞ : z ∈ D} is finite. After adding up all inequalities of forRΠ f (k)2dµ. the form (3.11) over all disks D = D(z1, γ1) ∈ Ξ, we obtain the required estimate Having Theorem 3.5 at our disposal, we introduce the classes of k-C measures. Definition 3.6. Fix a number γ ∈ (0, 1). The class Mk,γ consists of measures µ on Π such that (cid:3) k,γ(µ) := sup z∈Π{µ(D(z, γ))(Im z)−2(k+1)(k!)2γ−2k} < ∞. Theorem 3.5 implies that the class Mk,γ consists of k-C measures. This class, in fact, does not depend on the value of γ chosen, however the value k,γ(µ) does. It is convenient to extend the definition of Mk,γ to half-integer values of k: Definition 3.7. Let k ∈ Z+ + 1 consists of measures satisfying 2 be a half-integer. The class Mk,γ, γ ∈ (0, 1) k,γ(µ) := sup z∈Π(cid:8)µ (D(z, γ)) (Im z)−2(k+1)Γ(k + 1)2γ−2k(cid:9) < ∞. (3.12) 10 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI Further on, the parameter γ will be fixed and will be often omitted in our notations. The quantity k(µ) will be called the k-norm of the measure µ. According to Definition 3.7, the spaces Mk,γ for different k ∈ Z+/2 are related by Mk = (Im z)2(l−k)Ml, with k(µ) = γ2(l−k)(Γ(k + 1)/Γ(l + 1))2l(µ). Theorem 3.5 enables us to find conditions for boundedness of the operators defined by differential sesquilinear forms Fα,β,µ in (3.1). They look similar to the corresponding conditions for sesquilinear forms in the Bergman space on the unit disk and are derived from Theorem 3.5 in the same way as Proposition 6.8 is derived from Theorem 6.3 in [8], so we restrict ourselves to formulations only. Theorem 3.8. Let the measure µ satisfy (3.12) with some k ∈ Z+/2. Then for α, β ∈ Z+, α + β = 2k, the sesquilinear form Fα,β,µ[f, g] = (−1)α+βZΠ ∂αf ∂βgdµ is bounded in A1 and defines a bounded Toeplitz operator Tα,β,β in the Bergman space A1, moreover its norm is majorated by k(µ). Taking into account our above agreement concerning distributional derivatives of measures without compact support condition, Theorem 3.8 can be reformulated as Theorem 3.9. Under the conditions of Theorem 3.8, the sesquilinear form F∂α∂β µ[f, g] = (∂α∂βµ, f ¯g) is bounded in A1 and defines a bounded Toeplitz operator T∂α∂β µ in A1. As usual for Toeplitz type operators, boundedness conditions lead to compact- ness conditions, formulated in similar terms. Theorem 3.10. For R > 0, denote by QR the rectangle in Π: Suppose that α + β = 2k and QR = {z = x + y ∈ Π : x ∈ (−R, R), y ∈ (R−1, R)}. z∈Π\QR(cid:8)µ (D(z, γ)) (Im z)−2(k+1)Γ(k + 1)2γ−2k(cid:9) = 0. sup lim R→∞ Then the operator T∂α∂βµ is compact in A1. Proof. It goes in a standard way. Split the measure µ into two parts, µ = µR +µ′ R, where µR has support outside QR and µ′ R has compact support. Correspondingly, the operator T∂α∂βµ splits into two terms, TR + T′ R. The first operator, by Theorem 3.9, has small norm, as soon as R is chosen sufficiently large. The operator T′ (cid:3) R is compact by Theorem 3.1. Therefore, T∂α∂βµ is compact. (3.13) POLYANALYTIC BERGMAN SPACES 11 4. Examples We give some examples of symbols -- distributions and -- hyperfunctions. More examples can be constructed, following the pattern seen in [7], [8]. Example 4.1. Let the measure µ be supported on the lattice L = Z + iN. With an integer point n = (n1 + in2) we assign the weight mn > 0. Suppose that supn mn < ∞. Then the Toeplitz operator, with the measure µ =Pn mnδ(z−n) as symbol, is bounded. Example 4.2. In the setting of Example 4.1, consider the Toeplitz operator with distributional symbol µα,β,W = W (n)∂α∂βµ, where W (n) is a weight function, W (n1 + in2) = n2−α−β. Then the conditions of Theorem 3.9 are satisfied for the measure W (n)µ and, therefore, the Toeplitz operator with symbol µα,β,W is bounded. If W (n) is a function on the lattice satisfying W (n) → 0 as n → ∞ then, by Theorem 3.10, this Toeplitz operator is compact. Example 4.3. In the setting of Example 4.1, consider the, initially formal, sum a =Xn∈L W (n)∂αn∂βnµ({n}), where (αn, βn) is a collection of orders of differentiation and W (n) is a weight sequence. This is a sum of distributions supported at single points of the lattice L. To each of them, we can apply Theorem 3.9 and obtain an estimate of the norm of the corresponding sesquilinear form Fn, where the order kn = αn + βn is involved: Fn[f, g] ≤ CW (n)((αn!βn!)(γ/2)−(αn+βn)n−αn+βn 2 = CW (n)τ (n)kfkkgk. kfkkgk (4.1) A rough way to estimate the sesquilinear form Fa would be to consider the sum of the terms in (4.1), Fa[f, g] ≤X(W (n)τ (n))kfkkgk, (4.2) so the sesquilinear form is bounded as soon as the series in (4.2) converges. A more exact treatment uses the finite multiplicity covering by disks containing no more than, say, 10 points of the lattice L similarly to how this was done in Section 2. In this way, the sesquilinear form is majorated by a smaller quantity, Fa[f, g] ≤ sup{(W (n)τ (n))kfkkgk}, for which the finiteness condition requires a considerably milder decay require- ment for W (n) than the finiteness of the coefficient in (4.2). Now we consider some symbols with support touching the boundary of the upper half-plane Π. 12 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI µ =Pj µj is a measure on Π, which, however, does not have compact support in Example 4.4. Let Lj ⊂ Π be the straight line {z = x + iy : x ∈ R, y = 2−j}, and µj be the measure W (j)δ(Lj) with some weight sequence W (j). In other words, it is the Lebesgue measure on the line Lj with the weight factor W (j). The sum Π. By Definition 3.6, the measure µ belongs to the class M0,γ (say, for γ = 1 2), as soon as W (j) = O(2−2j). Example 4.5. For the same system of straight lines Lj we consider ς =Xj W (j)ςj ≡Xj W (j)∂jδ(Lj) =Xj (i/2)jW (j)(1 ⊗ ∂j yδ(y − 2−j)). (4.3) In (4.3), ς is a formal sum of distributions W (j)ςj, each being a derivative of a measure, of unbounded orders, and this formal sum corresponds to the sesquilin- ear form Fς [f, g] =Xj Fςj [f, g] =Xj (−1)jW (j)ZLj ∂jf · ¯gdx. (4.4) The sequence of weights W (j) should be chosen in such a way that the sum (4.4) converges for f, g ∈ A1 and, moreover, is a bounded sesquilinear form on A1. By Theorem 3.8, the measure µj = 1 ⊗ δ(y − 2−j) belongs to the class Mj,γ with estimate Thus, if the sum Pj W (j)(2/γ)2j(j!)2 is finite, the sesquilinear form (4.4) is bounded on A1. j,γ(µj) ≤ C(2/γ)2j(j!)2. 5. The structure of the Bergman spaces. j Along with the Bergman space A1 of analytic functions on Π, we consider spaces of polyanalytic functions. We denote by Aj, j = 1, 2, . . . the space of square integrable functions on Π satisfying the iterated Cauchy-Riemann equation ∂ f = 0 (the reader was, probably, intrigued by the subscript in the notation A1 -- now its use is justified). Of course, Aj ⊂ Aj ′ for j < j′, so, to get rid of these 'less polyanalytic' functions, the true polyanalytic Bergman spaces have been introduced (see [13]), by Worth mentioning is the following direct sum decomposition of L2(Π): j−1, j = 2, . . . ; A(1) = A1. A(j) = Aj ⊖ Aj−1 = Aj ∩ A⊥ L2(Π) =Mn∈N A(j) ⊕Mn∈N eA(j), where eA(j) are true poly-antianalytic Bergman spaces (see, for details, [13]). For these poly-Bergman spaces on the upper half-plane, there exists a system of creation and annihilation operators, described in [4, 14]. These operators are POLYANALYTIC BERGMAN SPACES 13 two-dimensional singular integral operators, (SΠu)(w) = − 1 πZΠ u(z)dA(z) (z − w)2 and (S∗ Πu)(w) = − 1 πZΠ u(z)dA(z) (¯z − ¯w)2 . Being understood in the principal value sense, they are bounded in L2(Π) and ad- joint to each other. They are, in fact, the Beurling -- Ahlfors operators compressed to the half-plane, and are surjective isometries, SΠ : A(j) → A(j+1), S∗ S∗ Π : A(j) → A(j−1), j > 1, j > 1, (5.1) while Π : eA(j) → eA(j+1), SΠ : eA(j) → eA(j−1), Π : A1 → {0}, SΠ : fA1 → {0}. Thus, in particular, we have surjective isometries (SΠ)jA(1) = A(j+1)(Π). S∗ Formulas (5.1), possessing the structure similar to the ones of the Landau subspaces for the Schrodinger equation with uniform magnetic field, justify calling SΠ, S∗ Π creation and annihilation operators. Remark 5.1. Here one can notice a certain discrepancy in notations: S denotes usually the annihilation operator in the poly-Fock spaces while SΠ denotes here the creation operator in the poly-Bergman spaces -- however, this is the tradition and we do not want to break it. The operators Sj Π, restricted to A1, admit a representation, found in [6], which is much more convenient for using in further reductions. Theorem 5.2 ([6, Theorem 3.3]). For u ∈ A1, (Sj Πu)(z) = ∂j[(z − ¯z)ju(z)] j! , j ≥ 0. (5.2) Note that the isometry U of the Bergman spaces on the upper half-plane Π and on the disk D is not carried over to the poly-analytic Bergman spaces. 6. Relations among the Toeplitz operators in the true poly-Bergman spaces. Let a be a distribution on Π, defined at least on C ∞( ¯Π) ∩ L1(Π). We consider the sesquilinear form u = Sj Fa[u, v] = (au, ¯v) = (a, u¯v), (6.1) with f, g being elements of the standard orthonormal basis in A(1). The sesquilin- ear form (6.1) is defined for f, g in the basis in A(1) and can be extended by sesquiliearity to the linear span of the basis. If it turns out that (6.1) is bounded on this span, it can be extended by continuity to the whole A(1) and thus it would Πf ∈ A(j+1), Πg ∈ A(j+1), v = Sj 14 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI define a bounded operator in A(1). On the other hand, since Sj Π is a unitary op- erator, Fa[u, v] can be therefore extended to a bounded sesquilinear form defined for u, v being arbitrary elements in A(j+1), thus defining a bounded operator in A(j+1). The present section is devoted to finding an explicit relation between these two operators. Further on, in this section we will only use the Hilbert space L2(Π, dA) and therefore we will suppress the notation of the space in the scalar product h., .i; the parentheses (., .) still denote the action of a distribution on a smooth function, without the complex conjugation. Since u, v are elements in the poly-Bergman space, the product u¯v is a smooth function in L1(Π). The following result leads to establishing a relation between Toeplitz operators in the true poly-Bergman space A(j) and the Bergman space A1. Proposition 6.1. Let a be a distribution in the half-plane Π. Then Πg) = h(K(j)a)f, gi, Fa[u, v] = Fa[Sj Πg] = (a, Sj Πf, Sj Πf Sj (6.2) Π defined in (5.2), where K(j) is a differential operator of order 2j having with Sj the form K(j) = K(j)(∆(y2·), ¯∂(y·), ∂(y·)), (6.3) and K(j) being a polynomial of degree j. Moreover, if we assign the weight −1 to the differentiation and the weight 1 to the multiplication by y, with weights adding under the multiplication, then all monomials in K(j) have weight 0. Proof. We demonstrate the reasoning for the case j = 1. The general case uses the same machinery with some tedious bookkeeping. We set u = SΠf , v = SΠg and consider the sesquilinear form Fa[f, g] = ha∂(yf ), ∂(yg)i ≡ (a, ∂(yf )∂(yg)) for f, g being some elements in the standard orthonormal basis in A(Π). Due to ∂∗ = − ¯∂: (a, ∂(yf )∂(yg)) = (a, (−if + y∂f )(−ig + y∂g)) = (a, f ¯g) + (a,−if y ¯∂g) + (a, i¯gy∂f ) + (a, y2(∂f ) ¯∂¯g) (6.4) (the last transformation uses ¯∂g = 0). For the second term on the right-hand side in (6.4), by the general rules of manipulation with distributions, we have (a,−if y ¯∂g) = (−iya, f ∂g) = (−iya, ∂( ¯f g)) − (−iyF, ( ¯∂f )g) = ( ¯∂(iya), f ¯g), because ¯∂f = 0. The third term on the right in (6.4) is transformed in a similar way, and for the last one, (a, y2∂f ¯∂¯g) = (y2a, ¯∂(∂f ¯g) − ¯∂(∂f )¯g) = −( ¯∂(y2a), (∂f )¯g) = −( ¯∂(y2a), ∂(f ¯g) − {f ∂¯g}) = (∂ ¯∂(y2a), f ¯g), again, the terms in curly bracket vanishing due to ¯∂g = 0. Collecting the terms in (6.4), after simple transformations, we obtain the required relation. For higher order, the procedure of transformation is similar, by means of for- Π in the expression mally commuting a and factors in the creation operators Sj POLYANALYTIC BERGMAN SPACES 15 Πf Sj Πf, Sj Πgi ≡ (a, Sj haSj Πg), so that the Cauchy-Riemann operator falls on the functions f, g, while any commutation with a produces a derivative of a. It remains to notice that when commuting the terms in the expression on the left- hand side in (6.2), the weight of the terms does not change. Alternatively, one can make the calculations similar to the ones shown above, again by moving the Cauchy-Riemann operator to the functions f, g and on F . To make the general reasoning more transparent, we present here our transfor- mations for the case j = 2. So, we set u = S2 Fa[u, v] = Fa[S2 Πf, S2 Πg] = ((S2 Πf )a, S2 Πf, v = S2 Πg. We start with Πg) = −2((∂2(y2f ) × a), ¯∂2(y2¯g)). (6.5) We expand in (6.5) the derivatives of the product by the Leibnitz formula, to obtain Fa[u, v] = (∂2((y2f )a) − 2∂((y2f )∂a) + y2f ∂2a, ¯∂2(y2¯g). (6.6) Now we carry over the derivatives ∂, ∂2 to the second factor in (6.6) (this is legal due to the definition of the derivatives of distributions): Fa[u, v] = ((y2f )a, ∂2( ¯∂2(y2¯g))) (6.7) +2((y2f )∂a, ∂( ¯∂2(y2¯g))) + (y2f ∂2a, ¯∂2(y2¯g). We consider then the terms in (6.7) separately. In the first term, we commute ∂2 and ¯∂2 in the second factor: ∂2( ¯∂2(y2¯g)) = ¯∂2(∂2(y2¯g)) = 1 2 ¯∂2(¯g), since ∂¯g = 0. Therefore, the first term in (6.7) transforms to 2 (f ¯∂2(y2F ), ¯g) 2 ((y2f )a, ¯∂2¯g) = 1 1 2 ( ¯∂2(y2f a), ¯g) = 1 = (f K1a, ¯g) = (K1a, f ¯g), with K1a being the distribution Next, for the second term in (6.7), we have K1a = 1 2 ¯∂2(y2a). 2((y2f )∂a, ∂( ¯∂2(y2¯g))) = −2((y2f )∂a, ¯∂2∂(y2¯g))) = 2((y2f )∂a, ¯∂2(y¯g)) = 2( ¯∂2(y2f ∂a), y¯g). Now, ¯∂2(y2f ∂a) = f ¯∂2(y2∂a) = f (2∂a + 2y ¯∂∂a + y2 ¯∂2∂a). Thus, the second term in (6.7) equals to 2((y2f )∂a, ∂( ¯∂2(y2¯g))) = (f K2a, ¯g) = (K2a, f ¯g), where K2a = y(2∂a + 2y ¯∂∂a + y2 ¯∂2∂a). 16 GRIGORI ROZENBLUM AND NIKOLAI VASILEVSKI Finally, the third term in (6.7) is transformed as (y2f ∂2a, ¯∂2(y2¯g) = ( ¯∂2f y2∂2a, y2¯g) = (f ¯∂2(y2a), y2¯g) = (f y2 ¯∂2(y2a), ¯g) = (f K3a, ¯g) = (K3a, f ¯g), where K3a = y2 ¯∂2(y2a). A simple bookkeeping shows that the operators K1, K2, K3 have the structure claimed by the theorem, K(2) = K1+K2+K3. Note again that although the distribution a does not necessarily have compact support in Π, our definition of derivatives of such distributions conserves the formal differentiation rules we used in these calculations. (cid:3) As explained above, the equality (6.2) extends to the whole of A1, as soon as we know that the right-hand side or on the left-hand side is a bounded sesquilinear form in A1. Thus, the statement of Proposition 6.1 can be formulated as the following theorem. Theorem 6.2. The operators Ta(A(j+1)) and TKa(A1) are unitarily equivalent (up to a numerical factor) as soon as one of them is bounded. In this case, if one of these operators is compact, or belongs to a Schatten class, or is of finite rank, or zero, then the same holds for the other one. The terms in the differential operator K(j) can be regrouped so that it takes the form K(j) = Xp+¯p+2q≤2j bp,¯p,q(∂)p( ¯∂)¯p∆qyp+¯p+2q. Now we can apply the boundedness conditions obtained earlier, in Section 3, for differential sesquilinear forms to obtain boundedness conditions for Toeplitz operators in true poly-Bergman spaces. Theorem 6.3. Let µ be a measure on Π such that ykµ are k-C measures for A1 defines a bounded Toeplitz operator in A(j+1). If, moreover, ykµ are vanishing k-C measures for A1, then the corresponding operator in A(j+1) is compact. for k = 0, 1, . . . , 2j. Then the sesquilinear formR f ¯gdµ is bounded in A(j+1) and References [1] A. Alexandrov, G. Rozenblum, Finite rank Toeplitz operators: some extensions of D.Luecking's theorem, J. Funct. Anal. 256 (2009) 2291 -- 2303. [2] H. Hedenmalm, B. Korenblum, and K. Zhu, Theory of Bergman spaces. Graduate Texts in Mathematics, 199. Springer-Verlag, New York, 2000. x+286 pp. [3] S.H.Kang, J.Y.Kim, Toeplitz operators on Bergman spaces defined on upper planes. Comm. Korean Math. Soc., 14 (1), 1999, 171 -- 177. [4] Yu. Karlovich, L. Pessoa, C ∗-Algebras of Bergman Type Operators with Piecewise Contin- uous Coefficients. Int. Equat and Oper. Theory, 57 (4) (2007) 521 -- 565. [5] O. Parfenov, The asymptotic behavior of singular numbers of imbedding operators of some classes of analytic functions. (Russian) Mat. Sb. (N.S.) 115(157) (1981), no. 4, 632 -- 641; translation: Mathematics of the USSR-Sbornik, 1982, 43:4, 563 -- 571. POLYANALYTIC BERGMAN SPACES 17 [6] L. Pessoa, A.M. Santos, Theorems of Paley -- Wiener type for spaces of polyanalytic functions, Current trends in analysis and its applications, 605 -- 613, Trends Math., Birkhauser/Springer, Cham, 2015. [7] G. Rozenblum, N. Vasilevski,Toeplitz operators defined by sesquilinear forms: Fock space case. Journal of Functional Analysis, 267 (2014), 4399 -- 4430. [8] G. Rozenblum, N. Vasilevski, Toeplitz operators defined by sesquilinear forms: Bergman space case, J. Math. Sci. (N.Y.) 213 (2016), no. 4, 582 -- 609. [9] G. Rozenblum, N. Vasilevski, Toeplitz operators in the Herglotz space. Integral Equations Operator Theory 86 (2016), no. 3, 409 -- 438 [10] G. Rozenblum, N. Vasilevski, Toeplitz operators in polyanalytic Bergman type spaces, Func- tional Analysis and Geometry. Selim Grigorievich Krein Centennial. Contemporary Math- ematics, 733, AMS, (2019). [11] A. S´anchez-Nungaray, N. Vasilevski, Toeplitz operators on the Bergman spaces with pseu- dodifferential defining symbols. Operator theory, pseudo-differential equations, and math- ematical physics. Operator Theory: Advances and Applications, v. 228 (2013), 355-374. [12] J. Taskinen, J. Virtanen, Toeplitz operators on Bergman spaces with locally integrable symbols. Rev. Mat. Iberoam. 26 (2010), no. 2, 693 -- 706. [13] N. Vasilevski, On the structure of Bergman and poly-Bergman spaces. Integral Equations Operator Theory 33 (1999), no. 4, 471 -- 488. [14] N. Vasilevski, Poly-Bergman spaces and two-dimensional singular integral operators. The Extended Field of Operator Theory, 349 -- 359, Oper. Theory Adv. Appl., 171, Birkhauser, Basel, 2007. [15] N. Vasilevski, Commutative Algebras of Toeplitz Operators on the Bergman Space. Birkhauser, 2008. [16] K. Zhu, Operator theory in function spaces. Second edition. Mathematical Surveys and Monographs, 138. American Mathematical Society, Providence, RI, 2007. [17] K. Zhu, Spaces of holomorphic functions in the unit ball. Graduate Texts in Mathematics, 226. Springer-Verlag, New York, 2005. [18] K. Zhu, Analysis on Fock Spaces, Graduate Texts in Mathematics, 263. Springer-Verlag, New York, 2012. . Chalmers University of Technology and The University of Gothenburg (Swe- den); St.Petersburg State University, Dept. Math. Physics (St.Petersburg, Russia) E-mail address: [email protected] Cinvestav (Mexico, Mexico-city) E-mail address: [email protected]
1711.05110	1	1711	2017-11-14T14:32:28	Operator inequalities implying similarity to a contraction	[ "math.FA" ]	Let $T$ be a bounded linear operator on a Hilbert space $H$ such that \[ \alpha[T^,T]:=\sum_{n=0}^\infty \alpha_n T^{n}T^n\ge 0. \] where $\alpha(t)=\sum_{n=0}^\infty \alpha_n t^n$ is a suitable analytic function in the unit disc $\mathbb{D}$ with real coefficients. We prove that if $\alpha(t) = (1-t) \tilde{\alpha} (t)$, where $\tilde{\alpha}$ has no roots in $[0,1]$, then $T$ is similar to a contraction. Operators of this type have been investigated by Agler, M\"uller, Olofsson, Pott and others, however, we treat cases where their techniques do not apply. We write down an explicit Nagy-Foias type model of an operator in this class and discuss its usual consequences (completeness of eigenfunctions, similarity to a normal operator, etc.). We also show that the limits of $\\|T^nh\\|$ as $n\to\infty$, $h\in H$, do not exist in general, but do exist if an additional assumption on $\alpha$ is imposed. Our approach is based on a factorization lemma for certain weighted $\ell^1$ Banach algebras.	math.FA	math	OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION GLENIER BELLO-BURGUET AND DMITRY YAKUBOVICH Abstract. Let T be a bounded linear operator on a Hilbert space H such that α[T ∗, T ] := ∞ X n=0 αnT ∗nT n ≥ 0. n=0 αntn is a suitable analytic function in the unit disc D with real coefficients. where α(t) = P∞ We prove that if α(t) = (1 − t) α(t), where α has no roots in [0, 1], then T is similar to a contraction. Operators of this type have been investigated by Agler, Muller, Olofsson, Pott and others, however, we treat cases where their techniques do not apply. We write down an explicit Nagy-Foias type model of an operator in this class and discuss its usual consequences (completeness of eigenfunctions, similarity to a normal operator, etc.). We also show that the limits of kT nhk as n → ∞, h ∈ H, do not exist in general, but do exist if an additional assumption on α is imposed. Our approach is based on a factorization lemma for certain weighted ℓ1 Banach algebras. Let H be a separable complex Hilbert space and denote by L(H) the set of bounded linear operators on H. Let T ∈ L(H) and let 1. Introduction α(t) = αntn ∞Xn=0 ∞Xn=0 be an analytic function on the open unit disc D = {z : z < 1} such that αn are real and Pn αnkT nk2 < ∞. Then we define the so-called hereditary calculus α[T ∗, T ] := αnT ∗nT n. This work is devoted to the study of operators T ∈ L(H) that satisfy an operator inequality (1) α[T ∗, T ] ≥ 0. Notice that for α(t) = 1 − t, this is just the class of all contractions on H. The study of operator inequalities of this type was originated in the work by Agler [1], where he studied more general inequalities of the form Pj,k αjkT ∗jT k ≥ 0. Suppose that k(w, z) = 1/(cid:0)Pj,k αjkwjzk(cid:1) is analytic in D×D and k( ¯w, z) is a reproducing kernel that defines it is unitarily equivalent to the operator L∞ a functional Hilbert space Hk(D) of functions on D. Let M be the operator M u(z) = zu(z), acting on Hk(D), and assume that it is bounded. Agler's main result in [1] asserts that an operator T whose spectrum σ(T ) is contained in D satisfies the above inequality if and only if j=1 M ∗ restricted to an invariant subspace. This is what Agler calls a coanalytic model of T . Date: July 11, 2018. First author acknowledge the Grant Severo-Ochoa La Caixa for undergraduate studies. Both authors are partially supported by Plan Nacional I+D grant no. MTM2015-66157-C2-1-P. The authors also acknowledge financial support from the Spanish Ministry of Economy and Competitiveness, through the "Severo Ochoa Programme for Centres of Excellence in R&D" (SEV-2015-0554). 1 OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 2 The condition σ(T ) ⊂ D is too restrictive; it has been shown in subsequent papers that in many cases it can be replaced by σ(T ) ⊂ D. Then, in general, instead of the operator j=1 M ∗ ⊕ U , where U is a unitary. For instance, in [2], Agler proved that an operator T is a hypercontraction of order n (that is, it satisfies (1) for α(t) = (1 − t)k, k = 1, . . . , n) iff it can be extended to an operator j=1 M ∗ ⊕ U , where M acts on the Bergman space, whose reproducing kernel is 1/(1 − ¯wz)n. L∞ j=1 M ∗, the above coanalytic model involves a direct sum L∞ L∞ If T is of class C0 • (that is, T n → 0 strongly), the above unitary summand is absent. The case where α is a polynomial, say α = p, was studied further by Muller in [16]. He considers the class C(p) of operators T ∈ L(H) such that (1) is satisfied and proves that T has a coanalytic model whenever p(1) = 0, 1/p(t) is analytic in D and 1/p( ¯wz) is a reproducing kernel. Notice that the last condition is equivalent to the fact that all Taylor coefficients of 1/p(t) at the origin are positive. In [21], Olofsson deals with a more general setting, when α is not a polynomial. Suppose an analytic function α(t) on D satisfies α 6= 0 in D and 1/α has positive Taylor coefficients at the origin. Olofsson studies contractions T on H that satisfy α[rT ∗, rT ] ≥ 0 for all r, 0 ≤ r < 1 (he imposes some more assumptions on α). He obtains the coanalytic model for this class of operators. Certain types of operator inequalities like (1) have also been studied for commuting tuples of operators in [3], [22], [8] and other papers. Pott in [22] considered positive regular polyno- mials. These are polynomials of several complex variables with non-negative coefficients such that the constant term is 0 and the coefficients of the linear terms are positive. Given such polynomial p, Pott constructed a dilation model for commuting tuples of operators satisfying the positivity conditions (1 − p)k[T ∗, T ] ≥ 0 for 1 ≤ k ≤ m (see Theorem 3.8 in [22]). In [8], Bhattacharyya and Sarkar define the characteristic function θT for this class of tuples and construct a functional model in the pure case (see Theorem 4.2 in [8]). A general framework of Agler's theory in case of operator inequalities for commuting tuples of operators has been given in [3] and further generalized in [4]; the latter work treats gen- j=1 M ∗ ⊕ U eral analytic models, which involve multi-dimensional analogues of operators L∞ attached to a domain in Cn. class of operator inequalities, Here we restrict ourselves to a single operator, but for this case, we can deal with a large for which the original Agler's approach does not seem to apply. In what follows, we will say that an analytic function α(t) is admissible if it has the form If α[T ∗, T ] ≥ 0, then T is similar to a contraction. Theorem 1.1. Let T ∈ L(H) be an operator whose spectrum is contained in the closed unit α(t) = (1 − t)eα(t), whereP∞ disc D. Let α(t) = (1 − t)eα(t) be an admissible function such thatP∞ Notice that P∞ n=0 eαn < ∞ and eα is positive on [0, 1] (in particular, α0 > 0). n=0 eαn(1 + kT nk2) < ∞. n=0 eαn(1 + kT nk2) < ∞ implies that P∞ If α(t) is an admissible function and an operator T on H is related to α as in the above theorem, then we will say that T belongs to the class Cα (see the definition at the beginning of Section 3). One of our main results is as follows. the operator α[T ∗, T ] is well-defined. n=0 αn(1 + kT nk2) < ∞, so that An important particular case is when α is analytic on a disc t < R of radius R > 1, in particular, if α is a polynomial or is rational. In this case, α is admissible whenever α(t) > 0 on [0, 1) and α(t) has a simple root at t = 1. We get that given a function α of this type and a Hilbert space operator T , whose spectral radius is less than or equal to 1, T is similar to a contraction whenever α[T ∗, T ] ≥ 0. We remark that the condition that α has no roots in [0, 1] has a clear spectral meaning. In- deed, the eigenvalues and, more generally, the approximate point spectrum of T are contained in {z ∈ D : α(z2) ≥ 0}. As it is seen from the example of normal operators, under the above OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 3 condition, the approximate point spectrum of T can be whatever closed subset of the closed unit disc. The main difference with the approach originated by Agler is that here we do not need to assume that 1/α( ¯wz) is a reproducing kernel. That is, we admit that some of the Taylor coefficients of 1/α(t) at t = 0 can be negative and we also allow α to have zeros in D (excluding the interval [0, 1]). Notice that we only get similarity to a contraction; in fact, there are many operators with kT k > 1, to which our results apply. Notice that the majority of papers based on Agler's approach (for the case of a single operator) deal only with contractions (see the Remark 4.3 below). Our main tool for proving Theorem 1.1 is related with Banach algebras. We say that a n=0 of positive real numbers is a good weight if sequence ω = {ωn}∞ (GW 1) ωn ≥ 1 for every n, (GW 2) ωnωm ≥ ωn+m (submultiplicative property) for every n, m, (GW 3) ω1/n n → 1. Given a good weight ω, we define the corresponding weighted Wiener algebra Aω as the following set of analytic functions: Aω :=nα(t) = For analytic functions f (t) = P∞ αntn : ∞Xn=0 ∞Xn=0 αnωn < ∞o. It is immediate to check that Aω is a commutative, unital Banach algebra of analytic functions in D. n=0 gntn, we will use the notation f < g when fn ≥ gn for every n ≥ 0 and the notation f ≻ g when f < g and f0 > g0. To prove Theorem 1.1, the following lemma on factorization in the algebra Aω will be used. n=0 fntn and g(t) = P∞ Lemma 1.2. Let ω be a good weight. If f ∈ Aω is a positive function on [0, 1], then there exists a function g ∈ Aω such that g ≻ 0 and f g ≻ 0. In Section 3, we collect some elementary properties of classes Cα; Proposition 3.1 and Lemma 3.2 give some examples. In particular, we show that any diagonalizable matrix with spectrum on the unit circle belongs to Cα for some admissible function α. Given an operator T of class Cα, in Section 6 we will write down its concrete coanalytic model, in other words, an explicit Nagy-Foias-like functional model of T up to similarity. To construct a functional model of a contraction, first one has to single out its unitary part (recall that the Nagy-Foias construction "forgets" this part). Section 5 is devoted to defining the unitary part of an operator T ∈ Cα, which is a necessary first step to passing to the Nagy-Foias transcription. The standard Nagy-Foias model of a contraction S ∈ L(H) makes use of its defect operator, which is defined as a nonnegative square root DS = (I − S∗S)1/2. This model is related with the following well-known identity (2) khk2 = ∞Xn=0 kDSSnhk2 + lim n→∞ kSnhk2, h ∈ H, valid for any contraction S (see [18, Section 1.10]). This motivates the next definition, which will be useful for us. Definition 1.3. Let T ∈ L(H) be a power bounded operator (that is, supn≥0 kT nk < ∞), and let D : H → F , where H, F are Hilbert spaces. We will say that D is an abstract defect operator for T if there are some positive constants c, C such that for any h ∈ H, (3) ckhk2 ≤ ∞Xn=0 kDT nhk2 + lim sup n→∞ kT nhk2 ≤ Ckhk2 . OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 4 By applying the Banach limit, we will show that a power bounded operator is similar to a contraction if and only if it has an abstract defect operator. More precisely, we will prove the following. Lemma 1.4. Let T ∈ L(H) be a power bounded operator. Then an operator D ∈ L(H) is an abstract defect operator for T if and only if there exists an invertible operator W ∈ L(H) such that eT := W T W −1 is a contraction and kDhk = kD eT W hk for any h ∈ H. The Nagy-Foias-like model we give is in some aspects close to [24]. However, here we deal with a general case and not only with a C0 • case, as in [24]. Since the model is only up to similarity, it is not unique, but we suggest a reasonable choice. It will be proven that for an operator T ∈ Cα, (α[T ∗, T ])1/2 can be taken as an abstract defect operator of T . This will permit us to write down explicitly an analogue of the characteristic function of T ∗. Our Nagy-Foias-like transcription implies the major part of usual consequences of the Nagy- Foias theory (such as criteria for completeness of eigenvectors of T in terms of the determinant of Θ∗, criteria for these to form a Riesz basis, similarity to a normal operator, etc). These criteria are formulated in terms of the determinant of Θ∗(z). In Section 7 we show that, roughly speaking, Θ∗ has a determinant whenever α[T ∗, T ] is of trace class and discuss briefly the above-mentioned consequences. In Section 8, we will give necessary and sufficient conditions for the inclusion of operator classes Cα ⊂ Cτ . It will follow, in particular, that there are many functions τ such that the class Cτ strictly contains C1−t, the class of all contractions on H. As compared with Agler's case, the construction of the Nagy-Foias model in our case has some extra difficulties. They are related with the fact that for operators of class Cα in general, the limit limn kT nhk2, h ∈ H, does not exist (and therefore we need in general Banach limits). In Section 9, we prove, that these limits do exist under the additional assumption that α has no roots on the unit circle (except for the root at t = 1). 2. Proof of Lemma 1.2 on factorization in the Banach algebra Aω If ωn = 1 for all n, then the algebra Aω is just the usual Wiener algebra (which we denote AW ) of analytic functions in D with absolutely summable Taylor coefficients. In fact, if we denote by H(D) the set of functions analytic on (a neighborhood of) D, then, obviously, H(D) ⊂ Aω ⊂ AW for every good weight ω. The first inclusion is due to the exponential decay of Taylor coeffi- cients of functions in H(D). Lemma 2.1. If q(t) = (t − λ)(t − λ) for some λ ∈ C \ R, then there exists a polynomial p such that p ≻ 0 and pq ≻ 0. Proof. Let m be the smallest nonnegative integer such that Re(λ2m ) ≤ 0. We define p(t) := (t2j m−1Yj=0 + λ2j )(t2j + ¯λ2j ) (so that p(t) = 1 if m = 0). Note that by the minimality of m, for each factor we have (t2j + λ2j )(t2j + ¯λ2j ) = t2j+1 + 2 Re(λ2j )t2j + λ2j ≻ 0. Therefore p ≻ 0. Moreover (pq)(t) = (t2m − λ2m )(t2m − ¯λ2m ) = t2m+1 − 2 Re(λ2m )t2m + λ2m ≻ 0. (cid:3) Corollary 2.2. If q is a real polynomial without real roots and q(0) > 0, then there exists a polynomial p such that p ≻ 0 and pq ≻ 0. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 5 Proof. Note that q = Cq1 · · · qk, where C is a positive constant and each factor qj has the form qj(t) = (t − λj)(t − λj) for some λj ∈ C \ R. Then for each factor qj we can construct (cid:3) the polynomial pj as in the previous lemma and we just take p = p1 · · · pk. Corollary 2.3. If q is a real polynomial such that q(t) > 0 for every t ∈ [0, 1], then there exists a function u ∈ H(D) such that u ≻ 0 and uq ≻ 0. Proof. Decompose q as the product of polynomials qnr, q+ and q−, where the roots of qnr are nonreal, the roots of q+ are positive and the roots of q− are negative. Without loss of generality, q+(0) = 1, q−(0) = 1 and qnr(0) = q(0) > 0. Therefore q− ≻ 0 and by Corollary 2.2, there exists a polynomial p such that p ≻ 0 and pqnr ≻ 0. Notice that 1/q+ ∈ H(D) and (cid:3) 1/q+ ≻ 0. Hence, we can take u := p/q+, and the statement follows. Proof of Lemma 1.2. Let f (t) > ε > 0, for t ∈ [0, 1]. Take N ∈ N such thatP∞ ε/2. Hence it is obvious thatP∞ NXn=0 + Xn≥N +1; fn<0 n=N +1 fn < ε/2. Put fN (t) = fntn − fntn. h(t) = ε 2 , ε 2 n=N +1 fnωn < Then fN is a polynomial and h ∈ Aω. Since fN (t) > ε − ε/2 − ε/2 = 0 for t ∈ [0, 1], we can apply Corollary 2.3 to obtain a function u ∈ H(D) such that u ≻ 0 and ufN ≻ 0. Hence u ∈ Aω. On the other hand, for t ∈ D we have (cid:12)(cid:12)(cid:12) Xn≥N +1; fn<0 fntn(cid:12)(cid:12)(cid:12) ≤ ∞Xn=N +1 fn < ε/2. Hence h(t) 6= 0 for t ∈ D. Notice that the properties of the weight imply that the characters of Aω are exactly the evaluation functionals at the points of D. It follows that v := 1/h ∈ Aω. Note that v ≻ 0, because it has the form c/(1 − a), where a ∈ Aω, a ≻ 0 and c = ε/2. Put g := uv ∈ Aω. Then g ≻ 0 and since f < fN + h, we have gf < g(fN + h) = vufN + u ≻ 0. (cid:3) 3. The classes Cα and Proof of Theorem 1.1 class of operators In what follows, α(t) = (1 − t)eα(t) will be an admissible function. We associate to it the For example, C1−t is just the set of all contractions in L(H). Notice that any admissible this case, any T ∈ L(H) with σ(T ) ⊂ D that satisfies α[T ∗, T ] ≥ 0 is in Cα. function α ∈ H(D) has a simple root at t = 1, and the corresponding eα is also in H(D). In Theorem 1.1 asserts that if T ∈ Cα for some admissible function α then T is similar to a contraction. Here are some elementary properties of the classes Cα. Proposition 3.1. (a) If N ∈ L(H) is a normal operator with kN k ≤ 1, then N ∈ Cα for every admissible function α. In particular, all unitary operators are in Cα. (b) If T1, T2 ∈ Cα, then the orthogonal sum T1 ⊕ T2 also is in Cα. (c) It T ∈ Cα, then ζT ∈ Cα for every ζ on the unit circle T = {z : z = 1}. (d) If T ∈ Cα, then T L ∈ Cα for every T -invariant subspace L ⊂ H. (e) If T is Hilbert space operator, whose spectral radius is less than one, then T ∈ C1−tn for any sufficiently large n > 0. Cα := {T ∈ L(H) : σ(T ) ⊂ D, ∞Xn=0 eαn(1 + kT nk2) < ∞, α[T ∗, T ] ≥ 0}. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 6 Proof. Statements (a)-(c) are obvious. (d) Put S := T L. The first two conditions for S ∈ Cα n=0 αnkT nhk2 ≥ 0 for every h ∈ H, in particular for every h ∈ L. (e) Note that σ(T ) ⊂ D implies that kT nk < 1 (cid:3) for n ≫ 0. follows from kSnk ≤ kT nk, and the positivity condition is obvious sinceP∞ Lemma 3.2. For any complex square matrix T without nontrivial Jordan blocks such that σ(T ) ⊂ T, there exists an admissible polynomial p(t) such that p[T ∗, T ] = 0. Proof. Suppose T is of size n × n. Let {vj} (1 ≤ j ≤ n) be a basis of eigenvectors of T in Cn and let λj ∈ T be the corresponding eigenvalues. Consider the polynomial p(t) = (1 − t) Y1≤k<ℓ≤n (1 − 2 Re(λk ¯λℓ)t + t2). Notice that it only has roots on the unit circle; it follows that p is admissible. We will prove that p[T ∗, T ] = 0. Each h ∈ Cn can be written as h =P hjvj. We get pjDXk hℓvℓE =Xj pjDT jXk λj khkvk,Xℓ λj ℓhℓvℓE (cid:10)p[T ∗, T ]h, h(cid:11) =Xj =Xk,ℓ Xj hkvk, T jXℓ k hk¯hℓhvk, vℓi =Xk,ℓ pjλj ℓλj p(λk ¯λℓ)hk¯hℓhvk, vℓi = 0, because p(λkλℓ) = 0 for all k, ℓ. (cid:3) For the proof of Theorem 1.1 we need some technical results. Let ω be a good weight and let f ∈ Aω. If T, B ∈ L(H) and T satisfies the condition kT nk2 . ωn (i.e., kT nk2 ≤ Cωn for some positive constant C), then the operator f [T ∗, T ](B) := fnT ∗nBT n ∞Xn=0 is well defined. Indeed, kf [T ∗, T ](B)k . kBkkf kAω . Note that, in particular, f [T ∗, T ](I) = f [T ∗, T ]. Lemma 3.3. Let ω be a good weight, f, g, h ∈ Aω such that f g = h and let T, B ∈ L(H). If kT nk2 . ωn then one has (i) h[T ∗, T ](B) = g[T ∗, T ](f [T ∗, T ](B)). (ii) h[T ∗, T ] = g[T ∗, T ](f [T ∗, T ]). Proof. Let us define g[N ](t) := gntn and hN := f g[N ] (N ≥ 0). NXn=0 Then, kg − g[N ]kAω −−−−→ N→∞ that 0 and kh − f g[N ]kAω = kf g − f g[N ]kAω −−−−→ N→∞ 0. It easily implies (4) k(g[N ]f )[T ∗, T ](B) − h[T ∗, T ](B)kL(H) −−−−→ N→∞ 0. Note that (znf )[T ∗, T ](B) = T ∗nf [T ∗, T ](B)T n for every n ≥ 0. Hence (g[N ]f )[T ∗, T ](B) = NXn=0 gnT ∗nf [T ∗, T ](B)T n, k(g[N ]f )[T ∗, T ](B) − g[T ∗, T ](f [T ∗, T ](B))kL(H) −−−−→ N→∞ 0. and therefore, (5) Formulas (4) and (5) give (i). To get (ii), one just has to put B = I. (cid:3) OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 7 If A, T ∈ L(H) and A is positive, it is said that T is an A-contraction if T ∗AT ≤ A. Remark 3.4. T is similar to a contraction if and only if T is an A-contraction for some A ≥ εI, where ε > 0. Proof of Theorem 1.1. Put ωn = 1+kT nk2. Since kT m+nk ≤ kT mkkT nk, the weight ω satisfies (GW 1) -- (GW 3). By Lemma 1.2, there exists a function eβ ∈ AW such that eβ ≻ 0 and f := eβeα ≻ 0. Then (1 − t)f = eβ(t)α(t) and by Lemma 3.3 (ii) we get ∞Xn=0eβnT ∗nα[T ∗, T ]T n ≥ 0. Hence T is a f [T ∗, T ]-contraction and the theorem follows from Remark 3.4. f [T ∗, T ] − T ∗f [T ∗, T ]T = (6) (cid:3) 4. The abstract defect operator of T Let us begin by recalling the notion of a Banach limit. If we denote by c the set of all convergent complex sequences then we can define the linear functional L : c → C given by n=1 ∈ c. It is immediate that kLk = 1, L(x′) = L(x) if L(x) = lim xn for every x = {xn}∞ x′ = {xn}∞ n=2, and also L(x) ≥ 0 if x ≥ 0 (i.e., xn ≥ 0 for every x). Using the Hahn-Banach Theorem, these properties of the limit functional can be extended to ℓ∞. Theorem A. There is a linear functional L : ℓ∞ → C such that (a) kLk = 1; (b) L(x) = lim xn for every x ∈ c; (c) L(x) ≥ 0 for every x ∈ ℓ∞ such that x ≥ 0; (d) L(x′) = L(x) if x ∈ ℓ∞ and x′ = {xn}∞ n=2; (e) lim inf xn ≤ L(x) ≤ lim sup xn if x ∈ ℓ∞ is a real sequence. Proof. Statements (a)-(d) are contained in [9, Theorem III.7.1] and assertion (e) is their easy (cid:3) consequence (and it is also standard). A functional L with the above properties is called a Banach limit. Proof of Lemma 1.4. We remark first that by a lemma by Gamal [10, Lemma 2.1], for any power bounded operator T , one has (7) lim inf n→∞ kT nhk2 ≍ lim sup n→∞ kT nhk2, h ∈ H (we say that two quantities A, B, depending on h or some other parameter, are comparable and write A ≍ B if there are two positive constants c, C such that cA ≤ B ≤ CA). Suppose first that there exists a linear isomorphism W ∈ L(H) with the properties stated in Lemma. Since eT = W T W −1 ∈ L(H) is a contraction, by (2) we have ∞Xn=0 Note that eT n = W T nW −1, and thus kD eTeT nhk2 + lim keT nhk2. kD eT W T nW −1hk2 + lim khk2 = kW T nW −1hk2. n→∞ n→∞ khk2 = ∞Xn=0 Since W is invertible and kDhk = kD eT W hk, we get (8) khk2 ≍ kW hk2 = kDT nhk2 + lim n→∞ kW T nhk2. ∞Xn=0 By (7), limn→∞ kW T nhk2 ≍ lim supn→∞ kT nhk2. We deduce that D is an abstract defect operator for T . OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 8 Conversely, suppose now that D is an abstract defect operator for T . Fix a Banach limit L and put (9) h2 := ∞Xn=0 kDT nhk2 + L(cid:0){kT nhk2}(cid:1). Notice that (7) and Theorem A (e) give that lim sup n→∞ kT nhk2 ≍ L ({kT nhk2}), h ∈ H. The relation (3) implies that h ≍ khk, h ∈ H. It follows that · is an equivalent Banach space norm on H. By applying the Cauchy-Schwarz inequality, it is easy to see that [x, y] := ∞Xn=0 hDT nx, DT nyi + L(cid:0){hT nx, T nyi}(cid:1) absolutely converges for any x, y ∈ H. It is a semi-inner product on H, which induces the norm · . So, in fact, · is a Hilbert space norm equivalent to k · k. (see [12] and [17] for a similar argument). Therefore there exists a linear isomorphism W : H → H such that kW hk = h. Observe that T h2 = h2 − kDhk2 ≤ h2. Let eT := W T W −1 ∈ L(H) (similar to T ). Take x ∈ H and put h := W −1x. We get so eT is a contraction. Since kDhk2 = h2 − T h2 and kD eT W hk2 = kW hk2 − keT W hk2 = h2 − T h2, we get kDhk = kD eT W hk for every h ∈ H. kW T hk ≤ kW hk (cid:3) Let α be an admissible function and let T ∈ Cα. We know already that T is similar to a B := (f [T ∗, T ])1/2, where the positive square root has been taken. Then B > εI for some ε > 0. We will assume, contraction. Since α ∈ AW , by Lemma 1.2, there exists a function eβ ∈ AW such that eβ ≻ 0 and f := eβeα ≻ 0. Hence (1 − t)f (t) = eβ(t)α(t). Set without loss of generality, thatP fk = kf kAW = 1. We put operator for T . More specifically, if eβ, f and B are as above, then the expression Theorem 4.1. If T ∈ Cα for some admissible function α ∈ AW , then D is an abstract defect D := (α[T ∗, T ])1/2. h2 := (10) (11) ∞Xn=0 kDT nhk2 + lim n→∞ kBT nhk2 defines an equivalent Hilbert space norm in H and T is a contraction with respect to this norm. In particular, the limit in (11) exists for every h ∈ H. Moreover, (12) h2 − T h2 = kDhk2 (∀h ∈ H). Proof. Since (1 − t)f (t) = eβ(t)α(t), we have B2 − T ∗B2T = βnT ∗nD2T n. ∞Xn=0 OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 9 Therefore, for every h ∈ H we have kBhk2 − kBT hk2 = Changing h by T jh we obtain βnkDT nhk2. ∞Xn=0 ∞Xn=0 kBT jhk2 − kBT j+1hk2 = βnkDT n+jhk2, for every j ≥ 0. Summing these equations for j = 0, 1, . . . , N − 1 we obtain kBhk2 − kBT N hk2 = N −1Xn=0 βnkDT nhk2 + βj kDT nhk2, j=0eβj. In particular, 0 < eβ0 ≤ βn and therefore by (13) we get nXj=n−N +1 ∞Xn=N kDT nhk2. kBhk2 ≥  N −1Xn=0 n=0 kDT nhk2 converges. On the other hand, since (13) where βn =Pn Hence the seriesP∞ βj ≤ ∞Xj=0 βnkDT nhk2 ≥ eβ0 N −1Xn=0 nXj=n−N +1 βj kDT nhk2 ≤ keβkAW ∞Xn=0 βnkDT nhk2 + lim N→∞ βj = keβkAW < ∞, ∞Xn=N kDT nhk2 → 0 we obtain that ∞Xn=N  nXj=n−N +1 when N → ∞. Therefore, taking limit in (13) when N goes to infinity we obtain that (14) kBhk2 = kBT N hk2 (and the limit in the right hand side exists for any h). Since eβ0 ≤ βn ≤ keβkAW < ∞, it follows that h defines an equivalent Hilbert space norm on H. Formula (12) is immediate from (cid:3) (11). Remark 4.2. Notice that Theorems 1.1 and 4.1 give two methods of finding an equivalent norm such that T is a contraction in this norm. With the method of Theorem 4.1 we obtained in addition that D is an abstract defect operator. at the origin. Then in the above calculations, we could set f (t) ≡ 1, so that B = I and j=1 M ∗ ⊕ U , where M is a weighted shift and U is unitary. In this situation, there is no need to assume that α is admissible and one can deal with more general functions. For the case when α is a polynomial, this a result by Muller [16, Theorem 3.10]. Most general result of this kind was given recently by Olofsson, see Theorem Remark 4.3. Assume that α(t) 6= 0 for t ∈ D, t 6= 1, and 1/eα has positive Taylor coefficients eβ = 1/eα ≻ 0. In this case, formula (14) yields a unitarily equivalent (coanalytic) model of T , which represent it as a part of an operatorL∞ 6.6 in [21]. On the other hand, in this case eβ ≻ 0, so that {βn} is an increasing sequence and In fact, in this setting, one has not to assume that α is admissible and can deal with more general functions. However, in both results cited above, T has to be a contraction. Theorem 4.1 requires α to be admissible, but applies to operators T that are not contractions. therefore the backward shift M ∗ has to be a contraction. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 10 Remark 4.4. Using the notation above, note that kBT nhk2 = ∞Xk=0 lim∗xn = lim n→∞ fkDT n+kh, T n+khE . ∞Xk=0 fkxn+k, kDT nhk2 + lim∗ n→∞kT nhk2. ∞Xn=0 So if we put (15) then (11) can be written as (16) h2 = So, on the contrary to the general formula (9), the use of a Banach limit is unnecessary in the context of an operator T in the class Cα, and one can use instead the "regularized" limit lim∗. We observe that lim∗ coincides with the usual limit when the sequence is convergent (here we use the above normalization assumption thatP fk = 1). On the other hand, the example of matrices T , meeting the requirements of Lemma 3.2, shows that in general, the limit lim kT nhk2 does not exist for T ∈ Cα. In Section 9, we will show that this limit does exist if α satisfies an extra requirement. In what follows, for T ∈ Cα, the operator D, given by (10), will be called the defect operator of T . 5. On definition of unitary part of a Cα operator Definition 5.1. Let K be a Hilbert space. We say that T ∈ L(K) is a completely nonunitary operator if there is no nonzero reducing subspace L for T such that T L is unitary. It is well-known that every contraction S can be decomposed into an orthogonal sum of a unitary operator and a completely nonunitary operator (called the unitary part and the completely nonunitary part of S, respectively). We recall that the standard construction of the Nagy-Foias model applies only to completely nonunitary contractions. If one takes an operator T in the class Cα and applies to it Theorem 4.1, then one gets a direct sum decomposition (17) H = H0 ∔ H1 such that T H0 is similar to a unitary operator and T H1 is similar to a completely non-unitary operator. For a general admissible function α, we cannot say much more. However, some extra properties hold if α is in following subclass. Definition 5.2. A function α ∈ AW will be called strongly admissible if it has the form α(t) = (1 − t)eα(t) for some function eα ∈ AW with real Taylor coefficients, which has no roots on the unit circle T and satisfies eα(0) = α(0) > 0. Notice that any strongly admissible function is admissible. For the sequel, let us recall the following characterization of the unitary part of a contraction. Theorem B (See [18], Theorem I.3.2). To every contraction S on the space H there cor- responds a decomposition of H into an orthogonal sum of two subspaces reducing S, say H = H0 ⊕ H1, such that the part of S on H0 is unitary, and the part of S on H1 is completely nonunitary; H0 or H1 may equal the trivial subspace {0}. This decomposition is uniquely determined. Indeed, H0 consists of those elements h of H for which (n = 1, 2, . . .). kSnhk = khk = kS∗nhk S0 = SH0 and S1 = SH1 are called the unitary part and the completely nonunitary part of S, respectively, and S = S0 ⊕ S1 is called the canonical decomposition of S. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 11 The next theorem is an analogue of this decomposition for our operators in the case of a strongly admissible function α. Theorem 5.3. Let T ∈ Cα, where α is strongly admissible. Denote by H0 the elements h ∈ H for which there exists a two-sided sequence {hn}n∈Z such that h0 = h, T hn = hn+1 and khnk = khk for every n ∈ Z. Let eT be the operator T acting on eH := (H, · ) (the new norm, which was given in (11)). Then H0 is a closed subspace of H and there exists a direct sum decomposition H = H0 ∔ H1 with the following properties: (i) T H0 is unitary; (ii) H0 and H1 are invariant subspaces of T ; (iii) H0 and H1 are orthogonal in eH; (iv) eT H0 and eT H1 are the unitary part and the completely nonunitary part of the con- Remark 5.4. In Lemma 3.2 we saw that if T is a finite matrix without Jordan blocks and σ(T ) ⊂ T, then T ∈ Cp for some admissible polynomial p. In this case, H0 = H and H1 = H in the decomposition (17), but T H0 = T is non-unitary, as a rule. As a consequence we get: traction eT , respectively. (i) Theorem 5.3 is not valid if we do require α to be strongly admissible; (ii) A non-unitary finite matrix T with σ(T ) ⊂ T cannot belong to Cα if α is strongly admissible. Observe that a completely nonunitary contraction can be similar to a unitary operator. For a general operator T , one cannot single out the largest direct summand which is similar to a unitary. Hilbert space.) (n = 1, 2, . . .), 2 eT nh = h = eT ∗nh Proof of Theorem 5.3. Since eT is a contraction on eH, Theorem B gives us that for the decom- position eH = eH0 ⊕ eH1, where eH0 consists of those elements h of eH for which we have that eT0 := eT eH0 is unitary and eT1 := eT eH1 is completely nonunitary. The goal of the following two claims is to prove that H0 = eH0. Claim 1. Let h ∈ H. Then h ∈ eH0 if and only in there exists a sequence {hn}n∈Z such that h0 = h,eT hn = hn+1 and hn = h for every n ∈ Z. (In fact, this claim is a variation of Theorem B and is valid for any contraction eT on a Indeed, suppose that h ∈ eH0. Define the sequence {hn}n∈Z by h0 := h, hn := eT nh and h−n := eT ∗nh, for n ≥ 1. Since eT eH0 is unitary we obtain that eT h−n = eTeT ∗nh = eT ∗(n−1) = Fix n ≥ 1. Since eT nh−n = h, we have that eT nh−n eT ∗neT n)h−n, h−ni = 0. But using that eT n is a contraction on eH it follows that (I−eT ∗neT n)h−n = 0. Therefore, using that eT nh−n = h, we obtain that eT ∗nh = h−n. Then h ∈ eH0. This finishes Claim 2. Let h ∈ H. Then h ∈ eH0 if and only in there exists a sequence {hn}n∈Z such h−n+1 for n ≥ 1. It follows that the sequence {hn}n∈Z satisfies the conditions of the statement. Reciprocally, suppose that a sequence {hn}n∈Z satisfies the conditions of the statement. = h2 = h−n2. Hence h(I − (Note that this claim is stated in terms of the original norm in H.) Indeed, we apply Claim 1. Let {hn}n∈Z be a sequence such that h0 = h and T hn = hn+1. We have to show that the sequence {hn2}n∈Z is constant if and only if the sequence {khnk2}n∈Z is constant. Since the two norms are equivalent, if either of these two sequences is constant, the other is in ℓ∞(Z). that h0 = h, T hn = hn+1 and khnk = khk for every n ∈ Z. the proof of Claim 1. By (12) we have that hn+12 = T hn2 = hn2 − kDhnk2. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 12 Therefore hn = h for every n ∈ Z if and only if kDhnk2 = 0 for every n ∈ Z. We will use the backward shift operator ∇, acting on ℓ∞(Z) by [∇a]n = an+1, a ∈ ℓ∞(Z). For any f ∈ AW , f (∇) is well-defined by(cid:0)f (∇)a(cid:1)n :=P∞ j=0 fjan+j. Denote by h the sequence {khnk2}n∈Z. Then it is easy to obtain that kDhnk2 = [α(∇)h]n. Hence hn = h for every n ∈ Z if and only if α(∇)h = (. . . , 0, 0, 0, . . .) = 0. So we have to show that h is constant if and only if α(∇)h = 0. The direct implication is obvious. For the converse, fix a factorization α(t) = (1 − t)q(t)γ(t), where q is a polynomial with zeroes in D and γ ∈ AW (D) without zeroes in D (recall that α is assumed to be strongly admissible). Then 1/γ ∈ AW (D). Hence γ(∇)[q(∇)(1− ∇)h] = 0 implies that q(∇)(1− ∇)h = 0 (just multiply by 1/γ). Now let g = (1 − ∇)h. We want to proof that g = 0. When q has just a single root, say q(t) = 1 − at for some a with a > 1, the result is immediate. For a general q we just need to apply induction on the number of roots of q. This finishes the proof of Claim 2. surjective isometry. The rest of items of the theorem follow immediately using Theorem B. (cid:3) Therefore we have that H0 = eH0. Put H1 := eH1. Now (i) is obvious, since T H0 is a Remark 5.5. In Lemma 3.2 we saw that if T is a finite matrix without Jordan blocks and σ(T ) ⊂ T, then T ∈ Cp for some admissible polynomial p. In this case, H0 = H and H1 = H in the decomposition (17), but T H0 = T is non-unitary, as a rule. As a consequence we get: (i) Theorem 5.3 is not valid if we do require α to be strongly admissible; (ii) A non-unitary finite matrix T with σ(T ) ⊂ T cannot belong to Cα if α is strongly admissible. 6. The Nagy-Foias¸ model of T Let T ∈ Cα for some α ∈ AW . In this section, for simplicity, we will assume that α is strongly admissible and T is a completely nonunitary operator. Then, using the notation of the previous section, we have that eT is a completely nonunitary contraction on eH. Let D eT and D eT ∗ be the defect operators of T and let D eT and D eT ∗ be its defect spaces. We recall that the defect operator of T has been defined by (10). We define the defect space of T as DT = clos DH, where the closure is taken with respect to k · k. Define V : D eT → DT by V (D eT h) = Dh, h ∈ H. Then, using equation (12), we obtain that V is an isometry from D T to DT . Let us define the functions Θ∗ ∈ H ∞(D eT ∗ → DT ) and ∆∗ : T → D eT ∗ given by Now put ∆∗(ζ) := (I − Θ∗(ζ)∗Θ∗(ζ))1/2, Θ∗(z)h := V (−eT ∗ + zD eT (I − zT )−1D eT ∗)h, clos ∆∗L2(D eT ∗)(cid:19) ⊖(cid:18)Θ∗ z−1v(z)(cid:19). M∗(cid:18)u v(cid:19) :=(cid:18) u(z)−u(0) KΘ∗ :=(cid:18) H 2(DT ) z h ∈ D eT ∗ , z ∈ D, ζ ∈ T. ∆∗(cid:19)H 2(D eT ∗ ). Finally, let us define Φ1 : H → H 2(DT ) by Φ1h(z) = D(I − zT )−1h and M∗ : KΘ∗ → KΘ∗ by Theorem 6.1. Let T ∈ Cα, where α is strongly admissible. With the notation used above, there exists a linear map Φ2 : H → clos ∆∗L2(D eT ∗) such that kΦ2h(z)k2 = lim∗ n→∞kT nhk2 for every h ∈ H, and (i) Φ :=(cid:0)Φ1 Φ2(cid:1) : H → KΘ∗ is an isometric isomorphism; (ii) ΦT = M∗Φ. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 13 It is a common point that this kind of result implies a variant of von Neumann inequality: if T satisfies the conditions of the above theorem, then for any f ∈ H(D), kf (T )k ≤ C max f , D where C is a constant depending only on T (in fact, C = k(kΦ)k(kΦ−1)). Here we should mention the works by Olofsson (see [20]), where he relates certain trans- fer functions associated with n-hypercontractions with Bergman inner functions, which are crucial in the description of invariant subspaces of Bergman spaces. His results were further generalized in the work by Ball and Bolotnikov [5], [6]. 7. Operators in Cα whose characteristic function has a determinant In what follows, Sp (0 < p ≤ ∞) will denote the Schatten-von Neumann class of operators. Lemma 7.1. Let T ∈ Cα for some admissible function α and let p ∈ [1, ∞]. (i) If I − T ∗T ∈ Sp, then D2 ∈ Sp. (ii) If D2 ∈ Sp and α has no zeros in D, then I − T ∗T ∈ Sp. Proof. (i) It is immediate, since D2 = α[T ∗, T ] = α[T ∗, T ](I − T ∗T ). (ii) Since α has no zeros in D, 1/α ∈ AW , and we obtain that (1/α)[T ∗, T ](D2) = (1/α)[T ∗, T ](α[T ∗, T ](I − T ∗T )) = I − T ∗T, which proves the result. (cid:3) Lemma 7.2. Let T ∈ Cα for some admissible function α and let p ∈ [1, ∞]. If σ(T ) 6= D, then D T ∈ Sp if and only if D T ∗ ∈ Sp. Proof. In the case when 0 /∈ σ(T ), D T is unitarily equivalent to D T ∗, see [18, the proof of Theorem VIII.1.1] or [13, Lemma 9]; this implies our assertion. The general case follows from this one. Indeed, take any λ ∈ D \ σ(T ) and consider the Mobius self-map of D, given by bλ(z) = (z − λ)/(1 − ¯λz). Let Tλ be the contraction, defined by Tλ = bλ( T ). Then the formula I − T ∗ λ Tλ = W ∗(I − T ∗ T )W with W = (1 − λ2)1/2(I − ¯λ T )−1 implies that D T ∈ Sp if and only if D Tλ ∈ Sp. Since 0 /∈ σ( Tλ), the general case follows. (cid:3) We remark that the previous lemma applies to a more general situation when T is a power bounded operator with σ(T ) 6= D and D is its abstract defect operator. Then, by Lemma 1.4, one gets a contraction eT , similar to T , and so for any p, D ∈ Sp iff D T ∈ Sp iff D T ∗ ∈ Sp. We recall the well-known fact that the characteristic function of a contraction S has the determinant whenever σ(S) 6= D and I − S∗S ∈ S1 (this is the so-called class of weak con- tractions). It follows that Θ∗ has a determinant whenever σ(T ) 6= D and D ∈ S2. Notice that det Θ∗ is an H ∞ function such that k det Θ∗k∞ ≤ 1. We obtain the following statement. Proposition 7.3. Suppose α is strongly admissible and T ∈ Cα is a completely nonunitary operator. Suppose also that σ(T ) 6= D and α[T ∗, T ] ∈ S1. Denote by σp(T ) the point spectrum of T . Then the following assertions are equivalent. (i) T is complete, that is, H = span{ker(λI − T )k : (ii) T ∗ is complete; (iii) det Θ∗(z) is a Blaschke product. k ≥ 1, λ ∈ σp(T )}; This follows from the above observations and from an analogous fact for completely nonuni- tary contractions, see [19, p. 134]. In a similar way, one can extend the results by Treil [23], Nikolski - Benamara [7], Kupin [14], [15] and others to the setting of operators in Cα, where α is strongly admissible. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 14 8. A result on the inclusion of classes Cα ⊂ Cτ The goal of this section is to prove the following result on the containment of classes. neighbourhood of the origin. Let α and τ be admissible functions and let γ := eτ /eα. Note that γ is analytic on a Theorem 8.1. Let α, τ and γ be as above. (i) If Cα ⊂ Cτ , then γ ≻ 0. (ii) If γ ∈ AW , then Cα ⊂ Cτ if and only if γ ≻ 0. (iii) If γ is not bounded in D (in particular, if γ has poles in D), then Cα 6⊂ Cτ . The following lemma is in the spirit of Pringsheim's Theorem (see [11]). Lemma 8.2. Let g be a meromorphic function in D, analytic in the origin, such that g < 0. If g is bounded on [0, 1), then g is a bounded analytic function on D. Proof. Since g < 0, we have (18) g(z) ≤ g(z) whenever the series for g(z) converges. Let r ∈ (0, 1] be the radius of convergence of g. If r < 1, then g has poles on the circle {z = r}, and this contradicts the boundedness of g on [0, r). Hence r = 1, and therefore by(18), g is bounded on D. (cid:3) n=0 be an orthonormal basis in H and let {λn}∞ Lemma 8.3. Let {en}∞ n=1 be a sequence of positive numbers which are eventually 1. Define the sequence Λ = {Λn}∞ n=0 by Λn := (λ1 · · · λn+1)2. Let T be the weighted shift operator, given by T en = λn+1en+1, and let α ∈ AW . Then T ∈ Cα if and only if α(∇)Λ ≻ 0 (here ∇ is the backward shift on one-sided sequences {An}n≥0). Proof. Note that Λ is a bounded sequence and that T is a power bounded operator. Hence α[T ∗, T ] and α(∇)Λ are well defined, so we need to prove that α[T ∗, T ] ≥ 0 if and only if α(∇)Λ ≻ 0. It is immediate that (19) α[T ∗, T ] ≥ 0 ⇐⇒ Next we observe that αnkT nhk2 ≥ 0 (∀h ∈ H). ∞Xn=0 (20) αnkT nhk2 ≥ 0 (∀h ∈ H) ⇐⇒ αnkT nejk2 ≥ 0 (∀j ≥ 0). Indeed, the direct implication is obvious. The converse is seen from the following formula, n=0 hnen ∈ H, where {hn}∞ n=0 ∈ ℓ2: valid for any vector h =P∞ ∞Xn=0 αnkT nhk2 = ∞Xn=0 αn(cid:18) ∞Xj=0 hj2kT nejk2(cid:19) = ∞Xj=0 hj2(cid:18) ∞Xn=0 αnkT nejk2(cid:19) ∞Xn=0 ∞Xn=0 (we can change the order of summation because the series converge absolutely). Fix j ≥ 0. For every n ≥ 0 we have T nej = λj+1λj+2 · · · λj+nej+n =s Λn+j Λj ej+n, so kT nejk2 = Λn+j/Λj and therefore αnkT nejk2 = ∞Xn=0 αn Λn+j Λj = 1 Λj [α(∇)Λ]j . ∞Xn=0 OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 15 Hence it follows that (21) ∞Xn=0 αnkT nejk2 ≥ 0 (∀j ≥ 0) ⇐⇒ α(∇)Λ ≻ 0. The statement now follows from (19), (20) and (21). (cid:3) Before starting the proof of Theorem 8.1, let us make a final observation. Remark 8.4. If T ∈ L(H) is a power bounded operator and α ∈ AW then α[T ∗, T ] ≥ 0 ⇐⇒ ⇐⇒ αnkT nhk2 ≥ 0 (∀h ∈ H) αnkT n+jhk2 ≥ 0 (∀j ≥ 0, ∀h ∈ H), ∞Xn=0 ∞Xn=0 where in the last equivalence we just change h by T jh. Therefore, if we fix h ∈ H and define the sequence Λ = {Λn}∞ n=0 by Λn := kT nhk2 then α[T ∗, T ] ≥ 0 implies that α(∇)Λ ≻ 0. Proof of Theorem 8.1. (i) Suppose that γ 6≻ 0 and let ℓ be the smallest index such that γℓ < 0. Note that ℓ ≥ 1 because γ0 > 0. By Lemma 8.3, we just need to find a sequence of positive numbers Λ = {Λn}∞ n=0 that is eventually constant such that α(∇)Λ ≻ 0 and τ (Λ) 6≻ 0, because in that case if we fix an orthonormal basis {en}∞ n=0 of H then the weighted shift operator T defined by T en =pΛn+1/Λn en+1 satisfies T ∈ Cα \ Cτ . Let us construct that sequence. Consider the sequence Γ := (γℓ, γℓ−1, . . . , γ0, 0, 0, . . .) and define the sequence Ψ by Ψ := (τ )−1(∇)Γ. Note that Ψ is well defined because Γ has only finitely many nonzero terms and τ is invertible in a neighbourhood of the origin. Since Γn = 0 for n ≥ ℓ + 1 we obtain that also Ψn = 0 for n ≥ ℓ + 1. Finally, let Λ be a the sequence that satisfies Ψ = (1 − ∇)Λ, with Λ0 large enough so that Λn > 0 for every n. Note that Λn is constant for n ≥ ℓ + 1 and it also satisfies α(∇)Λ = α(∇)Ψ =(cid:18) α τ(cid:19) (∇)Γ =(cid:18) 1 γ(cid:19) (∇)Γ = (0, . . . , 0, ℓ ⌣ 1 , 0, . . .) ≻ 0 and since Γ0 = γℓ < 0. Hence (i) is proved. (ii) It is clear that Cα ⊂ Cτ if γ ≻ 0. Indeed, if T ∈ Cα, then using Lemma 3.3 (ii) we have τ (∇)Λ = τ (∇)Ψ = Γ 6≻ 0, τ [T ∗, T ] = (γα)[T ∗, T ] = γnT ∗nα[T ∗, T ]T n ≥ 0, ∞Xn=0 hence T ∈ Cτ . The other implication follows from (i). (iii) Note that Lemma 8.2 implies that if γ is not bounded on D then γ 6≻ 0 (recall that γ has no poles in [0, 1]), so this statement also follows from (i) and the theorem is proved. (cid:3) Remark 8.5. Note that in general, for a rational admissible function r, it is not possible to find an admissible polynomial p such that Cr ⊂ Cp. For example, consider the rational admissible function r(t) = . 1 − t 1 − t/2 If such a p exists, say p(t) = (1−t)p(t), then by Theorem 8.1 (i) we should have p(t)(1−t/2) ≻ 0. But it is immediate to check that this is impossible for any real polynomial p. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 16 Remark 8.6. Let α be an admissible function. Since eα ∈ AW , by Theorem 8.1 (i) we have that C1−t ⊂ Cα if and only if eα ≻ 0. If moreover eα has zeroes in D, then by Theorem 8.1 (iii) we deduce that C1−t $ Cα. For example C1−t $ C(1−t)(t2+1/4). Remark 8.7. Consider the strongly admissible rational function α(t) = 1 − t 9 (t − 3 4 2 )2 . Then Cα 6⊂ C1−tn for every n ≥ 1. Indeed, suppose on the contrary that Cα ⊂ C1−tn for some n ≥ 1. Then, by Theorem 8.1 (i), we should have 4 9(cid:18)t − 3 2(cid:19)2 (1 + t + · · · + tn−1) ≻ 0; however, the coefficient of t in the above expression is −1/3. This shows that there are classes Cα that are not contained in the union of classes C1−tn over all n ≥ 1. Lemma 8.8. Let α and β be admissible functions, and consider the following conditions. (a) α, β ∈ H(D); (b) α(t)/(1 − t) or β(t)/(1 − t) has no zeros on D. If (a) or (b) holds, then there exists an admissible function γ such that Cα ∪ Cβ ⊂ Cγ. Proof. (a) Suppose that α, β ∈ H(D). Then α β = a b ϕ for some polynomials a and b without common roots and a function ϕ ∈ H(D) which is positive on [0, 1]. By Lemma 1.2, there exists a function ψ ∈ AW such that ψ ≻ 0 and ψϕ ≻ 0. Put a = a+a−anr where a+ contains the positive roots of a, a− contains the negative roots of a and anr contains the non-real roots of a. If for example a does not have any positive root, then we just put a− = 1. In the same way, put b = b+b−bnr. Applying Corollary 2.2 twice, notice that there exists a polynomial p without roots in D such that p ≻ 0, panr ≻ 0 and pbnr ≻ 0. Let v := , w := a−anrp b+ b−bnrp . a+ Note that v ≻ 0, w ≻ 0 and a/b = v/w. Now we simply put γ := awψ = bvψϕ. Since wψ ≻ 0 and vψϕ ≻ 0, the result follows from Theorem 8.1 (i). (b) Suppose that β(t)/(1 − t) has no zeros on D. Then α/β =: ϕ ∈ AW is positive on [0, 1]. By Lemma 1.2, there exists a function ψ ∈ AW such that ψ ≻ 0 and ψϕ ≻ 0. Now we put (cid:3) γ := αψ = βϕψ. Since ψ ≻ 0 and ϕψ ≻ 0, the result follows from Theorem 8.1 (ii). Corollary 8.9. Let T1 be a complex square matrix and T2 be a Hilbert space operator. If T1 has no Jordan blocks and σ(T1) ⊂ T, whereas the spectral radius of T2 is less than 1, then there exists an admissible function γ such that T1 ⊕ T2 ∈ Cγ. Proof. This follows immediately from Lemma 3.2, Proposition 3.1 (b) and (e), and Lemma 8.8. (cid:3) 9. Existence of the limit of kT nhk2 As it was explained at the end of Section 4, in general, the limit of norms kT nhk as n → ∞ (where h ∈ H) does not exist. In this section we prove the following result: Theorem 9.1. Let α(t) = (1 − t)α(t), where α is strongly admissible, and let T ∈ Cα. Then, for every h ∈ H, there exists the limit limn→∞ kT nhk2. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 17 {an}∞ every n ≥ 1. We will use the backward shift ∇ and the shift ∇−, acting on one-sided bounded sequences n=0. They are given by [∇a]n = an+1 for every n ≥ 0 and [∇−a]0 = 0, [∇−a]n = an−1 for If we identify the sequence a = {an}∞ can identify the operator ∇ and ∇− with the operators given by n=0 with the power series a(z) =P∞ n=0 anzn, then we (∇a)(z) = , (∇−a)(z) = za(z). a(z) − a(0) z It is clear that ∇∇− = I and ∇−∇a(z) = a(z) − a(0). f (∇−)a is given by cn :=Pn Given a function f ∈ AW , the operators f (∇) and f (∇−) are well-defined. Note that c = j=0 fjan−j. In terms of power series, one just has c(z) = f (z)a(z). We can say, in fact, that in the power series representation, f (∇−) is an analytic Toeplitz operator and f (∇) is an anti-analytic Toeplitz operator. The following formula will be useful: (22) ∇k −∇ja(z) = zk−j(a(z) − aj−1zj−1 − · · · − a1z − a0). We need some auxiliary lemmas. Lemma 9.2. Let f, g ∈ AW and let a ∈ ℓ∞. Then f (∇)[g(∇)a] = (f g)(∇)a. The proof is immediate just doing a change of summation indices. Lemma 9.3. Let f ∈ AW and let a ∈ ℓ∞ be a convergent sequence, say an → a∞. (i) If b = f (∇)a, then bn → f (1)a∞. (ii) The same is true for ∇− in place of ∇. Namely, if c = f (∇−)a, then also cn → f (1)a∞. Proof. Both statements are straightforward, and we will only check (i). Fix ε > 0 and let j=0 fj for every n ≥ N . Then bn − f (1)a∞ ≤ ∞Xj=0 fjan+j − a∞ < ε (cid:3) an − a∞ < ε/P∞ for every n ≥ N . We can rephrase part (ii) of last lemma in terms of formal power series as follows. Corollary 9.4. Let f ∈ AW and let a(z) = P∞ n=0 anzn be a formal power series where the sequence {an}∞ n=0 converges to some number a∞ ∈ R. If b(z) = f (z)a(z), then bn → f (1)a∞. Lemma 9.5. Let q be a real polynomial whose roots are in D. Let a ∈ ℓ∞ and put b = q(∇)a. If bn → b∞ ∈ R, then an → b∞/q(1). Proof. Put q(t) = qsts + · · · + q1t + q0. Then ∇s −b = (q0∇s − + q1∇s −∇ + · · · + qs∇s −∇s)a, which can be written in formal power series using (22) as zsb(z) = q0zsa(z) + q1zs−1(a(z) − a0) + · · · + qs(a(z) − as−1zs−1 − · · · − a1z − a0). So if we put q(t) = q0ts + q1ts−1 + · · · + qs, then zsb(z) = q(z)a(z) − r(z) for some polynomial r of degree at most s − 1. Note that q has no roots in D, hence 1/q ∈ AW and therefore a(z) = b(z) + zs q(z) r(z) q(z) . Since r/q ∈ AW , its n-th Taylor coefficient tends to 0. Now the statement follows using the (cid:3) previous corollary and that q(1) = q(1). OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 18 Lemma 9.6. Let q be a real polynomial whose roots are in D and put Q(t) = (1 − t)q(t). Let a ∈ ℓ∞. If b = Q(∇)a and bn ≥ 0, then there exists lim an. Proof. Put c := q(∇)a. Then b = q(∇)a − ∇q(∇)a, so bn = cn − cn+1 ≥ 0. Hence {cn}∞ n=0 is a decreasing sequence. Since kck∞ ≤ (q0 + · · · + qs)kak∞, the sequence c is bounded. Therefore {cn}∞ (cid:3) n=0 converges and by the previous lemma we obtain that an → c∞/q(1). Proof of Theorem 9.1. Let h ∈ H. Since T ∈ Cα, we obtain that P∞ Changing h by T jh for j ≥ 1 we get that P∞ sequence a by an = kT nhk2 then we have that b := α(∇)a satisfies bn ≥ 0 for every n ≥ 0. n=0 αnkT nhk2 ≥ 0. n=0 αnkT n+jhk2 ≥ 0. Hence, if we define the By Lemma 9.2 we have b = α(∇)a = (1 − ∇)α(∇)a. Since α does not vanish on T, we can split it as α = q β, where q is a polynomial with roots in D and β ∈ AW does not vanish on D. Therefore, if we put Q(t) = (1 − t)q(t) and c := β(∇)a, then b = Q(∇)c. By Lemma 9.6 we know that there exists lim cn, and since a = (1/ β)(∇)c and (1/ β) ∈ AW , the statement follows by Corollary (cid:3) 9.4. Corollary 9.7 (of Theorem 9.1). Assume the hypotheses of Theorem 4.1, in particular, that T ∈ Cα and that f , B are defined as in this theorem. If α is strongly admissible, then the norm defined in (11) can be alternatively expressed by h2 = ∞Xn=0 kDT nhk2 + lim n→∞ kT nhk2. This follows from Theorem 9.1 and formula (16). References [1] J. Agler. The Arveson extension theorem and coanalytic models. Integral Equations and Operator Theory, 5(1):608 -- 631, 1982. [2] J. Agler. Hypercontractions and subnormality. J. Operator Theory, 13(2):203 -- 217, 1985. [3] C. Ambrozie, M. Englis, and V. Muller. Operator tuples and analytic models over general domains in Cn. Journal of Operator Theory, 47(2):289 -- 304, 2002. [4] J. Arazy and M. Englis. Analytic models for commuting operator tuples on bounded symmetric domains. Transactions of the American Mathematical Society, 355(2):837 -- 864, 2003. [5] J. A. Ball and V. Bolotnikov. Weighted Bergman spaces: shift-invariant subspaces and input/state/output linear systems. Integral Equations Operator Theory, 76(3):301 -- 356, 2013. [6] J. A. Ball and V. Bolotnikov. Weighted Hardy spaces: shift invariant and coinvariant subspaces, linear systems and operator model theory. Acta Sci. Math. (Szeged), 79(3-4):623 -- 686, 2013. [7] N.-E. Benamara and N. Nikolski. Resolvent tests for similarity to a normal operator. Proc. London Math. Soc. (3), 78(3):585 -- 626, 1999. [8] T. Bhattacharyya and J. Sarkar. Characteristic function for polynomially contractive commuting tuples. J.Math. Anal. Appl., 321:242 -- 259, 2006. [9] J. B. Conway. A course in functional analysis, volume 96. Springer-Verlag, 2 edition, 1990. [10] M. F. Gamal. On power bounded operators that are quasiaffine transforms of singular unitaries. Acta Universitatis Szegediensis. Acta Scientiarum Mathematicarum, 77(3-4):589 -- 606, 2011. [11] E. Hille. Analytic function theory, Volume I, volume 269. American Mathematical Soc., 2012. [12] L. K´erchy. Isometric asymptotes of power bounded operators. Indiana University mathematics journal, pages 173 -- 188, 1989. [13] L. K´erchy. On the functional calculus of contractions with nonvanishing unitary asymptotes. Michigan Math. J., 37:323 -- 338, 1990. [14] S. Kupin. Linear resolvent growth test for similarity of a weak contraction to a normal operator. Ark. Mat., 39:95 -- 119, 2001. [15] S. Kupin. Operators similar to contractions and their similarity to a normal operator. Indiana University Mathematics Journal, 52(3):753 -- 768, 2003. [16] V. Muller. Models of operators using weighted shifts. J. Operator Theory, 20:3 -- 20, 1988. [17] B. d. S. Nagy. On uniformly bounded linear transformations in hilbert space. Acta Sci. Math.(Szeged), 11(19):7, 1947. OPERATOR INEQUALITIES IMPLYING SIMILARITY TO A CONTRACTION 19 [18] B. S. Nagy, C. Foias, H. Bercovici, and L. K´erchy. Harmonic analysis of operators on Hilbert space. Springer Science & Business Media, 2010. [19] N. K. Nikolski. Operators, functions, and systems: an easy reading. Vol. 2, volume 93 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2002. Model operators and systems, Translated from the French by Andreas Hartmann and revised by the author. [20] A. Olofsson. Operator-valued Bergman inner functions as transfer functions. St. Petersburg Mathematical Journal, 19(4):603 -- 623, 2008. [21] A. Olofsson. Parts of adjoint weighted shifts. Journal of Operator Theory, 74(2):249 -- 280, 2015. [22] S. Pott. Standard models under polynomial positivity conditions. J. Operator Theory, 41(2):365 -- 389, 1999. [23] S. Treil. Unconditional bases of invariant subspaces of a contraction with finite defects. Indiana University Mathematics Journal, 46(4):1021 -- 1054, 1997. [24] D. V. Yakubovich. Linearly similar model of Sz.-Nagy-Foias type in a domain. Algebra i Analiz, 15(2):180 -- 227, 2003. Glenier Bello-Burguet Instituto de Ciencias Matem´aticas (CSIC-UAM-UC3M-UCM), C/ Nicol´as Cabrera, no 13-15 Campus de Cantoblanco, UAM, 28049 Madrid, Spain E-mail address: [email protected] D. V. Yakubovich Departamento de Matem´aticas, Universidad Aut´onoma de Madrid, Cantoblanco, 28049 Madrid, Spain and Instituto de Ciencias Matem´aticas (CSIC-UAM-UC3M-UCM) E-mail address: [email protected]
1808.03335	1	1808	2018-08-08T16:01:07	Asymptotically almost periodic solutions of fractional relaxation inclusions with Caputo derivatives	[ "math.FA" ]	In the paper under review, we analyze asymptotically almost periodic solutions for a class of (semilinear) fractional relaxation inclusions with Stepanov almost periodic coefficients. As auxiliary tools, we use subordination principles, fixed point theorems and the well known results on the generation of infinitely differentiable degenerate semigroups with removable singularites at zero. Our results are well illustrated and seem to be not considered elsewhere even for fractional relaxation equations with almost sectorial operators.	math.FA	math	ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES MARKO KOSTI ´C Abstract. In the paper under review, we analyze asymptotically almost pe- riodic solutions for a class of (semilinear) fractional relaxation inclusions with Stepanov almost periodic coefficients. As auxiliary tools, we use subordination principles, fixed point theorems and the well known results on the generation of infinitely differentiable degenerate semigroups with removable singularites at zero. Our results are well illustrated and seem to be not considered elsewhere even for fractional relaxation equations with almost sectorial operators. 8 1 0 2 g u A 8 ] . A F h t a m [ 1 v 5 3 3 3 0 . 8 0 8 1 : v i X r a 1. Introduction and preliminaries The notion of an almost periodic function was introduced by Bohr in 1925 and [5], [9], [11] and [22]). later generalized by many other mathematicians (see e.g. Let I = R or I = [0, ∞), and let f : I → X be continuous. Given ǫ > 0, we call τ > 0 an ǫ-period for f (·) iff kf (t + τ ) − f (t)k ≤ ǫ, t ∈ I. The set consisted of all ǫ-periods for f (·) is denoted by ϑ(f, ǫ). It is said that f (·) is almost periodic, a.p. for short, iff for each ǫ > 0 the set ϑ(f, ǫ) is relatively dense in I, which means that there exists l > 0 such that any subinterval of I of length l meets ϑ(f, ǫ). The space consisted of all almost periodic functions from the interval I into X will be denoted by AP (I : X). The notion of an asymptotically almost periodic function was introduced by Fr´echet in 1941 (for further information concerning the vector-valued asymptot- ically almost periodic functions, see [4]-[5], [9] and references cited therein). A function f ∈ Cb([0, ∞) : X) is called asymptotically almost periodic iff for every ǫ > 0 we can find numbers l > 0 and M > 0 such that every subinterval of [0, ∞) of length l contains, at least, one number τ such that kf (t + τ ) − f (t)k ≤ ǫ for all t ≥ M. The space consisting of all asymptotically almost periodic functions from [0, ∞) into X is denoted by AAP ([0, ∞) : X). For a function f ∈ C([0, ∞) : X), the following statements are equivalent ([28]): (i) f ∈ AAP ([0, ∞) : X). (ii) There exist uniquely determined functions g ∈ AP ([0, ∞) : X) and φ ∈ C0([0, ∞) : X) such that f = g + φ. (iii) The set H(f ) := {f (· + s) : s ≥ 0} is relatively compact in Cb([0, ∞) : X). 2010 Mathematics Subject Classification. 34G25, 47D03, 47D06, 47D99. Key words and phrases. Abstract semilinear Cauchy inclusions, Asymptotic almost periodicity, Stepanov asymptotic almost periodicity. The author is partially supported by grant 174024 of Ministry of Science and Technological Development, Republic of Serbia. 1 2 MARKO KOSTI ´C Assume that 1 ≤ p < ∞. Then it is said that a function f ∈ Lp loc(I : X) is Stepanov p-bounded, Sp-bounded shortly, iff kf kSp := sup t∈I Z t+1 t kf (s)kp ds!1/p < ∞. The space Lp S(I : X) consisting of all Sp-bounded functions becomes a Banach space when equipped with the above norm. A function f ∈ Lp S(I : X) is said to be Stepanov p-almost periodic, Sp-almost periodic shortly, iff the function f : I → Lp([0, 1] : X), defined by f (t)(s) := f (t + s), t ∈ I, s ∈ [0, 1] is almost periodic (cf. Amerio, Prouse [1] for more details). We say that f ∈ Lp S([0, ∞) : X) is asymptotically Stepanov p-almost periodic, asymptotically Sp- almost periodic shortly, iff f : [0, ∞) → Lp([0, 1] : X) is asymptotically almost periodic. We use the shorthands AP Sp([0, ∞) : X) and AAP Sp([0, ∞) : X) to denote the vector spaces consisting of all Stepanov p-almost periodic functions and asymptotically Stepanov p-almost periodic functions, respectively. It is well-known that if f (·) is an almost periodic (respectively, a.a.p.) function then f (·) is also Sp-almost periodic (resp., Sp-a.a.p.) for 1 ≤ p < ∞. The converse statement is not true, in general. Concerning almost periodic and asymptotically almost periodic solutions of var- ious classes of abstract Volterra integro-differential equations in Banach spaces, the reader may consult [1], [4]-[5], [9]-[11], [18]-[25] and [32]. Let γ ∈ (0, 1), and let A be a multivalued linear operator on a Banach space X. Of importance is the following fractional relaxation inclusion and its semilinear analogue (DFP)f,γ :(cid:26) Dγ (DFP)f,γ,s :(cid:26) Dγ t u(t) ∈ Au(t) + f (t), t > 0, u(0) = x0, t u(t) ∈ Au(t) + f (t, u(t)), t > 0, u(0) = x0, where Dγ t denotes the Caputo fractional derivative of order γ, x0 ∈ X and f : [0, ∞) → X, resp. f : [0, ∞) × X → X, is Stepanov almost periodic. The main aim of this paper is to continue our recent research studies [18]-[21] by investigating asymptotically almost periodic solutions of fractional Cauchy inclusions (DFP)f,γ and (DFP)f,γ,s. We would like to note that the existence and uniqueness of (asymp- totically) quasi-periodic solutions of fractional relaxation equations with Caputo derivatives have not received much attention so far: in the existing literature, we have been able to find only one research paper (Li, Liang, Wang [23]) concern- ing similar problematic; that paper is devoted to the study of S-asymptotically ω-periodic solutions to fractional relaxation equations with finite delay. The es- tablished results of ours are completely new in degenerate case and, as already mentioned in the abstract, they seem to be new even for fractional relaxation equa- tions with almost sectorial operators ([26]). The paper is very simply organized and we deeply believe that our readers will fairly quickly move throughout it. We use the standard notation henceforth. By X and Y we denote two Banach spaces over the field of complex numbers. The symbol L(X, Y ) stands for the space consisting of all continuous linear mappings from X ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES3 into Y ; L(X) ≡ L(X, X). By I we denote the identity operator on X. Denote by Cb([0, ∞) : X), C0([0, ∞) : X) and BU C([0, ∞) : X) the vector spaces consisted of all bounded continuous functions from [0, ∞) into X, all bounded continuous func- tions from [0, ∞) into X vanishing at infinity, and all bounded uniformly continuous functions from [0, ∞) into X, respectively. Equipped with the usual sup-norm, any of these spaces becomes one of Banach's. Denote by C0([0, ∞) × Y : X) the space of all continuous functions h : [0, ∞) × Y → X satisfying that limt→∞ h(t, y) = 0 uniformly for y in any compact subset of Y. If f : I ×Y → X, then we define f : I ×Y → Lp([0, 1] : X) by f (t, y) := f (t+·, y), t ≥ 0, y ∈ Y. For the purpose of research of asymptotically almost periodic solutions of semi- linear fractional Cauchy inclusions, we need to recall the following well-known def- initions (see e.g. Zhang [32], Long, Ding [24], and Kosti´c [20]): Definition 1.1. Let 1 ≤ p < ∞. (i) A function f : I × Y → X is said to be almost periodic iff f (·, ·) is bounded, continuous as well as for every ǫ > 0 and every compact K ⊆ Y there exists l(ǫ, K) > 0 such that every subinterval J ⊆ I of length l(ǫ, K) contains a number τ with the property that kf (t + τ, y) − f (t, y)k ≤ ǫ for all t ∈ I, y ∈ K. The collection of such functions will be denoted by AP (I × Y : X). (ii) A function f : [0, ∞) × Y → X is said to be asymptotically almost periodic iff it is bounded continuous and admits a decomposition f = g + q, where g ∈ AP (R×Y : X) and q ∈ C0([0, ∞)×Y : X). Denote by AAP ([0, ∞)×Y : X) the vector space consisting of all such functions. (iii) A function f : I × Y → X is called Stepanov p-almost periodic, Sp-almost periodic shortly, iff f : I × Y → Lp([0, 1] : X) is almost periodic. (iv) A function f : [0, ∞) × Y → X is said to be asymptotically Sp-almost periodic iff f : [0, ∞)×Y → Lp([0, 1] : X) is asymptotically almost periodic. The collection of such functions will be denoted by AAP Sp([0, ∞)×Y : X). It could be of importance to remind ourselves of the following known facts ([32]): (i) Let f ∈ AP (I × Y : X) and h ∈ AP (I : Y ). Then the mapping t 7→ f (t, h(t)), t ∈ I belongs to the space AP (I : X). (ii) Let f ∈ AAP ([0, ∞)× Y : X) and h ∈ AAP ([0, ∞) : Y ). Then the mapping t 7→ f (t, h(t)), t ≥ 0 belongs to the space AAP ([0, ∞) : X). It can be easily proved that any asymptotically almost periodic two-parameter function is also asymptotically Stepanov p-almost periodic (1 ≤ p < ∞). More details about Stepanov p-almost periodic functions depending on two parameters can be found in [20]. Given s ∈ R in advance, set ⌊s⌋ := sup{l ∈ Z : s ≥ l} and ⌈s⌉ := inf{l ∈ Z : s ≤ l}. The Gamma function is denoted by Γ(·) and the principal branch is always used to take the powers. Define gα(t) := tα−1/Γ(α), t > 0 (α > 0). Fractional calculus and fractional differential equations are rapidly growing fields of research (see e.g. [3], [6], [14]-[15], [27], [29] and [31]). Let γ ∈ (0, 1). Then the Wright function Φγ(·) is defined by the formula Φγ(z) := ∞ Xn=0 (−z)n n!Γ(1 − γ − γn) , z ∈ C. 4 MARKO KOSTI ´C trΦγ(t) dt = Γ(1+r) R ∞ It is well known that Φγ(·) is an entire function, as well as that Φγ(t) ≥ 0, t ≥ 0, 0 e−ztΦγ(t) dt = Eγ(−z), z ∈ C, where 0 Eγ(·) denotes the Mittag-Leffler function. For more details about the Mittag-Leffler and Wright functions, we refer the reader to the doctoral dissertation of Bazhlekova [3] and references cited therein. Γ(1+γr) , r > −1 and R ∞ Let 0 < τ ≤ ∞, let m ∈ N, and let I = (0, τ ). Then we can introduce the Sobolev space W m,1(I : X) in the following way (see e.g. [3, p. 7]): W m,1(I : X) :=(f ∃ϕ ∈ L1(I : X) ∃ck ∈ C (0 ≤ k ≤ m − 1) f (t) = m−1 Xk=0 ckgk+1(t) +(cid:0)gm ∗ ϕ(cid:1)(t) for a.e. t ∈ (0, τ )). If this is the case, we have ϕ(t) = f (m)(t) in distributional sense, and ck = f (k)(0) (0 ≤ k ≤ m − 1). 1.1. Multivalued linear operators in Banach spaces. The multivalued linear operators approach to abstract degenerate differential equations with integer order derivatives has been obeyed in the fundamental monograph [8] by Favini and Yagi (cf. also Kamenskii, Obukhovskii and Zecca [12] for a slightly different approach to condensing multivalued mappings and semilinear differential inclusions in Banach spaces). In what follows, we will present a brief overview of definitions from the theory of multivalued linear operators in Banach spaces. A multivalued map (multimap) A : X → P (Y ) is said to be a multivalued linear operator (MLO) iff the following holds: (i) D(A) := {x ∈ X : Ax 6= ∅} is a linear subspace of X; (ii) Ax + Ay ⊆ A(x + y), x, y ∈ D(A) and λAx ⊆ A(λx), λ ∈ C, x ∈ D(A). If X = Y, then we say that A is an MLO in X. It is well known that, if x, y ∈ D(A) and λ, η ∈ C with λ + η 6= 0, then λAx + ηAy = A(λx + ηy). Assuming A is an MLO, we have that A0 is a linear subspace of Y and Ax = f + A0 for any x ∈ D(A) and f ∈ Ax. Define R(A) := {Ax : x ∈ D(A)}. Then the set A−10 = {x ∈ D(A) : 0 ∈ Ax} is called the kernel of A and it is denoted by N (A). The inverse A−1 of an MLO is defined by D(A−1) := R(A) and A−1y := {x ∈ D(A) : y ∈ Ax}. It can be simply shown that A−1 is an MLO in X, as well as that N (A−1) = A0 and (A−1)−1 = A; A is said to be injective iff A−1 is single-valued. Let A, B : X → P (Y ) be two MLOs. Then we define its sum A + B by D(A + B) := D(A) ∩ D(B) and (A + B)x := Ax + Bx, x ∈ D(A + B). Clearly, A + B is likewise an MLO. Assume that A : X → P (Y ) and B : Y → P (Z) are two MLOs, where Z is likewise a complex Banach space. The product of operators A and B is defined by D(BA) := {x ∈ D(A) : D(B) ∩ Ax 6= ∅} and BAx := B(D(B) ∩ Ax). It is well known that BA : X → P (Z) is an MLO and (BA)−1 = A−1B−1. It is said that an MLO operator A : X → P (Y ) is closed if for any sequences (xn) in D(A) and (yn) in Y such that yn ∈ Axn for all n ∈ N we have that the suppositions limn→∞ xn = x and limn→∞ yn = y imply x ∈ D(A) and y ∈ Ax. We need the following auxiliary lemma from [16]. ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES5 Lemma 1.2. Let Ω be a locally compact, separable metric space, and let µ be a locally finite Borel measure defined on Ω. Suppose that A : X → P (Y ) is a closed MLO. Let f : Ω → X and g : Ω → Y be µ-integrable, and let g(x) ∈ Af (x), x ∈ Ω. Then RΩ f dµ ∈ D(A) and RΩ g dµ ∈ ARΩ f dµ. Let A be an MLO in X, let C ∈ L(X) be injective, and let CA ⊆ AC. Then the C-resolvent set of A, ρC(A) for short, is defined as the union of those complex numbers λ ∈ C for which (i) R(C) ⊆ R(λ − A); (ii) (λ − A)−1C is a single-valued linear continuous operator on X. The operator λ 7→ (λ − A)−1C is called the C-resolvent of A (λ ∈ ρC (A)); the resolvent set of A is defined by ρ(A) := ρI (A), R(λ : A) ≡ (λ − A)−1 (λ ∈ ρ(A)). The basic properties of C-resolvents of single-valued linear operators continue to hold in the multivalued linear setting ([8], [16]). Concerning the abstract degenerate Volterra integro-differential equations and abstract degenerate fractional differential equations, the reader may consult au- thor's forthcoming monograph [16]. Now we will recall the basic things about fractional powers and interpolation spaces of multivalued linear operators. Assume that X is a Banach space, (−∞, 0] ⊆ ρ(A) as well as there exist finite numbers M ≥ 1 and β ∈ (0, 1] such that kR(λ : A)k ≤ M (1 + λ)−β , λ ≤ 0. Then it is not difficult to see that there exist two positive real constants c > 0 and M1 > 0 such that ρ(A) contains an open region Ω = {λ ∈ C : ℑλ ≤ (2M1)−1(c − ℜλ)β, ℜλ ≤ c} of complex plane around the half-line (−∞, 0], where we have the estimate kR(λ : A)k = O((1 + λ)−β), λ ∈ Ω. Designate by Γ′ the upwards oriented curve {ξ ± i(2M1)−1(c − ξ)β : −∞ < ξ ≤ c}. Following Favini and Yagi [8], we define the fractional power A−θ := 1 2πiZΓ′ λ−θ(cid:0)λ − A(cid:1)−1 dλ ∈ L(X) for θ > 1 − β. Set Aθ := (A−θ)−1 (θ > 1 − β). Then the semigroup properties A−θ1A−θ2 = A−(θ1+θ2) and Aθ1 Aθ2 = Aθ1+θ2 hold for θ1, θ2 > 1 − β (recall that the fractional power Aθ need not be injective and the meaning of Aθ is understood in the MLO sense for θ > 1 − β). We topologize the vector space D(A) by the norm k·k[D(A)] := inf y∈A· kyk. Then (D(A), k · k[D(A)]) is a Banach space and the norm k · k[D(A)] is equivalent with the following one k · k + k · k[D(A)]; (D(Aθ), k · k[D(Aθ)]) is likewise a Banach space and we have the equivalence of norms k · k[D(Aθ)] and k · k + k · k[D(Aθ)] for θ > 1 − β (cf. the proof of [8, Proposition 1.1]). Let θ ∈ (0, 1), let and let X θ ξ>0 A :=(x ∈ X : sup (cid:13)(cid:13)·(cid:13)(cid:13)X θ := k · k + sup ξ>0 A, k · kX θ < ∞), ξθ(cid:13)(cid:13)(cid:13) x − x(cid:13)(cid:13)(cid:13) ξ(cid:0)ξ + A(cid:1)−1 ξ(cid:0)ξ + A(cid:17)−1 ξθ(cid:13)(cid:13)(cid:13) · −·(cid:13)(cid:13)(cid:13) . A Then it is well known that (X θ continuously embedded in X. ) is a Banach space as well as that X θ A A We refer the reader to [7]-[8] and [16] for further information concerning inter- polation spaces and fractional powers of multivalued linear operators. 6 MARKO KOSTI ´C 2. Asymptotically almost periodic solutions of abstract fractional Cauchy inclusions (DFP)f,γ and (DFP)f,γ,s In order to formulate our main results, we need to recall two important lemmae regarding the composition principles for asymptotically Stepanov almost periodic two-parameter functions ([20]). Lemma 2.1. Let I = [0, ∞). Suppose that the following conditions hold: (i) g ∈ AP Sp(I ×X : X) with p > 1, and there exist a number r ≥ max(p, p/p− 1) and a function Lg ∈ Lr S(I : X) such that (2.1) kg(t, x) − g(t, y)k ≤ Lg(t)kx − yk, t ≥ 0, x, y ∈ X. (ii) y ∈ AP Sp(I : X), and there exists a Lebesgue's measurable set E ⊆ I with m(E) = 0 such that K = {y(t) : t ∈ I \ E} is relatively compact in X. (iii) f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0([0, ∞) × X : Lq([0, 1] : X)) and q := pr/p + r. (iv) x(t) = y(t) + z(t) for all t ≥ 0, where z ∈ C0([0, ∞) : Lp([0, 1] : X)). (v) There exists a Lebesgue's measurable set E′ ⊆ I with m(E′) = 0 such that K ′ = {x(t) : t ∈ I \ E′} is relatively compact in X. Then q ∈ [1, p) and f (·, x(·)) ∈ AAP Sq(I : X). In the case of consideration of usual Lipschitz type condition (2.2) kf (t, x) − f (t, y)k ≤ Lkx − yk, t ≥ 0, x, y ∈ X, the following result holds true: Lemma 2.2. Let I = [0, ∞). Suppose that the following conditions hold: (i) g ∈ AP Sp(I × X : X) with p ≥ 1, and there exists a constant L > 0 such that (2.2) holds with the function f (·, ·) replaced by the function g(·, ·) therein. (ii) y ∈ AP Sp(I : X), and there exists a Lebesgue's measurable set E ⊆ I with m(E) = 0 such that K = {y(t) : t ∈ I \ E} is relatively compact in X. (iii) f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0([0, ∞) × X : Lp([0, 1] : X)). (iv) x(t) = y(t) + z(t) for all t ≥ 0, where z ∈ C0([0, ∞) : Lp([0, 1] : X)). (v) There exists a Lebesgue's measurable set E′ ⊆ I with m(E′) = 0 such that K ′ = {x(t) : t ∈ I \ E′} is relatively compact in X. Then f (·, x(·)) ∈ AAP Sp(I : X). The following lemma can be proved in exactly the same way as [19, Proposition 2.11]. Lemma 2.3. Suppose that 1 ≤ p < ∞, 1/p + 1/q = 1 and (R(t))t>0 ⊆ L(X, Y ) is k=0 kR(·)kLq[k,k+1] < a strongly continuous operator family satisfying that M :=P∞ ∞. If f : R → X is Sp-almost periodic, then the function G : R → Y, given by (2.3) R(t − s)f (s) ds, t ∈ R, −∞ is well-defined and almost periodic. G(t) :=Z t Remark 2.4. Let p > 1, and let t 7→ kR(t)k, t ∈ (0, 1] be an element of the space k=0 kR(·)kLq[k,k+1] < ∞ holds provided that there exists a finite number ζ < 0 such that kR(t)k = O(tζ ), t → +∞ and ζ < (1/p) − 1. Lq[0, 1]. Then the inequality P∞ ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES7 2.1. Subordinated fractional resolvent families with removable singulari- ties at zero. In this subsection, we analyze the class of multivalued linear operators A satisfying the condition [8, (P), p. 47]: (P) There exist finite constants c, M > 0 and β ∈ (0, 1] such that Ψ := Ψc :=nλ ∈ C : ℜλ ≥ −c(cid:0)ℑλ + 1(cid:1)o ⊆ ρ(A) and kR(λ : A)k ≤ M(cid:0)1 + λ(cid:1)−β , λ ∈ Ψ. Suppose that the condition (P) holds. Then degenerate strongly continuous semi- group (T (t))t>0 ⊆ L(X) generated by A satisfies estimate kT (t)k ≤ M0e−cttβ−1, t > 0 for some finite constant M0 > 0 ([19]). Furthermore, (T (t))t>0 is given by the formula T (t)x = x dλ, t > 0, x ∈ X, 1 2πiZΓ eλt(cid:0)λ − A(cid:1)−1 where Γ is the upwards oriented curve λ = −c(η + 1) + iη (η ∈ R). Let 0 < γ < 1, and let ν > −β. Set Tγ,ν(t)x := tγνZ ∞ 0 sνΦγ(s)T(cid:0)stγ(cid:1)x ds, Following Bazhlekova [3] and Wang, Chen, Xiao [31], we set t > 0, x ∈ X. Sγ(t) := Tγ,0(t) and Pγ(t) := γTγ,1(t)/tγ, t > 0. Recall that (Sγ(t))t>0 is a subordinated (gγ, I)-regularized resolvent family gener- ated by A, which is not necessarily strongly continuous at zero. In [16], we have proved that there exists a finite constant M1 > 0 such that For our later purposes, we need to improve the growth order of these operator (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13) +(cid:13)(cid:13)Pγ(t)(cid:13)(cid:13) ≤ M1tγ(β−1), t > 0. families at infinity: Lemma 2.5. There exists a finite constant M2 > 0 such that (2.4) (2.5) Φγ(s)sβ−1 ds, Proof. By definition of (Sγ(t))t>0 and growth order of (T (t))t>0, we have: (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13) ≤ M2t−γ, t ≥ 1 (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13) ≤ M0Z ∞ 0 Φγ(s)ecstγ(cid:0)stγ(cid:1)β−1 and (cid:13)(cid:13)Pγ(t)(cid:13)(cid:13) ≤ M2t−2γ, t ≥ 1. ds = M0tγ(β−1)Z ∞ ecstγ 0 for any t > 0. Since Φγ(s) ∼ (Γ(1 − γ))−1, s → 0+, a Tauberian type theorem [2, Proposition 4.1.4; b)] immediately implies the first estimate in (2.5). We can prove the second estimate in (2.5) by using the same result, since (cid:13)(cid:13)Pγ(t)(cid:13)(cid:13) ≤ M0tγ(β−1)Z ∞ 0 ecstγ Φγ(s)sβ ds, t > 0. (cid:3) Suppose that x0 ∈ X belongs to the domain of continuity of (Sγ(t))t>0, that is limt→0+ Sγ(t)x0 = x0 (this holds provided that x0 ∈ X belongs to the space D((−A)θ) with 1 ≥ θ > 1 − β or that x ∈ X θ A with 1 > θ > 1 − β). In this subsection, we will employ the following definition of Caputo fractional t u(t) is defined derivatives of order γ ∈ (0, 1). The Caputo fractional derivative Dγ 8 MARKO KOSTI ´C for those functions u : [0, T ] → X for which u(0,T ](·) ∈ C((0, T ] : E), u(·) − u(0) ∈ L1((0, T ) : X) and g1−γ ∗ (u(·) − u(0)) ∈ W 1,1((0, T ) : X), by Dγ t u(t) = d dt"g1−γ ∗(cid:16)u(·) − u(0)(cid:17)#(t), t ∈ (0, T ]. Morover, we will use the following definition (cf. [16, Section 3.5] for more details on the subject): Definition 2.6. By a classical solution of (DFP)f,γ, we mean any function u ∈ C([0, ∞) : X) satisfying that the function Dγ t u(t) is well-defined on any finite interval (0, T ] and belongs to the space C((0, T ] : E), as well as that u(0) = u0 and Dγ t u(t) − f (t) ∈ Au(t) for t > 0. Set Rγ(t) := tγ−1Pγ(t), t > 0. Then we need the following lemma (cf. also [20, Lemma 2.7]): Lemma 2.7. Let f ∈ AAP Sq([0, ∞) : X) with some q ∈ (1, ∞), let 1/q + 1/q′ = 1, and let q′(γβ − 1) > −1. Define H(t) :=Z t 0 Rγ(t − s)f (s) ds, t ≥ 0. Then H ∈ AAP ([0, ∞) : X). Proof. Let the locally p-integrable functions g : R → X, q : [0, ∞) → X satisfy the conditions from [10, Lemma 1], with the number p replaced with q therein, and let the function G(·) be given by (2.3), with the function f (·) replaced by g(·) therein. Since q′(γβ − 1) > −1 and, due to (2.4), kRγ(t)k ≤ M1tγβ−1, t ∈ (0, 1], we can apply Lemma 2.3 (see also Remark 2.4) in order to see that G(·) is almost periodic. On the other hand, it is clear that there exists a number η ∈ (0, 1) such that (1 − η)(1 + γ) > 1. Put F (t) :=Z t 0 Rγ(t − s)q(s) ds −Z ∞ t Rγ(s)g(t − s) ds, t ≥ 0. Owing to Holder inequality, we have that H(·) is well-defined. Since H(t) = G(t) + F (t) for all t ≥ 0, we need to prove that F ∈ C0([0, ∞) : X). Evidently, kRγ(t)k ≤ M2t−γ−1, t ≥ 1 and kRγ(·)kLq′ [t+k,t+k+1]kgkSq ∞ Xk=0 ≤ kRγ(·)kL∞[t+k,t+k+1]kgkSq ≤ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Z ∞ t ∞ ≤ Rγ(s)g(t − s) ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=0 Xk=0 ≤ Const. ∞ kgkSq (t + k)γ+1 ∞ Xk=0 kgkSq tη(1+γ)k(1−η)(γ+1) ≤ Const. t−η(1+γ)kgkSq , t > 1. ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES9 we obtain that: On account of this, we have that limt→∞R ∞ Xk=0 Rγ(t − s)q(s) ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Z t/2 ≤ kgkSq ⌈t/2⌉ 0 ⌈t/2⌉ t Rγ(s)g(t−s) ds = 0. Arguing similarly, kRγ(t − ·)kLq′ [k,k+1] ≤ kgkSq kRγ(t − ·)kL∞[k,k+1] Xk=0 ≤ M2(1 + ⌈t/2⌉)(t − ⌈t/2⌉)−γ−1kgkSq, t ≥ 2; 0 Rγ(t − s)q(s) ds = 0. It remains to be proved that t/2 Rγ(t − s)q(s) ds = 0 (observe that the integral in this limit expres- sion converges by Holder inequality, the estimate q′(γβ − 1) > −1 and the Sq- boundedness of function q(·)). Let a number ǫ > 0 be fixed. Then there exists kq(s)kq ds < ǫq, t ≥ t0. Suppose that t > 2t0 + 6. Then the Holder inequality implies the existence of a finite constant c > 0 such that: t hence, limt→∞R t/2 limt→∞R t t0 > 0 such that R t+1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Rγ(t − s)q(s) ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=0 Xk=0 Z t ≤ c ≤ c t/2 ⌊t/2⌋−2 ⌊t/2⌋−2 ≤ cǫM kRγ(t − ·)kLq′ [t/2+k,t/2+k+1]ǫ + ǫ(cid:13)(cid:13)·β−1(cid:13)(cid:13)Lq′ [0,2] kRγ(t − ·)kL∞[t/2+k,t/2+k+1]ǫ + ǫ(cid:13)(cid:13)·β−1(cid:13)(cid:13)Lq′ [0,2] Xk=0 (t/2 + k)−γ−1 + ǫ(cid:13)(cid:13)·β−1(cid:13)(cid:13)Lq′ [0,2] ∞ ⌊t/2⌋−2 Xk=0 k(1−η)(1+γ) + ǫ(cid:13)(cid:13)·β−1(cid:13)(cid:13)Lq′ [0,2]. ≤ cǫM (t/2)−η(1+γ) This estimate simply completes the proof of lemma. (cid:3) Remark 2.8. Let f ∈ AAP Sq([0, ∞) : X) with some q ∈ (1, ∞), let 1/q + 1/q′ = 1, and let q′(γβ − 1) > −1. Let (Rγ(t))t>0 ⊆ L(X, Y ) be a strongly continuous opera- tor family, and let H : [0, ∞) → Y be defined as above. Then H ∈ AAP ([0, ∞) : Y ), provided that (Rγ(t))t>0 has the same growth rate at zero and infinity as the op- erator family (Rγ(t))t>0 considered above. Our first result reads as follows: Theorem 2.9. Suppose that 1 ≥ θ > 1−β and x0 ∈ D((−A)θ), resp. 1 > θ > 1−β and x0 ∈ X θ A, as well as there exists a constant σ > γ(1 − β) such that, for every T > 0, there exists a finite constant MT > 0 such that f : [0, ∞) → X satisfies kf (t) − f (s)k ≤ MT t − sσ, 0 ≤ t, s ≤ T. Let 1 ≥ θ > 1 − β, resp. 1 > θ > 1 − β, and let f ∈ L∞ loc(cid:16)(0, ∞) :(cid:2)D(cid:0)(−A)θ(cid:1)(cid:3)(cid:17), resp. f ∈ L∞ loc(cid:16)(0, ∞) : X θ A(cid:17). 10 MARKO KOSTI ´C Then there exists a unique classical solution u(·) of problem (DFP)f,γ. If, addi- tionally, f ∈ AAP Sq([0, ∞) : X) with some q ∈ (1, ∞), 1/q + 1/q′ = 1 and q′(γβ − 1) > −1, then u ∈ AAP ([0, ∞) : X). Proof. The first part of theorem is a simple consequence of [16, Theorem 3.5.3], which also shows that the classical solution of (DFP)f,γ is given by the formula u(t) = Sγ(t)x0 +Z t 0 (cid:0)t − s(cid:1)γ−1 Pγ(t − s)f (s) ds, t ≥ 0. The second part of theorem immediately follows from this representation, Lemma 2.5 and Lemma 2.7. (cid:3) Suppose now that 1 ≥ θ > 1 − β and x0 ∈ D((−A)θ), resp. 1 > θ > 1 − β and A. In [16, Section 3.5], we have proved that there exists a finite constant x0 ∈ X θ M1,θ > 0 such that (2.6) resp. (2.7) (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13)L(X,[D((−A)θ)]) +(cid:13)(cid:13)Pγ(t)(cid:13)(cid:13)L(X,[D((−A)θ)]) ≤ M1tγ(β−θ−1), (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13)L(X,X θ A) +(cid:13)(cid:13)Pγ(t)(cid:13)(cid:13)L(X,X θ A) ≤ M1tγ(β−θ−1), We need to improve the estimates (2.6)-(2.7) for long time behaviour: t > 0. t > 0, Lemma 2.10. There exists a finite constant M2,θ > 0 such that A) ≤ M2,θt−γ(θ+1), resp. t > 0, t > 0. A) ≤ M2,θt−γ(θ+1)−1, Let a(t) > 0 satisfy c2(a(t)2 + 1) + a(t)2 = t−2. Using Cauchy theorem, we can deform the path of integration Γ to the upwards oriented curve Γ′, obtained by replacing the union of segments [c, c(a(t) + 1) + ia(t)] ∪ [c(−a(t) + 1) − ia(t), c] of the curve Γ with the part of circle with radius 1/t and center at point c. Applying the computation contained in the proof of [2, Theorem 2.6.1], we get that 2,θecttγ(β−θ−1), (cid:13)(cid:13)T (t)(cid:13)(cid:13)L(X,[D((−A)θ )]) +(cid:13)(cid:13)T (t)(cid:13)(cid:13)L(X,[D((−A)θ)]) ≤ M ′ resp. Now the final conclusions follow similarly as in the proof of Lemma 2.5. (cid:3) (cid:13)(cid:13)T (t)(cid:13)(cid:13)L(X,X θ A) +(cid:13)(cid:13)T (t)(cid:13)(cid:13)L(X,X θ A) ≤ M ′ 2,θecttγ(β−θ−1). Using Lemma 2.10 and Remark 2.8, we can simply prove that the following holds true: Proof. Let t > 2/c. By the proof of [8, Proposition 3.2] and the well known integral computation similar to that appearing in the proof of [7, Lemma 7.1], we have that (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13)L(X,[D((−A)θ)]) +(cid:13)(cid:13)Sγ(t)(cid:13)(cid:13)L(X,X θ (cid:13)(cid:13)Pγ(t)(cid:13)(cid:13)L(X,[D((−A)θ)]) +(cid:13)(cid:13)Pγ(t)(cid:13)(cid:13)L(X,X θ T (t)x, (−λ)θeλt(cid:0)λ − A(cid:1)−1 2πiZΓ 1 as well as ξθhξ(ξ − A)−1 − IiT (t)x = 1 2πiZΓ ξθλ ξ − λ , x ∈ X x dλ! ∈(cid:0)−A(cid:1)θ eλt(cid:0)λ − A(cid:1)−1 x dλ, ξ > 0, x ∈ X. ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES11 Theorem 2.11. Let the requirements of the first part of Theorem 2.9 hold. Then there exists a unique classical solution u(·) of problem (DFP)f,γ. Assume, further, that the initial value x0 satisfies limt→0+ Sγ(t)x0 = x0 in [D((−A)θ)], resp. X θ A, as well as f ∈ AAP Sq([0, ∞) : X) and 1 > θ > 1 − β, resp. f ∈ AAP Sq([0, ∞) : X), with some q ∈ (1, ∞). Let 1/q + 1/q′ = 1 and q′(γ(β − θ) − 1) > −1. Then u ∈ AAP ([0, ∞) : [D((−Aθ))]), resp. u ∈ AAP ([0, ∞) : X θ A). Remark 2.12. (i) The condition q′(γ(β − θ) − 1) > −1 immediately forces that β − γ > 0 and therefore β > 1/2, which can be slightly restrictive in applications (see [8, Section 3.7] for more details). (ii) It is very simple to see that limt→0+ Sγ(t)x0 = x0 in [D((−A)θ)], resp. X θ A, holds provided that limt→0+ T (t)x0 = x0 in [D((−A)θ)], resp. X θ A. Concerning the space [D((−A)θ)], we have that the last equality holds for any x0 ∈ (−A)−θ(Ωc), where Ωc denotes the domain of continuity of semigroup (T (t))t>0; speaking-matter-of-factly, if y0 ∈ Ωc and x0 = (−A)−θy0, then T (t)y0 − y0 ∈ (−A)θ(T (t)x0 − x0), t > 0 and therefore kT (t)x0 − x0k[D((−A)θ )] ≤ kT (t)y0 − y0k, t > 0, which implies the claimed assertion. Concerning the space X θ A, the situation is much more compli- cated and, without any doubt, not well explored in the existing literature. For semilinear problems, we will use the following notion. Definition 2.13. By a mild solution of (DFP)f,γ,s, we mean any function u ∈ C([0, ∞) : X) satisfying that u(t) = Sγ(t)x0 +Z t 0 (cid:0)t − s(cid:1)γ−1 Set, for every x ∈ Cb([0, ∞) : X), Pγ(t − s)f (s, u(s)) ds, t ≥ 0. (Υx)(t) := Sγ(t)x0 +Z t 0 (cid:0)t − s(cid:1)γ−1 Pγ(t − s)f (s, x(s)) ds, t ≥ 0. Suppose that (2.1) holds for a.e. t > 0 (I = [0, ∞)), with locally integrable positive function Lf (·). Set, for every n ∈ N, An := sup 0 (cid:13)(cid:13)Rγ(t − xn)(cid:13)(cid:13) n 0 n × · · ·Z x2 t≥0Z t 0 Z xn Yi=2(cid:13)(cid:13)Rγ(xi − xi−1)(cid:13)(cid:13) ≤ An(cid:13)(cid:13)u − v(cid:13)(cid:13)∞, Yi=1 Then a simple calculation shows that (2.8) (cid:13)(cid:13)(cid:13)(cid:0)Υnu(cid:1) −(cid:0)Υnv(cid:1)(cid:13)(cid:13)(cid:13)∞ The following result is in a close relationship with [20, Theorem 2.10]: Theorem 2.14. Suppose that I = [0, ∞) and the following conditions hold: (i) g ∈ AP Sp(I ×X : X) with p > 1, and there exist a number r ≥ max(p, p/p− 1) and a function Lg ∈ Lr S(I : X) such that (2.1) holds. (ii) f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0(I × X : Lq([0, 1] : X)) and q = pr/p + r. Lf (xi) dx1 dx2 · · · dxn. u, v ∈ BU C([0, ∞) : X), n ∈ N. 12 MARKO KOSTI ´C Set q′ := ∞, provided r = p/p − 1 and q′ := pr pr − p − r , provided r > p/p − 1. Assume also that: (iii) q′(γβ − 1) > −1, (iv) (2.1) holds for a.e. t > 0, with locally bounded positive function Lf (·) satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. Proof. Due to (i)-(ii) and Lemma 2.1, we get that for each x ∈ AAP (I : X) one has f (·, x(·)) ∈ AAP Sq(I : X), where q = pr/p + r; here we would like to re- call only that the range of an X-valued asymptotically almost periodic function is relatively compact in X by [32, Theorem 2.4]. Owing to the assumption (iii), Lemma 2.7 and the obvious equality limt→+∞ Sγ(t)x0 = 0, we get that the mapping Υ : AAP (X) → AAP (X) is well-defined. Using now (2.8), (iv) and a well-known extension of the Banach contraction principle, we obtain the existence of an asymp- totically almost periodic solution of inclusion (DFP)f,γ,s. To prove the uniqueness of mild solutions, assume that u(·) and v(·) are two mild solutions of inclusion (DFP)f,γ,s. Then (2.4) yields that ku(t) − v(t)k ≤ M1Z t 0 (cid:0)t − s(cid:1)γβ−1 Lf (s)ku(s) − v(s)k ds, t ≥ 0. By [6, Lemma 6.19, p. 111], we get that u(s) = v(s) for all s ∈ [0, t] (t > 0 fixed). This completes the proof of theorem. (cid:3) If we employ Lemma 2.2 in place of Lemma 2.1, then we are in a position to formulate and prove the following analogue of Theorem 2.14 in the case of consid- eration of classical Lipschitz condition (2.2): Theorem 2.15. Let I = [0, ∞), and let p > 1. Suppose that the following conditions hold: (i) g ∈ AP Sp(I × X : X) with p ≥ 1, and there exists a constant L > 0 such that (2.2) holds with the function f (·, ·) replaced by the function g(·, ·) therein. (ii) f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0(I × X : Lp([0, 1] : X)). p p−1 (γβ − 1) > −1. (iii) (iv) (2.1) holds for a.e. t > 0, with locally bounded positive function Lf (·) satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. Since any function g ∈ AAP (I × X : X) belongs to the class AAP Sp(I × X : X) for all p > 1, and lim p→+∞ p p − 1 (γβ − 1) = γβ − 1 > −1, Theorem 2.15 immediately implies the following important corollary: ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES13 Corollary 2.16. Suppose that I = [0, ∞), the function f (·, ·) is asymptotically almost periodic and (2.1) holds for a.e. t > 0, with locally bounded positive function Lf (·) satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. It is not trivial to state a satisfactory criterion which would enable one to see that there exists an integer n ∈ N such that An < 1. On the other hand, a very simple calculation involving the estimates (2.4) and (2.5) shows that A1 ≤ L" M1 γβ + M2 γ #, provided the requirements of Corollary 2.16. Hence, we have the following: Corollary 2.17. Suppose that I = [0, ∞), the function f (·, ·) is asymptotically almost periodic and (2.2) holds for some L ∈ [0, ( M1 γ )−1). Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. γβ + M2 Now we would like to present the following illustrative example (see [20] for more details). Example 2.18. (i) (von Wahl [30]) We can simply incorporate Theorem 2.14- Theorem 2.15 in the analysis of existence and uniqueness of asymptotically almost periodic solutions of the following fractional semilinear equation with higher order differential operators in the Holder space X = Cα(Ω) : ( Dγ t u(t, x) = − Pβ≤2m u(0, x) = u0(x), x ∈ Ω, aβ(t, x)Dβu(t, x) − σu(t, x) + f (t, u(t, x)), t ≥ 0, x ∈ Ω; where α ∈ (0, 1), m ∈ N, Ω is a bounded domain in Rn with boundary of )βi, the functions aβ : Ω → C satisfy certain conditions and σ > 0 is sufficiently large. Then the condition (P) holds with the exponent β = 1 − α class C4m, Dβ = Qn 2m and this value is known to be sharp. i=1( 1 i ∂xi ∂ (ii) (Favini, Yagi [8, Example 3.6]) Concerning degenerate semilinear fractional differential equations, it is clear that the possible applications of Theorem 2.14-Theorem 2.15 can be given in the analysis of existence and unique- ness of asymptotically almost periodic solutions of the following fractional Poisson semilinear heat equation in the Lebesgue space X = Lp(Ω) : Dγ v(t, x) = 0, m(x)v(0, x) = u0(x), t [m(x)v(t, x)] = (∆ − b)v(t, x) + f (t, m(x)v(t, x)), t ≥ 0, x ∈ Ω; (t, x) ∈ [0, ∞) × ∂Ω, x ∈ Ω, where Ω is a bounded domain in Rn, b > 0, m(x) ≥ 0 a.e. x ∈ Ω, m ∈ L∞(Ω), 1 < p < ∞ and the operator ∆ − b acts on X with the Dirichlet boundary conditions.   2.2. The non-analyticity of semigroup (T (t))t>0. It is not difficult to observe that the analyticity of degenerate semigroup (T (t))t>0 examined in the previous subsection is a slightly redundant assumption. In this subsection, we investigate the case in which the operator C ∈ L(X) is injective and (T (t))t≥0 ⊆ L(X) is a C-regularized semigroup, i.e, the mapping t 7→ T (t)x, t ≥ 0 is continuous for every 14 MARKO KOSTI ´C fixed element x ∈ X, T (0) = C and T (t + s)C = T (t)T (s) for all t, s ≥ 0. The multivalued linear operator A :=((x, y) ∈ X × X : T (t)x − Cx =Z t 0 T (s)y ds for all t ≥ 0) is said to be the integral generator of (T (t))t≥0; see [14]-[16] for more details on the subject. Let kT (t)k = O(ect), t ≥ 0 for some negative constant c < 0. Then we 0 e−λtT (t)x dt = (λ − A)−1C, know that {λ ∈ C : ℜλ > c} ⊆ ρC (A) as well as R ∞ x ∈ X, ℜλ > c ([16]). Define the operator families Sγ(·), Pγ(·) and Rγ(·) as well as the sequence (An)n∈N and the operator Υ(·) as in the previous subsection. Then the opera- tor A = C−1AC is the integral generator of an exponentially bounded (gγ, C)- regularized resolvent family (Sγ(t))t≥0 (cf. [16, Definition 3.2.2] for the notion), so that A is closed, (Sγ(t))t≥0 is strongly continuous for t ≥ 0, Sγ(t)x − Cx ∈ 0 gγ(t − s)Sγ(s)y ds, t ≥ 0, AR t 0 gγ(t − s)Sγ(s)x ds, t ≥ 0, x ∈ X; Sγ(t)x − Cx =R t whenever y ∈ Ax; Sγ(t)A ⊆ ASγ(t), t ≥ 0 and Sγ(t)C = CSγ(t), t ≥ 0. In this subsection, we will employ the classical definition of Caputo fractional derivatives of order γ ∈ (0, 1). The Caputo fractional derivative Dγ t u(t) is defined for those continuous functions u : [0, ∞) → X for which g1−γ ∗ (u(·) − u(0)) ∈ C1([0, ∞) : X), by Dγ t u(t) = d dt"g1−γ ∗(cid:16)u(·) − u(0)(cid:17)#(t), t ≥ 0. We continue by observing that the arguments contained in the proof of Lemma 2.5 show that there exist two finite constants M3 > 0 and M4 > 0 such that (2.9) as well as (2.10) (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13) +(cid:13)(cid:13)Pγ(t)(cid:13)(cid:13) ≤ M3, t > 0 (cid:13)(cid:13)Sγ(t)(cid:13)(cid:13) ≤ M4t−γ, t ≥ 1 and (cid:13)(cid:13)Pγ(t)(cid:13)(cid:13) ≤ M4t−2γ, t ≥ 1. Definition 2.19. Let x0 ∈ R(C), and let f : [0, ∞) → X be continuous. Then we say that a continuous function u : [0, ∞) → X is a classical solution of (DFP)f,γ iff Dγ t u(t) is well-defined and continuous for t ≥ 0 as well as (DFP)f,γ holds identically for t ≥ 0. Now we state the following result: Theorem 2.20. Suppose that C−1x0 ∈ D(A) and C−1f ∈ W 1,1 loc ((0, ∞) : X). Then there exists a unique classical solution u(·) of problem (DFP)f,γ. If, additionally, f ∈ AAP Sq([0, ∞) : X) with some q ∈ (1, ∞), 1/q + 1/q′ = 1 and q′(γ − 1) > −1, then u ∈ AAP ([0, ∞) : X). Proof. The uniqueness of classical solutions is a simple consequence of the fact that A generates a global (gγ, C)-regularized resolvent family ([16]). To prove the existence of classical solutions, we first observe that the argumentation contained 0 e−λtRγ(t)x dt = (λγ − A)−1C, x ∈ X, λ > 0. By the uniqueness theorem for Laplace transform, we get that in the proof of [13, Theorem 5] (see also [17]) shows that R ∞ (2.11) (cid:0)g1−γ ∗ Rγ(·)x(cid:1)(t) = Sγ(t)x, t ≥ 0, x ∈ X. ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES15 and, on account of this, it is almost trivial to verify that Dγ Let C−1x0 = y0 and z0 ∈ Ay0. Then Sγ(t)y0 − Cy0 =R t t ≥ 0. Set Fγ(t) :=R t 0 gγ(t − s)Sγ(s)z0 ds, t ≥ 0 t Sγ(t)C−1x0 = Sγ(t)z0, 0 Rγ(t − s)C−1f (s) ds, t ≥ 0 and u(t) := Sγ(t)C−1x0 + Fγ(t), t ≥ 0. Using the proof of [2, Proposition 1.3.6], we get that F ′ (2.12) which simply implies that Dγ suffices to show that t Fγ(t) = (g1−γ ∗ F ′ γ(t) =(cid:0)Rγ ∗ (C−1f )′(cid:1)(t) + Rγ(t)(C−1f )(0), γ(cid:1)(t) − f (t) ∈ A(cid:2)Rγ ∗ C−1f(cid:3)(t), (cid:0)g1−γ ∗ F ′ i.e., by (2.11)-(2.12), t ≥ 0, t > 0, x ∈ X, γ)(t), t ≥ 0. By the foregoing, it (cid:0)Sγ ∗ (C−1f )′(cid:1)(t) + Sγ(t)C−1f (0) − f (t) =(cid:0)Sγ ∗ C−1f(cid:1)′ (t) − f (t) ∈ A(cid:2)Rγ ∗ C−1f(cid:3)(t), t ≥ 0. Due to the closedness of A, the only thing that remained to be proved is (cid:16)g1 ∗(cid:0)Sγ ∗ C−1f(cid:1)′(cid:17)(t) −(cid:0)g1 ∗ f(cid:1)(t) ∈ A(cid:2)g1 ∗ Rγ ∗ C−1f(cid:3)(t), i.e., due to (2.11), t ≥ 0, (cid:0)Sγ ∗ C−1f(cid:1)(t) −(cid:0)g1 ∗ f(cid:1)(t) ∈ A(cid:2)gγ ∗ Sγ ∗ C−1f(cid:3)(t), This follows from Lemma 1.2 and the inclusion Sγ(t)x − Cx ∈ A(cid:0)gγ ∗ Sγ(·)x(cid:1)(t), t ≥ 0, x ∈ X. The remaining part of theorem is a consequence of Lemma 2.7. (cid:3) t ≥ 0. Following our analyses from [16, Subsection 2.2.5], it will be convenient to intro- duce the following definition: Definition 2.21. We say that a continuous function t 7→ u(t), t ≥ 0 is a mild solution of the semilinear fractional Cauchy inclusion (DFP)f,γ,s iff u(t) = Sγ(t)C−1x0 +Z t 0 (t − s)γ−1Pγ(t − s)C−1f (s, u(s)) ds, t ≥ 0. Keeping in mind the estimates (2.9)-(2.10) and the foregoing arguments, it is almost straightforward to formulate and prove the following analogues of Theorem 2.14-Theorem 2.15 and Corollary 2.16-Corollary 2.17: Theorem 2.22. Suppose that I = [0, ∞) and the following conditions hold: (i) g ∈ AP Sp(I ×X : X) with p > 1, and there exist a number r ≥ max(p, p/p− 1) and a function Lg ∈ Lr S(I : X) such that (2.1) holds. (ii) C−1f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0(I × X : Lq([0, 1] : X)) and q = pr/p + r. Set q′ := ∞, provided r = p/p − 1 and q′ := pr pr − p − r , provided r > p/p − 1. Assume also that: (iii) q′(γ − 1) > −1, (iv) (2.1) holds with functions f (·, ·) and Lf (·) replaced therein by the functions C−1f (·, ·) and LC−1f (·), respectively, for a.e. t > 0, where LC−1f (·) is a locally bounded positive function satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. 16 MARKO KOSTI ´C Theorem 2.23. Let I = [0, ∞), and let p > 1. Suppose that the following conditions hold: (i) g ∈ AP Sp(I × X : X) with p ≥ 1, and there exists a constant L > 0 such that (2.2) holds with the function f (·, ·) replaced by the function g(·, ·) therein. (ii) C−1f (t, x) = g(t, x) + q(t, x) for all t ≥ 0 and x ∈ X, where q ∈ C0(I × X : Lp([0, 1] : X)). p p−1 (γ − 1) > −1. (iii) (iv) (2.1) holds with functions f (·, ·) and Lf (·) replaced therein by the functions C−1f (·, ·) and LC−1f (·), respectively, for a.e. t > 0, where LC−1f (·) is a locally bounded positive function satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. Corollary 2.24. Suppose that I = [0, ∞), the function C−1f (·, ·) is asymptotically almost periodic and (2.1) holds with functions f (·, ·) and Lf (·) replaced therein by the functions C−1f (·, ·) and LC−1f (·), respectively, for a.e. t > 0, where LC−1f (·) is a locally bounded positive function satisfying An < 1 for some n ∈ N. Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. A very simple calculation involving the estimates (2.4)-(2.5) shows that A1 ≤ L" M3 γ + M4 γ #, provided the requirements of Corollary 2.16. Hence, we have the following: Corollary 2.25. Suppose that I = [0, ∞), the function C−1f (·, ·) is asymptotically almost periodic and (2.2) holds for some L ∈ [0, ( M3 γ )−1). Then there exists a unique asymptotically almost periodic solution of inclusion (DFP)f,γ,s. γ + M4 Theoretical results established here can be applied to a class of abstract degen- erate fractional differential equations, see e.g. [8, Example 2.1] and [16, Example 3.2.14] (C = I). The most important applications with C 6= I can be given to abstract non-degenerate fractional differential equations: Example 2.26. Assume that k ∈ N, aα ∈ C, 0 ≤ α ≤ k, aα 6= 0 for some α with the spaces Lp(Rn) (1 ≤ p ≤ ∞), C0(Rn), Cb(Rn), BU C(Rn), α = k, P (x) = Pα≤k aαiαxα, x ∈ Rn, c′ := supx∈Rn ℜ(P (x)) < 0, X is one of P (D) := Xα≤k aαf (α) and D(P (D)) :=(cid:8)f ∈ E : P (D)f ∈ E distributionally(cid:9). 2 − 1 p , if X = Lp(Rn) for some p ∈ (1, ∞) and nX > n Set nX := n 1 2 , otherwise. Then it is well known (see e.g. [14, Example 2.8.6]) that the operator P (D) gener- ates a global (1 − ∆)−nX k/2-regularized semigroup (T (t))t≥0 satisfying the estimate kT (t)k = O(e−ct), t ≥ 0 for all c ∈ (c′, 0). This immediately implies that we can consider the existence and uniqueness of asymptotically almost periodic solutions of the following semilinear Cauchy problem: (cid:26) Dγ t u(t, x) =Pα≤k aαf (α)(t, x) + f (t, u(t, x)), t ≥ 0, x ∈ Rn, u(0, x) = u0(x), x ∈ Rn. ASYMPTOTICALLY ALMOST PERIODIC SOLUTIONS OF FRACTIONAL RELAXATION INCLUSIONS WITH CAPUTO DERIVATIVES17 We close the paper by providing basic information on asymptotically almost periodic solutions of abstract fractional inclusion Dγ t ∈ R, +u(t) ∈ Au(t) + f (t), (2.13) where Dγ +u(t) denotes the Weyl-Liouville fractional derivative of order γ and f : R → X ([25]). In this paper, Mu, Zhoa and Peng have considered various types of (asymptotically) generalized almost periodic and generalized almost automorphic solutions of (2.13) provided that the operator A is single-valued and generates an exponentially decaying C0-semigroup. By applying the Fourier transform in [25, Lemma 6], the authors have proposed the following definition of mild solution of (2.13): A continuous function u : R → X is said to be a mild solution of (2.13) iff it has the following form (t − s)γ−1Pγ(t − s)f (s) ds, t ∈ R; (t − s)γ−1Pγ(t − s)f (s, u(s)) ds, t ∈ R. a semilinear analogue is −∞ u(t) =Z t u(t) =Z t −∞ We would like to observe that the method proposed in the proof of [25, Lemma 6] is completely meaningful in the case that A is an MLO generating degenerate semigroup with removable singularity at zero or that A is an MLO generating an exponentially decaying C-regularized semigroup. Therefore, we are in a position to analyze the existence and uniqueness of asymptotically almost periodic solutions R 7→ X of (semilinear) fractional Cauchy inclusion (2.13) (cf. Zhang [32] for the notion) by using similar arguments to those employed in the proofs of [25, Theorem 8, Theorem 9, Theorem 17, Theorem 19]. Details can be left to the interested reader. References [1] M. Amerio, G. Prouse, Almost Periodic Functions and Functional Equations, Van Nostrand- Reinhold, New York, 1971. [2] W. Arendt, C. J. K. Batty, M. Hieber, F. Neubrander, Vector-valued Laplace Transforms and Cauchy Problems, Birkhauser/Springer Basel AG, Basel, 2001. [3] E. Bazhlekova, Fractional Evolution Equations in Banach Spaces, PhD Thesis, Eindhoven University of Technology, Eindhoven, 2001. [4] D. N. Cheban, Asymptotically Almost Periodic Solutions of Differential Equations, Hindawi Publishing Corporation, 2009. [5] T. Diagana, Almost Automorphic Type and Almost Periodic Type Functions in Abstract Spaces, Springer-Verlag, New York, 2013. [6] K. Diethelm, The Analysis of Fractional Differential Equations, Springer-Verlag, Berlin, 2010. [7] A. Favaron, A. Favini, Fractional powers and interpolation theory for multivalued linear operators and applications to degenerate differential equations, Tsukuba J. Math. 35 (2011), 259 -- 323. [8] A. Favini, A. Yagi, Degenerate Differential Equations in Banach Spaces, Chapman and Hall/CRC Pure and Applied Mathematics, New York, 1998. [9] G. M. N'Gu´er´ekata, Almost Automorphic and Almost Periodic Functions in Abstract Spaces, Kluwer Acad. Publ., Dordrecht, 2001. [10] H. R. Henr´ıquez, On Stepanov-almost periodic semigroups and cosine functions of operators, J. Math. Anal. Appl. 146 (1990), 420 -- 433. [11] Y. Hino, T. Naito, N. V. Minh, J. S. Shin, Almost Periodic Solutions of Differential Equations in Banach Spaces, Stability and Control: Theory, Methods and Applications, 15. Taylor and Francis Group, London, 2002. 18 MARKO KOSTI ´C [12] M. Kamenskii, V. Obukhovskii, P. Zecca, Condensing Multivalued Maps and Semilinear Differential Inclusions in Banach Spaces, de Gruyter, Berlin-New York, 2001. [13] V. Keyantuo, C. Lizama, M. Warma, Spectral criteria for solvability of boundary value prob- lems and positivity of solutions of time-fractional differential equations, Abstr. Appl. Anal., Vol. 2013, Article ID 614328, 11 pages, 2013. [14] M. Kosti´c, Generalized Semigroups and Cosine Functions, Mathematical Institute SANU, Belgrade, 2011. [15] M. Kosti´c, Abstract Volterra Integro-Differential Equations, CRC Press, Boca Raton, Fl., 2015. [16] M. Kosti´c, Abstract Degenerate Volterra Integro-Differential Equations: Linear Theory and Applications, Book Manuscript, 2016. [17] M. Kosti´c, A note on semilinear fractional equations governed by abstract differential oper- ators, An. S¸tiint¸. Univ. Al. I. Cuza Ia¸si. Mat. (N.S.), LCXII 2:3 (2016), 757 -- 762. [18] M. Kosti´c, Almost periodicity of abstract Volterra integro-differential equations, Adv. Oper. Theory 2 (2017), 353 -- 382. [19] M. Kosti´c, Abstract Volterra integro-differential equations: generalized almost periodicity and asymptotical almost periodicity of solutions, Electron. J. Differential Equations, vol. 2017, no. 239 (2017), 1 -- 30. [20] M. Kosti´c, The existence and uniqueness of almost periodic and asymptotically almost peri- odic solutions of semilinear Cauchy inclusions, Hacet. J. Math. Stat., submitted. [21] M. Kosti´c, The existence and uniqueness of pseudo-almost periodic solutions of fractional Sobolev inclusions, Appl. Math. Comp. Sci. 2 (2017), 1 -- 12. [22] M. Levitan, V.V. Zhikov, Almost Periodic Functions and Differential Equations, Cambridge Univ. Press, London, 1982. [23] F. Li, J. Liang, H. Wang, S-asymptotically ω-periodic solution for fractional differential equa- tions of order q ∈ (0, 1) with finite delay, Adv. Difference Equ. 2017:83, doi:10.1186/s13662- 017-1137-y. [24] W. Long, S.-H. Ding, Composition theorems of Stepanov almost periodic functions and Stepanov-like pseudo-almost periodic functions, Adv. Difference Equ., Vol. 2011, Article ID 654695, 12 pages, doi:10.1155/2011/654695. [25] J. Mu, Y. Zhoa, L. Peng, Periodic solutions and S-asymptotically periodic solutions to frac- tional evolution equations, Discrete Dyn. Nat. Soc., Volume 2017, Article ID 1364532, 12 pages, https://doi.org/10.1155/2017/1364532. [26] F. Periago, B. Straub, A functional calculus for almost sectorial operators and applications to abstract evolution equations, J. Evol. Equ. 2 (2002), 41 -- 68. [27] J. Pruss, Evolutionary Integral Equations and Applications, Monogr. Math. 87, Birkhauser, Basel, Boston, Berlin, 1993. [28] W. M. Ruess, W. H. Summers, Asymptotic almost periodicity and motions of semigroups of operators, Linear Algebra Appl. 84 (1986), 335 -- 351. [29] S. G. Samko, A. A. Kilbas, O. I. Marichev, Fractional Derivatives and Integrals: Theory and Applications, Gordon and Breach, New York, 1993. [30] W. von Wahl, Gebrochene Potenzen eines elliptischen Operators und parabolische Differen- tialgleichungen in Raumen holderstetiger Funktionen, Nachr. Akad. Wiss. Gottingen Math.- Phys. Kl. 11 (1972), 231 -- 258. [31] R.-N. Wang, D.-H. Chen, T.-J. Xiao, Abstract fractional Cauchy problems with almost sec- torial operators, J. Differential Equations 252 (2012), 202 -- 235. [32] C. Zhang, Ergodicity and asymptotically almost periodic solutions of some differential equa- tions, IJMMS 25 (2001), 787 -- 800. Faculty of Technical Sciences, University of Novi Sad, Trg D. Obradovi´ca 6, 21125 Novi Sad, Serbia E-mail address: [email protected]
1209.3042	1	1209	2012-09-13T21:12:17	A Banach space in which every injective operator is surjective	[ "math.FA", "math.GN" ]	We construct an infinite dimensional Banach space of continuous functions C(K) such that every one-to-one operator on C(K) is onto.	math.FA	math	A BANACH SPACE IN WHICH EVERY INJECTIVE OPERATOR IS SURJECTIVE ANTONIO AVIL´ES AND PIOTR KOSZMIDER Abstract. We construct an infinite dimensional Banach space of continuous functions C(K) such that every one-to-one operator T : C(K) −→ C(K) is onto. 1. Introduction Already S. Banach has asked if every infinite dimensional Banach space X has a proper subspace isomorphic to X ([5]). It took many decades until this problem has been solved in the negative by T. Gowers in [13]. Thus, in spaces X like in [13] all isomorphisms T : X → X must be onto X. This property is shared by spaces with few operators constructed and investigated for example in [12] or [3] based on hereditarily indecomposable Banach spaces (abbreviated HI) as well as in Banach spaces of the form C(K) with few operators first constructed in [19]. In this paper answering a question from [14] we strengthen this property proving the following: Theorem 1.1. There is an infinite dimensional Banach space X such that when- ever a bounded linear operator T : X → X is injective, then it is an isomorphism onto X. Moreover X is of the form C(K). Our approach is to construct a Banach space of the form C(K) with few op- erators in the sense of [19]. Additional special properties of the compact space present in our construction (K is an almost P -space) guarantee that every injective operator is an isomorphism and so previously known arguments for C(K) spaces with few operators can be used to conlude that the operator is onto. This seems the only possible approach of modifying known spaces with few operators that can work. Indeed, separable Banach spaces and HI spaces cannot serve for our purposes because the main property may be shared only by finite dimensional spaces or those which have nonseparable dual space in the weak∗ topology (see [14]): Proposition 1.2. Suppose X is an infinite dimensional Banach space whose dual ball which is separable in the weak∗ topology. Then there is an injective operator on X which is not an automorphism of X. In particular this property is shared by all separable infinite dimensional Banach spaces and all HI spaces. Proof. By a result from [11], if the dual ball of X is separable in the weak∗ topology, then there is a compact injective operator T : X → X. A compact operator cannot be onto an infinite dimensional Banach space. As any HI space embeds into ℓ∞ (A. First author by was supported by MEC and FEDER (Project MTM2011-25377), Ramon y Cajal contract (RYC-2008-02051) and an FP7-PEOPLE-ERG-2008 action. The second author was partially supported by the National Science Center research grant DEC-2011/01/B/ST1/00657. 1 2 ANTONIO AVIL ´ES AND PIOTR KOSZMIDER IV. 6 of [4]) whose dual ball is separable in the weak∗ topology, we conclude that such spaces have separable dual balls as well. (cid:3) We also prove (see 2.3, 2.4 and the remarks after it) that a Banach space of the form C(K) where all injective operators are automorphisms cannot be one of the known constructions as in [19, 25, 20, 10, 22]. Another remark is that our space satisfies a version of the invariant subspace property for nonseparable spaces, namely every operator T : X −→ X has an invariant subspace Y such that both Y and X/Y are of density c. Indeed, the subspace Y can be chosen to be complemented and such that still every injective operator Y −→ Y or X/Y −→ X/Y is an isomorphism. This is analogous to what happens in some separable Banach spaces with few operators [1, 2], where it can be proven that every operator has a proper infinite-dimensional invariant subspace. We wish to express our gratitude to Amin Kaidi. The discussion with him during a visit of the first author to the University of Almeria is in the origin of this paper. 2. Almost P -spaces and weak multipliers If K is a compact and Hausdorff and g ∈ C(K) we can define an operator Tg : C(K) → C(K) by Tg(f ) = f g, likewise for g a Borel bounded function on K we can define g∗ T : M (K) → M (K) by R f dg∗ T (µ) = R f gdµ for every µ ∈ M (K) and f ∈ C(K). Definition 2.1. Let T : C(K) → C(K) be a linear bounded operator. We say that T is a weak multiplication if and only if there is a g ∈ C(K) and a weakly compact operator S : C(K) → C(K) such We say that T is a weak multiplier if and only if there is a Borel g∗ : K → K and a weakly compact operator S : C∗(K) → C∗(K) such that T = Tg + S. T ∗ = g∗ T + S. We will say that a space C(K) has few operators if and only if all operators on C(K) are weak multiplications or weak multipliers. Of course a weak multiplication is a weak multiplier and the form of the weak multiplication is much nicer. However, the notion of a weak multiplier plays a more natural role for example having all operators weak multipliers is invariant under isomorphisms of C(K) spaces and having all operators weak multiplications is not ([28]). It is easy to construct a weak multiplier on C(K) which is not a weak multiplica- tion if we have a function which is discontinuous at one point and the discontinuity 2 ] is such cannot be removed by changing the value in this point. For example χ[0, 1 a function on [0, 1] and we get T : C([0, 1]) → C([0, 1]) defined by T (f ) = χ[0, 1 2 ](f − f ( 1 2 )), then T ∗ = g∗ T where g∗ = χ[0, 1 2 ] is the Borel function which is discontinuous at one point 1 2 . We say that an x ∈ K is C∗-embedded if and only every continuous g : K \ {x} → [0, 1] can be extended to a continuous function on K. It is proved in 2.7 of [19] that every weak multiplier is a weak multiplication on C(K) if and only if for every x ∈ K the subspace K \ {x} is C∗-embedded in K. 3 Definition 2.2 ([17]). A topological space is called an almost P -space if and only if every nonempty Gδ-set has non-empty interior. Proposition 2.3. If K is not a weak P -space, then there is a an injective operator T : C(K) → C(K) which is not an automorphism of C(K). Proof. Let Un ⊆ K be open subsets of K for n ∈ N such that Tn∈N Un = F 6= ∅ has empty interior. We may assume that Un+1 ⊆ Un for each n ∈ N, and so that F is closed. Let gn : K → [0, 1/2n] be continuous functions satisfying gn(x) = 1/2n whenever x ∈ K \ Un and gn(x) = 0 whenever x ∈ F . Let g = Σn∈Ngn. Note that g ∈ C(K) assumes value 0 in x ∈ K if and only if x ∈ F . Consider the operator Tg. The operator is injective, because any nonzero f ∈ C(K) has values separated from zero on an open set, i.e., not included in F . Tg is not an isomorphism because whenever n ∈ N and the support of f is inluded in Un we have Tg(f ) ≤ (1/2n−1)f . It follows from the open mapping theorem that the image of Tg is not closed, in particular, Tg is not onto C(K). (cid:3) Recall that a topological space satisfies the countable chain condition if and only if it does not admit an uncountable pairwise disjoint collection of open sets. Lemma 2.4. No infinite compact Hausdorff almost P -space K satisfies the count- able chain condition. Proof. We may assume that K has at most countably many isolated points. This excludes scattered spaces where isolated points form a dense open set whose com- plement would be Gδ under the above assumption. In any nonscattered compact space we can construct a family {Us : s ∈ {0, 1}n, n ∈ N} of nonempty open sets such that Us ⊆ Ut whenever t ⊆ s and Us ∩ Ut = ∅ whenever s ∪ t is not a function where s, t ∈ {0, 1}n and n ∈ N. If K is a weak P -space for every σ ∈ {0, 1}N there is nonempty open Uσ ⊆ Us for all s ∈ {0, 1}n with s ⊆ σ. It follows that {Uσ : σ ∈ {0, 1}N} is a pairwise disjoint family of nonempty open sets of K of cardinality continuum. (cid:3) The compact spaces constructed in [19, 25, 20, 10, 22] all satisfy the countable chain condition (which is essentially used), and so they are not almost P -spaces. It follows that the constructions like in [19, 25, 20, 10, 22] cannot be used to conclude the main result of this paper. The crucial properties of P -spaces in the context of multiplications and weakly compact operators are the following: Lemma 2.5. Suppose that K is a compact Hausdorff space which does not satisfy the countable chain condition. Then no weakly compact operator T : C(K) → C(K) is injective. Proof. Suppose that T : C(K) → C(K) is weakly compact. Then T (fn) → 0 for any sequence of pairwise disjoint (i.e., f . nfm = 0 for distinct n, m ∈ N) functions in C(K) ([7]). Let {Ui : i ∈ I} be an uncountable pairwise disjoint family of nonempty open subsets of K existing by Lemma 2.4 and let fi ∈ C(K) be nonzero functions with supports included in Ui for each i ∈ I. It follows that T (fi) = 0, that is T (fi) = 0, for all but countably many i ∈ I, and therefore T is not injective. (cid:3) Lemma 2.6. Suppose that K is a compact Hausdorff which is an almost P -space and g ∈ C(K). Tg is an isomorphism onto its range if and only if Tg is injective. 4 ANTONIO AVIL ´ES AND PIOTR KOSZMIDER Proof. If Tg is not an isomorphism onto its range, then g(x0) = 0 for some x0 ∈ K. Since {x ∈ K : g(x) = 0} is Gδ, it follows that it contains a nonempty open set U ⊆ K. Taking a nonzero function f ∈ C(K) with its support in U , we have Tg(f ) = 0 and hance Tg is not injective. (cid:3) Proposition 2.7. Suppose that K is a compact totally disconnected almost P -space such that every operator on C(K) is a weak multiplier. Then every injective linear bounded operator on C(K) is an automorphism of C(K). Proof. Let T be an injective operator on C(K). Since we assume that all operators on C(K) are weak multipliers, by Theorem 2.2. (b) of [19] there is a bounded borel g : K → R with at most countably many points of discontinuity and a weakly compact operator S : C(K)∗ → C(K)∗ such that T ∗ = g∗ T + S. Let X = {xn : n ∈ N} be the set of the points of discontinuity of g. First we will prove that there is no infinite sequence Y = {yn : n ∈ N} of points of K such that g(yn) < 1/n. Supose that there is such a Y and let us derive a contradiction. Let y ∈ K \ X be an accumulation point of Y . It exists as the set of accumulation points of Y is closed and with no isolated points since K cannot have nontrivial convergent sequences because this would give a complemented in C(K) copy of c0 yielding a noncommutativity of the quotent ring of all operators on C(K) divided by weakly compact operators, contradicting by 4.5. of [28] the fact that all operators on C(K) are weak multipliers. Moreover g(y) = 0 since g is continuous at y. Consider Vn = g−1[(−1/n, 1/n)] \ X. It is a relative open set in K \ X. Let Un be an open set of K such that Un \ X = Vn. Let U = T(Un \ {xk : k < n}). U ∋ y is a non-empty Gδ set, and so it has a nonempty interior W . We may assume that W is clopen. Of course gW = 0. Taking a function f ∈ C(K) with support in W and any µ ∈ M (K) we have T ∗(µ)(f ) = Z gf dµ + S(µ)(f ) = 0 + S(µ)(f ). So the dual of the restriction of T to functions with supports in W is weakly compact and so by Gantmacher theorem the restriction of T to functions with supports in W is weakly compact. In particular T is not injective by Lemma 2.5 as W is an almost P -space. It follows that T is not injective, a contradiction. This completes the proof of the nonexistence of Y as above. Hence there is k ∈ N such that g(x) > 1/k for all but finitely many points x ∈ K. By modifying g to g on this finite set of points we obtain T ∗ =g∗ T + S′ where g(x) > 1/k for all points x ∈ K and S′ = S +(g−g)∗ T is weakly compact on M (K). g∗ T is now an automorphism on M (K), S′ is strictly singular as weakly compact on a dual to C(K) by [8, VI.8.10 and p. 394]. By Propositon 2 (c) 10 of [24] T ∗ =g∗ T +S′ is Fredholm and so has a closed range and has the same Fredholm index as g∗ T that is zero. As T is injective, T ∗ is onto, hence as an operator of Fredholm index zero T ∗ has zero kernel an by the open mapping theorem it is an automorphism of M (K). Hence T is an automorphism of C(K) by Ex. 2.39 of [9] which completes the proof of the proposition. (cid:3) In this paper we shall construct a compact space K that satisfies the hypotheses of Proposition 2.7. We do not know if one can prove without extra set-theoretic axioms that such a K can exist so that every operator on C(K) is actually a weak 5 multiplication. In fact the countable chain condition was needed in the construc- tions like in [19, 25, 20, 10, 22] to ensure that K \ {x} is C∗-embedded in K for every x ∈ K and consequently (by 2.7 of [19]) that every weak multiplier is actually a weak multiplication. It should be noted however, that assuming an additional set-theoretic axiom ♦ (see [16, 23]) such construction seems possible. Let us say that a Boolean algebra is an almost P -algebra if its Stone space is an almost P -space. We will use the symbols ∧, ∨, − and ≤ to denote the usual operations and the order in a Boolean algebra. If a is an element of a Boolean algebra, we will denote by [a] the corresponding clopen subset in its Stone space. Lemma 2.8. Suppose that A is a Boolean algebra which is an almost P -algebra. Suppose that fn ∈ C(KA) for each n ∈ N and Pn∈N fn(x) > δ for some x ∈ KA and some δ ∈ R. Then there is a ∈ A such that fn(y) > δ Xn∈N for every y ∈ [a]. Proof. By the continuity of fn, for each n ∈ N there is an an ∈ A such that x ∈ [an] and fn(y) is close to fn(x) for each y ∈ [an]. So Tn∈N[an] is a nonempty Gδ set and so has a nonempty interior as KA is an almost P -space. It follows that there is a ∈ A as required. (cid:3) 3. Separation by submorphisms instead of separation by suprema We say that two subsets I and J of a Boolean algebra A are orthogonal if x∧y = 0 for all x ∈ I, y ∈ J. We say that I and J are separated in A if there exists c ∈ A such that x ≤ c, y ∧ c = 0 for all x ∈ I, y ∈ J. We will say that a function between Boolean algebras φ : A1 −→ A2 is a sub- morphism if φ(a ∧ b) = φ(a) ∧ φ(b), φ(a ∨ b) = φ(a) ∨ φ(b) and φ(0) = 0. If it also satisfies φ(1) = 1, then we call φ a morphism. Theorem 3.1. Let A be a Boolean algebra which is an almost P -algebra and KA its Stone space. Suppose that for every pairwise disjoint family a = {an : n < ω} ⊂ A \ {0} there exists a submorphism φa : Pa −→ A such that (1) Pa is a subalgebra of P(ω) that contains all finite sets. (2) φa({n}) = an (3) For any infinite σ ⊂ ω, and for any family of pairwise disjoint {bn, n ∈ σ} ⊂ A \ {0} orthogonal with {an : n ∈ N}, there exists τ ⊂ σ such that (a) τ ∈ Pa (b) {bn : n ∈ τ } and {bn : n ∈ σ \ τ } are not separated in A. Then every operator T : C(KA) −→ C(KA) is a weak multiplier. Proof. Suppose that A is a as above and that T : C(KA) → C(KA) is not a weak multiplier. By 2.1 and 2.2 of [19] there is an ε > 0, an antichain {an : n ∈ N} of A and points xn ∈ K such that xn 6∈ [an] and 1) T (1An)(xn) > ε. 6 ANTONIO AVIL ´ES AND PIOTR KOSZMIDER By a slight modification of the points xn we may assume that they are not in the closed and nowhere dense set F = Sn∈N[an] \ Sn∈N[an]. By applying the Ramsey theorem [16, Theorem 9.1] to a coloring c : [N]2 → {0, 1} where c(n, m) = 1 iff xn ∈ [am] or xm ∈ [an] and the fact that {an : n ∈ N} is an antichain we may assume, by taking an infinite subset of N that xn 6∈ [am] for any two distinct n, m ∈ N. By applying the Rosenthal lemma [26, Lemma 1.1] by taking an infinite subset of N we may assume that Σn∈N\{k}T (1[an])(xk) < ε/3 holds for all k ∈ N. Of course if T (1[an])(xk) < δ, for some δ > 0, then there is bk n 6= 0 in A such that xk ∈ [bk n] and T (1[an])[bk n for all k ∈ N, so by the choice of xn outside of F we may assume that {bn : n ∈ N} and {an : n ∈ N} are orthogonal and n] < δ. Since A is an almost P -algebra there are 0 6= bk ≤ bk 2) Σn∈N\{k}T (1[an])bk < ε/3. Moreover, by 1) and considering a bit smaller bks we may assume that for all n ∈ N and for all x ∈ [bn] we have 3) T (1an)(x) > ε. Let φa and Pa be as in the statement of the Theorem for a = {an : n ∈ N}. Case 1. There is an infinite σ ⊆ N and 0 6= ck ≤ bk for all k ∈ N such that T (1φa(τ ))[ck] = Xn∈τ T (1[an])[ck] holds for every τ ⊆ σ, τ ∈ Pa. In this case, by 2) and 3), if k ∈ τ , for every x ∈ [ck] we have T (1φa(τ ))(x) = Xn∈τ 1 T (1[an])(x) > 2ε/3 and if k 6∈ τ , for every x ∈ [ck] we have T (1φa(τ ))(x) = Xn∈τ 1 T (1[an])(x) < ε/3 Then (T (1φa(τ )))−1((−∞, ε/3]) and (T (1φa(τ )))−1([2ε/3, ∞)) separate {cn : n ∈ τ } from {cn : n 6∈ τ } whenever τ ⊂ σ, τ ∈ Pa. This contradicts condition (3) of the Theorem. Case 2. Case 1 does not hold. Assuming the negation of the condition from case 1, we will carry out a transfinite inductive construction which will contradict the boundedness of the operator T . Let {σξ : ξ < ω1} be an almost disjoint (i.e., such that pairwise intersections of its elements are finite) family of infinite subsets of N. For all ξ < ω1 construct: • an infinite τξ ⊆ σξ, • an antichain {cξ ξ ≤ ξ′ < ω1, k : k ∈ N} such that 0 6= cξ′ k ≤ cξ k ≤ bk for all k and all • nξ, mξ ∈ N \ {0} such that for all x ∈ [cξ nξ ] we have T (1φa(τξ))(x) − Xn∈τξ (4) T (1an)(x) > 1/mξ. The possiblity of such a construction follows from the fact that KA is an almost P -space Lemma 2.8 and the assumption that Case 1 does not hold. A single pair (n′, m′) has appeared infinitely many times as (nξ, mξ). Let i ∈ N be such that i/6m′ > T and consider ξ1 < ... < ξi such that (nξi , mξi ) = (n′, m′). Let n0 ∈ N be such that the pairwise intersections of all τξ1 , ..., τξi be included in {0, ..., n0 − 1}. Puting τ ′ ξ = τξ \ {0, ..., n0 − 1} and noting that 7 1φa(τξ) = 1φa(τ ′ ξ) + Xn∈τξ∩n0 1an we conclude from (4) that T (1φa(τ ′ ξj ))(x) − Xn∈τ ′ ξj T (1an)(x) > 1/m′ ξj ξj n′ ] and all 1 ≤ j ≤ i. By taking n′ T (1an)(x) < 1/3m′ and so T (1φa(τ ′ for all x ∈ [cξi 0 > n0 and using (2) we may assume ))(x) > 2/3m′ for all x ∈ [cξi that Pn∈τ ′ n′] for all 1 ≤ j ≤ i and consequently for some F ⊆ {1, ..., i} of cardinality not smaller than i/2 and some x ∈ [cξi Xj∈F ))(x) > (1/3m′)(i/2) > T . n′ ] we have T (1φa(τ ′ ξj However τ ′ Pj∈F 1φa(τ ′ ξs are pairwise disjoint, so φa(τ ′ ξ) = 1 and so T (Pj∈F 1φa(τ ′ ξj ξ)s are pairwise disjoint. It follows that )) ≤ T , a contradiction. (cid:3) Remarks. Theorem 3.1 holds true even without the assumption that A is an almost P -algebra. The proof in the general case would follow the steps of [27, Theorem 3.3.2], which corresponds to the case in which Pa is the algebra of all subsets τ ⊂ ω such that {an : n ∈ τ } has a supremum in A, and φa(τ ) = W{an : n ∈ τ }. We decided to include a self-contained proof, and the assumption that A is a P -agebra simplifies the argument and it is enough for our purposes. Lemma 3.2. Let A ⊂ B be Boolean algebras such that A < c ≤ B, let φ : P(ω) −→ B be a submorphism such that φ{n} ∈ A for every n < ω, let I be a set with I < c and suppose that for every i ∈ I, Xi, Yi ⊂ A are orthogonal sets which are nonseparated in A. Then, there exists an infinite set τ ⊂ ω such that Xi nd Yi remain unseparated in Ahφυi for all υ ⊂ τ and all i ∈ I. Proof. We follow a similar argument as in [19, p. 1659]. Let M be an almost disjoint family of subsets of ω with M = c. We shall prove that there exists τ ∈ M that satisfies the statement of the Lemma. We proceed by contradiction, so suppose that for every τ ∈ M there exist i ∈ I and υ ⊂ τ such that Yi and Xi are separated in Ahφυi. The separation means that there exist pairwise disjoint b, c, d ∈ A such that z = (b ∧ φυ) ∨ (c ∧ −φυ) ∨ d separates Yi and Xi. Since M = c, and there are less than c many choices for i, b, c and d, it follows that there exists a single i ∈ I and three fixed elements b, c and d such that for at 8 ANTONIO AVIL ´ES AND PIOTR KOSZMIDER least two different choices of τ ∈ M there exists υ ⊂ τ such that z(υ) = (b ∧ φυ) ∨ (c ∧ −φυ) ∨ d separates Yi and Xi. So pick τ 6= τ ′ in M, υ ⊂ τ , υ′ ⊂ τ ′ as above. Then, since both z(υ) and z(υ′) separate Yi from Xi, z0 = (b ∧ φυ ∧ φυ′) ∨ (c ∧ −(φυ ∧ φυ′)) ∨ d also separates Yi from Xi. But z0 ∈ A since φυ ∧ φυ′ = φ(υ ∩ υ′) = _n∈υ∩υ′ φ{n} as υ ∩ υ′ is finite. This is a contradiction with the hypothesis that Yi and Xi are nonseparated in A. (cid:3) 4. The construction Along this section, B will be the Boolean algebra of subsets of N modulo finite sets, B = P(N)/F in. Theorem 4.1. There exists an infinite Boolean algebra A ⊂ B which satisfies the hypotheses of Theorem 3.1. Proof. We shall construct A as a union of a c-chain of subalgebras of B, each subalgebra being of size less than c, in the form A = Sα<c Aα, Aα < c. The algebras Aα are constructed by induction. For the first step, A0 can be any arbitrary infinite subalgebra of B of cardinality less than c. Let us write c = [α<c Fα ∪ [α<β<c Gαβ as a disjoint union where Fα = Gαβ = c and Fα ∩ α = Gαβ ∩ β = ∅ for each α < β < c. We write Gαβ = {ξδγ αβ is a bijection between c × c and Gαβ . Once the algebra Aα is defined, we make the following enumerations: αβ : δ, γ < c} in such a way that (δ, γ) 7→ ξδγ (1) {(dξ n)n∈N : ξ ∈ Fα} are all strictly decreasing sequences of Aα. (2) {{aγ n(α) : n ∈ N} : γ < c} are all pairwise disjoint families in Aα \ {0}. We fix morphisms ψγ n(α). Such morphisms can be the following way: let σn ⊆ N be pairwise disjoint sets such that h(σn) = an where h : ℘(N) → ℘(N)/F in = B is the canonical surjective homomorphism; now define ψγ α : P(ω) −→ B such that ψγ α({n}) = aγ α(ρ) = h[Sn∈ρ σn]. (3) {{bδ n(α)) : n ∈ σδ(α)} : δ < c} are all possible families where σδ(α) is an n(α) : n ∈ σδ(α)} is a family of pairwise disjoint infinite subset of N and {bδ elements of Aα \ {0}. At limit steps we shall define Aα = Sβ<α Aβ. At a successor stage, Aξ+1 will be the algebra generated by Aξ and a certain element xξ that we will add, Aξ+1 = Aξhxξi. When ξ ∈ Fα for some α, xξ will be a lower bound of {dξ n : n < ω}. When ξ = ξγδ n(β), n ∈ σδ(β)} is pairwise αβ ∈ Gαβ for some α < β and the family {aγ n(α), bδ 9 + = {bδ disjoint, we shall find an infinite τ ξ ⊂ σδ(β) such that Bξ n(β) : n ∈ τ ξ} and Bξ − = {bδ n(β) : n ∈ σδ(β) \ τ ξ} are nonseparated in Aα and we shall define xξ = ψγ α(τ ξ), Aξ+1 = Aξhxξi. In this way, in the steps in Fα we take care that we obtain an almost P -space, while in the steps in Gαβ we take care that the hypothe- ses of Theorem 3.1 are satisfied1. In order this to work we must make sure that when we construct Aξ+1, those families of the form Bξ′ − that were chosen to be nonseparated in some previous step ξ′ < ξ, remain nonseparated in Aξ+1, assuming inductively that they were kept nonseparated in Aξ. More precisely, this is done as follows: + and Bξ′ (1) If ξ ∈ Gαβ for some α < β ≤ ξ, then ξ = ξγδ n(α) : n < ω} and ψ = ψγ n(β) : n ∈ σ = σδ(β)} on the other hand. αβ for some γ, δ < c. We have {an = aγ α : P(ω) −→ B on the one hand, and {bn = bδ If an ∧ bm 6= ∅ for some n, m ∈ σ we do nothing and just define Aξ+1 = Aξ (we call this a trivial step). Otherwise, we can apply Lemma 3.2 for B = B, A = Aξ, φ = ψP(σ) : P(σ) −→ B, and the sets Bζ − which are kept nonsep- arated in Aξ for ζ < ξ by the inductive hypothesis. Let τ ⊂ σ be given by Lemma 3.2. Since Aξ < c there are less than c many υ ⊂ τ such that {bn : n ∈ υ} and {bn : n ∈ σ \ υ} are separated. So choose τ ξ ⊂ τ such that {bn : n ∈ υ} and {bn : n 6∈ σ \ υ} are not separated, xξ = ψ(τ ξ) and Aξ+1 = Aξhxξi. + and Bζ (2) If ξ ∈ Fα for some α ≤ ξ, then {dξ n : n < ω} is a strictly decreasing sequence of elements of Aα ⊂ Aξ, and we must add an element below it. If there exists some element a ∈ Aξ such that a < dξ n for all n, then we do not need to add anything, and we make just Aξ+1 = Aξ. Otherwise, choose xξ = x ∈ B \ {0} such that x < dξ n for all n. Notice that if y ∈ Aξ and y ≤ x then y = 0. + and Bζ We have to check that Bζ − remain unseparated in Aξ+1 for ζ < ξ. So suppose for contradiction that they were separated by some element z = (b ∧ x) ∨ (c ∧ −x) ∨ d with b, c, d ∈ Aξ pairwise disjoint. We claim that z′ = (c ∨ d) also separate Bζ −, which is a contradiction since z′ ∈ Aξ. On the one hand, if u ∈ Bζ + and Bζ + then u < z = (b ∧ x) ∨ (c ∧ −x) ∨ d < b ∨ c ∨ d. But moreover, u ∧ b < z ∧ b < x and since u ∧ b ∈ Aξ, we get that u ∧ b = 0. This proves that u < z′ = (c ∨ d). On the other hand if v ∈ Bζ −, then v ∧ z = 0. This implies that v ∧ c ∧ −e = 0, so v ∧ c < e and since v ∧ c ∈ Aξ, we get that v ∧ c = 0. It is clear that we also have v ∧ d = 0, so v ∧ z′ = v ∧ (c ∨ d) = 0 as required. (cid:3) Corollary 4.2. There exists a Banach space X ⊂ ℓ∞/c0 of the form X = C(K) such that every injective operator T : X −→ X is surjective. 1When a = {aγ n(α) : n < ω}, we will have Pa = {x ⊂ ω : ψγ α(x) ∈ A} and φa will be the restriction of ψγ α to Pa. 10 ANTONIO AVIL ´ES AND PIOTR KOSZMIDER 5. Invariant subspaces Proposition 5.1. If K = KA is as in Theorem 3.1, then for every operator T : X −→ X there exists a proper nonempty clopen subset L of K such that {f ∈ C(K) : f K\L = 0} is an invariant subspace of T . Proof. By Theorem 3.1, T is a weak multiplier, so T ∗ =g∗ T +S where g∗ : K −→ K is Borel and S is weakly compact.It is enough to find a proper nonempty clopen L such that the set NL of measures whose support is disjoint from L is invariant under T ∗. The set NL is invariant under g∗ T for all clopens L and all Borel functions g, hence it is enough to find L for which NL is invariant under S. We prove that if {Lα : α < ω1} is a disjoint family of nonempty clopen subsets of K, then there exists α such that NLα is invariant under S. If it was not the case, then for every α < ω1 there exists a measure µα ∈ NLα, that we can take with kµαk = 1, such that SµαLα 6= 0. There is an ε > 0 such that Sµα(Lα) > ε for uncountably many α < ω1. This contradicts weak compactness by the Dieudonn-Grothendieck theorem [6, VII.14]. (cid:3) The invariant subspace Y = {f ∈ C(K) : f K\L = 0} is complemented and isomorphic to C(L), the complement being isomorphic to C(K \L). If L a nonempty clopen set of K, then L is still as in Theorem 3.1, so again the Banach space C(L) has the same properties as C(K): C(L) has density c and every injective operator C(L) −→ C(L) is an isomorphism. But note that C(L) is not isomorphic to C(K), or otherwise we could construct an injective operator C(K) −→ C(K) which is not surjective. References [1] S. Argyros, R. Haydon, A hereditarily indecomposable L∞-space that solves the scalar-plus- compact problem. Acta Math. 206 (2011), no. 1, 1-54. [2] S. Argyros, P. Motakis, A reflexive HI space with the hereditary Invariant Subspace Property, http://arxiv.org/abs/1111.3603, preprint 2011. [3] S. Argyros, A. Tolias, Methods in the theory of hereditarily indecomposable Banach spaces. Mem. Amer. Math. Soc. 170 (2004), no. 806. [4] S. Argyros, S. Todorcevic, Ramsey methods in analysis. Advanced Courses in Mathematics. CRM Barcelona. Birkhuser Verlag, Basel, 2005. [5] S. Banach, Th´eorie des op´erations lin´eaires, Monografje Matematyczne, Pa´nstwowe Wydawnictwo Naukowe, 1932. [6] J. Diestel, Sequences and series in Banach spaces, Graduate Texts in Mathematics, 92. Springer-Verlag, New York, 1984. [7] J. Diestel, J.J. Jr. Uhl, Vector measures. Mathematical Surveys, No. 15. American Mathe- matical Society, Providence, R.I., 1977. [8] N. Dunford, J. Schwartz, Linear Operators; Part I, General Theory. Interscience Pub- lishers, INC., New York, Fourth printing, 1967. [9] M. Fabian et al, Functional analysis and infinite-dimensional geometry, CMS Books in Math- ematics/Ouvrages de Math´ematiques de la SMC, 8, Springer-Verlag, New York, 2001. [10] R. Fajardo, An indecomposable Banach space of continuous functions which has small den- sity. Fund. Math. 202 (2009), no. 1, pp. 43 -- 63. [11] S. Goldberg, A. H. Kruse, The existence of compact linear maps between Banach spaces. Proc. Amer. Math. Soc. 13 1962 808 -- 811. [12] T. Gowers, B. Maurey, The unconditional basic sequence problem. J. Amer. Math. Soc. 6 (1993), no. 4, 851 -- 874. [13] T. Gowers, A solution to Banach's hyperplane problem. Bull. London Math. Soc. 26 (1994), no. 6, 523 -- 530. 11 [14] A. Haıly, A. Kaidi, A. Rodr´ıguez Palacios, Algebra descent spectrum of operators, Israel J. Math. 177 (2010), 349-368. [15] R. Haydon, A nonreflexive Grothendieck space that does not contain ℓ∞. Israel J. Math. 40 (1981), no. 1, 65 -- 73. [16] T. Jech, Set theory, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2003, The third millennium edition, revised and expanded. [17] R. Levy, Almost-P -spaces. Canad. J. Math. 29 (1977), no. 2, 284 -- 288. [18] S. Koppelberg, Handbook of Boolean algebras. Vol. 1. Edited by J. Donald Monk and Robert Bonnet. North-Holland Publishing Co., Amsterdam, 1989. [19] P. Koszmider, Banach spaces of continuous functions with few operators. Math. Ann. 330 (2004), no. 1, 151 -- 183. [20] P. Koszmider, A space C(K) where all nontrivial complemented subspaces have big densities. Studia Math. 168 (2005), no. 2, 109 -- 127. [21] P. Koszmider, A survey on Banach spaces C(K) with few operators. Rev. R. Acad. Cienc. Exactas Fs. Nat. Ser. A Math. RACSAM 104 (2010), no. 2, 309 -- 326. [22] P. Koszmider, On large indecomposable Banach spaces, http://arxiv.org/abs/1106.2916, preprint 2011. [23] K. Kunen, Set Theory. An Introduction to Independence Proofs. North-Holland 1980. [24] J. Lindenstrauss, L. Tzafriri, Classical Banach spaces. I. Sequence spaces. Springer-Verlag, Berlin-New York, 1977. [25] G. Plebanek, A construction of a Banach space C(K) with few operators. Topology Appl. 143 (2004), pp. 217 -- 239. [26] H. P. Rosenthal, On relatively disjoint families of measures with some applications to Banach space theory. Studia Math. 37 (1970), 13-36. [27] I. Schlackow, Classes of C(K) spaces with few operators D. Phil. Thesis. Univerity of Oxford, 2009, available at http://people.maths.ox.ac.uk/schlack1/ [28] I. Schlackow, Centripetal operators and Koszmider spaces Topology Appl. 155 (2008), no. 11, pp. 1227 -- 1236 E-mail address: [email protected] Departamento de Matem´aticas, Universidad de Murcia, 30100 Murcia (Spain) E-mail address: [email protected] Institute of Mathematics, Polish Academy of Sciences, ul. ´Sniadeckich 8, 00-956 Warszawa, Poland
1306.5516	1	1306	2013-06-24T06:24:49	Some Generalization of Hadamard's type Inequalities through Differentiability for s-Convex Function and their Applications	[ "math.FA", "math.CA" ]	In this paper, a general form of integral inequalities of Hermite-Hadamard's type through differentiability for s-Convex function in second sense and whose all derivatives are absolutely continuous are established. The generalized integral inequalities contributes some better estimates than some already presented. The inequalities are then applied to numerical integration and some special means.	math.FA	math	SOME GENERALIZATIONS OF HADAMARD'S-TYPE INEQUALITIES THROUGH DIFFERENTIABILITY FOR S-CONVEX FUNCTIONS AND THEIR APPLICATIONS MUHAMMAD MUDDASSAR* AND MUHAMMAD IQBAL BHATTI ABSTRACT. In this paper, a general form of integral inequalities of Hermite - Hadamard's type through differentiability for s-convex function in second sense and whose all deriva- tives are absolutely continuous are established. The generalized integral inequalities con- tributes some better estimates than some already presented. The inequalities are then ap- plied to numerical integration and some special means. . A F h t a m [ 1 v 6 1 5 5 . 6 0 3 1 : v i X r a 1. INTRODUCTION Let f : ∅ 6= I ⊆ R → R be a function defined on the interval I of real numbers. Then f is called convex if f (t x + (1 − t) y) ≤ t f (x) + (1 − t) f (y) for all x, y ∈ I and t ∈ [0, 1]. There are many results associated with convex functions in the area of inequalities, but one of those is the classical Hermite Hadamard inequality: f(cid:18) a + b 2 (cid:19) ≤ 1 b − aZ b a f (x)dx ≤ f (a) + f (b) 2 . (1.1) This inequality gives us an estimate, from below and from above, of the average value of the convex function f : [a, b] → R for a, b ∈ I, with a < b. H. Hudzik and L. Maligranda in [6], define the class of functions which are s−convex in the second sense. This class is defined as follows: A function f : R+ → R is said to be s−convex in the second sense if for any two non- negative real numbers x, y, for λ ∈ [0, 1] and adjust s ∈ (0, 1], we have the following inequality f (λ x + (1 − λ) y) ≤ λs f (x) + (1 − λ)s f (y) (1.2) Date: 18 March, 2012. 2000 Mathematics Subject Classification. 26D15, 26A51, 39A10. Key words and phrases. Hermite-Hadamard type inequality, s-Convex function, Beta function, Holder's In- tegral Inequality, Quadrature Rules, Special Means. 1 2 Muhammad Muddassar and M. I. Bhatti It is seen that from (1.2), for s = 1, s-convex function is convex. H. Hudzik et al. in the same paper [6] talked about some results associating with s−convex functions in second sense. We find some more new results about Hadamard's inequality for s−convex func- tions in [1, 2, 9, 10]. Although it is seen that many important inequalities connecting with 1-convex (convex) functions given in [5], but one of them is (1.1). S. S. Dragomir et al. gave a variant of Hermite-Hadamard's inequality for s−convex func- tions in second sense in [4]. Theorem 1. Let a function f : R+ → R+ be s−convex in the second sense, where s ∈ (0, 1], and a, b ∈ R+, with a < b. If f ∈ L1[a, b], then the following inequality holds (1.3) f (a) + f (b) f (x) dx ≤ s + 1 2s−1f(cid:18) a + b 2 (cid:19) ≤ 1 b − aZ b a In second inequality in (1.3), the constant k = 1 s+1 is the most suitable. In [8], where B. Jagers gave both the right and left bound for the constant c(s) in the inequality c(s) f(cid:18) a + b 2 (cid:19) ≤ and improved (1.3). He proved that 1 b − a Z b a f (x) dx 2s+1 − 1 s + 2 ≤ c(s) ≤ 2 s s+1 s−1 s+1 (cid:18) 2s − 1 s (cid:19) 2s+1 − 2s−1 − 1 s + 1 ≤ In [3, 5] S. S. Dragomir et al. discussed inequalities for differentiable and twice differ- entiable functions connecting with the H-H Inequality on the basis of the Lemma 2 and Lemma 3. S. S. Dragomir et al. in [5, pp. 65] generalized the lemma for n-time differentiable map- ping and he states in this way: Lemma 1. Let f : I ⊂ R → R be n times differentiable function on I o with f (n) ∈ L1[a, b], then (−1)nZ b a f (x)dx = n (−1)n−m+1(cid:20) (t − a)m − (t − b)m Xm=1 n!"Z t (x − a)nf (n)(x)dx +Z b (cid:21) f (m−1)(t) (x − b)nf (n)(x)dx# (1.4) m! + 1 a t Here we explore Lemma3 and 2 by different approach and then make their use for investigation for some more results which generalize the results explored by S. Hussain et. al. in [7]. Using n = 1 and t = a+b 2 in (1.4); we get lemma 2. Generalization of Hadamard-Type Inequality for s-Convex functions ... 3 Lemma 2. Let f : I ⊆ R → R be differentiable function on I ◦, a, b ∈ I ◦ with a < b and f ′ ∈ L1[a, b], then f(cid:18) a + b 2 (cid:19) − 1 (1 − t) (cid:20)f ′(cid:18)ta + (1 − t) b − aZ b 2 (cid:19) f (x) dx = a + b a 0 4 Z 1 (b − a) − f ′(cid:18)tb + (1 − t) a + b 2 (cid:19)(cid:21) dt For n = 2 we get from identity (1.4); 1 b − aZ b a f (x)dx = f (t) + f ′(t) 2 (b + a) − 2t 2(b − a)"Z t 1 a + (x − a)2f ′′(x)dx +Z b t (x − a)2f ′′dx# (1.5) For t = a the above equation (1.5) becomes; (b − a) f (x)dx = f (a) + 2 t = b the above equation (1.5) becomes; (b − a) f (x)dx = f (b) + 1 a 1 b − aZ b b − aZ b a By adding (1.6) and (1.7), we have a 1 2(b − a)Z b 2(b − a)Z b 1 a f ′(b) + 2 (x − a)2f ′′(x)dx f ′(a) + (x − a)2f ′′dx (1.6) (1.7) (1.8) 1 b − aZ b − a f (a) + f (b) 2 − f (x)dx = (b − a) 4 [f ′(b) − f ′(a)] 1 4(b − a)Z b a (cid:2)(x − a)2 + (x − b)2(cid:3) f ′′(x)dx Now for x = ta + (1 − t)b in (1.8). We get lemma 3 Lemma 3. Let f : I ⊆ R → R be twice differentiable function on I ◦ with f ′′ ∈ L1[a, b], then 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 In [7] S. Hussain et al. use the above lemmas and give some new improvements in the right classical Hermite Hadamard inequality. We give here definition of Beta function of Euler type which will be helpful in our next discussion, which is for x, y > 0 defined as β(x, y) = Γ(x).Γ(y) Γ(x + y) =Z 1 0 tx−1 (1 − t)y−1 dt In this paper, after this Introduction, in section 2 we generalize the results for some s−Hermite Hadamard type inequalities for n-differentiable functions discussed in [7]. In section 3 we give applications of the results from section 2 for quadrature rules and in section 4 we will discuss application for some special means. 4 Muhammad Muddassar and M. I. Bhatti 2. INEQUALITIES FOR n-DIFFERENTIABLE FUNCTIONS Theorem 2. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n) is s-convex on [a, b] for some fixed s ∈ (0, 1], then for every λ ∈ [a, b], we have f (x)dx + (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)nZ b ≤ a Proof. From Lemma 1, we have f (x)dx + (−1)nZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n n 1 m! Xm=1 (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n!hβ(s + 1, n + 1)(cid:16)(λ − a)n+1f (n)(a) + (b − λ)n+1f (n)(b)(cid:17) +β(1, n + s + 1)(cid:0)(λ − a)n+1 + (b − λ)n+1(cid:1)f (n)(λ)i . (cid:21) f (m−1)(t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dx +Z b (x − b)nf (n)(x)dx(cid:12)(cid:12)(cid:12) (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)nf (n)(tb + (1 − t)λ)dt(cid:21) (−1)n−m+2(cid:20) (t − a)m − (t − b)m Xi=1 n!"(cid:12)(cid:12)(cid:12)(cid:12) Z t (x − a)nf (n)(x)(cid:12)(cid:12)(cid:12)(cid:12) n!(cid:20)(λ − a)n+1Z 1 +(b − λ)n+1Z 1 # (cid:12)(cid:12)(cid:12) ≤ ≤ m! 1 1 a 0 0 t (2.9) (2.10) Since f (n) is s- convex on [a, b] for t ∈ [0, 1], so f (n)(ta + (1 − t)λ) ≤ tsf (n)(a) + (1 − t)sf (n)(λ) Inequality (2.10) becomes (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a (−1)nZ b ≤ f (x)dx + n Xi=1 m! (−1)n−m+2(cid:20) (t − a)m − (t − b)m (cid:21) f (m−1)(t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (1 − t)nhtsf (n)(a) + (1 − t)sf (n)(λ)i dt (1 − t)nhtsf (n)(b) + (1 − t)sf (n)(λ)i dt(cid:21) (2.11) 0 1 n!(cid:20)(λ − a)n+1Z 1 + (b − λ)n+1Z 1 0 Where, 0 Z 1 Z 1 0 ts(1 − t)ndt = β(s + 1, n + 1) (1 − t)n+sdt = β(1, n + s + 1) (2.12) (2.13) By combining (2.11), (2.12) and (2.13) we get (2.9). Generalization of Hadamard-Type Inequality for s-Convex functions ... 5 a n n! m! f (x)dx + (n + 1)− 1 Theorem 3. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-convex on [a, b] for some fixed s ∈ (0, 1], and q ≥ 1. Then for every λ ∈ [a, b], (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)nZ b Xm=1 (cid:20)(λ − a)n+1nβ(s + 1, n + 1)f (n)(a)q + β(1, n + s + 1)f (n)(λ)qo (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ +(b − λ)n+1nβ(s + 1, n + 1)f (n)(b)q + β(1, n + s + 1)f (n)(λ)qo (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m (−1)nZ b Proof. From Lemma1 we have (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)nf (n)(tb + (1 − t)λ)dt(cid:21) (2.15) Xm=1 n!(cid:20)(λ − a)n+1Z 1 +(b − λ)n+1Z 1 f (x)dx + q(cid:21) ≤ m! 1 n 0 0 p a 1 q 1 (2.14) From here, let's start with first integral of right side of (2.15) and using Holder's Integral Inequality Z 1 0 (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)n(1− 1 ≤(cid:18)Z 1 =Z 1 0 0 q )(1 − t)n( 1 q )f (n)(ta + (1 − t)λ)dt p (cid:18)Z 1 (1 − t)ndt(cid:19) 0 1 (1 − t)n(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) 1 q q dt(cid:19) n+1 and above inequality becomes where p = q q−1 . Now hereR 1 0 (1 − t)ndt = 1 Z 1 0 (1 − t)nf (n)(ta + (1 − t)λ)dt p (cid:18)Z 1 ≤(cid:18) 1 n + 1(cid:19) 0 1 Since f (n)q is s- convex on [a, b] for t ∈ [0, 1], so (1 − t)n(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) 1 q q dt(cid:19) (2.16) f (n)(ta + (1 − t)λ) ≤ tsf (n)(a)q + (1 − t)sf (n)(λ)q 6 Muhammad Muddassar and M. I. Bhatti now using equation (2.12) and (2.13) in (2.16), we have Z 1 0 dt (1 − t)n(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) n + 1(cid:19) ≤(cid:18) 1 p hβ(s + 1, n + 1)(cid:12)(cid:12)(cid:12) 1 f (n)(a)(cid:12)(cid:12)(cid:12) Similarly, we have 0 Z 1 (1 − t)n(cid:12)(cid:12)(cid:12) n + 1(cid:19) ≤(cid:18) 1 dt 1 f (n)(tb + (1 − t)λ)(cid:12)(cid:12)(cid:12) p hβ(s + 1, n + 1)(cid:12)(cid:12)(cid:12) f (n)(b)(cid:12)(cid:12)(cid:12) + β(1, n + s + 1)(cid:12)(cid:12)(cid:12) f (n)(λ)(cid:12)(cid:12)(cid:12)i (2.17) + β(1, n + s + 1)(cid:12)(cid:12)(cid:12) f (n)(λ)(cid:12)(cid:12)(cid:12)i (2.18) Using (2.17) and (2.18) in (2.15), we get (2.14). Remark 1. For n = 1, λ = a+b 2 in (2.14), we get theorem 4 of [7]. Theorem 4. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is concave on [a, b] for conjugate numbers p, q where q ≥ 1. Then for every λ ∈ [a, b], (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)nZ b Xm=1 f (x)dx + m! n a p (np + 1) (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m (cid:20)(λ − a)n+1(cid:12)(cid:12)(cid:12)(cid:12) + (b − λ)n+1(cid:12)(cid:12)(cid:12)(cid:12) ≤ n! −1 (2.19) (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + λ 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) b + λ (cid:21) (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)nf (n)(tb + (1 − t)λ)dt(cid:21) (2.20) n 1 Xm=1 n!(cid:20)(λ − a)n+1Z 1 +(b − λ)n+1Z 1 0 0 ≤ f (x)dx + (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m m! Proof. From Lemma 1, we have (−1)nZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Generalization of Hadamard-Type Inequality for s-Convex functions ... 7 From here, let's start with first integral of right side of (2.20) and using Holder's Integral Inequality Z 1 0 1 1 0 (1 − t)nf (n)(ta + (1 − t)λ)dt p (cid:18)Z 1 ≤(cid:18)Z 1 (1 − t)npdt(cid:19) 0 (cid:12)(cid:12)(cid:12) p (cid:18)Z 1 =(cid:18) 1 np + 1(cid:19) 0 (cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) Now using the concavity of (cid:12)(cid:12)f (n)(cid:12)(cid:12) Z 1 0 (cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) R 1 0 q q f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) dt(cid:19) 1 q q equality on the second integral in the left side of the inequality (2.21), we have 0 (ta + (1 − t)λ) dt 0 t0dt So inequality (2.21) can be written in this way f (n) R 1 t0dt(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dt ≤(cid:18)Z 1 =(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + λ p (cid:12)(cid:12)(cid:12)(cid:12) np + 1(cid:19) (1 − t)nf (n)(ta + (1 − t)λ)dt ≤(cid:18) 1 p (cid:12)(cid:12)(cid:12)(cid:12) (1 − t)nf (n)(tb + (1 − t)λ)dt ≤(cid:18) 1 np + 1(cid:19) 1 1 Z 1 0 Z 1 0 Using (2.23) and (2.24) in (2.20), we get (2.19). Similarly we have q on [a, b]. And by applying the Jensen's Integral In- 1 q q dt(cid:19) (2.21) q !(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (2.22) (2.23) (2.24) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + λ 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) b + λ Remark 2. For n = 1 and λ = a+b 2 in (2.19) , we get theorem 5 of [7]. n Theorem 5. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-convex on [a, b] for some fixed s ∈ (0, 1] and for conjugate numbers p, q where q ≥ 1 with p = q q−1 , then for every λ ∈ [a, b], Xm=1 (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m (cid:18) 1 s + 1(cid:19) + (b − λ)n+1(cid:16)(cid:12)(cid:12)(cid:12) f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:21) f (m−1)(λ) + (−1)nZ b q(cid:20)(λ − a)n+1(cid:16)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) f (n)(λ)(cid:12)(cid:12)(cid:12) q(cid:21) q(cid:17) f (n)(b)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) f (n)(λ)(cid:12)(cid:12)(cid:12) 1 q q(cid:17) (2.25) (np + 1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ m! n! −1 a q q p 1 1 8 Muhammad Muddassar and M. I. Bhatti Proof. From Lemma (1) we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a (−1)nZ b ≤ f (x)dx + n 1 m! (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m Xm=1 n!(Z t a (cid:12)(cid:12)(cid:12) (x − a)nf (n)(x)(cid:12)(cid:12)(cid:12) n!(cid:20)(λ − a)n+1Z 1 +(b − λ)n+1Z 1 (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dx) dx +Z b (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)nf (n)(tb + (1 − t)λ)dt(cid:21) (2.26) (cid:12)(cid:12)(cid:12) (x − t)nf (n)(x)(cid:12)(cid:12)(cid:12) = 1 0 0 t From here, let's start with first integral of right side of (2.26) and using Holder's Integral Inequality Z 1 0 (1 − t)nf (n)(ta + (1 − t)λ)dt (1 − t)npdt(cid:19) ≤(cid:18)Z 1 0 np+1 inequality (2.27), we have 0 (1 − t)npdt = 1 HereR 1 Now using the s-convexity of(cid:12)(cid:12)f (n)(cid:12)(cid:12) Z 1 dt ≤Z 1 0 (cid:12)(cid:12)(cid:12) ≤ f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) q The inequality in (2.27) becomes Z 1 0 (1 − t)nf (n)(ta + (1 − t)λ)dt ≤(cid:18) 1 np + 1(cid:19) In similar way we can prove 1 p (cid:18)Z 1 0 (cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) 1 q q dt(cid:19) (2.27) q on [a, b] in the second integral on the left side of the 0 (cid:16)ts(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) s + 1(cid:16)(cid:12)(cid:12)(cid:12) 1 q q + (1 − t)s(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) p (cid:12)(cid:12)f (n)(a)(cid:12)(cid:12) q(cid:17) dt f (n)(λ)(cid:12)(cid:12)(cid:12) q(cid:17) f (n)(λ)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)f (n)(λ)(cid:12)(cid:12) s + 1 q q 1 ! (2.28) 1 q (2.29) Z 1 0 (1 − t)nf (n)(tb + (1 − t)λ)dt ≤(cid:18) 1 np + 1(cid:19) 1 p (cid:12)(cid:12)f (n)(b)(cid:12)(cid:12) Using (2.29) and (2.30) in (2.26), we get (2.25). q +(cid:12)(cid:12)f (n)(λ)(cid:12)(cid:12) s + 1 q 1 q ! (2.30) Remark 3. For n = 1 and λ = a+b 2 in (2.25), we get theorem 6 of [7]. Generalization of Hadamard-Type Inequality for s-Convex functions ... 9 Corollary 1. (Midpoint's type Inequality) In Theorem 5, if we select λ = a+b 2 we obtain for n = 1, f(cid:18) a + b 2 (cid:19) − (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 b − aZ b a ≤ f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (b − a)2 4(p + 1) 1 q q(cid:19) 1 1 q"(cid:18)f ′(a)q +(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) s + 1(cid:19) f ′(cid:18) a + b p(cid:18) 1 q# q(cid:19) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:18)f ′(b)q +(cid:12)(cid:12)(cid:12)(cid:12) f ′(cid:18) a + b s + 1(cid:19) p (cid:18) 1 (b − a)2 2(p + 1) 1 q 1 1 (f ′(a) + f ′(b)) (2.31) ≤ Proof. Proof is very similar to the above theorem and at the end, the second inequality m=1 (αm)r + (βm)r for is found using the following inequalityPn 0 ≤ r < 1, where α1, α2, α3, ..., αn, β1, β2, β3, ..., βn ≥ 0. Theorem 6. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-convex on [a, b] for some fixed s ∈ (0, 1] and for conjugate numbers p, q where q ≥ 1 with p = q m=1 (αm + βm)r ≤Pn q−1 , then (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n q m! (−1)n−m+2 (b − a)m Xm=1 (b − a)n+1 (cid:18) 1 n + 1(cid:19) ≤ +nβ(s + 1, n + 1)(cid:12)(cid:12)(cid:12) Proof. Let's start with the lemma 1 n! f (x)dx + (−1)nZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ For λ = a in (2.33), we have (−1)nZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 q q a n f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) hf (m−1)(b) − (−1)mf (m−1)(a)i + 2(−1)nZ b q−1(cid:20)nβ(s + 1, n + 1)(cid:12)(cid:12)(cid:12) qo f (n)(a)(cid:12)(cid:12)(cid:12) + β(1, n + s + 1)(cid:12)(cid:12)(cid:12) f (n)(b)(cid:12)(cid:12)(cid:12) q(cid:21) qo f (n)(b)(cid:12)(cid:12)(cid:12) + β(1, n + s + 1)(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) (cid:21) f (m−1)(λ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m Xm=1 n!(Z λ dx +Z b a (cid:12)(cid:12)(cid:12) (x − a)nf (n)(x)(cid:12)(cid:12)(cid:12) λ (cid:12)(cid:12)(cid:12) (x − t)nf (n)(x)(cid:12)(cid:12)(cid:12) f (m−1)(a)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dx) (x − b)nf (n)(x)(cid:12)(cid:12)(cid:12) (−1)n−m+2 (−1)m+1(b − a)m n!(Z b a (cid:12)(cid:12)(cid:12) Xm=1 ≤ m! m! 1 1 n f (x)dx + dx) (2.33) 1 q (2.32) (2.34) 10 Muhammad Muddassar and M. I. Bhatti For λ = b in (2.33), we have n f (x)dx + (−1)nZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ Now combining (2.34) and (2.35), we have 2(−1)nZ b a f (x)dx + (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n Xm=1 we can write (2.36) in the form 2(−1)nZ b a f (x)dx + (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n Xm=1 ≤ 1 1 m! m! m! ≤ (2.35) dx (2.36) Xm=1 dx) (−1)n−m+2 (b − a)m n!Z b a {(x − a)n + (x − b)n}(cid:12)(cid:12)(cid:12) f (m−1)(b)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)n−m+2 (b − a)m n!(Z b a (cid:12)(cid:12)(cid:12) (x − a)nf (n)(x)(cid:12)(cid:12)(cid:12) hf (m−1)(b) − (−1)mf (m−1)(a)i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(x)(cid:12)(cid:12)(cid:12) hf (m−1)(b) − (−1)mf (m−1)(a)i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)n−m+2 (b − a)m (cid:26)Z 1 (b − a)n+1 (1 − t)n(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)b)(cid:12)(cid:12)(cid:12) +Z 1 dt(cid:27) (2.37) tn(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)b)(cid:12)(cid:12)(cid:12) n + 1(cid:19) p (cid:16)β(s + 1, n + 1)(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) q(cid:17) + β(1, n + s + 1)(cid:12)(cid:12)(cid:12) f (n)(b)(cid:12)(cid:12)(cid:12) p (cid:16)β(n + s + 1, 1)(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) q(cid:17) + β(n + 1, s + 1)(cid:12)(cid:12)(cid:12) f (n)(b)(cid:12)(cid:12)(cid:12) dt ≤(cid:18) 1 dt ≤(cid:18) 1 n + 1(cid:19) (2.39) (2.38) 1 q q n! dt 1 q 0 1 1 q 0 Now here we can easily prove the following result Z 1 0 (1 − t)n(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)b)(cid:12)(cid:12)(cid:12) And furthermore, we have Z 1 0 tn(cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)b)(cid:12)(cid:12)(cid:12) Using (2.38) and (2.39) in (2.37), we get (2.32). Remark 4. For n = 2 in (2.32), we get theorem 8 of [7]. Corollary 2. (Trapezoidal's type Inequality) In Theorem 6, if we select n = 2 for s = 1, we obtain f (a) + f (b) 2 − (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 b − aZ b a f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1 (b − a)2 2(6) (b − a)2 2(6) p (cid:18) 1 12(cid:19) p (cid:18) 1 12(cid:19) 1 q 1 ≤ 1 q (cid:2)f ′′(a)q + f ′′(a)q(cid:3) [f ′′(a) + f ′′(b)] 1 q (2.40) Generalization of Hadamard-Type Inequality for s-Convex functions ... 11 Proof. Proof is very similar as we did in corollary 1. Theorem 7. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-concave on [a, b] for some fixed s ∈ (0, 1] and for conjugate numbers p, q where q ≥ 1 with p = q q−1 , then for every λ ∈ [a, b], Xm=1 (−1)n−m+2(cid:20) (λ − a)m − (λ − b)m m! n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (np + 1) n! ≤ −1 p .2 s−1 a f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:21) f (m−1)(λ) + (−1)nZ b q (cid:20)(λ − a)n+1(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + λ 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) + (b − λ)n+1(cid:12)(cid:12)(cid:12)(cid:12) (cid:21) f (n)(cid:18) b + λ (2.41) (2.42) (2.43) Proof. To prove, we carry on in similar way as we did in theorem 5. q we obtain By s-concavity of(cid:12)(cid:12)f (n)(cid:12)(cid:12) Z 1 0 (cid:12)(cid:12)(cid:12) f (n)(ta + (1 − t)λ)(cid:12)(cid:12)(cid:12) Z 1 0 (cid:12)(cid:12)(cid:12) f (n)(tb + (1 − t)λ)(cid:12)(cid:12)(cid:12) In an analogous manner q q From (2.26), (2.42) and (2.43) instantly give (2.41). q q dt ≤ 2s−1(cid:12)(cid:12)(cid:12)(cid:12) dt ≤ 2s−1(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + λ 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) b + λ Remark 5. For n = 1 and λ = a+b 2 in (2.41), we get theorem 7 of [7]. Xm=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Theorem 8. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is concave on [a, b] for conjugate numbers p, q where q ≥ 1 with p = q (−1)n−m+2 (b − a)m q−1 , then 2m! n ≤ 1 a f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) hf (m−1)(b) − (−1)mf (m−1)(a)i + (−1)nZ b (b − a)n+1 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + b 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + b n! (b − a)n+1 (cid:18) 1 np + 1(cid:19) (β(np + 1, 1)) p (cid:12)(cid:12)(cid:12)(cid:12) n! = 1 (2.44) Theorem 9. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-convex on [a, b] for conjugate (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 12 Muhammad Muddassar and M. I. Bhatti numbers p, q where q ≥ 1 with p = q q−1 , then n Xm=1 ≤ 2m! (−1)n−m+2 (b − a)m (b − a)n+1 (cid:18) 1 np + 1(cid:19) n! (b − a)n+1 = n! a 1 p (s + 1)− 1 hf (m−1)(b) − (−1)mf (m−1)(a)i + (−1)nZ b q(cid:17) f (n)(b)(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) q (cid:16)(cid:12)(cid:12)(cid:12) f (n)(a)(cid:12)(cid:12)(cid:12) f (n)(b)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) q (cid:16)(cid:12)(cid:12)(cid:12) p (s + 1)− 1 1 q q q 1 f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q(cid:17) 1 q Theorem 10. Let f : I ⊂ [0,∞) → R be n-times differentiable function on I ◦ such that f (n) ∈ L1[a, b], where a, b ∈ I, a < b. If f (n)q is s-concave on [a, b] for conjugate numbers p, q where q ≥ 1 with p = q q−1 , then (β(np + 1, 1)) (2.45) (−1)n−m+2 (b − a)m 2m! n Xm=1 ≤ a (β(np + 1, 1)) f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) hf (m−1)(b) − (−1)mf (m−1)(a)i + (−1)nZ b (b − a)n+1 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18) a + b 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (n)(cid:18)a + b n! (b − a)n+1 (cid:18) 1 np + 1(cid:19) (cid:12)(cid:12)(cid:12)(cid:12) p 2 n! = s−1 s−1 2 1 p q q 1 (2.46) 3. APPLICATIONS TO COMPOSITE QUADRATURE RULES Let K be the partition {a = x0 < x1 < ... < xn−1 < xn = b} of the interval [a, b] and consider the quadrature formula f (x)dx = S(f, K) + R(f, K) (3.47) where Z b a S(f, K) = n−1 Xm=0 f(cid:18) xm + xm+1 2 (cid:19) (xm+1 − xm) for the midpoint version and R(f, K) denotes the related approximation error. S(f, K) = n−1 Xm=0 f (xm) + f (xm+1) 2 (xm+1 − xm) for the trapezoidal version and R(f, K) denotes the related approximation error. Proposition 1. Let f : I ⊆ R → R be a differentiable mapping on I o such that f ′ ∈ L1[a, b], where a, b ∈ I with a < b and f ′ is s-convex on [a, b], then R(f, K)≤ 1 (p + 1) p (cid:18) 1 s + 1(cid:19) 1 1 q n−1 Xm=0 (xm+1 − xm)3 2 [f ′(xm)+ f ′(xm+1)] (3.48) Generalization of Hadamard-Type Inequality for s-Convex functions ... 13 Proof. By applying subdivisions [xm, xm+1] of the division k for m = 0, 1, 2, ..., n − 1 on Corollary 1, we have 1 xm+1 − xm Z xm+1 xm (cid:12)(cid:12)(cid:12)(cid:12) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f (x)dx − f(cid:18) xm+1 + xm (xm+1 − xm)2 p (cid:18) 1 s + 1(cid:19) 2(p + 1) ≤ 2 1 1 q (f ′(xm+1) + f ′(xm)) (3.49) a 2 xm n−1 n−1 Taking sum over m from 0 to n − 1 and taking into account that f ′q is s-convex, we get (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (x)dx −S(f, K)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Z b (cid:19)(xm+1 − xm)(cid:27)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (x)dx−f(cid:18)xm+1 + xm (cid:19)(cid:27)(cid:12)(cid:12)(cid:12)(cid:12) f (x)dx−(xm+1 − xm) f(cid:18)xm+1 + xm =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xm=0(cid:26)Z xm+1 (cid:26)Z xm+1 Xm=0(cid:12)(cid:12)(cid:12)(cid:12) (xm+1 − xm)Z xm+1 (xm+1 − xm)(cid:12)(cid:12)(cid:12)(cid:12) (cid:26) Xm=0 (cid:19)(cid:27)(cid:12)(cid:12)(cid:12)(cid:12) −f(cid:18)xm+1 + xm By combining (3.49) and (3.50), we get (3.48). Which completes the proof. f (x)dx ≤ n−1 = xm xm 1 2 2 (3.50) Proposition 2. Let f : I ⊆ R → R be a twice differentiable mapping on I o such that f ′′ ∈ L1[a, b], where a, b ∈ I with a < b and f ′′ is s-convex on [a, b], then R(f, K) ≤ (6) 1 (6) ≤ 1 1 1 (s + 2)(s + 3)(cid:19) p (cid:18) p (cid:18) (s + 2)(s + 3)(cid:19) 1 1 1 1 2 q n−1 (xm+1 − xm)3 Xm=0 (xm+1 − xm)3 Xm=0 2 q n−1 1 q (cid:2)f ′′(xm)q + f ′′(xm+1)q(cid:3) [f ′′(xm) + f ′′(xm+1)] (3.51) Proof. Proof is very similar as that of Proposition 1 by using corollary 2. 4. APPLICATIONS TO SOME SPECIAL MEANS Let us recall the following means for any two positive numbers a and b. (1) The Arithmetic mean (2) The Harmonic mean A ≡ A(a, b) = a + b 2 H ≡ H(a, b) = 2ab a + b 14 Muhammad Muddassar and M. I. Bhatti 1 p , if a = b; if a 6= b. a, −ap+1 (3) The p−Logarithmic mean (p+1)(b−a)i h bp+1 (4) The Identric mean a, Lp ≡ Lp(a, b) =  I ≡ I(a, b) = aa(cid:17) e(cid:16) bb  L ≡ L(a, b) =  (5) The Logarithmic mean b−a ln b − ln a , a, 1 1 b−a , if a = b; if a 6= b. if a = b; if a 6= b. The following inequality is well known in the literature in [10]: H ≤ G ≤ L ≤ I ≤ A It is also known that Lp is monotonically increasing over p ∈ R, denoting L0 = I and L−1 = L. Now Here we find some new applications for special means of real numbers by using the results of Section 2. Proposition 3. Let p > 1, 0 < a < b and q = p A(a, b) − L(a, b) ≤ p−1 . Then one has the inequality. (b − a)2 A (a,b) 3 (4.52) Proof. By Theorem 6 applied for the mapping f (x) = ex for s = 1 we have the above inequality (4.52). A result which is connected with Geometric, Identric and Harmonic mean is the following one: Proposition 4. Let p > 1, 0 < a < b and q = p p−1 , then ≤ exp"− (b − a)2 2 p + 1(cid:19)2 (cid:18) 2 H−1 (a, b)# (cid:12)(cid:12)(cid:12)(cid:12) G(a, b) I(a, b) (cid:12)(cid:12)(cid:12)(cid:12) Proof. Follows by Theorem 5, setting f (x) = − ln(x) for n = 1 and s = 1. More results which are connected with p−Logarithmic mean Lp(a, b) is the following one: Proposition 5. Let p > 1, 0 < a < b and q = p p−1 , then (cid:12)(cid:12)(cid:12) A 1 2 (a, b) − L2 p(a, b)(cid:12)(cid:12)(cid:12) ≤ 1 q (b − a)2 2(p + 1) 2(cid:19) p (cid:18) 1 1 1 2 , b H−1(cid:16)a 1 2(cid:17) Proof. Follows by of Corollary 1 of Theorem 5, setting f (x) = √x, x ≥ 0 for n = 1 and s = 1. Generalization of Hadamard-Type Inequality for s-Convex functions ... 15 Proposition 6. Let p > 1, 0 < a < b and q = p p−1 , then n [(1 − a), (1 − b)] 1 A [(1 − a)n, (1 − b)n] − Ln q (cid:18) n(n − 1) q (cid:18) n(n − 1) (b − a)2 12 (b − a)2 12 ≤ q−1 ≤ q (cid:16)A1/qh1 − aq(n−1),1 − bq(n−1)i(cid:17) q (cid:16)Ah1 − a(n−1),1 − b(n−1)i(cid:17) Proof. Follows by Theorem 6, setting f (x) = (1 − x)n, n ≥ 2 and n ∈ Z. (s + 2)(s + 3)(cid:19) (s + 2)(s + 3)(cid:19) q−1 1 5. CONCLUSION Here we can further find some new relations in the same way as above associating with some special means by taking some other convex functions. For example choosing different convex functions like f (x) = − ln x, f (x) = 1 x and f (x) = − ln(1 − x) for different values of s in s-convexity(concavity), we get new relations relating to to some special means. REFERENCES [1] M. Alomari and M. Darus, Hadamard-type inequalities for s−convex functions, Inter. Math. Forum, 3(40) (2008) 1965-1970. [2] M. Alomari and M. Darus, On Co-ordinated s−convex functions, Inter. Math. Forum, 3(40) (2008) 1977- 1989. [3] S. S. Dragomir, On some inequalities for differentiable convex functions and applications, (submitted) [4] S. S. Dragomir, S. Fitzpatrick, The Hadamard's inequality for s−convex functions in the second sense, Demonstratio Math., 32 (4) (1999) 687-696. [5] S. S. Dragomir, C. E. M. Pierce, Selected Topics on Hermite-Hadamard Inequalities and Applications. RGMIA, Monographs, Victoria University 2000. (online: http://ajmaa.org/RGMIA/monographs.php/). [6] H. Hudzik, L. Maligranda, Some remarks on s−convex functions, Aequationes math., 48 (1994) 100-111. [7] S. Hussain, M. I. Bhatti and M. Iqbal, Hadamard-Type inequalities for s-Convex Functions I, Punjab Uni- versity Jouranl of Mathematics (ISSN 1016-2526) Vol. 41 (2009) pp51-60. [8] B. Jagers, On a Hadamard-type inequality for s−convex functions. http://wwwhome.cs.utwente.nl/∼jagersaa/alphaframes/Alpha.pdf. [9] U. S. Kirmaci, M. Klarici´c Bakula, M. E. Ozdemir, J. Pecari´c, Hadamard-type inequalities for s-convex functions, Appl. Math.Comput., 193(1), (2007) 26-35. doi:10.1016/j.amc.2007.03.030. [10] J. Pecari´c, F. Proschan, Y. L. Tong, Convex functions,partial orderings and statistical applications, Aca- demic Press, Inc., New York, 1992, p. 137. E-mail address: [email protected] E-mail address: [email protected] 16 Muhammad Muddassar and M. I. Bhatti DEPARTMENT OF MATHEMATICS, UNIVERSITY OF ENGINEERING AND TECHNOLOGY, LAHORE - PAK- ISTAN
1910.07243	1	1910	2019-10-16T09:38:17	Modular Riesz bases versus Riesz bases in Hilbert $C^*$-Modules	[ "math.FA" ]	In this paper, we give new characterizations of modular Riesz bases in Hilbert $C^*$-modules. We prove that modular Riesz bases share many properties with Riesz bases in Hilbert spaces. Moreover we show that there are also important differences; for example, there exist exact frames that are not modular Riesz bases.	math.FA	math	Modular Riesz bases versus Riesz bases in Hilbert C ∗-Modules Marzieh Hasannasab October 17, 2019 Abstract In this paper we give new characterizations of modular Riesz bases in Hilbert C ∗-modules. We prove that modular Riesz bases share many properties with Riesz bases in Hilbert spaces. Moreover we show that there are also important differences; for example, there exist exact frames that are not modular Riesz bases. 1 Introduction The theory of frames in Hilbert C ∗-modules was introduced by Frank and Larson [6]; more recent works include [3, 5, 7, 8, 11, 14]. Hilbert C ∗-modules are generalizations of Hilbert spaces in which the inner product takes values in a C ∗-algebra. Kasparov's Stabilization Theorem [9] shows that every finitely or countably generated Hilbert C ∗-module has a Parseval frame. The differences between a Hilbert space and a Hilbert C ∗-module have a significant impact on frame theory in these spaces too. In this paper, we focus on Riesz bases in Hilbert C ∗-modules. In the literature, two attempts to define Riesz bases in Hilbert C ∗-modules have been given. In [6], a Riesz basis in a Hilbert C ∗- module is defined as an ω-independent frame; in [11], modular Riesz bases are defined by a direct generalization of one of the equivalent definitions of Riesz bases in Hilbert spaces (see the end of the section for the exact definition). We show that modular Riesz bases in Hilbert C ∗-modules and Riesz bases in Hilbert spaces share many key features; indeed, we prove that some of the well-known characterization theorems for Riesz bases in Hilbert spaces have an analog in Hilbert C ∗-modules. Furthermore, characterizations of modular Riesz bases with respect to the canonical dual frame and ω-independent frames are given. As the last result we show that even though modular Riesz bases 1 behave similar to Riesz bases in Hilbert spaces, there are some important differences due to the special structure of the Hilbert C ∗-modules. Indeed we show that Riesz bases and in particular modular Riesz bases in Hilbert C ∗- modules are exact frames but there exist exact frames in Hilbert C ∗-modules that are not modular Riesz bases. The paper is organized as follows. In the rest of the introduction we will recall some basic definitions concerning Hilbert C ∗-modules, their frames and the two definitions of Riesz bases in such spaces. The new results appear in Section 2. Throughout this paper, A denotes a unital C ∗-algebra with identity 1A. An element a ∈ A is positive if a = b∗b for some b ∈ A. Considering a vector space H, assume that there exists an A-valued inner product h·, ·i : H × H → A with the following properties: (i) hx, xi ≥ 0 for every x ∈ H and hx, xi = 0 if and only if x = 0, (ii) hx, yi = hy, xi∗ for every x, y ∈ H, (iii) hax + y, zi = ahx, zi + hy, zi for every a ∈ A and x, y, z ∈ H. Under the conditions stated above, H is called a Hilbert A-module (or Hilbert C ∗-module in case we want to avoid the reference to the name of the underlying C ∗-algebra) if it is complete with respect to the norm kxk := khx, xik 2 . A Hilbert A-module H is countably generated if there exists a countable subset {xj : j ∈ N} of H such that the set of all its finite A-linear combinations is dense in H; in that case the set {xj : j ∈ N} is called a set of generators. 1 Let H and K be two Hilbert A-modules over a C ∗-algebra A. A linear operator T : H → K is called adjointable if the exists a linear operator T ∗ : K → H such that hT x, yi = hx, T ∗yi, for all x ∈ H, y ∈ K. Given any unital C ∗-algebra, every adjoinable operator is bounded and A- linear, [13]. For every countable index set J, the standard Hilbert A-module is defined by ℓ2(J, A) :=({aj}j∈J ⊂ A : Xj∈J a∗ j aj is norm convergent in A). For each j ∈ J, letting δij = 0 if i 6= j and δjj = 1, define ej = {δij1A}i∈J . The sequence {ej}j∈J is called the canonical orthonormal basis for ℓ2(J, A). Following [6], we will now give the key definition of frames in Hilbert C ∗- modules. 2 Definition 1.1 A sequence {xj}j∈J of elements in a Hilbert C ∗-module H is said to be a frame if the infinite series Pj∈J hx, xjihxj, xi converges in norm for all x ∈ H and there exist two constants 0 < C ≤ D < ∞ such that Chx, xi ≤Xj∈J hx, xjihxj, xi ≤ Dhx, xi, x ∈ H. (1.1) Appropriate choices for the numbers C and D are called frame bounds. {xj}j∈J is a Bessel sequence with bound D if the right-hand side inequality in (1.1) holds. A frame is called tight frame if we can choose C = D and it is called Parseval frame if it is a tight frame with bound 1. Kasparov's Stabilization Theorem [9] shows that every finitely or countably generated Hilbert C ∗-module over a unital C ∗-algebra has a frame. In in- equality (1.1) we are comparing the positive elements in A. Arambasi´c [1] and Jing in [8], independently showed that one can replace (1.1) with two inequal- ities in terms of the norm of elements. Indeed, it is proved that a sequence {xj}j∈J ⊂ H is a frame if and only if there exist positive constants C and D such that hx, xjihxj, xik ≤ Dkxk2, x ∈ H. Ckxk2 ≤ kXj∈J For every Bessel sequence {xj}j∈J , the operator T : H → l2(J, A) defined as T x = {hx, xji}j∈J , x ∈ H is called the analysis operator. The operator T is adjointable and its adjoint, the synthesis operator, is given by U{aj}j∈J =Pj∈J ajxj. The frame operator S : H → H, defined as Sx = U ∗U(x) = Pj∈Jhx, xjixj, is bounded, positive and invertible, see [6]. Thus the following reconstruction formula holds for frames in Hilbert C ∗-modules: x = SS−1x =Xj∈J hS−1x, xjixj =Xj∈J hx, S−1xjixj, x ∈ H. (1.2) We call {S−1xj}j∈J the canonical dual frame of {xj}j∈J. Also if {xj}j∈J and {yj}j∈J are frames and x = Pj∈Jhx, yjixj, for all x ∈ H, then {xj}j∈J and {yj}j∈J are called dual frames. Following [6], a frame {xj}j∈J for a Hilbert A-module H is a Riesz basis if the following two conditions are satisfied: (i) xj 6= 0 for all j ∈ J. 3 (ii) If Pj∈S ajxj = 0 for a finite set of coefficients {aj}j∈S ⊂ A, S ⊆ J, then ajxj = 0 for every j ∈ S. It is proved in [7] that a Riesz basis in a Hilbert C ∗-module may have many dual frames and it may even admit two different dual frames both of which are Riesz bases. This shows that the definition of Riesz bases in Hilbert C ∗- modules is not the analog of Riesz basis in Hilbert spaces. Later, in [10] modular Riesz bases were introduced: Definition 1.2 A sequence {xj}j∈J in Hilbert A-module H is called a modular Riesz basis for H, if there exists an invertible A-linear and adjointable operator U : l2(J, A) → H such that U ej = xj for each j ∈ J, where {ej}j∈J = {(δij1A)i∈J }j∈J is the standard orthonormal basis of l2(J, A). It is proved in [11] that a sequence {xi}j∈J ⊂ H is a modular Riesz basis if and only if {xj}j∈J is a set of generators for H (as a Banach A-module) and there exist C, D > 0 such that for every finite sequence {ai}i∈S in A, 2 C(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi∈S ai2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi∈S aixi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ D(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi∈S ai2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . (1.3) This gives a characterization of modular Riesz bases in terms of the analysis operator. As a consequence of (1.3), it is proved in [11] that the set of all modular Riesz bases coincides with the set of all frames that have a unique dual frame which is a modular Riesz basis. Also it is straightforward from (1.3), that the analysis operator of a modular Riesz basis is bijective and that an infinite series Pj∈J ajxj is convergent if and only if {aj}j∈J ∈ ℓ2(J, A). 2 Modular Riesz bases versus Riesz bases It is well-known that in a Hilbert space H, Riesz bases, w-independent frames, and exact frames are different names for the same class of sequences. We will now define the analogue version of the mentioned sequences in Hilbert C ∗-modules, and consider their interrelations. Given a C ∗-algebra A and a Hilbert A-module H, let A-span{xj}j∈J :=(Xj∈S ajxj : aj ∈ A, and S ⊂ J is a finite subset ) . 4 Definition 2.1 A sequence {xj}j∈J ⊂ H is said to be: (i) ω-independent, if wheneverPj∈J ajxj is convergent and equal to zero for some {aj}j∈J ⊂ A, then necessarily aj = 0 for all j ∈ J. (ii) biorthogonal with {yj}j∈J ⊂ H, if hxi, yji = δij1A for all i, j ∈ J. (iii) exact frame, if {xj}j∈J is a frame and for every ℓ ∈ J, {xj}j6=ℓ cease to be a frame for H. The following result generalizes Theorem 7.7.1 in [4] to Hilbert C ∗-modules and shows the connection between the sequences defined in part (i) and (ii) of Definition 2.1 and modular Riesz bases. Theorem 2.2 Let {xj}j∈J be a frame in a Hilbert A-module H. Then the following statements are equivalent: (i) {xj}j∈J is a modular Riesz basis, (ii) {xj}j∈J is ω-independent, (iii) {xj}j∈J and its canonical dual {S−1xj}j∈J are biorthogonal, (iv) {xj}j∈J has a biorthogonal sequence, Proof. (i) ⇒ (ii) If {xj}j∈J is a modular Riesz basis, then T ∗ : l2(J, A) → H Since T ∗ is injective, we have {aj}j∈J = 0. (ii) ⇒ (iii) Consider the canonical dual frame {S−1xj}j∈J , where S is the is bijective by (1.3). Assume that Pj∈J ajxj = 0 for some {aj}j∈J ⊆ A. Since Pj∈J ajxj is convergent, {aj}j∈J ∈ l2(J, A). Therefore {aj}j∈J ∈ KerT ∗. frame operator of {xj}j∈J. We have xℓ = Pj∈J hS−1xℓ, xjixj for all ℓ ∈ J. hS−1xℓ, xjixj + (hS−1xℓ, xℓi − 1A)xℓ = 0. Hence Xj6=ℓ Since {xj}j∈J is ω-independent, we have hS−1xℓ, xℓi = 1A, hS−1xℓ, xji = 0 f or each j 6= ℓ. (iii) ⇒ (iv) Clear. (iv) ⇒ (i) Let T ∗ be the synthesis operator for {xj}j∈J and let {aj}j∈J ∈ KerT ∗ and {yj}j∈J be biorthogonal to {xj}j∈J. Then for every ℓ ∈ J, aℓ = haℓxℓ, yℓi = hXj∈J ajxj, yℓi = 0. Therefore the synthesis operator is bijective. (cid:3) 5 Corollary 2.3 Let H be a Hilbert A-module H over a unital C ∗-algebra A and let {xj}j∈J be a frame in H with the canonical dual frame {S−1xj}j∈J. If hxj, S−1xji = 1A for each j ∈ J, then {xj}j∈J is a modular Riesz basis. Proof. By the reconstruction formula, for each j ∈ J we have xj =Xℓ∈J hxj, S−1xℓixℓ = xj +Xℓ6=j hxj, S−1xℓixℓ. ThusPℓ6=jhxj, S−1xℓixℓ = 0. Applying S−1, we obtainPℓ6=jhxj, S−1xℓiS−1xℓ = 0. From this relation, we conclude that hxj, S−1xℓihS−1xℓ, xji = 0, for all j ∈ J. (2.1) Xℓ6=j Since the element hxj, S−1xℓihS−1xℓ, xji = hxj, S−1xℓihxj, S−1xℓi∗ is positive for all j, ℓ ∈ J, (2.1) implies that khxj, S−1xℓik2 = khxj, S−1xℓihxj, S−1xℓi∗k = 0 for all ℓ ∈ J \ {j}. This means that {xj}j∈J and its canonical dual frame are biorthogonal. There- fore by Theorem 2.2, {xj}j∈J is a modular Riesz basis. (cid:3) In a Hilbert space, a frame is exact if and only if it is a Riesz basis, [4, Theorem 5.5.4]. We will now analyse the relationship between the exact frames and Riesz bases in Hilbert C ∗-modules; in particular the result with show that the situation is different compare to the Hilbert space case. We will need the following lemma, which yields a characterization for exact frames in Hilbert C ∗-modules, see [8] for the proof. Lemma 2.4 Let {xj}j∈J be a frame for a Hilbert A-module H and let 1A be the identity element of A. For each ℓ ∈ J, the sequence {xj}j6=ℓ is a frame for H if and only if 1A − hxℓ, S−1xℓi is invertible in A. The following result shows that not only modular Riesz basis but also the Riesz bases are indeed exact frames. Proposition 2.5 Let H be a Hilbert A-module H over a unital C ∗-algebra A. If {xj}j∈J is a Riesz basis then it is an exact frame. Proof. Assume that {xj}j∈J is a Riesz basis and that there exists some ℓ ∈ J such that {xj}j6=ℓ is a frame for H. Then there exists {aj}j∈J\{ℓ} ⊂ A such that xℓ = Pj6=ℓ ajxj. Hence letting aℓ = −1A, we have Pj∈J ajxj = 0. This leads to a contradiction with the definition of a Riesz basis. (cid:3) 6 Unfortunately, as we will see in the next example, the set of modular Riesz bases is smaller than the set of exact frames. Example 2.6 Consider the Banach space of all bounded sequences ℓ∞(N). This space forms a C ∗-algebra with respect to operations {uj}j∈N.{vj}j∈N = {ujvj}j∈N, {uj}∗ j∈N = {¯uj}j∈N. Let H = c0 denote the Banach space of all sequences vanishing at infinity. Then H is a Hilbert A-module with inner product h {uj}j∈N, {vj}j∈Ni = {uj ¯vj}j∈N, {uj}j∈N, {vj}j∈N ∈ H. Let δj ∈ H be the sequence in H that takes the value 1 at the jth coordinate and 0 everywhere else. Note that the sequence {δj}j∈J is a Parseval frame, since for every u = {uj}j∈J ∈ H (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj∈J hu, δjihδj, ui(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xj∈J = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) uj2δj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13){uj2}j∈J(cid:13)(cid:13) = sup uj2 = (sup uj)2 = kuk2. Thus the frame operator is S = IdH and for each j ∈ N, hδj, S−1δji = hδj, δji = δj. Therefore the sequence 1A − hδj, S−1δji takes the value 0 at the jth coordinate and 1 everywhere else. This shows that hδj, S−1δji 6= 1A and that also 1A − hδj, S−1δji is not invertible in A. By Lemma 2.4 we conclude that {δj}j∈N is an exact frame. On the other hand Theorem 2.2 implies that {δj}j∈N is not a modular Riesz basis. Moreover, the set {δj}j∈N is indeed a Riesz basis for H. To see this, let aj = {aij}i∈N ∈ A and S ⊂ N be a finite set and assume that Pj∈S ajδj = 0. We have Xj∈S ajδj =Xj∈S {aijδij}i∈N = {Xj∈S aijδij}i∈N = 0. This implies that ajj = 0 for all j ∈ S and therefore ajδj = 0. This shows that {δj}j∈N is a Riesz basis for H. (cid:3) 7 Acknowledgments After acceptance of the paper the author became award of the paper [2] which has some overlap with the results. Indeed, part of the results in Theorem 2.2 also follows from Cor. 4.20 in [2]. The author would like to thank Professor Arambasic for pointing this out. References [1] L. Arambasi´c, On frames for countably generated Hilbert C ∗-modules, Proc. Amer. Math. Soc. 135 (2007) 469 -- 478. [2] L. Arambasi´c, and D. Baki´c. Frames and outer frames for Hilbert-modules. Linear and multilinear algebra 65.2 (2017) 381 -- 431. [3] D. Baki´c, Weak frames in Hilbert C ∗-modules with application in Gabor analysis. arXiv preprint arXiv:1903.01952 (2019). [4] O. Christensen: An introduction to frames and Riesz bases, Second ex- panded edition. Birkhauser (2016). [5] M. Frank and D. R. Larson, A module frame concept for Hilbert C ∗- modules, Contemporary Math., 247 (1999) 207 -- 234. [6] M. Frank and D. R. Larson, Frames in Hilbert C ∗-modules and C ∗- algebras, J. Operator Theory 48 (2002) 273 -- 314. [7] D. Han, W. Jing, D. Larson and R. Mohapatra, Riesz bases and their dual modular frames in Hilbert C ∗-modules, J. Math. Anal. Appl. 343 (2008) 246 -- 256. [8] W. Jing, Frames in Hilbert C ∗-modules, PhD thesis, University of Central Florida, 2006. [9] I. Kaplansky, Algebras of type I, Ann. Math. 56 (1952) 460 -- 472. [10] A. Khosravi, B. Khosravi , Fusion frames and g-frames in Hilbert C ∗- modules, Int. J. Wavelet, Multiresolution and Inf. Processing 6 (2008) 433 -- 446. [11] A. Khosravi, B. Khosravi, g-frames and modular Riesz bases in Hilbert C ∗- modules, Int. J. Wavelets, Multiresolution and Inf. Processing. 10 (2012) 1250013. 8 [12] A. Khosravi, K. Musazadeh, Fusion frames and g-frames, J. Math. Anal. Appl. 342 (2008) 1068 -- 1083. [13] E. C. Lance, Hilbert C ∗-modules: A Toolkit for Operator Algebraists, Lon- don Math. Soc. Lecture Note Ser. vol. 210, Cambridge Univ. Press, 1995. [14] A. Austad, M. S. Jakobsen, and F. Luef, Gabor Duality Theory for Morita Equivalent C ∗-algebras. arXiv preprint arXiv:1905.01889 (2019). [15] N. Wegge-Olsen, K-theory and C ∗-algebras - A Friendly Approach, Oxford Univ. Press, Oxford, England, 1993. Marzieh Hasannasab, Technical University of Kaiserslautern Paul-Ehrlich Strasse Gebaude 31, 67663 Kaiserslautern, Germany [email protected] 9
1608.07859	2	1608	2017-03-29T17:07:29	On the non-triviality of certain spaces of analytic functions. Hyperfunctions and ultrahyperfunctions of fast growth	[ "math.FA", "math.CV" ]	We study function spaces consisting of analytic functions with fast decay on horizontal strips of the complex plane with respect to a given weight function. Their duals, so called spaces of (ultra)hyperfunctions of fast growth, generalize the spaces of Fourier hyperfunctions and Fourier ultrahyperfunctions. An analytic representation theory for their duals is developed and applied to characterize the non-triviality of these function spaces in terms of the growth order of the weight function. In particular, we show that the Gelfand-Shilov spaces of Beurling type $\mathcal{S}^{(p!)}_{(M_p)}$ and Roumieu type $\mathcal{S}^{\{p!\}}_{\{M_p\}}$ are non-trivial if and only if $$ \sup_{p \geq 2}\frac{(\log p)^p}{h^pM_p} < \infty, $$ for all $h > 0$ and some $h > 0$, respectively. We also study boundary values of holomorphic functions in spaces of ultradistributions of exponential type, which may be of quasianalytic type.	math.FA	math	ON THE NON-TRIVIALITY OF CERTAIN SPACES OF ANALYTIC FUNCTIONS. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH ANDREAS DEBROUWERE AND JASSON VINDAS Abstract. We study function spaces consisting of analytic functions with fast de- cay on horizontal strips of the complex plane with respect to a given weight function. Their duals, so called spaces of (ultra)hyperfunctions of fast growth, generalize the spaces of Fourier hyperfunctions and Fourier ultrahyperfunctions. An analytic rep- resentation theory for their duals is developed and applied to characterize the non- triviality of these function spaces in terms of the growth order of the weight function. In particular, we show that the Gelfand-Shilov spaces of Beurling type S(p!) (Mp) and Roumieu type S{p!} {Mp} are non-trivial if and only if (log p)p hpMp sup p≥2 < ∞, for all h > 0 and some h > 0, respectively. We also study boundary values of holomorphic functions in spaces of ultradistributions of exponential type, which may be of quasianalytic type. 1. Introduction The purpose of this paper is to introduce and analyze two families of spaces of analytic functions and provide an analytic representation theory for their duals. Our function spaces consist of analytic functions with very fast decay on strips of the complex plane with respect to a weight function. Their duals lead to new classes of hyperfunctions and ultrahyperfunctions of 'fast growth', and generalize the Fourier hyperfunctions and the Fourier ultrahyperfunctions. Fourier hyperfunctions were systematically studied by Kawai in [20], and their local theory includes that of Sato's hyperfunctions [19, 30]. In one dimension, their global sections on R = [−∞, ∞] are the dual of the space of analytic functions with expo- nential decay on some (horizontal) strip, the latter test function space coincides with the Gelfand-Shilov space S {1} {p!} of Roumieu type [4, 14]. Moreover, in the co- homological approach, this dual space can be represented as the quotient of the space {1} = S {p!} 2010 Mathematics Subject Classification. Primary 30D60, 46E10, 46F15. Secondary 46F05, 46F12, 46F20. Key words and phrases. spaces of analytic functions; hyperfunctions; ultrahyperfunctions; ultra- distributions; boundary values; analytic representations; non-triviality; Laplace transform; Gelfand- Shilov spaces. A. Debrouwere gratefully acknowledges support by Ghent University, through a BOF Ph.D.-grant. The work of J. Vindas was supported by Ghent University, through the BOF-grant 01N01014. 1 2 A. DEBROUWERE AND J. VINDAS of analytic functions defined outside the real line and having infra-exponential growth outside every strip containing the real line modulo its subspace of entire functions of infra-exponential type. We mention that the use of analytic functions for the repre- sentation of dual spaces has a long tradition, which goes back to the pioneer works of Kothe [24, 25] and Silva [37, 38]. We refer to the monographs [4, 5] for accounts on analytic representations of (ultra)distributions, see also [8, 11] for recent results. Interestingly, several basic problems in the theory of PDE naturally lead to ultra- distribution spaces whose elements are not hyperfunctions [15]. Important instances of such spaces are the spaces of tempered ultrahyperfunctions and Fourier ultrahy- perfunctions, introduced in one-dimension by Silva [38] and in several variables by Hasumi [16] and Park and Morimoto [33], see also [18, 28, 29, 40, 44]. More recently, microlocal analysis, edge of the wedge theorems, and Bochner-Schwartz theorems in the context of ultrahyperfunctions have been investigated in [3, 13, 43]; applications of tempered ultrahyperfunctions can be found e.g. in [10, 32]. Also in recent times, ultra- hyperfunctions have shown to be quite useful in mathematical physics, particularly as a framework for Wightman-type axiomatic formulations of relativistic quantum field theory with a fundamental length [2, 12, 31, 39]. In this article we are interested in the following generalization of Kawai's and Silva's works (see also [15, Chap. 1]). Let ω : [0, ∞) → [0, ∞) be a non-decreasing function. For h > 0 we denote by T h the horizontal strip of the complex plane Im z < h. We shall study the space U(ω)(C) of entire functions ϕ satisfying (1.1) ϕ(z)eω(λ Re z) < ∞, sup z∈T h for every h, λ > 0, and the space A{ω}(R) of analytic functions ϕ defined on some strip T hand satisfying the estimate (1.1) for some h, λ > 0. We call their duals U ′ (ω)(C) and A′ {ω}(R) the spaces of ultrahyperfunctions of (ω)-type and hyperfunctions of {ω}-type, respectively. When ω(t) = t, one recovers the spaces of Fourier ultrahyperfunctions and Fourier hyperfunctions. The first natural question to be addressed is whether these spaces are non-trivial. We shall provide a necessary and sufficient condition on the growth of ω characterizing the non-triviality of both spaces. One of our main results asserts that U(ω)(C) contains a non-identically zero function if and only if (1.2) e−µtω(t) = 0, lim t→∞ for each µ > 0, while the corresponding non-triviality assertion holds for A{ω}(R) if only if (1.2) is satisfied for some µ > 0. We remark that this characterization is of similar nature to the Denjoy-Carleman theorem in the theory of ultradifferentiable functions [1, 21]. In the case of A{ω}(R), the result will follow from complex analysis arguments. The analysis of U(ω)(C) requires a more elaborate treatment, involving duality arguments and analytic representations. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 3 It is worth pointing out that when ω is the associated function [21] of a logarithmi- cally convex weight sequence (Mp)p∈N, our test function spaces coincide with Gelfand- Shilov spaces of mixed type [4, 14], that is, U(ω)(C) = S (p!) (Mp) and A{ω}(R) = S {p!} {Mp}. Specializing our result, we obtain (cf. Proposition 2.7): S (p!) (Mp) and S {p!} {Mp} are non-trivial if and only if the weight sequence satisfies the mild lower bound (1.3) (log p)p hpMp sup p≥2 < ∞, for all h > 0 in the Beurling case and for some h > 0 in the Roumieu case; see Example 2.6 for instances of sequences that satisfy (1.3). Notice that finding precise conditions on two weight sequences that characterize the non-triviality of S (Np) {Mp} is a long-standing open question, raised by Gelfand and Shilov [15, Chap. 1]. Our result then solves this question when one fixes Np = p!. (Mp) and S {Np} Our second goal is to give an analytic representation theory for the dual spaces U ′ (ω)(C) and A′ {ω}(R). We will show that every ultrahyperfunction of (ω)-type (hy- perfunction of {ω}-type) can be represented as the boundary value of an analytic function defined outside some strip (outside the real line) and satisfying bounds of type O(eω(λ Re z)) on substrips of its domain. Furthermore, we prove a result concern- ing the analytic continuation of functions whose boundary values give rise to the zero functional, which can either be viewed as a weighted version of Painlev´e's theorem on analytic continuation or as a one-dimensional version of the edge of the wedge theorem. This allows us to express U ′ {ω}(R) as quotients of certain spaces of ana- lytic functions. We mention that in order to establish the non-triviality of U(ω)(C), we should already pass through the analytic representation theory of some intermediate duals of certain spaces of analytic functions. (ω)(C) and A′ As an application of our ideas, we shall study in the last part of the article bound- ary values of holomorphic functions in spaces of ultradistributions of exponential type. Such spaces are defined in terms of a weight function ω that is subadditive but not necessarily non-quasianalytic, and they are the duals of spaces of ultradifferentiable functions that generalize the Hasumi-Silva space K1(R) of exponentially rapidly de- creasing smooth functions [18, 16, 45]. We use here some variants of the theory of almost analytic extensions [4, 34] and Laplace transform characterizations of several interesting subspaces of Fourier (ultra)hyperfunctions. In order to study the Laplace transform, we introduce a notion of support for (ultra)hyperfunctions of fast growth in the spirit of Silva [38] and provide a support separation theorem. The paper is organized as follows. Section 2 is devoted to the study of some properties of weight functions. Many crucial arguments in the article depend upon the existence of analytic functions satisfying certain lower and upper bounds with respect to a weight function on a strip, Section 3 deals with the construction of such analytic functions. We also prove there a quantified Phragm´en-Lindelof type result for analytic functions on strips. Basic properties of the test function spaces U(ω)(C) and A{ω}(R) are discussed in Section 4, where we determine their images under the Fourier transform as well. The non-triviality theorem for A{ω}(R) is also shown in Section 4. In Section 5 we present 4 A. DEBROUWERE AND J. VINDAS (ω)(C) and A′ the analytic representation theory for U ′ {ω}(R), and, as an application, we characterize the non-triviality of the space U(ω)(C). We introduce the notion of (real) support in Section 6 and provide a support separation theorem. In Section 7, we give a variant of the theory from the previous sections that applies to spaces defined via subadditive weights. For a weight function ω, the modification consists in replacing (1.1) in the definition of the test functions spaces by estimates of the form ϕ(z)eλω( Re z) < ∞. sup z∈T h We mention that, in the Roumieu case, these test function spaces were investigated by Langenbruch [26] under a mild condition (much weaker than subadditivity) on the weight function ω (see Remark 7.3). If ω is subadditive, the resulting dual spaces are subspaces of the Fourier (ultra)hyperfunctions, which we employ to introduce spaces of ultradistributions of exponential type as their images under the Fourier transform, in analogy to the Beurling-Bjorck approach to ultradistribution theory [1]. We conclude the article with the study of boundary values of analytic functions in these ultradistri- bution spaces of exponential type in Section 8. 2. Weight functions In this preliminary section we prove some auxiliary results on weight functions that will be used later in this work. We also discuss the special case when the weight function arises as the associated function of a weight sequence. Another class of weight functions will be considered in Section 7. A weight function is simply a non-decreasing function ω : [0, ∞) → [0, ∞). Unless otherwise stated, we shall always assume throughout Sections 2-6 that ω satisfies (2.1) lim t→∞ ω(t) log t = ∞. We often consider the ensuing additional conditions on weight functions: (δ) 2ω(t) ≤ ω(Ht) + log A, t ≥ 0, for some A, H ≥ 1, We also introduce the following quantified version of (ǫ)0 and (ǫ)∞: We extend ω to the whole real line by setting ω(t) = ω(t), t ∈ R. Furthermore, for λ > 0 we employ the short-hand notation ωλ(t) = ω(λt). The relation ω ⊂ σ between two weight functions means that there are C, λ > 0 such that σ(t) ≤ ωλ(t) + C, t ≥ 0. 0 (ǫ)0 Z ∞ (ǫ)∞ Z ∞ (ǫ)µ Z ∞ 0 0 ω(t)e−µtdt < ∞ for all µ > 0, ω(t)e−µtdt < ∞ for some µ > 0. ω(t)e−µtdt < ∞, µ > 0. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 5 The stronger relation ω ≺ σ means that the latter inequality remains valid for every λ > 0 and suitable C = Cλ > 0. The reader should keep in mind that these two relations "reverse" orders of growth. We say that ω and σ are equivalent, denoted by ω ∼ σ, if both ω ⊂ σ and σ ⊂ ω hold. Examples 2.1. Some examples of weight functions are: s > 0, • ts, • exp(ts logr(1 + t)), • exp(cid:18) • et. t logs(e + t)(cid:19) , 0 ≤ s < 1, r ≥ 0, sr > 0, s > 0, They all satisfy (δ). Moreover, the first three of them fulfill (ǫ)0, while the last one satisfies (ǫ)∞ but not (ǫ)0. The next lemma gives alternative forms of (ǫ)0 and (ǫ)∞. Lemma 2.2. Let ν > µ > 0 and suppose ω is a weight function satisfying (ǫ)µ, then ω(t) = o(eνt). Consequently, ω satisfies (ǫ)0 ((ǫ)∞) if and only if et ≺ ω (et ⊂ ω). Proof. Suppose the opposite, then there would exist ε > 0 and a sequence of positive numbers (tn)n∈N such that ω(tn) ≥ εeνtn, tn+1 ≥ νtn µ , n ∈ N. Hence, Z ∞ 0 ω(t)e−µtdt ≥ ∞Xn=0Z tn+1 tn ω(t)e−µtdt ≥(cid:18) ν µ − 1(cid:19) t0 ω(tn)e−νtn. ∞Xn=0 (cid:3) Since the last series is divergent, this contradicts (ǫ)µ. We now show three useful lemmas. Lemma 2.3. Let ω be a weight function satisfying (δ) and (ǫ)0 ((ǫ)∞). Then, there is another weight function σ with ω ∼ σ that satisfies (δ), (ǫ)0 ((ǫ)∞), and the additional condition (ζ) lim t→∞ σ(λt) − σ(t) = ∞, ∀λ > 1. Proof. Let (tn)n∈N be a sequence of non-negative numbers such that t0 = 0 < t1 < t2 < . . . < tn → ∞ and Define ω(t) ≥ n log t, t ≥ tn. ρ(t) = n log t, if t ∈ [tn, tn+1), and σ(t) = ω(t) + ρ(t) ≤ 2ω(t). The condition (δ) thus implies that ω and σ are equivalent weight functions. Since (δ) and (ǫ)0 ((ǫ)∞) are invariant under the relation ∼, the weight σ satisfies these conditions as well. For λ > 1 and t ∈ [tn, tn+1), we have σ(λt) − σ(t) ≥ ρ(λt) − ρ(t) ≥ n(log(λt) − log t) = n log λ, whence (ζ) follows. (cid:3) 6 A. DEBROUWERE AND J. VINDAS Lemma 2.4. Let ω be a weight function satisfying (ǫ)0 ((ǫ)∞). Then, there is a weight function σ satisfying (ǫ)0 ((ǫ)∞), (δ), and ω(t) ≤ σ(t) for all t ≥ 0. Proof. We take σ(t) =R t+1 also have σ(t) ≥ R t+1 t obtain that σ is convex; therefore, 2σ(t) ≤ σ(2t) + σ(0). 0 ω(x)dx. The condition (ǫ)0 ((ǫ)∞) clearly holds for σ. We ω(x)dx ≥ ω(t). Furthermore, since ω(t) is non-decreasing, we (cid:3) Lemma 2.5. Let ω be a weight function satisfying (ǫ)0. Then, there is a weight function σ satisfying (ǫ)0 such that ωλ(t) = o(σ(t)) for all λ > 0. Proof. We inductively determine a sequence of non-negative numbers (tn)n∈Z+ with t1 = 0 that satisfies ω(t)e−t/n2 dt ≤ 1 2n , tn n ≥ tn−1 n − 1 + 1, n ≥ 2. Z ∞ tn We now define σ(t) = nω(nt), Clearly, σ is a weight function and ωλ(t) = o(σ(t)) for all λ > 0. Moreover, for each n0 ∈ Z+, for t ∈(cid:20)tn n , tn+1 n + 1(cid:19) . ∞Xn=n0Z tn+1/n+1 tn/n ω(t)e−t/n2 dt Z ∞ 0 σ(t)e−t/n0dt ≤Z tn0 /n0 0 σ(t)e−t/n0dt + ≤ tn0ω(tn0) + ≤ tn0ω(tn0) + tn ∞Xn=n0Z ∞ ∞Xn=n0 1 2n < ∞. nω(nt)e−t/ndt (cid:3) We now consider the case when the weight function arises as the associated function of a weight sequence [21]. Let (Mp)p∈N be a sequence of positive real numbers and define mp = Mp/Mp−1 for all p ≥ 1. We shall always assume that limp→∞ mp = ∞ and that the sequence Mp is log-convex, that is, M 2 p ≤ Mp−1Mp+1 for p ≥ 1, or, equivalently, that mp is non-decreasing. In the sequel, we consider the ensuing conditions for weight sequences: (M.2) Mp+q ≤ AH p+qMpMq, p, q ∈ N, for some A, H ≥ 1, (M.5)0 (M.5)∞ e−µmp < ∞ for all µ > 0, e−µmp < ∞ for some µ > 0. ∞Xn=1 ∞Xn=1 The associated function M of Mp is defined as M(t) = sup p∈N log(cid:18)tpM0 Mp (cid:19) , t > 0, HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 7 and M(0) = 0. Clearly, M is a weight function. We denote by m(t) = Xmp≤t 1, t ≥ 0, the counting function of the sequence (mp)p∈Z+. It is well known that [21, Eq. (3.11), p. 50] (2.2) M(t) =Z t 0 m(λ) λ dλ, t ≥ 0. As customary, the relation Mp ⊂ Np between two weight sequences means that there are C, h > 0 such that Mp ≤ ChpNp for all p ∈ N. The stronger relation Mp ≺ Np means that the latter inequality remains valid for every h > 0 and a suitable C = Ch > 0. We say that Mp and Np are equivalent, denoted by Mp ∼ Np, if both Mp ⊂ Np and Np ⊂ Mp hold. Denote by M and N the associated functions of Mp and Np, respectively. Then, Mp ⊂ Np (Mp ≺ Np) if and only if M ⊂ N (M ≺ N) [21, Lemmas 3.8 and 3.10] and therefore our use of the symbols ⊂, ≺, and ∼ for weight functions is consistent with that for weight sequences. Examples 2.6. Some examples of weight sequences are: s > 0, • p!s, • log(p + e)s(p+e)r, • log(p + e)(p+e). s, r ≥ 1, sr > 1, They all satisfy (M.2). Moreover, the first two of them fulfill (M.5)0 while the last one satisfies (M.5)∞ but not (M.5)0. The next proposition characterizes (M.5)0 and (M.5)∞ in terms of the associated function. Proposition 2.7. Let Mp be a weight sequence. Then, M satisfies (δ) if and only if Mp satisfies (M.2). Moreover, the following statements are equivalent: (i) M satisfies (ǫ)0 ((ǫ)∞), (ii) m satisfies (ǫ)0 ((ǫ)∞), (iii) et ≺ M (et ⊂ M), (iv) et ≺ m (et ⊂ m), (v) Mp satisfies (M.5)0 ((M.5)∞). (vi) log(p + e)(p+e) ≺ Mp (log(p + e)(p+e) ⊂ Mp). Proof. The equivalence between (M.2) and (δ) is shown in [21, Prop. 3.6]. Integration by parts yields e−µλdm(λ) = m(t)e−µt + µZ t 0 m(λ)e−µλdλ, µ > 0. Moreover, by (2.2), we obtain Xmp≤t 0 e−µmp =Z t Z t 0 M(λ)e−µλdλ = 1 µZ t 0 m(λ)e−µλ λ dλ − M(t)e−µt µ , µ > 0. 8 A. DEBROUWERE AND J. VINDAS Lemma 2.2 now implies that (i) -- (v) are equivalent to one another. Since the associated function of the sequence log(p + e)(p+e) is equivalent to et, conditions (iii) and (vi) are also equivalent to each other. (cid:3) 3. Analytic functions in a strip The goal of this section is to construct functions that are analytic and satisfy certain lower and upper bounds with respect to a weight function in a given horizontal strip of the complex plane. Since the functions we aim to construct are zero-free, we first study harmonic functions in a strip. We also show a Phragm´en-Lindelof type result for analytic functions defined on strips and having decay with respect to a weight function, Proposition 3.5 actually delivers a useful three lines type inequality. As usual, O(V ) stands for the space of analytic functions in an open set V ⊆ C. Given h > 0, we write T h = R + i(−h, h), T h − = R + i(−h, 0). Furthermore, we shall always write z = x + iy ∈ C for a complex variable. The Poisson kernel of the strip T π + = R + i(0, h), and T h + is well-known. It is given by sin y P (x, y) = cosh x − cos y , and has the ensuing properties [42]: (A) P (x, y) is harmonic on T 2π + , (B) P (x, y) > 0 on T π +, (C) P (x, y) ≤ , x ≥ 1, 0 < y < 2π, sin ye−x+1 cosh 1 − 1 (D) Z ∞ 0 We employ the notation P (x, y)dx = π − y, 0 < y < π. and, for a measurable function f on R, , h Ph(x, y) = P(cid:16) πx 2hZ ∞ 1 −∞ Ph{f ; x, y} = πy h (cid:17) , h > 0, Ph(t − x, y)f (t)dt, its Poisson transform with respect to the strip T h +. Lemma 3.1. Let ω be a weight function satisfying (ǫ)π/h.Then, Ph{ω; x, y} is harmonic on T h + and satisfies the lower bound If ω satisfies (ǫ)π/(2h), then the upper bound y ω(x) Ph{ω; x, y} ≥ h(cid:17) , 2 (cid:16)1 − π(cid:19)(cid:19)(cid:16)1 − Ph{ω; x, y} ≤(cid:18)ω(2x) + ω(cid:18) 2h x + iy ∈ T h +. y h(cid:17) + C, x + iy ∈ T h +, HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 9 holds as well, where C = Proof. Since e−πt/(2h)ω(t)dt < ∞. e 2h(cosh 1 − 1)Z ∞ Ph{ω; x, y} = Pπnωh/π; 0 πx h , πy h o , we may assume that h = π. Set P {ω; x, y} = Pπ{ω; x, y}. The fact that ω satisfies (ǫ)1 implies that P {ω; x, y} is harmonic on T π + [42, Thm. 1]. By the symmetry properties of the weight function ω and the Poisson kernel P , it suffices to show the inequalities for x ≥ 0. We have that 1 P {ω; x, y} = P (t, y)ω(t + x)dt ≥ P (t, y)ω(t + x)dt −∞ 2πZ ∞ 2π Z ∞ ω(x) 0 ≥ P (t, y)dt = 1 2πZ ∞ π(cid:17) , 2 (cid:16)1 − 0 y ω(x) where, in the last equality, we have used property (D) of the Poisson kernel P . Next, assume that ω satisfies (ǫ)1/2. Then, P {ω; x, y} ≤ = P (t, y)ω(t + x)dt P (t, y)ω(t + x)dt + P (t, y)ω(t + x)dt 0 1 πZ ∞ πZ x 1 0 π ≤ ω(2x) Z ∞ ≤ ω(2x)(cid:16)1 − 0 1 πZ ∞ πZ ∞ 1 x 0 P (t, y)dt + P (t, y)ω(2t)dt y π(cid:17) + 1 πZ ∞ 0 P (t, y)ω(2t)dt. Property (C) of the Poisson kernel P and condition (ǫ)1/2 imply that Z ∞ 0 P (t, y)ω(2t)dt =Z 1 0 P (t, y)ω(2t)dt +Z ∞ 1 P (t, y)ω(2t)dt ≤ ω(2)(π − y) + e 2(cosh 1 − 1)Z ∞ 0 e−t/2ω(t)dt < ∞. (cid:3) Lemma 3.1 has the following important consequence. Proposition 3.2. Let ω be a weight function satisfying (ǫ)π/(8hλ). Then, there is F ∈ O(T h) such that eω(λx) ≤ F (z) ≤ Ce4ω(2λx), z ∈ T h, for some C = Ch,λ > 0. If, in addition, ω satisfies (δ), then z ∈ T h, F (z) ≤ A3Ceω(2H 2λx), ω(x) 2 (cid:16)1 − with y h(cid:17) ≤ Ph{ω; x, y} ≤(cid:18)ω(x) + ω(cid:18) h h(cosh 1 − 1)Z ∞ C = e 0 π(cid:19)(cid:19)(cid:16)1 − y h(cid:17) + C, x + iy ∈ T h +, e−πt/hω(t)dt < ∞. 10 A. DEBROUWERE AND J. VINDAS where A, H are the constants occurring in (δ). Proof. Define U(x, y) = 4P4h(ωλ; x, y+h). Lemma 3.1 implies that F (z) = eU (x,y)+iV (x,y), with V the harmonic conjugate of U, satisfies all requirements. (cid:3) Remark 3.3. Let ω be a weight function (not necessarily satisfying (2.1)) that is sub- additive, i.e. ω(t1 + t2) ≤ ω(t1) + ω(t2), t1, t2 ≥ 0. Observe that subadditivity implies that ω(t) = O(t); in particular, (ǫ)0 holds. One can readily show the inequalities (cf. the proof of Lemma 3.1) Hence, for each λ > 0 and h > 0, there is F ∈ O(T h) such that eλω(x) ≤ F (z) ≤ Ce4λω(x), z ∈ T h, for some C > 0. Subadditive weight functions will play an important role in Sections 7 and 8 below. We end this section with a Phragm´en-Lindelof type result for analytic functions defined on strips. We need the following lemma. Lemma 3.4. Let ϕ be analytic and bounded on the strip T h Then, if ϕ is non-identically zero, + and continuous on T h +. (i) −∞ <Z ∞ −∞ log ϕ(x)e−πx/hdx and −∞ <Z ∞ −∞ (ii) log ϕ(z) ≤ Ph{log ϕ; x, y} + Ph{log ϕ(· + ih); x, h − y}, log ϕ(x + ih)e−πx/hdx, z ∈ T h +. Proof. We may assume that h = π. The results can be derived by considering the conformal mapping z → i − ez i + ez from the strip T π statements for bounded analytic functions on the unit disk [22]. + onto the the unit disk and using the well known counterparts of the (cid:3) Proposition 3.5. Let ω be a weight function satisfying (ǫ)π/h. Let ϕ be holomorphic on the strip T h +, and suppose that + and continuous on T h z ∈ T h +, ϕ(z) ≤ M, ϕ(x) ≤ Ce−ω(x), x ∈ R, for some M, C > 0. Then, ϕ(z) ≤ M y/hC 1−(y/h) exp(cid:18)− ω(x) 2 (cid:16)1 − y h(cid:17)(cid:19) , z = x + iy ∈ T h +. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 11 Proof. We may assume that ϕ is non-identically zero. By Lemma 3.4(ii) and property (D) of the Poisson kernel we have that ϕ(z)ePh{ω;x,y} ≤ M y/h exp(cid:18) 1 2hZ ∞ −∞ (log ϕ(t) + ω(t))Ph(t − x, y)dt(cid:19) . By applying Jensen's inequality to the probability measure Ph(t − x, y)/(2(h − y))dt for x ∈ R and 0 < y < h fixed, we obtain ϕ(z)ePh{ω;x,y} ≤ M y/h 2(h − y)Z ∞ −∞ ≤ M y/hC 1−(y/h). exp(cid:16)(cid:16)1 − y h(cid:17) (log ϕ(t) + ω(t))(cid:17) Ph(t − x, y)dt The result now follows from the first part of Lemma 3.1. (cid:3) 4. Spaces of analytic functions of rapid decay on strips We now introduce and discuss some basic properties of the spaces of analytic func- tions that we are mainly concerned with. They generalize the test function spaces for the Fourier hyperfunctions and ultrahyperfunctions [18, 19, 33, 38]. We will also determine their images under the Fourier transform, extending various results from [6]. Throughout the rest of the article the parameters h, k, λ, b, and R always stand for positive real numbers. Let ω be a weight function. We write Ah ω for the (B)-space (Banach space) consisting of all analytic functions ϕ ∈ O(T h) that satisfy kϕkh ω := sup z∈T h ϕ(z)eω(x) < ∞. We set Ah ωλ = Ah,λ ω and kϕkh ωλ = kϕkh,λ ω = kϕkh,λ, ϕ ∈ Ah,λ ω . Montel's theorem and the uniqueness property of holomorphic functions yield the following simple lemma. Lemma 4.1. Let ω and σ be two weight functions such that (4.1) lim t→∞ σ(t) − ω(t) = ∞. Then, for 0 < h < k, the restriction mapping Ak σ → Ah ω is injective and compact. As topological vector spaces, we define (4.2) and U(ω)(C) = lim ←− h→∞ lim ←− λ→∞ Ah,λ ω , A{ω}(R) = lim −→ h→0+ lim −→ λ→0+ Ah,λ ω , Ah Ak,λ ω , Ah (4.3) (ω) = lim ←− k→h− lim −→ λ→0+ If ω satisfies (δ), Lemma 4.1 implies that U(ω)(C) and Ah (ω) are (F S)-spaces, while A{ω}(R) and Ah {ω} are (DF S)-spaces. Let σ be another weight function such that ω ∼ σ, then U(ω)(C) = U(σ)(C) and A{ω}(R) = A{σ}(R), topologically. The same is {ω} = lim −→ k→h+ lim ←− λ→∞ Ak,λ ω . 12 A. DEBROUWERE AND J. VINDAS true for the spaces (4.3). We shall call the elements of the dual spaces U ′ (ω)(C) and A′ {ω}(R) ultrahyperfunctions of (ω)-type and hyperfunctions of {ω}-type, respectively. When no reference to ω is made, we will simply call them (ultra)hyperfunctions of fast growth. We endow these (and other) duals with the strong topology. Observe that U(t)(C) and A{t}(R) are the test function spaces for the Fourier ultrahyperfunc- tions [18, 33] and Fourier hyperfunctions [20], respectively. If ω = M is the associ- ated function of a weight sequence Mp, then U(Mp)(C) := U(M )(C) = S (p!) (Mp)(R) and A{Mp}(R) := A{M }(R) = S {p!} In particular, A{p!α}(R), α > 0, is equal to the Gelfand-Shilov space S {1} {Mp}(R), the mixed type Gelfand-Shilov spaces [36]. {α} [14]. the non-triviality of the spaces (4.3), and in particular that of A{ω}(R) =Sh>0 Ah As already pointed out in the introduction, it is a priori not obvious that the spaces (4.2) and (4.3) should contain non-identically zero functions. We address in this section {ω}. The analysis of the corresponding problem for U(ω)(C) is more demanding and is post- poned to Section 5. We begin with the following necessary condition for the non- triviality of Ah,λ ω . Proposition 4.2. Let ω be a weight function and suppose that Ah,λ which is non-identically zero. Then, ω satisfies (ǫ)π/(hλ). ω contains a function Proof. Let ϕ be a non-zero element of Ah,λ 0 < k < h, we obtain ω . By applying Lemma 3.4(i) to ϕ(· − ik), −∞ <Z ∞ −∞ log ϕ(x − ik)e−πx/hdx ≤ log kϕkh,λZ ∞ −∞ e−πx/hdx − 2 λZ ∞ 0 ω(x)e−πx/(hλ)dx, (cid:3) and soR ∞ 0 ω(x)e−πx/(hλ)dx < ∞. Proposition 4.3. Let ω be a weight function. The space Ah {ω}) is non-trivial if and only if ω satisfies (ǫ)0 ((ǫ)∞). Consequently, A{ω}(R) is non-trivial if and only if ω satisfies (ǫ)∞. (ω) (Ah Proof. The direct implication follows from Proposition 4.2. Moreover, if ω satisfies (ǫ)∞, Proposition 3.2 gives the non-trivality of Ah {ω}. Assume now that ω satisfies (ǫ)0. By Lemma 2.5, there is a weight function σ satisfying (ǫ)0 such that ωλ(t) = o(σ(t)) for all λ > 0. By Proposition 3.2 there is an analytic function F on T h such that F (z) ≥ eσ(x) for all z ∈ T h. Then, 1/F is an element of Ah (ω) which is non-identically zero. (cid:3) In the next section (Subsection 5.2) we shall show that U(ω)(C) is non-trivial if and only if ω satisfies (ǫ)0, this will be a consequence of a representation theorem for its dual (see Theorem 5.9). The rest of this section is devoted to computing the images of U(ω)(C) and A{ω}(R) under the Fourier transform. These spaces are the ultradifferentiable counterparts HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 13 of the classical Hasumi-Silva space K1(R) of exponentially rapidly decreasing smooth functions [18, 16, 45]. We fix constants in the Fourier transform as F (ϕ)(ξ) = bϕ(ξ) =Z ∞ −∞ ϕ(x)e−ixξdx. Let ω be a weight function. We write Kh ψ ∈ L1(R) such that 1,ω(R) for the function space consisting of all ρω(ψ) := sup x∈R F −1(ψ)(x)eω(x) < ∞, ρh(ψ) := sup ξ∈R ψ(ξ)ehξ < ∞; it becomes a (B)-space when endowed with the norm ρh ω(ψ) := max(ρω(ψ), ρh(ψ)), ψ ∈ Kh 1,ω(R). We set further Kh 1,ωλ(R) = Kh,λ 1,ω(R), ωλ(ψ) = ρh,λ ρh ω (ψ), ψ ∈ Kh,λ 1,ω(R), and K1,(ω)(R) = lim ←− λ→∞ lim ←− h→∞ Kh,λ 1,ω(R), K1,{ω}(R) = lim −→ λ→0+ lim −→ h→0+ Kh,λ 1,ω(R). Proposition 4.4. Let ω be a weight function satisfying (δ) and (ε0) ((ǫ)∞). The Fourier transform is a topological isomorphism of U(ω)(C) onto K1,(ω)(R) (of A{ω}(R) onto K1,{ω}(R)). Proof. For ϕ ∈ Ah,λ [18, p. 167]) ω we have the following formulas for its Fourier transform (see e.g. −∞ Z ∞ Z ∞ −∞  bϕ(ξ) = ϕ(x + ik)e−i(x+ik)ξdx, ξ ≤ 0, ϕ(x − ik)e−i(x−ik)ξdx, ξ ≥ 0, where 0 < k < h. This shows that the Fourier transform is a well defined continuous mapping in both the Beurling and Roumieu case. Conversely, let ψ ∈ Kh,λ 1,ω(R). Then, there is ϕ ∈ O(T h) with bϕ = ψ such that ϕ(z) ≤ ρh(ψ) π(h − k) , z ∈ T k, where 0 < k < h. Invoking Proposition 3.5 and condition (δ), we conclude that the inverse Fourier transform is also a well defined continuous mapping in both the Beurling and Roumieu case. (cid:3) 1,(ω)(R) and K′ We call the elements of K′ 1,{ω}(R) ultradistributions of class (ω) (of Beurling type) of exponential type and ultradistributions of class {ω} (of Roumieu type) of infra-exponential type, respectively. Observe that Proposition 4.4 allows one to define the Fourier transform from K′ 1,{ω}(R) onto A′ (ω)(C) and from K′ 1,(ω)(R) onto U ′ {ω}(R) via duality. 14 A. DEBROUWERE AND J. VINDAS Let us remark that when ω = M is the associated function of a weight sequence Mp satisfying (M.2), then K1,(Mp)(R) := K1,(M )(R) = S (Mp) (p!) (R) and K1,{Mp}(R) := K1,{M }(R) = S {Mp} {p!} (R), topologically. When Mp is non-quasianalytic [21], we have the continuous and dense inclusions D(Mp)(R) ֒→ K1,(Mp)(R) and D{Mp}(R) ֒→ K1,{Mp}(R) and therefore the inclusions K′ (R) and K′ 1,{Mp}(R) ⊂ D{Mp}′ 1,(Mp)(R) ⊂ D(Mp)′ (R). 5. Boundary values of analytic functions in spaces of (ultra)hyperfunctions of fast growth In this section we build an analytic representation theory for the spaces U ′ (ω)(C) and A′ {ω}(R). We show that every ultrahyperfunction of (ω)-type (hyperfunction of {ω}- type) can be represented as the boundary value of an analytic function defined outside some strip (outside the real line) and satisfying certain growth bounds with respect to the weight function ω. Silva obtained analytic representations of ultrahyperfunctions via a careful analysis of the Cauchy-Stieltjes transform [18, 38]. We shall follow a similar approach, the functions constructed in Section 3 are essential for our method. Furthermore, we present an (ultra)hyperfunctional version of Painlev´e's theorem on analytic continuation. These results allow us to express the dual spaces U ′ (ω)(C) and A′ {ω}(R) as quotients of spaces of analytic functions in a very precise fashion. As an application of these ideas, we characterize the non-triviality of the space U(ω)(C). 5.1. Analytic representations. Given 0 < b < R, we use the notation T b,R = T R\T b = R + i((−R, −b) ∪ (b, R)). Let ω be a weight function. We define Ob,R as the (B)-space consisting of all analytic functions F ∈ O(T b,R) that satisfy ω F b,R ω := sup z∈T b,R F (z)e−ω(x) < ∞, and P R ω as the (B)-space consisting of all P ∈ O(T R) such that P R := sup z∈T R P (z)e−ω(x) < ∞. We set Ob,R ωλ = Ob,R,λ ω , P R ωλ = P R,λ ω , ωλ = F b,R,λ F b,R ω = F b,R,λ, and ωλ = P R,λ As in Lemma 4.1, one easily obtains: P R ω = P R,λ, F ∈ Ob,R,λ ω , P ∈ P R,λ ω . Lemma 5.1. Let ω and σ be weight functions such that (4.1) holds. Then, for 0 < b < c < L < R, the restriction mappings Ob,R σ are injective and compact. σ and P R ω → Oc,L ω → OL Suppose that σ is another weight function such that both (4.1) and Z ∞ 0 eω(t)−σ(t)dt < ∞ HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 15 hold. If ω satisfies (δ), the above conditions are fulfilled for ω = ωλ and σ = ωHλ for each λ > 0. Let 0 < b < h < R. Given an analytic function F ∈ Ob,R ω , we associate to F an element of (Ah σ)′ via the boundary value mapping hbv(F ), ϕi = −ZΓk F (z)ϕ(z)dz, ϕ ∈ Ah σ, where b < k < h and Γk is the (counterclockwise oriented) boundary of T k. By Cauchy's integral theorem, the definition of bv(F ) is independent of the chosen k and f = bv(F ) is indeed a continuous linear functional on Ah σ. We say that F is an analytic representation of f . We have the following general result on the existence of analytic representations. Proposition 5.2. Let 0 < k < b < h < R and let ω, σ, and κ be three weight functions satisfying (5.1) and (5.2) lim t→∞ σ(t) − ω(t) = ∞, Z ∞ 0 eω(t)−σ(t)dt < ∞, lim t→∞ Z ∞ 0 κ(t) − σ(t) = ∞, eσ(t)−κ(t)dt < ∞. Furthermore, suppose there is P ∈ O(T R) such that C1eω(x) ≤ P (z) ≤ C2eσ(x), z ∈ T R, for some C1, C2 > 0. Then, every f ∈ (Ak Ob,R κ, that is, there is F ∈ Ob,R on Ah σ σ ω)′ is the boundary value of some element of such that bv(F ) = f on Ah κ. Proof. Cauchy's integral formula yields ϕ(ζ) = 1 2πiP (ζ)ZΓb ϕ(z)P (z) z − ζ dz, ζ ∈ T k, for each ϕ ∈ Ah κ. Let Rn(ζ) be a sequence of Riemann sums converging to the integral in the right-hand side of the above expression. Then, Rn(ζ)/(2πiP (ζ)) → ϕ(ζ), as n → ∞ in Ak ω, as one readily verifies. Hence, such that bv(F ) = f on Ah κ. (cid:3) Our next result shows that functions whose boundary value give rise to the zero functional can be analytically continued. hf (ζ), ϕ(ζ)i =ZΓb 1 P (z) 2πi (cid:28)f (ζ), 2πi (cid:28)f (ζ), (z − ζ)P (ζ)(cid:29) ϕ(z)dz. (ζ − z)P (ζ)(cid:29) 1 P (z) F (z) = Thus, is an element of Ob,R σ 16 A. DEBROUWERE AND J. VINDAS Proposition 5.3. Let 0 < b < h < R and let ω, σ, and κ be three weight functions satisfying (5.1) and (5.2). Furthermore, suppose there is P ∈ O(T R) such that for some C1, C2 > 0. If F ∈ Ob,R ω C1eσ(x) ≤ P (z) ≤ C2eκ(x), z ∈ T R, is such that bv(F ) = 0 on Ah σ, then F ∈ P R κ . Proof. Let 0 < b < k < h < L < R. It suffices to show that (5.3) F (z) = P (z) 2πi ZΓL F (ζ) (ζ − z)P (ζ) dζ, z ∈ T h,L. Fix z ∈ C with h < Im z < L; the case −L < Im z < −h is analogous. We denote by Γ+ (Γ−) the part of a contour Γ in the upper (lower) half-plane. Cauchy's integral formula yields Since ζ → 1/((ζ − z)P (ζ)) ∈ Ah σ, the assumption bv(F ) = 0 on Ah σ implies that F (z) = L P (z) 2πi ZΓ+ ZΓ+ k F (ζ) (ζ − z)P (ζ) F (ζ) (ζ − z)P (ζ) dζ −ZΓ+ k F (ζ) (ζ − z)P (ζ) dζ! . dζ = −ZΓ− k F (ζ) (ζ − z)P (ζ) dζ. Furthermore, because ζ → 1/((ζ − z)P (ζ)) is analytic on the horizontal strip −R < Im ζ < −b, we have by Cauchy's integral theorem that ZΓ− k F (ζ) (ζ − z)P (ζ) dζ =ZΓ− L F (ζ) (ζ − z)P (ζ) dζ. This shows (5.3). (cid:3) Combining these two results with Proposition 3.2, we obtain the following corollaries. Corollary 5.4. Let 0 < k < b < h < R and let ω be a weight function satisfying (δ) such that bv(F ) = f on Ah,2H 3λ and (ǫ)0. For every f ∈ (Ak,λ . ω )′ there is F ∈ Ob,R,2H 2λ ω ω Corollary 5.5. Let 0 < b < h < R and let ω be a weight function satisfying (ǫ)0 and (δ). If F ∈ Ob,R,λ is such that bv(F ) = 0 on Ah,Hλ , then F ∈ P R,2H 3λ . ω ω ω 5.2. Analytic representations of ultrahyperfunctions of (ω)-type and the non- triviality of U(ω)(C). We start by studying the analytic representations of the dual of Ah (ω). Let us introduce some further notation. For 0 < h < R we write Oh,R (ω) = lim −→ b→h− lim −→ L→R+ lim −→ λ→∞ Ob,L,λ ω , P L,λ ω . P R lim −→ λ→∞ (ω) = lim −→ L→R+ (ω) and P R (ω) are (DF S)-spaces. Lemma 5.1 implies that, if ω satisfies (δ), the spaces Oh,R Furthermore, the boundary value mapping bv : Oh,R (ω) → (Ah (ω))′ is well defined and continuous. We need the following lemma. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 17 Lemma 5.6. Let ω be a weight function satisfying (ǫ)0 and (δ). For each λ > 0 the space is dense in the space Ah,Hλ ω with respect to the norm k · kh,λ. Ah,∞ (ω) = lim ←− µ→∞ Ah,µ ω Proof. Let ϕ ∈ Ah,Hλ ψ(0) = 1. Define ϕn(z) = ϕ(z)ψ(z/n) ∈ Ah,∞ (ω) . By Proposition 4.3, Ah+1 (ω) ω for n ≥ 1. Then, is non-trival, so select ψ ∈ Ah+1 (ω) with kϕ − ϕnkh,λ = sup z∈T h ϕ(z)(1 − ψ(z/n))eω(λx) ≤ Akϕkh,Hλ sup z∈T h 1 − ψ(z/n)e−ω(λx), which proves the result since ψ(z/n) → 1, as n → ∞, uniformly on compact subsets of T h+1. (cid:3) We can now show that (Ah (ω))′ is isomorphic to the quotient space Oh,R (ω) /P R (ω). Proposition 5.7. Let 0 < h < R and let ω be a weight function satisfying (ǫ)0 and (δ). The following sequence 0 −→ P R (ω) −→ Oh,R (ω) bv −−→ (Ah (ω))′ −→ 0 is topologically exact. Moreover, for every f ∈ (Ah such that for every R > h there is F ∈ Ob,R,λ every bounded set B ⊂ (Ah ω (ω))′ one can find 0 < b < h and λ > 0 such that bv(F ) = f . In addition, for (ω) such that bv(A) = B. (ω))′ there is a bounded set A ⊂ Oh,R ω (ω) ⊂ Ah . Since Ah,∞ (ω) and suppose bv(F ) = 0 on Ah Proof. In view of the Pt´ak open mapping theorem [17, Chap. 3, Prop. 17.2], it suffices to show that the sequence is algebraically exact. It is clear that P R (ω) ⊆ ker bv. Conversely, let F ∈ Oh,R (ω). Let 0 < b < h, L > R, and λ > 0 be such that F ∈ Ob,L,λ (ω), Lebesgue's dominated convergence theorem and Lemma 5.6 imply that actually bv(F ) = 0 on Ah,H 2λ . Hence, by Corollary 5.5, we have that F ∈ P L,2H 4λ (ω). The second statement (and therefore also the surjectivity of the boundary value mapping) is a consequence of the Hahn-Banach theorem and Corollary 5.4. The last part follows from the general fact that for a surjective continuous linear mapping T : E → F between reflexive (DF )-spaces it holds that for every bounded set B ⊂ F there is a bounded set A ⊂ E such that T (A) = B; this follows from the fact that T t is an injective strict morphism and the bipolar theorem. (cid:3) ⊂ P R ω ω We now proceed to show that U(ω)(C) is non-trivial if and only if ω satisfies (ǫ)0. For it, we need some basic facts about projective spectra. Let (Xn)n∈N be a sequence of topological spaces and let un+1 : Xn+1 → Xn be a continuous mapping for each n ∈ N. Consider the projective limit X of the spectrum (Xn)n∈N, that is, n X = {(xn) ∈Yn∈N Xn : xn = un+1 n (xn+1), n ∈ N}, 18 A. DEBROUWERE AND J. VINDAS with the natural projection mappings uj : X → Xj : (xn) → xj. The spectrum is called reduced if uj has dense range for each j ∈ N. We shall employ the following well known result, due to De Wilde (see also [9] for the case of Banach spaces). Lemma 5.8 ([7]). Let (Xn)n∈N be a spectrum of complete metrizable topological spaces. If un+1 has dense range for each n ∈ N, then the spectrum (Xn)n∈N is reduced. n Theorem 5.9. The space U(ω)(C) is non-trivial if and only if ω satisfies (ǫ)0. If, in addition, ω satisfies (δ), then U(ω)(C) is dense in Ah (ω) for all h > 0. Proof. The direct implication follows from Proposition 4.2. By Lemma 2.4, we may assume that ω satisfies (δ) for the first assertion. So, for the converse, first notice that Proposition 4.3 ensures that the spaces Ah (ω) are non-trivial for each h > 0. Lemma 5.8 implies that it suffices to show that for all 0 < k < h the space Ah (ω), but this precisely follows from Proposition 5.7 and the Hahn-Banach theorem. (cid:3) (ω) is dense in Ak Combining Lemma 5.6 and Theorem 5.9, we obtain, (ω)(C). Corollary 5.10. The continuous inclusion U(ω)(C) ֒→ A{ω}(R) is dense if ω satis- fies (ǫ)0 and (δ). In particular, one also obtains the continuous and dense inclusion A′ {ω}(R) ֒→ U ′ Our next goal is to construct analytic representations of elements of U ′ (ω)(R) which are globally defined, namely, everywhere outside some closed horizontal strip. The basic idea is to paste together the analytic representations obtained in Proposition 5.7 with the aid of a Mittag-Leffler procedure. We define Oh (ω) = [λ>0[b<h\R>b Ob,R,λ ω , O(ω) = [h>0 Oh (ω), P(ω) = [λ>0\R>0 P R,λ ω . We use the union and intersection notation to emphasize that we do not topologize the latter spaces. We need the following lemma. Lemma 5.11. Let 0 < L < R and let ω be a weight function satisfying (ǫ)0 and (δ). Then, U(ω)(C) is dense in P R,λ ω with respect to the norm · L,Hλ. Proof. By Theorem 5.9, it suffices to show that AR,∞ (ω) to the norm · L,Hλ. Let P ∈ P R,λ Pn(z) = P (z)ϕ(z/n) ∈ AR,∞ (ω) for n ≥ 1. We have ω and choose ϕ ∈ AR,∞ is dense in P R,λ ω with respect (ω) with ϕ(0) = 1. Set P − PnL,Hλ = sup z∈T L P (z)(1 − ϕ(z/n))e−ω(Hλx) ≤ AP L,λ sup z∈T L 1 − ϕ(z/n)e−ω(λx). Since ϕ(z/n) → 1, as n → ∞, uniformly on compact subsets of T R, this proves the result. (cid:3) Proposition 5.12. Let ω be a weight function satisfying (ǫ)0 and (δ). The sequence 0 −→ P(ω) −→ Oh (ω) bv −−→ (Ah (ω))′ −→ 0 is exact. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 19 Proof. The fact that P(ω) = ker bv is clear from Proposition 5.7. We show that the (ω))′. By Proposition 5.7 there are boundary value mapping is surjective. Let f ∈ (Ah 0 < b < h and λ > 0 such that for every n ∈ N there is Gn ∈ Ob,b+n+1,λ such that bv(Gn) = f . Corollary 5.5 and Lemma 5.6 yield Gn+1 − Gn = Pn ∈ P b+n+1,2H 4λ and thus, by Lemma 5.11, there is ϕn ∈ U(ω)(C) such that Pn − ϕnb+n,2H 5λ ≤ 2−n. Define ω ω Fn(z) = Gn(z) − ϕk + (Pk − ϕk), z ∈ T b,b+n. n−1Xk=0 ∞Xk=n Then, Fn ∈ Ob,b+n,2H 5λ F (z) := Fn(z) for z ∈ T b,b+n, is a well defined element of Oh , bv(Fn) = f , and Fn+1(z) = Fn(z) for z ∈ T b,b+n. Hence, (ω) such that bv(F ) = f . (cid:3) ω Summarizing, the following representation theorem should now be clear. Theorem 5.13. Let ω be a weight function satisfying (ǫ)0 and (δ). The sequence 0 −→ P(ω) −→ O(ω) bv −−→ U ′ (ω)(C) −→ 0 is exact. 5.3. Analytic representations of hyperfunctions of {ω}-type. We now turn our attention to the representation of A′ {ω}(R). In analogy to the previous subsection, we begin our study with the dual of Ah {ω}. For 0 < h < R we set Oh,R,λ {ω} = lim ←− b→h+ lim ←− L→R− lim ←− µ→λ+ Ob,L,µ ω , Oh,R {ω} = lim ←− λ→0+ Oh,R,λ {ω} , and P R,λ {ω} = lim ←− L→R− lim ←− µ→λ+ P L,µ ω , P R P R,λ {ω} . {ω} = lim ←− λ→0+ {ω} and P R Lemma 5.1 implies that, if ω satisfies (δ), the spaces Oh,R If ω satisfies (ζ) from Lemma 2.3, this is also true for Oh,R,λ the boundary value mapping {ω} {ω} are (F S)-spaces. {ω} . Furthermore, and P R,λ bv : Oh,R {ω} → (Ah {ω})′ is well defined and continuous. We need to the following density lemma. Lemma 5.14. Let 0 < h < R and 0 < λ < µ. Let ω be a weight function satisfying (ǫ)0, (δ), and (ζ). Then, U(ω)(C) is dense in Oh,R,λ {ω} with respect to the topology of Oh,R,λ {ω} . {ω} ∩ P R,µ Proof. By Theorem 5.9 it is enough to check that AR respect to the topology of Oh,R,λ ϕ(0) = 1. Set Pn(z) = P (z)ϕ(z/n) ∈ AR be arbitrary. For λ < ν0 < ν we have P −Pnb,L,ν = sup z∈T b,L {ω} . Let P ∈ Oh,R,λ P (z)(1−ϕ(z/n))e−ω(νx) ≤ P b,L,ν0 sup z∈T b,L (ω) is dense in Oh,R,λ {ω} with (ω) with (ω) for all n ≥ 1. Let h < b < L < R and ν > λ {ω} ∩ P R,µ {ω} and choose ϕ ∈ AR {ω} ∩ P R,µ 1−ϕ(z/n)e−(ω(νx)−ω(ν0x)). 20 A. DEBROUWERE AND J. VINDAS This shows the lemma because ω(νt) − ω(ν0t) → ∞, as t → ∞, and ϕ(z/n) → 1, as n → ∞, uniformly on compact subsets of T R. (cid:3) Proposition 5.15. Let 0 < h < R and let ω satisfy (ǫ)0 and (δ). The sequence 0 −→ P R {ω} −→ Oh,R {ω} bv −−→ (Ah {ω})′ −→ 0 is topologically exact. Moreover, for every bounded set B ⊂ (Ah set A ⊂ Oh,R {ω} such that bv(A) = B. {ω})′ there is a bounded Proof. In view of the open mapping theorem it suffices to show that the sequence is algebraically exact. Corollary 5.5 implies P R {ω} = ker bv. We now show that the boundary value mapping is surjective. By Lemma 2.3 we may assume that ω satisfies (ζ). Let f ∈ (Ah {ω})′ and define Xn = {F ∈ Oh+(1/n),R,1/n {ω} : bv(F ) = f on U(ω)(C)}, n ≥ 1. By Corollary 5.4 Xn is a non-empty closed subspace of Oh+(1/n),R,1/n and therefore a complete metrizable topological space (with respect to the relative topology). Consider the projective spectrum (Xn)n∈Z+ (where the linking mappings are just the inclusion mappings) and let X be its projective limit. It suffices to show that X is non-empty, which would be implied by Lemma 5.8 if we verify that Xn+1 is dense in Xn. Since, by Corollary 5.5, every F ∈ Xn can be written as F = G + P where G ∈ Xn+1 and P ∈ Oh+(1/n),R,1/n , the density of Xn+1 in Xn follows from Lemma 5.14. The second part follows from the general fact that for an exact sequence of Fr´echet spaces ∩P R,4H 4/n {ω} {ω} {ω} 0 −→ E −→ F T −→ G −→ 0, with E an (F S)-space, it holds that every bounded set B ⊂ G can be written as T (A) = B for some bounded set A ⊂ F [27, Lemma 26.13]. (cid:3) We now construct global analytic representations. Set Oh {ω} = lim ←− R→∞ Oh,R {ω}, O{ω} = lim ←− h→0+ Oh {ω}, P{ω} = lim ←− R→∞ P R {ω}. Note that all of the above spaces are (F S)-spaces if ω satisfies (δ), as follows from Lemma 5.1. Proposition 5.16. Let ω be a weight function satisfying (ǫ)0 and (δ). The sequence 0 −→ P{ω} −→ Oh {ω} bv −−→ (Ah {ω})′ −→ 0 is topologically exact. Moreover, for every bounded set B ⊂ (Ah set A ⊂ Oh {ω} such that bv(A) = B. {ω})′ there is a bounded Proof. Set Xn = P h+n sequence of short sequences {ω} and Yn = Oh,h+n {ω} for n ≥ 1. Consider the following projective HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 21 0 0 X1 X2 ... Y1 Y2 ... bv bv (Ah ω)′ (Ah ω)′ ... 0 0 By Proposition 5.15 every horizontal line is a topologically exact sequence. Moreover, by Lemma 5.11, Xn+1 is dense in Xn for all n ≥ 1. Hence, the Mittag-Leffler lemma [21, Lemma 1.3] yields the topological exactness of the sequence 0 −→ lim ←− n Xn = P{ω} −→ lim ←− n Yn = Oh {ω} bv −−→ (Ah {ω})′ −→ 0. The second statement follows from [27, Lemma 26.13] (cf. the end of the proof of Proposition 5.15). (cid:3) We then have, Theorem 5.17. Let ω be a weight function satisfying (ǫ)0 and (δ). The sequence 0 −→ P{ω} −→ O{ω} bv −−→ A′ {ω}(R) −→ 0 is topologically exact. Moreover, for every bounded set B ⊂ A′ set A ⊂ O{ω} such that bv(A) = B. {ω}(R) there is a bounded Proof. We now set Yn = O1/n sequence of short sequences {ω} and Zn = (A1/n {ω})′ for n ≥ 1. Consider the projective 0 0 P{ω} P{ω} ... Y1 Y2 ... bv bv Z1 Z2 ... 0 0 By Proposition 5.16 every horizontal line is a topologically exact sequence. Hence, the Mittag-Leffler lemma [21, Lemma 1.3] yields the topologically exact sequence 0 −→ P{ω} −→ lim ←− n Yn = O{ω} bv −−→ lim ←− n Zn = A′ {ω}(R) −→ 0. 22 A. DEBROUWERE AND J. VINDAS For the second statement, we apply again [27, Lemma 26.13]. (cid:3) 6. Support Silva introduced in [38] a useful notion of real support for ultrahyperfunctions. We extend such considerations to ultrahyperfunctions and hyperfunctions of fast growth via the analytic continuation properties of their analytic representations. We will use these ideas to establish a support separation theorem, which in particular gives that every (ultra)hyperfunction of fast growth can be written as the sum of two (ul- tra)hyperfunctions of fast growth having support in the positive and negative half-axis, respectively. Based upon this result, we shall study in Section 8 analytic representa- tions of ultradistributions of class ω of exponential type via Laplace transforms. 6.1. Real support of ultrahyperfunctions of (ω)-type. Let R = R ∪ {−∞, ∞} be the extended real line endowed with its usual topology (two-point compactification of R). For 0 < b < R and A ⊆ R we set T R(A) = (A ∩ R) + i(−R, R) and T b,R(A) = T b,R ∪ T R(A). Let ω be a weight function. Given K ⋐ R a proper closed set with non- empty interior we denote by Ob,R ω (K) the (B)-space of functions F ∈ O(T b,R(int K)) that satisfy F b,R ω,K := sup{F (z)e−ω(x) : z ∈ T b,R(int K)} < ∞. For an open subset Ω of R, we then define Oh,R ω (Ω) = lim ←− K ⋐Ω lim ←− b→h+ lim ←− L→R− Ob,L ω (K). We also write Ob,R ωλ (K) = Ob,R,λ ω (K), Oh,R ωλ (Ω) = Oh,R,λ ω (Ω), and F b,R ωλ,K = F b,R,λ ω,K = F b,R,λ K , F ∈ Ob,R,λ ω (K). Furthermore, we set Oh,R (ω) (Ω) = [λ>0 Oh,R,λ ω (Ω), O(ω)(Ω + iR) = [λ,h>0\R>h Oh,R,λ ω (Ω). Suppose that ω satisfies (δ) and (ǫ)0. Let f ∈ U ′ (ω)(C) and Ω ⊆ R be open. We say that f vanishes on Ω if there is F ∈ O(ω)(Ω + iR) such that bv(F ) = f . Theorem 5.13 implies that in such a case this property holds for all analytic representations of f (in O(ω)). Moreover, the proof of Proposition 5.12 shows that for f to vanish on Ω it suffices that there is F ∈ Oh,R (ω) (Ω), for some 0 < h < R, such that bv(F ) = f . We define the real support of f ∈ U ′ (ω)(C), denoted by suppR f , as the complement of the largest open subset of R on which f vanishes. Given a closed subset K ⊆ R, we write U ′ (ω)[K + iR] = {f ∈ U ′ (ω)(C) : suppR f ⊆ K}, the subspace of ultrahyperfunctions of (ω)-type with real support in K. When I is a closed interval of R, f ∈ U ′ (ω) (R\I) is such that bv(F ) = f , (ω)[I + iR], and F ∈ Oh,R HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 23 Cauchy's integral theorem yields hf, ϕi = −ZΓb(J) F (z)ϕ(z)dz, ϕ ∈ U(ω)(C), where h < b < R, J an interval in R such that I ⋐ J, and Γb(J) the boundary of T b(J). More generally, if K is a proper closed subset of R and J1, J2, . . . , Jn is a finite covering of K by open intervals of the extended real line such that their finite end points do not belong to K, and f = bv(F ) ∈ U ′ (ω) (R\K), we have the representation (ω)[K + iR] with F ∈ Oh,R (6.1) hf, ϕi = −(cid:18)ZΓb(J1) +ZΓb(J2) + · · · +ZΓb(Jn)(cid:19) F (z)ϕ(z)dz, for each ϕ ∈ U(ω)(C). The space U ′ (ω)[K + iR] consists of analytic functionals if K is a compact subset of R. Indeed, if S ⊆ C is a closed set, we write O[S] = lim −→ S⋐V O(V ). The Silva-Kothe-Grothendieck theorem [30] now yields U ′ (ω)[K + iR] = [R>0 O′[K + i[−R, R]] = O′[K + iR], which is in fact a result due to Silva [38]. Our next goal is to show a support separation theorem for U ′ (ω)(C). For a, b ∈ R, we employ the special notations U ′ (ω),a+ = U ′ (ω)[[a, ∞] + iR], U ′ (ω),b− = U ′ (ω)[[−∞, b] + iR]. Theorem 6.1. Let −∞ < a ≤ b < ∞ and let ω be a weight function satisfying (δ) and (ǫ)0. The sequence 0 −→ U ′ (ω)[[a, b] + iR] −→ U ′ (ω),a+ ⊕ U ′ (ω),b− λ−−→ U ′ (ω)(C) −→ 0 is exact, where λ((f1, f2)) = f1 − f2. We need some extra notions to prove Theorem 6.1. Let Ω ⊆ R be open and define ω(Ω) as the (B)-space of analytic functions ϕ ∈ O(T h(Ω)) such that Ah kϕkh ω,Ω := sup z∈T h(Ω) ϕ(z)eω(x) < ∞. Set Ah ωλ(Ω) = Ah,λ ω (Ω) and kϕkh ωλ,Ω = kϕkh,λ ω,Ω = kϕkh,λ Ω , ϕ ∈ Ah,λ ω (Ω). We have the following refinement of Proposition 5.2. 24 A. DEBROUWERE AND J. VINDAS Proposition 6.2. Let 0 < k < b < h < R, let U ⋐ V ⋐ Ω be open subsets of R, and let ω, σ, and κ be three weight functions satisfying (5.1) and (5.2). Furthermore, suppose there is P ∈ O(T R) satisfying C1eω(x) ≤ P (z) ≤ C2eσ(x), z ∈ T R, for some C1, C2 > 0. Then, for every f ∈ (Ak bv(F ) = f on Ah κ(Ω). ω(U))′ there is F ∈ Ob,R σ (R\V ) such that In view of Proposition 3.2, we have the following corollary. Corollary 6.3. Let 0 < k < b < h < R, let U ⋐ V ⋐ Ω be open subsets of R, and let ω be a weight function satisfying (δ) and (ǫ)0. For every f ∈ (Ak,λ ω (U))′ there is F ∈ Ob,R,2H 2λ (R\V ) such that bv(F ) = f on Ah,2H 3λ (Ω). ω ω The proof of Theorem 6.1 is based on the following functional analytic method for checking the surjectivity of linear continuous mappings. Lemma 6.4. Let E1, E2, and E be semi-reflexive (DF )-spaces and let ρj : E → Ej, j = 1, 2, be continuous linear mappings. Then, 1, x′ 2 → E′ : (x′ 2) → ρt 1) + ρt 1 × E′ 2(x′ 2) ι : E′ 1(x′ is surjective if and only if ρ : E → E1 × E2 : x → (ρ1(x), ρ2(x)) is injective and has closed range. Proof. Since E1, E2, and E are semi-reflexive we may identify the transpose of ι with ρ. A continuous linear mapping between two Fr´echet spaces is surjective if and only if its transpose is injective and has weakly closed range [41, Thm. 37.3]. The result now follows from the fact that the closed convex sets of a Hausdorff locally convex space are the same for all topologies compatible with the dual pairing. (cid:3) Proof of Theorem 6.1. Theorem 5.13 implies that U ′ (ω)[[a, b] + iR] = U ′ (ω),a+ ∩ U ′ (ω),b−. It remains to show that λ is surjective. By Lemma 2.3 we may assume that ω satisfies (ζ). Let f ∈ U ′ (ω),b− such that f = f1 + f2. Choose h, λ > 0 such that f can be extended to an element of (Ah,λ ω )′ (we also write f for this extension). Define (ω)(C), we show that there are f1 ∈ U ′ (ω),a+ and f2 ∈ U ′ X h,λ = lim −→ µ→λ+ lim −→ k→h+ Ak,µ ω , and X h,λ a+ = lim −→ µ→λ+ lim −→ k→h+ lim −→ ε→0+ Ak,µ ω ((a − ε, ∞]), X h,λ b− = lim −→ µ→λ+ lim −→ k→h+ lim −→ ε→0+ Ak,µ ω ([−∞, b + ε)). Condition (ζ) implies that the three above spaces are (DF S)-spaces. Observe that f ∈ (X h,λ)′. Let g ∈ (X h,λ a+ )′ and choose R > h. By Corollary 6.3, we have that for every ε > 0 and k > h there is Gε,k = G ∈ Ok,R,4H 2λ ([−∞, a − ε]) such that bv(G) = g. By a similar Mittag-Leffler procedure as in the proof of Proposition 5.12, one can now show that there is G ∈ Oh,R,8H 7λ ([−∞, a)) such that bv(G) = g. Likewise, it holds that ω ω HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 25 b− )′ there is G ∈ Oh,R,8H 7λ for every g ∈ (X h,λ suffices to show that the mapping a+ )′ × (X h,λ (X h,λ ω b− )′ → (X h,λ)′ : (f1, f2) → f1 + f2 ((b, ∞]) such that bv(G) = g. Therefore, it is surjective; but the latter is a consequence of Lemma 6.4. (cid:3) 6.2. Support of hyperfunctions of {ω}-type. We now define the support of hy- perfunctions of {ω}-type. Given a compact subset K ⊂ R, we write A[K] = O[K] for the space of germs of analytic functions on K, so that A′[K] is the space of analytic functionals on K. Let ω be a weight function satisfying (δ) and (ǫ)0. Given a proper open subset Ω of R, we define O{ω}(Ω) = lim ←− λ→0+ an (F S)-space. We say that f ∈ A′ {ω}(R) vanishes on Ω if there is F ∈ O{ω}(Ω) such that bv(F ) = f . In view of Theorem 5.17, this definition is independent of the chosen analytic representation. We can therefore define the support of f , denoted by supp f , as the complement of the largest open set where f vanishes. lim ←− h→0+ lim ←− R→∞ (Ω), ω Oh,R,λ Given a closed subset K in R, we set A′ {ω}[K] = {f ∈ A′ {ω}(R) : supp f ⊆ K}, and for a, b ∈ R, we write A′ {ω},a+ = A′ {ω}[[a, ∞]], A′ {ω},b− = A′ {ω}[[−∞, b]]. If f = bv(F ) ∈ A′ {ω}[K] with F ∈ O{ω}(R\K) and J1, J2, . . . , Jn is a finite covering of K by open intervals of the extended real line (such that their finite end points do not belong to K), Cauchy's integral theorem also gives the contour integral representation (6.1), where b > 0 depends on ϕ (concretely, for ϕ ∈ Ab {ω}). In particular, for a compact set in K in R, A′ {ω}[K] = A′[K], as follows from the Silva-Kothe-Grothendieck theorem. In case K is unbounded, we can also represent A′ {ω}[K] as a dual space. We define the (DFS)-space A{ω}[K] = lim −→ λ→0+ lim −→ h→0+ lim −→ K ⋐Ω Ah,λ ω (Ω). Corollary 6.3 and the same method as in Section 5.3 yield: Theorem 6.5. Let K be a closed subset of R and let ω be a weight function satisfying (δ) and (ǫ)0. Then, (A{ω}[K])′ = A′ {ω}[K]. Furthermore, the sequence 0 −→ P{ω} −→ O{ω}(R \ K) bv−−−→ (A{ω}[K])′ −→ 0 is topologically exact. We also have the ensuing support separation theorem. 26 A. DEBROUWERE AND J. VINDAS Theorem 6.6. Let −∞ < a ≤ b < ∞ and let ω be a weight function satisfying (δ) and (ǫ)0. The sequence 0 −→ A′ {ω}[[a, b]] −→ A′ {ω},a+ ⊕ A′ {ω},b− λ−−→ A′ {ω}(R) −→ 0 is topologically exact, where λ((f1, f2)) = f1 − f2. Proof. Theorem 5.17 implies that A′ {ω}[[a, b]] = A′ {ω},a+ ∩ A′ {ω},b−. The surjectivity of λ follows from Lemma 6.4 and Theorem 6.5. (cid:3) 7. Spaces of (ultra)hyperfunctions defined via subadditive weight functions In this section we briefly indicate how the results from Sections 4 -- 6 can be extended to include spaces defined in terms of subadditive weight functions. Since the theory and methods are completely analogous to those already developed, we shall omit all proofs. These ideas will be applied in Section 8 to the study of analytic representations of ultradistributions of exponential type and Laplace transforms. 7.1. Subadditive weight functions. We collect here a number of properties of the weight functions that we shall employ in the rest of the paper. Let ω be a weight function (not necessarily satisfying (2.1)) such that ω(0) = 0 (this will be assumed from now on). We are interested in the following conditions [1]: (α) ω(t1 + t2) ≤ ω(t1) + ω(t2), t1, t2 ≥ 0, (γ) ω(t) ≥ c log(1 + t) + a, for some a ∈ R and c > 0. In the sequel we shall refer to condition (2.1) as (γ)0. We set λω(t) = λω(t). Let σ be another weight function, ω and σ are said to be ∗-equivalent, denoted by ω ≍ σ, if ω(t) = O(σ(t)) and σ(t) = O(ω(t)) (as t → ∞), or equivalently, λ1ω(t) − C1 ≤ σ(t) ≤ λ2ω(t) + C2, t ≥ 0, for some λ1, λ2, C1, C2 > 0. If ω and σ both satisfy (α) and (δ), then ω ≍ σ if and only if ω ∼ σ. We point out that subadditivity, that is, condition (α), always yields [23, p. 240] the existence of the limit lim t→∞ ω(t) t . Consequently, we either have ω(t) ≍ t or ω(t) = o(t). We shall need the following result of Petzsche and Vogt which allows one to replace a weight function by one enjoying better regularity properties. Lemma 7.1. [34, Prop. 1.2] Let ω be a weight function satisfying (α) and (γ). Then, there is another weight function σ with ω ≍ σ such that either σ(t) = t or σ satisfies the following conditions: (i) σ ∈ C ∞(0, ∞) and limt→0+ σ′(t) = ∞, (ii) σ is strictly concave, (iii) lim t→∞ σ′(t) = 0, HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 27 (iv) lim inf t→∞ σ′′(t)t2 > 0. If ω satisfies (γ)0, then condition (iv) may be replaced by the stronger property (iv)′ lim t→∞ σ′′(t)t2 = ∞. The Young conjugate of ω is defined as ω∗(s) = sup t≥0 (ω(t) − ts), s > 0, and is a convex non-increasing function. We set ω∗(s) = ω∗(s) for s ∈ R, s 6= 0. If ω(t) = o(t), then ω∗(s) < ∞ for all s > 0. Clearly, we have (λω)∗(s) = λω∗(cid:16) s λ(cid:17) , s > 0. 7.2. Test function spaces. Given a weight function ω, we set Ah λω = Ah,λ ω and kϕkh λω = kϕkh,λ ω , ϕ ∈ Ah,λ ω . Our basic spaces of entire and analytic functions are now defined as U(ω)(C) = lim ←− λ→∞ lim ←− h→∞ Ah,λ ω , A{ω}(R) = lim −→ λ→0+ lim −→ h→0+ Ah,λ ω . If ω(t) → ∞, as t → ∞, then Lemma 4.1 yields that U(ω)(C) is an (F S)-space while A{ω}(R) is a (DFS)-space. Let σ be another weight such that ω ≍ σ, then U(ω)(C) = U(σ)(C) and A{ω}(R) = A{σ}(R), topologically. If ω satisfies both conditions (α) and (δ), then U(ω)(C) = U(ω)(C) and A{ω}(R) = A{ω}(R), topologically. The elements of the dual spaces U ′ {ω}(R) are called ultrahyperfunctions of (ω)-type and hyperfunctions of {ω}-type, respectively. Observe that U(log(1+t))(C) = U(C) is the test function space for the Silva space of tempered ultrahyperfunctions [18, 16, 38] (called tempered ultradistributions there). If ω satisfies (α) and (γ) ((γ)0), we have the continuous and dense inclusions S (p!) {p!} = A{t}(R) = A{t}(R) → A{ω}(R) (density follows from Theorem 7.7 below). (p!) = U(t)(C) = U(t)(C) → U(ω)(C) and S {p!} (ω)(C) and A′ We mention that if a (log-convex) weight sequence Mp satisfies (M.2) and (M.4) Mp p! p! (cid:19)c ⊆(cid:18) Mp , Petzsche and Vogt have shown [34, Sect. 5] that there is a weight function ω satisfying (α), (γ)0, and ω ≍ M; under these circumstances, we thus have U(Mp)(C) = U(ω)(C) and A{Mp}(R) = A{ω}(R), topologically. Furthermore, they proved, under (M.2), that (M.4) is equivalent to the so-called Rudin condition: (M.4)′′ max q≤p (cid:18)Mq q! (cid:19) 1 q ≤ A(cid:18)Mp p! (cid:19) 1 p , p ∈ N, for some A > 0. Observe that strong non-quasianalyticity (i.e., Komatsu's condition (M.3) [21]) auto- matically yields (M.4), as shown by Petzsche [35, Prop. 1.1]. Theorem 7.2. Let ω be a weight function satisfying (γ)0. Then, U(ω)(C) (A{ω}(R)) is non-trivial if and only if ω satisfies (ǫ)0 ((ǫ)∞). 28 A. DEBROUWERE AND J. VINDAS Proof. The conditions are necessary by Proposition 4.2. In the Roumieu case sufficiency is a consequence of Proposition 3.2. For the Beurling case we notice that, by Lemma 2.4, we may assume that ω satisfies (δ). Hence, U(ω)(C) ⊆ U(ω)(C) and the result follows from Theorem 5.9. (cid:3) Remark 7.3. In [26] Langenbruch studied the space A{ω}(R) for weight functions ω satisfying (γ)0 and (7.1) lim sup t→∞ ω(t + 1) ω(t) < ∞. He showed that the space A{ω}(R) is (tamely) isomorphic to the strong dual of a finite type power series space [26, Thm. 4.4] and thus, in particular, is non-trivial. Under the extra assumption (7.1), Theorem 7.2 in the Roumieu case is therefore due to Langenbruch. We mention that one can readily show that condition (7.1) implies (ǫ)∞ (without passing through Theorem 7.2), whereas there exist weight functions satisfying (ǫ)∞ and (γ)0 but not (7.1). Finally, let us discuss the Fourier transform on our test function spaces. We set 1,λω(R) = Kh,λ 1,ω(R) and Kh ρh λω(ψ) = ρh,λ ω (ψ), ψ ∈ Kh,λ 1,ω(R); define then K1,(ω)(R) = lim ←− λ→∞ lim ←− h→∞ Kh,λ 1,ω(R), K1,{ω}(R) = lim −→ λ→0+ 1,(ω)(R) and K′ lim −→ h→0+ Kh,λ 1,ω(R). In analogy to the terminology from Section 4, we call K′ 1,{ω}(R) the spaces of ultradistributions of class (ω) of exponential type and ultradistributions of class {ω} of infra-exponential type, respectively. If ω satisfies (α), (γ), and is non- quasianalytic, then we have the continuous and dense inclusions D(ω)(R) ֒→ K1,(ω)(R) ֒→ K′ (ω)(R) (see [1] for the definition of the Beurling-Bjorck space D(ω)(R)). In particular, K1,(log(1+t))(R) = K1(R) is the space of exponentially rapidly decreasing smooth functions [18] and we have that D(R) ֒→ K1(R) ֒→ K′ 1,(ω)(R) ֒→ D′ 1(R) ֒→ D′(R). Proposition 7.4. Let ω be a weight function satisfying (α) and (γ) ((γ)0). The Fourier transform is a topological isomorphism from U(ω)(C) (A{ω}(R)) onto K1,(ω)(R) (K1,{ω}(R)). Proof. This can be shown in the same way as Proposition 4.4. (cid:3) The Fourier transform from K′ 1,(ω)(R) (K′ 1,{ω}(R)) onto U ′ (ω)(C) (A′ {ω}(R)) is there- fore well defined via duality. 7.3. Boundary values. Given 0 < b < R, we set Ob,R Furthermore, we define λω = Ob,R,λ ω and P R λω = P R,λ ω . O(ω) = [λ,h>0\R>h Oh,R,λ ω , P(ω) = [λ>0\R>0 P R,λ ω , HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 29 and O{ω} = lim ←− λ→0+ lim ←− h→0+ lim ←− R→∞ Oh,R,λ ω , P{ω} = lim ←− λ→0+ lim ←− R→∞ P R,λ ω . If ω(t) → ∞, as t → ∞, Lemma 5.1 implies that O{ω} and P{ω} are (F S)-spaces. The boundary value mappings bv : O(ω) → U ′ (ω)(C), bv : O{ω} → A′ {ω}(R) are well defined. In the Roumieu case, this mapping is continuous. Taking Remark 3.3 into account, Propositions 5.2 and 5.3 yield the following corollaries. Corollary 7.5. Let 0 < k < b < h < R and let ω be a weight function satisfying (α). (i) If ω satisfies (γ), then for every f ∈ (Ak,λ ω )′ there is F ∈ Ob,R,4λ+2c−1 ω bv(F ) = f on Ah,4λ+4c−1 ω . (ii) If ω satisfies (γ)0, then for every f ∈ (Ak,λ ω )′ there is F ∈ Ob,R,4λ ω bv(F ) = f on Ah,8λ ω . such that such that Corollary 7.6. Let 0 < b < h < R and let ω be a weight function satisfying (α). is such that bv(F ) = 0 on Ah,λ+2c−1 (i) If ω satisfies (γ) and F ∈ Ob,R,λ ω ω , then F ∈ P R,4λ+8c−1 ω . (ii) If ω satisfies (γ)0 and F ∈ Ob,R,λ ω P R,8λ ω . is such that bv(F ) = 0 on Ah,2λ ω , then F ∈ Using exactly the same technique as in Sections 5.2 and 5.3 and applying Corollaries 7.5 and 7.6 (instead of Corollaries 5.4 and 5.5), one can show the ensuing theorem. We leave the details to the reader. Theorem 7.7. Let ω be a weight function satisfying (α). (i) If ω satisfies (γ), then the sequence 0 −→ P(ω) −→ O(ω) bv −−→ U ′ (ω)(C) −→ 0 is exact. (ii) If ω satisfies (γ)0, then the sequence 0 −→ P{ω} −→ O{ω} bv −−→ A′ {ω}(R) −→ 0 is topologically exact. Moreover, for every bounded set B ⊂ A′ bounded set A ⊂ O{ω} such that bv(A) = B. {ω}(R) there is a Let us briefly discuss the notion of support in this new setting. Define Oh,R λω (Ω) = Oh,R,λ ω (Ω), and O(ω)(Ω + iR) = [λ,h>0\R>h Oh,R,λ ω (Ω), O{ω}(Ω) = lim ←− λ→0+ lim ←− h→0+ lim ←− R→∞ Oh,R,λ ω (Ω). We suppose that ω satisfies (α) and (γ) ((γ)0). Vanishing of f ∈ U ′ {ω}(R)) on an open set Ω ⊆ R means that there is F ∈ O(ω)(Ω + iR) (F ∈ O{ω}(Ω)) such that bv(F ) = f . The definition of suppR f (supp f ) should now be clear. (ω)(C) (f ∈ A′ 30 A. DEBROUWERE AND J. VINDAS We shall adopt the same kind of notations as in Section 6 for the rest of the spaces defined there by simply replacing ω by ω. Furthermore, all results from that section remain valid in our new context; in fact, due to Remark 3.3 and the general formu- lation of Proposition 6.2, same proofs apply here. In particular, we state the support separation theorem for future reference. Theorem 7.8. Let −∞ < a ≤ b < ∞ and let ω be a weight function satisfying (α). (i) If ω satisfies (γ), then the sequence 0 −→ U ′ (ω)[[a, b] + iR] −→ U ′ (ω),a+ ⊕ U ′ (ω),b− λ−−→ U ′ (ω)(C) −→ 0 is exact, where λ((f1, f2)) = f1 − f2. (ii) If ω satisfies (γ)0, then the sequence 0 −→ A′ {ω}[[a, b]] −→ A′ {ω},a+ ⊕ A′ {ω},b− λ−−→ A′ {ω}(R) −→ 0 is topologically exact. 8. Boundary values of holomorphic functions in spaces of ultradistributions of exponential type 1,(ω)(R) and K′ This last section is devoted to boundary values of holomorphic functions in the spaces K′ 1,{ω}(R). Our main result is a representation theorem for these two spaces as quotients of certain spaces of analytic functions (Theorem 8.4). Our arguments rely on the study of the Laplace transform of (ultra)hyperfunctions of fast growth supported in a proper interval of R and ideas from the theory of almost analytic extensions [4, 34]. Let us fix some notation. We shall write for complex variables ζ = ξ + iη, z = x + iy, and ∂ = ∂ ∂ ¯z = 1 2(cid:18) ∂ ∂x + i ∂ ∂y(cid:19) . The following condition for weight functions plays a role below: (NA) ω(t) = o(t). 8.1. Laplace transforms. We discuss here how to define the (Fourier-)Laplace trans- form of ultrahyperfunctions of (ω)-type and hyperfunctions of {ω}-type with support in a proper closed interval I of R. We assume that ω satisfies (α) and (γ). Given f ∈ U ′ (ω)[I + iR], we define its Laplace transform as (8.1) L{f ; ζ} = − F (z)eizζdz, 1 2πZΓb(J) for ζ ∈ C in a suitable domain to be specified below and where F ∈ Oh,R (ω)(R\I) is an analytic representation of f , that is, bv(F ) = f , h < b < R, and J an interval in R such that I ⋐ J. The definition is clearly independent of the chosen representative of F . In the rest of our discussion, we distinguish three cases. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 31 Case 1: a bounded interval I = [−a, a], 0 ≤ a < ∞. In this case (8.1) is defined for all ζ ∈ C. In fact, L{f ; ζ} is an entire function that satisfies the estimate: there is h > 0 such that for every ε > 0 (8.2) L{f ; ζ}e−(a+ε)η−(h+ε)ξ < ∞. sup ζ∈C Conversely, it is well known [30, Thm. 2.5.2] that if an entire function G satisfies the bound (8.2), then it is the Laplace transform of an analytic functional; more precisely, there is f ∈ O′[I + i[−h, h]] ⊂ U ′ (ω)[I + iR] such that G(ζ) = L{f ; ζ}. If ω satisfies (γ)0 and f ∈ A′ {ω}[I] = A′[I], then L{f ; ζ} satisfies (8.2) for every h, ε > 0, and the converse holds true: if an entire function G satisfies (8.2) for every h, ε > 0, then there is f ∈ A′[I] such that G(ζ) = L{f ; ζ}. Case 2: a left-bounded interval I = [−a, ∞], 0 ≤ a < ∞. As was pointed out in Subsection 7.1, either ω satisfies (NA) or ω(t) ≍ t. First assume that ω satisfies (NA). Then, L{f ; ζ} is analytic on the upper half-plane Im ζ = η > 0 and satisfies the bound: there are λ, h > 0 such that for every ε > 0 (8.3) L{f ; ζ}e−(a+ε)η−(h+ε)ξ−λω∗(η/λ) < ∞. sup η>0 Moreover, the Laplace transform has as boundary value on R the inverse Fourier trans- form of f , namely, lim η→0+ L{f ; · + iη} = g, in K′ 1,(ω)(R), λ, h, ε > 0 and where f =bg. If ω satisfies (γ)0 and f ∈ A′ L{f ; · + iη} = g, lim η→0+ in K′ 1,{ω}(R). {ω}[I], then L{f ; ζ} satisfies (8.3) for every Next, assume ω(t) ≍ t, so that K1,(ω)(R) = U(t)(C) and K1,{ω}(R) = A{t}(R) are invariant under the Fourier transform. In the Beurling case there are λ, h > 0 such that L{f ; ζ} is a holomorphic function on the half-plane {ζ = ξ + iη : η > λ} and satisfies (8.4) If we set L{f ; ζ}e−(a+ε)η−(h+ε)ξ < ∞. sup η>λ then bv(G) = g in U ′ Roumieu case L{f ; ζ} is analytic on the upper half-plane and satisfies (8.4) for every λ, h, ε > 0. Furthermore, bv(G) = g in A′ 0, η < −λ, L{f ; ζ}, η > λ, G(ζ) = (t)(C) (in the sense of Section 7.3), where again f = bg. In the G(ζ) = {t}(R) where now L{f ; ζ}, η > 0, 0, η < 0. Case 3: a right-bounded interval I = [−∞, a], 0 ≤ a < ∞. Here the treatment is completely analogous to case 2 but with the upper half-planes replaced by lower ones. 32 A. DEBROUWERE AND J. VINDAS 8.2. Analytic representations of ultradistributions of exponential type. Our next aim is to study boundary values of holomorphic function in the spaces K′ 1,(ω)(R) and K′ 1,{ω}(R). For it, we need the following modified version of a construction of almost analytic extensions by Petzsche and Vogt [34, Prop. 2.2]. Lemma 8.1. Let 0 < k < h and let ω be a weight function satisfying the conditions (i)-(iv) of Lemma 7.1, and let σ be another weight function such that Z ∞ 0 teω(t)−σ(t)dt < ∞. Then, for every ϕ ∈ Ah σ there is Ψ ∈ C ∞(C) with ΨR = bϕ such that ωe−kξ(ω∗)′′(η)e−ω∗(η), ζ ∈ C\R, ∂Ψ(ζ) ≤ kϕkh (8.5) (8.6) and (8.7) where Ψ(ζ) ≤ Ckϕkh σe−kξ, ζ ∈ C, C = 2Z ∞ 0 eω(t)−σ(t)dt. Proof. The assumptions on ω imply that ω′ is a smooth bijection on (0, ∞). Set H = (ω′)−1 ∈ C ∞(0, ∞) and observe that ω∗(s) = ω(H(s)) − sH(s), s > 0. Differentiation shows that (ω∗)′(s) = −H(s), s > 0. We set H(s) = H(s) for s ∈ R, s 6= 0. Furthermore, since ω is concave and increasing, we have that (8.8) tω′(t) ≤ ω(t), t ≥ 0. The rest of the proof is based on the following representation of the Fourier transform of ϕ [18, p. 167] bϕ(ξ) = For ζ ∈ C\R we define Ψ(ζ) = ϕ(x + ik)e−i(x+ik)ξdx, ξ ≤ 0, ϕ(x − ik)e−i(x−ik)ξdx, ξ ≥ 0. ϕ(x + ik)e−i(x+ik)ζdx, ξ ≤ 0, ϕ(x − ik)e−i(x−ik)ζdx, ξ ≥ 0, −∞ −∞ Z ∞ Z ∞ Z H(η) Z H(η) −H(η) −H(η)   HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 33 and Ψ(ξ) = bϕ(ξ) for ξ ∈ R. Clearly, Ψ ∈ C ∞(C\R). Employing (8.5) and (8.8) one can readily prove that Ψ ∈ C 1(C) with ∂Ψ(ξ) = 0, ξ ∈ R, and thus Ψ ∈ C ∞(C). We now show (8.6). Let ζ ∈ C\R and assume ξ ≤ 0, the case ξ ≥ 0 can be treated similarly. Using the remarks given at the beginning of the proof, we have ∂Ψ(ζ) ≤ 1 2 H ′(η)(ϕ(H(η) + ik)e−i(H(η)+ik)ζ + ϕ(−H(η) + ik)e−i(−H(η)+ik)ζ ) ≤ kϕkh = kϕkh ωe−kξH ′(η)e−ω(H(η))+H(η)η ωe−kξ(ω∗)′′(η)e−ω∗(η). It remains to establish (8.7). By continuity, it suffices to show this for ζ ∈ C\R. We assume that ξ ≤ 0, the case ξ ≥ 0 is similar. Then, Ψ(ζ) ≤ 2kϕkh e−σ(x)+xηdx. σe−kξZ H(η) 0 Notice that for 0 < x < H(η) we have ω′(x) > η. Therefore, applying (8.8), we obtain that Z H(η) 0 e−σ(x)+xηdx ≤Z H(η) 0 e−σ(x)+xω′(x)dx ≤Z ∞ 0 eω(x)−σ(x)dx, (cid:3) which completes the proof. Corollary 8.2. Let ω be a weight function satisfying (α) and (NA). (i) If ω satisfies (γ), then for every ψ ∈ K1,(ω)(R) and every λ, h > 0 there is Ψ ∈ C ∞(C) with ΨR = ψ satisfying the bounds (8.9) ∂Ψ(ζ)ehξ+λω∗(η/λ) < ∞, sup ζ∈C\R Ψ(ζ)ehξ < ∞. sup ζ∈C (ii) If ω satisfies (γ)0, then for every ψ ∈ K1,{ω}(R) there are λ, h > 0 and Ψ ∈ C ∞(C) with ΨR = ψ satisfying the inequalities (8.9). Proof. We may assume that ω satisfies the conditions (i)-(iv) from Lemma 7.1 (condi- tions (i)-(iv)′ in the Roumieu case). Petzsche and Vogt have shown [34, Lemma 2.3] that there is ε > 0 such that (ω∗)′′(s)e−εω∗(s) < ∞; sup s>0 if ω additionally satisfies (iv)′, the latter inequality holds for every ε > 0. Therefore the result follows by applying Lemma 8.1 to the weight µω, for a suitable µ > 0, and ϕ = F −1(ψ). (cid:3) The next result gives a sufficient condition for the existence of boundary values of 1,(ω)(R) and analytic functions in the ultradistribution spaces of exponential type K′ K′ 1,{ω}(R). Proposition 8.3. Let ω be a weight function satisfying (α) and (NA). 34 A. DEBROUWERE AND J. VINDAS (i) If ω satisfies (γ), then every G ∈ O(T R + ) satisfying G(ζ)e−hξ−λω∗(η/λ) < ∞, (8.10) sup ζ∈T R + for some λ, h > 0, has boundary values in K′ K′ 1,(ω)(R), such that 1,(ω)(R), that is, there is g ∈ g = lim η→0+ G(· + iη), in K′ 1,(ω)(R). (ii) If ω satisfies (γ)0, then every G ∈ O(T R every λ, h > 0, has boundary values in K′ + ) satisfying the inequality (8.10) for 1,{ω}(R). Proof. We only treat the Roumieu case, the Beurling case is completely analogous. Due to the Banach-Steinhaus theorem and the fact that the space K1,{ω}(R) is Montel, it suffices to show that η→0+Z ∞ lim −∞ G(ξ + iη)ψ(ξ)dξ exists and is finite for every ψ ∈ K1,{ω}(R). By Corollary 8.2 there is Ψ ∈ C ∞(C) with ΨR = ψ satisfying the inequalities (8.9) for some λ, h > 0. Choose 0 < L < R and fix 0 < η < R − L. Applying the Stokes theorem to the rectangle (−N, N) + i(0, L), N > 0, and the function eG(ξ + iv) = G(ξ + i(η + v))Ψ(ξ + iv) we obtain that G(ξ + i(η + L))Ψ(ξ + iL)dξ The second and third integral on the right hand side tend to zero, as N → ∞. Hence, 0 −N −N Z N G(ξ + iη)ψ(ξ)dξ =Z N −Z L G(N + i(η + v))Ψ(N + iv)dv +Z L + 2iZ N −NZ L Z ∞ G(ξ + iη)ψ(ξ)dξ =Z ∞ −∞ −∞ 0 G(ξ + i(η + v))∂Ψ(ξ + iv)dvdξ. 0 + 2iZ ∞ −∞Z L G(ξ + iη)ψ(ξ)dξ =Z ∞ −∞ 0 η→0+Z ∞ lim −∞ By Lebesgue's dominated convergence theorem, we obtain that G(−N + i(η + v))Ψ(−N + iv)dv G(ξ + i(η + L))Ψ(ξ + iL)dξ G(ξ + i(η + v))∂Ψ(ξ + iv)dvdξ. G(ξ + iL)Ψ(ξ + iL)dξ + 2iZ ∞ −∞Z L 0 G(ξ + iv)∂Ψ(ξ + iv)dvdξ. (cid:3) We now have all necessary tools to express the spaces K′ 1,{ω}(R) as quotients of spaces of analytic functions. Let a ≥ 0 and let ω be a weight function 1,(ω)(R) and K′ HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 35 satisfying (α), (γ), and (NA). We introduce the space Oexp,h,a G ∈ O(C\R) that satisfy ω,λ (C\R) consisting of all G(ζ)e−(a+ε)η−(h+ε)ξ−λω∗(η/λ) < ∞, sup ζ∈C\R for every ε > 0. We set O(exp),a (ω) (C\R) = [λ,h>0 Oexp,h,a ω,λ (C\R), O{exp},a {ω} (C\R) = \λ,h>0 Oexp,h,a ω,λ (C\R). We define the boundary value mapping as follows bv : O(exp),a (ω) (C\R) → K′ 1,(ω)(R), G 7→ lim η→0+ G(· + iη) − G(· − iη) Proposition 8.3 guarantees that bv is well defined. Moreover, if ω satisfies (γ)0, then bv(O{exp},a 1,{ω}(R). We also write Oexp,h,a(C) for the space of all entire functions G ∈ O(C) that satisfy (C\R)) ⊆ K′ {ω} G(ζ)e−(a+ε)η−(h+ε)ξ < ∞, sup ζ∈C for every ε > 0 and set O(exp),a(C) = [h>0 Oexp,h,a(C), O{exp},a(C) = \h>0 Oexp,h,a(C). Theorem 8.4. Let a ≥ 0 and let ω be a weight function satisfying (α) and (NA). (i) If ω satisfies (γ), then the sequence 0 −→ O(exp),a(C) −→ O(exp),a (ω) (C\R) bv −−→ K′ 1,(ω)(R) −→ 0 is exact. (ii) If ω satisfies (γ)0, then the sequence 0 −→ O{exp},a(C) −→ O{exp},a {ω} (C\R) bv −−→ K′ 1,{ω}(R) −→ 0 is exact. Proof. We only give the proof in the Roumieu case, the Beurling case is similar. The fact that ker bv = O{exp},a(C) follows from Theorem 5.17 (with ω(t) = t). So, we only need to show that the boundary value mapping is surjective. Let g ∈ K′ 1,{ω}(R) and {ω}(R). By Theorem 7.8 there are f+ ∈ A′ {ω},(−a)+ and f− ∈ A′ {ω},a− From the discussion in Subsection 8.1 on the Laplace transform, it is clear that G ∈ O{exp},a (cid:3) (C\R) and bv(G) = g. {ω} such that f = f+ − f−. Define set f = bg ∈ A′ G(ζ) = L{f+; ζ}, η > 0, L{f−; ζ}, η < 0. 36 A. DEBROUWERE AND J. VINDAS As an application of Theorem 8.4, we characterize in a precise fashion those an- alytic functions on the upper half-plane that are the Laplace transform of an (ul- tra)hyperfunction of ω-type supported on a fixed half-axis. The following result is a theorem of Paley-Wiener type. Theorem 8.5. Let a ≥ 0, let ω be a weight function satisfying (α) and (NA), and suppose that G is an analytic function on the upper half-plane. (i) If ω satisfies (γ), then G satisfies the estimate (8.11) G(ζ)e−(a+ε)η−(h+ε)ξ−λω∗(η/λ) < ∞, sup η>0 for some h, λ > 0 and for every ε > 0 if and only if there is f ∈ U ′ that G(ζ) = L{f ; ζ}. (ω),(−a)+ such (ii) If ω satisfies (γ)0, then G satisfies (8.11) for every h, λ, ε > 0 if and only if there is f ∈ A′ {ω},(−a)+ such that G(ζ) = L{f ; ζ}. Proof. We only treat the Roumieu case, the Beurling case is analogous. It has al- ready been pointed out in Subsection 8.1 that the Laplace transform of an element of A{ω},(−a)+ satisfies the required bounds. Conversely, let G be an analytic function on the upper half-plane satisfying (8.11) for every h, λ, ε > 0. By Proposition 8.3 there is g ∈ K′ 1,{ω}(R) such that g = lim η→0+ G(· + iη), in K′ 1,{ω}(R). {ω}(R). By Theorem 7.8, there are f+ ∈ A′ {ω},(−a)+ and f− ∈ A′ {ω},a− such that f = f+ − f−. Define Let f =bg ∈ A′ L{f+; ζ}, η > 0, eG(ζ) = Notice that eG ∈ O{exp},a (C\R) and bv(eG) = g. Hence, by Theorem 8.4, there is H ∈ O{exp},a(C) such that G = eG + H on the upper half-plane. Since there is {ω}[[−a, a]] such that H(ζ) = L{h; ζ} (Case 1 in Subsection 8.1), (cid:3) h ∈ A′[[−a, a]] = A′ we conclude that G(ζ) = L{f+ + h; ζ}. L{f−; ζ}, η < 0. {ω} If ω satisfies (α) but not (NA), we must have ω(t) ≍ t and thus K1,(ω)(R) = U(t)(C) and K1,{ω}(R) = A{t}(R). In this case, the counterparts of Theorems 8.4 and 8.5 go back to the work of Silva and Morimoto [29, 38]. For the sake of completeness, we end this article by stating these theorems. Let a ≥ 0. We define Oexp,h,a(C\T λ) as the space of all G ∈ O(C\T λ) such that sup G(ζ)e−(a+ε)η−(h+ε)ξ < ∞, ζ∈C\T λ for every ε > 0, and O(exp),a = [λ,h>0 Oexp,h,a(C\T λ), O{exp},a(C\R) = \λ,h>0 Oexp,h,a(C\T λ). HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 37 The proofs of the ensuing two results go along the same lines as those of Theorems 8.4 and 8.5, we therefore choose to omit them. Theorem 8.6. Let a ≥ 0. The sequences 0 −→ O(exp),a(C) −→ O(exp),a bv −−→ U ′ (t)(C) −→ 0 and 0 −→ O{exp},a(C) −→ O{exp},a(C\R) bv −−→ A′ {t}(R) −→ 0 are exact (the boundary value operator being interpreted in the sense of Section 7.3). Theorem 8.7. Let a ≥ 0. (i) Suppose G is analytic on the half-plane {ζ = ξ + iη ∈ C : η > λ} for some λ > 0, and satisfies (8.12) G(ζ)e−(a+ε)η−(h+ε)ξ < ∞, sup η>λ for some h > 0 and every ε > 0, then there is f ∈ U ′ (t),(−a)+ with G(ζ) = L{f ; ζ}. Conversely, the Laplace transform L{f ; ζ} of any f ∈ U ′ (t),(−a)+ is analytic on some half-plane {ζ = ξ + iη ∈ C : η > λ}, λ > 0, and satisfies (8.12) for some h > 0 and every ε > 0. (ii) A function G analytic on the upper half-plane satisfies (8.12) for every h, λ, ε > 0 if and only if there is f ∈ A′ {t},(−a)+ such that G(ζ) = L{f ; ζ}. References [1] G. Bjorck, Linear partial differential operators and generalized distributions, Ark. Mat. 6 (1966), 351 -- 407. [2] E. Bruning, S. Nagamachi, Relativistic quantum field theory with a fundamental length, J. Math. Phys. 45 (2004), 2199 -- 2231. [3] E. Bruning, S. Nagamachi, Edge of the wedge theorem for tempered ultra-hyperfunctions, Com- plex Var. Elliptic Equ. 59 (2014), 787 -- 808. [4] R. Carmichael, A. Kami´nski, S. Pilipovi´c, Boundary values and convolution in ultradistribution spaces, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2007. [5] R. Carmichael, D. Mitrovi´c, Distributions and analytic functions, Pitman Research Notes in Mathematics Series, 206, Longman Scientific & Technical, Harlow; John Wiley & Sons, Inc., New York, 1989. [6] S.-Y. Chung, D. Kim, S. Lee, Characterization for Beurling-Bjorck space and Schwartz space, Proc. Amer. Math. Soc. 125 (1997), 3229 -- 3234. [7] M. De Wilde, Crit`eres de densit´e et de s´eparation dans des limites projectives et inductives d´enombrables, Bull. Soc. Roy. Sci. Li´ege 41 (1973), 155 -- 162. [8] P. Dimovski, S. Pilipovi´c, J. Vindas, Boundary values of holomorphic functions in translation- invariant distribution spaces, Complex Var. Elliptic Equ. 60 (2015), 1169 -- 1189. [9] E. Dubinsky, Projective and inductive limits of Banach spaces, Studia Math. 42 (1972), 259 -- 263. [10] R. Estrada, J. Vindas, On Borel summability and analytic functionals, Rocky Mountain J. Math. 43 (2013), 895 -- 903. [11] C. Fern´andez, A. Galbis, M. C. G´omez-Collado, (Ultra)distributions of Lp-growth as boundary values of holomorphic functions, Rev. R. Acad. Cienc. Exactas F´ıs. Nat. Ser. A Mat. 97 (2003), 243 -- 255. [12] D. H. T. Franco, J. A. Louren¸co, L. H. Renoldi, The ultrahyperfunctional approach to non- commutative quantum field theory, J. Phys. A 41 (2008), 095402, 21 pp. 38 A. DEBROUWERE AND J. VINDAS [13] D. H. T. Franco, L. H. Renoldi, A note on Fourier-Laplace transform and analytic wave front set in theory of tempered ultrahyperfunctions, J. Math. Anal. Appl. 325 (2007), 819 -- 829. [14] I. M. Gelfand, G. E. Shilov, Generalized functions. Vol. 2. Spaces of Fundamental and Generalized Functions, Academic Press, New York-London, 1968. [15] I. M. Gelfand, G. E. Shilov, Generalized functions. Vol. 3. Theory of differential equations, Aca- demic Press, New York, 1967. [16] M. Hasumi, Note on the n-dimensional tempered ultra-distributions, Tohoku Math. J. 13 (1961), 94 -- 104. [17] J. Horv´ath, Topological vector spaces and distributions, Addison-Wesley, Reading, MA, 1966. [18] R. F. Hoskins, J. Sousa Pinto, Theories of generalised functions. Distributions, ultradistributions and other generalised functions, Horwood Publishing Limited, Chichester, 2005. [19] A. Kaneko, Introduction to hyperfunctions, Kluwer Academic Publishers Group, Dordrecht; SCI- PRESS, Tokyo, 1988. [20] T. Kawai, On the theory of Fourier hyperfunctions and its applications to partial differential equations with constant coefficients, J. Fac. Sci. Tokyo Sect. IA Math. 17 (1970), 467 -- 517. [21] H. Komatsu, Ultradistributions I. Structure theorems and a characterization, J. Fac. Sci. Tokyo Sect. IA Math. 20 (1973), 25 -- 105. [22] P. Koosis, Introduction to Hp spaces, London Mathematical Society Lecture Note Series, 40, Cambridge University Press, Cambridge-New York, 1980. [23] J. Korevaar, Tauberian theory. A century of developments, Grundlehren der Mathematischen Wissenschaften, 329, Springer-Verlag, Berlin, 2004. [24] G. Kothe, Die Randverteilungen analytischer Funktionen, Math. Z. 57 (1952), 13 -- 33. [25] G. Kothe, Dualitat in der Funktionentheorie, J. Reine Angew. Math. 191 (1953), 30 -- 49. [26] M. Langenbruch, Bases in spaces of analytic germs, Ann. Polon. Math. 106 (2012), 223 -- 242. [27] R. Meise, D. Vogt, Introduction to functional analysis, Clarendon Press, Oxford, 1997. [28] M. Morimoto, Sur les ultradistributions cohomologiques, Ann. Inst. Fourier (Grenoble) 19 (1969), 129 -- 153. [29] M. Morimoto, Analytic functionals with non-compact carrier, Tokyo J. Math. 1 (1978), 77 -- 103. [30] M. Morimoto, An introduction to Sato's hyperfunctions, A.M.S., Providence, 1993. [31] S. Nagamachi, E. Bruning, Frame independence of the fundamental length in relativistic quantum field theory, J. Math. Phys. 51 (2010), 022305, 18 pp. [32] Y. Oka, K. Yoshino, Solvability of Lewy equation in the space of the tempered ultrahyperfunctions, J. Pseudo-Differ. Oper. Appl. 3 (2012), 321 -- 328. [33] 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. [34] H.-J. Petzsche, D. Vogt, Almost analytic extension of ultradifferentiable functions and the bound- ary values of holomorphic functions, Math. Ann. 267 (1984), 17 -- 35. [35] H.-J. Petzsche, On E. Borel's theorem, Math. Ann. 282 (1988), 299 -- 313. [36] S. Pilipovi´c, B. Prangoski, J. Vindas, On quasianalytic classes of Gelfand-Shilov type. Parametrix and convolution, preprint, arXiv:1507.08331. [37] J. Sebastiao e Silva, Le calcul op´erationnel au point de vue des distributions, Portugal. Math. 14 (1956), 105 -- 132. [38] J. Sebastiao e Silva, Les fonctions analytiques comme ultra-distributions dans le calcul op´erationnel, Math. Ann. 136 (1958), 58 -- 96. [39] M. A. Soloviev, Quantum field theory with a fundamental length: a general mathematical frame- work, J. Math. Phys. 50 (2009), 123519, 17 pp. [40] J. Sousa Pinto, Silva tempered ultradistributions, Portugal. Math. 47 (1990), 267 -- 292. [41] F. Tr`eves, Topological vector spaces, distributions and kernels, Academic Press, New York, 1967. [42] D. V. Widder, Functions harmonic in a strip, Proc. Amer. Math. Soc. 12 (1961), 67 -- 72. [43] K. Yoshino, M. Suwa, The structure of positive definite Fourier ultra-hyperfunctions, Complex Var. Elliptic Equ. 51 (2006), 611 -- 624. HYPERFUNCTIONS AND ULTRAHYPERFUNCTIONS OF FAST GROWTH 39 [44] V. V. Zharinov, Fourier-ultrahyperfunctions, (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), 533 -- 570; English translation: Math. USSR-Izv. 16 (1981), 479 -- 511. [45] Z. Ziele´zny, On the space of convolution operators in K′ 1, Studia Math. 31 (1968), 111 -- 124. Department of Mathematics, Ghent University, Krijgslaan 281, 9000 Gent, Belgium E-mail address: [email protected] Department of Mathematics, Ghent University, Krijgslaan 281, 9000 Gent, Belgium E-mail address: [email protected]
1701.06226	1	1701	2017-01-22T22:31:59	The Szlenk index of $L_p(X)$ and $A_p$	[ "math.FA" ]	Given a Banach space $X$, a $w^$-compact subset of $X^$, and $1<p<\infty$, we provide an optimal relationship between the Szlenk index of $K$ and the Szlenk index of an associated subset of $L_p(X)^*$. As an application, given a Banach space $X$, we prove an optimal estimate of the Szlenk index of $L_p(X)$ in terms of the Szlenk index of $X$. This extends a result of H\'ajek and Schlumprecht to uncountable ordinals. More generally, given an operator $A:X\to Y$, we provide an estimate of the Szlenk index of the "pointwise $A$" operator $A_p:L_p(X)\to L_p(Y)$ in terms of the Szlenk index of $A$.	math.FA	math	THE SZLENK INDEX OF Lp(X) AND Ap RYAN M. CAUSEY Abstract. Given a Banach space X, a w∗-compact subset of X ∗, and 1 < p < ∞, we provide an optimal relationship between the Szlenk index of K and the Szlenk index of an associated subset of Lp(X)∗. As an application, given a Banach space X, we prove an optimal estimate of the Szlenk index of Lp(X) in terms of the Szlenk index of X. This extends a result of H´ajek and Schlumprecht to uncountable ordinals. More generally, given an operator A : X → Y , we provide an estimate of the Szlenk index of the "pointwise A" operator Ap : Lp(X) → Lp(Y ) in terms of the Szlenk index of A. 7 1 0 2 n a J 2 2 ] . A F h t a m [ 1 v 6 2 2 6 0 . 1 0 7 1 : v i X r a 1. Introduction and h ∈ BLq . Recall that these functions act on Lp(X) by hgh, f i = R 1 Throughout this work, X will be a fixed Banach space and K ⊂ X ∗ will be a w∗-compact, non-empty subset. For 1 < p < ∞, we let Kp denote the w∗-closure in Lp(X)∗ of all functions of the form gh ∈ Lq(X ∗) ⊂ Lp(X)∗, where g : [0, 1] → K is simple and Lebesgue measurable, 0 hg(), f ()ih()d for f ∈ Lp(X). Note that if R > 0 is such that K ⊂ RBX ∗, Kp ⊂ RBLp(X)∗, so that Kp is also w∗-compact. If K = BX ∗, Kp = BLp(X)∗ by the Hahn-Banach theorem. If A : X → Y is an operator, then there exists a "pointwise A" operator Ap : Lp(X) → Lp(Y ) given by (Apf )() = A(f ()) for all ∈ [0, 1]. Then if K = A∗BY ∗, Kp = (Ap)∗BLp(Y )∗, which follows from the Hahn-Banach theorem. Thus it is natural to examine what relationship exists between K and Kp. In particular, one may ask what relationship exists between the Szlenk indices of these sets. To that end, we obtain the optimal relationship. In what follows, ω denotes the first infinite ordinal. Theorem 1. Fix 1 < p < ∞. Suppose that ξ is an ordinal such that Sz(K) 6 ωξ. Then Sz(Kp) 6 ω1+ξ. If K is convex, Sz(Kp) 6 ωSz(K). If K is convex and Sz(K) > ωω, Sz(K) = Sz(Kp). Using the facts stated in the introduction that Kp = (Ap)∗BLp(Y )∗ if K = A∗BY ∗, we immediately deduce the following from Theorem 1. If A : X → Y is an operator and K = A∗BY ∗, then Corollary 2. Fix 1 < p < ∞. Sz(Ap) 6 ωSz(A), and if Sz(A) > ωω, Sz(Ap) = Sz(A). In particular, Ap is Asplund if and only if A is. Applying Corollary 2 to the identity of a Banach space, we extend the result of H´ajek and Schlumprecht from [8] to uncountable ordinals. 1 2 RYAN M. CAUSEY We recall that K is said to be w∗-fragmentable if for any non-empty subset L of K and any ε > 0, there exists a w∗-open subset U of X ∗ such that L ∩ U 6= ∅ and diam(L ∩ U) < ε. We recall that K is w∗-dentable if for any non-empty subset L of K and any ε > 0, there exists a w∗-open slice S of X ∗ such that L ∩ U 6= ∅ and diam(L ∩ S) < ε. We recall that a w∗-open slice is a subset of X ∗ of the form {x∗ ∈ X ∗ : Re x∗(x) > a} for some x ∈ X and a ∈ R. As mentioned in [5], a consequence of Corollary 2 is that if Sz(K) 6 ωξ, then Sz(K) 6 Dz(K) 6 ω1+ξ, where Dz(K) denotes the w∗-dentability index of K. Thus Corollary 2 implies that K is w∗-dentable if and only if it is w∗-fragmentable. In addition to considering the Szlenk index of a set, one may consider the ξ-Szlenk power type pξ(L) of the set L, which is important in ξ-asymptotically uniformly smooth renorm- ings of Banach spaces and operators. The concept of a ξ-asymptotically uniformly smooth operator was introduced in [6], and further sharp renorming results regarding the ξ-Szlenk power type of an operator were established in [4]. To that end, we have the following. Theorem 3. For any ordinal ξ and any 1 < p < ∞, if 1/p + 1/q = 1, p1+ξ(Kp) 6 max{q, pξ(K)}. In the case that ξ > ω and pξ(K) 6 p, pξ(K) = pξ(Kp), in showing that Theorem 3 is sharp in some cases. The author wishes to thank P.A.H. Brooker, P. H´ajek, N. Holt, and Th. Schlumprecht for helpful remarks during the preparation of this work. 2. Lp(X), Trees, Szlenk index, games 2.1. Trees, Γξ,n, Pξ,n, and stablization results. Given a set Λ, we let Λ<N denote the finite, non-empty sequences in Λ. Given two members s, t of Λ<N, we let sat denote the concatenation of s and t, s denotes the length of s, s (cid:22) t means s is an initial segment of t, and si denotes the initial segment of s having length i. Given t ∈ Λ<N, we let [(cid:22) t] = {s ∈ Λ<N : s (cid:22) t}. Any subset T of Λ<N which contains all non-empty initial segments of its members will be called a B-tree. We define by transfinite induction the derived B trees of T . We let MAX(T ′) denote the (cid:22)-maximal members of T and T ′ = T \ MAX(T ). We then define T 0 = T , T ξ+1 = (T ξ)′, and if ξ is a limit ordinal, T ξ = ∩γ<ξT γ. We let o(T ) denote the smallest ordinal ξ such that T ξ = ∅, provided such an ordinal exists. If no such ordinal exists, we write o(T ) = ∞. We say T is well-founded if o(T ) is an ordinal, and T is ill- founded if o(T ) = ∞. For convenience, we agree to the convention that if ξ is an ordinal ξ < ∞, and that ω∞ = ∞. Given a B-tree T and a Banach space Y , we let T.Y = {(ζi, Zi)k i=1 ∈ T, Zi ∈ codim(Y )}, where codim(Y ) denotes the closed subspaces of Y having finite codimension in Y . We let C denote the norm compact subsets of BX and i=1 : (ζi)k T.X.C = {(ζi, Zi, Ci)k i=1 : (ζi)k i=1, Zi ∈ codim(X), Ci ∈ C}. THE SZLENK INDEX OF Lp(X) AND Ap 3 We note that T.Y and T.X.C are B-trees. Furthermore, for any ordinal γ, (T.Y )γ = T γ.Y and (T.X.C)γ = T γ.X.C. In particular, T.Y and T.X.C have the same order as T . Given a B-tree T , a Banach space Y , and a collection (xt)t∈T.Y ⊂ Y , we say (xt)t∈T.Y is normally weakly null provided that for any t = (ζi, Zi)k i=1 ∈ T.Y , xt ∈ Zk. Given another B-tree S and a function σ : S.Y → T.Y , we say σ is a pruning provided that for every s, s1 ∈ S.Y with s ≺ s1, σ(s) ≺ σ(s1), and if s1 = sa(ζ, Z) and σ(s1) = ta(µ, W ) for some t ∈ T.Y , W 6 Z. If σ : S.Y → T.Y is a pruning and τ : MAX(S.Y ) → MAX(T.Y ) is such that for every s ∈ MAX(S.Y ), σ(s) (cid:22) τ (s), we say the pair (σ, τ ) is an extended pruning, and denote this by (σ, τ ) : S.Y → T.Y . For every ξ ∈ N and n ∈ N, a B-tree Γξ,n was defined in [3] so that o(Γξ,n) = ωξn. Furthermore, a function Pξ : Γξ → [0, 1] was defined so that for every t ∈ MAX(Γξ), Ps(cid:22)t Pξ(s) = 1. Furthermore, Γξ+1 is the disjoint union of Γξ,n, n ∈ N. For convenience, It follows from the definitions that we define Pξ,n : Γξ,n → [0, n] by Pξ,n(s) = nPξ+1(s). Γξ,1 = Γξ and Pξ,1 = Pξ. For every ξ and every n ∈ N, there exist disjoint subsets Λξ,n,1, . . ., Λξ,n,n of Γξ,n such that Γξ,n = ∪n i=1Λξ,n,i. It follows from the facts regarding Pξ+1 discussed in [3] that, with these definitions, for every ordinal ξ, every n ∈ N, every 1 6 i 6 n, and every t ∈ MAX(Γξ,n), PΛξ,n,i∋s(cid:22)t Pξ,n(s) = 1. For any Banach space Y , we may define Pξ,n on Γξ,n.Y and Γξ,n.X.C by letting Pξ,n((ζi, Zi)k i=1) = Pξ,n((ζi)k i=1) and Pξ,n((ζi, Zi, Ci)k i=1) = Pξ,n((ζi)k i=1). We say an extended pruning (σ, τ ) : Γξ,n.X → Γξ,n.X is level preserving provided that for every 1 6 i 6 n, σ(Λξ,n,i) ⊂ Λξ,n,i. The following theorem collects results from Theorem 3.3, Propositions 3.2, 3.3, and Lemma 3.4 of [4]. Theorem 4. Suppose ξ is an ordinal and n is a natural number. (i) If f : Π(Γξ.n.X) → R is bounded and λ ∈ R is such that λ < inf t∈M AX(Γξ.n.X)Xs(cid:22)t Pξ,n(s)f (s, t), then there exist a level preserving extended pruning (σ, τ ) : Γξ,n.X → Γξ,n.X and real i=1 bi and for every 1 6 i 6 n and every Λξ,n,i ∋ numbers b1, . . . , bn such that λ < Pn s (cid:22) t ∈ MAX(Γξ,n.X), bi 6 f (σ(s), τ (t)). (ii) If (M, d) is a compact metric space and f : Π(Γξ,n.X) → M is any function, then for any δ > 0, there exist x1, . . . , xn ∈ M and a level preserving extended pruning (σ, τ ) : Γξ,n.X → Γξ,n.X such that for every 1 6 i 6 n and every Λξ,n,i ∋ s (cid:22) t ∈ MAX(Γξ,n.X), d(xi, f (σ(s), τ (t))) < δ. 4 RYAN M. CAUSEY (iii) If F is a finite set and f : MAX(Γξ,n.X) → F is any function, there exists a level preserving extended pruning (σ, τ ) : Γξ,n.X → Γξ,n.X such that f ◦ τ M AX(Γξ,n.X) is constant. (iv) For any natural numbers k1 < . . . < kr 6 n, there exists an extended pruning (σ, τ ) : Γξ,r.X → Γξ,n.X such that for every 1 6 i 6 r, σ(Λξ,n,i) ⊂ Λξ,n,ki. 2.2. The Szlenk index, Szlenk power type. Given a w∗-compact subset L of X ∗ and ε > 0, we let sε(K) denote the set consisting of those x∗ ∈ L such that for every w∗- neighborhood V of x∗, diam(L ∩ V ) > ε. We define the transfinite derivations s0 ε(L) = L, sξ+1 ε (L) = sε(sξ ε(L)), and if ξ is a limit ordinal, sξ sζ ε(L). ε(L) = \ζ<ξ If there exists an ordinal ξ such that sξ ε(L) = ∅, we let Sz(L, ε) be the minimum such ε(L) is w∗-compact, we deduce that ordinal. Otherwise we write Sz(L, ε) = ∞. Since sξ Sz(L, ε) cannot be a limit ordinal. We agree to the conventions that ω∞ = ∞ and ξ < ∞ for any ordinal ξ. We let Sz(L) = supε>0 Sz(L, ε). If B : Z → W is an operator, we let Sz(B, ε) = Sz(B∗BW ∗, ε), Sz(B) = Sz(B∗BW ∗). If Z is a Banach space, Sz(Z, ε) = Sz(IZ, ε) and Sz(Z) = Sz(IZ). We recall that a set L ⊂ X ∗ is called w∗-fragmentable if for any ε > 0 and any w∗- compact, non-empty subset M of L, sε(M) ( M. This is equivalent to Sz(L) < ∞. We say an operator B : Z → W is Asplund if B∗BW ∗ is w∗-fragmentable, which happens if and only if Sz(B) < ∞. We say a Banach space Z is Asplund if IZ is Asplund. These are not the original definitions of Asplund spaces and operators, but they are equivalent to the original definitions (see [1]). If Sz(K) 6 ωξ+1, then for any ε > 0, Sz(K, ε) 6 ωξn for some n ∈ N. We let Szξ(K, ε) be the smallest n ∈ N such that Sz(K, ε) 6 ωξn. We define the ξ-Szlenk power type pξ(K) of K by pξ(K) = lim sup ε→0+ log Szξ(K, ε) log(ε) . This value need not be finite. By convention, we let pξ(K) = ∞ if Sz(K) > ωξ+1. We let pξ(A) = pξ(A∗BY ∗) and pξ(X) = pξ(BX ∗). The quantities pξ(X), pξ(A) are important for the renorming theorem of ξ-asymptotically uniformly smooth norms with power type modulus. Given a w∗-compact subset L of X ∗ and ε > 0, we let HL i=1 Ci such that Ci ∈ C for each 1 6 i 6 n and such that there exist (xi)n products Qn Qn i=1 Ci and x∗ ∈ K such that for each 1 6 i 6 n, Re x∗(xi) > ε. ε denote the set of Cartesian i=1 ∈ THE SZLENK INDEX OF Lp(X) AND Ap 5 2.3. The Szlenk index of Kp. Recall that for 1 < p < ∞, Lp(X) denotes the space of equivalence classes of Bochner integrable functions f : [0, 1] → X such that R kf kp < ∞, where [0, 1] is endowed with its Lebesgue measure. Recall also that if 1 < q < ∞, Lq(X ∗) is isometrically included in Lp(X)∗ by the action f 7→ Z hg, f i, for g ∈ Lq(X ∗). We also recall that if : X → R is any Lipschitz function, then for any f ∈ Lp(X), ◦ f ∈ Lp. We note that the Szlenk index and the ξ Szlenk power type of K are unchanged by scaling K by a positive scalar or by replacing K with its balanced hull. Moreover, for a positive scalar c, (cK)p = cKp, which has the same Szlenk index and ξ-Szlenk power type as Kp. If TK is the balanced hull of K, Kp ⊂ (TK)p and Sz(K) = Sz(TK) ([3, Lemma 2.2]) so that Theorem 1, Corollary 2, and Theorem 3 hold in general if they hold under the assumption that K ⊂ BX ∗ is balanced. Therefore we can and do assume throughout that K ⊂ BX ∗ and K is balanced. Let : X → R be given by (x) = maxx∗∈K Re x∗(x). Since we have assumed K is balanced, (x) = maxx∗∈K x∗(x). It is easy to see that for any 1 < p < ∞ and any f ∈ Lp(X), k(f )kLp = maxf ∗∈Kp Re f ∗(f ). Combining this fact with [4, Corollary 2.4] and the proof of that corollary, we obtain the following. Theorem 5. Fix 1 < p, α < ∞. (i) If for every B-tree T with o(T ) = ω1+ξ and every normally weakly null (ft)t∈T.Lp(X) ⊂ BLp(X), inf(cid:8)k(f )kLp : t ∈ T.Lp(X), f ∈ co(fs : ∅ ≺ s (cid:22) t)(cid:9) = 0, then Sz(Kp) 6 ω1+ξ. (ii) If there exists a constant C such that for every n ∈ N, every B-tree T with o(T ) = ω1+ξn, and every normally weakly null collection (ft)t∈T.Lp(X) ⊂ BLp(X), inf(cid:8)k(f )kLp : t ∈ T.Lp(X), f ∈ co(fs : ∅ ≺ s (cid:22) t)(cid:9) 6 Cn−1/α, then p1+ξ(Kp) 6 α. Proposition 6. Suppose T is a non-empty B-tree. Suppose also that (Cs)s∈T.X ⊂ C is fixed and for s = (ζi, Zi)k i=1 ∈ T.X, let λ(s) = Zk ∩ Cs. Suppose that S is a non-empty, well- founded B-tree and θ : S.X → T.X is a pruning. For s ∈ S.X, let s(s) = Qs i=1 λ(θ(si)). If ε > 0 is such that for every t ∈ S.X, s(s) ∈ HK 6= ∅, then for any 0 < δ < ε, any ε 0 6 γ < o(S), and any s ∈ Sγ.X, s(s) ∈ Hsγ 6= ∅. Moreover, for any 0 < δ < ε, Sz(K, δ) > o(S). δ (K) ε Proof. We induct on γ. The base case is the hypothesis. Assume γ + 1 < o(S) and the result holds for γ. Assume s ∈ Sγ+1.X, which means there exists ζ such that sa(ζ, Z) ∈ Sγ.X for all Z ∈ codim(X). Then for every Z ∈ codim(X), there exists Z > WZ ∈ codim(X) 6 RYAN M. CAUSEY such that s(sa(ζ, Z)) ⊂ s(s) × BWZ . From this and the inductive hypothesis, for every Z ∈ codim(X), we fix xZ ∈ BWZ , (xZ Z(xZ) > ε and Re x∗ i ) > ε for each 1 6 i 6 s. By compactness of s(s) × K with the product topology, where λ(θ(si)) has its norm topology and K has its w∗-topology, δ (K) such that Re x∗ i=1 ∈ s(s), and x∗ Z ∈ sγ Z(xZ i )s ∅ 6= \Z∈codim(X) {(xY 1 , . . . , xY s, x∗ Y ) : Z > Y ∈ codim(X)} ⊂ s(s) × K. Fix (x1, . . . , xs, x∗) lying in this intersection. Obviously x∗ ∈ sγ w∗-neighborhood V of x∗, there exists Z ∈ codim(X) such that ker(x∗) ⊂ Z and x∗ whence δ (K). Moreover, for any Z ∈ V , diam(sγ δ (K) ∩ V ) > kx∗ Z − x∗k > Re (x∗ Z − x∗)(xZ) = Re x∗ Z(xZ) > ε > δ. This implies x∗ ∈ sγ+1 that s(s) ∈ Hsγ+1 δ (K) δ ε and completes the successor case. (K). It is obvious that Re x∗(xi) > ε for all 1 6 i 6 s. This shows Finally, assume γ < o(S) is a limit ordinal and the result holds for all ordinals less than γ. Fix s ∈ Sγ.X and let s(s) × K be topologized as in the successor case. By the inductive hypothesis, for all β < γ, there exists (xβ ε (K) and for all 1 6 i 6 s, Re x∗ 1 , . . . , xβ β ∈ sβ s, x∗ β) ∈ s(s) × K such that x∗ i=1 λ(θ(si))(cid:1) × K, β(xβ i ) > ε. By compactness of (cid:0)Qs \β<γ 1 , . . . , xµ s, x∗ {(xµ µ) : µ > β} 6= ∅. Clearly any (x1, . . . , xs, x∗) lying in this intersection is such that x∗ ∈ sγ 1 6 i 6 s, Re x∗(xi) > ε. This shows that s(s) ∈ Hsγ δ (K) ε and completes the induction. δ (K) and for any We have shown that for any 0 < δ < ε, Sz(K, δ) > o(S). If o(S) is a limit ordinal, we deduce that Sz(K, δ) > o(S) since Sz(K, δ) cannot be a limit ordinal. If o(S) is a successor, say o(S) = ξ + 1, then there exists a length 1 sequence (ζ) ∈ Sξ. For every Z ∈ codim(X), s((ζ, Z)) = WZ ∩ Cθ((ζ,Z)) for some W ⊂ Z. The first part of the proof yields that for each Z ∈ codim(X), there exists xZ ∈ WZ ∩ Cθ((ζ,Z)) ⊂ WZ ∩ BX and some x∗ ζ(K) such that Z(xZ) > ε. Arguing as in the successor case, we deduce that any w∗-limit of a subnet Re x∗ Z)Z∈codim(X) lies in sξ+1 of (x∗ (K), whence Sz(K, δ) > ξ + 1 = o(S). Z ∈ sξ δ (cid:3) 2.4. Games. Suppose T ⊂ Λ<N is a well-founded, non-empty B-tree and E ⊂ MAX(T.X.C) is some subset. We define the game on T.X.C with target set E. Player I first chooses (ζ1, Z1) ∈ Λ × codim(X) such that (ζ) ∈ T and Player II then chooses C1 ∈ C. Assuming (ζi, Zi)n i=1 ∈ T.X and C1, . . . , Cn ∈ C have been chosen, the game terminates if (ζi, Zi)n i=1 ∈ MAX(T.X). Otherwise Player I chooses (ζn+1, Zn+1) ∈ Λ × codim(X) such that (ζi)n+1 i=1 ∈ T and Player II chooses Cn+1 ∈ C. Since T is well-founded, this game must terminate after finitely many steps. Suppose that the resulting choices are (ζi, Zi)n i=1 and C1, . . . , Cn ∈ C. We say that Player II wins if (ζi, Zi, Ci)n i=1 ∈ E, and Player I wins otherwise. THE SZLENK INDEX OF Lp(X) AND Ap 7 A strategy for Player I for the game on T.X.C with target set E is a function ψ : T ′.X.C ∪ i=1 ∈ T . We say ψ is a i=1 ∈ MAX(T.X.C) {∅} → Λ × codim(X) such that if ψ((ζi, Zi, Ci)n−1 winning strategy for Player I provided that for any sequence (ζi, Zi, Ci)n such that (ζi, Zi) = ψ((ζj, Zj, Cj)i−1 j=1) for every 1 6 i 6 n, (ζi, Zi, Ci)n i=1 ) = (ζn, Zn), (ζi)n i=1 /∈ E. A strategy for Player II for the game on T.X.C with target set E is a function ψ defined on the set {((ζi, Zi, Ci)n−1 i=1 , (ζn, Zn)) :(ζi, Zi, Ci)n−1 i=1 ∈ T } (ζi)n i=1 ∈ {∅} ∪ T.X.C, (ζn, Zn) ∈ Λ × codim(X), and taking values in C. We say ψ is a winning strategy for Player II provided that for any sequence (ζi, Zi, Ci)n j=1, (ζi, Zi)) for all 1 6 i 6 n, (ζi, Zi, Ci)n i=1 ∈ MAX(T.X.C) such that Ci = ψ((ζj, Zj, Cj)i−1 i=1 ∈ E. Proposition 7. [5, Proposition 3.1] For any non-empty, well-founded B-tree T and any E ⊂ T.X.C, either Player I or Player II has a winning strategy for the game on T.X.C with target set E. Proposition 8. Suppose that Player II has a winning strategy for a game on T.X.C with tar- get set E. Then there exists (Cs)s∈T.X ⊂ C such that for every t = (ζi, Zi)k i=1 ∈ MAX(T.X), (ζi, Zi, Cti)k i=1 ∈ E. Proof. Fix a winning strategy ψ for Player II in the game. We define Cs by induction on s. We let C(ζ,Z) = ψ(∅, (ζ, Z)). If s = k + 1, Csi has been defined for every 1 6 i 6 k, and s = sa k (ζ, Z), we let Cs = ψ(sk, (ζ, Z)). (cid:3) For the next proposition, if h ∈ Lp(X) is a simple function, we let h be the function in Lp(X) such that h() = 0 if h() = 0 and h() = h()/kh()k otherwise. Proposition 9. Let ξ be an ordinal, n a natural number, and let T be a B-tree with o(T ) > ω1+ξn. If ψ is a strategy for Player I for some game on Γξ,n.X.C, then for any 1 < p < ∞, any δ > 0, and any normally weakly null (ft)t∈T.Lp(X) ⊂ BLp(X), there exist s = (ζi, Zi)k i=1 ∈ MAX(Γξ,n.X), ∅ = t0 ≺ t1 ≺ . . . ≺ tk ∈ T.Lp(X), gi ∈ co(fu : ti−1 ≺ u (cid:22) ti), hi ∈ BLp(X), and Ci ∈ C such that for every 1 6 i 6 k, (i) hi is simple, (ii) range(hi) = Ci ⊂ BZi, (iii) kgi − hikLp(X) < δ, (iv) (ζi, Zi) = ψ((ζj, Zj, Cj)i−1 j=1). Remark 10. For a B-tree S on Λ and s ∈ S, we let S(s) denote those non-empty sequences u ∈ Λ<N such that sau ∈ S. An easy induction argument yields that for any ordinals ξ, ζ, Sξ(s) = (S(s))ξ for any ordinal ξ. From this it follows that s ∈ Sξ if and only if o(S(s)) > ξ. 8 RYAN M. CAUSEY Furthermore, another easy induction yields that if (Sξ)ζ = Sξ+ζ, from which it follows that if o(S) > ξ + ζ, o(Sξ) > ζ. Therefore if s ∈ Sξ+ω, o(Sξ(s)) > ω. Proof of Proposition 9. We first note that if Z ∈ codim(X), Lp(X)/Lp(Z) is either the zero vector space or isomorphic to Lp, and therefore has Szlenk index not exceeding ω. As explained in [5], this means that for any B-tree T with o(T ) > ω, any δ > 0, and any normally weakly null (ft)t∈T.Lp(X) ⊂ BLp(X), there exist t ∈ T.Lp(X), g ∈ co(fs : ∅ ≺ s (cid:22) t), and h ∈ BLp(Z) such that kg − hkLp(X) < δ. Moreover, by the density of simple functions, we may assume this h is simple. Let ψ be a strategy for Player I for a game on Γξ,n.X.C. Let T be a B-tree with o(T ) = ω1+ξn and define γ : Γξ,n.X ∪{∅} → [0, ωξn] by letting γ(t) = max{µ 6 ωξn : t ∈ (Γξ,n.X)µ} for t ∈ Γξ,n.X and γ(∅) = ωξn. Let s0 = t0 = ∅. Now assume that for some k ∈ N and all 1 6 i < k, si ∈ Γξ,n.X, ζi ∈ [0, ωξn], Zi ∈ codim(X), ti ∈ T.Lp(X), gi, hi ∈ BLp(X), and Ci ∈ C have been chosen such that for all 1 6 i < k, j=1, j=1), (i) hi is simple, (ii) si = (ζj, Zj)i (iii) t0 ≺ t1 ≺ . . . ≺ tk−1, (iv) ti ∈ (T.Lp(X))ωγ(si), (v) (ζi, Zi) = ψ((ζj, Zj, Cj)i−1 (vi) gi ∈ co(fu : ti−1 ≺ u (cid:22) ti), (vii) kgi − hikLp(X) < δ, (viii) range(hi) = Ci ⊂ BZi. If sk−1 is maximal in Γξ,n.X, we let s = sk−1, and one easily checks that the conclusions are satisfied. Otherwise let (ζk, Zk) = ψ((ζj, Zj, Cj)k−1 k−1(ζk, Zk). Let uk−1 be the sequence of first members of the pairs of tk−1 and let U denote the proper extensions of uk−1 in T ωγ(sk). Then (fta k−1u)u∈U.Lp(X) ⊂ BLp(X) is normally weakly null and o(U) > ω by the remark preceding the proof, so that the previous paragraph yields the existence of some u′ ∈ U.Lp(X), gk ∈ co(fu : tk−1 ≺ u (cid:22) ta k−1u′), and some simple function hk ∈ Lp(Zk) such that kgk − hkkLp(X) < δ. Let tk = ta k−1u′. In order to apply the remark before the proof, we note that since sk−1 ≺ sk, γ(sk−1) > γ(sk) + 1. Since j=1) and sk = sa ωγ(sk−1) > ω(γ(sk) + 1) = ωγ(sk) + ω, the remark preceding the proof applies. Note that Ck := range(hk) ⊂ BZk. This completes the recursive construction. Since Γξ,n.X is well-founded, eventually this process terminates. The resulting s = (ζi, Zi)k i=1 ∈ MAX(Γξ,n.X) clearly satisfies the conclusions. (cid:3) 3. Definition of an associated space and two games 3.1. The associated space and its properties. If E is a vector space with seminorm k · k, we say a sequence (ei)n i=1 in E is 1-unconditional provided that for any scalars (ai)n i=1 THE SZLENK INDEX OF Lp(X) AND Ap 9 and any (εi)n space E with seminorm k · k which is spanned by the 1-unconditional basis (ei)n p-concave provided there exists a constant C such that for any (fi)n i=1 aieik. Recall that for 1 < p < ∞, a vector i=1 is called i=1 ∈ {±1}n, kPn i=1 εiaieik = kPn i=1 ⊂ Lp, k n Xi=1 fieikLp(E) 6 Ck n Xi=1 kfikLpeikE. The smallest such constant C is denoted by M(p)(E). Given x ∈ span(ei : 1 6 i 6 n), where (ei)n space E, we write x = Pn i=1 is a Hamel basis for the seminormed 6= 0}. We say the vectors x1, . . . , xn ∈ span(ei : 1 6 i 6 n) are disjointly supported if the sets supp(x1), . . ., supp(xn) are pairwise disjoint. i=1 aiei and supp(x) = {i 6 n : ai For 1 < β < ∞, we say that an unconditional Hamel basis (ei)n i=1 for a seminormed space E satisfies an 1-lower ℓβ estimate provided that for any m ∈ N and any disjointly supported elements (xi)m i=1 ⊂ E, m m Xi=1 (cid:0) kxikβ(cid:1)1/β 6 k xik. Xi=1 Theorem 11. [7, Theorem 1.f.7] Fix 1 < β < p < ∞. There exists a constant C ′ = C ′(β, p) such that if (ei)n i=1 is a 1-unconditional basis for the seminormed space E which satisfies a 1-lower ℓβ estimate, then E is p-concave and M(p)(E) 6 C ′. For the remainder of this section, T is a fixed, non-empty B-tree. For a non-empty set J, we let c00(J) be the span of the canonical Hamel basis (ej)j∈J in the space of scalar-valued functions on J, where ej is the indicator of the singleton {j}. We let e∗ j denote the coordinate functional to ej. Given x ∈ c00(J), we may write x = Pj∈J ajej. Then we define x to be Pj∈J ajej. A suppression projection is an operator P from span(e∗ j : j ∈ J) into itself such that there exists a subset F of J such that P Pj∈J aje∗ j = Pj∈F aje∗ j . For 0 < φ < θ < 1, let Nθ,φ,T = {0} ∪ nθ k Xi=1 e∗ tji : t = (ζi, Zi, Ci)t i=1 ∈ T.X.C, 1 6 j1 < . . . < jk 6 t, k Yi=1 Zji ∩ Cji ∈ HK φ o ⊂ span(e∗ t : t ∈ T.X.C). For 0 < φ < θ < 1 and 1 < α < ∞, let Mθ,φ,α,T = n k Xi=1 aigi :gi ∈ ∪∞ n=1Nθn,φn,T , ai > 0, k Xi=1 i 6 1, supp(gi) are pairwise disjointo. aα Note that the set Mθ,φ,α,T is closed under suppression projections. We define the seminorm k · kθ,φ,α,T on c00(T.X.C) by kxkθ,φ,α,T = sup{f (x) : f ∈ Mθ,φ,α,T }. 10 RYAN M. CAUSEY Claim 12. Fix 1 < α < ∞ and 0 < φ < θ < 1. For any t ∈ T.X.C, (eti)t unconditional and satisfies a 1-lower ℓβ estimate in its span, where 1/α + 1/β = 1. i=1 is 1- : 1 6 i 6 t) Proof. Note that 1-unconditionality is obvious. Fix x1, . . . , xn ∈ span(eti with disjoint supports. That is, there exist pairwise disjoint subsets S1, . . . , Sn of {1, . . . , t} such that xi ∈ span(etj : j ∈ Si). Then there exist g1, . . . , gn ∈ Mθ,φ,α,T such that for each 1 6 i 6 n, gi(xi) = kxikθ,φ,α,T . Since Mθ,φ,α,T is closed under suppression projections, we may assume that supp(gi) ⊂ Si for each 1 6 i 6 n. Then if (ai)n i = 1, i=1 aα i=1 aikxikθ,φ,α,T = (Pn ai > 0, and Pn Xi=1 Xi=1 xikθ,φ,α,T > g(cid:0)(cid:12)(cid:12) i=1 kxikβ xi(cid:12)(cid:12)(cid:1) = k n n n θ,φ,α,T )1/β, g := Pn Xi=1 Xi=1 aigi(xi) = (cid:0) i=1 are such that Pn i=1 aigi ∈ Mθ,φ,α,T and n kxikβ θ,φ,α,T )1/β. (cid:3) Claim 13. Fix 1 < α < ∞ and 0 < φ < θ < 1. For any t = (ζi, Zi, Ci)k sequence (xi)k i=1 Zi ∩ Ci, and any sequence (ai)k i=1 of non-negative scalars, i=1 ∈ T.X.C, any i=1 ∈ Qk ( k Xi=1 aixi) 6 1 θ − φ k k Xi=1 aietikθ,φ,α,T . i=1 as in the statement, fix x∗ ∈ K such that Re x∗(Pk Proof. We recall that if C ∈ C, C ⊂ BX by the definition of C. With t, (xi)k i=1 Zi ∩Ci, and (ai)k i=1 aixi). For all j ∈ N, let Bj = {i 6 k : Re x∗(xi) ∈ (φj, φj−1]}. Note that for every j ∈ N, θj Pi∈Bj i=1 ∈ Qk i=1 aixi) = (Pk ∈ Nθj,φj,T , so e∗ ti k k φj−1 Xi∈Bj ai = φ−1(φ/θ)j(θj Xi∈Bj e∗ ti)( Xi=1 aieti) 6 φ−1(φ/θ)jk aietikθ,φ,α,T . Xi=1 Then ( k Xi=1 aixi) 6 ∞ Xj=1 Xi∈Bj aiRe x∗(xi) 6 ∞ Xj=1 φj−1 Xi∈Bj ai 6 ∞ Xj=1 φ−1(φ/θ)jk k Xi=1 aietikθ,φ,α,T = 1 θ − φ k k Xi=1 aietikθ,φ,α,T . Corollary 14. Fix 1 < p, α, β < ∞ with 1/α + 1/β = 1 and β < p. Let C ′ = C ′(β, p) be the constant from Theorem 11. Suppose that ξ is an ordinal, n is a natural number, ε > 0, and 0 < φ < θ < 1 are such that Player I has a winning strategy in the game with target set nt ∈ MAX(Γξ,n.X.C) : kXs(cid:22)t Pξ,n(s)eskθ,φ,α,Γξ,n > εo. (cid:3) THE SZLENK INDEX OF Lp(X) AND Ap 11 Then for any B-tree T with o(T ) > ω1+ξn and any normally weakly null (ft)t∈T.Lp(X) ⊂ BLp(X), inf(cid:8)k(f )kLp : t ∈ T.Lp(X), f ∈ co(fs : ∅ ≺ s (cid:22) t)(cid:9) 6 C ′ε n(θ − φ) . Proof. Recall for the proof that for a simple function h ∈ Lp(X), h is the function in Lp(X) such that h() = 0 if h() = 0 and h() = h()/kh()k otherwise. Fix a winning strategy ψ for Player I in the game with the indicated target set. Fix δ > 0. By Proposition 9, there exist s = (ζi, Zi)k i=1 ∈ MAX(Γξ,n.X), ∅ = t0 ≺ . . . ≺ tk, gi ∈ co(fu : ti−1 ≺ u (cid:22) ti), simple functions hi ∈ BLp(X), and Ci ∈ C such that kgi − hikLp(X) < δ, range(hi) = Ci ⊂ BZi, and (ζi, Zi) = ψ((ζj, Zj, Cj)i−1 j=1). This means that for any ∈ [0, 1], (hi())k i=1 Zi ∩ Ci, whence by Claim 13, for any non-negative scalars (ai)k i=1, i=1 ∈ Qk Xi=1 ( k aihi()) = ( k Xi=1 aikhi()khi()) 6 1 θ − φ k k Xi=1 aikhi()kesikθ,φ,α,Γξ,n. Since by Claim 13 (eu)u(cid:22)s satisfies a lower ℓβ estimate in its span, we deduce that k( k Xi=1 k ( n−1Pξ,n(si)hi)kLp = (cid:16)Z 1 0 (cid:12)(cid:12)(cid:12) Xi=1 n(θ − φ)(cid:16)Z 1 n(θ − φ)(cid:13)(cid:13)(cid:13)Xu(cid:22)s n−1Pξ,n(si)khi()khi())(cid:12)(cid:12)(cid:12) 0 (cid:13)(cid:13)(cid:13)Xu(cid:22)s Pξ,n(u)khu()keu(cid:13)(cid:13)(cid:13) Pξ,n(u)khukLp(X)eu(cid:13)(cid:13)(cid:13)θ,φ,α,Γξ,n C ′ 6 6 1 p p d(cid:17)1/p d(cid:17)1/p θ,φ,α,Γξ,n C ′ε 6 . n(θ − φ) Here we have used 1-unconditionality, khikLp(X) 6 1 for each 1 6 i 6 k, and the fact that since ψ is a winning strategy for Player I, kXu(cid:22)s Pξ,n(u)khukLp(X)eukθ,φ,α,Γξ,n 6 kXu(cid:22)s Pξ,n(u)eukθ,φ,α,Γξ,n 6 ε. Let g = n−1Pu(cid:22)s Pξ,n(u)gi ∈ co(fu : u (cid:22) tk) and h = n−1Pu(cid:22)s Pξ,n(u)hi. Since is 1-Lipschitz, it follows that k(g) − (g)kLp 6 kg − hkLp(X) < δ, so that k(g)kLp 6 δ + k(h)kLp 6 δ + C ′ε n(θ − φ) . Since δ > 0 was arbitrary, we are done. (cid:3) 12 RYAN M. CAUSEY 3.2. Particular games on Γξ,n.X.C. The statement of Proposition 6 is notationally cum- bersome. We isolate the following result as a way of using Proposition 6. Lemma 15. Fix 0 < φ < θ < 1. Suppose that ξ is an ordinal, m, n are natural numbers, (Cs)s∈Γξ,n.X ⊂ C, and (σ, τ ) : Γξ,m.X → Γξ,n.X is an extended pruning. For t = (ζi, Zi)k i=1 ∈ Γξ,n.X, let r(t) = (ζi, Zi, Cti)k i=1. If ν ∈ N is such that for every s ∈ MAX(Γξ,m.X), there exists a functional hs ∈ ∪ν l=1Nθi,φi,Γξ,n such that ∪t(cid:22)sr(σ(t)) ⊂ supp(hs), then Sz(K, φν/2) > ωξm. Proof. For s = (ζi, Zi)k i=1 ∈ Γξ,n.X, let λ(s) = Zk ∩ Cs. For s ∈ Γξ,m.X, let s(s) = i=1 λ(σ(si)). Fix s ∈ MAX(Γξ,m.X) and let hs ∈ ∪ν Qs and fix 1 6 l 6 ν such that hs ∈ Nθl,φl,Γξ,n. We will prove that s(s) ∈ HK 1 6 m 6 k and any C ′ i ∈ HK show that for any non-empty initial segment s1 of s, s(s1) ∈ HK Proposition 6 will finish the proof. l=1Nθl,φl,Γξ,n be as in the statement of the lemma φν . Since for any i ∈ HK φν , this will φν . From here, an appeal to k ∈ C such that Qk φν , Qm 1, . . . , C ′ i=1 C ′ i=1 C ′ i=1 Wji ∩ Cji ∈ HK Fix u = (µi, Wi, Ci)u and Qµ Note that for all 1 6 i 6 s, r(σ(si)) = (ζj, Zj, Ctj )li hypothesis, i=1 ∈ Γξ,n.X.C and 1 6 j1 < . . . < jµ 6 u such that hs = θlPµ φl. Let τ (s) = t = (ζi, Zi)η j=1 Zlj ∩ Ctlj uji i=1. For each 1 6 i 6 s, let li = σ(si). . By j=1 and s(s) = Qs i=1 e∗ {(ζj, Zj, Ctj )li j=1 : 1 6 i 6 s} = {r(σ(si)) : 1 6 i 6 s} ⊂ supp(hs) = {(µj, Wj, Cj)ji j=1 : 1 6 i 6 µ}. . Choose (xi)µ From this it follows that there exist m1 < . . . < ms such that for every 1 6 i 6 s, i=1 Wji ∩ Cji such that there exists x∗ ∈ K so that r(σ(si)) = ujmi Re x∗(xi) > φl for each 1 6 i 6 µ, which exists because Qµ φl. Since φl. Since Zli = Wjmi l 6 ν, HK i=1 Wji ∩ Cji ∈ HK , which shows that s(s) ∈ HK , (xmi)s φν , so that s(s) ∈ HK φν . i=1 ∈ Qs i=1 ∈ Qµ i=1 Zli ∩ Ctli and Ctli = Cjmi φl ⊂ HK (cid:3) Lemma 16. Fix 1 < α < ∞ and 0 < φ < θ < 1. If Sz(K) 6 ωξ, then for any ε > 0, Player I has a winning strategy in the game with target set nt ∈ MAX(Γξ.X.C) : kXs(cid:22)t Pξ(s)eskθ,φ,α,Γξ > εo. Proof. Suppose not. Then by Proposition 8, there exist ε > 0 and (Cs)s∈Γξ.X ⊂ C such that ε < infnkXs(cid:22)t Pξ(s)er(s)kθ,φ,α,Γξ : t ∈ MAX(Γξ.X)o. i=1 ∈ Γξ.X, let r(s) = (ζi, Zi, Csi)k For s = (ζi, Zi)k i=1. For every t ∈ MAX(Γξ.X), fix ft ∈ Mθ,φ,α,Γξ such that supp(ft) ⊂ [(cid:22) r(t)] and ft(Ps(cid:22)t Pξ(s)er(s)) = kPs(cid:22)t Pξ(s)er(s)kθ,φ,α,Γξ. THE SZLENK INDEX OF Lp(X) AND Ap 13 Define F : Π(Γξ.X) → R by letting F (s, t) = ft(er(s)). By Theorem 4, there exists an extended pruning (σ, τ ) : Γξ.X → Γξ.X such that ε < inf (s,t)∈Π(Γξ.X) F (σ(s), τ (t)). Fix ν ∈ N such that ε > θν and for each t ∈ MAX(Γξ.X), write fτ (t) = Pkt ai,t > 0, Pkt i=1 ai,tgi,t where n=1Nθn,φn,Γξ have pairwise disjoint supports. For each i,t 6 1, and gi,t ∈ ∪∞ i=1 aα t ∈ MAX(Γξ.X), let Rt = {i 6 kt : ai,t > ε}. i=1 aα Since Pkt i,t 6 1, Rt 6 ⌊1/εα⌋ =: k0. Note that since ε < fτ (t)(er(σ(s))) for any ∅ ≺ s (cid:22) t, r(σ(s)) ∈ ∪i∈Rtsupp(gi,t). We write Pi∈Rt ai,tgi,t = Plt i=1 is an enumeration of (ai,t)i∈Rt, and (hi,t)lt i=1 is the corresponding enumeration of (gi,t)i∈Rt. Define κ : Π(Γξ.X) → {1, . . . , k0} by letting κ(σ, τ ) be the unique i 6 lt such that r(σ(s)) ∈ supp(hi,t). By Theorem 4(ii), there exists an extended pruning (σ′, τ ′) : Γξ.X → Γξ.X and 1 6 l 6 k0 such that κ(σ′(s), τ ′(t)) = l for all (s, t) ∈ Π(Γξ.X). We now note that for any s ∈ MAX(Γξ.X), hl,τ ′(s) ∈ ∪ν i=1 bi,thi,t where lt 6 k0, (bi,t)lt i=1Nθi,φi,Γξ is such that {r(σ ◦ σ′(u)) : ∅ ≺ u (cid:22) s} ⊂ supp(hl,τ ′(s)), and an appeal to Lemma 15 yields that Sz(K, φν/2) > ωξ. This contradiction finishes the proof. To see that hl,τ ′(s) ∈ ∪ν i=1Nθi,φi,Γξ, we note that if hl,τ ′(s) ∈ Nθi,φi,Γξ, ε 6 hl,τ ′(t)(er(σ◦σ′ (s))) 6 khl,τ ′(t)k∞ 6 θi. This shows that i 6 ν by our choice of ν. (cid:3) Lemma 17. Fix 1 < α, β < ∞ and 0 < φ < 2−1/α and assume that 1/α + 1/β = 1. Assume that for some C > 1 and all i ∈ N, Szξ(K, φi/2) 6 C2i. Let θ = 2−1/α. Then for any n ∈ N and any C1 > C, Player I has a winning strategy in the game with target set nt ∈ MAX(Γξ,n.X.C) : kXs(cid:22)t Pξ,n(s)eskθ,φ,α,Γξ,n > C1n1/βo. Proof. Suppose not. Then for some n ∈ N, there exist (Cs)s∈Γξ,n.X ⊂ C and such that (ft)t∈M AX(Γξ,n.X) ⊂ Mθ,φ,α,Γξ,n Cn1/β < inf t∈M AX(Γξ,n .X) ft(cid:0)Xs(cid:22)t Pξ,n(s)er(s)(cid:1). We may assume as in Lemma 16 that supp(ft) ⊂ [(cid:22) r(t)] for each t ∈ MAX(Γξ,n.X). Then by Theorem 4(i), there exist a level preserving extended pruning (σ, τ ) : Γξ,n → Γξ,n i=1 bi and for all 1 6 i 6 n and all Λξ,n,i ∋ i=1 bi. Let and numbers, b1, . . . , bn such that Cn1/β < Pn s (cid:22) t ∈ MAX(Γξ,n), fτ (t)(er(σ(s))) > bi. Fix δ > 0 such that Cn1/β + nδ < Pn R = {i 6 n : bi > δ}. 14 RYAN M. CAUSEY Sublemma 18. There exist a level preserving extended pruning (σ0, τ0) : Γξ,n.X → Γξ,n.X, l, w ∈ N, (ai)l i=1 ⊂ {1, . . . , w}, and (gt)t∈M AX(Γξ,n.X) ⊂ Mθ,φ,α,Γξ,n such that α, (ki)i∈R ⊂ {1, . . . , l}, (wi)l i=1 ∈ Bℓl (i) for each t ∈ MAX(Γξ,n.X), kgt − fτ ◦τ0(t)k∞ < δ, (ii) for any t ∈ MAX(Γξ,n.X), there exist disjointly supported functionals h1,t, . . . , hl,t such that hi,t ∈ Nθwi ,φwi ,Γξ,n and gt = Pl i=1 aihi,t, (iii) for i ∈ R and Λξ,n,i ∋ s (cid:22) t ∈ MAX(Γξ,n.X), r(σ ◦ σ0(s)) ∈ supp(hki,t), We first finish the proof of the lemma and then return to the proof of the sublemma. Note that item (iii) of the sublemma implies that for i ∈ R and Λξ,n,i ∋ s (cid:22) t ∈ MAX(Γξ,n), From this and our choice of δ we deduce that bi 6 gt(er(σ◦σ0(s))) + δ = akiθwki + δ. n Cn1/β + δn < Xi=1 bi 6 δn +Xi∈R akiθwki . Partition R into sets R1, . . . , Rl, where Rj = {i ∈ R : ki = j}, so that l akiθwki = ajθwj Rj. Cn1/β < Xi∈R Indeed, suppose Rj > C2wj for some j. By We claim that for each j, Rj 6 C2wj . Theorem 4(iv), if Rj = {r1, . . . , rm}, with r1 < . . . < rm, there exists extended pruning (σ′, τ ′) : Γξ,m.X → Γξ,n.X such that σ′(Λξ,m,i) ⊂ Λξ,n,ri. We now use Lemma 15 to deduce that Szξ(K, φwj/2) > C2wj , which is a contradiction. Thus we deduce that Rj 6 C2wj for each j. This means that for each 1 6 j 6 l, Xj=1 θwj = (2−1/α)wj = (2wj )−1/α 6 C 1/αRj−1/α 6 CRj−1/α. Then l Xj=1 ajθwj Rj 6 C l Xj=1 ajRj1−1/α = C l l Xj=1 ajRj1/β 6 C(cid:0) Xj=1 ajα(cid:1)1/α(cid:0) n Xj=1 Rj(cid:1)1/β 6 CR1/β 6 Cn1/β. Thus we reach a contradiction. t ∈ MAX(Γξ,n), write fτ (t) = Pkt and ai,t > 0 such that Pkt i=1 aα Pkt i=1 aα Plt i=1 a′ We now return to the proof of the sublemma. First fix w ∈ N such that θw < δ. For each j=1Nθj ,φj,Γξ,n i,t 6 1. Let St = {i 6 kt : kai,tfi,tk∞ > δ}. Note that since i=1 ai,tfi,t for some disjointly supported fi,t ∈ ∪∞ i,t 6 1, St 6 ⌊1/δα⌋ =: k0. As in the previous lemma, we write Pi∈St ai,tfi,t = i,tf ′ i,t for some lt 6 k0. Considering the function from MAX(Γξ,n.X) given by t 7→ lt ∈ {1, . . . , k0}, we use Theorem 4(iii) to obtain l ∈ N and a level preserving extended pruning (σ′, τ ′) : Γξ,n.X → Γξ,n.X such that for all t ∈ MAX(Γξ,n.X), lτ ′(t) = l. Note that since ka′ i,τ ′(t) ∈ Nθj ,φj,Γξ,n, i,τ ′(t)k∞ > δ for every 1 6 i 6 l and t ∈ MAX(Γξ,n), if f ′ i,τ ′(t)f ′ THE SZLENK INDEX OF Lp(X) AND Ap 15 j 6 w. Let wi,τ ′(t) be the value j ∈ {1, . . . , w} such that f ′ the map from MAX(Γξ,n.X) into Bℓl i,τ ′(t) ∈ Nθj,φj,Γξ,n. By considering α × {1, . . . , w}l given by i=1(cid:1), i=1, (wi,τ ′(t))l t 7→ (cid:0)(ai,τ ′(t))l we use Theorem 4(iii) again to find another level preserving extended pruning (σ′′, τ ′′) : Γξ,n.X → Γξ,n.X, (ai)l i=1 ⊂ {1, . . . , w} such that for all t ∈ MAX(Γξ,n), k(ai,τ ′◦τ ′′(t))l i=1 − (ai)l i,τ ′◦τ ′′(t) ∈ Nθwi ,φwi ,Γξ,n. Note that for all t ∈ MAX(Γξ,n.X), α < δ and for all 1 6 i 6 l, f ′ i=1 ∈ Bℓl i=1kℓl α and (wi)l kfτ ◦τ ′◦τ ′′(t) − l Xi=1 aif ′ i,τ ′◦τ ′′(t)k∞ < δ. This implies that for any i ∈ R and any Λξ,n,i ∋ s (cid:22) t, since δ 6 bi 6 fτ ◦τ ′◦τ ′′(er(σ◦σ′ ◦σ′′(s))), j=1supp(f ′ r(σ ◦ σ′ ◦ σ′′(s)) ∈ ∪l j,τ ′◦τ ′′(t)). Thus we may let κ(s, t) be the unique j ∈ {1, . . . , l} such that r(σ ◦ σ′ ◦ σ′′(s)) ∈ supp(fj,τ ′◦τ ′′(t)) if s ∈ ∪i∈RΛξ,n,i, and κ(s, t) = 0 otherwise. Applying Theorem 4(ii), we deduce the existence of (ki)i∈R ⊂ {1, . . . , l} and a level preserving extended pruning (σ′′′, τ ′′′) : Γξ,n.X → Γξ,n.X such that setting σ0 = σ′ ◦ σ′′ ◦ σ′′′, τ0 = τ ′ ◦ τ ′′ ◦ τ ′′′, hi,t = f ′ i=1 aihi,t, finishes the proof. i,τ0(t), and gt = Pl (cid:3) 4. Proof of the main results 1 Proof of Theorem 1. Let φ = 1/3 and θ = 2/3, so that θ−φ = 3. Fix 1 < p < ∞. Fix any 1 < α, β < ∞ such that β < p and 1/α + 1/β = 1. Let C ′ = C ′(β, p) be the constant from Theorem 11. Fix ε > 0. By Lemma 16, Player I has a winning strategy in the game with target set nt ∈ MAX(Γξ.X.C) : kXs(cid:22)t Pξ(s)eskθ,φ,α,Γξ > εo. By Corollary 14, for any B-tree T with o(T ) = ω1+ξ and any normally weakly null collection (ft)t∈T.Lp(X) ⊂ BLp(X), infnk(f )kLp : t ∈ T.Lp(X), f ∈ co(fs : ∅ ≺ s (cid:22) t)o 6 3C ′ε. We deduce Sz(Kp) 6 ω1+ξ by Theorem 5(i). It is clear that Sz(K) 6 Sz(Kp) for any 1 < p < ∞. If K is convex, then either Sz(K) = ∞, in which case Sz(Kp) = ∞ = ω∞ = Sz(K), or there exists an ordinal ξ such that Sz(K) = ωξ [2, Proposition 4.2]. We deduce that Sz(Kp) 6 ω1+ξ = ωSz(K) by the previous paragraph. In the case that ξ > ω, 1 + ξ = ξ. (cid:3) 16 RYAN M. CAUSEY Proof of Theorem 3. If pξ(K) = ∞, there is nothing to show, so assume pξ(K) < ∞. Fix 1 < p, q < ∞ with 1/p + 1/q = 1. Fix 1 < α, β, γ < ∞ such that max{pξ(K), q} < γ < α and 1/α + 1/β = 1. Let C ′ = C ′(β, p) be the constant from Theorem 11. Let φ = 2−1/γ and note that supi∈N εγSzξ(K, φi/2)/2i < ∞. By Lemma 17, with θ = 2−1/α, there exists a constant C1 such that for every n ∈ N, Player I has a winning strategy in the game with target set nt ∈ MAX(Γξ,n.X.C) : kXs(cid:22)t Pξ,n(s)eskθ,φ,α,Γξ,n > C1/n1/βo. By Corollary 14, for every n ∈ N, every B-tree T with o(T ) = ω1+ξn, and every normally weakly null (ft)t∈T.Lp(X) ⊂ BLp(X), infnk(f )kLp : t ∈ T.Lp(X), f ∈ co(∅ ≺ s (cid:22) t)o 6 C1C ′ n(θ − φ) n1/β = C1C ′ n1/αθ − φ . By Theorem 5(ii), p1+ξ(Kp) 6 α. Since α > max{pξ(K), q} was arbitrary, we deduce that p1+ξ(Kp) 6 max{pξ(K), q}. (cid:3) References [1] P.A.H. Brooker, Asplund operators and the Szlenk index, Operator Theory 68 (2012), 405-442. [2] R.M. Causey, An alternate description of the Szlenk index with applications, Illinois J. Math. 59 (2) (2015), 359-390. [3] R.M. Causey, The Szlenk index of injective tensor products and convex hulls, to appear in Journal of Functional Analysis. DOI 10.1016/j.jfa.2016.12.017. [4] R.M. Causey, Power type ξ-asymptotically uniformly smooth norms, submitted. [5] R.M. Causey, A note on the relationship between the Szlenk and w∗-dentability indices of arbitrary w∗-compact sets, to appear in Positivity. [6] R.M. Causey, S.J. Dilworth, ξ-asymptotically uniformly smooth, ξ-asymptotically uni- formly convex, and (β) operators, submitted. [7] 4] J. Lindenstrauss, L. Tzafriri, Classical Banach spaces. II, Ergebnisse der Mathematik und ihrer Grenzgebiete, Vol. 92. Springer-Verlag, Berlin, 1977. [8] P. H´ajek, Th. Schlumprecht, The Szlenk index of Lp(X), Bull. Lond. Math. Soc. 46 (2014), no. 2, 415-424.
1812.06842	1	1812	2018-12-17T15:35:41	Multi-Toeplitz operators and free pluriharmonic functions	[ "math.FA", "math.OA" ]	We initiate the study of weighted multi-Toeplitz operators associated with noncommutative regular domains in B(H)^n. These operators are acting on the full Fock space with n generators and have as symbols free pluriharmonic functions. Several classical results from complex analysis concerning harmonic functions have analogues in our noncommutative setting. In particular, we show that the bounded free pluriharmonic functions are precisely those which are noncommutative Berezin transforms of weighted multi-Toeplitz operators, and solve the Dirichlet extension problem in this setting. Using noncommutative Cauchy transforms, we provide a free analytic functional calculus for n-tuples of operators, which extends to free pluriharmonic functions.	math.FA	math	MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS GELU POPESCU Abstract. We initiate the study of weighted multi-Toeplitz operators associated with noncommutative regular domains Dm q (H) ⊂ B(H)n, m, n ≥ 1, where B(H) is the algebra of all bounded linear operators on a Hilbert space H. These operators are acting on the full Fock space with n generators and have as symbols free pluriharmonic functions on the interior of the domain Dm q (H). We prove that the set of all weighted multi-Toeplitz operators coincides with A(Dm q )∗ + A(Dm q ) W OT , where the domain algebra A(Dm q ) is the norm-closed unital non-selfadjoint algebra generated by the universal model (W1, . . . , Wn) of the noncommutative domain Dm q (H). These results are used to study the class of free pluriharmonic functions on Dm q (H)◦. Several classical results from complex analysis concerning harmonic functions have analogues in our noncommutative setting. In particular, we show that the bounded free pluriharmonic functions are precisely those which are noncommutative Berezin transforms of weighted multi-Toeplitz operators, and solve the Dirichlet extension problem in this setting. Using noncommutative Cauchy transforms, we provide a free analytic functional calculus for n-tuples of operators, which extends to free pluriharmonic functions. Our study of weighted multi-Toeplitz operators on Fock spaces is a blend of multi-variable operator theory, noncommutative function theory, operator spaces, and harmonic analysis. Introduction Let H 2(D) be the Hardy space of all analytic functions on the open unit disc D := {z ∈ C : with square-sumable coefficients. An operator T ∈ B(H 2(D)) is called Toeplitz if z < 1} T f = P+(ϕf ), f ∈ H 2(T), for some ϕ ∈ L∞(T), where P+ is the orthogonal projection of the Lebesgue space L2(T) onto the Hardy space H 2(T), which is identified with H 2(D). Brown and Halmos [3] proved that a necessary and sufficient condition that an operator on the Hardy space H 2(D) be a Toeplitz operator is that its matrix [λij ] with respect to the standard basis ek(z) = zk, k ∈ {0, 1, . . .}, be a Toeplitz matrix, i.e λi+1,j+1 = λij , i, j ∈ {0, 1, . . .}, which is equivalent to S∗T S = T , where S is the unilateral shift on H 2(D). In this case, λij = ai−j, where ϕ =Pk∈Z akχk is the Fourier expansion of the symbol ϕ ∈ L∞(T). The class of Toeplitz operators originates with O. Toeplitz [36] and has been studied extensively over the years, starting with Hartman and Wintner [10] and the seminal paper of Brown and Halmos [3]. The study of Toeplitz operators on the Hardy space H 2(D) was extended to Hilbert spaces of holomorphic functions on the unit disc (see [11]) such as the Bergman space and weighted Bergman space, and also to higher dimensional setting involving holomorphic functions in several complex variables on various classes of domains in Cn (see Upmeier's book [37]). The class of Toeplitz operators is one of the most important classes of non-selfadjoint operators having applications in index theory and noncommutative geometry, prediction theory, boundary values problems Date: November 15, 2018. 2000 Mathematics Subject Classification. Primary: 47B35; 47A56; 47A13; Secondary: 46L52; 46L07; 47A60. Key words and phrases. Multivariable operator theory, Multi-Toeplitz operator, Full Fock space, Free pluriharmonic function, Noncommutative domain, Berezin transform. Research supported in part by NSF grant DMS 1500922. 1 2 GELU POPESCU for analytic functions, probability, information theory and control theory, and several other fields. We refer the reader to [2], [7], [33], and [11] for a comprehensive account on Toeplitz operators. A polynomial q ∈ ChZ1, . . . , Zni in n noncommutative indeterminates is called positive regular if all its coefficients are positive, q(0) = 0, and the coefficients of the linear terms Z1, . . . , Zn are different from zero. If q =Pα aαZα and X = (X1, . . . , Xn) ∈ B(H)n, we define the completely positive map aαXαY X ∗ α. Φq,X : B(H) → B(H), Φq,X (Y ) :=Xα Dm For each m ≥ 1, we define the noncommutative regular domain q (H) :=(cid:8)X := (X1, . . . , Xn) ∈ B(H)n : (id − Φq,X )k(I) ≥ 0 for 1 ≤ k ≤ m(cid:9) . According to [28] and [27], each such a domain has a universal model (W1, . . . , Wn) consisting of weighted left creation operators acting on the full Fock space with n generators. We mention a few remarkable particular cases. Single variable case: n = 1. (i) If m = 1 and q = Z, the corresponding domain Dm q (H) coincides with the closed unit ball [B(H)]1 := {X ∈ B(H) : kXk ≤ 1}, the study of which has generated the Nagy-Foia¸s theory of contractions (see [35]). In this case, the universal model is the unilateral shift S acting on the Hardy space H 2(D). The Toeplitz operators on the Hardy space H 2(D) have been studied extensively (see for example [7], [33]) (ii) If m ≥ 2 and q = Z, the corresponding domain coincides with the set of all m-hypercontractions studied by Agler in [1], and recently by Olofsson [17], [18]. The corresponding universal model is the unilateral shift acting on the weighted Bergman space Am(D), the Hilbert space of all analytic functions on the unit disc D with kfk2 := m − 1 π ZD f (z)2(1 − z2)m−2dz < ∞. In [16], Louhichi and Olofsson obtain a Brown-Halmos type characterization of Toeplitz operators with harmonic symbols on Am(D), which can be seen as a reproducing kernel Hilbert space with reproducing kernel given by κm(z, w) := (1−z ¯w)−m, z, w ∈ D. Their result was recently extended by Eschmeier and Langendorfer [9] to the analytic functional Hilbert space Hm(B) on the unit ball B ⊂ Cn given by the reproducing kernel κm(z, w) := (1 − hz, wi)−m for z, w ∈ B, where m ≥ 1. Multivariable noncommutative case: n ≥ 2. (i) When m = 1 and q = Z1 + ··· + Zn, the noncommutative domain Dm q (H) coincides with the closed unit ball [B(H)n]1 := {(X1, . . . , Xn) : X1X ∗ n ≤ I}, the study of which has generated a free analogue of Nagy-Foia¸s theory. The corresponding universal model is the n-tuple of left creation operators (S1, . . . , Sn) acting on the full Fock space with n generators. A study of unweighted multi-Toeplitz operators on the full Fock space with n generators was initiated in [21], [22] and has had an important impact in multivariable operator theory and the structure of free semigroups algebras (see [4], [5], [6], [25], [26], [14], [15]). 1 + ··· + XnX ∗ (ii) When m ≥ 1, n ≥ 1, and q is any positive regular polynomial the domain Dm q (H) was studied in [28] (when m = 1), and in [27] (when m ≥ 2). In this case, the corresponding universal model is an n-tuple of weighted left creation operators acting on the full Fock space. We remark that, in the particular case when m ≥ 2 and q = Z1 + ··· + Zn, the corresponding domain can be seen as a noncommutative m-hyperball, the elements of which can be viewed as multivariable noncommutative analogues of Agler's m-hypercontractions. As far as we know, Toeplitz operators have not been introduced or studied in this very general setting. The goal of the present paper is to initiate the study of weighted multi-Toeplitz operators associated q (H) ⊂ B(H)n, m, n ≥ 1, when q ∈ ChZ1, . . . , Zni is any with noncommutative regular domains Dm positive regular polynomial in noncommutative indeterminates. This is accompanied by the study of their symbols which are free pluriharmonic functions on the interior of the domain Dm q (H). MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 3 In Section 1, we present some background from [28] and [27] on the noncommutative domains Dm their universal models, and the associated noncommutative Berezin transforms. q (H), In Section 2, we introduce the weighted multi-Toeplitz operators which are acting on the full Fock q (H) ⊂ B(H)n. space F 2(Hn) with n generators and are associated with the noncommutative domain Dm We show that they are uniquely determined by their free pluriharmonic symbols ϕ(X1, . . . , Xn) = ∞Xk=1 Xα∈F+ n ,α=k bαX ∗ α + ∞Xk=0 Xα∈F+ n ,α=k aαXα, aα, bα ∈ C, where F+ operator norm topology for any n-tuple (X1, . . . , Xn) in the interior of Dm all weighted multi-Toeplitz operators coincides with n is the unital free semigroup with n generators and the convergence of the series is in the q (H). We prove that the set of A(Dm q )∗ + A(Dm q ) W OT , where the domain algebra A(Dm q ) is the norm-closed unital non-selfadjoint algebra generated by the universal model (W1, . . . , Wn) of the noncommutative domain Dm q (H). In the particular case when n = 1 and q = Z, we obtain a characterization of the Toeplitz operators with harmonic symbol on the Bergman space Am(D), which should be compared with the corresponding result from [16]. In Section 3, we provide basic results concerning the free pluriharmonic functions on the noncommu- q (H)◦ and show that they are characterized by a mean value property. This result is tative domain Dm used to obtain an analogue of Weierstrass theorem for free pluriharmonic functions and to show that the set Har((Dm q )◦) of all pluriharmonic functions is a complete metric space with respect to an appropriate metric ρ. We also obtain, in this section, a Schur type result [34] characterizing the free pluriharmonic functions with positive real parts in terms of positive semi-definite weighted multi-Toeplitz kernels. q )◦) of all bounded free pluriharmonic functions on Dm Section 4 concerns the space Har∞((Dm q (H)◦. q )◦) if and only if it is the noncommutative Berezin One of the main results states that F ∈ Har∞((Dm transform of a weighted multi-Toeplitz operator. Moreover, we prove that the map Φ : Har∞((Dm q )◦) → A(Dm q )∗ + A(Dm q ) W OT defined by Φ(F ) := SOT- limr→1 F (rW ) is a completely isometric isomorphism of operator spaces. A noncommutative version of the Dirichlet extension problem for harmonic functions (see [13]) is also q )◦) has a continuous extension in the operator norm topology to provided. We prove that F ∈ Har((Dm Dm q (H) if and only if there exists a multi-Toeplitz operator ψ ∈ A(Dm such that F is the noncommutative Berezin transform of ψ. In Section 5, using noncommutative Cauchy transforms associated with the domain Dm q (H), we provide a free analytic functional calculus for n-tuples of operators X = (X1, . . . , Xn) ∈ B(H)n with the spectral radius of the reconstruction operator Rq,X strictly less than 1. This extends to free pluriharmonic functions, proving that the map q )∗ + A(Dm q ) k·k Ψq,X :(cid:0)Har((Dm q )◦), ρ(cid:1) → (B(H),k · k) defined by Ψq,X (G) := G(X) is continuous and its restriction Ψq,XHol((Dm algebra homomorphism. Several consequences of this result are also provided. q )◦) is a continuous unital We should mention that our results are presented in the more general setting of weighted multi-Toeplitz matrices with operator-valued entries and free pluriharmonic functions with operator-valued coefficients, while the noncommutative domain Dm f (H) is generated by any positive regular free holomorphic functions f in a neighborhood of the origin. In a forthcoming paper [32], we obtain a Brown-Halmos characterization of the weighted multi-Toeplitz operators associated with the noncommutative m-hyperball (the case when q = Z1+···+Zn, m ≥ 2) which is a noncommutative version of Eschmeier and Langendorfer recent commutative result [9]. This result shows that the weighted multi-Toeplitz are characterized by an algebraic equation involving the universal model (W1, . . . , Wn) of the noncommutative m-hyperball. It remains to be seen if this characterization extends to the more general domains Dm q , where q is any positive regular polynomial. 4 GELU POPESCU 1. Noncommutative domains, universal models, and Berezin transforms This section contains some definitions and the necessary background from [28] and [27] on the noncom- f (H), their universal models, and the associated noncommutative Berezin mutative regular domains Dm transforms. Let F+ n be the unital free semigroup on n generators g1, . . . , gn and the identity g0. The length of α ∈ F+ n is defined by α := 0 if α = g0 and α := k if α = gi1 ··· gik , where i1, . . . , ik ∈ {1, . . . , n}. If Z1, . . . , Zn are noncommutative indeterminates, we denote Zα := Zi1 . . . Zik if α = gi1 . . . gik ∈ F+ n , i1, . . . ik ∈ {1, . . . , n}, and Zg0 := 1. Similarly, if X := (X1, . . . , Xn) ∈ B(H)n, where B(H) is the algebra of all bounded linear operators on the Hilbert space H, we denote Xα := Xi1 ··· Xik and Xg0 := IH. A formal power series f :=Pα∈F+ aαZα, aα ∈ C, in noncommutative indeterminates Z1, . . . , Zn, is called free holomorphic function on the noncommutative ball [B(H)n]ρ for some ρ > 0, where [B(H)n]ρ := {(X1, . . . , Xn) ∈ B(H)n : kX1X ∗ nk1/2 < ρ}, if the series P∞ k=0Pα=k aαXα is convergent in the operator norm topology for any (X1, . . . , Xn) ∈ [B(H)n]ρ. According to [24], f is a free holomorphic function on [B(H)n]ρ for any Hilbert space H if and only if 1 + ··· + XnX ∗ n lim sup k→∞ Xα=k aα2 1/2k 1 ρ . ≤ Throughout this paper, we assume that aα ≥ 0 for any α ∈ F+ n , ag0 = 0, and agi > 0 if i ∈ {1, . . . , n}. A function f satisfying all these conditions on the coefficients is called positive regular free holomorphic function on [B(H)n]ρ. Let Φf,X : B(H) → B(H) be the completely positive linear map given by Φf,X (Y ) :=Pα≥1 aαXαY X ∗ α for Y ∈ B(H), where the convergence is in the week operator topology, and define the noncommutative regular domain Dm f (H) :=(cid:8)X := (X1, . . . , Xn) ∈ B(H)n : (id − Φf,X )k(I) ≥ 0 for 1 ≤ k ≤ m(cid:9) . We saw in [27], that X ∈ Dm noncommutative domain Dm with the abstract domain Dm b(m) g0 := 1 and (1.1) b(m) α = f (H) if and only if Φf,X (I) ≤ I and (id − Φf,X )m(I) ≥ 0. The abstract f is the disjoint union`H Dm f (H), over all Hilbert spaces H. We associate f a unique n-tuple (W1, . . . , Wn) of weighted shifts, as follows. Define aγ1 ··· aγj(cid:18)j + m − 1 m − 1 (cid:19) αXj=1 Xγ1 ···γj =α if α ∈ F+ n ,α ≥ 1. γ1 ≥1,...,γj ≥1 Let Hn be an n-dimensional complex Hilbert space with orthonormal basis e1, e2, . . . , en, where n ∈ N := {1, 2, . . .}. We consider the full Fock space of Hn defined by n := C1 and H ⊗k n where H ⊗0 i ∈ {1, . . . , n}, be the diagonal operators defined by setting is the Hilbert tensor product of k copies of Hn. Let Di : F 2(Hn) → F 2(Hn), H ⊗k n , F 2(Hn) :=Mk≥0 Dieα :=vuut b(m) α b(m) giα eα, α ∈ F+ n , where {eα}α∈F+ operators Wi : F 2(Hn) → F 2(Hn) associated with Dm the left creation operators on the full Fock space F 2(Hn), i.e. is the orthonormal basis of the full Fock space F 2(Hn). The weighted left creation f are defined by Wi := SiDi, where S1, . . . , Sn are n Siϕ := ei ⊗ ϕ, ϕ ∈ F 2(Hn), i ∈ {1, . . . , n}. MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 5 A simple calculation reveals that (1.2) eβγ and W ∗ γ Wβeγ = qb(m) qb(m) βγ qb(m) γ√b(m) 0 α eγ if α = βγ otherwise β eα = for any α, β ∈ F+ following properties: n . We recall from [27] that the weighted left creation operators W1, . . . , Wn have the aβWβW ∗ β ≤ I, where the convergence is in the strong operator topology; (i) Pβ≥1 (ii) (id − Φf,W )m (I) = PC, where PC is the orthogonal projection of F 2(Hn) on C1 ⊂ F 2(Hn), and the map Φf,W : B(F 2(Hn)) → B(F 2(Hn)) is defined by Φf,W (Y ) := Xα≥1 aαWαY W ∗ α, where the convergence is in the weak operator topology; (iii) W := (W1, . . . , Wn) is a pure element of the domain Dm f (F 2(Hn)), i.e. lim p→∞ Φp f,W (I) = 0 in the strong operator topology. The right creation operators are defined by Riϕ := ϕ ⊗ ei, i ∈ {1, . . . , n}. We can also define the weighted right creation operators Λi : F 2(Hn) → F 2(Hn) by setting Λi := RiGi, i = 1, . . . , n, where each diagonal operator Gi, i = 1, . . . , n, is given by Gieα :=vuut b(m) α b(m) αgi eα, α ∈ F+ n , where the coefficients b(m) n , are described by relation (1.1). In this case, we have for any α, β ∈ F+ weighted left creation operators, one can show that n , where β denotes the reverse of β = gi1 ··· gik , i.e., β = gik ··· gi1. As in the case of where f (Z) := Pα≥1 a αZα, α denotes the reverse of α, and Φ f ,Λ(Y ) := Pα≥1 a αΛαY Λ∗ Y ∈ B(F 2(Hn)), with the convergence is in the weak operator topology. Let X := (X1, . . . , Xn) ∈ Dm f,X : H → F 2(Hn) ⊗ ∆m,X (H) be the noncommutative α for any Berezin kernel defined by α eα ⊗ ∆m,X X ∗ αh, h ∈ H, where ∆m,X := [(I − Φf,X )m(I)]1/2 and the coefficients b(m) α are given by relation (1.1). We know that K (m) f,X X ∗ i = (W ∗ i ⊗ I)K (m) f,X i ∈ {1, . . . , n}. Assume that X is a pure n-tuple, i.e. Φk f,X is an isometry and the n-tuple W := (W1, . . . , Wn) plays the role of the universal model for the noncommutative domain Dm f . f,X (I) → 0 strongly, as k → ∞. Then K (m) (1.3) (1.4) eγ α qb(m) γ√b(m) 0 βeα = (cid:16)id − Φ f ,Λ(cid:17)m if α = γ β otherwise (I) = PC, α , α ∈ F+ Λβeγ = qb(m) qb(m) γ β γ eγ β and Λ∗ Xβ≥1 a βΛβΛ∗ β ≤ I and f (H) and let K (m) nqb(m) f,X h := Xα∈F+ K (m) 6 GELU POPESCU Let ϕ(W ) := Pβ∈F+ In [27], we proved that Pβ∈F+ polynomial in eα, α ∈ F+ n . If n n cβWβ, cβ ∈ C, be a formal sum with the property that Pβ∈F+ < ∞. cβWβ(p) ∈ F 2(Hn) for any p ∈ P, where P ⊂ F 2(Hn) is the set of all n cβ2 1 b(m) β then there is a unique bounded operator acting on F 2(Hn), which we should also denote by ϕ(W ), such that cβWβ(p)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) < ∞, sup p∈P,kpk≤1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ ϕ(W )p = Xβ∈F+ n n cβWβ(p) for any p ∈ P. f ). One can prove that F ∞(Dm The set of all operators ϕ(W ) ∈ B(F 2(Hn)) satisfying the above-mentioned properties is denoted by F ∞(Dm f ) is a Banach algebra, which we call Hardy algebra associated with the noncommutative domain Dm f ) to be the norm closure of all polynomials in the weighted left creation operators W1, . . . , Wn and the identity. Using the weighted right creation operators associated with Dm f . We introduce the domain algebra A(Dm In a similar manner, using the weighted right creation operators Λ := (Λ1, . . . , Λn) associated with f , one can define the corresponding the Hardy algebra R∞(Dm c βΛβ f , one can also define the corresponding domain algebra R(Dm f ). f ). More precisely, if g(Λ) = Pβ∈F+ Dm n b(m) β < ∞ and such that is a formal sum with the property thatPβ∈F+ n cβ2 1 p∈P,kpk≤1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ g(Λ)p = Xβ∈F+ c βΛβ(p) sup n n c βΛβ(p)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) < ∞, for any p ∈ P. then there is a unique bounded operator on F 2(Hn), which we also denote by g(Λ), such that The set of all operators g(Λ) ∈ B(F 2(Hn)) satisfying the above-mentioned properties is denoted by R∞(Dm f ), where ′ stands for the commutant. f ). We proved in [27] that F ∞(Dm f ) and F ∞(Dm f )′′ = F ∞(Dm f )′ = R∞(Dm The noncommutative Berezin transform at X ∈ Dm B(m) X : B(F 2(Hn)) → B(H) defined by f (H), where X is a pure element, is the map B (m) X [g] := K (m) f,X ∗ (g ⊗ IH)K (m) f,X , g ∈ B(F 2(Hn)), where the K (m) polynomials p(W ) in the operators Wi, i ∈ {1, . . . , k}, and the identity. If g is in the operator space f,X : H → F 2(Hn) ⊗ H is noncommutative Berezin kernel. Let P(W ) be the set of all where the closure is in the operator norm, we define the Berezin transform at X ∈ Dm f (H), by S := span{p(W )q(W )∗ : p(W ), q(W ) ∈ P(W )}, B(m) X [g] := lim r→1 K (m) f,rX (g ⊗ IH)K (m) f,rX , g ∈ S, ∗ where the limit is in the operator norm topology. In this case, the Berezin transform at X is a unital completely positive linear map such that If, in addition, X is a pure n-tuple in Dm mutative Berezin transforms and their applications can be found in [23], [27], [28], [29], and [30]. f (H), then limr→1 B X [g], g ∈ S. More on noncom- B(m) X (WαW ∗ β ) = XαX ∗ β, α, β ∈ Fn. (m) rX [g] = B (m) MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 7 2. Wieghted multi-Toeplitz operators on Fock spaces In this section, we introduce the weighted multi-Toeplitz operators associated with the noncommutative f (H) ⊂ B(H)n. We show that they are uniquely determined by their free pluriharmonic domain Dm symbols and provide a characterization in terms of the domain algebra A(Dm f ). n such that ω = σγ. In this case we set ω\rγ := σ. If σ 6= g0 we write ω >r γ. We say that ω and γ are comparable if either ω ≥r γ or γ >r ω. If ω, γ ∈ F+ n , we say that ω ≥r γ if there is σ ∈ F+ n are comparable, we consider the weights In what follows, we need some notation. If ω, γ ∈ F+ r b(m) r b(m) γ b(m) ω ω b(m) γ λω,γ := , , where the coefficients b(m) let [Cω,γ]F+ n ×F+ n α , α ∈ F+ if ω ≥r γ, if γ >r ω, be the operator matrix representation of T ∈ B(E ⊗ F 2(Hn)), i.e. n , are given by relation (1.1). Let E be a separable Hilbert space and hCω,γx, yi := hT (x ⊗ eγ), y ⊗ eωi n and x, y ∈ E. for any ω, γ ∈ F+ Definition 2.1. We say that T is a weighted right multi-Toeplitz operator if for each i ∈ {1, . . . , n} and ω, γ, α, β ∈ F+ n , and Cα,β = 0 if α, β are not comparable. λωgi,γgi Cωgi,γgi = λω,γ Cω,γ , if ω, γ are comparable, We remark that when n = m = 1, f = Z, and E = C we recover the classical Toeplitz operators on the Hardy space H 2(D). Also if n ≥ 2, m = 1, and f = Z1 + ··· + Zn we obtain the unweighted right multi-Toeplitz operators on the full Fock space F 2(Hn) (see [21], [22] and [26]). In this case, we have b(m) α = 1 for any α ∈ F+ n and the condition above becomes Cωgi,γgi =(Cω,γ , 0, if ω ≥r γ or γ >r ω, otherwise, and Cα,β = 0 if α, β are not comparable. For an equivalent and more transparent definition of weighted right multi-Toeplitz operators on the full Fock space F 2(Hn) see the remarks following the next theorem. Theorem 2.2. Any weighted right multi-Toeplitz operator T ∈ B(E ⊗ F 2(Hn)) has a formal Fourier representation where {A(α)}α∈F+ n and {B(α)}α∈F+ n \{g0} are some operators on the Hilbert space E, such that T q = ϕ(W )q, ϕ(W ) := Xα≥1 B(α) ⊗ W ∗ A(α) ⊗ Wα, α + A(0) ⊗ I + Xα≥1 q = Xα≤k hα ⊗ eα, for any hα ∈ E and k ∈ N. If T1, T2 are weighted right multi-Toeplitz operators having the same formal Fourier representation, then T1 = T2. Proof. First, we note that, using Definition 2.1, one can prove that T ∈ B(E ⊗ F 2(Hn)) is a weighted right multi-Toeplitz operator if and only if the entries of its matrix representation [Cω,γ]F+ satisfy the following relations: n ×F+ n σ b(m) γ b(m) σγ (i) Cσγ,γ =r b(m) (ii) Cγ,σγ =r b(m) (iii) Cα,β = 0 if (α, β) ∈ F+ σ b(m) γ b(m) σγ Cσ,g0 for any σ, γ ∈ F+ n ; Cg0,σ for any σ, γ ∈ F+ n ; n × F+ n is not of the form (σγ, γ) or (γ, σγ) for σ, γ ∈ F+ n . 8 GELU POPESCU Consequently, T ∈ B(E ⊗ F 2(Hn)) is a weighted right multi-Toeplitz if and only if (2.1) hT (x ⊗ eγ), y ⊗ eωi = We define the formal Fourier representation of T by setting qb(m) qb(m) γ ω ω\r γqb(m) √b(m) γ\r ω√b(m) qb(m) ω γ 0, (cid:10)T (x ⊗ 1), y ⊗ eω\rγ(cid:11) , (cid:10)T (x ⊗ eγ\rω), y ⊗ 1(cid:11) , if ω ≥r γ, if γ >r ω, otherwise.  where the coefficients are given by (2.2) ϕ(W ) := Xα≥1 (cid:10)A(α)x, y(cid:11) :=qb(m) (cid:10)B(α)x, y(cid:11) :=qb(m) B(α) ⊗ W ∗ α + A(0) ⊗ I + Xα≥1 A(α) ⊗ Wα, α hT (x ⊗ 1), y ⊗ eαi , α ∈ F+ n , α hT (x ⊗ eα), y ⊗ 1i , α ∈ F+ n\{g0}, for any x, y ∈ E. We also set A(0) := A(g0). Hence, we deduce that 1 and qb(m) qb(m) for any x ∈ E. As a consequence, we can see that Pα≥1 T (x ⊗ 1) = Xα∈F+ T ∗(x ⊗ 1) = Xα∈F+ WOT convergent series. We note that n ,α≥1 1 α α n A(α)x ⊗ eα ϕ(W )(x ⊗ eβ) := Xα≥1 (B(α) ⊗ W ∗ α)(x ⊗ eβ) + Xα∈F+ n B∗ (α)x ⊗ eα 1 b(m) α A∗ (α)A(α) and Pα≥1 (A(α) ⊗ Wα)(x ⊗ eβ), 1 b(m) α B(α)B∗ (α) are is well-defined as a vector in E ⊗ F 2(Hn). Indeed, the first sum consists of finitely many terms, while the second one is equal to Pα∈F+ α , one can eαβ. Using the definition of the coefficients b(m) n easily see that b(m) α b(m) β b(m) αβ A(α)x ⊗r b(m) m − 1 (cid:19) b(m) β ≤(cid:18)β + m − 1 Xα∈F+ kA(α)xk2 b(m) β b(m) αβ n αβ . This implies ≤(cid:18)β + m − 1 m − 1 (cid:19) Xα∈F+ n kA(α)xk2 1 b(m) α < ∞. Since T is a weighted right multi-Toeplitz operator, we can use relations (2.1) and (2.2), to obtain (2.3) hT (x ⊗ eγ), y ⊗ eωi = Now, note that (cid:10)Aω\r γx, y(cid:11) , (cid:10)Bγ\rωx, y(cid:11) , if ω ≥r γ, if γ >r ω, otherwise. ω qb(m) γ√b(m) √b(m) qb(m) 0, ω γ  hϕ(W )(x ⊗ eγ), y ⊗ eωi = Xα≥1(cid:10)B(α)x, y(cid:11) hW ∗ αeγ, eωi + Xα∈F+ n(cid:10)A(α)x, y(cid:11)hWαeγ , eωi . MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 9 Due to the definition of the weighted left creation operators W1, . . . , Wn, we have for any α ∈ F+ n , and for any α ∈ F+ n with α ≥ 1. Using these relations, we deduce that (2.4) hϕ(W )(x ⊗ eγ), y ⊗ eωi = γ qb(m) qb(m) 0, αγ √b(m) ω√b(m) 0, αω , if ω = αγ, otherwise. , if γ = αω, otherwise hW ∗ hWαeγ, eωi = αeγ, eωi = qb(m)  (cid:28)√b(m) qb(m) ω√b(m) 0, αω αγ γ A(α)x, y+ , B(α)x, y(cid:29) , if ω = αγ, α ∈ F+ n , if γ = αω, α ∈ F+ otherwise. n with α ≥ 1, Comparing these relations with (2.3), we conclude that for any x, y ∈ E and γ, ω ∈ F+ theorem is now straightforward. The proof is complete. hT (x ⊗ eγ), y ⊗ eωi = hϕ(W )(x ⊗ eγ), y ⊗ eωi n . Consequently, we obtain T (x⊗ eγ) = ϕ(W )(x ⊗ eγ). The last part of the (cid:3) Let F 2 f,m be the Hilbert space of formal power series in noncommutative indeterminates Z1, . . . , Zn with complete orthogonal basis {Zα : α ∈ F+ n} and such that kZαkf,m := 1√b(m) α . It is clear that F 2 f,m = ϕ := Xα∈F+ n aαZα : aα ∈ C and kϕk2 f,m := Xα∈F+ n 1 b(m) α aα2 < ∞ . The left multiplication operators L1, . . . , Ln are defined by Liξ := Ziξ operator Uf,m : F 2(Hn) → F 2 i = L(m) i Uf,m for any i ∈ {1, . . . , n}. A straightforward calculation reveals that T ∈ B(E ⊗ F 2(Hn)) is a weighted right multi-Toeplitz operator if and only if A := Uf,mT U ∗ f,m defined by Uf,m(eα) :=qb(m) for all ξ ∈ F 2 n , is unitary and Uf,mW (m) α Zα, α ∈ F+ f,m. Note that the f,m satisfies the condition hA(x ⊗ eγ), y ⊗ eωi = ω 1 1√b(m) qb(m) 0, γ (cid:10)Aω\r γ x, y(cid:11) , (cid:10)Bγ\rωx, y(cid:11) , if ω ≥r γ, if γ >r ω, otherwise, n n \{g0} in B(H). Note that the Hilbert space F 2 for some operators {A(α)}α∈F+ f,m can be and {B(α)}α∈F+ seen as a weighted Fock space. In the particular case when n = 1 and q = Z, it coincides with the weighted Bergman space Am(D), while A is a Toeplitz operator with operator-valued bounded harmonic symbol (see [16] for the scalar case when E = C). All the results of the present paper can be written in the setting of multi-Toeplitz operators on weighted Fock spaces. However, we preferred this time to put the weights on the left creation operators instead on the full Fock space. f ) the spatial tensor product B(E) ⊗min A(Dm f ), where A(Dm f ) is the noncom- mutative domain algebra. Let P ⊂ F 2(Hn) be the set of all polynomials in eα, α ∈ F+ n . We denote by AE (Dm The main result of this section is the following characterization of the weighted right multi-Toeplitz operators in terms of their Fourier representations, which can be viewed as their symbols. 10 GELU POPESCU Theorem 2.3. Let {A(α)}α∈F+ Then n n \{g0} be two sequences of operators on a Hilbert space E. and {B(α)}α∈F+ B(α) ⊗ W ∗ ϕ(W ) := Xα≥1 α + A(0) ⊗ I + Xα≥1 A(α) ⊗ Wα is the formal Fourier representation of a weighted right multi-Toeplitz operator T ∈ B(E ⊗ F 2(Hn)) if and only if (i) Pα≥1 (ii) 1 b(m) α A∗ (α)A(α) and Pα≥1 sup 0≤r<1kϕ(rW )k < ∞. Moreover, in this case, 1 b(m) α B(α)B∗ (α) are WOT convergent series, and (a) for each r ∈ [0, 1), the operator ϕ(rW ) := B(α) ⊗ rαW ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k ∞Xk=1 Xα=k A(α) ⊗ rαWα f )∗ + AE(Dm f ), where the series are convergent in the operator is in the operator space AE (Dm norm topology; (b) T = SOT- lim r→1 (c) kTk = sup ϕ(rW ), and 0≤r<1kϕ(rW )k = lim r→1kϕ(rW )k = sup q∈E⊗P,kqk≤1kϕ(W )qk. Proof. Assume that T ∈ B(E ⊗ F 2(Hn)) is a weighted right multi-Toeplitz operator and that ϕ(W ) is its formal Fourier representation. Note that part (i) was proved in the proof of Theorem 2.2. Using the definition of the weighted left creation operators and the fact that we deduce that α b(m) b(m) kWαk ≤ 1 αβ , β ≤(cid:18)β + m − 1 m − 1 (cid:19) b(m) α (cid:18)α + m − 1 m − 1 (cid:19)1/2 qb(m) α(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) b(m) α WαW ∗ . n ,α=k Consequently Since the operators Wα, α ∈ F+ n ,α = k have orthogonal ranges, we deduce that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα=k ≤(cid:18)k + m − 1 m − 1 (cid:19) . (α)A(α)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα∈F+ A(α) ⊗ rαWα(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = rk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα=k k=0(cid:13)(cid:13)(cid:13)Pα=k A(α) ⊗ rαWα(cid:13)(cid:13)(cid:13). A similar result holds for the which implies the convergence of the seriesP∞ operators B(α). Using part (i), we can easily see that ϕ(rW ) is in AE (Dm f ), where the series in the definition of ϕ(rW ) are convergent in the operator norm topology. This shows that part (a) holds. Now, we prove that, for any r ∈ [0, 1), (IE ⊗ K (m) (2.5) f,rW ξ = Xβ∈F+ f,rW : F 2(Hn) → F 2(Hn) ⊗ DrW is the noncommutative Berezin kernel defined by 1/2(cid:18)k + m − 1 m − 1 (cid:19)1/2 f,rW )∗(T ⊗ IF 2(Hn))(IE ⊗ K (m) ξ ∈ F 2(Hn), eβ ⊗ ∆1/2 f,rW ) = ϕ(rW ), f )∗ + AE (Dm where K (m) m,rW W ∗ β ξ, 1 b(m) α A∗ K (m) n MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 11 and DrW := ∆1/2 m,rW (F 2(Hn)). Let γ, ω ∈ F+ ϕq(W ) := X1≤α≤q n , set q := max{γ,ω} and define B(α) ⊗ W ∗ A(α) ⊗ Wα. α + Xα∈F+ n ,α≤q Note that W ∗ careful computation reveals that α(eγ) = 0 if α > q and, similarly, W ∗ β (eω) = 0 if β > q. Consequently, using Theorem 2.2, f,rW )(x ⊗ eγ), y ⊗ eωE f,rW eωE α(eγ) , Xβ∈F+ m,rW W ∗ n y ⊗ eβ ⊗ ∆1/2 m,rW W ∗ m,rW W ∗ α(eγ), ∆1/2 m,rW W ∗ y ⊗ eβ ⊗ ∆1/2 β (eω)+ β (eω)E m,rW W ∗ β (eω)+ f,rW eγ), y ⊗ K (m) x ⊗ eα ⊗ ∆1/2 α(eγ), Xβ∈F+ n n n m,rW W ∗ f,rW )∗(T ⊗ IF 2(Hn))(IE ⊗ K (m) D(IE ⊗ K (m) =D(T ⊗ IF 2(Hn))(x ⊗ K (m) =(T ⊗ IF 2(Hn))Xα∈F+ =Xα∈F+ T (x ⊗ eα) ⊗ ∆1/2 hT (x ⊗ eα), y ⊗ eβiD∆1/2 = Xα∈F+ n Xβ∈F+ n ,α≤q Xβ∈F+ = Xα∈F+ n ,α≤q Xβ∈F+ = Xα∈F+ = Xα∈F+ n Xβ∈F+ =D(ϕq(W ) ⊗ IF 2(Hn))(x ⊗ K (m) =D(IE ⊗ K (m) n ,β≤q n ,β≤q n n = hϕq(rW )(x ⊗ eγ), y ⊗ eωi = hϕ(rW )(x ⊗ eγ), y ⊗ eωi m,rW W ∗ hT (x ⊗ eα), y ⊗ eβiD∆1/2 hϕq(W )(x ⊗ eα), y ⊗ eβiD∆1/2 α(eγ), ∆1/2 m,rW W ∗ β (eω)E m,rW W ∗ α(eγ), ∆1/2 m,rW W ∗ β (eω)E hϕq(W )(x ⊗ eα), y ⊗ eβiD∆1/2 m,rW W ∗ f,rW eγ), y ⊗ K (m) f,rW )∗(ϕq(W ) ⊗ IF 2(Hn))(IE ⊗ K (m) m,rW W ∗ α(eγ), ∆1/2 β (eω)E f,rW eωE f,rW )(x ⊗ eγ), y ⊗ eωE for any x, y ∈ E. This shows that relation (2.5) holds. Since K (m) f,rW is an isometry for any r ∈ [0, 1), we deduce that (2.6) kϕ(rW )k ≤ kTk, r ∈ [0, 1), which completes the proof of part (ii). Now, we show that T = SOT- lim r→1 due to part (i) and the proof of Theorem 2.2, we deduce that ϕ(rW ). Indeed, first note that, 2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα∈F+ n (A(α) ⊗ Wα)(x ⊗ eβ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:18)β + m − 1 m − 1 (cid:19) Xα∈F+ n kA(α)xk2 1 b(m) α < ∞ and also a similar result involving the other series in the definition of ϕ(W ). Consequently, (2.7) for any p :=Pα≤k h(α) ⊗ eα, where h(α) ∈ E and k ∈ N. Let x ∈ E ⊗ F 2(Hn) and choose p as above such that kx − pk ≤ ǫ 2kT k . Using relation (2.6) and Theorem 2.2, we obtain kϕ(rW )p − ϕ(W )pk → 0, as r → 1, kϕ(rW )x − T xk ≤ kϕ(rW )(x − p)k + kϕ(rW )p − ϕ(W )pk + kϕ(W )p − T xk ≤ 2kTkkx − pk + kϕ(rW )p − ϕ(W )pk ≤ ǫ + kϕ(rW )p − ϕ(W )pk. 12 GELU POPESCU Now, relation (2.7) implies lim supr→1 kϕ(rW )x−T xk ≤ ǫ for any ǫ > 0. Hence limr→1 kϕ(rW )x−T xk = 0 for any x ∈ E ⊗ F 2(Hn), which proves part (b). To prove part (c), let ǫ > 0 and choose p ∈ E ⊗ F 2(Hn) be a polynomial such that kpk = 1 and kT pk > kTk − ǫ. Theorem 2.2 and relation (2.7) imply that there is t ∈ (0, 1) such that kϕ(tW )pk > kTk− ǫ. This shows that supr∈[0,1) kϕ(rW )k ≥ kTk. Since the reverse inequality holds due to relation (2.6), we deduce that (2.8) sup r∈[0,1)kϕ(rW )k = kTk. Now, let t1, t2 ∈ [0, 1) such that t1 < t2. Since ϕ(t2W ) is in AE (Dm Berezin transform to deduce that f,rW )∗(cid:0)ϕ(t2W ) ⊗ IF 2(Hn)(cid:1) (IE ⊗ K (m) (IE ⊗ K (m) for any r ∈ [0, 1). Taking r := t1 and employing the fact that K (m) t2 f,rW ) = ϕ(t2rW ) f,rW is an isometry, we obtain f ) we can use the noncommutative which together with relation (2.8) show that kTk = limr→1 kϕ(rW ). On the other hand, the fact that kTk = q∈E⊗P,kqk≤1kϕ(W )qk is a consequence of Theorem 2.2. sup kϕ(t1W )k ≤ kϕ(t2W )k, It remains to prove the converse of the theorem. Assume that {A(α)}α∈F+ n \{g0} are two sequences of operators on a Hilbert space E satisfying conditions (i) and (ii), where ϕ(W ) is the formal series and {B(α)}α∈F+ n Xα≥1 B(α) ⊗ W ∗ α + Xα∈F+ n A(α) ⊗ Wα. As is the first part of the proof, we can show that ϕ(rW ) is in the operator space AE(Dm and ϕ(W )p makes sense for any polynomial in E ⊗ F 2(Hn). Note that item (ii) implies (2.9) sup q∈E⊗P,kqk≤1kϕ(W )qk < ∞. f )∗ + AE (Dm f ), Indeed, if M > 0 and there is a polynomial p0 ∈ E ⊗ P such that kϕ(W )p0k > M . Using the fact that kϕ(rW )p0 − ϕ(W )p0k → 0 as r → 1, we find t ∈ (0, 1) such that kϕ(tW )p0k > M , which implies that kϕ(W )k > M . Since M is arbitrary, we get a contradiction. Therefore, relation (2.9) holds. Consequently, there is a unique operator T ∈ B(E ⊗F 2(Hn)) such that T p = ϕ(W )p for any polynomial p ∈ E ⊗F 2(Hn). Now one can easily see that relation (2.4) holds and hT (x ⊗ eγ), y ⊗ eωi = hϕ(W )(x ⊗ eα), y ⊗ eωi γ qb(m) qb(m) √b(m) qb(m) αγ αγ ω A(α)x, y+ , B(α)x, y+ , 0, if ω = αγ, if γ = αω, otherwise. qb(m) qb(m) γ ω ω\r γqb(m) √b(m) γ\r ω√b(m) qb(m) ω γ 0, (cid:10)T (x ⊗ 1), y ⊗ eω\rγ(cid:11) , (cid:10)T (x ⊗ eγ\rω), y ⊗ 1(cid:11) , if ω ≥r γ, if γ >r ω, otherwise. = =   This shows that T is a weighted right multi-Toeplitz operator and completes the proof. Corollary 2.4. The set of all weighted right multi-Toeplitz operators on E ⊗ F 2(Hn) coincides with (cid:3) AE (Dm f )∗ + AE (Dm f ) = AE (Dm f )∗ + AE (Dm f ) W OT SOT , where AE (Dm f ) := B(E) ⊗min A(Dm f ) and A(Dm f ) is the noncommutative domain algebra. MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 13 Proof. Let MT be the set of all weighted right multi-Toeplitz operators and note that the inclusion MT ⊆ AE (Dm holds due to Theorem 2.3. Since f )∗ + AE (Dm f ) SOT AE (Dm f )∗ + AE (Dm f ) SOT ⊆ AE (Dm f )∗ + AE(Dm f ) W OT , it remains to show that AE (Dm f )∗ + AE (Dm f ) W OT To this end, note that, for any operator A ∈ B(E) and α ∈ F+ α and A ⊗ Wα are multi-Toeplitz. On the other hand, if {Ti} is a net of weighted right multi-Toeplitz operators such that Ti → T in the weak operator topology, passing to the limit in relation (2.1), written for Ti, shows that T is a weighted right multi-Toeplitz operator as well. This completes the proof. (cid:3) ⊆ MT . n , the operators A ⊗ W ∗ Next, we show that for certain noncommutative domains Dm f , the corresponding set of weighted right multi-Toeplitz operators does not contain any nonzero compact operator. Theorem 2.5. Let Dm the condition f be a noncommutative domain where the coefficients b(m) α associated to f satisfy b(m) giα b(m) α sup α∈F+ n < ∞, i ∈ {1, . . . , n}. Then there is no nonzero compact weighted right multi-Toeplitz operator on the full Fock space F 2(Hn). Proof. First, note that (2.10) b(m) σα b(m) α sup α∈F+ n < ∞ for each σ ∈ F+ n . Indeed, if σ = gi1 ··· giq , i1, . . . , iq ∈ {1, . . . , n}, then bσγ bγ = bgi1 ···giq γ bgi2 ···giq γ = bgi2 ···giq γ bgi3 ···giq γ ··· bgiq γ bγ . Using now the condition in the theorem, the assertion follows. Assume that T is a compact weighted right multi-Toeplitz operator on F 2(Hn). Then, we have (2.11) hT eγ, eσγi =vuut b(m) hT eσγ, eγi =vuut b(m) hT eg0 , eσi hT eσ, eg0i σ b(m) γ b(m) σγ σ b(m) γ b(m) σγ n , and hT eα, eβi = 0 if (α, β) ∈ F+ for any σγ ∈ F+ n is not of the form (σγ, γ) or (γ, σγ) for σ, γ ∈ F+ n . Since a compact operator maps weakly convergent sequences to norm convergent sequences, we deduce that hT eγ, eσγi → 0 and hT eσγ, eγi → 0 as γ → ∞. Consequently, using relations (2.10) and (2.11), we conclude that hT eγ, eσγi = hT eσγ, eγi = 0 for any σ ∈ F+ n . Now, using again relation (2.11), we deduce that T = 0, which completes the proof. (cid:3) n × F+ We shall present a concrete class of noncommutative domains for which the theorem above holds. Con- f is the noncommutative sider the case when f = Z1 +··· + Zn and m ∈ N. The corresponding domain Dm m-hyperball, which is defined by (cid:8)X := (X1, . . . , Xn) ∈ B(H)n : (id − ΦX )k(I) ≥ 0 for 1 ≤ k ≤ m(cid:9) , 14 GELU POPESCU where ΦX : B(H) → B(H) is defined by ΦX (Y ) :=Pn g0 = 1 and b(m) b(m) α =(cid:18)α + m − 1 n ,α ≥ 1. Consequently, m − 1 (cid:19) if α ∈ F+ = (cid:18)α + m m − 1(cid:19) (cid:18)α + m − 1 m − 1 (cid:19) → 1, b(m) αgi b(m) α as α → ∞. i=1 XiY X ∗ i for Y ∈ B(H). In this case, we have This shows that Theorem 2.5 holds for the weighted right multi-Toeplitz operators associated with the noncommutative m-hyperball. 3. Free pluriharmonic functions on the noncommutative domain Dm f,rad In this section, we provide basic results concerning the free pluriharmonic functions on the noncom- mutative domain Dm f,rad(H) and show that they are characterized by a mean value property. This result is used to obtain an analogue of Weierstrass theorem for free pluriharmonic functions and to show that the set of all pluriharmonic functions is a complete metric space with respect to an appropriate metric. A Schur type result in this setting is also presented. Since the domain Dm we can introduce the radial part of the domain Dm f is radial (see [30]), i.e. rX ∈ Dm f (H), i.e. f (H) for any X ∈ Dm f (H) and any r ∈ [0, 1), Note that, in general, we have tDm f (H) ⊆ Dm f (H). Dm f,rad(H) := [0≤t<1 f (H) ⊆ Dm IntDm f,rad(H) ⊆ Dm f (H) ⊆ Dm f,rad(H)−. In the particular case when q is a positive regular noncommutative polynomial, we have IntDm q (H) = Dm q,rad(H) and Dm q,rad(H)− = Dm q (H) = Dm q (H)−. Zg0 := 1. function on the abstract domain Dm Let Z1, . . . , Zn be noncommutative indeterminates, set Zα := Zi1 ··· Zik if α = gi1 ··· gik ∈ F+ Definition 3.1. A formal power series F := Pα∈Fn f,rad :=`H Dm ∞Xk=1 Xα∈F+ A(α) ⊗ Zα with A(α) ∈ B(E) is called free holomorphic f,rad(H), if the series A(α) ⊗ Xα F (X) := n , and n ,α=k is convergent in the operator norm topology for any X ∈ Dm f,rad(H) and any Hilbert space H. We remark that it is enough to assume in Definition 3.1 that H is an arbitrary infinite dimensional separable Hilbert space. Unless otherwise specified, we assume throughout this paper that H has this property. f ) the set of all free holomorphic functions on the abstract domain Dm f,rad with A(α) ⊗ Zα with A(α) ∈ B(E) be a formal power series and We denote by HolE (Dm define γ ∈ [0,∞] by setting operator coefficients in B(E). Let F := Pα∈Fn k∈Z+ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα∈F+ := lim sup 1 γ n ,α=k 1 k . A(α) ⊗ Wα(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 15 According to [31], the series ∞Xk=1 is convergent. Moreover, if γ > 0 and r ∈ [0, γ), then the convergence is uniform on rDm if γ ∈ [0,∞) and s > γ, then there is a Hilbert space H and Y ∈ sDm f (H) such that the series f (H). In addition, is divergent in the operator norm topology. As a consequence of these results, we deduce the following. Proposition 3.2. F = Pα∈Fn A(α) ⊗ Zα is free holomorphic on Dm f,rad if and only if , X ∈ γDm f,rad(H), (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xα=k A(α) ⊗ Xα(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞Xk=1 Xα=k A(α) ⊗ Yα ωβA∗ n ,β=k 1/2k ≤ 1, (β)A(β)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) where ωβ := supγ∈F+ n b(m) γ b(m) βγ and b(m) is given by relation (1.1). Proof. Due to the results preceding the proposition, we have that F is free holomorphic on Dm f,rad if and γ 1 k lim sup k∈Z+ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ A(α) ⊗ Wα(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) β eα = =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ Wβ W ∗ n ,β=k only if lim supk∈Z+(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Pα∈F+ n ,α=k which implies kWβW ∗ we deduce that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ n ,β=k A(β) ⊗ Wβ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) The proof is complete. ≤ 1. On the other hand, due to relation (1.2), we have qb(m) γ√b(m) 0 α eα if α = βγ, otherwise, βk = ωβ. Since the operators Wβ , with β ∈ F+ n and β = k, have orthogonal ranges, A∗ (β)A(β) ⊗ WβW ∗ 1/2 β(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ n ,β=k ωβA∗ 1/2 (β)A(β)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . (cid:3) Definition 3.3. A map G : Dm f,rad(H) → B(E) ⊗min B(H) is called self-adjoint free pluriharmonic function on Dm f,rad(H) such that G = ℜF . Any linear combination of self-adjoint free pluriharmonic functions is called free pluriharmonic function. f,rad(H) with coefficients in B(E) if there is a free holomorphic function F on Dm n A(α) ⊗ Zα, the relation above implies We remark that if G = ℜF as in the latter definition, then G determines F up to an operator A(g0) ∈ B(E) with ℜA(g0) = 0. Indeed, assume that ℜF = 0. Then F (rW ) = −F (rW )∗, r ∈ [0, 1). If F has the representation F =Pα∈F+ Hence, we deduce that A(α) = 0 if α ∈ F+ n with α ≥ 1 and ℜA(g0) = 0. Therefore, F = A(g0) ⊗ I. On the other hand, it is easy to see that any free pluriharmonic function H has a representation of the form H = H1 + iH2, where H1 and H2 are self-adjoint free pluriharmonic functions. Note also that any free 2 + i F −F ∗ holomorphic function F is a free pluriharmonic function, due to the decomposition F = F +F ∗ . Using Proposition 3.2, one can easily prove the following characterization of free pluriharmonic func- F (rW )(x ⊗ 1) = −F (rW )∗(x ⊗ 1) = −A∗ x ∈ E. (g0)x, 2i tions on Dm f,rad. 16 GELU POPESCU f,rad(H) with coefficients in B(E) if and only if there exist two sequences {A(α)}α∈F+ f,rad(H) → B(E) ⊗min B(H) is a free pluriharmonic function on n ⊂ B(E) and Proposition 3.4. A map G : Dm Dm {B(α)}α∈F+ n \{g0} ⊂ B(E) such that lim sup k∈Z+ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and Xβ∈F+ n ,β=k ωβA∗ (β)A(β)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞Xk=1 Xα=k 1/2k ≤ 1 and ωβB(β)B∗ n ,β=k 1/2k ≤ 1 (β)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lim sup k∈Z+ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xβ∈F+ ∞Xk=1 Xα=k G(X) = A(α) ⊗ Xα, where the series are convergent in the operator norm topology for any X ∈ Dm space H. Moreover, the representation of G is unique. α + A(0) ⊗ I + B(α) ⊗ X ∗ f,rad(H) and any Hilbert f,X (m) where the K (m) f (H), where X is a pure element, is the X [g] :=(cid:16)IE ⊗ K (m) X : B(E ⊗ F 2(Hn)) → B(E ⊗ H) defined by eB f,X : H → F 2(Hn) ⊗ H is noncommutative Berezin kernel. The extended noncommutative Berezin transform at X ∈ Dm ∗(cid:17) (g ⊗ IH)(cid:16)IE ⊗ K (m) f,X(cid:17) , map eB(m) We denote by PE (W ) the set of all operators of the form Pα≤k A(α) ⊗ Wα, where k ∈ N and A(α) ∈ B(E). The following result extends Theorem 2.4 from [30]. We include it for completeness. Theorem 3.5. If X ∈ Dm tractive linear map Φf,X : SE → B(E) ⊗min B(H) such that rX [ϕ], f (H) and SE := PE (W )∗ + PE (W ) , then there is a unital completely con- g ∈ B(E ⊗ F 2(Hn)), Φf,X (ϕ) = lim k·k ϕ ∈ SE , where the limit is in the operator norm topology. If, in addition, X is a pure n-tuple in Dm f (H), then r→1eB(m) rX [ϕ] = eB(m) lim r→1eB(m) X [ϕ], ϕ ∈ SE . Proof. Let ϕ ∈ SE and let {qk(W, W ∗)}∞ k=1 ⊂ PE(W )∗ + PE (W ) be such that qk(W, W ∗) → ϕ in the operator norm, as k → ∞. For any X ∈ Dm f (H), the noncommutative von Neumann inequality (see [27]) implies kqk(X, X ∗) − qj(X, X ∗)k ≤ kqk(W, W ∗) − qj(W, W ∗)k for any k, j ∈ N. Consequently, since {qk(W, W ∗)}∞ (3.1) k=1 is a Cauchy sequence, so is the sequence {qk(X, X ∗)}∞ X ∈ Dm exists in the operator norm and kΦf,X (ϕ)k ≤ kϕk. Now, we show that ϕ ∈ SE , k=1. Therefore, f (H), Φf,X (ϕ) := lim k→∞ Φf,X (ϕ) = lim qk(X, X ∗), (m) rX [ϕ], where the limit is in the operator norm topology. Since r→1eB relation (3.1) implies rX [qk(W, W ∗)] = (IE ⊗ K (m) eB(m) f,rX )∗(qk(W, W ∗) ⊗ IH)(IE ⊗ K (m) f,rX ) = qk(rX, rX ∗), (3.2) Given ǫ > 0, let N ∈ N be such that kqN (W, W ∗) − ϕk < ǫ Φf,rX (ϕ) = eB (m) rX [ϕ], r ∈ [0, 1). 3 . Note that Φf,rX(ϕ) − qN (rX, rX ∗) = lim k→∞ qk(rX, rX ∗) − qN (rX, rX ∗) and (3.3) kΦf,rX (ϕ) − qN (rX, rX ∗)k ≤ kϕ − qN (W, W ∗)k < ǫ 3 , r ∈ [0, 1]. MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 17 Note also that there is a δ ∈ (0, 1) such that (3.4) kqN (rX, rX ∗) − qN (X, X ∗)k < Using the relations (3.1), (3.2), (3.3) and (3.4), we obtain rX [ϕ]k = kΦf,X (ϕ) − Φf,rX (ϕ)k kΦf,X(ϕ) −eB(m) ǫ 3 , r ∈ (δ, 1). ≤ kΦf,X (ϕ) − qN (X, X ∗)k + kqN (X, X ∗) − qN (rX, rX ∗)k for any r ∈ (δ, 1). Therefore, where the limit is in the operator norm topology. Similarly, one can prove that, for any k × k matrix [ϕij]k×k with entries in SE , in the operator norm. Using the properties of the noncommutative Berezin kernel, we deduce that Φf,X (ϕ) = lim + kqN (rX, rX ∗) − Φf,rX (ϕ)k < ǫ r→1eB ϕ ∈ SE , (m) rX [ϕ], [Φf,X (ϕij )]k×k = lim r→1heB(m) rX [ϕij ]ik×k (cid:13)(cid:13)(cid:13)[Φf,X (ϕij )]k×k(cid:13)(cid:13)(cid:13) ≤ k[ϕij]k×kk . This proves that Φf,X is a unital completely contractive linear map. Now, assume that X ∈ Dm pure n-tuple of operators. Then we have f (H) is a X [qk(W, W ∗)] = (IE ⊗ K (m) eB(m) f,X )∗(qk(W, W ∗) ⊗ IH)(IE ⊗ K (m) f,X ) = qk(X, X ∗). Since qk(W, W ∗) → ϕ in the operator norm, as k → ∞, we can use the relation above and (3.1) to deduce that The proof is complete. eB(m) X [ϕ] = lim k→∞ qk(X, X ∗) = Φf,X (ϕ) = lim rX [ϕ]. r→1eB(m) (cid:3) Next, we show that the free pluriharmonic functions are characterized by a mean value property. Theorem 3.6. If G : Dm mean value property, i.e f,rad(H) → B(E) ⊗min B(H) is a free pluriharmonic function, then it has the r X [G(rW )], X ∈ rDm f )∗ + An(Dm f ) k·k f (H), r ∈ (0, 1). satisfies the relation Conversely, if a function ϕ : [0, 1) → An(Dm t W [ϕ(t)], r then the map F : Dm f,rad(H) → B(E) ⊗min B(H) defined by for any 0 ≤ r < t < 1, r X [ϕ(r)], 1 X ∈ rDm f (H), r ∈ (0, 1), 1 G(X) = eB(m) ϕ(r) = eB(m) F (X) := eB(m) is a free pluriharmonic function. Moreover, F (rW ) = ϕ(r) for any r ∈ [0, 1). In particular, F ≥ 0 if and only if ϕ ≥ 0. Proof. Assume that G : Dm tion f,rad(H) → B(E) ⊗min B(H) is a free pluriharmonic function with representa- G(X) = ∞Xk=1 Xα=k B(α) ⊗ X ∗ α + A(0) ⊗ I + A(α) ⊗ Xα, ∞Xk=1 Xα=k where the series are convergent in the operator norm topology for any X ∈ Dm space H. Since the universal model W = (W1, . . . , Wn) is in Dm f (F 2(Hn)), for any r ∈ [0, 1), we have f,rad(H) and any Hilbert G(rW ) = ∞Xk=1 Xα=k B(α) ⊗ rαW ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k A(α) ⊗ rαWα, 18 GELU POPESCU where the convergence is in the operator norm topology. If X ∈ Dm and f,rad(H), r ∈ (0, 1), then 1 r X ∈ Dm f (H) 1 eB(m) r X [G(rW )] = lim = lim δ→1 δ→1eB(m) δ r X [G(rW )] G(δX) = G(X), where the limits are in the operator norm topology. The latter equality is due to the continuity of G on rDm the relation f (H). To prove the converse, assume that the function ϕ : [0, 1) → AE (Dm f )∗ + AE(Dm f ) satisfies k·k Due to Theorem 2.3 and Corollary 2.4, ϕ(r) is a weighed right multi-Toeplitz operator and has a unique formal Fourier representation t W [ϕ(t)], for any 0 ≤ r < t < 1. r ϕ(r) = eB(m) Xα≥1 rαB(α)(r) ⊗ W ∗ rαA(α)(r) ⊗ Wα for some operators {A(α)(r)}α∈F+ n and {B(α)(r)}α∈F+ ϕδ(r) := Xα≥1 rαB(α)(r) ⊗ δαW ∗ n \{g0}. Moreover, setting rαA(α)(r) ⊗ δαWα, δ ∈ [0, 1), where the convergence of the series is in the operator norm topology, we have (3.5) ϕ(r) = SOT- lim δ→1 ϕδ(r) and sup δ∈[0,1)kϕδ(r)k = kϕ(r)k. Due to similar reasons, ϕ(t) is a weighted right multi-Toeplitz operator and has a unique formal Fourier representation Xα≥1 tαB(α)(t) ⊗ W ∗ tαA(α)(t) ⊗ Wα for some operators {A(α)(t)}α∈F+ n and {B(α)(t)}α∈F+ ψδ(t) := Xα≥1 tαB(α)(t) ⊗ δαW ∗ n \{g0}. Moreover, setting tαA(α)(t) ⊗ δαWα, δ ∈ [0, 1), where the convergence of the series is in the operator norm topology, we have ϕ(t) = SOT- lim δ→1 ψδ(t) and sup δ∈[0,1)kψδ(t)k = kϕ(t)k. Now, since the map Y → Y ⊗ I is SOT-continuous on bounded sets, so is the noncommutative Berezin t W . Using the results above, we deduce that α + Xα≥0 α + Xα≥0 α + Xα≥0 α + Xα≥0 r transform eB(m) eB = SOT- lim (m) t W [ψδ(t)] r (m) t W [ϕ(t)] = SOT- lim r δ→1eB δ→1Xα≥1 Consequently, using the fact that ϕ(r) = eB(m) rαB(α) ⊗ W ∗ B(α)(r) for any α ∈ F+ r rαB(α)(t) ⊗ δαW ∗ α + Xα≥0 rαA(α)(t) ⊗ δαWα . t W [ϕ(t)] and relation (3.5), we can easily see that B(α)(t) = n with α ≥ 1 and A(α)(t) = A(α)(r) for any α ∈ F+ ϕ(r) = Xα≥1 for some operators {A(α)}α∈F+ topology. Now, for any X ∈ rDm and {B(α)}α∈F+ f (H), r ∈ (0, 1), we define α + Xα≥0 rαA(α) ⊗ Wα, n n . Therefore, r ∈ [0, 1), n \{g0}, where the convergence is in the operator norm F (X) := eB(m) 1 r X [ϕ(r)]. MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS Hence, we deduce that F (X) = Xα≥1 B(α) ⊗ X ∗ α + Xα≥0 A(α) ⊗ Xα, X ∈ Dm f,rad(H), and F (rW ) = ϕ(r) for any r ∈ [0, 1). The proof is complete. 19 (cid:3) E (Dm A(α) ⊗ Zα, we associated the kernel ΓrF : F+ f,rad) be the set of all free holomorphic functions F ∈ HolE (Dm n → B(E) defined by n × F+ f,rad) such that ℜF ≥ 0. If Let Hol+ F := Pα∈Fn ΓrF (ω, γ) := rαA(α), γ αγ qb(m) qb(m) A(g0) + A∗ √b(m) ω√b(m) 0, αω (g0), rαA∗ (α), if ω = αγ, α ≥ 1, if ω = γ, if γ = αω, α ≥ 1, otherwise.  We will use the notation S+ E (Dm such that the weighted multi-Toeplitz kernels ΓrF : F+ f,rad) for the set of all free holomorphic functions F ∈ HolE (Dm f,rad) n → B(E), r ∈ [0, 1), are positive semidefinite. n × F+ Now, we prove a Schur type result for free pluriharmonic functions with positive real parts. A(α) ⊗ Zα and let hβ ∈ E, for n Consequently, F (rW )∗ + F (rW ) ≥ 0 for any r ∈ [0, 1) if and only if ΓrF is a positive semidefinite kernel for any r ∈ [0, 1). Theorem 3.7. Hol+ E (Dm E (Dm f,rad). f,rad) = S+ E (Dm Proof. Assume that F ∈ Hol+ β ∈ F+ hγ ⊗ eγ+ hγ ⊗ eγ+ eαβ, eγ+ β = ∞Xs=1 Xα=s n with β ≤ q. Note that, using relation (1.2), we have f,rad) has the representation F =Pα∈F+ hβ ⊗ eβ , Xγ≤q A(α) ⊗ rαWαXβ≤q ∞Xs=1 Xα=sXβ≤q A(α)hβ ⊗ rαWαeβ, Xγ≤q rα(cid:10)A(α)hβ, hγ(cid:11)qb(m) n ,α≥1 Xβ,γ≤q = Xα∈F+ qb(m) rγ\rβ(cid:10)A(γ\rβ)hβ, hγ(cid:11)qb(m) = Xβ,γ≤q,γ>rβ qb(m) = Xβ,γ≤q,γ>rβ αXβ≤q * ∞Xs=1 Xα=s (F (rW )∗ + F (rW ))Xβ≤q hβ ⊗ eβ , Xγ≤q hβ ⊗ eβ , Xγ≤q hγ ⊗ eγ+ = Xβ,γ≤q hΓrF (γ, β)hβ, hγi . A(α) ⊗ rαW ∗ αβ β γ hγ ⊗ eγ+ = Xβ,γ≤q,γ>rβ In a similar manner, one can prove that hΓrF (γ, β)hβ, hγi . hΓrF (γ, β)hβ, hγi . Note also that, for any β ∈ Fn, ΓrF (β, β) = ΓrF (g0, g0) = Ag0 + A∗ above, we deduce that g0 .Taking into account the relations 20 GELU POPESCU On the other hand, if X ∈ Dm f,rad(H), then there is r ∈ (0, 1) such that X ∈ rDm f (H). Due to Theorem 3.6, we have Since the noncommutative Berezin transform is a positive map, we deduce that F (X)∗ + F (X) ≥ 0 for any X ∈ Dm f,rad(H) whenever F (rW )∗ + F (rW ) ≥ 0 for any r ∈ [0, 1). The converse is obviously true. Putting all these things together we complete the proof. (cid:3) F (X)∗ + F (X) = eB(m) r X [F (rW )∗ + F (rW )] . 1 The next result is an analogue of Weierstrass theorem for free pluriharmonic functions on the noncom- mutative domain Dm f,rad(H). Theorem 3.8. Let Fk : Dm f,rad(H) → B(E) ⊗min B(H), k ∈ N, be a sequence of free pluriharmonic functions such that, for any r ∈ [0, 1), the sequence {Fk(rW}∞ k=1 is convergent in the operator norm topology. Then there is a free pluriharmonic function F : Dm f,rad(H) → B(E) ⊗min B(H) such that Fk(rW ) converges to F (rW ), as k → ∞, for any r ∈ [0, 1). In particular, Fk converges to F uniformly on any domain rDm f (H), r ∈ [0, 1). Proof. Assume that Fk has the representation B(α)(k) ⊗ X ∗ Fk(X) = Xα≥1 α + Xα≥0 A(α)(k) ⊗ Xα, X ∈ Dm f,rad(H), for some operators {A(α)(k)}α∈F+ topology. According to Theorem 2.3, for any r ∈ [0, 1), Fk(rW ) is in the operator space AE (Dm AE (Dm n \{g0}, where the convergence is in the operator norm f )∗ + and {B(α)(k)}α∈F+ f ). Define the function ϕ : [0, 1) → AE (Dm by setting k·k n f )∗ + AE (Dm f ) Fk(rW ), r ∈ [0, 1). (3.6) ϕ(r) := lim k→∞ Let 0 ≤ r < t < 1 and note that t W [Fk(tW )] = lim k→∞ Fk(rW ) = ϕ(r), r k→∞eB(m) r t W [ϕ(t)] = lim eB(m) F (X) := eB(m) 1 where the limits are in the operator norm topology. According to Theorem 3.6 the map F : Dm B(E) ⊗min B(H) defined by f,rad(H) → r X [ϕ(r)], X ∈ rDm f (H), r ∈ (0, 1), is a free pluriharmonic function. and F (rW ) = ϕ(r) for any r ∈ [0, 1). Using relation (3.6), we obtain F (rW ) = limk→∞ Fk(rW ), r ∈ [0, 1). Since sup f (H)kF (X) − Fk(X)k = kF (rW ) − Fk(rW )k, X∈rDm we deduce that Fk converges to F uniformly on any domain rDm Corollary 3.9. Let Fk : Dm functions such that {Fk(0)} is a convergent sequence in the operator norm topology and f (H), r ∈ [0, 1). The proof is complete. (cid:3) f,rad(H) → B(E) ⊗min B(H), k ∈ N, be a sequence of free pluriharmonic F1 ≤ F2 ≤ ··· . Then Fk converges to a free pluriharmonic function on Dm f,rad(H). Proof. We may assume that F1 ≥ 0, otherwise we take Gk := Fk − F1, k ∈ N. Due to Harnack type inequality for positive free pluriharmonic functions on Dm f,rad(H) (see [31]), if k ≥ q, then we have kFk(X) − Fq(X)k ≤ kFk(0) − Fq(0)k 1 − r 1 + r for any X ∈ rDm f (H). Since {Fk(0)} is a Cauchy sequence in the operator norm, we deduce that {Fk} is a uniformly Cauchy sequence on rDm f (H). Hence {Fk(rW )} is a Cauchy sequence and, therefore, convergent in the operator norm topology. Applying Theorem 3.8, we find a free pluriharmonic function F : Dm f,rad(H) → B(E) ⊗min B(H) such that Fk(rW ) converges to F (rW ), as k → ∞, for any r ∈ [0, 1). In particular, Fk converges to F uniformly on any domain rDm f (H), r ∈ [0, 1). The proof is complete. (cid:3) MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 21 Let HarE (Dm f,rad) denote the set of all free pluriharmonic functions F : Dm f,rad(H) → B(E)⊗min B(H). If F, G ∈ HarE (Dm f,rad) and 0 < r < 1, we define If H is an infinite dimensional Hilbert space, the noncommutative von Neumann inequality for the n-tuples in the domain Dm f (H) implies dr(F, G) := kF (rW ) − G(rW )k. dr(F, G) = sup f (H)kF (X) − G(X)k. X∈rDm Let {rm}∞ we define m=1 be an increasing sequence of positive numbers convergent to 1. For any F, G ∈ HarE (Dm f,rad), ρ(F, G) := 2(cid:19)k ∞Xk=1(cid:18) 1 drk (F, G) 1 + drk (F, G) . Using standards arguments, one can show that ρ is a metric on HarE (Dm f,rad). Theorem 3.10. (cid:16)HarE (Dm f,rad), ρ(cid:17) is a complete metric space. Proof. It is easy to see that if ǫ > 0, then there exists δ > 0 and N ∈ N such that, for any F, G ∈ HarE(Dm f,rad), drN (F, G) < δ =⇒ ρ(F, G) < ǫ. Conversely, if δ > 0 and N ∈ N are fixed, then there is ǫ > 0 such that, for any F, G ∈ HarE(Dm f,rad), ρ(F, G) < ǫ =⇒ drN (F, G) < δ. Let {Gk}∞ k=1 ⊂ HarE (Dm f,rad) be a Cauchy sequence in the metric ρ. A consequence of the remark above is that {Gk(rN W )}∞ k=1 is a Cauchy sequence in B(E ⊗ F 2(Hn)), for any N ∈ N. Consequently, for each N ∈ N, the sequence {Gk(rN W )}∞ k=1 is convergent in the operator norm. Using Theorem 3.8, we find a free pluriharmonic function G ∈ HarE (Dm f,rad) such that Gk(rW ) converges to G(rW ) for any r ∈ [0, 1). By the observation made at the beginning of this proof, we conclude that ρ(Gk, G) → 0, as k → ∞, which completes the proof. (cid:3) 4. Bounded free pluriharmonic functions and Dirichlet extension problem In this section, we characterize the bounded pluriharmonic functions on Dm f,rad(H) as noncommutative Berezin transforms of weighted right multi-Toeplitz operators and present a noncommutative version of Dirichlet extension problem. Let us recall some definitions concerning completely bounded maps on operator spaces. We identify Mk(B(H)), the set of k × k matrices with entries in B(H), with B(H(k)), where H(k) is the direct sum of k copies of H. If X is an operator space, i.e., a closed subspace of B(H), we consider Mk(X ) as a subspace of Mk(B(H)) with the induced norm. Let X ,Y be operator spaces and u : X → Y be a linear map. Define the map uk : Mk(X ) → Mk(Y) by uk([xij ]k) := [u(xij )]k. We say that u is completely bounded if kukcb := supk≥1 kukk < ∞. When kukcb ≤ 1 (resp. uk is an isometry for any k ≥ 1) then u is completely contractive (resp. isometric). We call u completely positive if uk is positive for all k ≥ 1. For more information on completely bounded (resp. positive) maps, we refer to [19] and [20]. A free pluriharmonic function G : Dm f,rad(H) → B(E) ⊗min B(H) is called bounded if kGk := supG(X)k < ∞, where the supremum is taken over all n-tuples X ∈ Dm noncommutative von Neumann inequality for elements in Dm this section, that the Hilbert space H is separable and infinite dimensional. Denote by Har∞ the set of all bounded free pluriharmonic functions on Dm f,rad(H) and any Hilbert space H. Due to the f,rad(H), it is enough to assume, throughout f,rad) f,rad with coefficients in B(E), where E is a E (Dm 22 GELU POPESCU by setting separable Hilbert space. For each k = 1, 2, . . ., we define the norms k·kk : Mk(cid:16)Har∞ where the supremum is taken over all n-tuples X ∈ Dm that the norms k · kk, k = 1, 2, . . ., determine an operator space structure on Har∞ of Ruan (see e.g. [8]). k[Fij ]kkk := supk[Fij(X)]kk, f,rad(H) and any Hilbert space H. It is easy to see f,rad), in the sense E (Dm E (Dm f,rad)(cid:17) → [0,∞) Theorem 4.1. If F : Dm f,rad(H) → B(E) ⊗min B(H), then the following statements are equivalent: (i) F is a bounded free pluriharmonic function on Dm (ii) there exists ψ ∈ AE (Dm f )∗ + AE (Dm f ) SOT f,rad(H); X is the noncommutative Berezin transform at X. In this case, ψ = SOT- lim r→1 such that F (X) = eB(m) X [ψ] for X ∈ Dm f,rad(H), F (rW ). More- where eB(m) over, the map Φ : Har∞(Dm f,rad) → AE(Dm f )∗ + AE (Dm f ) SOT defined by Φ(F ) := ψ f ) is the noncommutative domain algebra. is a completely isometric isomorphism of operator spaces, where AE (Dm A(Dm Proof. Let F ∈ Dm f,rad(H) and note that, due to Proposition 3.4, it has a representation f ) := B(E) ⊗min A(Dm f ) and F (X) = ∞Xk=1 Xα=k B(α) ⊗ X ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k A(α) ⊗ Xα, where the series are convergent in the operator norm topology for any X ∈ Dm have F (rW ) ∈ AE (Dm 2.3, we find a unique weighted right multi-Toeplitz operator T ∈ B(E ⊗ F 2(Hn)) such that (4.1) f,rad(H). Consequently, we f ) for any r ∈ [0, 1), and supr∈[0,1) kF (rW )k < ∞. Applying Theorem f )∗ + AE (Dm F (rW ) and T = SOT- lim r→1 kTk = sup r∈[0,1)kF (rW )k. Therefore, T ∈ AE(Dm Indeed, since F (rW ) ∈ AE (Dm f )∗ + AE (Dm f ) f )∗ + AE (Dm SOT f ), we have . Now, we prove that F (X) = eB(m) f,X )∗(F (rW ) ⊗ IH)(IE ⊗ K (m) f,X ) F (rX) = (IE ⊗ K (m) X [T ] for X ∈ Dm f,rad(H). for X ∈ Dm hand, since F is continuous on Dm f,rad(H) and r ∈ [0, 1). Since the map Y 7→ Y ⊗ I is SOT-continuous on bounded sets, we use relation (4.1) to deduce that SOT- limr→1 F (rX) = eB(m) f,rad(H). On the other f,rad(H) with respect to the operator norm topology, we conclude that F (X) = eB(m) f,rad(H), which shows that item (ii) holds. Conversely, assume that (ii) holds. Then F (X) = eB . Due to Corollary 2.4, ψ is a weighted right multi-Toeplitz operator on E ⊗ F 2(Hn). Applying Theorem 2.3, we find be two sequences {A(α)}α∈F+ n \{g0} of operators on a Hilbert space E such that, ψ = SOT- limr→∞ G(rW ), where f,rad(H) and some ψ ∈ AE (Dm X [ψ] for any X ∈ Dm X [T ] for X ∈ Dm X [T ] for X ∈ Dm and {B(α)}α∈F+ f )∗ + AE (Dm f ) SOT (m) n G(rW ) := ∞Xk=1 Xα=k B(α) ⊗ rαW ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k A(α) ⊗ rαWα, with the convergence is in the operator norm topology. Moreover, we have supr∈[0,1) kG(rW )k = kψk. Define the bounded free pluriharmonic function G : Dm f,rad(H) → B(E) ⊗min B(H) by setting G(X) = ∞Xk=1 Xα=k B(α) ⊗ X ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k A(α) ⊗ Xα, , X ∈ Dm f,rad(H), MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 23 where the series are convergent in the operator norm topology. Note that X [ψ] = SOT- lim X [G(rW )] eB(m) r→1eB(m) r→1 ∞Xk=1 Xα=k = SOT- lim r→1 G(rX) = G(X), = SOT- lim rαB(α) ⊗ X ∗ α + A(0) ⊗ I + ∞Xk=1 Xα=k A(α) ⊗ rαXα where the last equality is due to the continuity of G. Therefore, G(X) = F (X) for any X ∈ Dm f,rad(H). f,rad). As in the case when k = 1, we can Now, let [Fij ]k×k be a k × k matrix with entries in Har∞ use the noncommutative von Neumann inequality for the domain Dm r∈[0,1)k[Fij (rW )]k×kk k[Fij ]k×kk = sup E (Dm f , to show that and that Tij := SOT- limr→1 Fij (rW ) are weighted right multi-Toeplitz operators. Since (IE ⊗ K (m) f,rW )∗(Tij ⊗ IH)(IE ⊗ K (m) f,rW ) = Fij (rW ), r ∈ [0, 1), we deduce that k[Fij(rW )]k×kk ≤ k[Tij]k×kk, r ∈ [0, 1), which, due to the convergence above, implies k[Fij(rW )]k×kk = k[Tij]k×kk. This completes the proof. (cid:3) A consequence of Theorem 4.1 and Corollary 2.4 is the following noncommutative version of Herglotz theorem (see [12], [13]). Corollary 4.2. Any non-negative bounded free pluriharmonic function on Dm of a positive weighted right multi-Toeplitz operator on E ⊗ F 2(Hn). Corollary 4.3. If F : Dm Y ∈ Dm Proof. Assume that F has the representation f (H) is a pure n-tuple of operators, then limr→1 F (rY ) exists in the strong operator topology. f,rad(H) → B(E) ⊗min B(H) is a bounded free pluriharmonic function and f,rad is the Berezin transform F (X) = ∞Xk=1 Xα=k B(α) ⊗ X ∗ α + ∞Xk=0 Xα=k A(α) ⊗ Xα, X ∈ Dm f,rad(H), where the series are convergent in the operator norm topology. Due to Theorem 4.1, we find a unique weighted right multi-Toeplitz operator T ∈ B(E ⊗ F 2(Hn)) such that (4.2) kTk = sup T = SOT- lim r→1 and Let Y ∈ Dm F (rW ) r∈[0,1)kF (rW )k. f (H) be a pure n-tuple of operators and let r ∈ [0, 1). Then we have A(α) ⊗ rαYα B(α) ⊗ rαY ∗ F (rY ) = α + ∞Xk=0 Xα=k ∞Xk=1 Xα=k = B(m) = (IE ⊗ K (m) Y [F (rW )] f,Y )∗(F (rW ) ⊗ IH)(IE ⊗ K (m) f,Y ), where the convergence of the series is in the operator norm topology. Consequently, since the map A 7→ A⊗ I is SOT-continuous on bounded sets, relation (4.2) implies that SOT- limr→1 F (rY ) exists and it is equal to (IE ⊗ K (m) f,Y )∗(T ⊗ IH)(IE ⊗ K (m) Corollary 4.4. Given a function F : Dm equivalent: f,rad(H) → B(E) ⊗min B(H), the following statements are f,Y ). The proof is complete. (cid:3) (i) F is a bounded free plurihamonic function. t W [F (tW )], 0 ≤ r < t < 1. F (X) = lim r k·k F (rW ) = eB(m) δ→1 ∞Xk=1 Xα=k δ→1eB(m) [ϕ(r)] = eB(m) δ r X 1 = lim B(α) ⊗ δαX ∗ α + r X [ϕ(r)]. ∞Xk=0 Xα=k A(α) ⊗ δαXα 24 GELU POPESCU (ii) There is a bounded function ϕ : [0, 1) → AE (Dm f )∗ + AE (Dm f ) k·k which satisfies the relation t W [ϕ(t)], r ϕ(r) = eB(m) r X [ϕ(r)] for any X ∈ rDm f (H) and r ∈ (0, 1). for any 0 ≤ r < t < 1, and F (X) := eB(m) 1 Moreover, F and ϕ uniquely determine each other and F (rW ) = ϕ(r) for any r ∈ [0, 1). Proof. Assume that F is a bounded free pluriharmonic function and has representation F (X) = ∞Xk=1 Xα=k B(α) ⊗ X ∗ α + ∞Xk=0 Xα=k A(α) ⊗ Xα, X ∈ Dm f,rad(H), where the series are convergent in the operator norm topology. Then supr∈[0,1) kF (rW )k < ∞ and Define ϕ : [0, 1) → AE (Dm f )∗ + AE (Dm f ) by setting ϕ(r) := F (rW ). Note that, if X ∈ rDm f (H), then Conversely, assume that item (ii) holds. Applying Theorem 3.6 to ϕ, we deduce that F is a free pluriharmonic function and F (rW ) = ϕ(r) for any r ∈ [0, 1). Since ϕ is bounded, we also have kFk ≤ supr∈[0,1) kF (rW )k < ∞. This completes the proof. (cid:3) E (Dm We denote by Harc f,rad(H) with operator- valued coefficients in B(E), which have continuous extensions (in the operator norm topology) to the domain Dm f (H). Here is our noncommutative version of the Dirichlet extension problem for harmonic functions [13]. f,rad) the set of all free pluriharmonic functions on Dm Theorem 4.5. If F : Dm f,rad(H) → B(E) ⊗min B(H), then the following statements are equivalent: (i) F is a free pluriharmonic function on Dm f (H); norm topology) to the domain Dm (ii) F is a free pluriharmonic function on Dm f,rad(H) which has a continuous extension (in the operator f,rad(H) such that F (rW ) converges in the operator norm topology, as r → 1. (iii) there exists ψ ∈ AE (Dm f )∗ + AE (Dm f ) k·k such that F (X) = eB(m) X [ψ] for X ∈ Dm f,rad(H); In this case, ψ = lim r→1 F (rW ), where the convergence is in the operator norm. Moreover, the map Φ : Harc f,rad) → AE(Dm morphism of operator spaces. E (Dm f )∗ + AE (Dm f ) k·k defined by Φ(F ) := ψ is a completely isometric iso- k·k f )∗ + AE (Dm f ) and bounded. Setting ψ := ϕ(1) and using Theorem 4.1, we deduce that F (X) = eB(m) Proof. The implication (i) =⇒ (ii) is clear. Assume that (ii) holds and note that the function ϕ : [0, 1] → AE (Dm given by ϕ(r) := F (rW ) if r ∈ [0, 1) and ϕ(1) := limr→1 F (rW ) is continuous X (ψ) for X ∈ Dm f,rad(H). Therefore, the implication (ii) =⇒ (iii) holds true. Now, we prove the implication (iii) =⇒ (i). (m) Assume that item (iii) holds. Thus there exists ψ ∈ AE(Dm X (ψ) for X ∈ Dm f,rad(H), kFk = kψk, and ψ = SOT- limr→1 F (rW ). In what follows, we show that ψ = limr→1 F (rW ) in the such that F (X) = eB f,rad(H). According to Theorem 4.1, F is a bounded free pluriharmonic function on Dm f )∗ + AE (Dm f ) k·k MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 25 representation of ψ and note that operator norm topology. Indeed, letP∞ B(α) ⊗ X ∗ F (X) = α + ∞Xk=1 Xα=k ∞Xk=0 Xα=k k=1Pα=k B(α) ⊗ W ∗ α +P∞ k=0Pα=k A(α) ⊗ Wα be the Fourier A(α) ⊗ Xα, X ∈ Dm f,rad(H), k·k f )∗ + AE (Dm f )∗ + AE (Dm f ) operator norm topology, which proves our assertion. where the series are convergent in the operator norm topology. Then for any r ∈ [0, 1), F (rW ) ∈ AE (Dm , Theorem 3.5 implies rW (ψ) in the operator norm topology. Consequently, limr→1 F (rW ) = ψ in the rW (ψ). Since ψ ∈ AE (Dm f ) and F (rW ) = eB(m) f (H) and define eF (Y ) := eB that ψ = limr→1eB(m) Let Y ∈ Dm limit exists in the operator norm topology. It remains to prove that eFDm Indeed, if X ∈ Dm limr→1eB(m) X (ψ). Consequently, eF (X) = F (X) for any X ∈ Dm continuity of eF on Dm exists r0 ∈ (0, 1) such that kψ − F (r0W )k < ǫ use again Theorem 3.5 and deduce that f,rad(H), then X is a pure n-tuple and Theorem 3.5 implies that f,rad(H). Now, we prove the f (H). Since ψ = limr→1 F (rW ) in the operator norm topology, for any ǫ > 0 there , we can f,rad(H) = F an eF is contin- (m) rY (ψ) We remark that, due to Theorem 3.5, the latter 3 . Since ψ − F (r0W ) ∈ AE (Dm f (H). rX (ψ) = B(m) f )∗ + AE (Dm f ) uous on Dm k·k and lim r→1eB(m) rY [ψ − F (r0W )] = eF (Y ) − F (r0Y ) Y ∈ Dm ǫ 3 , keF (Y ) − F (r0Y )k ≤ kψ − F (r0W )k < f (H). On the other hand, since F is continuous on Dm for any Z ∈ Dm f (H) such that kY − Zk < δ. Using the estimations above, we note that f,rad(H), there is δ > 0 such that kF (r0Y )− F (r0Z)k < ǫ 3 keF (Y ) − eF (Z)k ≤ keF (Y ) − F (r0Y )k + kF (r0Y ) − F (r0Z)k + kF (r0Z) − eF (Z)k < ǫ f (H) such that kY − Zk < δ. The last part of the theorem follows from Theorem 4.1. (cid:3) for any Y, Z ∈ Dm The proof is complete. Corollary 4.6. Given a function F : Dm equivalent: f,rad(H) → B(E) ⊗min B(H), the following statements are (i) F is a free plurihamonic function which has continuous extension to Dm (ii) There is a continuous function ϕ : [0, 1] → AE(Dm f )∗ + AE (Dm f ) k·k f (H). which satisfies the relation in the operator norm topology t W [ϕ(t)], r ϕ(r) = eB(m) r X [ϕ(r)] for any X ∈ rDm f (H) and r ∈ (0, 1). for any 0 ≤ r < t < 1, and F (X) := eB(m) 1 Moreover, F and ϕ uniquely determine each other and F (rW ) = ϕ(r) for any r ∈ [0, 1). Proof. The proof is similar to that of Corollary 4.4, but uses Theorem 4.5. We leave it to the reader. (cid:3) 5. Cauchy transforms and functional calculus for noncommuting operators In this section, we use noncommutative Cauchy transforms associated with the domain Dm f (H), to provide a free analytic functional calculus for n-tuples of operators X = (X1, . . . , Xn) ∈ B(H)n with the spectral radius of the reconstruction operator strictly less than 1. This extends to free pluriharmonic functions and has several consequences. 26 GELU POPESCU Let f = Pα∈F+ X := (X1, . . . , Xn) ∈ B(H)n such that Pα≥1 n radius of X with respect to the noncommutative domain Dm aαZα, α ∈ C, be a positive regular free holomorphic. For any n-tuple of operators α is SOT-convergent, we define the joint spectral aαXαX ∗ rf (X) := lim k→∞ kΦk f to be f,X (I)k1/2k, where the positive linear map Φf,X : B(H) → B(H) is given by Φf,X (Y ) := Xα≥1 aαXαY X ∗ α, Y ∈ B(K), and the convergence is in the week operator topology. In the particular case when f := Z1 + ··· + Zn, we obtain the usual definition of the joint operator radius for n-tuples of operators. α is SOT convergent, one can easily see that the series Pα≥1 a αΛα ⊗ X ∗ α is SOT- the reconstruction operator associated with the n-tuple X := (X1, . . . , Xn) and the noncommutative domain Dm a αΛαΛ∗ Since Pα≥1 convergent in B(F 2(Hn) ⊗ H). We call the operator R f ,X := Xα≥1 f ,X(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)Φk (cid:13)(cid:13)(cid:13)Rk f ,Λ(I)(cid:13)(cid:13)(cid:13) f,X (I)(cid:13)(cid:13)1/2 a αZα and Φ f,Λ(Y ) := Pα≥1 r(cid:16)R f ,X(cid:17) ≤ r f (Λ)rf (X), 1/2(cid:13)(cid:13)Φk a αΛα ⊗ X ∗ f . Note that α a αΛαY Λ∗ where f := Pα≥1 where r(A) denotes the usual spectral radius of an operator A. Since(cid:13)(cid:13)(cid:13)Φ f ,Λ(I)(cid:13)(cid:13)(cid:13) ≤ 1 (see relation (1.4)), we deduce that r f (Λ) ≤ 1. This implies α. Consequently, we deduce that that , k ∈ N, r(R f ,X) ≤ rf (X). Assume now that X := (X1, . . . , Xn) ∈ B(H)n is an n-tuple of operators with r(R f ,X ) < 1. Note that the latter condition holds if rf (X) < 1. We introduce the Cauchy kernel associated with X to be the operator which is well-defined and , C(m) f,X :=(cid:16)I − R f ,X(cid:17)−m f,X = ∞Xk=0 f ,X!m Rk C(m) , Λβ ⊗ b(m) β X ∗ β , C(m) f,X = Xβ∈F+ f ) ¯⊗minB(H). n C(m) f,X : B(F 2(Hn)) → B(H) where the convergence is in the operator norm topology. We remark that C(m) f,X ∈ R∞(Dm f ) ¯⊗B(H), the W OT -closed operator algebra generated by the spatial tensor product. Moreover, its Fourier representation is (5.1) where the coefficients b(m) polynomial, the Cauchy kernel is in R(Dm α , α ∈ F+ n are given by relation (1.1). In the particular case when f is a Given an n-tuple of operators X := (X1, . . . , Xn) ∈ B(H)n with r(R f ,X ) < 1, we define the Cauchy transform at X to be the mapping MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 27 defined by The operator C(m) DC(m) f,X (A)x, yE :=D(A ⊗ IH)(1 ⊗ x), C(m) f,X (1 ⊗ y)E , f,X (A) is called the Cauchy transform of A at X. x, y ∈ H. In what follows, we provide a free analytic functional calculus for n-tuples of operators X ∈ B(H)n with r(R f ,X) < 1. Theorem 5.1. Let p ∈ ChZ1, . . . , Zni be a positive regular noncommutative polynomial and let X := (X1, . . . , Xn) ∈ B(H)n be an n-tuple of operators with with r(R p,X ) < 1. If is a free pluriharmonic function on the noncommutative domain Dm p (H), then G := Xα≥1 ∞Xs=1 Xα=s G(X) := dαZ ∗ α + Xα∈F+ n cαZα dαX ∗ α + ∞Xs=0 Xα=s cαXα is convergent in the operator norm of B(H) and the map Ψp,X : (Har(Dm p,rad), ρ) → (B(H),k · k) defined by Ψp,X (G) := G(X) is a continuous. In particular, Ψp,XHol(Dm the free analytic functional calculus on Hol(Dm Zi 7→ Xi, i ∈ {1, . . . , n}. p,rad ) is a continuous unital algebra homomorphism. Moreover, p,rad) is uniquely determined by the map Proof. Note that, using relations (5.1), (1.2), (1.3), we obtain p,X (Wα)x, yE =D(Wα ⊗ IH)(1 ⊗ x), C(m) DC(m) √b = 1 =* 1 √b (m) α (m) α eα ⊗ x, Xβ∈F+ eα ⊗ x, Xβ∈F+ n p,X (1 ⊗ y)E (Λβ ⊗ X ∗ b(m) β nrb(m) β e β ⊗ X ∗ β)(1 ⊗ y)+ y+ β for any x, y ∈ H. Hence we deduce that, for any polynomial q ∈ ChZ1, . . . , Zni, = hXαx, yi hq(X)x, yi =D(q(W ) ⊗ IH)(1 ⊗ x), C(m) p,X (1 ⊗ y)E cαZα is a free holomorphic function on Dm kq(X)k ≤ kq(W )kkC(m) p,Xk. p,rad, the series F (rW ) := n Since F := Pα∈F+ ∞Ps=0 Pα=s and (5.2) (5.3) and (5.4) r ∈ [0, 1), converges in the operator norm topology. Now, using relation (5.2), we deduce that F (rX) := cαrαXα converges in the operator norm topology of B(H), cαrαWα, ∞Pk=0 Pα=k kF (rX)k ≤ kF (rW )kkC(m) p,Xk, hF (rX)x, yi =D(F (rW ) ⊗ IH)(1 ⊗ x), C(m) p,X (1 ⊗ y)E for any x, y ∈ H and r ∈ [0, 1). ∞Pk=0 Pα=k x, y ∈ H. dαZα. Combining the results, we convergent in the operator norm topology and t (5.5) deduce that Hence, we obtain hF (X)x, yi =(cid:28)(F(cid:18) 1 p,tX (1 ⊗ y)(cid:29) , W(cid:19) ⊗ IH)(1 ⊗ x), C(m) kF (X)k ≤(cid:13)(cid:13)(cid:13)(cid:13)F(cid:18) 1 W(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)kC(m) Similar results hold true for the free holomorphic function E := Pα∈F+ ∞Xs=1 Xα=s ∞Xs=1 Xα=s kG(X)k ≤(cid:18)(cid:13)(cid:13)(cid:13)(cid:13)E(cid:18) 1 W(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)(cid:19)kC(m) W(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)F(cid:18) 1 is convergent in the operator norm of B(H) and (5.6) To prove the continuity of Ψp,X , let Gk and G be in Har(Dm p,tXk. G(X) := dαX ∗ cαXα α + t t t n p,tXk. 28 GELU POPESCU In what follows, we prove that if r(R p,X) < 1, then there is t > 1 suxh that r(R p,tX ) < 1. Indeed, since the spectrum of an operator is upper continuous, so is the spectral radius. Consequently, for any δ > 0, there is ǫ > 0 such that if kX − tXk < ǫ, then r(R p,tX ) < r(R p,X ) + δ. Hence, using the fact that r(R p,X ) < 1, we deduce that there is t > 1 such that r(R p,tX ) < 1. Using relations (5.3) and (5.4) in the particular case when r = 1 t and when X is replaced by tX, we deduce that F (X) := cαXα is p,rad) such that Gk → G, as m → ∞, in the metric ρ of Har(Dm p,rad). This is equivalent to the fact that, for each r ∈ [0, 1), where the convergence is in the operator norm of B(F 2(Hn)). Employing relation (5.6), when G is replaced by Gk − G, we deduce that Gk(rW ) → G(rW ), as k → ∞, which proves the continuity of Ψp,X . kGk(X) − G(X)k → 0, as k → ∞. Let Fj := c(j) α Zα, j ∈ {1, 2}, be free holomorphic functions on Dm ∞Ps=0 Pα=s c(j) α Zα, we have pj,k(X) → Fj(X), as k → ∞, in the operator norm for any X ∈ Dm the noncommutative domain algebra and F1(rW )F2(rW ) = (F1F2)(rW ) for any r ∈ [0, 1). Setting pj,k := p,rad(H). f,rad. Recall that A(Dm f ) is kPs=0 Pα=s Using relation (5.5), we obtain hp1,k(X)p2,k(X)x, yi =(cid:28)(p1,k(cid:18) 1 t W(cid:19) (p2,k(cid:18) 1 t W(cid:19) ⊗ IH)(1 ⊗ x), C(m) Passing to the limit as k → ∞ and using again relation (5.5), we obtain hF1(X)F2(X)x, yi =(cid:28)(F1(cid:18) 1 =(cid:28)(F1F2)(cid:18) 1 W(cid:19) (F2(cid:18) 1 W(cid:19) ⊗ IH)(1 ⊗ x), C(m) W(cid:19) ⊗ IH)(1 ⊗ x), C(m) p,tX (1 ⊗ y)(cid:29) t t t = h(F1F2)(X)x, yi x, y ∈ H. p,tX (1 ⊗ y)(cid:29) , p,tX (1 ⊗ y)(cid:29) for any x, y ∈ H. Consequently, Ψp,XHol(Dm p,rad ) is a unital algebra homomorphism. To prove the uniqueness of the free analytic functional calculus, assume that Φ : Hol(Dm is a continuous unital algebra homomorphism such that Φ(Zi) = Ti, i = 1, . . . , n. It is clear that p,rad) → B(H) (5.7) Ψp,X(q) = Φ(q) MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 29 for any polynomial q ∈ ChZ1, . . . , Zni. Let F =P∞ s=0Pα=s cαZα be an element in Hol(Dm let Qk :=Pk ∞Xs=0 Xα=s s=0Pα=s cαZα, k ∈ N. Since s=0 rs(cid:13)(cid:13)(cid:13)Pα=s cαWα(cid:13)(cid:13)(cid:13) converges, we deduce that Qk(rW ) → F (rW ) in the operator and the series P∞ norm, as k → ∞, which shows that Qk → F in the metric ρ of Hol(Dm and the continuity of Φ and Ψp,T , we deduce that Φ = Ψp,T . This completes the proof. Corollary 5.2. Let X := (X1, . . . , Xn) ∈ B(H)n be an n-tuple of operators with with r(R p,X ) < 1 and let F ∈ Hol(Dm p,rad). Hence, using relation (5.7) (cid:3) p,rad). If t > 1 is such that r(R p,tX ) < 1, then p,rad) and rscαWα F (rW ) = F (X) = C(m) p,tX(cid:20)F(cid:18) 1 t W(cid:19)(cid:21) , where F (X) is defined by the free analytic functional calculus. If, in addition, F is bounded, then Proof. The first part of the corollary is due to Theorem 5.1 (see relation (5.5)). Now, we assume that F F (X) = C(m) p,X (eF ), where eF = SOT- lim r→1 F (rW ). is bounded and has the representation F := cαZα. Then, we have F (rW ) := lim k→∞ cαWα, 0 < r < 1, in the operator norm of B(F 2(Hn)), and lim k→∞ cαXα = F (rX) ∞Pk=0 Pα=k kXs=0 rs Xα=s kXs=0 rs Xα=s in the operator norm of B(H). Now, due to the continuity of the noncommutative Cauchy transform in the operator norm, we deduce that F (rX) = Cp,X (F (rW )), r ∈ [0, 1). F (rW ) exists in the strong operator topology. Since Since F is bounded, we know that eF := lim kF (rW )k ≤ keFk, r ∈ [0, 1), we deduce that r→1 According to the proof of Theorem 5.1, we have t p,tXk. SOT- lim r→1 [F (rW ) ⊗ IH] = eF ⊗ IH. kF (X)k ≤(cid:13)(cid:13)(cid:13)(cid:13)F(cid:18) 1 W(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)kC(m) kF (X) − F (δX)k ≤(cid:13)(cid:13)(cid:13)(cid:13)F(cid:18) 1 W(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)kC(m) W(cid:19) − F(cid:18) δ t W(cid:1)(cid:13)(cid:13) → 0, as δ → 1, we obtain lim hF (rX)x, yi =D(F (rW ) ⊗ IH)(1 ⊗ x), C(m) t p,X (1 ⊗ y)E Using this relation, we deduce that Since (cid:13)(cid:13)F(cid:0) 1 since t W(cid:1) − F(cid:0) δ p,tXk, δ ∈ (0, 1). t δ→1 kF (X) − F (δX)k = 0. On the other hand, for any x, y ∈ H and r ∈ [0, 1), we can pass to the limit as r → 1 and obtain F (X) = C(m) completes the proof. p,X [eF ]. This (cid:3) 30 GELU POPESCU Corollary 5.3. Let X := (X1, . . . , Xn) ∈ B(H)n be an n-tuple of operators with with r(R p,X ) < 1. k=1 and G be bounded free holomorphic functions on Dm k=1 and G are free pluriharmonic functions in Har(Dm as k → ∞, then Gk(X) → G(X) in the operator norm of B(H). be the corresponding boundary operators in the noncommutative Hardy algebra F ∞(Dm (i) If {Gk}∞ k=1 and eG (ii) Let {Gk}∞ eGk → eG in the w∗-topology (or strong operator topology) and kGkk∞ ≤ M for any k ∈ N, then Gk(X) → G(X) in the weak operator topology. p,rad and let {eGk}∞ p,rad) such that kGk − Gk∞ → 0, p ). If Using Theorem 5.1 one can deduce the following. Corollary 5.4. For any n-tuple of operators (X1, . . . , Xn) ∈ Dm calculus coincides with the F ∞(Dm p )-functional calculus (see [27]). p,rad(H), the free analytic functional Using Corollary 5.2, we can obtain the following. Corollary 5.5. Let X := (X1, . . . , Xn) ∈ B(H)n be an n-tuple of operators with with r(R p,X) < 1. Then, the map Ψp,X : F ∞(Dm p ) → B(H) defined by n (Dm for any eG ∈ F ∞ α ∈ F+ n . Moreover, p ), is a unital WOT continuous homomorphism such that Ψf,X (Wα) = Xα for any Ψp,X(eG) := C(m) p,X [eG], kΨp,X(eG)k ≤ ∞Xk=0(cid:13)(cid:13)(cid:13)Rk f ,X(cid:13)(cid:13)(cid:13)!m keGk. Definition 5.6. Let H1 and H2 be two self-adjoint free pluriharmonic functions on Dm f,rad with scalar coefficients. We say that H2 is the pluriharmonic conjugate of H1, if H1 + iH2 is a free holomorphic function on Dm f,rad. Proposition 5.7. The free pluriharmonic conjugate of a self-adjoint free pluriharmonic function on Dm f,rad is unique up to an additive real constant. Proof. Assume that G = ℜF with F ∈ Hol(Dm f,rad), and let H be a self-adjoint free pluriharmonic function such that G + iH = Λ ∈ Hol(Dm and the equality H = H ∗ implies ℜ(Λ − F ) = 0. Consequently, due to the remarks following Definition 3.3, we have Λ − F = λ with λ is an imaginary complex number. Now, it is clear that H = F −F ∗ Theorem 5.8. Let X := (X1, . . . , Xn) ∈ B(H)n be an n-tuple of operators with with r(R p,X ) < 1 and let F ∈ Hol(Dm p,rad) be such that F (0) is real. If G = ℜF and t > 1 is such that r(R p,tX ) < 1, then 2i − iλ. The proof is complete. f,rad). Then H = 2Λ−F −F ∗ (cid:3) 2i p,tX − Ii (1 ⊗ y)(cid:29) , p,X − Ii (1 ⊗ y)E , x, y ∈ H. x, y ∈ H, If, in addition, F is bounded, then where eG := SOT- limr→1 G(rW ). Proof. Using the proof of Theorem 5.1, we deduce that t hF (X)x, yi =(cid:28)(G(cid:18) 1 W(cid:19) ⊗ IH)(1 ⊗ x),h2C(m) hF (X)x, yi =D(eG ⊗ IH)(1 ⊗ x),h2C(m) W(cid:19) ⊗ IH)(1 ⊗ x),h2C(m) p,tX − Ii (1 ⊗ y)(cid:29) = 2(cid:28)(F(cid:18) 1 W(cid:19) ⊗ IH)(1 ⊗ x), C(m) t = 2 hF (X)x, yi − F (0)hx, yi (cid:28)(F(cid:18) 1 t p,tX (1 ⊗ y)(cid:29) −(cid:28)F(cid:18) 1 t W(cid:19) ⊗ IH)(1 ⊗ x), 1 ⊗ y(cid:29) MULTI-TOEPLITZ OPERATORS AND FREE PLURIHARMONIC FUNCTIONS 31 and (cid:28)(F(cid:18) 1 t W(cid:19)∗ ⊗ IH)(1 ⊗ x),h2C(m) =D(F (0) ⊗ IH)(1 ⊗ x),h2C(m) p,tX − Ii (1 ⊗ y)(cid:29) p,tX − Ii (1 ⊗ y)E = F (0)hx, yi . Taking into account that F (0) ∈ R and adding the relations above, we obtain 2 hF (X)x, yi =(cid:28)(cid:20)(cid:18)F(cid:18) 1 t W(cid:19)∗ + F(cid:18) 1 t W(cid:19)(cid:19) ⊗ IH(cid:21) (1 ⊗ x),h2C(m) p,tX − Ii (1 ⊗ y)(cid:29) , which proves the first part of the theorem. In a similar manner, but using Corollary 5.2, one can prove the second part of the theorem. (cid:3) We remark that the free pluriharmonic conjugate H of G an be expressed in terms of G, due to the fact that H = F −F ∗ 2i − iλ, where λ is an imaginary complex number. References [1] J. Agler, Hypercontractions and subnormality, J. Operator Theory 13 (1985), 203 -- 217. [2] A. Bottcher and B. Silbermann, Analysis of Toeplitz operators, Springer-Verlag, Berlin, 1990. [3] A. Brown and P.R. Halmos, Algebraic properties of Toeplitz operators, J. Reine Angew. Math. 213 1963/1964 89 -- 102. [4] K. R. Davidson and D. Pitts, The algebraic structure of non-commutative analytic Toeplitz algebras, Math. Ann. 311 (1998), 275 -- 303. [5] K.R. Davidson, E. Katsoulis, and D. Pitts, The structure of free semigroup algebras, J. Reine Angew. Math. 533 (2001), 99 -- 125. [6] K. R. Davidson, J. Li, and D.R. Pitts, Absolutely continuous representations and a Kaplansky density theorem for free semigroup algebras, J. Funct. Anal. 224 (2005), no. 1, 160 -- 191. [7] R. G. Douglas, Banach algebra techniques in operator theory, Second edition. Graduate Texts in Mathematics, 179, Springer-Verlag, New York, 1998. xvi+194 pp. [8] E.G. Effros and Z.J. Ruan, Operator spaces, London Mathematical Society Monographs. New Series, 23. The Clarendon Press, Oxford University Press, New York, 2000. [9] J. Eschmeier and S. Langendorfer, Toeplitz operators with pluriharmonic symbol, preprint 2017. [10] P. Hartman and A. Wintner, The spectra of Toeplitz's matrices, Amer. J. Math. 76 (1954), 867-882. [11] H. Hedenmalm, B. Korenblum, and K. Zhu, Theory of Bergman spaces, Graduate Texts in Mathematics, 199, Springer-Verlag, New York, 2000. x+286 pp. [12] G. Herglotz, Uber Potenzreien mit positiven, reelen Teil im Einheitkreis, Berichte uber die Verhaundlungen der koniglich sachsischen Gesellschaft der Wissenschaften zu Leipzig, Math.-Phys. Klasse 63 (1911), 501 -- 511. [13] K. Hoffman, Banach Spaces of Analytic Functions, Englewood Cliffs: Prentice-Hall, 1962. [14] M. Kennedy, Wandering vectors and the reflexivity of free semigroup algebras, J. Reine Angew. Math. 653 (2011), 47 -- 73. [15] M. Kennedy, The structure of an isometric tuple, Proc. Lond. Math. Soc. (3) 106 (2013), no. 5, 1157 -- 1177. [16] I. Louhichi and A. Olofsson, Characterizations of Bergman space Toeplitz operators with harmonic symbols, J.Reine Angew. Math. 617 (2008), 1 -- 26. [17] A. Olofsson, A characteristic operator function for the class of n-hypercontractions, J. Funct. Anal. 236 (2006), no. 2, 517 -- 545. [18] A. Olofsson, An operator-valued Berezin transform and the class of n-hypercontractions, Integral Equations Operator Theory 58 (2007), no. 4, 503 -- 549. [19] V.I. Paulsen, Completely Bounded Maps and Dilations, Pitman Research Notes in Mathematics, Vol.146, New York, 1986. [20] G. Pisier, Similarity Problems and Completely Bounded Maps, Springer Lect. Notes Math., Vol.1618, Springer-Verlag, New York, 1995. [21] G. Popescu, Multi-analytic operators and some factorization theorems, Indiana Univ. Math. J. 38 (1989), 693 -- 710. [22] G. Popescu, Multi-analytic operators on Fock spaces, Math. Ann. 303 (1995), 31 -- 46. [23] G. Popescu, Poisson transforms on some C ∗-algebras generated by isometries, J. Funct. Anal. 161 (1999), 27 -- 61. [24] G. Popescu, Free holomorphic functions on the unit ball of B(H)n, J. Funct. Anal. 241 (2006), 268 -- 333. [25] G. Popescu, Entropy and multivariable interpolation, Mem. Amer. Math. Soc. 184 (2006), no. 868, vi+83 pp. [26] G. Popescu, Noncommutative transforms and free pluriharmonic functions, Adv. Math. 220 (2009), 831-893. [27] G. Popescu, Noncommutative Berezin transforms and multivariable operator model theory, J. Funct. Anal. 254 (2008), no. 4, 1003 -- 1057. 32 GELU POPESCU [28] G. Popescu, Operator theory on noncommutative domains, Mem. Amer. Math. Soc. 205 (2010), no. 964, vi+124 pp. [29] G. Popescu, Berezin transforms on noncommutative varieties in polydomains, J. Funct. Anal. 265 (2013), no. 10, 2500 -- 2552. [30] G. Popescu, Berezin transforms on noncommutative polydomains, Trans. Amer. Math. Soc. 368 (2016), no. 6, 4357 -- 4416. [31] G. Popescu, Bohr inequalities on noncommutative polydomains, preprint. [32] G. Popescu, Brown-Halmos characterization of multi-Toeplitz operators associated with noncommutative hyperballs, preprint. [33] M. Rosenblum and J. Rovnyak, Hardy classes and operator theory, Oxford University Press-New York, 1985. [34] I. Schur, Uber Potenzreihen die im innern des Einheitshreises beschrankt sind, J. Reine Angew. Math. 148 (1918), 122 -- 145. [35] B. Sz.-Nagy, C. Foias¸, H. Bercovici, and L. K´erchy, Harmonic Analysis of Operators on Hilbert Space, Second edition. Revised and enlarged edition. Universitext. Springer, New York, 2010. xiv+474 pp. [36] O. Toeplitz, Zur Theorie der quadratischen und bilinearen Formen von unendlichvielen Veranderlichen, Math. Ann. 70 (1911), no. 3, 351 -- 376. [37] H. Upmeier, Toeplitz operators and index theory in several complex variables, Operator Theory: Advances and Appli- cations 81, Birkhauser Verlag, Basel, 1996. Department of Mathematics, The University of Texas at San Antonio, San Antonio, TX 78249, USA E-mail address: [email protected]
1905.12870	2	1905	2019-06-07T04:24:43	On the Operator Jensen Inequality for Convex Functions	[ "math.FA" ]	This paper is mainly devoted to studying operator Jensen inequality. More precisely, a new generalization of Jensen inequality and its reverse version for convex (not necessary operator convex) functions have been proved. Several special cases are discussed as well.	math.FA	math	ON THE OPERATOR JENSEN INEQUALITY FOR CONVEX FUNCTIONS MOHSEN SHAH HOSSEINI, HAMID REZA MORADI, AND BAHARAK MOOSAVI Abstract. This paper is mainly devoted to studying operator Jensen inequality. More pre- cisely, a new generalization of Jensen inequality and its reverse version for convex (not necessary operator convex) functions have been proved. Several special cases are discussed as well. 1. Introduction Let B (H) be the C ∗ -- algebra of all bounded linear operators on a Hilbert space H. As customary, we reserve m, M for scalars and 1H for the identity operator on H. A self-adjoint operator A is said to be positive (written A ≥ 0) if hAx, xi ≥ 0 holds for all x ∈ H also an operator A is said to be strictly positive (denoted by A > 0) if A is positive and invertible. If A and B are self-adjoint, we write B ≥ A in case B −A ≥ 0. The Gelfand map f (t) 7→ f (A) is an isometrical ∗ -- isomorphism between the C ∗ -- algebra C (sp (A)) of continuous functions on the spectrum sp (A) of a selfadjoint operator A and the C ∗ -- algebra generated by A and the identity operator 1H. If f, g ∈ C (sp (A)), then f (t) ≥ g (t) (t ∈ sp (A)) implies that f (A) ≥ g (A). For A, B ∈ B (H), A ⊕ B is the operator defined on B (H ⊕ H) by A 0 0 B!. A linear map Φ : B (H) → B (K) is positive if Φ (A) ≥ 0 whenever A ≥ 0. It's said to be unital if Φ (1H) = 1K. A continuous function f defined on the interval J is called an operator convex function if f ((1 − v) A + vB) ≤ (1 − v) f (A) + vf (B) for every 0 < v < 1 and for every pair of bounded self-adjoint operators A and B whose spectra are both in J. The well known operator Jensen inequality states (sometimes called the Choi -- Davis -- Jensen inequality): (1.1) f (Φ (A)) ≤ Φ (f (A)) . It holds for every operator convex f : J → R, self-adjoint operator A with spectra in J, and unital positive linear map Φ [3, 5]. Hansen et al. [8] gave a general formulation of (1.1). The discrete version of their result reads as follows: If f : J → R is an operator convex function, A1, . . . , An ∈ B (H) are self-adjoint 2010 Mathematics Subject Classification. Primary 47A63, Secondary 26A51, 26D15, 26B25, 39B62. Key words and phrases. Jensen's inequality, convex functions, self-adjoint operators, positive operators. 1 2 M. Shah hosseini, H. R. Moradi & B. Moosavi operators with the spectra in J, and Φ1, . . . , Φn : B (H) → B (K) are positive linear mappings i=1 Φi (1H) = 1K, then such that Pn (1.2) f n Xi=1 Φi (Ai)! ≤ n Xi=1 Φi (f (Ai)). Though in the case of convex function the inequality (1.2) does not hold in general (see [3, Remark 2.6]), we have the following estimate [6, Lemma 2.1]: (1.3) f * n Xi=1 Φi (Ai)x, x+! ≤* n Xi=1 Φi (f (Ai))x, x+ for any unit vector x ∈ K. For recent results treating the Jensen operator inequality, we refer the reader to [9, 10, 11]. As a converse of (1.2), in [8] (see also [12]), it has been shown that if f : [m, M] → R is a convex function and A1, . . . , An are self-adjoint operators with the spectra in [m, M], then (1.4) where n Xi=1 Φi (f (Ai)) ≤ β1K + f n Xi=1 Φi (Ai)! β = max(cid:26) f (M) − f (m) M − m t + M f (m) − mf (M) M − m − f (t) : m ≤ t ≤ M(cid:27) . A monograph on the reverse of Jensen inequality and its consequences is given by Furuta et al. in [7]. In this paper, we prove an inequality of type (1.2) without operator convexity assumption. Furthermore, as we can see in (1.4), the constant β is dependent on m and M. In this paper, we establish another reverse of operator Jensen inequality by dropping this restriction. 2. Operator Jensen-type inequalities without operator convexity Let f : J → R be a convex function, A ∈ B (H) self-adjoint operator with the spectra in J, and let x ∈ H be a unit vector. Then from [13], f (hAx, xi) ≤ hf (A) x, xi . Replace A with Φ (A), where Φ : B (H) → B (K) is a unital positive linear map, we get (2.1) f (hΦ (A) x, xi) ≤ hf (Φ (A)) x, xi for any unit vector x ∈ K. Assume that A1, . . . , An are self-adjoint operators on H with i=1 Φi (1H) = 1K. Now apply inequality (2.1) to the self-adjoint operator A on the Hilbert space H ⊕ · · · ⊕ H spectra in J and Φ1, . . . , Φn : B (H) → B (K) are positive linear maps with Pn On the operator Jensen inequality for convex functions 3 defined by A = A1 ⊕ · · · ⊕ An and the positive linear map Φ defined on B (H ⊕ · · · ⊕ H) by Φ (A) = Φ1 (A1) ⊕ · · · ⊕ Φn (An). Thus, (2.2) f * n Xi=1 Φi (Ai)x, x+! ≤f n Xi=1 Φi (Ai)! x, x+ . Let us also recall that if f is a convex function on an interval J, then for each point (s, f (s)), there exists a real number Cs such that (2.3) Cs (t − s) + f (s) ≤ f (t) , (t ∈ J) . Inequality (2.2), together with (2.3) yield the following theorem. Theorem 2.1. Let f : J → R be a monotone convex function, A1, . . . , An ∈ B (H) self-adjoint operators with the spectra in J, and let Φ1, . . . , Φn : B (H) → B (K) be positive linear mappings i=1 Φi (1H) = 1K. Then such that Pn (2.4) where n Xi=1 Φi (f (Ai)) ≤ f n Xi=1 Φi (Ai)! + δ1K δ = sup( n Xi=1 Φi (CAiAi)x, x+ −* n Xi=1 Φi (Ai)x, x+* n Xi=1 Φi (CAi)x, x+ : x ∈ K; kxk = 1) . Proof. Fix t ∈ J. Since J contains the spectra of the Ai for i = 1, . . . , n, we may replace s in the inequality (2.3) by Ai, via a functional calculus to get Applying the positive linear mappings Φi and summing on i from 1 to n, this implies f (Ai) ≤ f (t) 1H + CAiAi − tCAi. (2.5) Φi (f (Ai)) ≤ f (t) 1K + Φi (CAiAi) − t n n Xi=1 Φi (CAi). n Xi=1 Xi=1 The inequality (2.5) easily implies, for any x ∈ K with kxk = 1, (2.6) (2.7) SincePn we may replace t by hPn * n Xi=1 i=1 Φi (Ai)x, xi in (2.6). This yields Φi (f (Ai))x, x+ ≤ f (t) +* n Xi=1 * n Xi=1 i=1 Φi (1H) = 1K we have hPn Φi (f (Ai))x, x+ ≤ f * n Φi (Ai)x, x+! Xi=1 +* n Φi (CAiAi)x, x+ −* n Xi=1 Xi=1 Φi (CAiAi)x, x+ − t* n Xi=1 Φi (CAi)x, x+ . i=1 Φi (Ai)x, xi ∈ J where x ∈ K with kxk = 1. Therefore, Φi (Ai)x, x+* n Xi=1 Φi (CAi)x, x+ . 4 M. Shah hosseini, H. R. Moradi & B. Moosavi On ther other hand, 0 ≤* n Xi=1 ≤ sup x∈K kxk=1 Φi (CAiAi)x, x+ −* n Xi=1 (* n Xi=1 Φi (CAiAi)x, x+ −* n Xi=1 Φi (Ai)x, x+* n Xi=1 Φi (CAi)x, x+ Φi (Ai)x, x+* n Xi=1 Φi (CAi)x, x+) thanks to (1.3). Therefore, * n Xi=1 Φi (f (Ai))x, x+ ≤ f * n Xi=1 ≤f n Xi=1 Φi (Ai)x, x+! + δ Φi (Ai)! x, x+ + δ (by (2.2)). This completes the proof. (cid:3) Remark 2.1. Inequality (2.7) provides the reverse of the inequality (1.3). In the next theorem, we aim to present operator Jensen-type inequality without operator convexity assumption. Theorem 2.2. Let all the assumptions of Theorem 2.1 hold, then (2.8) where f n Xi=1 Φi (Ai)! ≤ n Xi=1 Φi (f (Ai)) + ζ1K Φi (Ai)x, x+DCPn i=1 Φi(Ai)x, xE : x ∈ K; kxk = 1) . n Xi=1 i=1 Φi(Ai) Φi (Ai)x, x+ − n Xi=1 ζ = sup(CPn Proof. Fix t ∈ J. Since J contains the spectra of the Ai for i = 1, . . . , n andPn so the spectra of Pn (2.3) by Pn f n Xi=1 Φi (Ai)! ≤ f (t) 1K + CPn i=1 Φi (Ai), via a functional calculus to get Xi=1 n i=1 Φi(Ai) Φi (Ai) − tCPn i=1 Φi(Ai). i=1 Φi (1H) = 1K, i=1 Φi (Ai) is also contained in J. Then we may replace s in the inequality This inequality implies, for any x ∈ K with kxk = 1, (2.9) f n Xi=1 Φi (Ai)! x, x+ ≤ f (t) +CPn i=1 Φi(Ai) n Xi=1 Φi (Ai)x, x+ − t(cid:10)CPn i=1 Φi(Ai)x, x(cid:11) . On the operator Jensen inequality for convex functions 5 i=1 Φi (Ai)x, xi in (2.9). Thus, Φi (Ai)x, x+! (2.10) Substituting t with hPn f n Xi=1 Φi (Ai)! x, x+ ≤ f * n Xi=1 +CPn On the other hand, i=1 Φi(Ai)x, xE . i=1 Φi(Ai) Φi (Ai)x, x+DCPn n Xi=1 Φi (Ai)x, x+ − n Xi=1 Φi (Ai)x, x+(cid:10)CPn Φi (Ai)x, x+ −* n Xi=1 i=1 Φi(Ai)x, x(cid:11) Φi (Ai)x, x+(cid:10)CPn Φi (Ai)x, x+ −* n Xi=1 n Xi=1 i=1 Φi(Ai)x, x(cid:11)) 0 ≤CPn ≤ sup x∈K kxk=1 n i=1 Φi(Ai) Xi=1 (CPn i=1 Φi(Ai) thanks to (2.2). Consequently, f n Xi=1 Φi (Ai)! x, x+ ≤ f n Φi (Ai)x, x+! + ζ Xi=1 Φi (f (Ai))x, x+ + ζ ≤* n Xi=1 and the proof is complete. (by (1.3)) (cid:3) Remark 2.2. Notice that inequality (2.10) can be considered as a converse of inequality (2.2). 3. Some Applications In this section, we collect some consequences of Theorems 2.1 and 2.2. (I) Suppose, in addition to the assumptions in Theorem 2.1, f is differentiable on J whose derivative f ′ is continuous on J, then (2.4) and (2.8) hold with and δ = sup(* n Φi(cid:0)f ′ (Ai) Ai(cid:1)x, x+ −* n Xi=1 Xi=1 Φi (Ai)! n ζ = sup(f ′ n Xi=1 Xi=1 Φi (Ai)x, x+f ′ n −* n Xi=1 Xi=1 Φi (Ai)x, x+* n Xi=1 Φi (Ai)x, x+ Φi (Ai)! x, x+ : x ∈ K; kxk = 1) . Φi(cid:0)f ′ (Ai)(cid:1)x, x+ : x ∈ K; kxk = 1) (II) By setting f (t) = tp (p ≥ 1) in Theorems 2.1 and 2.2 we find that: (3.1) n Xi=1 Φi (Ap i ) ≤ n Xi=1 Φi (Ai)!p + pδ1K 6 where δ = sup(* n Xi=1 and (3.2) where ζ = sup  M. Shah hosseini, H. R. Moradi & B. Moosavi Φi (Ai)x, x+* n Xi=1 i Φi(cid:0)Ap−1 (cid:1)x, x+ : x ∈ K; kxk = 1) Φi (Ap i )x, x+ −* n Xi=1 Φi (Ai)!p n Xi=1 ≤ n Xi=1 Φi (Ap i ) + pζ1K * n Xi=1 Φi (Ai)!p x, x+ −* n Xi=1 Φi (Ai)x, x+* n Xi=1 Φi (Ai)!p−1 x, x+ : x ∈ K; kxk = 1  whenever A1, . . . , An ∈ B (H) are positive operators and Φ1, . . . , Φn : B (H) → B (K) positive i=1 Φi (1H) = 1K. If the operators A1, . . . , An are strictly positive, then (3.1) and (3.2) are also true for p < 0. linear mappings such that Pn (III) Assume that w1, . . . , wn are positive scalars such that Pn If we apply Theorems 2.1 and 2.2 for positive linear mappings Φi : B (H) → B (H) determined by Φi : T 7→ wiT (i = 1, . . . , n), we get i=1 wi = 1. n Xi=1 wif (Ai) ≤ f n Xi=1 wiAi! + δ1K wiAix, x+* n Xi=1 wiCAix, x+ : x ∈ K; kxk = 1) where δ = sup(* n Xi=1 and where wiCAiAix, x+ −* n Xi=1 wiAi! ≤ f n Xi=1 ζ = sup(CPn i=1 wiAi n Xi=1 wiAix, x+ − n Xi=1 Choi's inequality [4, Proposition 4.3] says that wif (Ai) + ζ1K n Xi=1 wiAix, x+(cid:10)CPn i=1 wiAix, x(cid:11) : x ∈ K; kxk = 1) . (3.3) Φ (B) Φ(A)−1Φ (B) ≤ Φ(cid:0)BA−1B(cid:1) whenever B is self-adjoint and A is positive invertible. We shall show the following comple- mentary inequality of (3.3): On the operator Jensen inequality for convex functions 7 Proposition 3.1. Let A, B ∈ B (H) such that B is self-adjoint and A is positive invertible, and let Φ : B (H) → B (K) be a unital positive linear mapping. Then (3.4) where δ = sup(cid:26)DΦ(A)− 1 Φ(cid:0)BA−1B(cid:1) ≤ Φ (B) Φ(A)−1Φ (B) + 2δΦ (A) 2 x, xE2 2 x, xE −DΦ(A)− 1 2 Φ (B) Φ(A)− 1 2 Φ(cid:0)BA−1B(cid:1) Φ(A)− 1 Proof. It follows from Theorem 2.1 that : x ∈ K; kxk = 1(cid:27) . (3.5) where Ψ(cid:0)T 2(cid:1) ≤ Ψ(T )2 + 2δ1K δ = sup(cid:8)(cid:10)Ψ(cid:0)T 2(cid:1) x, x(cid:11) − hΨ (T ) x, xi2 : x ∈ K; kxk = 1(cid:9) . To a fixed positive A ∈ B (H) we set Ψ (X) = Φ(A)− 1 1 2 XA 2 Φ(cid:16)A 2 1 2(cid:17) Φ(A)− 1 and notice that Ψ : B (H) → B (K) is a unital linear map. Now, if T = A− 1 from (3.5) that 2 BA− 1 2 , we infer Φ(A)− 1 2 ≤ Φ(A)− 1 2 Φ (B) Φ(A)−1Φ (B) Φ(A)− 1 2 + 2δ1K where δ = sup(cid:26)DΦ(A)− 1 2 Φ(cid:0)BA−1B(cid:1) Φ(A)− 1 2 Φ(cid:0)BA−1B(cid:1) Φ(A)− 1 2 x, xE −DΦ(A)− 1 2 Φ (B) Φ(A)− 1 2 x, xE2 : x ∈ K; kxk = 1(cid:27) . 1 By multiplying from the left and from the right with Φ(A) 2 we obtain (3.4). (cid:3) The parallel sum of two positive operators A, B is defined as the operator A simple calculation shows that (see, e.g., [2, (4.6) and (4.7)]) A : B =(cid:0)A−1 + B−1(cid:1)−1 . (3.6) A : B = A − A(A + B)−1A = B − B(A + B)−1B. If Φ is any positive linear map, then (see [2, Theorem 4.1.5]) (3.7) Φ (A : B) ≤ Φ (A) : Φ (B) . The following result gives a reverse of inequality (3.7). 8 M. Shah hosseini, H. R. Moradi & B. Moosavi Proposition 3.2. Let A, B ∈ B (H) positive invertible operators and let Φ : B (H) → B (K) be unital positive linear mapping. Then Φ (A) : Φ (B) ≤ Φ (A : B) + 2δΦ (A + B) where δ = supnDΦ(A + B)− 1 2 Φ(cid:0)A(A + B)−1A(cid:1) Φ(A + B)− 1 2 x, xE 2 Φ (A) Φ(A + B)− 1 2 x, xE2 : x ∈ K; kxk = 1(cid:27) . −DΦ(A + B)− 1 Proof. Proposition 3.1 easily implies (3.8) where Then we have Φ(cid:0)A(A + B)−1A(cid:1) ≤ Φ (A) Φ(A + B)−1Φ (A) + 2δΦ (A + B) 2 x, xE δ = supnDΦ(A + B)− 1 2 Φ(cid:0)A(A + B)−1A(cid:1) Φ(A + B)− 1 2 Φ (A) Φ(A + B)− 1 2 x, xE2 : x ∈ K; kxk = 1(cid:27) . −DΦ(A + B)− 1 Φ (A) : Φ (B) = Φ (A) − Φ (A) (Φ (A) + Φ (B))−1Φ (A) (by (3.6)) = Φ (A) − Φ (A) Φ(A + B)−1Φ (A) (by the linearity of Φ) ≤ Φ (A) − Φ(cid:0)A(A + B)−1A(cid:1) + 2δΦ (A + B) = Φ(cid:0)A − A(A + B)−1A(cid:1) + 2δΦ (A + B) = Φ (A : B) + 2δΦ (A + B) . (by (3.8)) (by the linearity of Φ) Hence the conclusions follow. (cid:3) Remark 3.1. A function f : [0, ∞) → R is called superquadratic (see [1, Definition 1]) if for each s ≥ 0, there exists a real constant Cs such that (3.9) f (t − s) + Cs (t − s) + f (s) ≤ f (t) for all t ≥ 0. By applying the same arguments as in Theorems 2.1 and 2.2 for definition (3.9), one can obtain stronger estimates than (2.4) and (2.8). We leave the elaboration of this idea to the interested reader. On the operator Jensen inequality for convex functions 9 References [1] S. Abramovich, G. Jameson and G. Sinnamon, Refining Jensen's inequality, Bull. Math. Soc. Sci. Math. Roumanie., 47 (2004), 3 -- 14. [2] R. Bhatia, Positive definite matrices, Princeton Series in Applied Mathematics, Princeton, 2007. [3] M. D. Choi, A Schwarz inequality for positive linear maps on C ∗ -- algebras, Illinois J. Math., 18 (1974), 565 -- 574. [4] M. D. Choi, Some assorted inequalities for positive linear maps on C ∗ -- algebras, J. Operator Theory., 4 (1980), 271 -- 285. [5] C. Davis, A Schwarz inequality for convex operator functions, Proc. Amer. Math. Soc., 8 (1957), 42 -- 44. [6] S. Furuichi, H. R. Moradi and A. Zardadi, Some new Karamata type inequalities and their applications to some entropies, Rep. Math. Phys., (2019) (accepted). arXiv:1811.07277. [7] T. Furuta, J. Mi´ci´c, J. Pecari´c and Y. Seo, Mond -- Pecari´c method in operator inequalities, Element, Zagreb, 2005. [8] F. Hansen, J. Pecari´c and I. Peri´c, Jensen's operator inequality and it's converses, Math. Scand., 100 (2007), 61 -- 73. [9] L. Horv´ath, K. A. Khan and J. Pecari´c, Cyclic refinements of the different versions of operator Jensen's inequality, Electron. J. Linear Algebra., 31(1) (2016), 125 -- 133. [10] J. Mi´ci´c, H. R. Moradi and S. Furuichi, Choi -- Davis -- Jensen's inequality without convexity, J. Math. Inequal., 12(4) (2018), 1075 -- 1085. [11] J. Mi´ci´c and J. Pecari´c, Some mappings related to Levinson's inequality for Hilbert space operators, Filomat., 31 (2017), 1995 -- 2009. [12] J. Mi´ci´c, J. Pecari´c and Y. Seo, Complementary inequalities to inequalities of Jensen and Ando based on the Mond -- Pecari´c method, Linear Algebra Appl., 318 (2000), 87 -- 108. [13] B. Mond and J. Pecari´c, On Jensen's inequality for operator convex functions, Houston J. Math., 21 (1995), 739 -- 753. (M. Shah Hosseini) Department of Mathematics, Shahr-e-Qods Branch, Islamic Azad University, Tehran, Iran. E-mail address: mohsen [email protected] (H. R. Moradi) Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran. E-mail address: [email protected] (B. Moosavi) Department of Mathematics, Safadasht Branch, Islamic Azad University, Tehran, Iran. E-mail address: baharak [email protected]
1504.02595	3	1504	2015-12-29T20:38:17	Error Estimates for Approximating Best Proximity Points for Cyclic Contractive Maps	[ "math.FA" ]	We find a priori and a posteriori error estimates of the best proximity point for the Picard iteration associated to a cyclic contraction map, which is defined on a uniformly convex Banach space with modulus of convexity of power type. We find the rate of convergence for the Picard sequence.	math.FA	math	Error Estimates for Approximating Best Proximity Points for Cyclic Contractive Maps Boyan Zlatanov Abstract: We find a priori and a posteriori error estimates of the best proximity point for the Picard iteration associated to a cyclic contraction map, which is defined on a uniformly convex Banach space with modulus of convexity of power type. Keywords: best proximity points, uniformly convex Banach space, modulus of convexity, a priori error estimate, a posteriori error estimate AMS Subject Classification: 41A25, 47H10, 54H25, 46B20 1 Introduction A fundamental result in fixed point theory is the Banach Contraction Principle. Fixed point theory is an important tool for solving equations T x = x for mappings T defined on subsets of metric spaces or normed spaces. One of the advantage of Banach fixed point Theorem is the error estimates of the successive iterations and the rate of convergence. There are equations T x = x for which the exact solution is not easy to find or even is not possible to find. The error estimate is very useful in these cases. An extensive study about approximations of fixed points can be found in [2]. One kind of a generalization of the Banach Contraction Principle is the notation of cyclical maps [7], i.e. T (A) ⊆ B and T (B) ⊆ A. Because a non-self mapping T : A → B does not necessarily have a fixed point, one often attempts to find an element x which is in some sense closest to T x. Best proximity point theorems are relevant in this perspective. The notation of best proximity point is introduced in [5]. This definition is more general than the notation of cyclical maps, in sense that if the sets intersect, then every best proximity point is a fixed point. A sufficient condition for existence and the uniqueness of best proximity points in uniformly convex Banach spaces is given in [5]. Since the publication [5] the problem for existence and uniqueness of best proximity point was widely investigated see for example [8, 11] and the research on this problem continues. In contrast with all the results about fixed points for self maps and cyclic maps, where "a priori error estimates" and "a posteriori error estimates" are obtained there are no such results about best proximity points. We have obtained "a priori error estimates" and "a posteriori error estimates" for the cyclic contractions from [5]. 2 Preliminaries In this section we give some basic definitions and concepts which are useful and related to the best proximity points. Let (X, ρ) be a metric space. Define a distance between two subset A, B ⊂ X by dist(A, B) = inf{ρ(x, y) : x ∈ A, y ∈ B}. For simplicity of the notations we will denote dist(A, B) with d. Let A and B be nonempty subsets of a metric space (X, ρ). The map T : AS B → AS B is called a cyclic map if T (A) ⊆ B and T (B) ⊆ A. A point ξ ∈ A is called a best proximity point of the cyclic map T in A if ρ(ξ, T ξ) = dist(A, B). Let A and B be nonempty subsets of a metric space (X, ρ). The map T : AS B → AS B is called a cyclic contraction map if T is a cyclic map and for some k ∈ (0, 1) there holds the inequality ρ(T x, T y) ≤ kρ(x, y)+(1−k)d for any x ∈ A, y ∈ B. The definition for cyclic contraction is introduced in [5]. The best proximity results need norm-structure of the space X. When we investigate a Banach space (X,k·k) we will always consider the distance between the elements to be generated by the norm 1 k·k i.e. ρ(x, y) = kx− yk. We will denote the unit sphere and the unit ball of a Banach space (X,k·k) by SX and BX respectively. The assumption that the Banach space (X,k · k) is uniformly convex plays a crucial role in the investigation of best proximity points. Definition 2.1. Let (X,k · k) be a Banach space. For every ε ∈ (0, 2] we define the modulus of convexity of k · k by δk·k(ε) = inf (cid:26)1 −(cid:13)(cid:13)(cid:13)(cid:13) x + y 2 (cid:13)(cid:13)(cid:13)(cid:13) : x, y ∈ BX ,kx − yk ≥ ε(cid:27) . The norm is called uniformly convex if δX (ε) > 0 for all ε ∈ (0, 2]. The space (X,k · k) is then called uniformly convex space. The results from [5] and [6] are summarized in the next theorem. Theorem 1. ([5, 6]) Let A and B be nonempty closed and convex subsets of a uniformly convex Banach space. Let T : A ∪ B → A ∪ B be a cyclic contraction map. Then there is a unique best proximity point ξ of T in A, T ξ is a unique best proximity point of T in B and ξ = T 2ξ = T 2nξ. Further if x0 ∈ A and xn+1 = T xn, then {x2n}∞n=1 converges to ξ and x2n+1 converges to T ξ. For any uniformly convex Banach space X there holds the inequality (cid:13)(cid:13)(cid:13)(cid:13) x + y 2 − z(cid:13)(cid:13)(cid:13)(cid:13) ≤ (cid:16)1 − δX (cid:16) r R(cid:17)(cid:17) R (1) for any x, y, z ∈ X, R > 0, r ∈ [0, 2R], kx − zk ≤ R, ky − zk ≤ R and kx − yk ≥ r. If (X,k·k) is a uniformly convex Banach space, then δX(ε) is strictly increasing function. Therefore if (X,k·k) is a uniformly convex Banach space then there exists the inverse function δ−1 of the modulus of convexity. If there exist constants C > 0 and q > 0, such that the inequality δk·k(ε) ≥ Cεq holds for every ε ∈ (0, 2] we say that the modulus of convexity is of power type q. It is well known that for any Banach space and for any norm there holds the inequality δ(ε) ≤ Kε2. The modulus of convexity with respect to the canonical norm k·kp in ℓp or Lp is δk·kp(ε) = 1− pq1 −(cid:0) ε 2(cid:1)p for p ≥ 2 and for 1 < p < 2 p = 2. It is 2(cid:1)p +(cid:12)(cid:12)1 − δ − ε the modulus of convexity δk·kp(ε) is the solution of the equation (cid:0)1 − δ + ε 2(cid:12)(cid:12) well known that the modulus of convexity with respect to the canonical norm in ℓp or Lp is of power type and there holds the inequalities δk·kp(ε) ≥ εp for p ∈ (1, 2) [9]. An extensive study of the Geometry of Banach spaces can be found in [1, 3, 4]. The next lemma is easy to get and it is used without stating it in most of the articles about best proximity points. p2p for p ≥ 2 and δk·kp(ε) ≥ (p−1)ε2 8 Lemma 2.1. Let A and B be nonempty subsets of a metric space (X, ρ) and let T : A ∪ B → A ∪ B be a cyclic contraction map. Then for every x ∈ A∪ B there holds the inequality ρ(T nx, T n+1x)− d ≤ kn (ρ(x, T x) − d). 3 Error estimates for best proximity points Theorem 2. Let A and B be nonempty, closed and convex subsets of a uniformly convex Banach (X,k · k) space, such that d = dist(A, B) > 0, and let there exist C > 0 and q ≥ 2, such that δk·k(ε) ≥ Cεq. Let T : A ∪ B → A ∪ B be a cyclic contraction map. Then (i) there exists a unique best proximity point ξ of T in A, T ξ is a unique best proximity point of T in B and ξ = T 2ξ = T 2nξ; (ii) for any x0 ∈ A the sequence {x2n}∞n=1 converges to ξ and {x2n+1}∞n=1 converges to T ξ, where xn+1 = T xn, n = 0, 1, 2, . . . ; 2 (iii) a priori error estimate holds (cid:13)(cid:13)ξ − T 2nx(cid:13)(cid:13) ≤ kx − T xk 1 − q√k2 qrkx − T xk − d Cd (cid:16) q√k(cid:17)2n ; (iv) a posteriori error estimate holds (cid:13)(cid:13)T 2nx − ξ(cid:13)(cid:13) ≤ kT 2n−1x − T 2nxk 1 − q√k2 qrkT 2n−1x − T 2nxk − d Cd q√k. (2) (3) Proof. The proof of (i) and (ii) follows from Theorem 1. in the text field. We will use the notation Sn,m(x) = kT nx − T mxk − d, just to be able to fit some of the formulas (iii) For any x ∈ A, n ∈ N and l ≤ 2n there holds the inequality δk·k(cid:18) kT 2nx − T 2n+2xk d + klS2n−l,2n+1−l(x)(cid:19) ≤ klS2n−l,2n+1−l(x) d + klS2n−l,2n+1−l(x) . Indeed let x ∈ A be arbitrary chosen. From Lemma 2.1 we have the inequalities kT 2nx − T 2n+1xk ≤ d + klS2n−l,2n+1−l(x), kT 2n+2x − T 2n+1xk ≤ d + kl+1S2n−l,2n+1−l(x) < d + klS2n−l,2n+1−l(x) and kT 2n+2x − T 2nxk ≤ kT 2n+2x − T 2n+1xk + kT 2n+1x − T 2nxk ≤ 2(cid:0)d + klS2n−l,2n+1−l(x)(cid:1) . After a substitution in (1) with x = T 2nx, y = T 2n+2x, z = T 2n+1x, r = kT 2n+2x − T 2nxk and R = d + kl(cid:0)kT 2n−lx − T 2n+1−lxk − d(cid:1) = d + klS2n−l,2n+1−l(x) and using the convexity of the set A we get the chain of inequalities 2 T 2nx+T 2n+2x − T 2n+1x(cid:13)(cid:13)(cid:13) d ≤ (cid:13)(cid:13)(cid:13) ≤ (cid:16)1 − δk·k(cid:16) kT 2nx−T 2n+2xk d+klS2n−l,2n+1−l(x)(cid:17)(cid:17)(cid:0)d + klS2n−l,2n+1−l(x)(cid:1) . From (4) we obtain the inequality δk·k(cid:18) kT 2nx − T 2n+2xk d + klS2n−l,2n+1−l(x)(cid:19) ≤ klS2n−l,2n+1−l(x) d + klS2n−l,2n+1−l(x) . (4) (5) From the uniform convexity of X is follows that δk·k is strictly increasing and therefore there exists its inverse function δ−1 k·k , which is strictly increasing too. From (5) we get kT 2nx − T 2n+2xk ≤ (cid:16)d + klS2n−l,2n+1−l(x)(cid:17) δ−1 (t) ≤ (cid:0) t By the inequality δk·k(t) ≥ Ctq it follows that δ−1 k·k d + klS2n−l,2n+1−l(x) ≤ kT 2n−lx − T 2n+1−lxk we obtain d + klS2n−l,2n+1−l(x)(cid:19) . k·k (cid:18) klS2n−l,2n+1−l(x) C(cid:1)1/q. From (6) and the inequalities d ≤ (6) kT 2nx − T 2n+2xk ≤ (cid:0)d + klS2n−l,2n+1−l(x)(cid:1) qr klS2n−l,2n+1−l(x) (cid:16) q√k(cid:17)l ≤ kT 2n−lx − T 2n+1−lxk qq S2n−l,2n+1−l(x) C.(d+klS2n−l,2n+1−l(x)) Cd . (7) 3 From (i) and (ii) there exists a unique ξ, such that kξ − T ξk = d, T 2ξ = ξ and ξ is a limit of the After a substitution with l = 2n in (7) we get the inequality sequence {T 2nx}∞n=1 for any x ∈ A. P∞n=1(cid:13)(cid:13)T 2nx − T 2n+2x(cid:13)(cid:13) ≤ kx − T xk qqkx−T xk−d = kx − T xk qqkx−T xk−d Cd Cd P∞n=1(cid:16) q√k(cid:17)2n q√k2 1− q√k2 · and consequently the series P∞n=1(T 2nx− T 2n+2x) is absolutely convergent. Thus for any m ∈ N there holds ξ = T 2mx −P∞n=m(cid:0)T 2nx − T 2n+2x(cid:1) and therefore we get the inequality Xn=m(cid:13)(cid:13)T 2nx − T 2n+2x(cid:13)(cid:13) ≤ kx − T xk qrkx − T xk − d (cid:13)(cid:13)ξ − T 2mx(cid:13)(cid:13) ≤ · (cid:16) q√k(cid:17)2m 1 − q√k2 Cd ∞ . (iv) We will use the notation Pn,m(x) = kT nx − T mxk, just to be able to fit some of the formulas in the text field. After a substitution with l = 1 + 2i in (7) we obtain P2n+2i,2n+2(i+1)(x) ≤ P2n−1,2n(x) qr P2n−1,2n(x) − d Cd (cid:16) q√k(cid:17)1+2i . From (8) we get that there holds the inequality P2n,2(n+m)(x) ≤ Pm−1 ≤ Pm−1 = P2n−1,2n(x) qq P2n−1,2n(x)−d = P2n−1,2n(x) qq P2n−1,2n(x)−d i=0 P2n+2i,2n+2(i+1)(x) (cid:16) q√k(cid:17)1+2i i=0 P2n−1,2n(x) qq P2n−1,2n(x)−d i=0 (cid:16) q√k(cid:17)1+2i Pm−1 1−( q√k)2m q√k · 1− q√k2 Cd Cd Cd (8) (9) and after letting m → ∞ in (9) we obtain the inequality (cid:13)(cid:13)T 2nx − ξ(cid:13)(cid:13) ≤ kT 2n−1x − T 2nxk qrkT 2n−1x − T 2nxk − d Cd q√k 1 − q√k2 . 4 Remarks and an Example Following [2] we would like to say a few words about the error estimates. The a priori estimate (2) shows that, when starting from an initial guess x ∈ A the upper bound of approximation error for the 2n iterate is completely determined by the cyclic contraction coefficient k and the initial displacement kx − T xk. Similarly, the a posteriori estimate shows that, in order to obtain the desired error approximation kT 2n − ξk < ε of the fixed point by means of Picard iteration we need to stop the iterative process at the first step 2n for which the displacement between two consecutive iterates satisfies the inequality q√k < ε. Thus the a posteriori estimation offers a direct stopping kT 2n−1x−T 2nxk criterion for the iterative approximation of fixed points by Picard iteration, while the a priori estimation indirectly gives a stopping criterion. qqkT 2n−1x−T 2nxk−d 1− q√k2 Cd We will illustrate Theorem 2 with the next example. Example 1: Let consider the space R2 = {(x, y) : x, y ∈ R} endowed with the norms kxkp = ppxp + yp, for p > 1. The space (R,k · kp) is uniformly convex with modulus of convexity of power 4 type, provided that p > 1. Let us consider the sets A = {(x, y) ∈ R2 : y − x + 1 ≤ 0, y + x − 1 ≥ 0} and B = {(x, y) ∈ R2 : y − x − 1 ≥ 0, y + x + 1 ≤ 0}. p by T (x, y) = (−((1 − λ)sign(x) + λx),−λy), where sign(x) = 1 if x > 0, sign(x) = −1 if x < 0 and sign(x) = 0 if x = 0. It is easy to calculate dist(A, B) = 2. Let λ ∈ (0, 1) . Let us define a map T : R2 p → R2 We will show that the map T : A ∪ B → A ∪ B is a cyclic contraction with k = λ. Let z = (x, y) ∈ A. From x, y ≥ 0 we get −λy − (1 − λ + λx) + 1 = −(λy + λx − λ) ≤ 0 and −λy + (1 − λ + λx) − 1 = −(λy − λx + λ) ≥ 0. Therefore T (A) ⊆ B. The inclusion T (B) ⊆ A is proven in a similar fashion. Let us put u1 = (x1, y1) ∈ A, u2 = (x2, y2) ∈ B and e1 = (1, 0) ∈ A. It is easy to observe that e1 is a best proximity point of T in A, T (e1) = −e1 and T 2(e1) = T (−e1) = e1. We get the chain of inequalities kT (x1, y1) − T (x2, y2)kp ≤ pp2(1 − λ) + λ(x1 + x2)p + λ(y1 + y2)p ≤ k2(1 − λ)e1 + λ(u1 − u2)kp ≤ λku1 − u2kp + 2(1 − λ)ke1kp ≤ λku1 − u2kp + (1 − λ)d. Thus we can apply Theorem 2 to get error estimates of the successive iterations {x2n}∞n=1, where xn+1 = T xn. We will consider a numeric example with λ = 2−1. From [9] we get C = p − 1 C = , q = 2 for p ∈ (1, 2]. 8 1 p2p , q = p for p ≥ 2 and Table 1: Number 2n of iterations, needed by the a posteriori estimate for λ = 2−1 with an initial point x0 = (1000, 8) ε \ p 10−2 10−4 10−6 10−8 10−10 1.1 34 48 60 74 88 1.5 32 46 58 72 84 2 30 44 58 70 84 3 42 62 82 102 122 5 66 100 132 166 200 20 266 398 532 664 798 Table 2: Number 2n of iterations, needed by the a priori estimate for λ = 2−1 with an initial point x0 = (1000, 8) ε \ p 10−2 10−4 10−6 10−8 10−10 1.1 54 66 80 94 106 1.5 50 64 78 90 104 2 46 58 72 86 98 3 64 84 104 124 144 5 104 138 170 204 238 20 428 560 694 826 960 5 Conclusion and open questions We would like to mention that the error estimates give much larger number of the iterations that are needed. It is due to the fact that we use the modulus of convexity, which is the infinum of 1 −(cid:13)(cid:13) 2 (cid:13)(cid:13) among all x, y ∈ Sx, such that kx − yk ≥ ε. It may happen that the modulus of convexity is greater in the direction of the best proximity point ξ than in the other directions but for the estimation of x+y 5 the error we do not use it. We would like to pose the following question is it possible to get better estimates if we use the directional modulus of convexity δk·k(x, ε)? For the estimations we use geometric progression and that is why we impose the condition for the modulus of convexity to be of power type. Is it possible to obtain error estimates if the modulus of convexity is not of power type? Is it possible to obtain error estimates for the sequence of successive iterates for weak cyclic Kannan contractions [11] and for cyclic φ -- contractions [8]? References [1] Beauzamy, B., Introduction to Banach Spaces and their Geometry, North -- Holland Publishing Company, Amsterdam, 1979. [2] Berinde, V., Iterative Approximation of Fixed Points, Springer, Berlin, 2007. [3] Deville, R., Godefroy, G., and Zizler, V., Smothness and renormings in Banach spaces, Pitman Monographs and Surveys in Pure and Applied Mathematics, 1993. [4] Fabian, M., Habala, P., H´ajek, P., Montesinos, V., Pelant, J. and Zizler, V., Functional Analysis and Infinite-Dimensional Geometry, Springer-Verlag, New York, 2011. [5] Eldred, A. and Veeramani, P., Existence and convergence of best proximity points, J. Math. Anal. Appl., 323 (2006), No. 2, 1001 -- 1006. [6] Karpagam, S. and Sushama Agrawal, Existence of best proximity Points of P -- cyclic contractions, Fixed Point Theory, 13 (2012), No. 1, 99 -- 105. [7] Kirk, W., Srinivasan, P. and Veeramani, P., Fixed points for mappings satisfying cyclical contrac- tive conditions, Fixed Point Theory, 4 (2003), No. 1, 79 -- 89. [8] Madalina, P., Rus, I. A., Fixed point theory for cyclic φ -- contractions, Nonlinear Anal., 72 (2010), No. 3 -- 4, 11811187. [9] Meir, A., On the Uniform Convexity of Lp Spaces, 1 < p < 2, Illinois J. Math., 28 (1984), No. 3, 420 -- 424. [10] Nordlander, G., The modulus of convexity in normed linear spaces, Ark. Mat., 4 (1960), No. 1, 15 -- 17. [11] Petric, M., Best proximity point theorems for weak cyclic Kannan contractions, Filomat, 25 (2011), No. 2, 145 -- 154. 6
1206.4848	1	1206	2012-06-21T12:12:28	Une remarque sur les espaces d'interpolation faiblement localement uniform\'ement convexes	[ "math.FA" ]	Let $(A_0, A_1)$ be an interpolation couple, and let $B_j$ be the closure of $A_0^\ast \cap A_1^\ast$ in $A_j^\ast$, $j = 0, 1$. For every $\theta \in \, ]0, 1[$, there exists a natural one to one contraction $R^\theta : A^\theta \rightarrow (B_0^\ast, B_1^\ast)^\theta$. For some $\beta \in \, ]0, 1[$, the closure of $R^\beta (A^\beta)$ in $(B_0^\ast, B_1^\ast)^\beta$ is supposed to be weakly LUR. Then $A^\theta = A_\theta$ for every $\theta \in \, ]0, 1[$.	math.FA	math	Une remarque sur les espaces d'interpolation faiblement localement uniform´ement convexes Daher Mohammad D´epartement de Math´ematiques, Universit´e Paris 7 e-mail: [email protected] June 10, 2018 1 dans A∗ R´esum´e: Soient (A0, A1) un couple d'interpolation, et Bj l'adh´erence 0 ∩ A∗ j , j = 0, 1. Pour tout θ ∈ ]0, 1[, il existe une contraction 1)θ. On suppose que, pour un β ∈ ]0, 1[, 1)β est faiblement LUR. Alors Aθ = Aθ de A∗ injective naturelle Rθ : Aθ → (B∗ l'adh´erence de Rβ(Aβ) dans (B∗ pour tout θ ∈ ]0, 1[. 0, B∗ 0, B∗ 0 ∩ A∗ Abstract: Let (A0, A1) be an interpolation couple, and let Bj be the j , j = 0, 1. For every θ ∈ ]0, 1[, there exists a natural 1)θ. For some β ∈ ]0, 1[, the closure 1)β is supposed to be weakly LUR. Then Aθ = Aθ for closure of A∗ one to one contraction Rθ : Aθ → (B∗ of Rβ(Aβ) in (B∗ every θ ∈ ]0, 1[. 1 in A∗ 0, B∗ 0, B∗ AMS Classification: 46B70 Mots cl´es: Interpolation, espace faiblement-LUR Avertissement: Une premi`ere version de ce travail a ´et´e publi´ee dans Col- loq. Math. Vol. 113, No. 2, 197-204, (2011). Le lemme 1 de cette version est malheureusement faux, ce qui oblige `a corriger l'ensemble. L'auteur remer- cie Sten Kaijser de lui avoir signal´e l'erreur dans la preuve. Un rectificatif indiquant les corrections a ´et´e envoy´e `a Colloq. Math.. Il nous a cependant sembl´e utile de pr´esenter une version r´evis´ee compl`ete. 1 1 Introduction et notations On note X ∗ le dual d'un espace de Banach X. Soit A = (A0, A1) un couple d'interpolation complexe, au sens de [BL]. Soit S = {z ∈ C ; 0 ≤ Re z) ≤ 1}. Rappelons d'abord la d´efinition de l'espace d'interpolation Aθ, o`u θ ∈ ]0, 1[ [BL, chapitre 4]. On note F (A) l'espace des fonctions F `a valeurs dans A0 + A1, continues born´ees sur S, holomorphes `a l'int´erieur de S, telles que, pour j ∈ {0, 1}, F (j + iτ ) prend ses valeurs dans Aj et kF (j + iτ )kAj → 0, quand τ → +∞. On munit F (A) de la norme kF kF (A) = max(sup τ ∈R kF (iτ )kA0, sup τ ∈R kF (1 + iτ )kA1). L'espace Aθ = (A0, A1)θ = {F (θ); F ∈ F (A)} est un Banach [BL, theorem 4.1.2] pour la norme d´efinie par kakAθ = inf nkF kF (A); F (θ) = ao . Rappelons maintenant la d´efinition de l'espace d'interpolation Aθ [BL, chapitre 4]. On note G(A) l'espace des fonctions g `a valeurs dans A0 + A1, continues sur S, holomorphes `a l'int´erieur de S, telles que z → (1 + z)−1kg(z)kA0+A1 est born´ee sur S, g(j + iτ ) − g(j + iτ ′) ∈ Aj pour tous τ, τ ′ ∈ R, j ∈ {0, 1}, et la quantit´e suivante est finie: k gkQG(A) = max  sup τ,τ ′∈R τ 6=τ ′ (cid:13)(cid:13)(cid:13)(cid:13) g(iτ ) − g(iτ ′) τ − τ ′ (cid:13)(cid:13)(cid:13)(cid:13)A0 , sup τ,τ ′∈R τ 6=τ ′ (cid:13)(cid:13)(cid:13)(cid:13) g(1 + iτ ) − g(1 + iτ ′) τ − τ ′ .   (cid:13)(cid:13)(cid:13)(cid:13)A1 Cette quantit´e d´efinit une norme sur l'espace QG(A), quotient de G(A) par les applications constantes `a valeurs dans A0 ∩ A1, et QG(A) est complet pour cette norme [BL, lemma 4.1.3]. On rappelle [BL, p. 89] que, pour g ∈ G(A), kg′(z)kA0+A1 ≤ k gkQG(A), z ∈ S. (1) C'est une cons´equence imm´ediate de l'in´egalit´e 2 (cid:13)(cid:13)(cid:13)(cid:13) g(z + it) − g(z) t (cid:13)(cid:13)(cid:13)(cid:13)A0+A1 ≤ k gkQG(A), z ∈ S, t ∈ R∗, qui d´ecoule de la d´efinition de k gkQG(A) et du th´eor`eme des trois droites [BL, lemma 1.1.2] appliqu´e aux fonctions z → h(g(z + it) − g(z))/t, a∗i, t r´eel fix´e, o`u a∗ parcourt la boule unit´e de A∗ 0 ∩ A∗ 1. L'espace Aθ = { g′(θ); g ∈ G(A) } est un Banach [BL, theorem 4.1.4] pour la norme d´efinie par kakAθ = inf nk gkQG(A) ; g′(θ) = ao . D'apr`es (1) kakA0+A1 ≤ kakAθ. La contraction Aθ → A0 + A1 est injective par d´efinition de Aθ. D'apr`es [B], Aθ s'identifie isom´etriquement `a un sous espace de Aθ. D'apr`es [BL, theorem 4.2.2], A0 ∩ A1 est toujours dense dans Aθ, 0 < 0 + A∗ 1, 1 = (A0 + A1)∗ [BL, theorem 2.7.1], on peut appliquer le th´eor`eme 1)θ, θ ∈ ]0, 1[ [BL, theorem θ < 1. Si A0 ∩ A1 est dense dans A0 et A1, on a (A0 ∩ A1)∗ = A∗ A∗ d'it´eration [BL, theorem 4.6.1] et (Aθ)∗ = (A∗ 4.5.1]. On fait cette hypoth`ese dans la suite. 0 ∩ A∗ 0, A∗ D´efinition 1 [DGZ] Un espace de Banach X est localement uniform´ement convexe, ce qu'on note LUR (resp. faiblement LUR) si, pour tout x ∈ X et pour toute suite (xn)n≥0 dans X satisfaisant kxnk2/2 + kxk2/2 − k(xn + x)/2k2 →n→∞ 0, alors xn →n→∞ x en norme (resp. faiblement). 2 R´esultats Notons Bj B0 ∩ B1 = A∗ isom´etriquement, pour θ ∈ ]0, 1[, l'adh´erence de A∗ 0 ∩ A∗ Il est clair que 1, isom´etriquement. D'apr`es [BL, Theorem 4.2.2 b)] on a j , j = 0, 1. 1 dans A∗ 0 ∩ A∗ (B0, B1)θ = (A∗ 0, A∗ 1)θ . (2) Comme B0 ∩ B1 est dense dans Bj, le dual de Bθ = (B0, B1)θ est (B∗ [BL, theorem 4.5.1] et, d'apr`es [BL, theorem 2.7.1], 0, B∗ 1)θ 3 B∗ 1)∗ = (A0 + A1)∗∗. 0 + B∗ 1 = (B0 ∩ B1)∗ = (A∗ 0 ∩ A∗ 0 + B∗ 1. En particulier, A0 + A1 s'identifie isom´etriquement `a un sous espace ferm´e de B∗ Soit ij : Bj → A∗ Aj → B∗ j l'application identit´e; la restriction de son adjoint i∗ j j , j = 0, 1, est contractante. : Lemme 2 Soit R : QG(A0, A1) → QG(B∗ par g(j + i ·) → i∗ et induit une contraction injective 1) l'application qui est d´efinie j (g(j + i ·)), j = 0, 1. L'application R est une contraction 0, B∗ Rθ : Aθ → (B∗ 0, B∗ 1)θ, θ ∈ ]0, 1[. Il est clair que R est une contraction (non injective en D´emonstration: 1)θ `a des quotients de QG(A0, A1) et g´en´eral). On identifie Aθ et (B∗ 1) respectivement. Notant que (R( g))′(θ) = Rθ(g′(θ)), R induit QG(B∗ une contraction Rθ sur ces quotients. Notons que, pour a ∈ Aθ, pour b ∈ B0 ∩ B1 = A∗ 1 = (A0 + A1)∗ (espace dense dans Bθ), 0 ∩ A∗ 0, B∗ 0, B∗ hRθ(a), bi = ha, bi. Si Rθ(a) = 0, alors ha, bi = 0 pour tout b ∈ B0 ∩ B1 = (A0 + A1)∗, d'o`u a = 0 dans A0 + A1, et dans Aθ. (cid:4) Th´eor`eme 3 Soient (A0, A1) un couple d'interpolation complexe et Bj l'ad- j , j = 0, 1, β ∈ ]0, 1[, Rβ d´efinie comme ci-dessus. h´erence de A∗ Soit Z β l'adh´erence de Rβ(Aβ) dans (B∗ 1)β. Supposons que Z β est un espace faiblement-LUR. Alors Aθ = Aθ, pour tout θ ∈ ]0, 1[. 1 dans A∗ 0 ∩ A∗ 0, B∗ La d´emonstration n´ecessite les lemmes suivants. Lemme 4 Pour tout θ ∈ ]0, 1[, Rθ est une isom´etrie: Aθ → (B∗ 0, B∗ 1)θ. D´emonstration: Comme Aθ s'identifie `a un sous-espace de Aθ [B], Rθ est 1)θ par le lemme 2. Comme A0 ∩ A1 contractante: Aθ = (A0, A1)θ → (B∗ est dense dans Aθ, il suffit de montrer que kakAθ ≤ kRθ(a)k(B∗ 1 )θ lorsque a ∈ A0 ∩ A1. Soit ε > 0; comme (Aθ)∗ = (A∗ 1)θ, il existe g ∈ G(A∗ 1) tel que 0, B∗ 0, A∗ 0, A∗ 0 ,B∗ 4 kakAθ < ha, g′(θ)i + ε, Soient k gkQG(A∗ 0,A∗ 1) ≤ 1. (3) Fn(z) = in [g(z + i/n) − g(z)], z ∈ S et Fn,δ(z) = eδz2 Fn,δ tend vers 0 `a l'infini sur le bord, d'o`u Fn,δ ∈ F (A∗ Fn(z), pour δ > 0. Comme Fn est born´ee sur le bord de S, 1). Par d´efinition 0, A∗ kFn,δ(θ)k(A∗ 0,A∗ 1)θ ≤ kFn,δkF (A∗ 0 ,A∗ 1) ≤ eδ sup Fn(z) ≤ eδk gkQG(A∗ 1) ≤ eδ. 0,A∗ z∈S D'o`u, pour tout n, par (2), kFn(θ)k(B0,B1)θ = ke−δθ2 Fn,δ(θ)k(A∗ 0,A∗ 1)θ = lim δ→0 ke−δθ2 Fn,δ(θ)k(A∗ 0,A∗ 1)θ ≤ 1. Comme g est holomorphe `a valeurs dans A∗ ha, g′(θ)i. Il existe n0 assez grand tel que, d'apr`es (3), 0+A∗ 1 = (A0∩A1)∗, ha, Fn(θ)i → n→∞ −2ε + kakAθ < ha, Fn0(θ)i ≤ kRθ(a)k(B∗ 1 )θ kFn0(θ)k(B0,B1)θ ≤ kRθ(a)k(B∗ 0 ,B∗ 1 )θ , 0 ,B∗ d'o`u l'in´egalit´e cherch´ee lorsque ε → 0. (cid:4) Lemme 5 Soient g ∈ G(A), θ ∈ ]0, 1[. L'application: τ → Rθ(g′(θ + iτ )) 1)θ, l'application: τ → est born´ee de R dans (B∗ 1)θ. Pour tout c ∈ (B∗ 0, B∗ 0, B∗ 1 )θ est s.c.i. sur R. 0 ,B∗ (cid:13)(cid:13)c + Rθ(g′(θ + iτ ))(cid:13)(cid:13)(B∗ D´emonstration: Par d´efinition de Aθ, g′(θ) ∈ Aθ ; par le lemme 2 kRθ(g′(θ))k(B∗ 1 )θ ≤ kg′(θ)kAθ ≤ k gkQG(A). 0 ,B∗ La fonction giτ d´efinie par giτ (z) = g(z+it), z ∈ S, τ ∈ R, v´erifie k giτ kQG(A) = k gkQG(A), donc (cid:13)(cid:13)Rθ(g′ iτ (θ))(cid:13)(cid:13)(B∗ D'apr`es (2), et comme B0 ∩ B1 = A∗ 0 ,B∗ 1 )θ ≤ k gkQG(A). 0 ∩ A∗ 1 est dense dans Bθ, on a 5 kc + Rθ(g′(θ + iτ ))k(B∗ 0 ,B∗ 1 )θ = sup(cid:8)(cid:12)(cid:12)hb, c + Rθ(g′(θ + iτ ))i(cid:12)(cid:12) ; kbk(B0,B1)θ ≤ 1(cid:9) = sup(cid:8)ha∗, c + g′(θ + iτ )i ; a∗ ∈ A∗ 0 ∩ A∗ 1, ka∗k(A∗ 0 ,A∗ 1)θ ≤ 1(cid:9) . 0 ∩ A∗ Comme g est holomorphe `a valeurs dans A0 + A1, pour tout a∗ ∈ A∗ 1 = (A0 + A1)∗, les applications τ → ha∗, c + g′(θ + iτ )i sont continues sur R. Leur supremum est donc s.c.i.. (cid:4) Lemme 6 Soient C = (C0, C1) un couple d'interpolation, β ∈ ]0, 1[, Z β un sous-espace ferm´e faiblement-LUR de C β, g ∈ G(C). On suppose que l'application φβ : τ ∈ R → g′(β + iτ ) est born´ee `a valeurs dans Z β et que l'application kc + φβkC β est s.c.i., pour tout c ∈ Z β fix´e. Alors φβ est p.s. ´egale `a une fonction fortement mesurable: R → Z β ⊂ C β. Preuve: Comme s → kφβ(τ ) + φβ(s)kZ β est s.c.i. sur R, si τn → τ , 0 ≤ limn→+∞En = lim(cid:8)2kφβ(τ )k2 ≤ 2 kφβ(τ )k2 ≤ 2 kφβ(τ )k2 = 2 limkφβ(τn)k2 Z β + 2kφβ(τn)k2 Z β + 2 limkφβ(τn)k2 Z β + 2 limkφβ(τn)k2 Z β − 2 kφβ(τ )k2 Z β − kφβ(τ ) + φβ(τn)k2 Z β(cid:9) Z β − limkφβ(τ ) + φβ(τn)k2 Z β Z β − 4 kφβ(τ )k2 Z β Z β . Comme kφβkZ β est mesurable born´ee, pour tout N et ε > 0, il existe, d'apr`es le th´eor`eme de Lusin, un compact KN,ε ⊂ [−N, N], de mesure > 2N − ε, sur lequel kφβkZ β est continue. Soit (τn)n≥0 une suite dans KN,ε convergeant vers τ . D'apr`es ce qui pr´ec`ede limn→+∞En = 0. Comme En ≥ 0, En →n→∞ 0. Par d´efinition de la propri´et´e faiblement-LUR de Z β, cela entraıne que φβ(τn) → φβ(τ ) faiblement dans Z β, c`ad φβ est faiblement continue sur KN,ε. Soit Y le sous espace ferm´e de Z β engendr´e par φβ(KN,ε). Alors Y est s´eparable: sinon, ´etant donn´ee une suite (sn)n≥1 dense dans KN,ε, il existe, d'apr`es le th´eor`eme de Hahn-Banach, z ∈ Y ∗, non nul, tel que (φβ(sn), z) = 0 pour tout n; par continuit´e s → (φβ(s), z) est nulle sur KN,ε, d'o`u z = 0 et la contradiction. Par le th´eor`eme de Pettis [DU, theorem II 2], φβ est fortement mesurable: KN,ε → Y ⊂ Z β. Cela montre le r´esultat annonc´e. (cid:4) 6 Lemme 7 Soient g ∈ G(A), φθ(t) = g′(θ + it), t ∈ R. i) Si φθ est `a valeurs dans un sous espace ferm´e s´eparable Z de Aθ, elle est fortement mesurable: R → Aθ. ii) Si φθ est `a valeurs dans un sous espace ferm´e s´eparable Z de Aθ, elle est fortement mesurable: R → Aθ. Dans la suite on utilise seulement i), dans la preuve du lemme 8 d). On donne deux preuves de i) (noter que ii) implique i)). 1)θ et, Preuve: i) D'apr`es le lemme 4, Z est un sous espace ferm´e de (B∗ d'apr`es le lemme 5, l'application t → kφθ(t) − ckZ est s.c.i. pour tout c ∈ Z. L'image r´eciproque par φθ de toute boule ouverte de Z est donc un bor´elien. Comme Z est s´eparable, tout ouvert de Z est r´eunion d´enombrable de boules, donc φθ est bien mesurable `a valeurs dans Z. (cid:4) 0, B∗ ii) Soient J l'injection canonique: Z → A0 + A1, et Y l'adh´erence de J(Z) dans A0 + A1. Comme Z et Y sont des espaces polonais, comme J est continue, J −1 est bor´elienne: J(Z) → Z, voir par exemple [A]. Comme J ◦ φθ : R → A0 + A1 est continue et `a valeurs dans J(Z), comme φθ = J −1 ◦ (J ◦ φθ), alors φθ est bor´elienne: R → Z. (cid:4) Lemme 8 Soient g ∈ G(A), β ∈ ]0, 1[ et φβ(·) = g′(β + i ·). a) On suppose que Rβ ◦ φβ est p.s. ´egale `a une fonction fortement mesurable: R → (B∗ 0, B∗ 1)β. Alors φβ est p.s. `a valeurs dans Aβ. On suppose d´esormais que φβ est p.s. ´egale `a une fonction fortement mesurable: R → Aβ. Alors b) pour θ 6= β, g′(θ) ∈ Aθ. c) pour tout θ 6= β, φθ est `a valeurs dans un sous espace s´eparable de Aθ. d) g′(β) ∈ Aβ. On a not´e Rβ ◦ φβ la fonction: t → Rβ(g′ Preuve: a) ´etape 1: Comme g est holomorphe `a l'int´erieur de S, pour tous t ∈ R, h > 0, θ ∈ ]0, 1[, on a, dans A0 + A1, it(β)). g(θ + i(t + h)) − g(θ + it) = Z t+h t g′(θ + iτ ) dτ (4) Posons o`u g(1) − g(0) = α0 + α1 (αj ∈ Aj, j = 0, 1), avec g1 = g − g(0) − α0 7 kg(1) − g(0)kA0+A1 = kα0kA0 + kα1kA1. D'apr`es l'in´egalit´e des accroissements finis et (1) kg(1) − g(0)kA0+A1 ≤ k gkQG(A). Alors g1 : S → A0 + A1 est continue sur S et holomorphe `a l'int´erieur de S. Comme g ∈ G(A), pour tout τ ∈ R et j ∈ {0, 1}, on a kg1(j + iτ )kAj ≤ kg(j + iτ ) − g(j)kAj + kαjkAj ≤ (1 + τ )k gkQG(A). L'application z → Gε(z) = eεz2 g1(z) est donc dans F (A) pour tout ε > 0. En particulier, pour tout t ∈ R, Gε(θ + it) ∈ Aθ, donc g1(θ + it) ∈ Aθ. D'o`u g1(θ + i(t + h)) − g1(θ + it) = g(θ + i(t + h)) − g(θ + it) ∈ Aθ. Alors, d'apr`es (4), R t+h t g′(θ + iτ ) dτ est dans Aθ, pour t et h r´eels. ´etape 2: Par hypoth`ese Rβ ◦ φβ est p.s. ´egale `a une fonction forte- ment mesurable: R → Z β, o`u Z β est l'adh´erence de Aβ dans (B∗ 1)β. Le th´eor`eme de diff´erentiabilit´e de Lebesgue [DU, chap. II theorem 9 p 48] entraınent que, p.s., on a dans Z β l'´egalit´e 0, B∗ iRβ◦φβ(it) = lim h→0 1 h Z t+h t Rβ◦φβ(iτ ) dτ = lim h→0 Rβ(cid:0) 1 h Z t+h t g′(β+iτ ) dτ(cid:1), (5) o`u h est r´eel. D'apr`es la fin de l'´etape 1 appliqu´ee en β et le lemme 4, cette limite dans Z β est en fait une limite dans Aβ, c`ad p.s. g′(β + i ·) ∈ Aβ. b) On suppose d'abord θ > β. ´etape 1: Soit V (z) = g1(β + (1 − β)z), z ∈ S. Cette fonction `a valeurs dans A0 + A1 est holomorphe `a l'int´erieur de S et continue sur S, donc s'exprime `a l'aide de la mesure harmonique sur le bord de S. Pour v´erifier que V , vue comme fonction `a valeurs dans Aβ + A1, est holomorphe `a l'int´erieur de S et continue sur S, il suffira donc de voir que V est continue sur l'axe imaginaire, `a valeurs dans Aβ. 8 On va montrer que V ∈ G(Aβ, A1) avec une norme ≤ (1 − β)k gkQG(A). L'in´egalit´e correspondante sur la droite Re = 1 est ´evidente. Pour la v´erifier sur l'axe imaginaire, posons, pour τ, τ ′ r´eels fix´es, Fτ,τ ′(ξ) = g(ξ + i(1 − β)τ ) − g(ξ + i(1 − β)τ ′) τ − τ ′ , ξ ∈ S, d'o`u Fτ,τ ′(β) = (V (iτ ) − V (iτ ′))/(τ − τ ′), et Fτ,τ ′(1) = (V (1 + iτ ) − V (1 + iτ ′))/(τ − τ ′). Pour tout t ∈ R, on a kFτ,τ ′(j + it)kAj ≤ (1 − β)k gkQG(A), j ∈ {0, 1}. Comme dans l'´etape 1 de a), pour tout ε > 0, l'application ξ → Hε,τ,τ ′(ξ) = eεξ2 Fτ,τ ′(ξ) v´erifie kHε,τ,τ ′kF (A) ≤ eε(1 − β)k gkQG(A), d'o`u kFτ,τ ′(β)kAβ ≤ (1 − β)k gkQG(A). On a donc, pour tous τ, τ ′ r´eels, kV (iτ ) − V (iτ ′)kAβ ≤ τ − τ ′(1 − β)k gkQG(A), ce qui prouve la continuit´e de V sur l'axe imaginaire, `a valeurs dans Aβ, et l'assertion annonc´ee. ´etape 2: d'apr`es la preuve de a), pour h r´eel, p.s. (V (i(τ + h)) − V (iτ ))/h = (1 − β)g′(β + (1 − β)iτ ) dans Aβ. lim h→0 D'apr`es [BL, lemma 4.3.3], on a alors V ′(η) ∈ (Aβ, A1)η, η ∈ ]0, 1[. ´etape 3: Choisissons η tel que θ = (1 − η)β + η. D'apr`es le th´eor`eme de r´eit´eration [BL, theorem 4.6.1], (Aβ, A1)η = Aθ, donc V ′(η) = (1 − β)g′(θ) ∈ Aθ, ce qui ach`eve la preuve lorsque β < θ. 9 Si 0 < θ < β le raisonnement est analogue, en rempla¸cant V par W (z) = g1(βz) ∈ G(A0, Aβ), telle que limh→0(W (1 + i(τ + h)) − W (1 + iτ ))/h existe dans Aβ, pour presque tout τ , avec h r´eel. c) Soit A′ 0 est s´eparable, ainsi que (A′ R}. Comme g1 est continue sur S, A′ son adh´erence Y dans Aβ. Par l'´etape 2 de b) appliqu´ee au couple (A′ g′(β + it) est p.s. dans (A′ 0 ⊂ A0 le sous espace ferm´e s´eparable engendr´e par {g1(it), t ∈ 0, A1)β et 0, A1), 0, A1)β, donc p.s. dans Y , ce qui r`egle le cas θ = β. Pour le cas β < θ, rempla¸cons la fonction V de l'´etape 1 de b) par Vt(z) = V (z + it), avec t fix´e r´eel. Comme en b), Vt ∈ G(Y, A1), V ′ t (η) ∈ (Y, A1)η, η ∈ ]0, 1[ et (Y, A1)η est s´eparable. Soit η d´efini comme dans l'´etape 3 de b). Comme ci-dessus, V ′ t (η) = (1 − β)g′(θ + i(1 − β)t). Soit Zθ l'adh´erence de (Y, A1)η dans (Aβ, A1)η = Aθ; Zθ est donc s´eparable et φθ = g′(θ + i.) est `a valeurs dans Zθ. On raisonne de fa¸con analogue si 0 < θ < β en consid´erant Wt(z) = W (z + it): Wt est dans G(A′ 0, Y ). d) Soit θ > β. Par c) et le lemme 7 i), φθ est fortement mesurable `a valeurs dans Aθ. Alors b) appliqu´e en ´echangeant les roles de β et θ donne g′(β) ∈ Aβ. (cid:4) D´emonstration du th´eor`eme 1: Soient a ∈ Aβ et g ∈ G(A) tel que a = g′(β). D'apr`es le lemme 5, l'application Rβ ◦ φβ : τ ∈ R → Rβ(g′(β + iτ )) est `a valeurs dans Z β ⊂ (B∗ 1)β et v´erifie les hypoth`eses du lemme 6 pour C β = (B∗ 1)β. Grace `a l'hypoth`ese sur Z β, on peut appliquer le lemme 6, donc Rβ ◦ φβ est p.s. ´egale `a une fonction fortement mesurable `a valeurs 1)β, et φβ est p.s. ´egale `a une fonction fortement mesurable `a dans (B∗ valeurs dans Aβ par le lemme 8 a). D'apr`es le lemme 8 b) g′(θ) ∈ Aθ pour tout θ 6= β. Il en r´esulte que Aθ = Aθ, pour tout θ 6= β. Enfin par le lemme 8 d) g′(β) = a ∈ Aβ, d'o`u Aβ = Aβ. (cid:4) 0, B∗ 0, B∗ 0, B∗ Proposition 9 Soient A0, A1 deux espaces de Banach tels que A0 s'injecte continuement dans A1, et β ∈ ]0, 1[. Si Aβ a la propri´et´e de Radon-Nikodym analytique (d´efinie par exemple dans [DU]) pour un β ∈ ]0, 1[, alors Aθ = Aθ pour tout θ ∈ ]0, 1[. Pour β = 1 ce r´esultat est [HP, Proposition 3.1]; appliqu´e au couple (A0, Aβ), il donne la conclusion pour θ ∈ ]0, β[. Preuve: D'apr`es le lemme 8 b), d), il suffit de montrer que pour toute g ∈ G(A), φβ est p.s. mesurable `a valeurs dans Aβ. 10 On a mentionn´e dans la preuve du lemme 8 b) que la fonction z → W (z) = g1(βz) est dans G(A0, Aβ). `A l'int´erieur de S, W ′ est donc holomorphe `a valeurs dans A0 + Aβ = Aβ; par (1) elle est born´ee. Comme Aβ poss`ede la propri´et´e de Radon-Nikodym analytique, W ′ admet p.s. des limites non tangentielles au bord de S. Soit ψ la limite p.s. (dans Aβ) de W ′ sur la droite Re = 1; ψ est donc p.s. mesurable `a valeurs dans Aβ. Comme g′ est continue (`a valeurs dans A0 +A1 = A1) sur S, ψ coincide p.s. avec la fonction t → βg′(β + iβt), ce qui ach`eve la preuve. (cid:4) Corollaire 10 Si A0 s'injecte continuement dans A1 avec image dense, si 1)β admet une norme ´equivalente Aβ est un treillis de Banach, et si (A∗ LUR pour un β ∈ ]0, 1[, alors (A∗ 1)θ pour tout θ ∈ ]0, 1[. 1)θ = (A∗ 0, A∗ 0, A∗ 0, A∗ 0, A∗ Preuve: Comme ℓ∞ n'admet aucune norme ´equivalente LUR [DGZ, Chap. 1)β ne contient pas ℓ∞ isomorphiquement. II, theorem 7.10], (Aβ)∗ = (A∗ Alors, d'apr`es un r´esultat de Bessaga-Pelczy´nski [DU, Corollary I 6], l'espace 1)β ne contient pas c0 isomorphiquement; comme c'est un treillis, il (A∗ poss`ede la propri´et´e de Radon-Nikodym analytique [E]. Son sous-espace isom´etrique (A∗ 1)β conserve cette propri´et´e. La proposition pr´ec´edente appliqu´ee `a A∗ 0 ach`eve la preuve. (cid:4) 0, A∗ 1, A∗ 0, A∗ Remerciement: Je remercie chaleureusement F. Lust-Piquard pour ses conseils lors de la r´edaction de ce travail. References [A] W. Arveson: An invitation to C ∗-algebras, Graduate Texts in Maths 39, Springer (1976). [BL] [B] J. Bergh, J. Lofstrom: Interpolation spaces an introduction, Springer- Verlag-Berlin Heidelberg New York, (1976). J. Bergh: On the relation between the two complex methods of inter- polation, Indiana Univ. Math. J. 28, p. 775-777, (1979). [Da] M. Daher: Une remarque sur l'espace Aθ, C. R. Acad. Sci. Paris, t. 322, s´erie I, n 0 7, 641-644, (1996). [DGZ] R. Deville, G. Godefroy, V. Zizler: Smoothness and renorming in Banach spaces, Pitman, Monographs and Surveys 64, Longman Sci- entific, (1993). 11 [DU] J. Diestel, J. J. Uhl: Vector measures, Math. Surveys 15 A.M.S, (1977). [E] G. A. Edgar: Banach spaces with the analytic Radon-Nikodym prop- erty and abelian groups, Proc. Intern. Conf. On Almost Everywhere Convergence in Probability and Ergodic Theory (Columbus, Ohio), 195-213, (1989). [HP] U. Haagerup, G. Pisier: Factorization of analytic functions with val- ues in non commutative L1 spaces and applications, Can. J. Math. XLI n 05, 882-906, (1989). 12
1505.04731	1	1505	2015-05-18T17:40:03	Hypercyclic behavior of some non-convolution operators on $H(\mathbb{C}^N)$	[ "math.FA", "math.CV", "math.DS" ]	We study hypercyclicity properties of a family of non-convolution operators defined on spaces of holomorphic functions on $\mathbb{C}^N$. These operators are a composition of a differentiation operator and an affine composition operator, and are analogues of operators studied by Aron and Markose on $H(\mathbb{C})$. The hypercyclic behavior is more involved than in the one dimensional case, and depends on several parameters involved.	math.FA	math	HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS ON H(CN ) SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Abstract. We study hypercyclicity properties of a family of non-convolution operators defined on spaces of holomorphic functions on CN . These operators are a composition of a differentiation operator and an affine composition operator, and are analogues of operators studied by Aron and Markose on H(C). The hypercyclic behavior is more involved than in the one dimensional case, and depends on several parameters involved. 5 1 0 2 y a M 8 1 ] . A F h t a m [ 1 v 1 3 7 4 0 . 5 0 5 1 : v i X r a Introduction If T is a continuous linear operator acting on some topological vector space X, the orbit under T of a vector x ∈ X is the set Orb(x, T ) := {x, T x, T 2x, . . . }. The operator T is said to be hypercyclic if there exists some vector x ∈ X, called hypercyclic vector, whose orbit under T is dense in X. In the Fr´echet space setting, an operator T is hypercyclic if and only if it is topologically transitive, that is, if for every pair of non empty open sets U and V , there exists a integer n0 ∈ N such T n0U ∩ V 6= ∅. An operator is said to be mixing if T nU ∩ V 6= ∅ for every integer n ≥ n0. Recently, some stronger forms of hypercyclicity have gained the attention of researchers, specially the concepts of frequently hypercyclic operators and strongly mixing operators with respect to some invariant probability measure on the space. The first examples of hypercyclic operators were found by Birkhoff [5] and MacLane [12], whose research was focused in holomorphic functions of one complex variable and not in prop- erties of operators. Birkhoff's result implies that the translation operator τ : H(C) → H(C) defined by τ (h)(z) = h(1+z) is hypercyclic. Likewise, MacLane's result states that the differenti- ation operator on H(C) is hypercyclic. In a seminal paper, Godefroy and Shapiro [9] unified and generalized both results, by showing that every continuous linear operator T : H(CN ) → H(CN ) which commutes with translations and which is not a multiple of the identity is hypercyclic. This operators are called non-trivial convolution operators. Another important class of operators on H(CN ) are the composition operators Cφ, induced by symbols φ which are automorphisms of CN . The hypercyclicity of composition operators 2010 Mathematics Subject Classification. 47A16, 32Axx. Key words and phrases. non-convolution operators, differentiation operators, composition operators, hyper- cyclic operators, strongly mixing operators. Partially supported by PIP 2010-2012 GI 11220090100624, PICT 2011-1456, UBACyT 20020100100746, AN- PCyT PICT 11-0738 and CONICET. 1 2 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY induced by affine automorphisms was completely characterized in terms of properties of the symbol by Bernal-Gonz´alez [3]. Besides operators belonging to some of these two classes, there are not many examples of hypercyclic operators on H(CN ). Motivated by this fact, Aron and Markose [1] studied the hypercyclicity of the following operator on H(C), T f (z) = f ′(λz + b), with λ, b ∈ C. The operator T is not a convolution operator unless λ = 1. They showed that T is hypercyclic for any λ ≥ 1 (a gap in the proof was corrected in [8]) and that it is not hypercyclic if λ < 1 and b = 0. Thus, they gave explicit examples of hypercyclic operators which are neither convolution operators nor composition operators. Recently, this operators were studied in [11], where the authors showed that the operator is frequently hypercyclic when b = 0, λ ≥ 1 and asked whether it is frequently hypercyclic for any b. In Section 2, we give a different proof of the result of [1, 8], but for any λ, b ∈ C. We conclude in Proposition 2.3 that T is hypercyclic if and only if λ ≥ 1, and that in this case, T is even strongly mixing with respect to some Borel probability measure of full support on H(C). In Sections 3 and 4 we define N -dimensional analogues of the operators considered by Aron and Markose and study the dynamics they induce in H(CN ). These operators are a compo- sition between a partial differentiation operator and a composition operator induced by some automorphism of CN . It turns out that its behavior is more complicated than its one variable analogue. One possible reason is that, while the automorphisms of C have a very simple struc- ture and hypercyclicity properties, the automorphisms of CN are much more involved. Even, the characterization of hypercyclic affine automorphisms is nontrivial (see [3]). In Section 3, we consider the case in which the composition operators are induced by a diagonal operator plus a translation, that is, for f ∈ H(CN ) and z = (z1, . . . , zN ) ∈ CN , we study operators of the form T f (z) = Dαf ((λ1z1, . . . , λN zN ) + b), where α is a multi-index and b and λ = (λ1, . . . , λN ) are vectors in CN . In this case we completely characterize the hypercyclicity of these non-convolution operators which, contrary to the one dimensional case studied in [1], does not only depend on the size of λ. In the last section, we study the operators which are a composition of a directional differentiation operator with a general affine automorphism of CN and determine its hypercyclicity in some cases. 1. Preliminaries In this section we state some known conditions which ensure that a linear operator is strongly mixing with respect to an invariant Borel probability measure of full support. First we recall the following definitions. Definition 1.1. A linear operator T on X is called frequently hypercyclic if there exists a vector x ∈ X, called a frequently hypercyclic vector, whose T -orbit visit each non-empty open set along a set of integers having positive lower density. HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 3 Definition 1.2. A Borel probability measure on X is Gaussian if and only if it is the distribution 0 gnxn, where (xn) ⊂ X and of an almost surely convergent random series of the form ξ = P∞ (gn) is a sequence of independent, standard complex Gaussian variables. Definition 1.3. We say that an operator T ∈ L(X) is strongly mixing in the Gaussian sense if there exists some Gaussian T -invariant probability measure µ on X with full support such that any measurable sets A, B ⊂ X satisfy µ(A ∩ T −n(B)) = µ(A)µ(B). lim n→∞ We will use the following result, which is a corollary of a theorem due to Bayart and Math- eron (see [2]). Essentially this theorem says that a large supply of eigenvectors associated to unimodular eigenvalues that are well distributed along the unit circle implies that the operator is strongly mixing in the Gaussian sense. Theorem 1.4 (Bayart, Matheron). Let X be a complex separable Fr´echet space, and let T ∈ L(X). Assume that for any set D ⊂ T such that T \ D is dense in T, the linear span of Sλ∈T−D ker(T − λ) is dense in X. Then T is strongly mixing in the Gaussian sense. The following result, proved by Murillo-Arcila and Peris in [13, Theorem 1], shows that operators defined on Fr´echet spaces which satisfy the Frequent Hypercyclicity Criterion are strongly mixing with respect to an invariant Borel measure with full support. Theorem 1.5 (Murillo-Arcila, Peris). Let X be a separable Fr´echet space and T ∈ L(X). for all x ∈ X0. Suppose further that there exists a sequence of maps Sk : X0 → X such that Suppose that there exists a dense subset X0 ⊂ X such thatPn T nx is unconditionally convergent T ◦ S1 = Id, T ◦ Sk = Sk−1 and Pk Sk(x) is unconditionally convergent for all x ∈ X0. Then there exists a Borel probability measure µ in X, T -invariant, such that the operator T is strongly mixing respect to µ. It can be shown that the hypothesis of the Theorem 1.5 implies the corresponding ones of the Theorem 1.4. So, in any case, both theorems allow us to conclude the existence of an invariant Gaussian probability measure for linear operators of full support which are strongly mixing. Finally, the next proposition states that the existence of such measures is preserved by linear conjugation. It's proof is standard. Proposition 1.6. Let X and Y be separable Fr´echet spaces and T ∈ L(X), S ∈ L(Y ). Suppose that SJ = JT for some linear mapping J : X → Y of dense range then, if T has an invariant Borel measure then so does S. Moreover, if T has an invariant Borel measure that is Gaussian, strongly mixing, ergodic or of full support, then so does S. 4 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY 2. Non-Convolution operators on H(C) Let us denote by D and τa the derivation and translation operators on H(C), respectively. Namely, for an entire function f , we have D(f )(z) = f ′(z) and τa(f )(z) = f (z + a). MacLane's theorem [12] says that D is a hypercyclic operator, and Birkhoff's theorem [5] states that τa is hypercyclic provided that a 6= 0. The translation operators is a special class of composition operators on H(C). By a composition operator we mean an operator Cφ such that Cφ(f ) = f ◦ φ, where φ is some automorphism of C. The hypercyclicity of the composition operators on H(C) has been completely characterized in terms of properties of the symbol function φ. Precisely, the relevant property of φ is the following. Definition 2.1. A sequence {φn}n∈N of holomorphic maps on C, is called runaway if, for each compact set K ⊂ C, there is an integer n ∈ N such that φn(K) ∩ K = ∅. In the case where φn = φn for every n ∈ N, we will just say that φ is runaway. This definition was first given by Bernal Gonz´alez and Montes-Rodr´ıguez in [4], where they also proved the following (see also [10, Therorem 4.32]). Theorem 2.2. Let φ be an automorphism of C. Then Cφ is hypercyclic if and only if φ is runaway. It is known that the automorphisms of C are given by φ(z) = λz + b, with λ 6= 0 and b ∈ C. In addition, φ is runaway if and only if λ = 1 and b 6= 0 (see [10, Example 4.28]). This means that the hypercyclic composition operators on H(C) are exactly Birkhoff's translation operators. Aron and Markose in [1] studied the hypercyclicity of the following operator on H(C), T f (z) = f ′(λz + b), with λ, b ∈ C, which is a composition of MacLane's derivation operator and a composition operator, i.e., T = Cφ ◦ D with φ(z) = λz + b. The main motivation for the study of this operator was the wish to understand the behavior of a concrete operator belonging neither to the class of convolution operators nor to the class of composition operators. As mentioned before, in [1] (see also [8]) the authors proved that T is hypercyclic if λ ≥ 1, and that it is not hypercyclic if λ < 1 and b = 0. In this section we give a simple proof of the result by Aron and Markose, for the full range on λ, b. This will allow us to illustrate some of the main ideas used in the next section to prove the more involved N -variables case. Suppose that λ 6= 1. The key observation is that T is conjugate to an operator of the same type, but with b = 0. Indeed, define T0f (z) = f ′(λz), then we have that the following diagram HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 5 commutes. H(C) T H(C) τ [ b 1−λ ] τ [ b λ−1 ] H(C) / H(C) T0 Note that b 1−λ is the fixed point of φ. This observation will be important later. Proposition 2.3. Let T be the operator defined on H(C) by T f (z) = f ′(λz + b). Then T is hypercyclic if and only if λ ≥ 1. In this case, T is also strongly mixing with respect to some Borel probability measure of full support on H(C). Proof. If λ = 1, then T is a non-trivial convolution operator, thus it is hypercyclic. Moreover, by the Godefroy and Shapiro's theorem and its extensions (see [9, 6, 14]), T is strongly mixing in the Gaussian sense. Hence, by Proposition 1.6, it suffices to prove the case b = 0 and λ 6= 1, i.e. for the operator T0. Suppose first that λ < 1 and let f ∈ H(C). Note that T n 0 f (z) = λ n(n−1) 2 f (n)(λnz). By the Cauchy's estimates we obtain that T n 0 f (0) ≤ λ n(n−1) 2 n! sup kzk≤1 f (z) −→ n→∞ 0. Since the evaluation at 0 is continuous, the orbit of f under T0 can not be dense. Suppose now that λ > 1. Let us see that we can apply the Murillo-Arcila and Peris crite- rion, Theorem 1.5. Let X0 be the set of all polynomials, which is dense in H(C). Then, for 0 f is actually a finite sum, thus it is unconditionally each polynomial f ∈ X0, the series Pn T n convergent. For n ∈ N we define a sequence of linear maps Sn : X0 → X as Sn(zk) = k! zk+n (k + n)! λnk+ n(n−1) 2 . It is easy to see that Sn satisfy the hypothesis of Theorem 1.5. • T0 ◦ S1 = I : • T0 ◦ Sn = Sn−1 : k + 1 T0 ◦ S1(zk) = T0(cid:18) 1 T0 ◦ Sn(zk) = T0(cid:18) k! (k + n)! zk+1 λk (cid:19) = zk. 2 (cid:19) zk+n λnk+ n(n−1) λk+n−1zk+n−1 λnk+ n(n−1) zk+n−1 2 k! (k + n − 1)! k! = = = Sn−1(zk). (k + n − 1)! λ(n−1)k+ (n−1)(n−2) 2 / / / O O 6 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY • The series Pn Sn(f ) is unconditionally convergent for each f ∈ X0. If z ≤ R, we get that, Xn Sn(zk) ≤Xn k! (k + n)! Rk+n ≤ k!eR. Thus, the operator T0 is strongly mixing in the Gaussian sense. (cid:3) We can summarize the results of this section in the following table. It is worth noticing that nor the hypercyclicity of Cφ nor the hypercyclicity of D imply the hypercyclicity of Cφ ◦ D. Cφ D λ < 1 λ = 1 λ > 1 Not Hypercyclic Hypercyclic ⇔ b 6= 0 Not Hypercyclic Hypercyclic Cφ ◦ D Not Hypercyclic Hypercyclic Hypercyclic Hypercyclic Hypercyclic 3. Non-Convolution operators on H(CN ) - the diagonal case The operators considered in the previous section were differentiation operators followed by a composition operator. In this section we consider N -dimensional analogues of those operators. First, we will be concerned with symbols φ : CN → CN , which are diagonal affine automorphism of the form φ(z) = λz + b = (λ1z1 + b1, . . . , λN zN + bN ), where λ, b ∈ CN ; and the differentiation operator is a partial derivative operator given by a multi-index α = (α1 . . . , αN ) ∈ NN 0 , Dαf = ∂αf 2 . . . ∂zαN N ∂zα1 1 ∂zα2 . Thus in this section T will denote the operator on H(CN ) defined by T f (z) = Cφ ◦ Dα(f )(z) = Dαf (λ1z1 + b1, . . . , λN zN + bN ). Note that, in the definition of T , we allow α to be zero. In this case, the operator is just a composition operator and its hypercyclicity is determined by the symbol φ. These symbol functions are special cases of affine automorphisms of CN . The existence of universal functions for composition operators with affine symbol on CN has been completely characterized by Bernal- Gonzalez in [3], where he proved that the hypercyclicity of the composition operator depends on whether or not the symbol is runaway. Recall that an automorphism ϕ of CN is said to be runaway if for any compact subset K there is some n ≥ 1 such that ϕn(K) ∩ K = ∅. Theorem 3.1 (Bernal-Gonz´alez). Assume that ϕ : CN → CN is an affine automorphism of CN , say ϕ(z) = Az + b. Then, the composition operator Cϕ is hypercyclic if and only if ϕ is a runaway automorphism if and only if the vector b is not in ran(A − I) and det(A) 6= 0. The proof of this result is based on the following N -variables generalization of Runge's ap- proximation theorem, which will be useful for us later. HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 7 Theorem 3.2. If K and L are disjoint convex compact sets in CN and f is a holomorphic func- tion in a neighborhood of K ∪ L, then there is a sequence of polynomials on CN that approximate f uniformly on K ∪ L. Remark 3.3. It is easy to prove that the mapping φ(z) = (λ1z1 + b1, . . . , λN zN + bN ) is runaway if and only if some coordinate is a translation, that is, for some i = 1, . . . , N we have, simultaneously, that λi = 1 and bi 6= 0. If λj = 0 for some j, then we have that the differential d(T nf )(ej) = 1, for every n ∈ N. Since, the application d(·)(ej ) is continuous, we conclude that the orbit of f under T can not be dense. The next result completely characterizes the hypercyclicity of the operator T f = Cφ ◦ Dαf , with λ 6= 0 and α 6= 0 (the case α = 0 is covered in [3], and as mentioned above T is not hypercyclic if λj = 0 for some j). Write λα =Qi≤N λαi i . Theorem 3.4. Let T be the operator on H(CN ), defined by T f (z) = Cφ ◦ Dαf (z), where α 6= 0, φ(z) = (λ1z1 + b1, . . . , λN zN + bN ) and λi 6= 0 for all i, 1 ≤ i ≤ N . Then, a) If λα ≥ 1 then T is strongly mixing in the Gaussian sense. b) If for some i = 1, . . . , N we have that bi 6= 0 and λi = 1, then T is mixing. c) In any other case, T is not hypercyclic. Remark 3.5. The item c) above includes the following cases: c − i) λα < 1 and b = 0. c − ii) λα < 1 and λi 6= 1 for every i, 1 ≤ i ≤ N . c − iii) λα < 1 and bi = 0 for every i such that λi = 1. In all three cases we have that the application φ(z) = λz + b has a fixed point and thus φ is not runaway. Also in case b) the application φ has one coordinate which is a translation, thus it is runaway. So, in particular, Theorem 3.4 implies that T = Cφ ◦ Dα is hypercyclic if and only if either λα ≥ 1 or φ is runaway. We can summarize our main theorem in the following table. λα < 1 and λα < 1 and λα ≥ 1 no coord. of φ is a translation a coord. of φ is a translation Cφ Dα Cφ ◦ Dα Not Hypercyclic Hypercyclic Not Hypercyclic Hypercyclic Hypercyclic Hypercyclic depends on φ Hypercyclic Hypercyclic We will divide the proof of part (a) of Theorem 3.4 in two lemmas. Through a change in the order of the variables, we may suppose that the first j variables, 0 ≤ j ≤ N , correspond to the coordinates in which λi = 1. The operator T is then of the form (1) T f (z) = Dαf (z1 + b1, . . . , zj + bj, λj+1zj+1 + bj+1, . . . , λN zN + bN ). 8 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Moreover, we can assume that bi = 0 for all i > j, because T is topologically conjugate to (2) T0f (z) = Dαf (z1 + b1, . . . , zj + bj, λj+1zj+1, . . . , λN zN ). through a translation. Indeed, defining c ∈ CN by cl = 0 if l ≤ j, and cl = bl 1−λl that T0 ◦ τc = τc ◦ T . if l > j, we get We first study the case in which for some i, we have λi 6= 1 and αi 6= 0 (note that if all λi = 1, then T is a convolution operator and it is thus strongly mixing in the Gaussian sense [6, 14]). Lemma 3.6. Let T be as in (1). Suppose that λα ≥ 1 and αi 6= 0 for some i > j. Then T is strongly mixing in the Gaussian sense. Proof. By the above comments, we may suppose that bi = 0 for i > j, so the operator T is as in (2). We apply Theorem 1.5 with X0 = spanneγzβ := eγ1z1+···+γj zj zβ with βi = 0 for i ≤ j and γ ∈ Cjo . The set X0 ⊂ H(CN ) is dense. Indeed, since the set {eγ : γ ∈ Cj} generates a dense subspace in H(Cj) (see for example [6, Proposition 2.4]), given a monomial zθ1 j , ǫ > 0 and R > 0, there is f ∈ span{eγ : γ ∈ Cj} with 1 . . . zθj We obtain sup kzk≤R (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) sup kzk≤R(cid:12)(cid:12)(cid:12)f (z1, . . . , zj) − zθ1 1 . . . zθj j (cid:12)(cid:12)(cid:12) < ǫ. N (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (z1, . . . , zj )zβj+1 j+1 . . . zβN N −zθ1 1 . . . zθj j zβj+1 j+1 . . . zβN < ǫRβ. ∈X0 {z } Therefore we can approximate any monomial in H(CN ) by functions of X0 uniformly on com- pacts sets. The series Pn T n(eγzβ) is unconditionally convergent because the operator T differentiates in some variable zi with i > j, and so it is a finite sum. On the other hand, if we denote by α(1) := (α1, . . . , αj) and α(2) := (αj+1, . . . , αN ) 6= 0, we obtain T n(eγzβ) = γnα(1)enhγ,biλnβ− n(n+1) 2 α(2) β! (β − nα(2))! eγzβ−nα(2). Now, we define a sequence of maps Sn : X0 → X0. First, we do that on the set {eγzβ} and then extending them by linearity Sn(eγ zβ) = β! γnα(1)enhγ,biλnβ+ n(n−1) 2 α(2)(β + nα(2))! eγzβ+nα(2). The following assertions hold: HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 9 • T ◦ S1 = I : T ◦ S1(eγzβ) = 1 β! γα(1)ehγ,biλβ (β + α(2))! T (eγzβ+α(2)) = 1 β! γα(1)ehγ,biλβ (β + α(2))! γα(1)ehγ,bieγ (β + α(2))! β! zβλβ = eγzβ. • T ◦ Sn = Sn−1 : T ◦ Sn(eγzβ) = T (eγzβ+nα(2)) 1 β! γnα(1)enhγ,biλnβ+ n(n−1) β!γα(1) ehγ,biλβ+(n−1)α(2) (β + nα(2))! α(2) (β + nα(2))! 2 = = γnα(1)enhγ,biλnβ+ n(n−1) 2 α(2)(β + nα(2))!(β + (n − 1)α(2))! γ(n−1)α(1) e(n−1)hγ,biλ(n−1)β+ (n−1)(n−2) 2 α(2)(β + (n − 1)α(2))! β! eγzβ+(n−1)α(2) eγzβ+(n−1)α(2) = Sn−1(eγzβ). • Given R > 0, α(1) ehγ,bi . We have Sn(eγzβ) ≤ (β+nα(2))! for some constant M > 0 not depending on n. Since, α(2) 6= 0, we get let z ≤ R and denote C = λβ γ M Cn that for each γ ∈ Cj and β ∈ CN with βi = 0 for i ≤ j, Pn Sn(eγzβ) is uniformly convergent on compacts sets. α(2) R We have thus shown that the hypothesis of Theorem 1.5 are fulfilled. Hence T is strongly mixing in the Gaussian sense, as we wanted to prove. (cid:3) The other case we need to prove is when T does not differentiate in the variables zi with i > j. This means that αi = 0 for all i > j. To prove this case we will use Theorem 1.4. Lemma 3.7. Let T be as in (1). Suppose that λα ≥ 1 and αi = 0 for every i > j. Then T is strongly mixing in the Gaussian sense. Proof. We may suppose that bi = 0 for i > j, so the operator T is as in (2). The functions eγzβ, with γi = 0 for all i > j and βi = 0 for every i ≤ j, are eigenfunctions of T . Indeed, T (eγzβ) = γα(1) eP γi(zi+bi)(λz)β = γα(1)λβehγ,bieγzβ, where, as in the proof of the last lemma, α(1) = (α1, . . . , αj) 6= 0 (note that in this case α(2) = (αj+1, . . . , αN ) = 0). By Theorem 1.4 it is enough to show that for every set D ⊂ T such that T \ D is dense in T, the set (3) neγzβ; β ∈ CN with βi = 0 for i ≤ j and γi = 0 for i > j, such that γαλβehγ,bi ∈ T \ Do , spans a dense subspace on H(CN ). Fix β ∈ CN with βi = 0 for every i ≤ j and consider the map 10 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY fβ : Cj → C γ 7→ γαλβehγ,bi. The application fβ is holomorphic and non constant. So there exists γ0 ∈ Cj such that αλβehγ0,bi = 1. Since, T \ D is a dense set in T, the vector γ0 is an accumulation point of γ0 T \ D. Thus, by [6, Proposition 2.4], we get that the set neγ; with γ such that γαλβehγ,bi ∈ T \ Do , spans a dense subspace in H(Cj). It is then easy to see that the set defined in (3) spans a dense subspace in H(CN ). In particular, we have shown that the set of eigenvectors of T associated to eigenvalues belonging to T \ D span a dense subspace in H(CN ). So, the hypothesis of Theorem 1.4 are satisfied and hence T is strongly mixing in the Gaussian sense. (cid:3) The following remark will be useful for the next proof and in the rest of the article. Remark 3.8. Recall the Cauchy's formula for holomorphic functions in CN , Dαf (z1, . . . , zN ) = α! (2πi)N Zw1−z1=r1 . . .ZwN −zN =rN f (w1, . . . , wN ) i=1(wi − zi)αi+1 dw1 . . . dwN . QN Therefore, we can estimate the supremum of Dαf over a set of the form B(z1, r1) × · · · × B(zN , rN ), where B(zj, rj) denotes the closed disk of center zj ∈ C and radius rj. Fix positive real numbers ε1, . . . , εN , then (4) kDαf k∞,B(z1,r1)×···×B(zN ,rN ) ≤ α! kf k∞,B(z1,r1+ε1)×···×B(zN ,rN +εN ) (2π)N εα1+1 1 . . . εαN +1 N . Proof. (of Theorem 3.4) Part a) is proved by Lemmas 3.6 and 3.7. b) Suppose that bl 6= 0 for some l such that λl = 1. We will prove that T is a mixing operator, i.e., that for every pair U and V of non empty open sets for the local uniform topology of H(CN ), there exists n0 ∈ N such that T n(U ) ∩ V 6= ∅ for all n ≥ n0. Let f and g be two holomorphic functions on H(CN ), L be a compact set of CN and θ a positive real number. We can assume that U = {h ∈ H(CN ) : kf − hk∞,L < θ} and V = {h ∈ H(CN ) : kg − hk∞,L < θ}, and that g is a polynomial and that L is a closed ball of (CN , k · k∞). We do so because we can define a right inverse map over the set of polynomials. Since T = Cφ ◦ Dα, we can define I α(zβ) = β! (α + β)! zα+β. Thus, S = I α ◦ Cφ−1 is a right inverse for T when restricted to polynomials. Hence, we assume that L = B(0, r) × B(0, r) × · · · × B(0, r), for some r > 0 and denote φi(z) = λiz + bi, for z ∈ C. We get that φ(z1, . . . , zN ) = (φ1(z1) . . . , φN (zN )) and φi(B(zi, ri)) = B(φi(zi), λiri). Now, suppose that P is a polynomial in CN . Applying the inequality (4) several times, in which each time we use it we divide each εi by 2, we get that HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 11 kg − T nP k∞,L =(cid:13)(cid:13)Cφ ◦ Dα(Sg − T n−1P )(cid:13)(cid:13)∞,L =(cid:13)(cid:13)Dα(Sg − T n−1P )(cid:13)(cid:13)∞,φ(L) =(cid:13)(cid:13)Dα(Sg − T n−1P )(cid:13)(cid:13)∞,Q B(bi,λir) ≤ 1 α! (2π)N εα1+1 α! (2π)N εα1+1 α! (2π)N εα1+1 1 1 ≤ ≤ ≤ . . . εαN +1 N . . . εαN +1 N . . . εαN +1 N 2α+N α!2 (cid:13)(cid:13)Sg − T n−1P(cid:13)(cid:13)∞,Q B(bi,λir+εi) (cid:13)(cid:13)Cφ ◦ Dα(S2g − T n−2P )(cid:13)(cid:13)∞,Q B(bi,λir+εi) (cid:13)(cid:13)Dα(S2g − T n−2P )(cid:13)(cid:13)∞,Q B((λi+1)bi,λi(λir+εi)) (cid:13)(cid:13)S2g − T n−2P(cid:13)(cid:13)∞,Q B((λi+1)bi,λi(λir+εi)+ εi 2n−k−1(cid:19) . kSng − P k i (0),λinr+εiPn−1 ∞,Q B(cid:18)φn λik 2 ) k=0 (2π)2N ε2(α1+1) 1 . . . ε2(αN +1) N Thus following, we get that kg − T nP k∞,L ≤ 2(n(n+1)/2)(α+N )α!n (2π)nN εn(α1+1) 1 . . . εn(αN +1) N Let us denote by l, the coordinate of φ that is a translation in C. Thus, we have that λl = 1 and bl 6= 0. This implies that B φn l (0), λlnr + εl n−1 Xk=0 λlk 2n−k−1! = B nbl, r + εl 1 2k! ⊂ B (nbl, r + 2εl) . n−1 Xk=0 Fix n0 ∈ N, such that B(0, r) ∩ B (nbl, r + 2εl) = ∅ for all n ≥ n0. Now, take δn > 0 and Λn a ball of (CN , k · k∞), such that [L + δn] ∩ [Λn + δn] = ∅ for all n ≥ n0 and N Yi=1 B φn l (0), λlnr + εl λlk 2n−k−1! ⊂ Λn. n−1 Xk=0 Also, denote by Kn = 2(n(n+1)/2)(α+N)α!n n(αN +1) N χΛn+δnSng. We get a polynomial Pn such that n(α1+1) 1 (2π)nN ε ...ε . Then, use Theorem 3.2 with hn = χL+δnf + Hence, kf − PnkL < θ and kSng − PnkΛn < θ Kn . kf − PnkL < θ and kg − T nPnkL < θ. Thus, Pn ∈ U ∩ T −nV for all n ≥ n0 and T is a mixing operator as we wanted to prove. c) Let b 1−λ = ( b1 1−λ1 bj 1−λj bN 1−λN , . . . , = 0. Then b ) where, if bj = 0 and λj = 0 for some j = 1, . . . , N , we will 1−λ is a fixed point of φ, and thus understand that Applying the Cauchy estimates we obtain n(n−1) 2 αDnαf(cid:18) b 1 − λ(cid:19) . T nf(cid:18) b 1 − λ(cid:19) = λ (cid:12)(cid:12)(cid:12)(cid:12) 1 − λ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Dnαf(cid:18) b n(n−1) 2 1 − λ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) T nf(cid:18) b ≤ λα λα ≤ n(n−1) 2 rnα (nα)! sup kzk≤r f (z) −→ n→∞ 0. 12 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Since the evaluation at the vector f under T is not dense. b 1−λ is a continuous functional, this implies that the orbit of (cid:3) Notice that in case b) of Theorem 3.4 we do not know if the operator Cφ ◦ Dα is strongly mixing in the Gaussian sense or even frequently hypercyclic. If λi ≤ 1 for 1 ≤ i ≤ N , we are able to show that the operator is frequently hypercyclic. To achieve this we prove that Cφ ◦ Dα is Runge transitive. Definition 3.9. An operator T on a Fr´echet space X is called Runge transitive if there is an increasing sequence (pn) of seminorms defining the topology of X and numbers Nm ∈ N, Cm,n > 0 for m, n ∈ N such that: (1) for all m, n ∈ N and x ∈ X, pm(T nx) ≤ Cm,npn+Nm(x) (2) for all m, n ∈ N, x, y ∈ X and ε > 0 there is some z ∈ X such that pn(z − x) < ε and pm(T n+Nmz − y) < ε. The concept of Runge transitivity was introduced by Bonilla and Grosse-Erdmann, were they proved in [7, Theorem 3.3], that every Runge transitive operator on a Fr´echet space is frequently hypercyclic. They also show that every translation operator on H(C) is Runge transitive. However, the differentiation operator on H(C) is not Runge transitive, even though we know that it is strongly mixing in the Gaussian sense. Now, we prove that some of the operators which are included in the case b) are frequently hypercyclic. Proposition 3.10. Let T be the operator on H(CN ), defined by T f (z) = Cφ ◦ Dαf (z), with α 6= 0, φ(z) = (λ1z1 + b1, . . . , λN zN + bN ) and λi 6= 0 for all i, 1 ≤ i ≤ N . Then, if λi ≤ 1 for every i, 1 ≤ i ≤ N and we have that bj 6= 0 and λj = 1 for some j, 1 ≤ j ≤ N , then T is Runge transitive. Proof. Define the increasing sequence of seminorms QN where the radius ri(m) are defined as follows: i=1 B(0,ri(m)) pm(f ) = sup f (z), ri(m) =( bim if bi 6= 0 m if bi = 0 We will prove that both conditions of the Definition 3.9 are satisfied with Nm = m + 1. For the first condition, we proceed as in the proof of part c) of Theorem 3.4. We will apply several times the Cauchy inequalities (4) with εi defined as εi =( bi 2 1 2 if bi 6= 0 if bi = 0 HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 13 and in each step we divide it by 2. So, we get that pm(T nf ) ≤ 2(n(n+1)/2)(α+N )α!n (2π)nN εn(α1+1) 1 . . . εn(αN +1) N f (z), sup Λ where Λ =Q B(cid:16)φn Since λi ≤ 1 for every i, 1 ≤ i ≤ N , we obtain that k=0 i (0), λinri(m) + εiPn−1 i (0) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xk=0 λinri(m) + εi φn n−1 bi n−1 λik 2n−k−1(cid:17). i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xk=0 λk and that ≤ bin, λik 2n−k−1 ≤ ri(m) + 2εi. From here it is easy to prove that Λ ⊆ Q B(0, ri(n + m + 1)). Thus, if we denote Cm,n = 2(n(n+1)/2)(α+N)α!n , we get that (2π)nN ε n(α1+1) 1 ...ε n(αN +1) N pm(T nf ) ≤ Cm,npn+m+1(f ). Suppose that ε is a positive number, n and m are two integer numbers and that f , g are two holomorphic functions on H(CN ), we want to prove that there exists some function h ∈ H(CN ) such that pn(f − h) < ε and pm(T n+m+1h − g) < ε. Similarly, for the second condition we can estimate pm(T n+m+1h − g) in the same way we did previously by making use of the right inverse for T . We get that pm(T n+m+1h − g) ≤ C sup Γ Sn+m+1g − h To assure the existence of such function h, by Runge's Theorem 3.2, it is enough to prove where C is some positive constant and Γ =Q B(cid:16)φn+m (0), λin+m+1ri(m) + εiPn+m that Γ ∩Q B(0, ri(n)) = ∅. We study this sets in the j-th coordinate. We get that 2n−k−1(cid:17). λik k=0 i Γj = B(bj(n + m), rj(m) + 2εj ) = B(bj(n + m), bj(m + 1)), which is disjoint from B(0, bj n). Then, we have proved that the operator T is Runge transitive, hence it is frequently hypercyclic. (cid:3) 4. The non-diagonal case We are now interested in the case in which the automorphism φ(z) = Az + b, is given by any invertible matrix A ∈ CN ×N . Let v 6= 0 be any vector in CN and let T be the operator on H(CN ) defined by where Dvf is the differential operator in the direction of v, T f (z) = Cφ ◦ Dvf (z) = Dvf (Az + b), Dvf (z0) = lim s→0 f (z0 + sv) − f (z0) s = ∇f (z0) · v = df (φ(z0))(v). 14 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY The next two remarks show that we may consider a simplified version of the operator T . Remark 4.1. We can assume that the matrix A is given in its Jordan form. Indeed, let Q be an invertible matrix in CN ×N such that A = QJQ−1, where J is the Jordan form of A. Also let c = Q−1b and denote Q∗(f )(z) = f (Qz) for f ∈ H(CN ). Thus, we have that If we denote ψ(z) = Jz + c and w = Q−1v then, Q∗(Cφ ◦ Dvf )(z) = ∇f (AQz + b) · v. (Cψ ◦ Dw)Q∗(f )(z) = ∇f (Q(Jz + c)) · Qw = ∇f (AQz + b) · v. We have proved that the following diagram commutes Cφ◦Dv H(CN ) H(CN ) Q∗ Q∗ H(CN ) / H(CN ) Cψ◦Dw This shows that Cφ ◦ Dv is linearly conjugate to Cψ ◦ Dw. Remark 4.2. We can assume that b = 0 if the affine linear map φ has a fixed point z0 = φ(z0). Indeed, if we denote ϕ(z) = Az then, τz0(Cφ ◦ Dv)(f )(z) = Dv(f )(A(z + z0) + b) = τz0Dv(f )(Az) = (Cϕ ◦ Dv)τz0(f )(z). We have that the following diagram commutes Cφ◦Dv H(CN ) H(CN ) τz0 τz0 H(CN ) / H(CN ) Cϕ◦Dv We conclude that Cφ ◦ Dv is linearly conjugate to Cϕ ◦ Dv. The first two results of this section deal with affine transformations that have fixed points. Proposition 4.3. Let A ∈ CN ×N be an invertible matrix and let v be a nonzero vector in CN . Suppose that the affine linear map φ(z) = Az + b has a fixed point and that lim k→∞ k! kAivk < +∞. k−1 Yi=0 Then the operator Cφ ◦ Dv acting on H(CN ) is not hypercyclic. Consequently, Cφ ◦ Dv is not hypercyclic if v belongs to an invariant subspace M of A such that the spectral radius of the restriction, r(AM ), is less than 1. This happens in particular if r(A) < 1 or if v is an eigenvector of A associated to an eigenvalue of modulus strictly less than 1. / / / / / / HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 15 Proof. We denote by dkf (z) to the k-th differential of a function f at z, which is a k-homogenous polynomial, and we denote by (cid:0)dkf(cid:1)∨(z) to the associated symmetric k-linear form. It is not difficult to see that the orbits of the operator Cφ ◦ Dv are determined by Assume that z0 is a fixed point of φ, then applying the Cauchy's inequalities we get (Cφ ◦ Dv)kf (z) =(cid:0)dkf(cid:1)∨(φkz)(v, Av, . . . , Ak−1v). (Cφ ◦ Dv)kf (z0) = (cid:0)dkf(cid:1)∨(φkz0)(v, Av, . . . , Ak−1v) = (cid:0)dkf(cid:1)∨(z0)(v, Av, . . . , Ak−1v) k−1 ≤ k! kAivk sup f (z). z−z0<1 Yi=0 Therefore {(Cφ ◦ Dv)kf (z0)} is a bounded set of C. Since the evaluation at z0 is continuous, Cφ ◦ Dv cannot have dense orbits. For the last assertion, first note that if J = Q−1AQ is the Jordan form of A, we have that w = Q−1v belongs to the invariant subspace Q−1M of J and that r := r(JQ−1M ) < 1. By Remarks 4.1 and 4.2 it suffices to prove that CJ ◦ Dw is not hypercyclic. It is not difficult to show that for every i ≥ N , where c is a constant that depends only on r and N . Therefore, kJ iwk ≤ cri−N iN kwk, kJ iwk ≤ k! kJ iwk cri−N iN kwk k! k−1 Yi=0 N −1 Yi=0 k−1 Yi=N ≤ (k!)N +1kJk(N +1)N/2ck−N kwkkr(k−N )(k−N −1)/2 → 0, which implies that CJ ◦ Dw is not hypercyclic by the first part of the proposition. (cid:3) In opposition to the previous result, if the matrix A is expansive when restricted to an invariant subspace then the operator is strongly mixing in the Gaussian sense. This assumption is similar to the hypothesis of the results in the previous sections. Indeed, in the one dimensional case we have that φ(z) = λz + b and if λ ≥ 1, then the operator Cφ ◦ D is strongly mixing in the Gaussian sense. Here, the linear part of the composition operator is expansive. This situation still holds in the diagonal case in H(CN ). In this last case, we have that φ(z1, . . . , zN ) = (λ1z1 + b1, . . . , λN zN + bN ). Suppose that α is a multi-index of modulus one, i.e. that Dα is a partial derivative, then the hypothesis λα ≥ 1 turns out to be exactly the same as imposing that the linear part of φ is expansive on the subspace spanned by α. The proper result reads as follows. Proposition 4.4. Let A ∈ CN ×N be an invertible matrix and let v 6= 0 be a vector in CN . Suppose that the affine linear map φ(z) = Az + b has a fixed point and that v belongs to a subspace M that reduces A and such that k(AM )−1k < 1. Then the operator Cφ ◦ Dv acting on H(CN ) is strongly mixing in the Gaussian sense. 16 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Proof. We will show that the hypothesis of the Theorem 1.5 are fulfilled, taking as dense sets the polynomials in N complex variables. It is clear thatPn T nf converges unconditionally for every polynomial f . Now we will define a right inverse for Cφ ◦ Dv, but first we set some notation. Let us denote the fixed point of φ by z0. Let us denote by π1 to the orthogonal projection over M , π2 = I − π1 the orthogonal projection over M ⊥. Set µ(z) = hz,vi kvk2 . We have that z 7→ µ(z)v is the orthogonal projection over span{v}, and we denote π = π1 − µ(z)v. Finally, set φi(z) = Az + πi(b), for i = 1, 2. Since, M reduces A, we have that φi is invertible and that πi(z0) is a fixed point of φi, for i = 1, 2. We define now for each g ∈ H(CN ), Rg(z) =Z µ(z) µ(z0) g(φ−1 1 (tv + π(z)) + π2(z))dt, and C(g)(z) = g(π1(z) + φ−1 that, 2 (π2(z))). Note that R ◦ C = C ◦ R. Finally, let S = C ◦ R. Observe Sg(z) =Z µ(z) µ(z0) g(φ−1(tv + π(z) + π2(z)))dt. Sg(z + sv) − Sg(z) g(φ−1(tv + π(z) + π2(z)))dt −Z µ(z) µ(z0) g(φ−1(tv + π(z) + π2(z)))dt# g(φ−1(tv + π(z) + π2(z)))dt We have that DvSg(z) = lim s→0 = lim s→0 = lim s→0 1 1 s µ(z0) s"Z µ(z+sv) sZ µ(z)+s µ(z) = g(φ−1(µ(z)v + π(z) + π2(z))) = g(φ−1z). Thus, [Cφ ◦ Dv] ◦ Sg = g for every g ∈ H(CN ). To conclude the proof we need to show that Pn Sng converges unconditionally for every polynomial g. First we will bound the supremum of Rg on B(π1z0, r) × B(π2z0, s), for a fixed polynomial g. Suppose that z ∈ B(π1z0, r) × B(π2z0, s) and that t ∈ [µ(z0), µ(z)] i.e. t lives in the complex segment from µ(z0) to µ(z). Then we have that ktv + π(z) − π1z0k2 = k(t − µ(z0))v + π(z − z0)k2 = t − µ(z0)2kvk2 + kπ(z − z0)k2 ≤ µ(z) − µ(z0)2kvk2 + kπ(z − z0)k2+ = kπ1(z − z0)k2 < r2. HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 17 Also, suppose that σ := k(AM )−1k < 1. We get that kφ−1 1 (π1(z)) − π1(z0)k = kφ−1 1 (π1(z)) − φ−1 1 (π1(z0))k = kA−1(π1(z) − π1(b)) − A−1(π1(z0) − π1(b))k Gathering the previous statements we get that ≤(cid:13)(cid:13)(AM )−1(cid:13)(cid:13) kπ1(z) − π1(z0)k = σr. Rg(z) ≤ µ(z) − µ(z0) sup t∈[µ(z0),µ(z)] ≤ r kvk2 sup w∈B(π1z0,r)×B(π2z0,s) Thus, we have proved that g(φ−1 1 (tv + π(z)) + π2(z)) g(φ−1 1 (π1(w)) + π2(w)) ≤ r kvk2 sup g(w). w∈B(π1z0,σr)×B(π2z0,s) sup Rg ≤ B(π1z0,r)×B(π2z0,s) r kvk2 sup g. B(π1z0,σr)×B(π2z0,s) Following by induction we obtain that sup Rng ≤ B(π1z0,r)×B(π2z0,s) ≤ r kvk2 rn kvk2n σ sup Rn−1g B(π1z0,σr)×B(π2z0,s) n(n−1) 2 sup g. B(π1z0,σnr)×B(π2z0,s) Finally, to conclude the proof we compute supB(π1z0,r)×B(π2z0,s) Sng(z): sup Sng(z) = sup RnC ng(z) B(π1z0,r)×B(π2z0,s) B(π1z0,r)×B(π2z0,s) ≤ ≤ rn kvk2n σ rn kvk2n σ n(n−1) 2 n(n−1) 2 sup C ng(z) B(π1z0,σnr)×B(π2z0,s) sup B(π1z0,σnr)×φ−n 2 (B(π2z0,s)) g(z) Since σ < 1, we have proved that Pn Sng converges unconditionally for every polynomial g. Hence the operator Cφ ◦ Dv is strongly mixing in the Gaussian sense, as we wanted to prove. (cid:3) We turn now our discussion to the cases in which the affine linear map φ(z) = Az + b does not have a fixed point. This is equivalent to say that b /∈ Ran(I − A). Thus, 1 belongs to the spectrum of A. Then the Jordan form of A, which we denote by J, has a sub-block with ones in the principal diagonal and the first sub-diagonal and zeros elsewhere. It is easy to see that there exists some k ∈ N, k ≤ N such that the canonical vector ek does not belong to Ran(I − J) and such that bk 6= 0. This argument will be the key to show that φ is a runaway map, hence the operator Cφ ◦ Dv is topologically transitive. The proof of this result is in the spirit of part b) of Theorem 3.4. 18 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Proposition 4.5. Let A ∈ CN ×N be an invertible matrix and let v 6= 0 be a vector in CN . Suppose that the affine linear map φ(z) = Az + b does not have a fixed point. Then the operator Cφ ◦ Dv acting on H(CN ) is mixing. Proof. Due to the previous observations it is enough to prove that Cψ ◦ Dw is topologically transitive if ψ(z) = Jz + b with b /∈ Ran(I − J) and w ∈ CN , w 6= 0. We will denote T = Cψ ◦ Dw. Given KU , KV two compact sets of CN , hU , hV two holomorphic functions in H(CN ) and θ a positive real number, we want to prove that there exists k ∈ N and g ∈ H(CN ) such that (5) kg − hU kKU < θ and k(Cψ ◦ Dw)kg − hV kKV < θ. We will use Runge's theorem to show the existence of such function g. As before, we denote by S the right inverse of Dw. We have that Cψ ◦ Dwg(z) − hV (z) = sup sup KV KV (cid:12)(cid:12)Cψ(cid:0)Dwg(z) − Cψ−1hV (z)(cid:1)(cid:12)(cid:12) Cψ(KV )(cid:12)(cid:12)Dwg(z) − Cψ−1 hV (z))(cid:12)(cid:12) J(KV )+b(cid:12)(cid:12)Dw(cid:0)g(z) − S ◦ Cψ−1hV (z)(cid:1)(cid:12)(cid:12) kwkN = sup = sup ≤ εN 1 sup J(KV )+Bε1 (b)(cid:12)(cid:12)g(z) − S ◦ Cψ−1hV (z)(cid:12)(cid:12) . Following in this way inductively, we will get an estimate of k(Cψ ◦ Dw)kg − hV kKV , sup KV (cid:12)(cid:12)(cid:12)(Cψ ◦ Dw)lg(z) − hV (z)(cid:12)(cid:12)(cid:12) ≤ α(l) sup Al (cid:12)(cid:12)(cid:12)g(z) − (S ◦ Cψ−1)lhV (z)(cid:12)(cid:12)(cid:12) , with α(l) > 0 and Al = J l(KV ) +Pl i=1 J i(B(0, εi)) +Pl It is enough to find some l ∈ N such that KU ∩ Al = ∅. Without loss of generality we can assume that e1 /∈ Ran(J − I) and b1 6= 0 (see the comments before the proposition). This means that J acts like the identity in the first coordinate. i=1 J i(b). Suppose that KV ⊂QN εi > 0 we obtain i=1 B(0, ri), then if we project in the first coordinate and choose proper [Al]1 = [J l(KV )]1 + [J i(B(0, εi))]1 + [J i(b)]1 l Xi=1 l Xi=1 Xi=1 l ⊂ B(0, r1) + B(0, εi) + lb1 ⊂ B(0, R) + lb1. HYPERCYCLIC BEHAVIOR OF SOME NON-CONVOLUTION OPERATORS 19 Thus, we will able to find l0 ∈ N such that [KU ]1 ∩ [Al]1 = ∅ for all l ≥ l0. Therefore, by Runge's Theorem, there exists some gl ∈ H(CN ) such that (5) is satisfied for all l ≥ l0. We have proved that the operator Cψ ◦ Dw is mixing, as we wanted to prove. (cid:3) References [1] Richard M. Aron and Dinesh Markose. On universal functions. J. Korean Math. Soc., 41(1):65 -- 76, 2004. Satellite Conference on Infinite Dimensional Function Theory. [2] Fr´ed´eric Bayart and ´Etienne Matheron. Mixing operators and small subsets of the circle. Preprint, 2011. [3] Luis Bernal-Gonz´alez. Universal entire functions for affine endomorphisms of cn. Journal of mathematical analysis and applications, 305(2):690 -- 697, 2005. [4] Luis Bernal Gonz´alez and Alfonso Montes-Rodr´ıguez. Universal functions for composition operators. Complex Variables Theory Appl., 27(1):47 -- 56, 1995. [5] George D. Birkhoff. D´emonstration d'un th´eor`eme ´el´ementaire sur les fonctions enti`eres. C. R., 189:473 -- 475, 1929. [6] Antonio Bonilla and Karl-G. Grosse-Erdmann. On a theorem of Godefroy and Shapiro. Integral Equations Operator Theory, 56(2):151 -- 162, 2006. [7] Antonio Bonilla and Karl-G. Grosse-Erdmann. Frequently hypercyclic operators and vectors. Ergodic Theory Dynam. Systems, 27(2):383 -- 404, 2007. [8] Gustavo Fern´andez and Andr´e Arbex Hallack. Remarks on a result about hypercyclic non-convolution oper- ators. J. Math. Anal. Appl., 309(1):52 -- 55, 2005. [9] Gilles Godefroy and Joel H. Shapiro. Operators with dense, invariant, cyclic vector manifolds. J. Funct. Anal., 98(2):229 -- 269, 1991. [10] Karl-G. Grosse-Erdmann and Alfred Peris Manguillot. Linear chaos. Universitext. Berlin: Springer. xii, 386 p. EUR 53.45 , 2011. [11] Manjul Gupta and Aneesh Mundayadan. q-frequently hypercyclic operators. Banach J. Math. Anal., to appear. [12] Gerald R. MacLane. Sequences of derivatives and normal families. J. Analyse Math., 2:72 -- 87, 1952. [13] Marina Murillo-Arcila and Alfred Peris. Strong mixing measures for linear operators and frequent hyper- cyclicity. J. Math. Anal. Appl., 398(2):462 -- 465, 2013. [14] Santiago Muro, Dami´an Pinasco, and Mart´ın Savransky. Strongly mixing convolution operators on Fr´echet spaces of holomorphic functions. Integral Equations Operator Theory, to appear. Santiago Muro Departamento de Matem´atica - Pab I, Facultad de Cs. Exactas y Naturales, Universidad de Buenos Aires, (1428), Ciudad Aut´onoma de Buenos Aires, Argentina and CONICET E-mail address: [email protected] Dami´an Pinasco Departamento de Matem´aticas y Estad´ıstica, Universidad Torcuato Di Tella, Av. Figueroa Alcorta 7350, (1428), Ciudad Aut´onoma de Buenos Aires, Argentina and CONICET E-mail address: [email protected] 20 SANTIAGO MURO, DAMI ´AN PINASCO, MART´IN SAVRANSKY Mart´ın Savransky Departamento de Matem´atica - Pab I, Facultad de Cs. Exactas y Naturales, Universidad de Buenos Aires, (1428), Ciudad Aut´onoma de Buenos Aires, Argentina and CONICET E-mail address: [email protected]
1208.2573	2	1208	2012-10-26T15:07:21	Some Companions of Ostrowski type inequality for s-convex and s-concave functions with applications	[ "math.FA", "math.CA" ]	In this paper, we obtain some companions of Ostrowski type inequality for absolutely continuous functions whose second derivatives absolute value are s-convex and s-concave.	math.FA	math	SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY FOR s−CONVEX AND s−CONCAVE FUNCTIONS WITH APPLICATIONS. M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ Abstract. In this paper, we obtain some companions of Ostrowski type in- equality for absolutely continuous functions whose second derivatives absolute value are s−convex and s−concave. Finally, we gave some applications for special means. 1. introduction In [4], Hudzik and Maligranda considered among others the class of functions which are s−convex in the second sense. This class is defined in the following way: Definition 1. A function f : R+ → R, where R+ = [0, ∞), is said to be s−convex in the second sense if f (αx + βy) ≤ αsf (x) + βsf (y) for all x, y ∈ [0, ∞), α, β ≥ 0 with α + β = 1 and for some fixed s ∈ (0, 1]. The class of s−convex functions in the second sense is usually denoted by K 2 s . It can be easily seen that for s = 1, s−convexity reduces to ordinary convexity of functions defined on [0, ∞). In [3], Dragomir and Fitzpatrick proved a variant of Hadamard's inequality which holds for s−convex functions in the second sense: Theorem 1. Suppose that f : [0, ∞) → [0, ∞) is an s−convex function in the second sense, where s ∈ (0, 1) and let a, b ∈ [0, ∞), a < b. If f ∈ L1[a, b], then the following inequalities hold: (1.1) 2s−1f(cid:18) a + b 2 (cid:19) ≤ 1 b − aZ b a f (x)dx ≤ f (a) + f (b) s + 1 . The constant k = 1 above inequalities are sharp. s+1 is the best possible in the second inequality in (1.1). The The following inequality is well known as Ostrowski's inequality in the literature [2]: Key words and phrases. s−convex function, Ostrowski inequality, Holder inequality. ♦Corresponding Author. 1 2 M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ Theorem 2. Let f : I ⊂ R → R be a differentiable mapping on I ◦, the interior of the interval I, such that f ′ ∈ L[a, b], where a, b ∈ I with a < b. If f ′(x) ≤ M, then the following inequality, f (x) − 1 b − aZ b a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (x)dx(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) holds for all x ∈ [a, b]. The constant 1 not be replaced by a smaller constant. ≤ M (b − a)" 1 4 2 (cid:1)2 (b − a)2 # + (cid:0)x − a+b 4 is the best possible in the sense that it can In [6], Set et al. proved some inequalities for s−concave and concave functions via following Lemma: Lemma 1. Let f : I ⊆ R → R be a twice differentiable function on I ◦ with f ′′ ∈ L1[a, b], then f (u)du − f (x) +(cid:18)x − a 1 b − aZ b 2 (b − a)Z 1 (x − a)3 0 a + b 2 (cid:19) f ′(x) 2 (b − a)Z 1 (b − x)3 0 = t2f ′′ (tx + (1 − t) a) dt + t2f ′′ (tx + (1 − t) b) dt. Theorem 3. Let f : I ⊆ [0, ∞) → R be a twice differentiable function on I ◦ such that f ′′ ∈ L1[a, b] where a, b ∈ I with a < b. If f ′′q is s−concave in the second sense on [a, b] for some fixed s ∈ (0, 1], p, q > 1 and 1 q = 1, then the following inequality holds: p + 1 (1.2) ≤ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 b − aZ b a (s−1) 2 a + b f (u)du − f (x) +(cid:18)x − p (b − a) (x − a)3(cid:12)(cid:12)f ′′(cid:0) x+a 2 (cid:19) f ′(x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:1)(cid:12)(cid:12) + (b − x)3(cid:12)(cid:12)f ′′(cid:0) x+a 2 (cid:1)(cid:12)(cid:12) 2 1 q ! (2p + 1) for each x ∈ [a, b] . Corollary 1. If in (1.2), we choose x = a+b 2 , then we have For instance, if s = 1, then we have 1 b − aZ b a 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) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (u)du − f(cid:18) a + b 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (u)du − f(cid:18) a + b ≤ (s−1) q 2 (b − a)2 16 (2p + 1) ≤ (b − a)2 16 (2p + 1) (cid:21) . 1 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:20)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 3b f ′′(cid:18) 3a + b 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) p (cid:20)(cid:12)(cid:12)(cid:12)(cid:12) (cid:21) . f ′′(cid:18) a + 3b f ′′(cid:18) 3a + b 1 In [1], Liu introduced some companions of an Ostrowski type inequality for functions whose first derivative are absolutely continuous. In this paper, we established some companions of Ostrowski type inequality for absolutely continuous functions whose second derivatives absolute value are s−convex and s−concave which are reduce the results proved in [5], for s = 1. In order to prove our main results we need the following Lemma [1]: SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY 3 Lemma 2. Let f : [a, b] → R be such that the derivative f ′ is absolutely continuous on [a, b]. Then we have the inequality f (t)dt − 1 2 [f (x) + f (a + b − x)] 3a + b 4 (cid:19) [f ′(x) − f ′(a + b − x)] (t − a)2 f ′′(t)dt +Z a+b−x x (t − b)2 f ′′(t)dt# a 1 1 1 = + b − aZ b 2(cid:18)x − 2 (b − a)"Z x +Z b for all x ∈(cid:2)a, a+b 2 (cid:3) . a+b−x a 2. main results (cid:18)t − a + b 2 (cid:19)2 f ′′(t)dt Theorem 4. Let f : [a, b] → R be a function such that f ′ is absolutely continuous on [a, b], f ′′ ∈ L1[a, b]. If f ′′ is s−convex on [a, b], then we have the following inequality: (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a 1 b − aZ b 2(cid:18)x − 1 + f (t)dt − 1 2 [f (x) + f (a + b − x)] 3a + b 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) (x − a)3 [f ′′(a) + f ′′(b)] ≤ (s + 1) (s + 2) (s + 3) (b − a) + 4(cid:0)s2 + 3s + 2(cid:1) (x − a)3 +(cid:0)s2 + s + 2(cid:1) (a + b − 2x)3 8 (s + 1) (s + 2) (s + 3) (b − a) [f ′′(x) + f ′′(a + b − x)] Proof. Using Lemma 2 and the property of the modulus we have a 1 for all x ∈(cid:2)a, a+b 2 (cid:3) . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 2(cid:18)x − 2(b − a)"Z x +Z b a+b−x + ≤ 1 1 f (t)dt − 1 2 [f (x) + f (a + b − x)] 3a + b 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) (t − a)2 f ′′(t) dt +Z a+b−x (t − b)2 f ′′(t) dt# . x a (cid:18)t − a + b 2 (cid:19)2 f ′′(t) dt Since f ′′ is s−convex on [a, b], we have x − a(cid:19)s f ′′(t) ≤(cid:18) t − a x − a(cid:19)s f ′′(x) +(cid:18) x − t f ′′(a) , t ∈ [a, x]; 4 M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ f ′′(t) ≤(cid:18) t − x a + b − 2x(cid:19)s a + b − 2x (cid:19)s f ′′(a + b − x)+(cid:18) a + b − x − t f ′′(x) , t ∈ (x, a+b−x] and f ′′(t) ≤(cid:18) t − a − b + x x − a (cid:19)s x − a(cid:19)s f ′′(b)+(cid:18) b − t Therefore we can write f ′′(a + b − x) , t ∈ (a+ b − x, b]. 1 2 1 a 1 + 3a + b f (t)dt − [f (x) + f (a + b − x)] (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) 2(cid:18)x − x − a(cid:19)s 2(b − a)(cid:26)Z x (t − a)2(cid:20)(cid:18) t − a 2 (cid:19)2(cid:20)(cid:18) t − x a + b − 2x(cid:19)s +Z a+b−x (cid:18)t − (cid:19)s +Z b (t − b)2(cid:20)(cid:18) t − a − b + x 2(b − a)(cid:26) 1 (x − a)3 f ′′(x) + x − a a + b 1 x 1 a+b−x a (3 + s) s2 + s + 2 + + 4 (s + 1) (s + 2) (s + 3) 2 (s + 1) (s + 2) (s + 3) (x − a)3 (s + 1) (s + 2) (s + 3) (b − a) ≤ = = x − a(cid:19)s f ′′(x) +(cid:18) x − t f ′′(a)(cid:21) dt a + b − 2x (cid:19)s f ′′(a + b − x) +(cid:18) a + b − x − t f ′′(a + b − x)(cid:21) dt) x − a(cid:19)s f ′′(b) +(cid:18) b − t f ′′(x)(cid:21) dt 2 (s + 1) (s + 2) (s + 3) (x − a)3 f ′′(a) s2 + s + 2 (a + b − 2x)3 f ′′(a + b − x) + (a + b − 2x)3 f ′′(x) 4 (s + 1) (s + 2) (s + 3) (x − a)3 f ′′(b) + 1 (3 + s) (x − a)3 f ′′(a + b − x)(cid:27) [f ′′(a) + f ′′(b)] + 4(cid:0)s2 + 3s + 2(cid:1) (x − a)3 +(cid:0)s2 + s + 2(cid:1) (a + b − 2x)3 8 (s + 1) (s + 2) (s + 3) (b − a) [f ′′(x) + f ′′(a + b − x)] , which is the desired result. (cid:3) Corollary 2. Let f as in Theorem 4. Additionally, if f ′(x) = f ′(a + b − x), we have 1 b − aZ b a 1 2 f (t)dt − (x − a)3 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) [f (x) + f (a + b − x)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12) [f ′′(a) + f ′′(b)] (s + 1) (s + 2) (s + 3) (b − a) ≤ + 4(cid:0)s2 + 3s + 2(cid:1) (x − a)3 +(cid:0)s2 + s + 2(cid:1) (a + b − 2x)3 8 (s + 1) (s + 2) (s + 3) (b − a) [f ′′(x) + f ′′(a + b − x)] . SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY 5 Corollary 3. In Corollary 2, if f is symmetric function, f (a + b − x) = f (x), for all x ∈(cid:2)a, a+b b − aZ b 1 2 (cid:3) we have f (t)dt − f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (x − a)3 a (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ (s + 1) (s + 2) (s + 3) (b − a) [f ′′(a) + f ′′(b)] + 4(cid:0)s2 + 3s + 2(cid:1) (x − a)3 +(cid:0)s2 + s + 2(cid:1) (a + b − 2x)3 8 (s + 1) (s + 2) (s + 3) (b − a) [f ′′(x) + f ′′(a + b − x)] , which is an Ostrowski type inequality. Corollary 4. In Theorem 4, if we choose (1) x = a+b 2 , we have ≤ (2) x = 3a+b a 4 1 , we have (b − a)2 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (t)dt − f(cid:18) a + b (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 8 (s + 1) (s + 2) (s + 3)(cid:20)f ′′(a) +(cid:0)s2 + 3s + 2(cid:1)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:21)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 4 (cid:19) + f(cid:18) a + 3b 128 (s + 1) (s + 2) (s + 3)(cid:20)2 f ′′(a) + (3s2 + 5s + 6)(cid:20)(cid:12)(cid:12)(cid:12)(cid:12) 2(cid:20)f(cid:18) 3a + b (b − a)2 f (t)dt − 1 1 a 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b + f ′′(b)(cid:21) . ≤ 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) 3a + b f ′′(cid:18) a + 3b on [a, b], f ′′ ∈ L1[a, b]. If f ′′q is s−convex on [a, b], for all x ∈(cid:2)a, a+b 2 (cid:3) and q > 1, Theorem 5. Let f : [a, b] → R be a function such that f ′ is absolutely continuous then we have the following inequality: (cid:21) + 2 f ′′(b)(cid:21) . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a 1 b − aZ b 2(cid:18)x − 1 + ≤ f (t)dt − 1 2 [f (x) + f (a + b − x)] 3a + b 1 1 1 p (s + 1) 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) (cid:0)f ′′(x)q + f ′′(a + b − x)q(cid:1) qi , 1 1 q + (x − a)3(cid:0)f ′′(a + b − x)q + f ′′(b)q(cid:1) 2 (b − a) (2p + 1) (a + b − 2x)3 + 4 q h(x − a)3(cid:0)f ′′(a)q + f ′′(x)q(cid:1) 1 q where 1 p + 1 q = 1. 6 M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ Proof. Using Lemma 2, Holder inequality and s−convexity of f ′′q , we have ≤ ≤ = 1 1 1 q 1 2 a+b−x 1 q 1 1 a a + b 1 1 x x a a a+b−x 1 1 + a 1 3a + b f (t)dt − f ′′(t)q f ′′(t)q f ′′(t)q dt(cid:19) dt! [f (x) + f (a + b − x)] (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) 2(cid:18)x − 2 (b − a)((cid:18)Z x p (cid:18)Z x (t − a)2p dt(cid:19) + Z a+b−x p Z a+b−x dt! 2 (cid:19)2p (cid:18)t − q + Z b p Z b (t − b)2p dt! dt!  2 (b − a)((cid:18)Z x x − a(cid:19)s x − a(cid:19)s p (cid:18)Z x f ′′(x)q +(cid:18) x − t (t − a)2p dt(cid:19) a (cid:20)(cid:18) t − a p Z a+b−x dt! + Z a+b−x 2 (cid:19)2p a + b − 2x(cid:19)s a + b − 2x (cid:19)s (cid:20)(cid:18) t − x f ′′(a + b − x)q +(cid:18) a + b − x − t (cid:18)t − p Z b + Z b (t − b)2p dt! f ′′(a + b − x)q(cid:21) dt! (cid:19)s a+b−x(cid:20)(cid:18) t − a − b + x 2 (b − a) (2p + 1) ! (x − a)2p+1 p (cid:18) x − a s + 1(cid:19) (cid:0)f ′′(a)q + f ′′(x)q(cid:1)  + 2 − x(cid:19)2p+1! p (cid:18) a + b − 2x 2p + 1(cid:18) a + b s + 1 (cid:19) + (x − a)2p+1 (2p + 1) ! p (cid:18) x − a s + 1(cid:19) (cid:0)f ′′(a + b − x)q + f ′′(b)q(cid:1) (cid:0)f ′′(x)q + f ′′(a + b − x)q(cid:1) x − a(cid:19)s f ′′(b)q +(cid:18) b − t f ′′(a)q(cid:21) dt(cid:19) 2 1 1 1 x − a 1 1 q x x a + b 1 1 q a+b−x 1 1 q 1 q 1 q 1 q . 1 q  When we arrange the statements above, we obtain the desired result. (cid:3) Corollary 5. Let f as in Theorem 5. Additionally, if f ′(x) = f ′(a + b − x), we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1 b − aZ b a f (t)dt − 1 2 (b − a) (2p + 1) + (a + b − 2x)3 4 1 1 1 2 p (s + 1) [f (x) + f (a + b − x)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q h(x − a)3(cid:0)f ′′(a)q + f ′′(x)q(cid:1) (cid:0)f ′′(x)q + f ′′(a + b − x)q(cid:1) qi , 1 q 1 1 q + (x − a)3(cid:0)f ′′(a + b − x)q + f ′′(b)q(cid:1) 1 f ′′(x)q(cid:21) dt! q  SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY 7 Corollary 6. In Corollary 5, if f is symmetric function, f (a + b − x) = f (x), we have for all x ∈(cid:2)a, a+b 2 (cid:3) . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1 b − aZ b a 1 1 1 p (s + 1) f (t)dt − f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:0)f ′′(x)q + f ′′(a + b − x)q(cid:1) qi , 1 1 q q h(x − a)3(cid:0)f ′′(a)q + f ′′(x)q(cid:1) 1 q 2 (b − a) (2p + 1) (a + b − 2x)3 + 4 1 (1) x = a+b 2 , we have b − aZ b Corollary 7. In Theorem 5, if we choose + (x − a)3(cid:0)f ′′(a + b − x)q + f ′′(b)q(cid:1) for all x ∈(cid:2)a, a+b 2 (cid:3) . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:18)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (t)dt − f(cid:18) a + b p 16"(cid:18)f ′′(a)q +(cid:12)(cid:12)(cid:12)(cid:12) q# . + f ′′(b)q(cid:19) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b (2) x = 3a+b 1 q (2p + 1) (b − a)2 , we have (s + 1) ≤ 4 a q 1 1 ≤ 1 p (s + 1) a 1 f (t)dt − (b − a)2 128 (2p + 1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +2(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) 3a + b +(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 3b (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (a) + f (b) 2 q 1 b − aZ b a f (t)dt − (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 q q(cid:19) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b 1 1 4 (cid:19)(cid:21)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19) + f(cid:18) a + 3b 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) 3a + b q(cid:19) 2(cid:20)f(cid:18) 3a + b q ((cid:18)f ′′(a)q +(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) 4 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 3b q) . + f ′′(b)q(cid:19) 1 q q 1 1 q q(cid:19) Corollary 8. In Corollary 5, if we choose x = a we have ≤ (b − a)2 8 (2p + 1) 1 p (s + 1) 1 q (cid:2)f ′′(a)q + f ′′(b)q(cid:3) 1 q . Remark 1. Using the well-known power-mean integral inequality one may get in- equalities for functions whose second derivatives absolute value are convex. The details are omitted. We obtain the following result for s−concave functions. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a 1 b − aZ b 2(cid:18)x − 1 + s−1 q 2 8 M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ Theorem 6. Let f : [a, b] → R be a function such that f ′ is absolutely continuous on [a, b], f ′′ ∈ L1[a, b]. If f ′′q is s−concave on [a, b], for all x ∈(cid:2)a, a+b then we have the following inequality: 2 (cid:3) and q > 1, f (t)dt − 1 2 [f (x) + f (a + b − x)] ≤ 2 (b − a) (2p + 1) (a + b − 2x)3 + 4 3a + b 1 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) p (cid:20)(x − a)3(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) x + a + (x − a)3(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 2b − x f ′′(cid:18) a + b 2 # (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) where 1 p + 1 q = 1. Proof. From Lemma 2 and using Holder inequality, we have ≤ = 1 a x 1 q 1 x x 1 1 a a 1 1 1 2 a 1 a+b−x a+b−x a + b 1 + 3a + b f (t)dt − f ′′(t)q f ′′(t)q f ′′(t)q dt(cid:19) [f (x) + f (a + b − x)] (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) 2(cid:18)x − 2 (b − a)((cid:18)Z x p (cid:18)Z x (t − a)2p dt(cid:19) + Z a+b−x p Z a+b−x dt! 2 (cid:19)2p (cid:18)t − + Z b p Z b (t − b)2p dt! 2 (b − a)((cid:18)Z x p (cid:18)Z x a (cid:12)(cid:12)(cid:12)(cid:12) (t − a)2p dt(cid:19) dt! + Z a+b−x p Z a+b−x 2 (cid:19)2p (cid:18)t − + Z b p Z b (t − b)2p dt! a+b−x(cid:12)(cid:12)(cid:12)(cid:12) dt ≤ 2s−1 (x − a)(cid:12)(cid:12)(cid:12)(cid:12) a(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) x(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) q dt!  f ′′(cid:18) t − a (cid:12)(cid:12)(cid:12)(cid:12) Z x a (cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) t − a f ′′(cid:18) t − x (a + b − x) + a + b − x − t x − t x − a a + b − 2x a + b − 2x x − a x − a a + b a+b−x (cid:12)(cid:12)(cid:12)(cid:12) x + x + 1 1 1 q Since f ′′q is s−concave, from (1.1), we have Z a+b−x x 1 q dt! x − t x − a q 1 q dt(cid:19) a(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) a + b − 2x x f ′′(cid:18) t − x f ′′(cid:18) t − a − b + x x − a b + (a + b − x) + a + b − x − t a + b − 2x b − t x − a q dt! (a + b − x)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 1 q  q 1 q dt! x(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) . q q , 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + x dt ≤ 2s−1 (a + b − 2x)(cid:12)(cid:12)(cid:12)(cid:12) q 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY 9 and f ′′(cid:18) t − a − b + x x − a b + b − t x − a Z b a+b−x(cid:12)(cid:12)(cid:12)(cid:12) q (a + b − x)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) dt ≤ 2s−1 (x − a)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 2b − x 2 . q (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Combining all above inequalities, we obtain 1 2 [f (x) + f (a + b − x)] a 1 1 1 + f (t)dt − 3a + b (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b 2(cid:18)x − 2 (b − a)  + 2 2p + 1(cid:18) a + b + (x − a)2p+1 (2p + 1) ! 2 s−1 q 2 2 (b − a) (2p + 1) 1 (a + b − 2x)3 + 4 q 1 q 1 p 1 1 p 2 2 s−1 s−1 (x − a) (a + b − 2x) 4 (cid:19) [f ′(x) − f ′(a + b − x)](cid:12)(cid:12)(cid:12)(cid:12) (2p + 1) ! (x − a)2p+1 q (cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) x + a − x(cid:19)2p+1! 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) q (cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b  (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) q (cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 2b − x  2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) x + a + (x − a)3(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 2b − x p (cid:20)(x − a)3(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b # (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) (x − a) (cid:12)(cid:12)(cid:12)(cid:12) s−1 2 2 2 1 p 1 q ≤ ≤ for all x ∈(cid:2)a, a+b 2 (cid:3) and 1 p + 1 q = 1. (cid:3) Corollary 9. Let f as in Theorem 6. Additionally, if f ′(x) = f ′(a + b − x), we have f (t)dt − 1 b − aZ b a s−1 q 2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2 (b − a) (2p + 1) + (a + b − 2x)3 4 for all x ∈(cid:2)a, a+b 2 (cid:3) . (cid:12)(cid:12)(cid:12)(cid:12) 1 1 2 [f (x) + f (a + b − x)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:20)(x − a)3(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) x + a # , 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b f ′′(cid:18) a + 2b − x 2 +(cid:12)(cid:12)(cid:12)(cid:12) (cid:19) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 10 M.EMIN OZDEMIR(cid:7) AND MERVE AVCI ARDIC⋆♦ Corollary 10. In Corollary 9, if f is symmetric function, f (a + b − x) = f (x), we have 1 b − aZ b a s−1 q 2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2 (b − a) (2p + 1) (a + b − 2x)3 + 4 1 f (t)dt − f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:20)(x − a)3(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) x + a # , (cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + b f ′′(cid:18) a + 2b − x 2 +(cid:12)(cid:12)(cid:12)(cid:12) (cid:19) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Corollary 11. In Theorem 6, if we choose x = 3a+b 4 we have for all x ∈(cid:2)a, a+b 2 (cid:3) . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) b − aZ b s−1 1 2 ≤ a q (b − a)2 128 (2p + 1) f (t)dt − 1 4 (cid:19)(cid:21)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2(cid:20)f(cid:18) 3a + b 4 (cid:19) + f(cid:18) a + 3b p (cid:20)(cid:12)(cid:12)(cid:12)(cid:12) 8 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 2 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) + 2(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) 7a + b f ′′(cid:18) a + b 1 8 (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) f ′′(cid:18) a + 7b (cid:21) . Remark 2. IIn Theorem 6, if we choose x = a+b Corollary 1. 2 , we have the first inequality in 3. applications for special means We consider the means for nonnegative real numbers α < β as follows: (1) The arithmetic mean: A (α, β) = α + β 2 , α, β ∈ R+. (2) The generalized logarithmic mean: Ln (α, β) =(cid:20) βn+1 − αn+1 (β − α) (n + 1)(cid:21) 1 n , α, β ∈ R+, α 6= β, n ∈ Z\ {−1, 0} . Proposition 1. Let 0 < a < b and 0 < s < 1. Then we have Ls s (a, b) − As (a, b) ≤ (b − a)2 s (s − 1) and ≤ Ls 8 (s + 1) (s + 2) (s + 3)(cid:2)as−2 +(cid:0)s2 + 3s + 2(cid:1) As−2 (a, b) + bs−2(cid:3) 4 (cid:19)s (cid:12)(cid:12)(cid:12)(cid:12) s (a, b) − A(cid:18)(cid:18) 3a + b 8 (s + 1) (s + 2) (s + 3)"2(cid:0)as−2 + bs−2(cid:1) +(cid:0)3s2 + 5s + 6(cid:1) (cid:18) 3a + b 4 (cid:19)s(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ,(cid:18) a + 3b 4 (cid:19)s−2 (b − a)2 s (s − 1) Proof. The assertion follows from Corollary 4 applied to the s−convex mapping f : [0, 1] → [0, 1], f (x) = xs. (cid:3) +(cid:18) a + 3b 4 (cid:19)s−2!# . SOME COMPANIONS OF OSTROWSKI TYPE INEQUALITY 11 Proposition 2. Let 0 < a < b and 0 < s < 1. Then for all q > 1, we have Ls s (a, b) − As (a, b) ≤ (b − a)2 s (s − 1) 16 (s + 1) 1 q (2p + 1) 1 p (cid:20)(cid:16)a(s−2)q + A(s−2)q (a, b)(cid:17) 1 q where 1 p + 1 q = 1. +(cid:18)(cid:16)A(s−2)q (a, b) + b(s−2)q(cid:17) 1 q(cid:19)(cid:21) Proof. The assertion follows from first inequality in Corollary 7 applied to the s−convex mapping f : [0, 1] → [0, 1], f (x) = xs. (cid:3) Proposition 3. Let 0 < a < b and 0 < s < 1. Then for all q > 1, we have Ls s (a, b) − A (as, bs) ≤ where 1 p + 1 q = 1. (b − a)2 s (s − 1) 8 (s + 1) 1 q (2p + 1) 1 p ha(s−2)q + b(s−2)qi 1 q , Proof. The assertion follows from first inequality in Corollary 8 applied to the s−convex mapping f : [0, 1] → [0, 1], f (x) = xs. (cid:3) References [1] Zheng Liu, Some companions of an Ostrowski type inequality and applications, JIPAM, vol. 10, iss. 2, art. 52, 2009. [2] A. Ostrowski: Uber die Absolutabweichung einer di erentierbaren Funktionen von ihren Inte- gralmittelwort. Comment. Math. Helv. 10 (1938), 226-227. [3] S.S. Dragomir and S. Fitzpatrick, The Hadamard's inequality for s−convex functions in the second sense, Demonstratio Math., 32 (4) (1999), 687-696. [4] H. Hudzik, L. Maligranda, Some remarks on s-convex functions, Aequationes Math. 48 (1994) 100 -- 111. [5] M. E. Ozdemir and M. Avci, Some Companions of Ostrowski type inequality for functions whose second derivatives are convex and concave , arXiv: 1207.6577. [6] E. Set, M. Z. Sarikaya and M. E. Ozdemir, Some Ostrowski's type inequalities for func- tions whose second derivatives are s−convex in the second sense and applications, arXiv: 1006.2488v1. (cid:7)Ataturk University, K.K. Education Faculty, Department of Mathematics, Erzurum 25240, Turkey E-mail address: [email protected] ⋆Adiyaman University, Faculty of Science and Arts, Department of Mathematics, Adiyaman 02040, Turkey E-mail address: [email protected]
1005.4318	1	1005	2010-05-24T12:20:02	A Solution of variational inequality problem for a finite family of nonexpansive mappings in Hilbert spaces	[ "math.FA", "math.NA", "math.OC" ]	In this paper we prove the strong convergence of the explicit iterative process to a common fixed point of the finite family of nonexpansive mappings defined on Hilbert space, which solves the the variational inequality on the fixed points set.	math.FA	math	A SOLUTION OF VARIATIONAL INEQUALITY PROBLEM FOR A FINITE FAMILY OF NONEXPANSIVE MAPPINGS IN HILBERT SPACES FARRUKH MUKHAMEDOV AND MANSOOR SABUROV Abstract. In this paper we prove the strong convergence of the explicit iterative process to a common fixed point of the finite family of nonexpansive mappings defined on Hilbert space, which solves the the variational inequality on the fixed points set. Mathematics Subject Classification: 46B20; 47H09; 47H10 Key words: Nonexpansive mappings, explicit iteration process; common fixed point; variational inequality. 1. Introduction An iterative approximation of fixed points and zeros of nonlinear operators has been studied extensively by many authors to solve nonlinear operator equations as well as variational in- equality problems (see e.g., [3, 4], [7], [9]-[15], [17]-[23]). A very important class of mappings is nonexpansive mappings. In particulary, iterative approximation of fixed points of nonexpansive mappings is an important subject in nonlinear functional analysis, and has many applications in various fields of mathematics, physics etc (see e.g., [2, 16, 28]). Let H be a real Hilbert space with inner product h·, ·i, and induced norm k·k. Let T : H → H be a mapping. Denote by F (T ) the set of fixed points of T , that is, F (T ) = {x ∈ H : T x = x}. Now let us recall some well-known definitions, in which will be used in this paper. A mapping f : H → H is said to be contraction if there exists α ∈ [0, 1) such that for all x, y ∈ H, A mapping T : H → H is called nonexpansive if for all x, y ∈ H, kf (x) − f (y)k ≤ αkx − yk. kT x − T yk ≤ kx − yk. A mapping A : H → H is called η−strongly monotone if for all x, y ∈ H, A mapping A : H → H is called k−Lipschitzian if for all x, y ∈ H, Dx − y, Ax − AyE ≥ ηkx − yk2. kAx − Ayk ≤ kkx − yk. 0 1 0 2 y a M 4 2 ] . A F h t a m [ 1 v 8 1 3 4 . 5 0 0 1 : v i X r a Iterative methods for nonexpansive mappings are now also applicable in solving convex min- imization problems (see, for example, [26] and references therein). Let K be a closed convex nonempty subset of H and T : K → K be a nonexpansive mapping such that F (T ) 6= ∅. Given u ∈ K and a real sequence {αn}∞ n=1 in the interval (0, 1), starting with an arbitrary initial x0 ∈ K, let a sequence {xn}∞ n=1 be defined by xn+1 = αnu + (1 − αn)T xn, n ∈ N. (1.1) Under appropriate conditions on the iterative parameter {αn}, it has been shown by Halpern [6], Lions [8], Wittmann [25] and Bauschke [1] that {xn}∞ n=1 converges strongly to PF (T )u, the projection of u to the fixed point set, F (T ) of T. This means that the limit of the sequence {xn}∞ n=1 solves the following minimization problem: 1 2 FARRUKH MUKHAMEDOV AND MANSOOR SABUROV find x∗ ∈ F (T ) such that kx∗ − uk = min x∈F (T ) kx − uk. A typical minimization problem is to minimize a quadratic function over the set of fixed points of nonexpansive mappings in a real Hilbert space and such problems go thus: find x∗ ∈ F (T ) such that 1 2DAx∗, x∗E −Dx∗, uE = min x∈F (T )(cid:18)1 2DAx, xE −Dx, uE(cid:19) (1.2) where u ∈ H is a fix point and A is a mapping on H which could be monotone, strongly monotone or even bounded linear strongly positive operator as the case may be. Let K1, K2, · · · , KN be N closed convex subsets of a real Hilbert space H having a nonempty intersection K. Suppose also that each Ci is a fixed point set of nonexpansive mappings Ti : H → H, i = 1, N . Xu [26] proved strong convergence of the iterative algorithm xn+1 = αnu + (I − αnA)Tnxn, n ∈ N, (1.3) where, I is an identity mapping and Tn := Tn(modN ), to a unique solution of the quadratic minimization problem where A is a bounded linear strongly positive operator on H and u is a given point in H. min x∈K (cid:18)1 2DAx, xE −Dx, uE(cid:19) , Marino and Xu [10], proved that the iteration scheme given by xn+1 = αnγf (xn) + (I − αnA)T xn, n ∈ N, converges strongly to a unique solution x′ ∈ F (T ) of the variational inequality which is the optimality condition for the minimization problem D(γf − A)x′, y − x′E ≤ 0, ∀y ∈ F (T ), (1.4) (1.5) x∈F (T )(cid:18)1 min 2DAx, xE − h(x)(cid:19) , where h is a potential function for γf (that is, h′(x) = γf (x) for all x ∈ H); provided f : H → H is a contraction, T : H → H is nonexpansive and the iterative parameter {αn} satisfies appropriate conditions. In [27], Yamada introduced the following hybrid iterative method xn+1 = (I − αnµA)T xn, n ∈ N, (1.6) where T is nonexpansive, A is L−Lipschitzian and strongly monotone operator with constant η and 0 < µ < 2η n=1 converges strongly to a unique solution x′ ∈ F (T ) of the variational inequality L2 . He proved that if {αn} satisfies appropriate conditions, then {xn}∞ DAx′, y − x′E ≥ 0, ∀y ∈ F (T ). Recently, motivated by the iteration schemes (1.4) and (1.6), M. Tian [24] introduced the following iterative method: xn+1 = αnγf (xn) + (I − αnµA)T xn, n ∈ N. (1.7) Tian [24] proved that if f : H → H is a contraction, A : H → H is an η−strongly monotone mapping, T : H → H a nonexpansive mapping and the parameter {αn} satisfies appropriate n=1 converges strongly to a unique solution x′ ∈ F (T ) of the conditions, then the sequence {xn}∞ variational inequality D(γf − µA)x′, y − x′E ≤ 0, ∀y ∈ F (T ). SOLUTION OF VARIATIONAL INEQUALITY ON FIXED POINTS SET 3 His results generalized and improved the corresponding results of Marino and Xu [10] and Yamada [27]. In this paper, we extend Tian's results [24] to a finite family of nonexpansive mappings. More precisely, we consider the following iterative algorithm where I : H → H is an identity mapping, A : H → H is a k−Lipschitzian η−strongly xn+1 = αnγf (xn) + (I − αnµA)Tnxn, n ∈ N, (1.8) monotone mapping, {Ti}N i=1 : H → H are nonexpansive mappings with F := F (Ti) 6= ∅, and Tn := Tn(modN ). Under appropriate conditions, we shall prove strongly convergence of the sequence {xn}∞ n=1 to a unique solution of the variational inequality N Ti=1 D(γf − µA)x′, y − x′E ≤ 0, ∀y ∈ F. Our results generalize the corresponding results of Marino and Xu [10], Tian [24], Xu [26], Yamada [27]. 2. Preliminaries The following lemmas play an important role in proving our main results. Lemma 2.1. [26] Let an be a sequence of nonnegative numbers satisfying the following condition an+1 ≤ (1 − αn)an + αnβn where {αn}∞ n=1, {βn}∞ n=1 are sequences of real numbers such that: (i) 0 < αn < 1 and lim n→∞ αn = 0; ∞ (ii) αn = ∞; Pn=1 (iii) lim sup βn ≤ 0. n→∞ Then lim n→∞ an = 0 Lemma 2.2. [26] In a real Hilbert space H, there holds the following inequality kx + yk2 ≤ kxk2 + 2hy, x + yi, for all x, y ∈ H. Lemma 2.3. [5] Let H be a Hilbert space and T : H → H be a nonexpansive mapping with F ix(T ) 6= ∅. If {xn}∞ n=1 converges to 0, then T x = x. n=1 converges weakly to x and {kxn−T xnk}∞ Lemma 2.4. [24] Let H be a real Hilbert space. Let I : H → H be an identity mapping, f : H → H be a contraction mapping with 0 < α < 1, A : H → H be a k−Lipschitzian η−strongly monotone mapping, and T : H → H be a nonexpansive mapping with F (T ) 6= ∅. Let xt be a unique solution of the following equation xt = tγf (xt) + (I − tµA)T xt, where 0 < µ < 2η inequality k2 , τ := µ(η − µk2 2 ), 0 < γ < τ α , and 0 < t < 1. Then the following variational has a unique solution x′ on the set F (T ) and xt converges strongly to x ′ as t → 0. D(γf − µA)x, y − xE ≤ 0, ∀y ∈ F (T ). (2.1) 4 FARRUKH MUKHAMEDOV AND MANSOOR SABUROV 3. Main results In this section we shall prove our main results. To formulate ones, we need some auxiliary results. Lemma 3.1. Let H be a real Hilbert space. Let I : H → H be an identity mapping, f : H → H be a contraction mapping with 0 < α < 1, A : H → H be a k−Lipschitzian η−strongly monotone mapping, and {Ti}N i=1 : H → H be nonexpansive mappings with F (Ti) 6= ∅, for all i = 1, N . If 0 < µ < 2η k2 and 0 < αn < 1 for all n ∈ N then we have the following inquality ∀n ∈ N, k(I − αnµA)Tnx − (I − αnµA)Tnyk ≤ (1 − αnτ )kx − yk, for all x, y ∈ H, where, Tn := Tn(modN ) and τ := µ(η − µk2 2 ). Proof. Let us denote by Sn := (I − αnµA)Tn. Then, we have for all x, y ∈ H and n ∈ N kSnx − Snyk2 ≤ kTnx − Tnyk2 − 2αnµ(cid:10)Tnx − Tny, ATnx − ATnyE +α2 nµ2kATnx − ATnyk2 nµ2k2(cid:1) kTnx − Tnyk2 2 (cid:19)(cid:19) kx − yk2 µk2 ≤ (cid:0)1 − 2αnµη + α2 ≤ (cid:18)1 − 2αnµ(cid:18)η − = (1 − 2αnτ )kx − yk2 ≤ (1 − αnτ )2kx − yk2, which means This completes the proof. kSnx − Snyk ≤ (1 − αnτ )kx − yk. (3.1) (3.2) (cid:3) Theorem 3.2. Let H be a real Hilbert space. Let I : H → H be an identity mapping, f : H → H be a contraction mapping with 0 < α < 1, A : H → H be a k−Lipschitzian η−strongly monotone mapping, and {Ti}N i=1 : H → H be nonexpansive mappings with F := F (Ti) 6= ∅. N Ti=1 α , and the sequence {αn}∞ n=1 ⊂ (0, 1) Suppose that 0 < µ < 2η satisfies the following condition: k2 , τ := µ(η − µk2 2 ), 0 < γ < τ (i) lim n→∞ ∞ αn = 0; (ii) αn = ∞; Pn=1 (iii) lim n→∞ αn αn+N = 1. If F = F (TN TN −1 · · · T1) = F (T1TN TN −1 · · · T2) = F (T2T1TN TN −1 · · · T3) = · · · = F (TN −1TN −2 · · · T1TN ) then the sequence {xn}∞ x′ ∈ F of the variational inequality n=1 which is defined by (1.8), converges strongly to a unique solution D(γf − µA)x′, y − x′E ≤ 0, ∀y ∈ F. (3.3) Proof. Since the mapping TN TN −1 · · · T1 : H → H is nonexpansive, then due to Lemma 2.4, the variational inequality (3.3) has a unique solution x′ on the set F. We will show that the sequence {xn}∞ n=1 given by (1.8) converges strongly to x′. SOLUTION OF VARIATIONAL INEQUALITY ON FIXED POINTS SET 5 Step 1. The sequences {xn}∞ n=1, {f (xn)}∞ n=1, {Tnxn}∞ n=1, and {ATnxn}∞ n=1, are bounded. Indeed, let p ∈ F and Sn := (I − αnµA)Tn. Using (3.2), we then have kxn+1 − pk = kαnγf (xn) + Snxn − pk = kαnγf (xn) − αnµAp + Snxn − Snpk ≤ αnγαkxn − pk + αnkγf (p) − µApk + (1 − αnτ )kxn − pk kγf (p) − µApk = (1 − αn(τ − γα))kxn − pk + αn(τ − γα) . τ − γα (3.4) Since 0 < γ < τ αn(τ − γα) < 1 for all n ≥ n0. We then obtain α , 0 < αn < 1, and lim n→∞ αn = 0 then there exists n0 > 0 such that 0 < kxn+1 − pk ≤ max(cid:26)kxn − pk, kγf (p) − µApk τ − γα (cid:27) , ∀n ≥ n0. Therefore, we get kxn+1 − pk ≤ max(cid:26)kxn0 − pk, kγf (p) − µApk τ − γα (cid:27) , ∀n ≥ n0, (3.5) (3.6) which means, {xn}∞ n=1 is a bounded sequence. From the following inequality kf (xn) − f (p)k ≤ αkxn − pk, kTnxn − pk = kTnxn − Tnpk ≤ kxn − pk, kATnxn − Apk = kATnxn − ATnpk ≤ kkTnxn − Tnpk ≤ kkxn − pk, it follows that the sequences {f (xn)}∞ Step 2. One has lim n→∞ follows that n=1, {ATnxn}∞ kxn+1 − Tnxnk = 0. Indeed, from lim n→∞ n=1, {Tnxn}∞ n=1, are bounded. αn = 0, (1.8), and Step 1 it kxn+1 − Tnxnk = lim n→∞ Step 3. For the sequence {xn} we have lim n→∞ lim n→∞ Tn−1 and αnkγf (xn) − µATnxnk = 0 kxn+N − xnk = 0. Indeed, noting that Tn+N −1 = we then obtain Sn+N −1xn−1 − Sn−1xn−1 = µ(αn+N −1 − αn−1)ATn−1xn−1 kxn+N − xnk ≤ kαn+N −1γf (xn+N −1) − αn−1γf (xn−1)k +kSn+N −1xn+N −1 − Sn−1xn−1k ≤ αn+N −1γαkxn+N −1 − xn−1k + γαn+N −1 − αn−1kf (xn−1)k +kSn+N −1xn+N −1 − Sn+N −1xn−1k +µαn+N −1 − αn−1kATn−1xn−1k ≤ (1 − αn+N −1(τ − γα))kxn+N −1 − xn−1k +αn+N −1 − αn−1(cid:16)γkf (xn−1)k + µkATn−1xn−1k(cid:17) = (1 − αn+N −1(τ − γα))kxn+N −1 − xn−1k +αn+N −1(τ − γα)βn+N −1, where βn−N +1 := αn+N −1 − αn−1 αn+N −1 · γkf (xn−1)k + µkATn−1xn−1k τ − γα . 6 FARRUKH MUKHAMEDOV AND MANSOOR SABUROV Letting an+N := kxn+N − xnk, one then gets an+N ≤ (1 − αn+N −1(τ − γα))an+N −1 + αn+N −1(τ − γα)βn+N −1, (3.7) where lim sup βn = lim sup n→∞ n→∞ αn+N −1 − αn−1 αn+N −1 · γkf (xn−1)k + µkATn−1xn−1k τ − γα = 0 According to Lemma 2.1, we obtain Step 4. We have lim n→∞ nonexpansive mapping and using Step 2 one has the following lim n→∞ an+N = lim n→∞ kxn+N − xnk = 0. kxn − Tn+N −1Tn+N −2 · · · Tnxnk = 0. Indeed, noting that Tn is a kxn+N − Tn+N −1xn+N −1k → 0, kTn+N −1xn+N −1 − Tn+N −1Tn+N −2xn+N −2k → 0, kTn+N −1Tn+N −2xn+N −2 − Tn+N −1Tn+N −2Tn+N −3xn+N −3k → 0, ... kTn+N −1Tn+N −2 · · · Tn+1xn+1 − Tn+N −1Tn+N −2Tn+N −3 · · · Tnxnk → 0, as n → ∞. Hence, we obtain kxn+N − Tn+N −1Tn+N −2Tn+N −3 · · · Tnxnk → 0, n → ∞. Since kxn+N − xnk → 0 as n → ∞ (see Step 3) we then get lim n→∞ kxn − Tn+N −1Tn+N −2 · · · Tnxnk = 0. Step 5. We want to show that lim sup n→∞ D(γf −µA)x′, xn−x′E ≤ 0, where x′ is a unique solution of the variational inequality (3.3) for the nonexpansive mapping TN TN −1 · · · T1 : H → H on the set F. Let {xnk}∞ k=1 be a subsequence of {xn}∞ n=1 such that lim sup n→∞ D(γf − µA)x′, xn − x′E = lim k→∞D(γf − µA)x′, xnk − x′E. m=1 of {xnk}∞ Since {xnk}∞ k=1 is bounded and H is a real Hilbert space, then there exists a subsequence m=1 converges weakly to some point y ∈ H. Without {xnkm }∞ loss any generality, we may assume that nkm are such kind of numbers that Tnkm = Ti0 for some i0 ∈ {1, 2, · · · , N} and for all m ∈ N. Then, from Step 4, it follows that k=1 such that {xnkm }∞ lim n→∞ kxnkm − Ti0+N −1Ti0+N −2 · · · Ti0xnkm k = 0. So due to Lemma 2.3, we have y ∈ F (Ti0+N −1Ti0+N −2 · · · Ti0) = F. Therefore, lim sup n→∞ D(γf − µA)x′, xn − x′E = lim m→∞D(γf − µA)x′, xnkm − x′E = D(γf − µA)x′, y − x′E ≤ 0. SOLUTION OF VARIATIONAL INEQUALITY ON FIXED POINTS SET 7 Step 6. The sequence {xn} converges to x′, i.e. lim n→∞ kxn − x′k = 0. Indeed, using Lemma 2.2 and (3.4) we obtain kxn+1 − x′k2 = kSnxn − Snx′ + αn(γf (xn) − µAT x′)k2 ≤ kSnxn − Snx′k2 + 2αnDγf (xn) − µAx′, xn+1 − x′E ≤ (1 − αnτ )2kxn − x k2 + 2αnγDf (xn) − f (x′), xn+1 − x′E ′ +2αnD(γf (xn) − µA)x′, xn+1 − x′E ≤ (1 − αnτ )2kxn − x ′ k2 + 2αnγαkxn − x′kkxn+1 − x′k +2αnD(γf (xn) − µA)x′, xn+1 − x′E ≤ (1 − αnτ )2kxn − x k2 + 2αn(1 − αn(τ − γα))γαkxn − x′k2 ′ +2α2 nγαkγf (x′) − µAx′k + 2αnD(γf (xn) − µA)x′, xn+1 − x′E ≤ (1 − 2αn(τ − γα))kxn − x ′ k2 + 2αn(τ − γα)βn, where βn = D(γf (xn) − µA)x′, xn+1 − x′E τ − γα τ 2 2(τ − γα) +αn(cid:20)(cid:18) − γα(cid:19) kxn − x′k2 + γα τ − γα kγf (x′) − µAx′k(cid:21) . Letting an := kxn − x′k2, we then have an+1 ≤ (1 − 2αn(τ − γα))an + 2αn(τ − γα)βn. (3.8) Since lim sup n→∞ D(γf − µA)x′, xn+1 − x′E ≤ 0 (see Step 5) one can get Therefore, according to Lemma 2.1, we obtain that lim sup βn ≤ 0. n→∞ lim n→∞ an = lim n→∞ kxn − x′k2 = 0, which implies that lim n→∞ kxn − x′k = 0. This completes the proof. (cid:3) Note that if f (x) = γu for some γ ∈ (0, 1) and u ∈ H we recover Xu's [26] result. Our results also generalize the corresponding results of Marino and Xu [10], Tian [24], Yamada [27]. Acknowledgement This work was done while the first named author (F.M.) was visiting the Abdus Salam International Centre for Theoretical Physics, Trieste, Italy as a Junior Associate. He would like to thank the Centre for hospitality and financial support. 8 FARRUKH MUKHAMEDOV AND MANSOOR SABUROV References [1] Bauschke H. H. The Approximation of fixed points of compositions of nonexpansive mappings in Hilbert spaces, J. Math. Anal. Appl. 202 (1996) 150-159. [2] Byrne C. A unified treatment of some iterative algorithms in signal processing and image construction, Inverse problems 20 (2004) 103-120. [3] Chidume C. E. and Ofoedu E. U. A new iteration process for finite families of generalized Lipschitz pseudo- contractive and generalized Lipschitz accretive mappings, Nonlinear Analysis; TMA., In press, accepted manuscript, available online 27 June 2007. [4] Chidume C. E.and Ofoedu E. U. Approximation of common fixed points for finite families of total asymp- totically nonexpansive mappings, J. Math. Anal. Appl., 333(1) (2007), 128-141. [5] Goebel K. Kirk W. A. Topics in metric fixed point theory, Cambridge Studies in Advanced Mathematics 28, CUP, Cambridge, New York, Port Chester, Melbourne, Sydney, 1990. [6] Halpern B. Fixed points of nonexpanding maps, Bulletin of the American Mathematical Society, 73 (1967), 957-961. [7] Ishikawa S. Fixed point by a new iteration method, Proc. Amer. Math. Soc. 44 (1974), 147-150. [8] Lions P. L. Approximation de points fixes de contractions, Computes rendus de lacademie des sciences, serie I-mathematique, 284 (1997) 1357-1359. [9] Mann W. R. Mean value methods in iteration, Proc. Amer. Math. Soc. 4 (1953) 506-510. [10] Marino G. and Xu H. K. A general iterative method for nonexpansive mappings in Hilbert spaces, J. Math. anal. Appl. 318 (2006) 43-52. [11] Megginson R. E. An introduction to Banach space theory, Springer-Verlag, New York, Inc., 1998. [12] Moudafi A. Viscosity approximation metheds for fixed-points problems, J. Math. Anal. Appl. 241 (2000) 46-55. [13] Ofoedu E. U. Iterative approximation of a common zero of a countably infinite family of m-accretive operators in Banach spaces, Hindawi-Fixed point theory and applications, doi:10.1155/2008/325792. [14] Ofoedu E. U. and Shehu Y. Iterative construction of a common fixed point of finite families of nonlinear mappings, in press, to appear in Vol. 16 No 1 of Journal of Applied Analysis (2010). [15] Ofoedu E. U., Shehu Y., and Ezeora J. N. Solution by iteration of nonlinear variational inequalities involving finite family of asymptotically nonexpansive mappings and monotone mappings, PanAmer. Math. J. 18 (2008) (4), 61-75. [16] Polilchuk C. I. and Mammone R. J. Image recovery by convex projections using a least-square constraint, J. Opt. Soc. Amer. A7 (1990) 517-521. [17] Reich S. Weak convergence theorems for nonexpansive mappings in Banach spaces, J. Math. Anal. Appl. 67 (1979), 274-276. [18] Reich S. Extension problems for accretive sets in Banach spaces, J. Functional Analysis 26 (1977), 378-395. [19] S. Reich, Strong convergence theorems for resolvents of accretive operators in Banach spaces, J. Math. Anal. Appl., 75 (1980), 287-292. [20] Rhoades B. E. Fixed point iterations for certain nonlinear mappings, J Math. Anal. Appl. 183 (1994) 118-120. [21] Rockefellar R. T. Monotone operators and proximal point algorithm, SIAM, J. Control Optim. 14 (1976), 877-898. [22] Schu J. Iterative construction of fixed points of asymptotically nonexpansive mappings, J. Math. Anal. Appl. 158 (1991), 407-413. [23] Schu J. Weak and strong convergence of fiexd points of asymptotically nonexpansive mappings, Bull. Austral. Math. Soc. 43 (1991), 153-159. [24] Tian M. A general iterative algorithm for nonexpansive mappings in Hilbert spaces, Nonlinear Analysis (2010), doi: 10.1016/j.na.2010.03.058. [25] Wittmann R. Approximation of fixed points of nonexpansive mappings, Archiv der mathematik, 58 (1992) 486-491. [26] Xu H. K. An iterative approach to quadratic optimization, Journal of Optimization Theory and Applica- tions, 116 (2003) no.3, 659-678. [27] Yamada I. The hybrid steepest decent for the variational inequality problems over the itersection of fixed in D. Butnariu, Y. Censor, S. Reich (Eds), Inherently parallel points sets of nonexpansive mapping: algorithme in feasibility and optimization and their application, Elsevier, New York, (2001) 473-504. [28] Youla D. On deterministic convergence of iterations of related projection operators, J. Vis. Commun. Image Represent 1 (1990), 12-20. SOLUTION OF VARIATIONAL INEQUALITY ON FIXED POINTS SET 9 Farrukh Mukhamedov, Department of Computational & Theoretical Sciences, Faculty of Sciences, International Islamic University Malaysia, P.O. Box, 141, 25710, Kuantan, Pahang, Malaysia E-mail address: [email protected] Mansoor Saburov, Department of Computational & Theoretical Sciences, Faculty of Sci- ence, International Islamic University Malaysia, P.O. Box, 141, 25710, Kuantan, Pahang, Malaysia E-mail address: [email protected]
1502.01957	1	1502	2015-02-06T17:02:53	Functional calculus for $C_0$-semigroups using infinite-dimensional systems theory	[ "math.FA" ]	In this short note we use ideas from systems theory to define a functional calculus for infinitesimal generators of strongly continuous semigroups on a Hilbert space. Among others, we show how this leads to new proofs of (known) results in functional calculus.	math.FA	math	FUNCTIONAL CALCULUS FOR C0-SEMIGROUPS USING INFINITE-DIMENSIONAL SYSTEMS THEORY FELIX L. SCHWENNINGER AND HANS ZWART Dedicated to Charles Batty on the occasion of his sixtieth birthday. Abstract. In this short note we use ideas from systems theory to define a functional calculus for infinitesimal generators of strongly continuous semi- groups on a Hilbert space. Among others, we show how this leads to new proofs of (known) results in functional calculus. 5 1 0 2 b e F 6 ] . A F h t a m [ 1 v 7 5 9 1 0 . 2 0 5 1 : v i X r a 1. Introduction Let A be a linear operator on the linear space X. In essence, a functional calculus provides for every (scalar) function f in the algebra A a linear operator f (A) from (a subspace of) X to X such that • f 7→ f (A) is linear; • f (s) ≡ 1 is mapped on the identity I; • If f (s) = (s − r)−1, then f (A) = (A − rI)−1; • For f = f1 · f2 we have f (A) = f1(A)f2(A). As the domains of the operators f (A) might differ, the above properties have to be seen formally, and, in general, need to be made rigorous. It is well-known that self- adjoint (or unitary operators) on a Hilbert space have a functional calculus with A being the set of continuous functions from R (or the torus T respectively) to C, (von Neumann [10]). This theory has been further extended to different operators and algebra's, see e.g. [7], [3], and [2]. For an excellent overview, in particular on the H∞-calculus, we refer to the book by Markus Haase, [5]. For the algebra of bounded analytic functions on the left half-plane and A the infinitesimal generator of a strongly continuous semigroup, we show how to build a functional calculus using infinite-dimensional systems theory. 2. Functional calculus for H− ∞ We choose our class of functions to be H− ∞, i.e., the algebra of bounded analytic functions on the left half-plane. For A we choose the generator of an exponentially stable strongly continuous semigroup on the Hilbert space X. This semigroup will be denoted by (cid:0)eAt(cid:1)t≥0. We refer to [4] for a detailed overview on C0-semigroups. Date: 5 February 2015. 2010 Mathematics Subject Classification. 47A60 (primary), 93C25 (secondary). Key words and phrases. Functional calculus; H∞-calculus; C0-semigroups; Infinite- dimensional systems theory; Admissibility. The first named author has been supported by the Netherlands Organisation for Scientific Research (NWO), grant no. 613.001.004. 1 2 FELIX L. SCHWENNINGER AND HANS ZWART In the following all semigroups are assumed to be strongly continuous. To explain our choice/set-up we start with the following observation. Let h be an integrable function from R to C which is zero on (0,∞) and let t 7→ 1(t) denote the indicator function of [0,∞), i.e., 1(t) = 1 for t ≥ 0 and 1(t) = 0 for t < 0. Then for t > 0 −∞ (cid:0)h ∗ eA·x0 1(·)(cid:1) (t) = Z ∞ = (cid:20)Z t = (cid:20)Z 0 h(τ )eA(t−τ )x0 1(t − τ )dτ h(τ )e−Aτ dτ(cid:21) eAtx0 h(τ )e−Aτ dτ(cid:21) eAtx0. −∞ −∞ Hence the convolution of h with the semigroup gives an operator times the semi- group. We denote this operator by g(A), with g the Laplace transform of h. Now we want to extend the mapping g 7→ g(A). Therefore we need the Hardy space H 2(X) = H 2(C+; X), i.e., the set of X-valued functions, analytic on the right half-plane which are uniformly square integrable along every line parallel to the imaginary axis. By the (vector-valued) Paley-Wiener Theorem, this space is isomorphic to L2((0,∞); X) under the Laplace transform, see [1, Theorem 1.8.3]. ∞ and f ∈ L2((0,∞); X) we Definition 2.1. Let X be a Hilbert space. For g ∈ H− define the Toeplitz operator Mg(f ) = L−1 [Π(g (L (f ))] , (1) where L and L−1 denotes the Laplace transform and its inverse, respectively, and Π is the projection from L2(iR, X) onto H 2(X). Remark 2.2. If we take f (t) = eAtx0, t ≥ 0, and "g = L(h)", then this extends the previous convolution. The following norm estimate is easy to see. Lemma 2.3. Under the conditions of Definition 2.1 we have that Mg is a bounded linear operator from L2((0,∞); X) to itself with norm satisfying (2) kMgk ≤ kgk∞. To show that Definition 2.1 leads to a functional calculus, we need the following concept from infinite dimensional systems theory, see e.g. [16]. Definition 2.4. Let Y be a Hilbert space, and C a linear operator bounded from D(A), the domain of A, to Y . C is an admissible output operator if the mapping x0 7→ CeA·x0 can be extended to a bounded mapping from X to L2([0,∞); Y ). Since in this paper only admissible output operators appear, we shall sometimes omit "output". In [17] the following was proved. Theorem 2.5. Let A be the generator of an exponentially stable semigroup on the Hilbert space X. For every g ∈ H− ∞ there exists a linear mapping g(A) : D(A) 7→ X such that Furthermore, ((cid:0)Mg(eA·x0)(cid:1) (t) = g(A)eAtx0, x0 ∈ D(A). • g(A) is an admissible operator; FUNCTIONAL CALCULUS USING SYSTEMS THEORY 3 • g(A)eAt extends to a bounded operator for t > 0; • g(A) commutes with the semigroup; • g(A) can be extended to a closed operator gΓ(A) such that g 7→ gΓ(A) has • This (unbounded) calculus extends the Hille-Phillips calculus. the properties of an (unbounded) functional calculus; Hence in general the functional calculus constructed in this way will contain unbounded operators. However, they may not be "too unbounded", as the product with any admissible operator is again admissible. Theorem 2.6 (Lemma 2.1 in [17]). Let A be the generator of an exponentially stable semigroup on the Hilbert space X and let C be an admissible operator, then (cid:0)Mg(CeA·x0)(cid:1) (t) = Cg(A)eAtx0, x0 ∈ D(A2). Moreover, Cg(A) extends to an admissible output operator. 3. Analytic semigroups From Theorem 2.5 we know that g(A)eAt is a bounded operator for t > 0. In this section we show that for analytic semigroups the norm of g(A)eAt behaves like log(t) for t close to zero. Let A generate an exponentially stable, analytic semigroup on the Hilbert space X. Then there exists a M, ω > 0 such that, see [11, Theorem 2.6.13], (3) k(−A) 1 2 eAtk ≤ M 1 √t e−ωt, t > 0. Using this inequality, we prove the following estimate. Theorem 3.1. Let A generate an exponentially stable, analytic semigroup on the Hilbert space X. There exists m, ε0 > 0 such that for every g ∈ H− (4) If we assume that (−A∗) (5) If both (−A∗) Proof. For y ∈ D(A∗), x ∈ D(A2) we have kg(A)eAεk ≤ mkgk∞ log(ε). 2 is admissible, then 2 or (−A) kg(A)eAεk ≤ mkgk∞p log(ε) ε ∈ (0, ε0). 2 are admissible, then g(A) is bounded. ∞, ε ∈ (0, ε0) 2 and (−A) for 1 1 1 1 1 2hy, g(A)eA2εxi = Z ∞ = Z ∞ 0 0 hy, (−A)eA2tg(A)eA2εxidt h(−A∗) 2 eA∗εeA∗ty, g(A)(−A) 1 1 2 eAεeAtxidt, where we used that g(A) commutes with the semigroup. Using Cauchy-Schwarz's inequality, we find (6) 1 2hy, g(A)eA2εxi ≤ k(−A∗) = k(−A∗) ≤ k(−A∗) 1 1 1 2 eA∗εeA∗·ykL2kg(A)(−A) 2 eA∗εeA∗·ykL2 · kMg(cid:16)(−A) 2 eA∗εeA∗·ykL2 · kgk∞ · k(−A) 2 eAεeA·xkL2 2 eAεeA·x(cid:17)kL2 2 eAεeA·xkL2, 1 1 1 4 FELIX L. SCHWENNINGER AND HANS ZWART where we used Lemma 2.3. Hence it remains to estimate the two L2-norms. Since X is a Hilbert space (cid:0)eA∗t(cid:1)t≥0 is an analytic semigroup as well. Hence both L2-norms behave similarly. We do the estimate for eAt. For ωε < 1/4, k(−A) 1 2 eAεeA·xk2 1 1 ε 0 k(−A) k(−A) 2 eAεeAtxk2dt 2 eAtxk2dt kxk2dt e−2εωt L2 = Z ∞ = Z ∞ ≤ M 2Z ∞ = M 2kxk2Z ∞ ≤ M 2kxk2m1 log(εω), e−2ωt dt t t 1 ε where we used (3) and m1 is an absolute constant. Combining the estimates and using the fact that ω is fixed, we find that there exists a constant m3 > 0 such that for all x ∈ D(A2) and y ∈ D(A∗) there holds hy, g(A)eA2εxi ≤ m3 log(ε)kgk∞kxkkyk. Since D(A2) and D(A∗) are dense in X, we have proved the estimate (4). We continue with the proof of inequality (5). If (−A∗) implies that 1 2 is admissible, then (6) 1 2hy, g(A)eA2εxi ≤ k(−A∗) 1 2 eA∗εeA∗·ykL2kg(A)(−A) 2 eAεeA·x(cid:17) kL2. ≤ m2kyk · kMg(cid:16)(−A) 1 1 2 eAεeA·xkL2 The estimate follows as shown previously. Let us now assume that (−A) missible. Then by Theorem 2.6 there holds 1 2 is ad- kg(A)(−A) 1 2 eAεeA·xkL2 ≤ kg(A)(−A) = kMg(cid:16)(−A) ≤ kgk∞k(−A) ≤ kgk∞mkxk, 1 1 2 eA·xkL2 2 eA·x(cid:17) kL2 2 eA·xkL2 1 where we have used Lemma 2.3 and the admissibility of (−A) (5) follows similarly as in the first part. 1 2 . Now the proof of 1 1 If (−A) 2 and (−A∗) 2 are both admissible, then we see from the above that the epsilon disappears from the estimate, and since the semigroup is strongly continu- ous, g(A) extends to a bounded operator. (cid:3) In [13], it is shown that for any δ ∈ (0, 1) there exists an analytic, exponentially stable semigroup on a Hilbert space, and g ∈ H− 2 is admissible and kg(A)eAεk ∼ (p log(ε))1−δ. Similarly, the sharpness of (4) is shown. In the next section we relate the above theorem to results in the literature. ∞ such that (−A) 1 FUNCTIONAL CALCULUS USING SYSTEMS THEORY 5 4. Closing remarks A natural question is whether the calculus above coincides with other definitions ∞-calculus. As the construction extends the Hille-Phillips calculus, the of the H− answer is "yes", see [14]. In [15], Vitse showed a similar estimate as in (4) for analytic semigroups on general Banach spaces by using the Hille-Phillips calculus. The setting there is slightly different since bounded analytic semigroups and functions g ∈ H− ∞ with bounded Fourier spectrum are considered. In [13], the authors improve Vitse's result with a more direct technique. In the course of that work, the approach to Theorem 3.1 via the calculus construction used here was obtained. Moreover, the techniques here and in Vitse's work [15] require that the functions f are bounded, analytic on a half-plane. In [13] it is shown that the corresponding result is even true for functions f that are only bounded, analytic on sectors which are larger than the sectorality sector of the generator A. Furthermore, Haase and Rozendaal proved that (4) holds for general (exponen- tially stable) semigroups on Hilbert spaces, see [6]. Their key tool is a transference principle. More general, they show that on general Banach spaces one has to con- sider the analytic multiplier algebra AM2(X), as the function space to obtain a corresponding result. Note that AM2(X) is continuously embedded in H− ∞ with equality if X is a Hilbert space. The difference in the transference principle and the approach followed here is that in the transference principle, estimates are first proved for "nice" functions and than extended to the whole space H− ∞. Whereas we prove the result first for "nice" elements in X, and then extend the operators g(A). 1 1 1 The fact that the calculus is bounded for analytic semigroups when both (−A) 2 are admissible, can already be found in [8]. However, as the admissi- 2 is equivalent to A satisfying square function estimates, the result is and (−A∗) bility of (−A) much older and goes back to McIntosh, [9]. 2 The construction of the H− Banach spaces, see [12, 14]. ∞-calculus followed here can be adapted to general References [1] W. Arendt, C. J. K. Batty, M. Hieber, and F. Neubrander. Vector-valued Laplace transforms and Cauchy problems, volume 96 of Monographs in Mathematics. Second Edition. Birkhauser Verlag, Basel, 2011. [2] D. Albrecht, X. Duong and A. McIntosch, Operator theory and harmonic analysis, appeared in: Workshop on Analysis and Geometry, 1995, Part III, Proceedings of the Centre for Math- ematics and its Applications, ANU, Canberra, 34 (1996) 77-136. [3] N. Dunford and J.T. Schwartz, Linear Operators, Part III: Spectral Operators, Wiley, 1971. [4] K.-J. Engel and R. Nagel, One-parameter Semigroups for Linear Evolution Equations, Grad- uate Texts in Mathematics, vol. 194, Springer-Verlag, New York, 2000. [5] M. Haase, The Functional Calculus for Sectorial Operators, Operator Theory, Advances and Applications, Vol. 169, Birkhauser, Basel, 2006. [6] M. Haase and J. Rozendaal, Functional calculus for semigroup generators via transference, Journal of Functional Analysis, 265 (2013) 3345 -- 3368. [7] E. Hille and R.S. Phillips, Functional Analysis and Semi-Groups, AMS, 1957. [8] C. Le Merdy, The Weiss conjecture for bounded analytic semigroups, J. London Math. Soc., 67 (2003), 715 -- 738. [9] A. McIntosh, Operators which have an H∞ functional calculus, Miniconference on Opera- tor Theory and Partial Differential Equations, Proceedings of the Centre for Mathematical Analysis, Australian National University, 14 (1986) 220-231. 6 FELIX L. SCHWENNINGER AND HANS ZWART [10] J. von Neumann, Mathematische Grundlagen der Quantummechanic, zweite Auflage, Springer Verlag, reprint 1996. [11] A. Pazy. Semigroups of linear operators and applications to partial differential equations, volume 44 of Applied Mathematical Sciences. Springer-Verlag, New York, 1983. [12] F.L. Schwenninger and H. Zwart, Weakly admissible H− ∞-calculus on reflexive Banach spaces, Indag. Math. (N.S.), 23(4) (2012), 796 -- 815. [13] F.L. Schwenninger, On measuring unboundedness of the H∞-calculus for generators of ana- lytic semigroups, submitted 2015. [14] F.L. Schwenninger and H. Zwart, The (weak) admissibility of the H∞-calculus for semigroup generators, in preparation, 2014. [15] P. Vitse, A Besov class functional calculus for bounded holomorphic semigroups, Journal of Functional Analysis, 228 (2005), 245 -- 269. [16] G. Weiss, Admissible observation operators for linear semigroups, Israel Journal of Mathe- matics, 65-1 (1989) 17 -- 43. [17] H. Zwart, Toeplitz operators and H∞ calculus, Journal of Functional Analysis, 263 (2012) 167 -- 182. Felix L. Schwenninger, Department of Applied Mathematics, P.O. Box 217, 7500 AE Enschede, The Netherlands E-mail address: [email protected] Hans Zwart, Department of Applied Mathematics, P.O. Box 217, 7500 AE Enschede, The Netherlands E-mail address: [email protected]
1008.2651	1	1008	2010-08-16T13:29:33	Factorization property and Arens regularity	[ "math.FA" ]	In this paper, we study the Arens regularity properties of module actions and we extend some proposition from Baker, Dales, Lau and others into general situations. For Banach $A-bimodule$ $B$, let $Z_1(A^{})$, ${Z}^\ell_{B^{}}(A^{})$ and ${Z}^\ell_{A^{}}(B^{*})$ be the topological centers of second dual of Banach algebra $A$, left module action $\pi_\ell:~A\times B\rightarrow B$ and right module action $\pi_r:~B\times A\rightarrow B$, respectively. We establish some relationships between them and factorization properties of $A^$ and $B^$. We search some necessary and sufficient conditions for factorization of $A^$, $B$ and $B^*$ with some results in group algebras. We extend the definitions of the left and right multiplier for module actions.	math.FA	math	ARENS REGULARITY AND FACTORIZATION PROPERTY KAZEM HAGHNEJAD AZAR Abstract. In this paper, we study the Arens regularity properties of module actions and we extend some proposition from Baker, Dales, Lau and others into general situations. For Banach A − bimodule B, let Z1(A∗∗), Z ℓ B∗∗ (A∗∗) and Z ℓ A∗∗ (B∗∗) be the topological centers of second dual of Banach algebra A, left module action πℓ : A × B → B and right module action πr : B × A → B, respectively. We establish some relationships between them and factorization properties of A∗ and B∗. We search some necessary and sufficient conditions for factorization of A∗, B and B∗ with some results in group algebras. We extend the definitions of the left and right multiplier for module actions. 1. Preliminaries and Introduction In 1951 Arens shows that the second dual A∗∗ of Banach algebra A endowed with the either Arens multiplications is a Banach algebra, see [1]. The constructions of the two Arens multiplications in A∗∗ lead us to definition of topological centers for A∗∗ with respect to both Arens multiplications. The topological centers of Banach algebras, module actions and applications of them were introduced and discussed in [3, 5, 6, 9, 15, 16, 17, 18, 19, 24, 25]. In this paper, we extend some problems from [3, 5, 6, 16, 22] to the general criterion on module actions with some applications in group algebras. Now we introduce some notations and definitions that we used in this paper. Throughout this paper, A is a Banach algebra and A∗, A∗∗, respectively, are the first and second dual of A. For a ∈ A and a′ ∈ A∗, we denote by a′a and aa′ respectively, the functionals on A∗ defined by ha′a, bi = ha′, abi = a′(ab) and haa′, bi = ha′, bai = a′(ba) for all b ∈ A. The Banach algebra A is embedded in its second dual via the identification ha, a′i - ha′, ai 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∗. The extension of bilinear maps on normed space and the concept of regularity of bilinear maps were studied by [1, 2, 5, 6, 9]. We start by recalling these definitions as follows. 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 hm∗(z′, x), yi = hz′, m(x, y)i where x ∈ X, y ∈ Y , z′ ∈ Z ∗, 2. m∗∗ : Y ∗∗ × Z ∗ → X ∗, given by hm∗∗(y′′, z′), xi = hy′′, m∗(z′, x)i where x ∈ X, 2000 Mathematics Subject Classification. 46L06; 46L07; 46L10; 47L25. Key words and phrases. Arens regularity, bilinear mapping, topological center, module action, factorization, weakly compact. 1 2 y′′ ∈ Y ∗∗, z′ ∈ Z ∗, 3. m∗∗∗ : X ∗∗ × Y ∗∗ → Z ∗∗, given by hm∗∗∗(x′′, y′′), z′i = hx′′, m∗∗(y′′, z′)i 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 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 hz′, m(xi, yj)i = lim j hz′, m(xi, yj)i, lim i whenever both limits exist for all bounded sequences (xi)i ⊆ X , (yi)i ⊆ Y and z′ ∈ Z ∗, see [5]. The mapping m is left strongly Arens irregular if Z1(m) = X and m is right strongly Arens irregular if Z2(m) = Y . Let now 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, respectively. 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 left and right module actions of A on B as follows: ZB∗∗(A∗∗) = Z(πℓ) = {a′′ ∈ A∗∗ : the map b′′ → π∗∗∗ ℓ (a′′, b′′) : B∗∗ → 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} ZA∗∗(B∗∗) = Z(πr) = {b′′ ∈ B∗∗ : the map a′′ → π∗∗∗ r (b′′, a′′) : A∗∗ → B∗∗ is weak∗ − to − weak∗ continuous} 3 Z t A∗∗(B∗∗) = Z(πt ℓ) = {b′′ ∈ B∗∗ : the map a′′ → πt∗∗∗ ℓ (b′′, a′′) : A∗∗ → 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. We write ab = πℓ(a, b), ba = πr(b, a), πℓ(a1a2, b) = πℓ(a1, a2b), πr(b, a1a2) = πr(ba1, a2), π∗ r (b′a, b) = π∗ for all a1, a2, a ∈ A, b ∈ B and b′ ∈ B∗ when there is no confusion. Regarding A as a Banach A − bimodule, the operation π : A × A → A extends to π∗∗∗ and πt∗∗∗t defined on A∗∗ × A∗∗. These extensions are known, respectively, as the first(left) and the second (right) Arens products, and with each of them, the second dual space A∗∗ becomes a Banach algebra. In this situation, we shall also simplify our notations. So the first (left) Arens product of a′′, b′′ ∈ A∗∗ shall be simply indicated by a′′b′′ and defined by the three steps: ℓ (a1b′, a2) = π∗ ℓ (b′, a2a1), π∗ r (b′, ab), ha′a, bi = ha′, abi, ha′′a′, ai = ha′′, a′ai, ha′′b′′, a′i = ha′′, b′′a′i. for every a, b ∈ A and a′ ∈ A∗. Similarly, the second (right) Arens product of a′′, b′′ ∈ A∗∗ shall be indicated by a′′ob′′ and defined by : haoa′, bi = ha′, bai, ha′oa′′, ai = ha′′, aoa′i, ha′′ob′′, a′i = hb′′, a′ob′′i. for all a, b ∈ A and a′ ∈ A∗. 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 Goldstine,s Theorem [4, 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 ha′, π(aα, bβ)i = ha′′b′′, a′i lim β ha′, π(aα, bβ)i = ha′′ob′′, a′i, lim α where a′′b′′ and a′′ob′′ are the first and second Arens products of A∗∗, respectively, see [16, 20]. We find the usual first and second topological center of A∗∗, which are Z1(A∗∗) = ZA∗∗(A∗∗) = Z(π) = {a′′ ∈ A∗∗ : b′′ → a′′b′′ is weak∗ − to − weak∗ Z2(A∗∗) = Z t A∗∗(A∗∗) = Z(πt) = {a′′ ∈ A∗∗ : a′′ → a′′ob′′ is weak∗ − to − weak∗ continuous}, continuous}. 4 Recall that a left approximate identity (= LAI) [resp. right approximate identity (= RAI)] in Banach algebra A is a net (eα)α∈I in A such that eαa −→ a [resp. aeα −→ a]. We say that a net (eα)α∈I ⊆ A is a approximate identity (= AI) for A if it is LAI and RAI for A. If (eα)α∈I in A is bounded and AI for A, then we say that (eα)α∈I is a bounded approximate identity (= BAI) for A. Let A have 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. 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. 2. Factorization property and topological centers of module actions Baker, Lau and Pym in [3] proved that for Banach algebra A with bounded right approximate identity, (A∗A)⊥ is an ideal of right annihilators in A∗∗ and A∗∗ ∼= (A∗A) ⊕ (A∗A)⊥. In the following, for a Banach A − bimodule B, we study the similar discussion on the module actions and for Banach A − bimodule B, we show that B∗∗ = (B∗A)∗ ⊕ (B∗A)⊥. Theorem 2-1. Let B be a Banach A − bimodule and A has a BRAI. Then the following assertions are hold: i) (B∗A)⊥ = {b′′ ∈ B∗∗ : π∗∗∗ ii) (B∗A)∗ is isomorphism with HomA(B∗, A∗). Proof. i) Let b′′ ∈ (B∗A)⊥. Then for all b′ ∈ B∗ and a ∈ A, we have (a′′, b′′) = 0 f or all a′′ ∈ A∗∗}. ℓ hπ∗∗ ℓ (b′′, b′), ai = hb′′, π∗ ℓ (b′, a)i = 0, it follows that for all a′′ ∈ A∗∗, we have hπ∗∗∗ ℓ (a′′, b′′), b′i = ha′′, π∗∗ ℓ (b′′, b′)i = 0. (a′′, b′′) = 0 for all a′′ ∈ A∗∗. Then for all Conversely, let b′′ ∈ B∗∗ such that π∗∗∗ a ∈ A and b′ ∈ B∗, we have ℓ (b′, a)i = hπ∗∗ hb′′, π∗ ℓ ℓ (b′′, b′), ai = ha, π∗∗ ℓ (b′′, b′)i = hπ∗∗∗ ℓ (a, b′′), b′i = 0, which implies that b′′ ∈ (B∗A)⊥. ii) Suppose that b′′ ∈ B∗∗. We define Tb′′ ∈ HomA(B∗, A∗), that is, Tb′′ b′ = ℓ (b′′, b′). Then Λ = b′′ → Tb′′ is linear continuous map from B∗∗ into HomA(B∗, A∗) π∗∗ such that KerΛ = {b′′ ∈ B∗∗ : π∗∗ Consequently, b′′ ∈ KerΛ if and only if ℓ (b′, b′′) = 0 f or all b′ ∈ B∗}. hb′′, π∗ ℓ (b′, a) = hπ∗∗ ℓ (b′′, b′), ai = 0, where b′ ∈ B∗ and a ∈ A. It follows that b′′ ∈ (B∗A)⊥. Since (B∗A)∗ ∼= B∗∗ continuous linear mapping Λ from (B∗A)∗ into HomA(B∗, A∗) is injective. Conversely, suppose that T ∈ HomA(B∗, A∗) and e′′ ∈ A∗∗ is any right identity for A∗∗. We define b′′ T ∈ B∗∗ such that for all b′, we have (B∗A)⊥ , the It is clear that the linear mapping T → b′′ T is continuous. For all a ∈ A, we have hb′′ T , b′i = he′′, T b′i. 5 hπ∗∗ ℓ (b′′ T , b′), ai = hb′′ T , π∗ ℓ (b′, a)i = he′′, T π∗ = hae′′, T b′i = hT b′, ai. ℓ (b′, a)i = he′′, (T b′)ai Consequently, π∗∗ T → Tb′′ is the identity map and consequently the isomorphism between HomA(B∗, A∗) and (B∗A)∗ is established. T , b′) = T b′. It follows that the linear mapping T → b′′ ℓ (b′′ T By using proceeding theorem we observe that HomA(B∗, A∗) has an right identity. Also if B = A, we obtain Theorem 1-1 from [3]. (cid:3) Corollary 2-2. Let B be a Banach A − bimodule and let e′′ be any right iden- tity of A∗∗. Then e′′B∗∗ ∼= (B∗A)∗ and (B∗A)⊥ = {b′′ − e′′b′′ : b′′ ∈ B∗∗}. Thus B∗∗ = (B∗A)∗ ⊕ (B∗A)⊥. Example 2-3. Let G be a locally compact group. Let 1 ≤ p < ∞ and 1 Then by using Theorem 2-1, we conclude that p + 1 q = 1. (Lp(G) ∗ L1(G))⊥ = {b ∈ Lq(G) : a′′b = 0 f or every a′′ ∈ L∞(G)}, (Lp(G) ∗ L1(G))∗ ∼= HomL1(G)(Lp(G), L∞(G)). Theorem 2-4. Assume that B is a left Banach A − module and A has a BAI. Then we have the following assertions. (1) If B∗ factors on the left and B∗∗ has a left unit as A∗∗ − module, then (B∗)⊥ = 0. (2) If B∗ not factors on the left and e′′ is a left unit as A∗∗ − module in B∗∗, then e′′ /∈ Z ℓ B∗∗(A∗∗). Proof. (1) Let a ∈ A, b′ ∈ B∗ and b′′ ∈ (B∗A)⊥. Then ℓ (b′, a)i = 0. ℓ (b′′, b′), ai = hb′′, π∗ hπ∗∗ Thus, for all b′′ ∈ A∗∗, we have hπ∗∗∗ ℓ (a′′, b′′), b′i = ha′′, π∗∗ ℓ (b′′, b′)i = 0. 6 It follows that π∗∗∗ Now let e′′ ∈ A∗∗ be a left unit as A∗∗ − module for B∗∗, then we have (a′′, b′′) = 0. ℓ b′′ = π∗∗∗ ℓ (e′′, b′′) = 0. Thus proof hold. (2) Assume a contradiction that e′′ ∈ ZB∗∗(A∗∗). Take b′′ ∈ (B∗A)⊥. Then Let (eα)α ⊆ A such that eα b′′ = π∗∗∗ ℓ (e′′, b′′) = πt∗∗∗t w∗ → e′′ and let b ∈ B, b′ ∈ B∗. Then (e′′, b′′). ℓ hπt∗∗ ℓ (e′′, b′), bi = he′′, πt∗ ℓ (b′, b)i = limαhπt∗ ℓ (b′, b), eαi = limαhb′, πt ℓ(b, eα)i = limαhb′, πℓ(eα, b)i = limαhπ∗ ℓ (b′, eα), bi It follows that Let (bβ)β ⊆ B such that bβ ℓ (b′, eα) = πt∗∗ w∗ − limαπ∗ w∗ → b′′. Then ℓ (e′′, b′). hb′′, b′i = hπt∗∗∗t ℓ (e′′, b′)i (e′′, b′′), b′i = hπt∗∗∗ (b′′, e′′), b′i = hb′′, πt∗∗ = limβhπt∗∗ ℓ (b′, eα), bβi = limβlimαhb′, πℓ(eα, bβ)i = limαlimβhb′, πℓ(eα, bβ)i (e′′, b′), bβi = limβlimαhπ∗ ℓ ℓ ℓ = limαhb′, π∗∗∗ ℓ = limαheα, π∗∗ (eα, b′′)i = limαhπ∗∗∗ ℓ (b′′, b′)i = limαhb′′, π∗ (eα, b′′), b′i ℓ (b′, eα)i = 0. ℓ It follows that (B∗A)⊥ = 0. By using Corollary 2-2, we have B∗∗ = (B∗A)⊥ ⊕ (B∗A)∗, and so B∗∗ = (B∗A)∗. Now since B∗A 6= B∗, by Hahn Banach theo- rem, there is 0 6= b′′ ∈ B∗∗ such that b′′B∗A = 0. It follows that b′′ ∈ (B∗A)⊥ which is contradiction. Corollary 2-5. For a left Banach A − module B, we have the following statements. (1) If B∗A = B∗ and B∗∗ has a left unit as A∗∗ − module, then (B∗)⊥ = 0. (2) If B∗A 6= B∗ and e′′ is a left unit as A∗∗−module in B∗∗, then e′′ /∈ Z ℓ B∗∗(A∗∗). Proof is similar to Theorem 2-4. (cid:3) Corollary 2-6. Assume that B is a left Banach A − module and B∗∗ has a left unit A∗∗ − module. If B∗A 6= B∗, then Z ℓ B∗∗(A∗∗) 6= A∗∗. In the proceeding corollary, take B = A and assume that A∗ not factors on the left. Then, if A has a LBAI, we conclude that A is not Arens regular. Example 2-7. Let G be a locally compact group, by using Corollary 2-6, we have the following inequality L1(G)∗∗(M (G)∗∗) 6= M (G)∗∗ , Z ℓ Z ℓ M(G)∗∗(L1(G)∗∗) 6= L1(G)∗∗. Thus L1(G) and M (G) are not Arens regular. 7 Baker, Dales, Lau, Pym and Ulger in [3, 5, 6, 16] have studied topological center of some Banach algebras and they have solved some problems related to module homomorphism and topological centers of Banach algebras. In the following we extend some problems of them into module actions with some new result and applications in group algebras. Throughout this paper, the notations W SC is used for weakly sequentially complete Banach space A, that is, A is said to be weakly sequentially complete, if every weakly Cauchy sequence in A has a weak limit in A. 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. Suppose that A is a Banach algebra and B is a Banach A − bimodule. According to [28] B∗∗ is a Banach A∗∗ − bimodule, where A∗∗ is equipped with the first Arens product. We define B∗B as a subspace of A, that is, for all b′ ∈ B∗ and b ∈ B, we define hb′b, ai = hb′, bai; We similarly define B∗∗∗B∗∗ as a subspace of A∗∗ and we take A(0) = A and B(0) = B. Let B be a left Banach A − module and (eα)α ⊆ A be a LAI [resp. weakly left ap- proximate identity(=WLAI)] for A. We say that (eα)α is left approximate identity (= LAI)[ resp. weakly left approximate identity (=W LAI)] for B, if for all b ∈ B, we have πℓ(eα, b) → b ( resp. πℓ(eα, b) w→ b). The definition of the right approximate identity (= RAI)[ resp. weakly right approximate identity (= W RAI)] is similar. We say that (eα)α is a approximate identity (= AI)[ resp. weakly approximate iden- tity (W AI)] for B, if B has left and right approximate identity [ resp. weakly left and right approximate identity ] that are equal. Let B be a left Banach as A − module and e be a left unit element of A. Then we say that e is a left unit (resp. weakly left unit) as A − module for B, if πℓ(e, b) = b (resp. hb′, πℓ(e, b)i = hb′, bi for all b′ ∈ B∗) where b ∈ B. The definition of right unit (resp. weakly right unit) as A − module is similar. We say that a Banach A − bimodule B is an unital, if B has the same left and right unit as A − module. Lemma 2-8. Let B be a Banach A−bimodule. Suppose that A has a BAI (eα)α ⊆ A. Then (1) B factors on the left if and only if πr(b, eα) w→ b for every b ∈ B. (2) B factors on the right if and only if πℓ(eα, b) w→ b for every b ∈ B. (3) If B∗ factors on the right, then πr(b, eα) w→ b for every b ∈ B. Proof. (1) Suppose that B factors on the left. Then for every b ∈ B, there are y ∈ B and a ∈ A such that b = ya. Thus for every b′ ∈ B∗, we have r (b′, y), aeαi hb′, πr(b, eα)i = hb′, πr(ya, eα)i = hb′, πr(y, aeα)i = hπ∗ 8 → hπ∗ r (b′, y), ai = hb′, yai = hb′, bi. It follows that πr(b, eα) w→ b. Conversely, by Cohen,s factorization Theorem, since BA is a closed subspace of B, the proof is hold. (2) Proof is similar to (1). (3) Assume that B∗ factors on the right with respect to A. Then for every b′ ∈ B∗, there are y′ ∈ B and a ∈ A such that b′ = ay′. Consequently for every b ∈ B, we have hb′, πr(b, eα)i = hay′, πr(b, eα)i = hy′, πr(b, eα)ai = hy′, πr(b, eαa)i = hπ∗ r (y′, b), eαai → hπ∗ r (y′, b), ai = hy′, πr(b, a)i = hay′, bi It follows that πr(b, eα) w→ b. = hb′, bi. In the proceeding theorem, if we take B = A, then we obtain Lemma 2.1 from [16]. (cid:3) Theorem 2-9. Let B be a Banach A − bimodule and A has a sequential W BAI. Then we have the following assertions. (i) Let B∗ be a W SC and A∗ factors on the left. Then (1) If B factors on the right, it follows that B∗ factors on the left. (2) If B∗ factors on the right, it follows that B factors on the left. (ii) Let B∗∗B∗ = A∗∗A∗. Then A∗ factors on the left if and only if B∗ factors on the left. (iii) Suppose that A is W SC and B factors on the left (resp. right). If B∗B = A∗, then we have the following assertions. (1) A is an unital and B has a right (resp. left) unit as Banach A − module. (2) A∗ factors on the both side and B∗ factors on the right (resp. left). (3) B∗∗ ∼= (AB∗)∗ (resp. B∗∗ ∼= (B∗A)∗). Proof. i) 1) Assume that b′′ ∈ B∗∗ and b′ ∈ B∗. Since A∗ factors on the left, there are a′ ∈ A∗ and a ∈ A such that b′′b′ = a′a. Suppose that (en)n ⊆ A is a sequential W BAI for A. Then we have hb′′, b′eni = hb′′b′, eni = ha′a, eni = ha′, aeni → ha′, ai. It follows that the sequence (b′en)n is weakly Cauchy sequence in B∗. Since B∗ is w→ x′. On the other hand, since B factors on W SC, there is x′ ∈ B∗ such that b′en the right, by using Lemma 2-8, for each b ∈ B, we have enb w→ b. Then we have hx′, bi = lim n hb′en, bi = lim n hb′, enbi = hb′, bi. It follows that x′ = b′, and so by Lemma 2-8, B∗ factors on the left. i) 2) Proof is similar to part (1). 9 ii) Let a′′ ∈ A∗∗ and a′ ∈ A∗. Then there are b′′ ∈ B∗∗ and b′ ∈ B∗ such that b′′b′ = a′′a′. Then ha′′, a′eni = ha′′a′, eni = hb′′b′, eni = hb′′, b′eni. Thus, by Cohen,s factorization theorem, proof hold. iii) 1) Suppose that (ek)k ⊆ A is a sequential W BAI for A. Let a′ ∈ A∗. Since B∗B = A∗, there are b′ ∈ B∗ and b ∈ B such that b′b = a′. Since B factors on the left, there are y ∈ B and a ∈ A such that b = ya. Then we have ha′, eki = hb′b, eki = hb′, beki = hb′, yaeki = hb′y, aeki → hb′y, ai = hb′, yai = hb′, bi. This shows that the sequence (ek)k ⊆ A is weakly sequence in A. Since A is WSC, it convergence weakly to some element e of A. Then, for each x ∈ A, we have xe = x(w − lim k ek) = w − lim k xek = a. It is similar to that ex = x, and so A is unital. Now let b ∈ B, then hb′, bei = hb′b, ei = lim k hb′b, eki = lim k hb′, yaeki = lim k hb′y, aeki → hb′y, ai = hb′, bi. Thus be = b for all b ∈ B. iii) 2) By using part (1) and [16, Theorem 2.6], it is clear that A∗ factors on the both side. Now let b′ ∈ B∗ and b ∈ B. By part (1), set e ∈ A as a left unite element of B. Then heb′, bi = hb′, bei = hb′, bi. It follows that eb′ = b′. Thus B∗ factors on the right. iii) 3) Now let b′′ ∈ (AB∗)⊥. By using part (2), since B∗ factors on the right, for every b′ ∈ B∗ there are x′ ∈ B∗ and a ∈ A such that b′ = ax′. Then hb′′, b′i = hb′′, ax′i = 0. It follows that b′′ = 0. It follows that (AB∗)⊥ = {0}. Therefore by using Corollary 2-2, we are done. (cid:3) In the proceeding theorem if we take A = B, then we have the following statements: Let A has a sequential W BAI and Suppose that A is a W SC. Then (1) A∗ factors on the left (resp. right) if and only if A∗ factors on the right (resp. left), see [16]. (2) A∗ factors on the left (resp. right) if and only if A is an unital, see [16]. (3) If A∗ factors on the left (resp. right), then (AA∗)⊥ ∼= A∗∗ (resp. (A∗A)⊥ ∼= A∗∗). 10 Example 2-10. i) Let G be a locally compact groups. Take B = c0(G) and B∗ = A = ℓ1(G). Since ℓ1(G)ℓ1(G) = ℓ1(G), by proceeding theorem we have c0(G)ℓ1(G) = c0(G). ii) Let ω : G → R+ \ {0} be a continuous function on a locally compact group G. Suppose that L1(G, ω) = {f is Borel measurable : k f kω=ZG f (s)ω(s) ds < ∞}, see [6]. Let X be a closed separable subalgebra of L1(G, ω). Then by using [6, Theorem 7.1] and [24, Lemma 3.2], X is W SC and it has a sequential BAI, respectively. Therefore by above results if X ∗X = X ∗, then X is unital and (XX ∗)⊥ = {0}. On the other hand, if the identity of G has a countable neighborhood base, then L1(G, ω) has a sequential BAI. Now let L1(G, ω)∗ = L∞(G, 1 ω ) factors on the left. Hence LU C(G, 1 ω ) = L∞(G, 1 ω )L1(G, ω) = L∞(G, 1 ω ), where L∞(G, 1 Consequently by above results, L1(G, ω) is unital and RU C(G, 1 ω ) = {f is Borel measurable : k f k∞,ω= ess sups∈G f (s) ω(s) < ∞}. ω )⊥ = {0}, where RU C(G, 1 ω ) = L1(G, ω)L∞(G, 1 ω ). Thus L1(G, ω)∗∗ = RU C(G, 1 ω )∗. Theorem 2-11. Suppose that B is a left Banach A − module and it has a W LBAI (eα)α ⊆ A. Then we have the following assertions. (1) B factors on the left. (2) If A∗ factors on the left, then B∗ factors on the left. Proof. (1) By using Lemma 2-8, proof hold. (2) Let b′′ ∈ B∗∗ and b′ ∈ B∗. Since π∗∗ right, there are a′ ∈ A∗ and a ∈ A such that π∗∗ ℓ (b′′, b′) ∈ A∗ and A∗ factors on the ℓ (b′′, b′) = a′a. Without loss w∗ → e′′ where e′′ left unit for A∗∗. Then for every b ∈ B, generality, we let eα we have hπ∗∗∗∗ ℓ (b′, e′′), bi = hb′, π∗∗∗ ℓ (e′′, b)i = lim α hb′, πℓ(eαb)i = hb′, bi. It follows that π∗∗∗∗ (b′, e′′) = b′. Now, we have the following equality ℓ hb′′, π∗ = hπ∗∗∗∗∗ ℓ ℓ (b′, eα) − b′i = hb′′, π∗∗∗∗ (b′′, b′), (eα − e′′)i = hπ∗∗ = ha′a, (eα − e′′)i = ha′, aeα − ae′′i ℓ (b′, (eα − e′′))i ℓ (b′′, b′), (eα − e′′)i It follows that π∗ ℓ (b′, eα) w→ b′, and so by Cohen,s factorization, we are done. = ha′, aeα − ai → 0. (cid:3) Theorem 2-12. Suppose that B is a right Banach A − module and it has a RBAI (eα)α ⊆ A. Then we have the following assertions. (1) B factors on the right. (2) Let Z ℓ A∗∗(B∗∗) = B∗∗. If A∗ factors on the right, then B∗ factors on the right. Proof. (1) By using Lemma 2-8, proof hold. 11 (2) Let b′′ ∈ B∗∗ and b′ ∈ B∗. First we show that π∗∗∗∗ r w∗ → a′′. Since Z ℓ (b′, b′′) ∈ A∗. Suppose A∗∗(B∗∗) = B∗∗, for each that (a′′ b′′ ∈ B∗∗, we have π∗∗∗ α)α ⊆ A∗∗ such that a′′ α α) w∗ → π∗∗∗ α), b′i → hπ∗∗∗ αi = hπ∗∗∗ (b′, b′′), a′′ (b′′, a′′ (b′′, a′′ r r hπ∗∗∗∗ r (b′′, a′′). Then (b′′, a′′), b′i = hπ∗∗∗∗ (b′, b′′), a′′i. r r r Consequently π∗∗∗∗ right, there are a′ ∈ A∗ and a ∈ A such that π∗∗∗∗ generality, we let eα we have hπ∗∗ r (e′′, b′), bi = hb′, π∗∗∗ (b′, b′′) ∈ (A∗∗, weak∗)∗ = A∗. Since A∗ factors on the (b′, b′′) = a′a. Without loss w∗ → e′′ where e′′ right unit for A∗∗. Then for each b ∈ B, hb′, πr(b, eα)i = hb′, bi. r r r (b, e′′)i = lim α It follows that π∗∗ r (e′′, b′) = b′. Now we have the following equality hb′′, π∗∗ r (eα, b′) − b′i = hb′′, π∗∗ = hπ∗∗∗∗ r (b′, b′′), eα − e′′i = ha′a, eα − e′′i r (eα, b′) − π∗∗ r (e′′, b′)i = ha′, aeα − ae′′i = ha′, aeα − ai → 0. Thus π∗∗ r (eα, b′) w→ b′. Consequently by Cohen,s factorization, we are done. Corollary 2-13. Suppose that B is a Banach A − bimodule and it has a BAI (eα)α ⊆ A. Then we have the following assertions. (1) B factors. (2) Let Z ℓ A∗∗(B∗∗) = B∗∗. If A∗ factors on the both side, then B∗ factors on the both side. (cid:3) Example 2-14. Assume that G is a locally compact group. We know that L1(G) is a M (G) − bimodule. Since M (G)L1(G) 6= M (G) and L1(G)M (G) 6= M (G), by using Theorem 2-11 and 2-12, we conclude that every LBAI or RBAI for L1(G) is not LBAI or RBAI for M (G), respectively. Definition 2-15. Let B be a left Banach A − module. Then B is said to be left weakly completely continuous (= Lwcc), if for each a ∈ A, the mapping b → πℓ(a, b) from B into B is weakly compact. The definition of right weakly completely contin- uous (= Rwcc) is similar. We say that B is a weakly completely continuous (= wcc), if B is Lwcc and Rwcc. 12 Theorem 2-16. Let B be a left Banach A − module and for all a ∈ A, La be the linear mapping from B into itself such that Lab = πℓ(a, b) for all b ∈ B. Then AB∗∗ ⊆ B if and only if La is weakly compact. Proof. Assume that AB∗∗ ⊆ B. We take L∗ that L∗ ab′ = π∗ ℓ (b′, a) for all b′ ∈ B∗. Then for every b′′ ∈ B∗∗, we have hL∗∗ ℓ (b′′, b′), ai a b′′, b′i = hb′′, L∗ ℓ (b′, a)i = hπ∗∗ ab′i = hb′′, π∗ a as the adjoint of La. It is easy to show = ha, π∗∗ ℓ (b′′, b′)i = hπ∗∗∗ ℓ (a, b′′), b′i. It follows that L∗∗ a b′′ = π∗∗ ℓ (a, b′′). w∗ → b′. Since π∗∗∗ ℓ ℓ ℓ Let (b′ (a, b′′), b′ αi = hL∗∗ αi = hπ∗∗∗ (a, b′′) ∈ B, we have α)α ⊆ B∗ such that b′ α hb′′, L∗ a b′′, b′ a is weak∗ − to − weak continuous, so La is weakly compact. We conclude that L∗ Conversely, assume that b′′ ∈ B∗∗. Then by Goldstine,s theorem [9, P.424-425], w∗ → b′′. Since for all a ∈ A, the operator La is there is a net (bα)α ⊆ B such that bα weakly compact, there is a subnet (bαβ )β from (bα)α such that (La(bαβ ))β is weakly w∗ → b′′, (La(bαβ ) = π(a, bαβ ))β is weakly (a, b′′), b′i = hb′′, L∗ αi = hπ∗∗∗ convergence to some point of B. Since bα convergence to π∗∗∗ (a, b′′). Consequently, for all b′ ∈ B∗, we have ab′i. ab′ ℓ hπ∗∗∗ ℓ (a, b′′), b′i = limβhb′, π(a, bαβ )i = limβhb′, Labαβ )i. It follows that π∗∗∗ ℓ (a, b′′) ∈ B. (cid:3) For a right Banach A − module B, we can write Theorem 2-16 as follows. Assume for all a ∈ A, Ra be the linear mapping from B into itself such that Rab = πr(b, a) for all b ∈ B. Then B∗∗A ⊆ B if and only if Ra is weakly com- pact. Proof of this assertion is similar to proof of Theorem 2-16. Corollary 2-17. i) Suppose that B is a left Banach A − module. Then AB∗∗ ⊆ B if and only if B is Lwcc. ii) Suppose that B is a right Banach A − module. Then B∗∗A ⊆ B if and only if B is Rwcc. Corollary 2-18. Let A be a W SC Banach algebra with a BAI. If A is Arens regular and A is a left ideal in its second dual, then A is reflexive. Proof. Since A is a left ideal in A∗∗, by using proceeding corollary, A is Lwcc. Then by using Corollary 2.8 from [16], we are done. (cid:3) 13 Example 2-19. i) Let G be a compact group. Then we know that L1(G) is a left ideal L1(G)∗∗ (resp. M (G)∗∗), and so by Corollary 2-18 (resp. Corollary 2-17), L1(G) (resp. M (G)) is a Lwcc. ii) Corollary 2-18 shows that if G is a finite group. Then L1(G) and M (G) are reflex- ive. iii) Let G be a locally compact Hausdorff group. Let X be a subsemigroup of G which is the clouser of an open subset, and which contains the identity e of G. Let Z be a closed two-sided proper ideal in X with the property that X \ Z is relatively compact. Let S be the quotient of X obtained by identifying all points of Z, more formally, for x, y ∈ X write x ∼ y if either x = y or both x ∈ Z and y ∈ Z and write S = X/ ∼. By using Corollary 3.3 from [21], L1(S) is an ideal L1(S)∗∗, and so by using Corollary 2-17, L1(S) is a Lwcc. Theorem 2-20. Let A be a W SC Banach algebra with a BAI. If A is Arens regular and A is a right ideal in its second dual, then A is reflexive. Proof. Proof is similar to Corollary 2-18. (cid:3) Definition 2-21. Suppose that B is a left Banach A − module. Let (eα)α ⊆ A be left approximate identity for A. We say that (eα)α is weak∗ left approximate identity (= W ∗LAI) for B∗, if for all b′ ∈ B∗, we have πℓ(eα, b′) w∗ → b′. The definition of the weak∗ right approximate identity (= W ∗RAI) is similar. We say that (eα)α is a weak∗ approximate identity (= W ∗AI) for B∗, if B∗ has weak∗ left and right approximate identity that are equal. Ulger in [22] shows that for a Banach algebra A with a BAI, if A is a bisded ideal in its second dual, then AA∗ = A∗A and if A is Arens regular, then A∗ factors on the both side. In the following, we extend these problems for module actions with some results in group algebras. 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 [5, 16, 20], the definition of wapℓ(B) is equivalent to the following (a′′, b′′), b′i = hπt∗∗∗t for all a′′ ∈ A∗∗ and b′′ ∈ B∗∗. Thus, we can write hπ∗∗∗ ℓ ℓ (a′′, b′′), b′i wapℓ(B) = {b′ ∈ B∗ : hπ∗∗∗ (a′′, b′′), b′i = hπt∗∗∗t (a′′, b′′), b′i ℓ f or all a′′ ∈ A∗∗, b′′ ∈ B∗∗}. ℓ 14 By using [20], b′ ∈ wapℓ(B) if and only if for each sequence (an)n ⊆ A and (bm)m ⊆ B and each b′ ∈ B∗, we have lim m lim n hb′, πℓ(an, bm)i = lim n lim m hb′, πℓ(an, bm)i, whenever both the iterated limits exist. It is clear that wapℓ(B) = B∗ if and only if Z ℓ A∗∗(B∗∗) = B∗∗. Definition 2-22. Let B be a left Banach A − module and A has a BAI as (eα)α. We introduce the following subspace of B∗. ℓ(B∗) = {b′ ∈ B∗ : π∗ ℓ (b′, eα) w→ b′}. Let B be a right Banach A − module and A has a BAI as (eα)α. Such as proceeding definition, we introduce the following subspace of B∗. r (b′, eα) ℜ(B∗) = {b′ ∈ B∗ : πt∗ w → b′}. If ℓ(B∗) = B∗ (resp. ℜ(B∗) = B∗), then it is clear that (eα)α ⊆ A is a weakly right (resp. left) approximate identity for B∗. Therefore by using Lemma 2-8, ℓ(B∗) = B∗ (resp. ℜ(B∗) = AB∗) if and only if B∗ factors on the left (resp. right). Theorem 2-23. Let B be a left Banach A − module and A has a RBAI as (eα)α. Then we have the following assertions. i) ℓ(B∗) = B∗A. ii) wapℓ(B) ⊆ ℓ(B∗), if B∗ has W ∗LAI as A − module (eα)α. Proof. i) Let πℓ : A × B → B be the left module action such that πℓ(a, b) = ab for all a ∈ A and b ∈ B. Thus for every a ∈ A, b′ ∈ B∗ and b′′ ∈ B∗∗, we have hb′′, π∗ ℓ (b′′, b′), aeαi → hπ∗∗ ℓ (b′′, b′), ai ℓ (b′a, eα)i = hb′′, π∗ ℓ (b′, aeα)i = hπ∗∗ = hb′′, π∗ ℓ (b′, a)i = hb′′, b′ai. ℓ (b′a, eα) w→ b′a and so b′a ∈ ℓ(B∗). For reverse inclusion, since by It follow that π∗ Cohen,s factorization theorem, we have B∗A is a closed subspace of B∗, ℓ(B∗) ⊆ B∗A. w∗ → b′. Also the set ii) Let b′ ∈ wapℓ(B). Since (eα)α is W ∗LAI for B∗, π∗ ℓ (b′, eα) w→ b′. {π∗ ℓ (b′, eα) : α ∈ I} is relatively weakly compact which implies that π∗ ℓ (b′, eα) Corollary 2-24. Let B be a left Banach A − module and A has a RBAI and ZA∗∗(B∗∗) = B∗∗. If B∗ has W ∗LAI, then B∗ factors on the left. (cid:3) Example 2-25. i) Let G be a finite group. Then, by using proceeding corollary, we conclude that L∞(G)M (G) = RU C(G) = L∞(G). ii) Let G be an infinite compact group. Since L1(G) has a BAI, L∞(G) has a W ∗BAI. By using Proposition 4.4 from [22], we know that wap(L1(G)) = C(G). Thus, by us- ing proceeding theorem, C(G) ⊆ ℓ(L∞(G)). 15 Theorem 2-26. Let B be a right Banach A − module and A has a LBAI as (eα)α. Then we have the following assertions. i) ℜ(B∗) = AB∗. ii) wapr(B) ⊆ ℜ(B∗), if B∗ has W ∗RAI A − module (eα)α. Proof. Proof is similar to Theorem 2-23. (cid:3) Corollary 2-27. Let B be a right Banach A − module and A has a BAI as (eα)α. If B factors on the left, then wapr(B) ⊆ ℜ(B∗). Proof. Let b ∈ B and b′ ∈ B∗. Since B factors on the left, there are a ∈ A and y ∈ B such that b = ya. Then hπt∗ r (b′, eα), bi = hb′, πt r(eα, b)i = hb′, πr(b, eα)i = hb′, πr(ya, eα)i It follows that πt∗ → b′. Then by using Theorem 2-26, we are done. r (b′, y), aeαi → hπ∗ r (b′, y), ai = hb′, yai = hb′, bi. = hπ∗ r (b′, eα) w∗ Corollary 2-28. Let B be a right Banach A − module and A has a BAI and Z t B∗∗(A∗∗) = B∗∗. If B∗ has W ∗RAI A − module, then B∗ factors on the right. (cid:3) Theorem 2-29. We have the following statements. (1) Let B be a left Banach A − module. If AB∗∗ ⊆ B, then B∗A ⊆ wapℓ(B). (2) Let B be a right Banach A − module. If B∗∗A ⊆ B, then AB∗ ⊆ wapr(B). Proof. (1) By Theorem 2-17, we know that B is Lwcc. Let a ∈ A and suppose that La is the mapping from B into itself by definition La(b) = πℓ(a, b) for each b ∈ B. By easy calculation, it is clear that (La)∗(b′) = π∗ ℓ (b′, a). Since La is weakly compact, (La)∗ is weakly compact. Then the set {(La)∗(π∗ ℓ (b′, x)) : x ∈ A1}, is weakly compact. Now let x ∈ A1 and y ∈ B. Then we have the following equality h(La)∗(π∗ ℓ (b′, x)), yi = hπ∗ = hπ∗ ℓ (b′, x), La(y)i = hπ∗ ℓ (π∗ ℓ (b′, x), a), y)i. ℓ (b′, x), πℓ(a, y)i It follows that the mapping π∗ and b′ ∈ B∗. Hence π∗ ℓ (π∗ ℓ (b′, x), a) is weakly compact for each a ∈ A ℓ (b′, x) ∈ wapℓ(B), and so B∗A ⊆ wapℓ(B). (2) Proof is similar to proceeding proof. (cid:3) 16 Example 2-30. i) Let G be a locally compact group and 1 ≤ p ≤ ∞. We know that Lp(G) is the left Banach L1(G) − module under convolution as multiplication. Assume that for all f ∈ L1(G), Lf : Lp(G) → Lp(G) be the linear mapping such that Lf g = f ∗ g whenever g ∈ Lp(G). Then, since L1(G)Lp(G)∗∗ = L1(G)Lp(G) ⊆ Lp(G) for all 1 < p < ∞, by Theorem 2-28, Lf is weakly compact. It is the same that for all µ ∈ M (G), the mapping Lµ from Lp(G) into itself with Lµf = µ ∗ f is weakly compact whenever 1 < p < ∞. ii) Let G be an infinite compact group. Then we know that L1(G) is an ideal in its second dual, L1(G)∗∗. Therefore, by using proceeding theorem and Proposition 3.3, from [22], we have LU C(G) = L∞(G)L1(G) ⊆ wap(L1(G)) and RU C(G) = L1(G)L∞(G) ⊆ wap(L1(G)). By using Proposition 4.4 from [22], since wap(L1(G)) = C(G), we conclude that LU C(G) ∩ RU C(G) ⊆ C(G), and so LU C(G) ∩ RU C(G) = C(G). Let B be a Banach A − bimodule and a′′ ∈ A∗∗. We define the locally topological centers 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′′)}. It is clear that \a′′∈A∗∗ \a′′∈A∗∗ Z t a′′ (B∗∗) = Z t A∗∗(B∗∗) = Z(πt ℓ), Za′′ (B∗∗) = ZA∗∗(B∗∗) = Z(πr). The definition of Z t b′′ (A∗∗) and Zb′′ (A∗∗) for some b′′ ∈ B∗∗ are the same. Theorem 2-31. Let B be a Banach left A − module and A has a LBAI (eα)α ⊆ A e′′ (B∗∗) = such that eα B∗∗. Then, B factors on the right with respect to A if and only if e′′ is a left unit for B∗∗. w∗ → e′′ in A∗∗ where e′′ is a left unit for A∗∗. Suppose that Z t Proof. Assume that B factors on the right with respect to A. Then for every b ∈ B, there are x ∈ B and a ∈ A such that b = ax. Then for every b′ ∈ B∗, we have hπ∗ ℓ (b′, eα), bi = hb′, πℓ(eα, b)i = hπ∗∗∗ ℓ (eα, b), b′i ℓ = hπ∗∗∗ = heαa, π∗∗ (eα, ax), b′i = hπ∗∗∗ ℓ (x, b′)i = hπ∗∗ (eαa, x), b′i ℓ (x, b′), eαai ℓ → hπ∗∗ ℓ (x, b′), ai = hb′, bi. 17 It follows that π∗ in B∗∗. Since Z t ℓ (b′, eα) w∗ e′′ (B∗∗) = B∗∗, for every b′ ∈ B∗, we have the following equality → b′ in B∗. Let b′′ ∈ B∗∗ and (bβ)β ⊆ B such that bβ w∗ → b′′ hπ∗∗∗ ℓ (e′′, b′′), b′i = lim α hb′, πℓ(eα, bβ)i lim β = lim β lim α hb′, πℓ(eα, bβ)i = lim β hb′, bβi = hb′′, b′i. It follows that π∗∗∗ Conversely, let e′′ be a left unit for B∗∗ and suppose that b ∈ B. Thren for every b′ ∈ B∗, we have (e′′, b′′) = b′′, and so e′′ is a left unit for B∗∗. ℓ hb′, π(eα, b)i = hπ∗∗∗(eα, b), b′i = heα, π∗∗(b, b′)i = hπ∗∗(b, b′), eαi = he′′, π∗∗(b, b′)i = hπ∗∗∗(e′′, b), b′i = hb′, bi. Then we have π∗ done. ℓ (b′, eα) w→ b′ in B∗, and so by Cohen factorization theorem we are (cid:3) Corollary 2-32. Let B be a Banach left A − module and A has a LBAI (eα)α ⊆ A e′′ (B∗∗) = such that eα B∗∗. Then π∗ w∗ → e′′ in A∗∗ where e′′ is a left unit for A∗∗. Suppose that Z t ℓ (b′, eα) w→ b′ in B∗ if and only if e′′ is a left unit for B∗∗. Hu, Neufang and Ruan in [15], have been studied multiplier on new class of Banach algebras. They showed that how a multiplier on Banach algebra A to be implemented by an element from A is determined by its behavior on A∗ and A∗∗, respectively. In the following we study some of these discussion on module actions with some results, see [15, Theorem 3]. For a Banach algebra A, we recall that a bounded linear operator T : A → A is said to be a left (resp. right) multiplier if, for all a, b ∈ A, T (ab) = T (a)b (resp. T (ab) = aT (b)). We denote by LM (A) (resp. RM (A)) the set of all left (resp. right) multipliers of A. The set LM (A) (resp. RM (A)) is normed subalgebra of the algebra L(A) of bounded linear operator on A. Let B be a Banach left [resp. right] A − module and T ∈ B(A, B). Then T is called extended left [resp. right] multiplier if T (a1a2) = πr(T (a1), a2) [resp. T (a1a2) = πℓ(a1, T (a2))] for all a1, a2 ∈ A. We show by LM (A, B) [resp. RM (A, B)] all of the Left [resp. right] multiplier ex- tension from A into B. Let a′ ∈ A∗. Then the mapping Ta′ : a → a′a [resp. Ra′ a → aa′] from A into A∗ is left [right] multiplier, that is, Ta′ ∈ LM (A, A∗) [Ra′ ∈ RM (A, A∗)]. Ta′ is weakly compact if and only if a′ ∈ wap(A). So, we can write wap(A) as a subspace of LM (A, A∗). Theorem 2-33. Let B be a Banach A − bimodule with a BAI (eα)α ⊆ A. Then (1) If T ∈ LM (A, B), then T (a) = π∗∗∗ (2) If T ∈ RM (A, B), then T (a) = π∗∗∗ r (b′′, a) for some b′′ ∈ B∗∗. (a, b′′) for some b′′ ∈ B∗∗. ℓ 18 Proof. (1) Since (T (eα))α ⊆ B is bounded, it has weakly limit point in B∗∗. Let b′′ ∈ B∗∗ be a weakly limit point of (T (eα))α and without loss generally, take T (eα) w→ b′′. Then for every b′ ∈ B∗ and a ∈ A, we have hπ∗∗∗ r (b′′, a), b′i = lim α hb′, T (eα)ai = lim α hb′, T (eαa)i = lim α hT ∗(b′), eαai = hT ∗(b′), ai = hb′, T (a)i. It follows that π∗∗∗ r (b′′, a) = T (a). (2) Proof is similar to (1). In the proceeding theorem, if we take B = A, then we have the following statements (1) If T ∈ LM (A), then T (a) = a′′a for some a′′ ∈ A∗∗. (2) If T ∈ RM (A), then T (a) = aa′′ for some a′′ ∈ A∗∗. Definition 2-34. Let B be a Banach left A − module and b′′ ∈ B∗∗. Suppose that (bα)α ⊆ B such that bα w∗ → b′′. We define the following set (cid:3) eZb′′(A∗∗) = {a′′ ∈ A∗∗ : π∗∗∗ ZB∗∗(A∗∗) = \b′′∈B∗∗ which is subspace of A∗∗. It is clear that Zb′′(A∗∗) ⊆ eZb′′ (A∗∗), and so Zb′′ (A∗∗) ⊆ \b′′∈B∗∗ eZb′′ (A∗∗). (a′′, b′′)}, (a′′, bα) w∗ → π∗∗∗ ℓ ℓ For a Banach right A − module, the definition of eZ t a′′ (B∗∗) is similar. Theorem 2-35. Let B be a left Banach A − module and T ∈ B(A, B). Consider the following statements. ℓ (1) T = ℓb, for some b ∈ B. (2) T ∗∗(a′′) = π∗∗∗ (3) T ∗(B∗) ⊆ BB∗. Then (1) ⇒ (2) ⇒ (3). Assume that B has W SC. If we take T ∈ RM (A, B) and B has a sequential BAI, then (1), (2) and (3) are equivalent. (a′′, b′′) for some b′′ ∈ B∗∗ such that eZb′′ (A∗∗) = A∗∗. Proof. (1) ⇒ (2) Let T = ℓb, for some b ∈ B. Then T ∗∗(a′′) = ℓ∗∗ and so proof is hold. (2) ⇔ (3) Take a′′ ∈ (BB∗)⊥. Assume that b′′ ∈ B∗∗ and (bα)α ⊆ B such that bα every b′ ∈ B∗∗, we have the following equality b (a′′) = π∗∗∗ ℓ (a′′, b) for every a′′ ∈ A∗∗, w∗ → b′′. For ha′′, T ∗(b′)i = hT ∗∗(a′′), b′i = hπ∗∗∗ ℓ (a′′, b′′), b′i = lim α hπ∗∗∗ ℓ (a′′, bα), b′i 19 = lim α ha′′, π∗∗ ℓ (bα, b′)i = 0. It follows that T ∗(B∗) ⊆ BB∗. Take T ∈ RM (A, B) and suppose that B is W SC with sequential BAI. It is enough, we show that (3) ⇒ (1). Assume that (en)n ⊆ A is a BAI for B. Then for every b′ ∈ B∗, we have hb′, T (en)i − hb′, T (em)i = hT ∗(b′), en − emi = hπ∗∗ ℓ (b, b′), en − emi = hb, π∗ ℓ (b′, en − em)i = hb′, πℓ(en − em, b)i → 0. It follows that (T (en))n is weakly Cauchy sequence in B and since B is W SC, there is b ∈ B such that T (en) w→ b in B. Let a ∈ A. Then for every b′ ∈ B∗, we have hb′, πℓ(a, b)i = hπ∗ ℓ (b′, a), b)i = lim hπ∗ ℓ (b′, a), T (en)i = lim n = lim n n hb′, πℓ(a, T (en))i = lim n hb′, T (aen)i hT ∗(b′), aeni = hT ∗(b′), ai = hb′, T (a)i. Thus ℓb(a) = πℓ(a, b) = T (a). (cid:3) Example 2-36. Let G be a locally compact group. Then by convolution multiplica- tion, M (G) is a L1(G) − bimodule. Let f ∈ L1(G) and T (µ) = µ ∗ f for all µ ∈ M (G). Then T ∗(L∞(G)) ⊆ M (G)M (G)∗. Also if we take T (µ) = f ∗ µ for all µ ∈ M (G), then we have T ∗(L∞(G)) ⊆ M (G)∗M (G). 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, Banach algebra and automatic continuity, Oxford 2000. 6. H. G. Dales, A. T-M. Lau, The second dual of Beurling algebras, Mem. Amer. Math. Soc 177 (2005). 7. H. G. Dales, A. Rodrigues-Palacios, M.V. Velasco, The second transpose of a derivation, J. London. Math. Soc. 2 64 (2001) 707-721. 8. N. Dunford, J. T. Schwartz, Linear operators.I, Wiley, New york 1958. 9. M. Eshaghi Gordji, M. Filali, Arens regularity of module actions, Studia Math. 181 3 (2007), 237-254. 10. M. Eshaghi Gordji, M. Filali, Weak amenability of the second dual of a Banach algebra, Studia Math. 182 3 (2007), 205-213. 11. K. Haghnejad Azar, A. Riazi, Arens regularity of bilinear forms and unital Banach module space, arXive. math. 12. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol I 1963. 13. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol II 1970. 14. Z. Hu, M. Neufang and Z.-J. Ruan, Multipliers on new class of Banach algebras, locally compact quantum groups, and toplogical centers, Proc. London Math. Soc. 3 100 (2010) 429 - 458. 20 15. A. T. Lau, V. Losert, On the second Conjugate Algebra of locally compact groups, J. London Math. Soc. 37 (2)(1988), 464-480. 16. A. T. Lau, A. Ulger, Topological center of certain dual algebras, Trans. Amer. Math. Soc. 348 (1996), 1191-1212. 17. 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. 18. M. Neufang, Solution to a conjecture by Hofmeier-Wittstock, Journal of Functional Analysis. 217 (2004), 171-180. 19. M. Neufang, On a conjecture by Ghahramani-Lau and related problem concerning topological center, Journal of Functional Analysis. 224 (2005), 217-229. 20. J. S. Pym, The convolution of functionals on spaces of bounded functions, Proc. London Math Soc. 15 (1965), 84-104. 21. J. S. Pym, A. Saghafi Volterra-like Banach algebras and their second duals, Mh. Math, 127 (1999), 203-217. 22. A. Ulger, Arens regularity sometimes implies the RNP, Pacific Journal of Math. 143 (1990), 377-399. 23. A. Ulger, Some stability properties of Arens regular bilinear operators, Proc. Amer. Math. Soc. (1991) 34, 443-454. 24. A. Ulger, Arens regularity of weakly sequentialy compact Banach algebras, Proc. Amer. Math. Soc. 127 (11) (1999), 3221-3227. 25. A. Ulger, Arens regularity of the algebra A ⊗B, Trans. Amer. Math. Soc. 305 (2) (1988) 623-639. 26. P. K. Wong, The second conjugate algebras of Banach algebras, J. Math. Sci. 17 (1) (1994), 15-18. 27. N. J. Young The irregularity of multiplication in group algebra Quart. J. Math. Soc., 24 (2) (1973), 59-62. 28. Y. Zhang, Weak amenability of module extentions of Banach algebras Trans. Amer. Math. Soc. 354 (10) (2002), 4131-4151. Department of Mathematics, University of Mohghegh Ardabili, Ardabil, Iran Email address: [email protected]
1501.00464	1	1501	2015-01-02T19:22:23	The solution to the Kadison--Singer Problem: yet another presentation	[ "math.FA" ]	In the summer of 2013 Marcus, Spielman, and Srivastava gave a surprising and beautiful solution to the Kadison--Singer problem. The current presentation is slightly more didactical than other versions that have appeared since; it hopes to contribute to a thorough understanding of this amazing proof.	math.FA	math	THE SOLUTION TO THE KADISON -- SINGER PROBLEM: YET ANOTHER PRESENTATION DAN TIMOTIN ABSTRACT. In the summer of 2013 Marcus, Spielman, and Srivastava gave a surprising and beautiful solution to the Kadison -- Singer problem. The current presentation is slightly more didactical than other versions that have appeared since; it hopes to contribute to a thorough understanding of this amazing proof. 5 1 0 2 n a J 2 ] . A F h t a m [ 1 v 4 6 4 0 0 . 1 0 5 1 : v i X r a CONTENTS 1. Introduction 2. The Kadison -- Singer problem 2.1. Pure states 2.2. The Kadison -- Singer conjecture 2.3. The Paving Conjecture 3. Intermezzo: what we will do next and why 3.1. General plan 3.2. Sketch of the proof 4. Analytic functions and univariate polynomials 4.1. Preliminaries 4.2. Nice families 5. Several variables: real stable polynomials 5.1. General facts 5.2. The barrier function 6. Characteristic and mixed characteristic polynomials 6.1. Mixed characteristic polynomial 6.2. Decomposing in rank one matrices and the characteristic polynomial 7. Randomisation 7.1. Random matrices and determinants 7.2. Probability and partitions 8. Proof of the Paving Conjecture 9. Final Remarks References 1 2 2 2 3 4 6 6 6 7 7 9 10 10 12 15 15 17 18 18 20 21 23 24 2 DAN TIMOTIN 1. INTRODUCTION The Kadison -- Singer Problem has been posed in [7] in the fifties, probably in relation to a statement of Dirac concerning the foundations of quantum mechanics. It has soon acquired a life of its own. On one hand, there have been several notable attempts to prove it. On the other hand, it has been shown that it is equivalent to various problems in Hilbert space theory, frame theory, geometry of Banach spaces, etc. However, for five decades the problem has remained unsolved. It is therefore very remarkable that in 2013 a proof has been given by Marcus, Spelman and Sri- vastava in [10]. The methods used were rather unexpected; moreover, they had shown their strength in some totally unrelated areas (Ramanujan graphs). They also have a very elementary flavour: most of the proof is based on a delicate analysis of the behavior of polynomials in one or several variables. In the year and a half that has passed a better grasp of the proof has been achieved, most no- tably through Terence Tao's entry in his blog [14] (but see also [11]). It still remains an astonishing piece of research, obtaining spectacular results on a long standing conjecture through some not very complicated and apparently unrelated arguments. The purpose of these notes is to contribute towards a better understanding of the MSS proof. There is of course no pretense to any originality: the content is essentially in [10], with some supplementary simplification due to [14] (and occasionally to [15]). But we have tried to make it more easy to follow by separating clearly the different steps and emphasizing the main arguments; also, in various places we have gone into more details than in the other presentations. It is to be expected that the methods of [10] might lead to new fruitful applications, and so it seemed worth to analyze them in detail. It is clear from the above that the notes concentrate on the MMS proof, so there will be very little about the Kadison -- Singer problem itself and about the plethora of research that had evolved in the last fifty years on its relations to other domains. In particular, with one exception that we need to use (the paving conjecture), we will not discuss the different reformulations and equivalent statements that have been obtained. For all these matters, one may consult former beautiful presentations, as for instance [4]. We will give in the next section a brief presentation of the original problem, as well as of another assertion, the paving conjecture, which has been shown soon afterwards to imply it. The description of the remaining part of the paper is postponed to Section 3, where the reader will have a general overview of the development of the proof. These notes have been written for a series of lectures given in December 2014 at the Indian Statis- tical Institute in Bangalore, in the framework of the meeting Recent Advances in Operator Theory and Operator Algebras. We thank B.V.R. Bath, J. Sarkar, and V.S. Sunder for the excellent work done in organizing the workshop and the conference, as well as for the invitation to present the lectures. 2. THE KADISON -- SINGER PROBLEM 2.1. Pure states. The material in this subsection is contained in standard books on C∗ -algebras (see, for instance, [6]). THE SOLUTION TO THE KADISON -- SINGER PROBLEM 3 We denote by B(H) the algebra of all bounded linear operators on the Hilbert space H. A C∗- algebra A ⊂ B(H) is a norm closed subalgebra of B(H), closed to the operation of taking the adjoint, and containing the identity. A state on a C∗-algebra A is a linear continuous map φ : A → C, which is positive (meaning that φ(a∗a) ≥ 0 for any a ∈ A), and such that φ(I) = 1. One proves then that kφk = 1 and that φ satisfies the Cauchy -- Schwarz inequality (2.1) φ(b∗a)2 ≤ φ(a∗a)φ(b∗b) for all a, b ∈ A. The set S(A) of all states on A is a convex, w∗-compact subset of A∗. A state φ is called pure if it is an extreme point of S(A). Example 2.1. If A is commutative, then by Gelfand's Theorem it is isomorphic to C(X), the algebra of continuous functions on the compact space X of all characters (multiplicative homomorphisms) χ : A → C. The dual C(X)∗ is formed by all Borel measures on X, and S(C(X)) is the set of probability measures on X. Pure states are precisely Dirac measures. In particular (and this is a fact that we will use below) a pure state on a commutative C∗-algebra is multiplicative. Example 2.2. If A = B(H), ξ ∈ H, and kξk = 1, then one can prove that φξ (T) := hTξ, ξi is a pure state. This fact will not be used in the sequel. By a theorem of Krein (see, for instance, [12, Ch.I.10]) any state φ on a C∗-algebra A extends to a state φ on B(H). The set Kφ of all extensions of φ is a convex w∗-compact subset of B(H)∗. Lemma 2.3. If φ is a pure state on A ⊂ B(H), then the extreme points of Kφ are pure states of B(H). Proof. Suppose φ is an extreme point of Kφ. If φ = 1 φ = 1 ψ1 = ψ2 = φ. 2 (ψ1 + ψ2), with ψ1, ψ2 states on B(H), then 2 (ψ1A + ψ2A). Since φ is pure, we must have ψ1A = ψ2A = φ, so ψ1, ψ2 ∈ Kφ, and therefore (cid:3) Consequently, a pure state φ on A has a unique extension to a state on B(H) if and only if it has a unique pure extension to a state on B(H). 2.2. The Kadison -- Singer conjecture. From now on we will suppose that the Hilbert space H is ℓ2 = ℓ2(N) and we will consider matrix representations of operators on B(ℓ2) with respect to the usual canonical basis of ℓ2. We define D to be the C∗-algebra of operators on ℓ2 whose matrix is diagonal. Note that the map diag : B(ℓ2) → D which sends an operator T to the diagonal operator having the same diagonal entries is continuous, positive, of norm 1. We may now state the Kadison -- Singer Problem: Does any pure state on D extend uniquely to a state on B(ℓ2)? Although Kadison and Singer originally thought a negative answer to this question as more prob- able, in view of its eventual positive answer we will state the conjecture in the affirmative form. Kadison -- Singer Conjecture (KS). Any pure state on D extends uniquely to a state on B(ℓ2). The first thing to note is that any state φ ∈ S(D) has a "canonical" extension to S(B(ℓ2)), given DAN TIMOTIN 4 by (2.2) φ(T) = φ(diag(T)). So the problem becomes whether φ is or not the unique extension of φ to B(ℓ2). extension of φ and T ∈ B(ℓ2), then If ψ is another ψ(T − diag T) = ψ(T) − φ(diag T) = ψ(T) − φ(T). So ψ = φ if and only if ψ(T − diag T) = 0 for any T ∈ B(ℓ2), which is equivalent to say that ψ(T) = 0 for any T ∈ B(ℓ2) with diag T = 0. As a consequence, we have the following simple lemma: Lemma 2.4. (KS) is true if and only if any extension ψ ∈ S(B(ℓ2)) of a pure state on A satisfies diag T = 0 =⇒ ψ(T) = 0. In fact, pure states of D can be described more precisely. Indeed, being a commutative algebra, D is isomorphic to C(X) (as noted in Example 2.1). One can identify X precisely: it is βN, the Stone- Cech compactification of N. We do not need this fact, but will use only a simple observation. Lemma 2.5. If φ is a pure state on D and P ∈ D is a projection, then φ(P) is either 0 or 1. Proof. It has been noted above (see Example 2.1) that φ is multiplicative. Then φ(P) = φ(P2) = φ(P)2, whence φ(P) is either 0 or 1. (cid:3) Remark 2.6. As hinted in the introduction, although in the original paper [7] there is no mention of quantum mechanics, in subsequent papers the authors state as source for the problem the work of Dirac on the foundation of quantum mechanics [5]. For some comments on this, see Subsection 9.1 below. 2.3. The Paving Conjecture. Instead of dealing directly with the Kadison -- Singer conjecture, we in- tend to prove a statement about finite dimensional matrices, which is usually known as Anderson's paving conjecture [1]. We use the notation Dm to indicate diagonal m × m matrices and diagm the corresponding map from Mm(C) to Dm. Paving Conjecture (PC). For any ǫ > 0 there exists r ∈ N such that the following is true: For any m ∈ N and T ∈ B(Cm) with diagm T = 0, there exist projections Q1, . . . , Qr ∈ Dm, with i=1 Qi = Im, and ∑r for all i = 1, . . . , r. kQiTQik ≤ ǫkTk A diagonal projection Q ∈ Dm has its entries 1 or 0, so it is defined by a subset S ⊂ {1, . . . , m}. i=1 Qi = Im correspond to partitions {1, . . . , m} = Thus diagonal projections Q1, . . . Qr ∈ Dm with ∑r S1 ∪ · · · ∪ Sr, Si ∩ Sj = ∅ for i 6= j. It is important that in the statement of (PC) the number r does not depend on m. This allows us to deduce from (PC) a similar statement, in which Cm is replaced with the whole ℓ2, is also true. We formulate this as a lemma. THE SOLUTION TO THE KADISON -- SINGER PROBLEM 5 Lemma 2.7. If (PC) is true, then for any ǫ > 0 there exists r ∈ N such that, for any T ∈ B(ℓ2) with diag T = 0 one can find projections Q1, . . . , Qr ∈ D, with ∑r kQiTQik ≤ ǫkTk i=1 Qi = I, and for all i = 1, . . . , r. Proof. Embed Cm canonically into ℓ2 on the first m coordinates and denote by Em the corresponding orthogonal projection. For T ∈ B(ℓ2) denote Tm = EmTEm. Applying (PC), one finds diagonal projections Q , such that ∑r , . . . , Q (m) (m) r (m) 1 Now, diagonal projections in B(ℓ2) can be identified with subsets of N, and therefore with ele- ments in the compact space {0, 1}N. In this compact space any sequence has a convergent subse- quence; therefore a diagonal argument will produce an increasing subsequence of positive integers mk, such that for each i = 1, . . . , r we have Q (mk) i → Qi for some Qi. We have i=1 Q i = Im and kQ (m) i TmQ (m) i k ≤ ǫkTmk. r ∑ i=1 Qi = lim k→∞ r ∑ i=1 Q (mk) i = lim k→∞ Imk = I. If ξ, η ∈ ℓ2 are vectors with finite support, then ξ, η ∈ Cdk for some k, and then hQiTQiξ, ηi = hTQiξ, Qiηi = hTmk Q (mk) i ξ, Q (mk) i ηi = hQ (mk) i Tmk Q (mk) i ξ, ηi ≤ kQ (mk) i Tmk Q (mk) i kkξkkηk ≤ ǫkTkkξkkηk. (cid:3) The Paving Conjecture is actually equivalent to the Kadison -- Singer Conjecture, but we will need (and prove) only one of the implications. Proposition 2.8. The Paving Conjecture implies the Kadison -- Singer Conjecture. Proof. Fix ǫ > 0, and suppose that r satisfies the conclusion of Lemma 2.7. Take a pure state ψ ∈ S(B(ℓ2)) and an operator T ∈ B(ℓ2) with diag T = 0. By Lemma 2.4 we have to show that ψ(T) = 0. Let Qi be the diagonal projections associated to T by Lemma 2.7. By Lemma 2.5, ψ(Qi) = φ(Qi) i=1 φ(Qi), it follows that there exists some i0 for which is 0 or 1 for each i. Since 1 = φ(I) = ∑r φ(Qi0) = 1, while φ(Qi) = 0 for i 6= i0. If i 6= i0, then (2.1) implies ψ(QiR) ≤ ψ(Q∗i Qi)ψ(R∗R) = ψ(Qi)ψ(R∗R) = 0, and similarly ψ(RQi) = 0 for all R ∈ B(ℓ2). Therefore ψ(T) = r ∑ i=1 r ∑ j=1 ψ(QiTQj) = ψ(Qi0TQi0). But the projections Qi have been chosen such as to have kQi0 TQi0k ≤ ǫkTk, so ψ(T) ≤ kQi0 TQi0k ≤ ǫkTk. Since this is true for any ǫ > 0, it follows that ψ(T) = 0, and the proposition is proved. (cid:3) 6 DAN TIMOTIN 3. INTERMEZZO: WHAT WE WILL DO NEXT AND WHY 3.1. General plan. As noted above, we intend to prove the Paving Conjecture. The proof will lead us on an unexpected path, so we explain here its main steps. The Paving Conjecture asks us to find, for a given matrix T, diagonal projections Qi that achieve certain norm estimates (namely, kQiTQik ≤ ǫkTk). Among the different ways to estimate the norm, the proof in [10] choses a rather unusual one: it uses the fact that the norm of a positive operator is its largest eigenvalue. So we have to consider characteristic polynomials of matrices -- in fact, the largest part of the proof is dedicated to estimating roots of such polynomials. (Although it has nothing to do with (KS), one should note the added benefit that we find a way to control with no extra effort all eigenvalues of the matrix, not only the largest one.) On the other hand, to achieve this control we need to make an unexpected detour: though the characteristic polynomial depends on a single variable, in order to control it one has to go through multivariable polynomials and to use the theory of real stable polynomials as developed by Borcea and Brand´en [2]. This may seem unnatural, but it should be mentioned that Borcea and Brand´en have already obtained through their methods spectacular results, in particular solving long-standing conjectures in matrix theory that also seemed at first sight to involve just a single complex variable [2, 3]. So maybe one should not be so surprised after all. A second feature of the proof is its use, at some point, of a random space. After obtaining certain results about eigenvalues of usual matrices, suddenly random matrices appear on the scene. In fact, the use of randomness is not really essential; it rather provides a convenient notation for computing averages. As noted in the previous section, to prove (PC) we need to find a partition of a finite set {1, . . . , m} into r subsets with certain properties. The random space eventually considered is finite; its elements are all different such partitions, and no subtle probability is used: all decompositions are assumed to be equally probable. What we will achieve eventually is an estimate on the average of the largest eigenvalue, which will lead to an individual estimate for at least one point of the random space -- that is, for one partition. This will be the desired partition. 3.2. Sketch of the proof. We summarize here the development of the proof. As announced above, we intend to discuss the eigenvalues of positive matrices, which are roots of the characteristic poly- nomial. So we need some preparation concerning polynomials and their roots; this is done first in one variable in Section 4. The main result here is Theorem 4.9, that shows that certain families of polynomials have roots that behave unexpectedly well with respect to averages. This will be used in Section 7 to link eigenvalues of random matrices to their averages. But we have to go to polynomials in several variables, namely real stable polynomials, which are defined by a condition on their roots. Section 5 is dedicated to real stable polynomials; after present- ing their main properties, we are especially interested in some delicate estimate on the location of the roots, which is done through an associated function called the barrier function. The properties of the barrier function represent the most technical and not very transparent part of the proof. The main thing to be used in the sequel is Theorem 5.8, that puts some restriction on the roots of a real stable polynomial. THE SOLUTION TO THE KADISON -- SINGER PROBLEM 7 We apply these facts to characteristic polynomials in Section 6. The voyage through several vari- ables done for polynomials has a correspondent here in the introduction of the mixed characteristic polynomial, which depends on several matrices. It happens to be the restriction to one variable of a real stable polynomial, and so Theorem 5.8 can be used in Theorem 6.4 to bound the roots of a mixed characteristic polynomial. Further, this bound translates in a bound for a usual characteristic poly- nomial in the particular case when the matrices have rank one, since then the mixed characteristic polynomial is precisely the characteristic polynomial of their sum. Section 7 introduces random matrices; as discussed above, the probability space in view is that of all possible partitions. The main result, Theorem 7.2, uses the results of Section 4 to show that for a sum of independent random matrices of rank one, the eigenvalues of its average yield estimates for the averages of its eigenvalues, and thus for the eigenvalues of at least one point of the probability space. In particular, applying this fact in conjunction with the bound on eigenvalues obtained in Section 6, we will obtain a partition with certain norm properties in Theorem 7.5. Finally, this last fact is put to good use in Section 8 to obtain a proof of the Paving Conjecture. The first step, that uses Theorem 7.5, obtains for orthogonal projections a quantitative version of (PC). To go from projections to general operators is well known since several decades and may be done in different ways. Here we use a dilation argument taken from [15] to obtain the Paving Conjecture for selfadjoint matrices; going to general matrices is then immediate. 4. ANALYTIC FUNCTIONS AND UNIVARIATE POLYNOMIALS 4.1. Preliminaries. The next theorem in complex function theory is a consequence of Cauchy's ar- gument principle. Theorem 4.1. Suppose ( fn) is a sequence of analytic functions on a domain D ⊂ C, which converges uni- formly on compacts to the function f 6≡ 0. If Γ is a simple contour contained in D such that f has no zeros on Γ, then there is n0 ∈ N such for n ≥ n0 the number of zeros of fn and of f in the interior of Γ coincide. The next corollary is usually called Hurwitz's Theorem if m = 1. The general case follows simply by induction (exercise!). Corollary 4.2. Suppose pn(z1, . . . , zm) are polynomials in m variables, such that pn → p uniformly on compacts in some domain D ⊂ Cm. If pm has no zeros in D for all m, then either p is identically zero, or it has no zeros in D. If f is a polynomial of degree n with all coefficients and all roots real, we denote its roots by ρn( f ) ≤ · · · ≤ ρ1( f ). Corollary 4.3. Suppose ps(z) = ∑n i=1 ai(s)zi, with ai : I → R continuous functions on an interval I ⊂ R, an(s) 6= 0 on I. If ps has real roots for all s ∈ I, then the roots ρ1(ps), . . . , ρn(ps) are continuous functions of s ∈ I. Proof. We use induction with respect to n. The case n = 1 is obvious. Then, for a general n, we prove first that ρ1(ps) is continuous, say in s0 ∈ I. Take ǫ > 0, and suppose also that ps0(s0 ± ǫ) 6= 0. By 8 DAN TIMOTIN continuity of ai, ps(s0 ± ǫ) 6= for s sufficiently close to s0, and so ps(z) 6= 0 for z on the circle Γ of diameter [s0 − ǫ, s0 + ǫ] (since all ps have real roots). By Theorem 4.1 all ps have at least one root inside Γ for s sufficiently close to s0. A similar argument, using a circle at the right of s0 + ǫ, shows that the ps have no roots larger than b. It follows that ρ1(ps) ∈ (a, b) for s close to s0. If we write now ps(z) = (z − ρ1(ps))qs(z), then qs has degree n − 1 and continuous coefficients, so its roots are continuous by the induction hypothesis. But we have ρi(p) = ρi−1(q) for i ≥ 2. Remark 4.4. Even without the assumption that the roots are real, one can prove that there exist continuous functions ρi(s) : I → C, i = 1, . . . , n, such that the roots of ps are ρ1(s), . . . , ρn(s) for all s ∈ I. The proof is more involved; see, for instance, [8, II.5.2]. (cid:3) We prove next two lemmas about polynomials with real coefficients and real roots. Lemma 4.5. Suppose the polynomial p of degree n has real coefficients, real roots, and the leading term positive. Moreover, assume that there exist real numbers an+1 < an < · · · < a1 such that ρj(p) ∈ [aj+1, aj] for all j = 1, . . . , n. Then (−1)j−1 p(aj) ≥ 0 for all j = 1, . . . , n. In other words, p changes signs (not necessarily strictly) on each of the intervals [aj+1, aj]. Proof. We will use induction with respect to n. For n = 1 the claim is obviously true. Suppose it is true up to n − 1, and let p be a polynomial of degree n as in the statement of the lemma. There are two cases to consider. Suppose first that the roots of p are exactly all points aj except some aj0 . Then p has only simple roots, so it changes signs in each of them. As p(x) > 0 for x > a1, we have p(x) < 0 on (a2, a1), etc, up to (−1)j0−1 p(x) > 0 on (aj0+1, aj0−1). Therefore (−1)j0−1 p(aj0) > 0; the other inequalities are trivial. In the remaining case, there is at least one root α of p that is not among the points aj; suppose α ∈ (aj0, aj0−1). If p(z) = (z − α)q(z), then q has degree n − 1 and satisfies the hypotheses of the lemma with respect to the points aj with j 6= 0. Then p(aj) has the same sign as q(aj) for j < j0 and opposite sign for j > j0; from here it follows easily that the correct signs for q (which we know true by the induction hypothesis) produce the correct signs for p. (cid:3) Lemma 4.6. Suppose the polynomial p has real coefficients and all roots real. Then (−1)k(cid:18) d dx(cid:19)k p′ p (x) > 0 for all k ∈ N and x > ρ1(p). In particular, p′ Proof. If p(z) = ∏n p is positive, nonincreasing, and convex for x > ρ1(p). i=1(z − ρi(p)), then p′ (−1)k(cid:18) d dx(cid:19)k p′ z−ρi(p) , and p (z) = ∑n (x) = k! n ∑ i=1 1 1 i=1 p (z − ρi(p))k+1 . All terms in the last sum are positive for x > ρ1(p), so the lemma is proved. (cid:3) THE SOLUTION TO THE KADISON -- SINGER PROBLEM 9 4.2. Nice families. Suppose F = { f1, . . . fm} is a family of polynomials of the same degree n. We denote ρ+ j (F ) := max 1≤i≤m ρj( fi), ρ−j (F ) := min 1≤i≤m ρj( fi) Definition 4.7. For a family of polynomials F = { f1, . . . fm} of the same degree n a nice family iff: (1) the coefficient of the dominant term of every fj is positive; (2) every fj has all roots real; (3) for all j = 2, . . . , n we have (4.1) ρ+ j (F ) ≤ ρ−j−1(F ). The usual formulation (including [10]) is that the fis have a common interlacing. Since the actual interlacing polynomial never enters our picture, we prefer this simpler phrasing. Lemma 4.8. (i) { f1, . . . fm} is nice iff every pair { fr, fs}, r 6= s, is nice. (ii) Every subfamily of a nice family is nice. (iii) If a ∈ R, then F = { f1, . . . fm} is nice if and only if G = {(x − a) f1, . . . (x − a) fm} is nice. Proof. (i) and (ii) are immediate. For (iii), there are several cases to consider: (1) If a ∈ [ρ−j0 j0 (F ), ρ+ (F )] for some j0, then ρ±j (G) = ρ±j (F ) for j < j0, (G) = ρ+ ρ+ j0 ρ±j (G) = ρ±j−1(F ) for j > j0 + 1. (F ), ρ−j0−1(F )) for some j0, then (G) = ρ+ (F ), ρ−j0 j0 j0+1(G) = 1, ρ−j0+1(G) = ρ−j0 (F ), (2) If a ∈ (ρ+ j0 ρ±j (G) = ρ±j (F ) for j < j0, ρ±j0 ρ±j (G) = ρ±j−1(F ) for j > j0. (G) = a, The formulas in (1) are also valid if a > ρ+ all these cases one can easily check that (iii) is true. 1 (F ) (taking j0 = 1) or a < ρ−n (F ) (taking j0 = n + 1). In (cid:3) As a consequence of Lemma 4.8, in order to check that a family is nice we can always assume that it has no common zeros. The main theorem of this section is the characterization of nice families that follows. Theorem 4.9. Suppose f1, . . . , fm are all polynomials of degree n, with positive dominant coefficients. The following are equivalent: (1) F = { f1, . . . fm} is a nice family. (2) Any convex combination of f1, . . . , fm has only real roots. 10 DAN TIMOTIN If these conditions are satisfied, then for any j = 1, . . . , n we have (4.2) i for any convex combination f = ∑k tk fk. min ρj( fi) ≤ ρj( f ) ≤ max i ρj( fi) Proof. (1) =⇒ (2). We may suppose by (ii) and (iii) of Lemma 4.8 that all coefficients tk are positive and that the family has no common zeros. In particular, if we denote ρ±j = ρ±j (F ), this implies ρ−j < ρ+ We will apply Lemma 4.5 to each of the polynomials fi and the points ρ−n j ≤ ρ−j−1 for all j. < ρ+ 1 . We obtain then, for each i = 1, . . . , m, that (−1)j fi(ρ−j ) ≥ 0 for all j, and fi(ρ+ Fix j; since the family F has no common zero, at least one of fi is nonzero in ρ−j , and so (−1)j f (ρ−j ) > 0. Similarly, f (ρ+ 1 ), f changes sign (strictly), and therefore must have a root in the interior. Since there are n intervals, we have thus found n roots of f , and so all its roots are real. Moreover, we have obtained ρj( f ) > ρ−j 1 ) > 0. Therefore on each of the intervals (ρ−j , ρ−j−1), as well as on (ρ−1 , ρ+ for all j. < · · · < ρ−1 < ρ−n−1 1 ) ≥ 0. On the other hand, we might have used, in applying Lemma 4.5 to the polynomials fi, the points < ρ+ 1 . A similar argument yields then n n−1 · · · < ρ+ < · · · < ρ−1 for all j. Therefore the inequalities (4.2) are proved. 1 instead of ρ−n < ρ−n−1 < ρ+ < ρ+ ρ−n ρj( f ) < ρ+ j (2) =⇒ (1). According to Lemma 4.8 it is enough to prove the implication for two functions f1, f2, and we may also suppose that they have no common roots. Fix 2 ≤ j ≤ n; we have to prove that j ≤ ρ−j−1. Denote ft = t f1 + (1 − t) f2 (0 ≤ t ≤ 1). By Corollary 4.3 the function t 7→ ρj( ft) is ρ+ continuous on [0, 1] and takes only real values; so its values for 0 < t < 1 cover the interval (ρ−j , ρ+ j ). It follows that this interval cannot contain a root of either f1 or f2, since a common root of, say, f1 and ft is also a root of f2. j ] and [ρ−j−1, ρ+ j−1] j ), this would contra- Suppose then first that f1 and f2 have only simple roots. Then the intervals [ρ−j , ρ+ < ρ−j−1. If ρ−j−1 ∈ (ρ−j , ρ+ > ρ+ j and (4.1) is proved. have all four endpoints disjoint, and by definition ρ−j dict the conclusion of the preceding paragraph. So ρ−j−1 To obtain the general case, note first that ft has all roots simple for 0 < t < 1. Indeed, a multiple solution x of ft = 0 would also be a multiple solution of f2 t−1 . But it is easy to see (draw the f1 graph!) that then f2 t′−1 has a single root in some interval (x − ǫ, x + ǫ) for at least some t′ close f1 to t (slightly larger or slightly smaller). However, from Theorem 4.1 it follows that ft′ has more than one root in the disc z − x < ǫ, and so ft′ would not have all roots real. = t′ To end the proof, we apply the first step to fǫ and f1−ǫ (ǫ > 0), which have only simple roots. = t Then we let ǫ → 0 and use Corollary 4.3 to obtain inequality (4.1). (cid:3) 5. SEVERAL VARIABLES: REAL STABLE POLYNOMIALS 5.1. General facts. Denote H = {z ∈ C : ℑz > 0}. Definition 5.1. A polynomial p(z1, . . . , zm) is called real stable it has real coefficients and it has no zeros in Hm. In case m = 1 a real stable polynomial is a polynomial that has real coefficients and real zeros. Genuine examples in several variables are produced by the next lemma. THE SOLUTION TO THE KADISON -- SINGER PROBLEM 11 Lemma 5.2. If A1, . . . , Am ∈ Md(C) are positive matrices, then the polynomial (5.1) q(z, z1, . . . , zm) = det(zId + m ∑ i=1 zi Ai) is real stable. Proof. It is immediate from the definition that whence the coefficients of q are real. q( ¯z, ¯z1, . . . , ¯zm) = q(z, z1, . . . , zm), Assume that q(z, z1, . . . , zm) = 0, and ℑz, ℑzi > 0. Since zId + ∑m i=1 zi Ai is not invertible, there exists ξ ∈ Cd, ξ 6= 0, such that 0 = h(zId + m ∑ i=1 zi Ai)ξ, ξi = zkξk2 + m ∑ i=1 zihAiξ, ξi, and so 0 = ℑzkξk2 + m ∑ i=1ℑzihAiξ, ξi. This is a contradiction, since ℑzkξk2 > 0 and ℑzihAiξ, ξi ≥ 0 for all i. (cid:3) The next theorem gives the basic properties of real stable polynomials. Denote, for simplicity, by ∂i the partial derivative ∂ ∂zi . Theorem 5.3. Suppose p is a real stable polynomial. (i) If m > 1 and t ∈ R, then p(z1, . . . , zm−1, t) is either real stable or identically zero. (ii) If t ∈ R, then (1 + t∂m)p is real stable. Proof. (i) Obviously p(z1, . . . , zm−1, t) has real coefficients. Suppose it is not identically zero. If ℑw > 0 is fixed, then the polynomial p(z1, . . . , zm−1, w) is real stable by definition. Therefore all polynomials p(z1, . . . , zm−1, t + i n ) (for n ∈ N) are real stable. We let then n → ∞ and apply Corol- lary 4.2 to D = Hm to obtain the desired result. (ii) We may assume t 6= 0 (otherwise there is nothing to prove). Suppose (1 + t∂m)p(z1, . . . , zm) = 0 for some (z1, . . . , zm) ∈ Hm. Since p is real stable, p(z1, . . . , zm) 6= 0. The one-variable polynomial q(z) := p(z1, . . . , zm−1, z) has no roots with positive imaginary part (in particular, q(zm) 6= 0), so we may write q(z) = c n ∏ i=1 (z − wi), ℑwi ≤ 0. Therefore 0 = (1 + t∂m)p(z1, . . . , zm) = (q + tq′)(zm) = q(zm)(cid:18)1 + t q′(zm) q(zm)(cid:19) , and, since q(zm) 6= 0, 0 = 1 + t Taking the imaginary part, we obtain n ∑ i=1 1 zm − wi = 1 + t n ∑ i=1 zm − wi zm − wi2 . t n ∑ i=1 ℑwi − ℑzm zm − wi2 = 0 12 DAN TIMOTIN which is a contradiction, since t 6= 0 and ℑwi − ℑzm < 0 for all i. (cid:3) We will also need a lemma that uses a standard result in algebraic geometry, namely B´ezout's Theorem (which can be found in any standard text). Lemma 5.4. Suppose p(z, w) is a nonconstant polynomial in two variables, of degree n in w, which is irre- ducible over R. There is a finite set F ∈ C such that, if p(z0, w0) = 0 and z0 6∈ F, then: (1) the equation p(z0, w) = 0 has n distinct solutions; (2) for each of these solutions (z0, w0) we have ∂p ∂w (z0, w) 6= 0. Proof. First, if p(z, w) = q(z)wn + . . . , then the roots of q form a finite set F1. Secondly, if p is irreducible, then p and ∂p ∂w are coprime over R, and hence also over C. B´ezout's Theorem in algebraic geometry states that two curves defined by coprime equations have only a finite number of common points, so this is true about the sets defined by p(z, w) = 0 and ∂p ∂w (z, w) = 0. Let F2 be the set of the projections of these points onto the first coordinate. The set F = F1 ∪ F2 has the required properties. (cid:3) 5.2. The barrier function. Our eventual purpose in this subsection is to obtain estimates on the roots of real stable polynomials; more precisely, we want to show that a restriction on the roots of a real stable polynomial p may imply a restriction on the roots of (1 − ∂i)p (which is also real stable by Theorem 5.3). We will often use the restriction of a polynomial in m complex variables to Rm ⊂ Cm. To make things easier to follow, we will be consistent in this subsection with the following notation: z, w will belong to Cm (and corresponding subscripted letters in C), while x, y, s, t will be in Rm (and corresponding subscripted letters in R). If x = (x1, . . . , xm) ∈ Rm, then {y ≥ x} will denote {y = (y1, . . . , ym) ∈ Rm : yi ≥ xi for all i = 1, . . . , m}. The main tool is a certain function associated to p called the barrier function, whose one-dimensional p ; if p(x) > 0 p(x) = ∂i(log p)(x). The argument of the barrier function will always version has already been met in Lemma 4.6. It is defined wherever p 6= 0 by Φi it can also be written as Φi actually be in Rm. p = ∂i p The connection of the barrier function with our problem is given by the simple observation that if p(x) 6= 0 and (1 − ∂i)p(x) = 0, then Φi < 1 does not contain zeros of (1 − ∂i)p. To determine such sets, the basic result is the next lemma, which is a multidimensional extension of Lemma 4.6. p(x) = 1. So, in particular, a set on which 0 ≤ Φi p Lemma 5.5. Suppose x ∈ Rm, and p(z1, . . . , zm) is a real stable polynomial that has no roots in {y ≥ x}, then (−1)k ∂k ∂zk j p(x′) ≥ 0 Φi for any k ≥ 0, 1 ≤ i, j ≤ m, and x′ ≥ x. ing and convex on [0, ∞]. In particular, if ej is one of the canonical basis vectors in Cm, then t 7→ Φi p(x + tej) is positive, nonincreas- THE SOLUTION TO THE KADISON -- SINGER PROBLEM 13 Proof. The assertion reduces to Lemma 4.6 for m = 1 or for k = 0; and also for i = j, since then fixing all variables except the ith reduces the problem to the one variable case. In the general case, it is enough to do it for y = x, since if p has no roots in {y ≥ x}, then it has no roots in {y ≥ x′} for all x′ ≥ x. By fixing all variables except i and j, we may assume that m = 2, i = 1, j = 2, k ≥ 1. Moreover, we may also assume that p > 0 on {y ≥ x} (otherwise we work with −p, since Φi So we have to prove that, if p(z1, z2) is a real stable polynomial which has no zeros in {y1 ≥ p = Φi −p). x1, y2 ≥ x2}, then 0 ≤ (−1)k ∂k ∂zk 2 Φ1 p(x1, x2) = (−1)k ∂k ∂zk 2 (cid:18) ∂ ∂z1 log p(cid:19) (x1, x2) = ∂ ∂z1 (−1)k ∂k ∂zk 2 log p! (x1, x2). We will in fact prove that the map t 7→ (−1)k ∂k ∂zk 2 log p(t, x2) is increasing for t ≥ x1. It is enough to achieve this for p irreducible over R, since, if p = p1 p2 is real stable and has no roots in {y ≥ x}, then the same is true for p1 and p2, and obviously (−1)k ∂k ∂zk 2 log p(t, x2) = (−1)k ∂k ∂zk 2 log p1(t, x2) + (−1)k ∂k ∂zk 2 log p2(t, x2). Suppose then that p is irreducible. For t ≥ x1 fixed, the polynomial p(t, z) is real stable, and thus has all roots real; denote them, as in Section 4, by ρ1(t) ≥ · · · ≥ ρn(t). Applying to p Lemma 5.4, take t ≥ x1 that does not belong to the finite set F therein. The functions ρi(t) are therefore differentiable in t, and we have (5.2) Therefore (5.3) p(t, z) = c(t) n ∏ i=1 (z − ρi(t)) (−1)k ∂k ∂zk 2 log p! ((−1)k ∂k ∂zk 2 log p)(t, x2) = (−1)k ∂k log(z − ρi(t))!(cid:12)(cid:12)(cid:12)z=x2 ∑ i=1 2 n ∂zk (k − 1)! (x2 − ρi(t))k . = − n ∑ i=1 If t ≥ x1, we cannot have ρi(t) ≥ x2 since then (t, ρi(t)) would be a root of p in {y ≥ x}, contrary to the assumption. Thus x2 − ρi(t) > 0, and in order to show that the function in (5.3) is increasing, it is enough to show that t 7→ ρi(t) is decreasing for t ≥ x1 and all i. Now all ρis are differentiable for t ≥ x1, t /∈ F. To show that they are decreasing, it is enough to show that ρ′i(t) ≤ 0 for such t. Suppose then that there exists i ∈ {1, . . . , n} and t ≥ x1 such that ρ′i(t) > 0; let s = ρi(t). Since ∂p (t, s) 6= 0, we may apply the (complex) implicit function ∂z2 theorem in a neighborhood of (t, s) (in C2). We obtain that the solutions of p(z1, z2) = 0 therein are of the form (z1, g(z1)) for some locally defined analytic function of one variable g, which by analytic 14 DAN TIMOTIN continuation has to be an extension of ρ to a complex neighborhood of t. So g′(t) = ρ′i(t), and in the neighborhood of t we have g(z1) = t + ρ′i(t)(z1 − t) + O(z1 − t2). If ℑz1 > 0 and small, one also has ℑg(z1) > 0. We obtain thus the zero (z1, g(z1)) of p in H2, contradicting the real stability of p. This ends the proof of the lemma. (cid:3) Corollary 5.6. Suppose x ∈ Rm, and p is a real stable polynomial, without zeros in {y ≥ x}. Then Φj Φj p(x) for any y ≥ x and j = 1, . . . , m. Proof. If p has no zeros in {y ≥ x}, obviously it has no zeros in {y ≥ x′} for any x′ ≥ x. Therefore, by Lemma 5.5, the function t 7→ Φj p(x′ + tei) is nonincreasing on [0, ∞) for any i = 1, . . . , m. We have p(y) ≤ then Φj p(x1, . . . , xm) ≥ Φj p(y1, x2, . . . , xm) ≥ Φj p(y1, y2, x3, . . . , xm) ≥ · · · ≥ Φj p(y1, . . . , ym) (cid:3) The main monotonicity and convexity properties of Φi a restriction on the location of zeros of (1 − ∂j)p. As noted above, we will use the condition Φj p are put to work in the next lemma to obtain p < 1, but in a more precise variant which will lends itself to iteration. Lemma 5.7. Let x ∈ Rm, and p a real stable polynomial, without zeros in {y ≥ x}. Suppose also that Φj p(x) + for some j ∈ {1, . . . , m} and δ > 0. Then: (i) (1 − ∂j)p has no zeros in {y ≥ x}. (ii) For any i = 1, . . . , m we have 1 δ ≤ 1 Proof. By Corollary 5.6 we have Φi (1−∂j)p(x + δej) ≤ Φi p(x). ∂j p(y) p(y) so ∂j p(y) 6= p(y), or (1 − ∂j)p(y) 6= 0. = Φ(y) ≤ Φ(x) ≤ 1 − 1 δ < 1, To prove (ii), note first that (1 − ∂j)p = p(1 − Φj p), whence log[(1 − ∂j)p] = log p + log(1 − Φj p), so, by differentiating, Φi (1−∂j)p = Φi p − ∂iΦj p 1 − Φj p . The required inequality becomes then ∂iΦj 1 − Φj (5.4) − p(x + δej) p(x + δej) ≤ Φi p(x) − Φi p(x + δej). By Corollary 5.6 we have Φj p(x + δej) ≤ Φj p(x) ≤ 1 − 1 δ , THE SOLUTION TO THE KADISON -- SINGER PROBLEM 15 or 1 p(x + δej) ≤ δ. 1 − Φj Further on, p has no zeros in {y ≥ x + δej}, so Lemma 5.5 (applied in x + δej) implies, in particular, that −∂iΦj p(x + δej) ≥ 0, whence − ∂iΦj 1 − Φj p(x + δej) p(x + δej) ≤ −δ∂iΦj p(x + δej). To prove (5.4), it is then enough to show that −δ∂iΦj p(x + δej) ≤ Φi p(x) − Φi p(x + δej). Using ∂iΦj p(x + δej) = ∂jΦi p(x + δej), the inequality can be written Φi p(x + δej) ≤ Φi p(x) + δ∂jΦi p(x + δej). This, however, is an immediate consequence of the convexity of the function t 7→ Φi has been proved in Lemma 5.5. p(x + tej), that (cid:3) Finally, the next theorem is the main result of this section that we will use in the sequel. Theorem 5.8. Let x ∈ Rm, and p a real stable polynomial, without zeros in {y ≥ x}. Suppose also that for some δ > 0 and j = 1, . . . , m. Then Φj p(x1, . . . , xm) + 1 δ ≤ 1 m ∏ i=1 (1 − ∂i)p has no zeros in {y ≥ x + δ}, where δ := (δ, . . . , δ) ∈ Rm. Proof. The proof follows by applying Lemma 5.7 successively to j = 1 and x, then to j = 2 and x + δe1, etc. (cid:3) 6. CHARACTERISTIC AND MIXED CHARACTERISTIC POLYNOMIALS 6.1. Mixed characteristic polynomial. We intend now to apply the results of Section 5 to polynomi- als related to matrices. Our final goal is to estimates eigenvalues; that is, roots of the characteristic polynomial. But we will first consider another polynomial, attached to a tuple of matrices. Definition 6.1. If A1, . . . , Am ∈ Md(C), then the mixed characteristic polynomial of the matrices Ai is defined by the formula (6.1) µ[A1, . . . , Am](z) = m ∏ i=1 (1 − ∂i) det(zId + m ∑ i=1 zi Ai)(cid:12)(cid:12)(cid:12)z1=···=zm=0 . It is easily seen that if we fix m − 1 of the matrices A1, . . . , Am, then µ[A1, . . . , Am](z) is of degree 1 in the entries of the remaining matrix. Indeed, if we develop the determinant that enters (6.1), then any term that contains a product of, say, k entries of Aj has also the factor zk j . If we apply (1 − ∂j), we are left with zk−1 , and if k ≥ 2 this terms becomes 0 if zj = 0. j 16 DAN TIMOTIN Example 6.2. For one or two matrices we have µ[A1](z) = zd − zd−1 Tr A1 if m = 1, µ[A1, A2](z) = zd − zd−1(Tr A1 + Tr A2) + zd−2(Tr A1 Tr A2 − Tr(A1 A2)) if m = 2. In the general case, the coefficients of µ[A1, . . . , Am](z) are certain expressions in the traces of monomials in A1, . . . , Am that are well known in the invariant theory of matrices (see [13]). The results in Section 5 have consequences for the mixed characteristic polynomials. Theorem 6.3. Suppose A1, . . . , Am ∈ Md(C) are positive matrices. Then µ[A1, . . . , Am](z) has only real roots. Proof. We have seen in Lemma 5.2 that the polynomial q defined by (5.1) is real stable. But µ[A1, . . . , Am] is obtained from q by first applying (1 − ∂i) for i = 1, . . . , m and then specializing to z1 = · · · = zm = 0. By Theorem 5.3, these operations preserve the real stable character. So µ[A1, . . . , Am] is a real stable polynomial of one variable, which means exactly that it has real roots. (cid:3) Remember Jacobi's formula for the derivative of the determinant of an invertible matrix: (6.2) (det M(t))′ det M(t) = Tr(cid:16)M(t)−1M′(t)(cid:17) . Theorem 6.4. Suppose A1, . . . , Am ∈ Md(C) are positive matrices, such that ∑m for each i = 1, . . . , m. Then any root of µ[A1, . . . , Am] is smaller than (1 + √ǫ)2. i=1 Ai = Id and Tr Ai ≤ ǫ Proof. The polynomial p(z) := det( m ∑ i=1 zi Ai). is real stable, being the specialization of the polynomial q in (5.1) to z = 0. (t, . . . , t) ∈ Cd, then, for y ≥ t we have ∑m i=1 tAi = tId. Therefore ∑m and p(y) 6= 0. We may apply Jacobi's formula (6.2) in order to compute the barrier function Φj i=1 yi Ai ≥ ∑m If t > 0 and t := i=1 yi Ai is invertible, p, and we obtain In particular, if t > 0, then Φj p(x1, . . . , xm) = Tr(( m ∑ i=1 zi Ai)−1 Aj). Φj p(t, . . . , t) = Tr(t−1 Aj) ≤ ǫ t . It follows then from Theorem 5.8 that, if we t, δ > 0 are such that ǫ i=1(1 − ∂i)p has no zeros in {y ≥ (t + δ, . . . , t + δ)}. The choice t = ǫ + √ǫ, δ = 1 + √ǫ (which can easily be shown to be optimal) yields t + δ = (1 + √ǫ)2, and therefore p has no roots y with yi ≥ (1 + √ǫ)2} for all i. δ ≤ 1, then ∏m t + 1 THE SOLUTION TO THE KADISON -- SINGER PROBLEM 17 Now, using the relation ∑m i=1 Ai = 1, one obtains µ[A1, . . . , Am](z) = = = m ∏ i=1 m ∏ i=1 m ∏ i=1 (1 − ∂i) det(zId + m ∑ i=1 (1 − ∂i) det( m ∑ i=1 (1 − ∂i)p(z, z, . . . , z), zi Ai)(cid:12)(cid:12)z1=···=zm=0 wi Ai)(cid:12)(cid:12)w1=···=wm=z which cannot be zero if z ≥ (1 + √ǫ)2. Therefore all roots of µ are smaller than (1 + √ǫ)2. (cid:3) 6.2. Decomposing in rank one matrices and the characteristic polynomial. In an important par- ticular case the mixed characteristic polynomial coincides with a usual characteristic polynomial. Remember this is defined, for A ∈ Md(C), by pA(z) = det(zId − A). Lemma 6.5. Suppose B, A1, . . . , Am ∈ Md(C), and A1, . . . , Am have rank one. Then the polynomial (z1, . . . , zm) 7→ det(B + z1 A1 + · · · + zm Am) is of degree ≤ 1 separately in each variable. Proof. By fixing all the variables except one, we have to show that, for any B, A1 ∈ Md(C), A1 of rank one, the function z 7→ det(B + zA1) is of degree at most 1. This is obvious if we choose a basis in which the first vector spans the image of A1, and we develop the determinant with respect to the first row. (cid:3) Suppose now p(z1, . . . , zm) is a polynomial of degree ≤ 1 separately in each variable. Then p is equal to its Taylor expansion at the origin of order 1 in each variable, that is: p(z1, . . . , zm) = ∑ ǫi∈{0,1} cǫ1,...,ǫm zǫ1 1 · · · zǫm m , with Therefore cǫ1,...,ǫm = ∂ǫ1 1 · · · ∂ǫm m p(w1, . . . , wm)(cid:12)(cid:12)w1=···=wm=0. p(z1, . . . , zm) = ∑ zǫ1 1 · · · zǫm m ∂ǫ1 1 · · · ∂ǫm m p(w1, . . . , wm)(cid:12)(cid:12)w1=···=wm=0 ǫi∈{0,1} m ∏ i=1 = (1 + zi∂i)p(w1, . . . , wm)(cid:12)(cid:12)w1=···=wm=0. In the case of the polynomial in Lemma 6.5, this formula becomes det(B + m ∑ i=1 zi Ai) = m ∏ i=1 (1 + zi∂i) det(B + m ∑ i=1 wi Ai)(cid:12)(cid:12)w1=···=wm=0. In fact, we are interested by this last formula precisely when B = zId and all zi = −1. We obtain then the next theorem. 18 DAN TIMOTIN Theorem 6.6. Suppose A1, . . . , Am ∈ Md(C) have rank one. If A = A1 + · · · + Am, then pA(z) = µ[A1, . . . , Am](z). Remark 6.7. The mixed characteristic polynomial and the usual characteristic polynomial are invari- ant with respect to a change of basis. So, although we have spoken about matrices for convenience, the statements of Theorems 6.4 and 6.6 can be stated for A1, . . . , Am ∈ L(V), where L(V) denotes the space of linear operators on the finite dimensional vector space V. 7. RANDOMISATION 7.1. Random matrices and determinants. Let (Ω, p) be a finite probability space. If X is a random variable on Ω, the expectation (or average) of E(X) is defined, as usually, by E(X) := ∑ ω∈Ω p(ω)X(ω). If X1, . . . , Xm are independent random variables, then, in particular, we have (7.1) E(X1 · · · Xm) = E(X1) · · · E(Xm). We will use random matrices A(ω) ∈ Md(C), whose entries are random variables; then E(A) is the matrix whose entries are the expectations of the corresponding entries of A. The random matrices A1, A2 are called independent if any entry of A1 is independent of every entry of A2. Also, when we say that a random matrix A(ω) has rank one, this means that A(ω) has rank one for any ω ∈ Ω. The characteristic polynomial pA of a random matrix A is also a random variable, by which we mean that its coeficients are random variables. Then the polynomial E(pA) has as coefficients the expectations of the coefficients of A. Theorem 7.1. Suppose A1(ω), . . . , Am(ω) are independent rank one random matrices in Md(C), and A = A1 + · · · + Am. Then E(pA) = µ[E(A1), . . . , E(Am)]. Proof. By Theorem 6.6 we have, for each ω ∈ Ω, pA(ω) = µ[A1(ω), . . . , Am(ω)]. By taking expecta- tions, E(pA) = E(µ[A1(ω), . . . , Am(ω)]). Now independence of Ais combined with the multilinearity of µ implies that E(µ[A1(ω), . . . , Am(ω)]) = µ[E(A1), . . . , E(Am)], which ends the proof. (cid:3) We can say more if we also assume that the Ais are all positive. Theorem 7.2. Suppose A1(ω), . . . , Am(ω) are independent rank one positive random matrices in Md(C), and A = A1 + · · · + Am. Then, for any j = 1, . . . , d, we have min ω∈Ω ρj(pA(ω)) ≤ ρj(µ[E(A1), . . . , E(Am)]) ≤ max ω∈Ω ρj(pA(ω)). THE SOLUTION TO THE KADISON -- SINGER PROBLEM 19 Proof. We prove only the left hand side inequality; the right hand side is similar. It is enough to show that for any i = 1, . . . we have (7.2) min ω∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), Ai(ω), E(Ai+1), . . . , E(Am)]) ≤ min ω∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), E(Ai), E(Ai+1), . . . , E(Am)]). Indeed, for i = m the left hand side coincides with minω∈Ω ρj(pA(ω)) by Theorem 6.6, while for i = 1 the right hand side is precisely ρj(µ[E(A1), . . . , E(Am)]). The chain of inequalities corresponding to i = 1, 2, . . . , m proves then the theorem. Fix then i and ω ∈ Ω, and consider the family of polynomials fω′ = µ[A1(ω), . . . , Ai−1(ω), Ai(ω′), E(Ai+1), . . . , E(Am)], ω′ ∈ Ω. Take cω′ ≥ 0, with ∑ω′∈Ω cω′ = 1. By the multilinearity of the mixed characteristic polynomial, we have ∑ ω′∈Ω cω′ fω′ = ∑ ω′∈Ω cω′µ[A1(ω), . . . , Ai−1(ω), Ai(ω′), E(Ai+1), . . . , E(Am)] cω′ Ai(ω′), E(Ai+1), . . . , E(Am)]. = µ[A1(ω), . . . , Ai−1(ω), ∑ ω′∈Ω Since the last polynomial is the mixed characterstic polynomial of positive matrices, it has all roots : ω′ ∈ Ω} is a nice family. Moreover, if real by Theorem 6.3. It follows by Theorem 4.9 that { fω′ we take as coefficients of the convex combination cω′ = p(ω′) and so for any j = 1, . . . , d we have ∑ω′∈Ω cω′Ai(ω′) = E(Ai). Applying the last part of Theorem 4.9 it follows that for any j = 1, . . . , d, ρj(µ[A1(ω), . . . , Ai−1(ω), Ai(ω′), E(Ai+1), . . . , E(Am)]) min ω′∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), E(Ai), E(Ai+1), . . . , E(Am)]). Taking the minimum with respect to ω ∈ Ω, we obtain (7.3) min ω∈Ω min ω′∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), Ai(ω′), E(Ai+1), . . . , E(Am)]) min ω∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), E(Ai), E(Ai+1), . . . , E(Am)]). Suppose the minimum in the left hand side is attained in ω = ω0, ω′ = ω′0. By independence of the random matrices Ai, we have p({σ ∈ Ω : A1(σ) = A1(ω0), . . . , Ai−1(σ) = Ai−1(ω0), Ai(σ) = A(ω′0)}) p({σ ∈ Ω : A1(σ) = A1(ω0), . . . , Ai−1(σ) = Ai−1(ω0)})p({σ ∈ Ω : Ai(σ) = A(ω′0)}) > 0. Taking σ0 ∈ Ω in the set in the left hand side, we obtain min σ∈Ω ρj(µ[A1(σ), . . . , Ai−1(σ), Ai(σ), E(Ai+1), . . . , E(Am)]) ≤ ρj(µ[A1(σ0), . . . , Ai−1(σ0), Ai(σ0), E(Ai+1), . . . , E(Am)]) = ρj(µ[A1(ω0), . . . , Ai−1(ω0), Ai(ω′0), E(Ai+1), . . . , E(Am)]) = min ω∈Ω ρj(µ[A1(ω), . . . , Ai−1(ω), Ai(ω′), E(Ai+1), . . . , E(Am)]). This inequality, together with (7.3), implies (7.2), finishing thus the proof of the theorem. min ω′∈Ω (cid:3) 20 DAN TIMOTIN Remark 7.3. The point of Theorem 7.2 is that the middle term might be easier to compute or to estimate. But, since the matrices E(A1), . . . , E(Am) are not of rank one, Theorem 6.6 does not apply, and µ[E(A1), . . . , E(Am)] is not a characteristic polynomial. However, Theorem 7.2 tells us that its roots can be used to estimate the eigenvalues of A(ω) for at least some value of ω. Corollary 7.4. Let A1(ω), . . . , Am(ω) be independent rank one positive random matrices in Md(C), and A = A1 + · · · + Am. Suppose E(A) = Id and E(Tr Ai) ≤ ǫ for some ǫ > 0. Then min ω∈ΩkA(ω)k ≤ (1 + √ǫ)2. Proof. Since Tr(E(Ai)) = E(Tr Ai) ≤ ǫ, the matrices E(A1), . . . , E(Am) satisfy the hypotheses of Theorem 6.4, all roots of µ[E(A1), . . . , E(Am)] are smaller than (1 + √ǫ)2. By Theorem 7.2 we obtain, in particular, min ω∈ΩkA(ω)k = min ω∈Ω ρ1(pA(ω)) ≤ ρ1(µ[E(A1), . . . , E(Am)]) ≤ (1 + √ǫ)2. (cid:3) 7.2. Probability and partitions. The last theorem of this section gets us closer to the paving conjec- ture. It is here that we make the connection between the probability space and the partitions. Let us first note that, similarly to Remark 6.7, one can see that the independence condition is not affected by a change of basis. So in Theorem 7.2 and in Corollary 7.4 we may assume that Ai take values in L(V) for some finite dimensional vector space V. This observation will be used in the proof of the next theorem. Theorem 7.5. Suppose A1, . . . , Am ∈ Md(C) are positive rank one matrices, such that ∑m i=1 Ai = Id and kAik ≤ C for all i = 1, . . . , m. Then for every positive integer r there exists a partition S1, . . . , Sr of {1, . . . , m}, such that for any j = 1, . . . , r. ∑ i∈Sj (cid:13)(cid:13)(cid:13) √r Ai(cid:13)(cid:13)(cid:13) ≤(cid:18) 1 + √C(cid:19)2 Proof. Since the purpose is to find a partition with certain properties, we will take as a random space Ω precisely the space of all partitions of {1, . . . , m} in r sets, with uniform probability p. Such a partition is determined by an element ω = (ω1, . . . , ωm), where ωj ∈ {1, . . . , r}, and Sj = {k : ωk = j}; so Ω = {1, . . . , r}m. Also, the different coordinates, that is the maps ω 7→ ωi, are independent scalar random variables on Ω. We consider the space V := Cd ⊕ · · · ⊕ Cd and define the random matrices Ai (i = 1, . . . , m) with values in L(V) by (7.4) Ai(ω) = 0 ⊕ · · · ⊕ rAi ⊕ · · · ⊕ 0, where rAi appears in position ωi. These are independent random matrices (since the coordinates ωi are independent). If we fix 1 ≤ j ≤ r, then ωi = j with probability 1/r, and so rAi appears in position j with probability 1/r. Therefore E(Ai) = rAi ⊕ · · · ⊕ rAi = Ai ⊕ Ai ⊕ · · · ⊕ Ai. 1 r 1 r THE SOLUTION TO THE KADISON -- SINGER PROBLEM 21 If A = A1 + · · · + Am, then E(A) = m ∑ i=1 E(Ai = m ∑ i=1 (Ai ⊕ Ai ⊕ · · · ⊕ Ai) = IV. Since Tr Ai(ω) = r Tr Ai for all i, we have E(Tr Ai) = E(r Tr Ai) = rE()kAik) ≤ rC. Corollary 7.4 yields the existence of ω ∈ Ω such that kA(ω)k ≤ (1 + √rC)2. But, according to (7.4), we have A(ω) = r ∑ ωi=1 Ai! ⊕ r ∑ ωi=2 Ai! ⊕ · · · ⊕ r ∑ ωi=r Ai.! We define then Sj = {i : ωi = j}. It follows that kr ∑i∈Sj Aik ≤ (1 + √rC)2 for all j, and dividing by r ends the proof of the theorem. (cid:3) 8. PROOF OF THE PAVING CONJECTURE We may now proceed to the proof of the paving conjecture; from this point on all we need from the previous sections is Theorem 7.5. We first deal with orthogonal projections. For such operators the paving conjecture is trivially verified (exercise: if P is an orthogonal projection and diag P = 0, then P = 0). But we will prove a quantitative version of the paving conjecture, in which one does not assume zero diagonal. Lemma 8.1. Suppose P ∈ Mm(C) is an orthogonal projection. For any r ∈ N there exists diagonal orthogo- nal projections Q1, . . . , Qr ∈ Mm(C), with ∑r j=1 Qj = Im, such that for all j = 1, . . . , r. kQj PQjk ≤(cid:18) 1 √r +qk diag Pk(cid:19)2 Proof. Denote by V the image of P, and d = dim V. Let (ei)m rank one positive operators Ai by Ai(v) = hv, P(ei)iP(ei). We have (8.1) kAik ≤ kP(ei)k2 = hP(ei), eii ≤ k diag Pk, i=1 a basis in Cm, and define on V the and, for v ∈ V, (8.2) Consequently, whence (8.3) hAiv, vi = hv, P(ei)ihP(ei), vi = hv, Peii2 = hv, eii2. m ∑ i=1 h Aiv, vi = m ∑ i=1hAiv, vi = m i=1hv, eii2 = kvk2, ∑ r ∑ i=1 Ai = IV. 22 DAN TIMOTIN From (8.1) and (8.3) it follows that Ai satisfy the hypotheses of Theorem 7.5, with C = k diag Pk∞. There exists therefore a partition S1, . . . , Sr of {1, . . . , m}, such that +qk diag Pk∞(cid:19)2 √r ∑ i∈Sj (cid:13)(cid:13)(cid:13) Ai(cid:13)(cid:13)(cid:13) ≤(cid:18) 1 for any j = 1, . . . , r. Define then Qj ∈ Mm(C) to be the diagonal orthogonal projection on the span of {ei : i ∈ Sj}. Then kQjPQjk = kQjP(QjP)∗k = kQj Pk2 = kQVk2. But, if v ∈ V, then, applying (8.2), kQjvk2 = ∑ i∈Sj hv, eii2 = ∑ i∈Sj So kQjPQjk = kQVk2 ≤ k ∑ i∈Sj and the lemma is proved. Aik · kvk2. i∈Sj hAiv, vi = h(cid:0) ∑ Aik ≤(cid:18) 1 √r i∈Sj Ai(cid:1)v, vi ≤ k ∑ +qk diag Pk∞(cid:19)2 , (cid:3) Theorem 8.2 (The paving conjecture). For any ǫ > 0 there exists r ∈ N such that, for any m ∈ N and T ∈ Mm(C) with diag T = 0 there exist diagonal orthogonal projections Q1, . . . , Qr ∈ Mm(C), with ∑r j=1 Qj = Im, such that for all j = 1, . . . , r. kQjTQjk ≤ ǫkTk Proof. Suppose first that T = T∗ and kTk ≤ 1. The 2m × 2m matrix 2 (Im − T2)1/2 Im+T 1 2 ! Im−T 2 P = 1 2 (Im − T2)1/2 2 , . . . , 1 + 1√2(cid:17)2 is an orthogonal projection and diag P = (cid:16) 1 − 1 ≤ ǫ. It follows from Lemma 8.1 that there exist diagonal projections Q′′1 , . . . , Q′′r ∈ M2m(C) with + 1√2(cid:17)2 i=1 Q′′i = I2d and kQ′′i PQ′′i k ≤(cid:16) 1√r ∑r Let Q′′i = Qi + Q′i be the decomposition of Q′′i in the diagonal projections corresponding to the first 2(cid:17). Choose r large enough to have 2(cid:16) 1√r for all i = 1, . . . , r. m and the last m vectors of the basis of C2m. So ∑r i=1 Qi = ∑r i=1 Qi = Im and, for each i = 1, . . . , m, (8.4) kQi(I + T)Qik ≤ 2(cid:18) 1 √r + 1 √2(cid:19)2 , kQ′i(I − T)Q′ik ≤ 2(cid:18) 1 √r + 1 √2(cid:19)2 . The first inequality implies that Qi(I + T)Qi ≤ 2(cid:16) 1√r − Qi ≤ QiTQi ≤"2(cid:18) 1 √r + + 1√2(cid:17)2 √2(cid:19)2 1 Qi, so − 1# Qi ≤ ǫQi (8.5) (the left inequality being obvious). Similarly, the second inequality in (8.4) yields (8.6) − ǫQ′i ≤ "1 − 2(cid:18) 1 √r + 1 √2(cid:19)2# Q′i ≤ Q′iTQ′i ≤ Q′i. THE SOLUTION TO THE KADISON -- SINGER PROBLEM 23 If we define Qij = QiQ′j (i, j = 1, . . . , r), then ∑r i,j=1 Qij = Im, and it follows from (8.5) and (8.6) that −ǫQij ≤ QijTQij ≤ ǫQij, or QijTQij ≤ ǫ. The theorem is thus proved for T a selfadjoint contraction, and it is immediate to extend it to arbitrary selfadjoint matrices. If we take now an arbitrary T ∈ Mm(C), with diag T = 0, we may write it as T = A + iB, with A, B selfadjoint, kAk, kBk ≤ kTk, and diag A = diag B = 0. Applying the first step, one finds diagonal projections Q′1, . . . , Q′r, Q′′1 , . . . , Q′′r ∈ Mm(C), with ∑r 2kTk and kQ′′i BQ′′i k ≤ ǫ i,j=1 Qi,j = Im, and kQijTQijk ≤ ǫkTk for i, j = 1, . . . , r. 2kTk for i = 1, . . . , r. If we define Qij = Q′iQ′′j , then ∑r (cid:3) i=1 Q′i = ∑r i=1 Q′′i = Im, kQ′i AQ′ik ≤ ǫ By writing carefully the estimates in the proof, one sees also that we may take r of order ǫ−4. 9. FINAL REMARKS 1. As noted in Remark 2.6, there is a connection between the Kadison -- Singer problem and quan- tum mechanics. We will give here a very perfunctory account. In the von Neumann picture of quan- tum mechanics, states (in the common sense) of a system correspond to states φ (in the C∗-algebra sense) of B(H), while observables of the system correspond to selfadjoint operators A ∈ B(H). The value of an observable in a state is precisely φ(A). A maximal abelian C∗-algebra A ⊂ B(H) corresponds to a maximal set of mutually compatible observables. If the extension of any pure state on A to a state on B(H) is unique, then one can say that the given set of observables determines completely all other observables. This seems to have been assumed by Dirac implicitely. Now, there are various maximal abelian subalgebras of B(H), but the problem can essentially be reduced to two different basic types: continuous (that are essentially isomorphic to L∞ acting as multiplication operators on L2) and discrete (that are isomorphic to D acting in ℓ2). The main topic of the original paper [7] is to prove that extension of pure states is not unique in general for continuous subalgebras. They suspected that the same thing happens for the discrete case, but could not prove it, and so posed it as an open problem. 2. We have said in the introduction that there are many statements that had been shown to be equivalent to (KS), besides (PC) that we have used in an essential way. We have thus, among others: (1) Weaver's conjectures in discrepancy theory. The original proof in [10] goes actually through one of these; the shortcut using (PC) is due to Tao [14]. (2) Feichtinger's conjecture in frame theory. (3) Bourgain -- Tzafriri conjecture. All these conjectures have in fact different forms, weaker or stronger variants, etc -- a detailed account may be found in [4]. It is worth noting that up to 2013 most specialists believed that they are not true, and that a counterexample will eventually be found. So it was a surprise when all these statements were simultaneously shown true by [10]. 24 DAN TIMOTIN 3. The method used in [10] is even stronger than described above. Actually, its first application was to a completely different problem in graph theory: the existence of certain infinite families of so-called Ramanujan graphs [9] (see also [11] for an account). 4. The most tedious proof in the above notes is that of Lemma 5.5. The original argument in [10] is more elegant, but uses another result of Borcea and Brand´en [3] that represents real stable poly- nomials in two variables as determinants of certain matrices -- a kind of converse to Lemma 5.2. The direct argument we use appears in [14]. REFERENCES [1] J. Anderson: Extensions, restrictions and representations of states on C∗-algebras, Trans. AMS 249 (1979), 195 -- 217. [2] J. Borcea, P. Brand´en: Applications of stable polynomials to mixed determinants: Johnson's conjectures, unimodality, and symmetrized Fischer products, Duke Math. Journal 143 (2008), 205 -- 223. [3] J. Borcea, P. Brand´en: Multivariate P ´olya -- Schur classification problems in the Weyl algebra, Proc. Lond. Math. Soc. (3) 101 (2010), 73-104. [4] P.G. Cassazza, M. Fickus, J.C. Tremain, E. Weber: The Kadison -- Singer problem in mathematics and engineering: a de- tailed account, Operator Theory, Operator Algebras, and Applications, 299 -- 355, Contemp. Math. 414, AMS, 2006. [5] P.A.M. Dirac: The Principles of Quantum Mecahnics, Oxford University Press, 1958. [6] R.V. Kadison, J.R. Ringrose: Fundamentals of the Theory of Operator Algebras, Academic Press, 1983. [7] R.V. Kadison, I.M. Singer: Extensions of pure states, American Jour. Math. 81 (1959), 383 -- 400. [8] T. Kato, Perturbation Theory for Linear Operators, Springer-Verlag, 1980. [9] A.Marcus, D.A. Spielman, N. Srivastava: Interlacing families I: Bipartite Ramanujan graphs of all degrees, arXiv:1304.4132. [10] A.Marcus, D.A. Spielman, N. Srivastava: Interlacing families II: Mixed characteristic polynomials and the Kadison -- Singer problem, arXiv:1306.3969. [11] A.Marcus, D.A. Spielman, N. Srivastava: Ramanujan graphs and the solution to the Kadison -- Singer Problem, arXiv:1408.4421v1, to appear in Ann. of Math. [12] M.A. Naimark: Normed Algebras, Wolters -- Noordhoff, Groningen, 1972. [13] C. Procesi: The invariant theory of n × n matrices, Adv. in Math. 19 (1976), 306 -- 381. [14] T. Tao: Real stable polynomials and the Kadison -- Singer problem, https://terrytao.wordpress.com/tag/kadison-singer-problem/. [15] A. Valette: Le probl`eme de Kadison -- Singer, arXiv:1409.5898v.
1504.02445	1	1504	2015-04-09T19:31:03	Rolewicz-type chaotic operators	[ "math.FA" ]	In this article we introduce a new class of Rolewicz-type operators in l_p, $1 \le p < \infty$. We exhibit a collection F of cardinality continuum of operators of this type which are chaotic and remain so under almost all finite linear combinations, provided that the linear combination has sufficiently large norm. As a corollary to our main result we also obtain that there exists a countable collection of such operators whose all finite linear combinations are chaotic provided that they have sufficiently large norm.	math.FA	math	Rolewicz-type chaotic operators D. Bongiornod,∗, U.B. Darjie, L. Di Piazzaf aDipartimento di Energia, Ingegneria dell'informazione e Modelli matematici (DEIM), bDepartment of Mathematics, University of Louisville, Louisville, KY 40292, USA cDipartimento di Matematica e Informatica, University of Palermo, Via Archirafi 34, University of Palermo, Palermo, Italy 90123 Palermo, Italy Keywords: operator chaotic operators, hypercyclic operators, lineable, Rolewicz ✩ This research was supported by the grant Cori 2013 of the University of Palermo. The second Author thanks the hospitality of the Department of Mathematics of University of Palermo ∗Corresponding author Email addresses: [email protected] (D. Bongiorno), [email protected] (U.B. Darji), [email protected] (L. Di Piazza) Preprint submitted to J. Math. Anal. and Appl. March 5, 2018 ✩ Rolewicz-type chaotic operators D. Bongiornod,∗, U.B. Darjie, L. Di Piazzaf dDipartimento di Energia, Ingegneria dell'informazione e Modelli matematici (DEIM), eDepartment of Mathematics, University of Louisville, Louisville, KY 40292, USA fDipartimento di Matematica e Informatica, University of Palermo, Via Archirafi 34, 90123 University of Palermo, Palermo, Italy Palermo, Italy Abstract In this article we introduce a new class of Rolewicz-type operators in ℓp, 1 ≤ p < ∞. We exhibit a collection F of cardinality continuum of operators of this type which are chaotic and remain so under almost all finite linear combinations, provided that the linear combination has sufficiently large norm. As a corollary to our main result we also obtain that there exists a countable collection of such operators whose all finite linear combinations are chaotic provided that they have sufficiently large norm. Introduction Hypercyclic operators are generalizations of cyclic operators which have been studied in operator theory for many years. If X is a Banach space and T : X → X is a bounded linear operator, we say that T is hypercyclic if there exists x ∈ X such that Orb(x, T ), the orbit of x under T , is dense in X. It is a well- known result of Rolewicz [21] that no finite dimensional Banach space admits a hypercyclic operator. However, every infinite dimensional Banach space admits a hypercyclic operator. This was shown independently by Ansari [1] and Bernal- Gonz`alez [6]. From the dynamical systems point of view, hypercyclic operators are closely connected to transitive maps. A continuous self-mapping f of a metric space X is transitive if for all non-empty open sets U, V of X, there exists n ≥ 1 such that f n(U ) ∩ V 6= ∅. If X is a separable metric space without isolated points, then f is transitive if and only if Orb(x, f ) is dense in X for some x ∈ X. Hence, in the case of a separable Banach space, a linear operator is hypercyclic ✩ This research was supported by the grant Cori 2013 of the University of Palermo. The second Author thanks the hospitality of the Department of Mathematics of University of Palermo ∗Corresponding author Email addresses: [email protected] (D. Bongiorno), [email protected] (U.B. Darji), [email protected] (L. Di Piazza) Preprint submitted to J. Math. Anal. and Appl. ✩ if and only if it is transitive. The notion of chaos in the sense of Devaney [12] also applies in the setting of Banach spaces. We say that a bounded linear operator T : X → X is chaotic in the sense of Devaney if T is transitive and the set of periodic points of T is dense in X. In this article, when we say an operator is chaotic, we mean that the operator is chaotic in the sense of Devaney. Whereas every infinite dimensional Banach space admits a hypercyclic operator, Bonet, Mart´ınez-Gim´enez, and Peris [10] showed that the separable hereditarily indecomposable Banach space constructed by Gowers and Maury [15] admits no chaotic operators. When one encounters an exotic or an unusual object, a natural question one asks is how many such objects are there. In topological setting, one asks if the collection of objects in questions forms a meager or comeager set. In every sep- arable infinite dimensional Hilbert space, the set of hypercyclic operators forms a nowhere dense set with respect to the norm topology [4, Thm 2.24]. Hence, in some sense there are very few chaotic operators in infinite dimensional Hilbert space. It is easy to construct two chaotic operators whose sum is not chaotic. In fact, Grivaux [16] showed that every bounded linear operator on a separable infinite dimensional Hilbert space is the sum of two chaotic operators. Hence being chaotic is far from being stable under finite linear combinations. How- ever, in recent years a new algebraic notion of largeness has been popularized and exploited [19], [2], [3]. The idea is to exhibit a large algebraic structure in a setting where no algebraic structure is apparent. If X is a Banach space and A ⊆ X, we say that A is lineable if there is a vector space V ⊆ X of dimen- sion continuum such that every non-zero element of V is in A. Recently, many sets of classical importance are shown to be lineable. For example, the set of nowhere differentiable functions is lineable [18]. For other examples and survey on the subject refer to [11] [13]. Clearly, the set of chaotic operators cannot be lineable as no operator with norm less than one is hypercyclic. However, as we will see from the main result of this article, some type of algebraic largeness can be inferred. The notion of shift is a basic yet fundamental concept in dynamical systems. In symbolic dynamics they are of utmost importance and have been extensively studied. In 1969 Rolewicz [21] made an important observation. If one considers the shift T : ℓ2 → ℓ2, defined by T (x1, x2, . . . ) = (x2, x3, . . . ), then for all λ > 1, λT is a transitive operator on ℓ2 whose set of periodic points is dense in ℓ2. Operators of these type are called Rolewicz operators. They have been generalized to what are called weighted shift operators or weighted Rolewicz operators. For a suitable choice of sequence {wn}, the shift operator with the weight sequence {wn} is defined by T (x1, x2, . . . ) = (w1x2, w2x3, . . . ). Weighted shift operators have been well-studied. In this note, we view the shift operator in a different light. We think of the shift operator as the projection of the input vector on coordinates {2, 3, . . . }. In general, given an increasing 3 function f : N → N, we define Tf : ℓp → ℓp, 1 ≤ p < ∞, by Tf (x1, x2, . . . ) = (xf (1), xf (2), . . . ). Hence, the usual shift operator equals Tf , where f (n) = n + 1. As in the Rolewicz operator, if λ > 1, then we have that λTf is chaotic. We call operators of this type Rolewicz-type operators.. We study and exploit these operators. We i=1 ciTfi is chaotic. The precise formulation is given in Theorem 7. As corollaries, we obtain the following results which show that set of chaotic operators contains large algebraic structure. prove that for suitable f1, . . . , ft, c1, . . . , ct, and sufficiently large λ, λPt Corollary 11 There exists an infinite family T of chaotic operators such i=1 ciTfi i=1 ciTfi is that for all T1, T2, . . . , Tt ∈ T and c1, c2, . . . , ct ∈ R the operator λPt is chaotic for sufficiently large λ, provided that the following sum Pt not the zero operator. Corollary 14 There exists a family T of cardinality continuum of chaotic operators such that for almost all (c1, c2, . . . , ct) ∈ Rt and T1, T2, . . . , Tt ∈ T the operator λPt i=1 ciTfi is chaotic for sufficiently large λ. We would like to point out that recently other notions of chaos have enjoyed attention in linear dynamics as well. These notions include Li-Yorke chaos and distributional chaos. A characterization of Li-Yorke chaos in Banach spaces in terms of irregular vectors was given in [5] and later extended to the setting of Fr´echet spaces in [8]. As an easy consequence of these results one obtains that all hypercyclic operators are Li-Yorke chaotic. Hence, all operators constructed in our main results are Li-Yorke chaotic. A characterization of distributional chaos in the setting of Fr´echet spaces in terms of distributionally irregular vectors was given in [9]. We conjecture that all operators constructed in our main results are distributionally chaotic. Indeed, we conjecture that if f1, . . . ft are strictly increasing pairwise almost disjoint functions and c1, . . . , ct are real numbers, i=1 ciTfi being chaotic (in the sense of Devaney) implies that T is distributionally chaotic. Indeed, for weighted backward shifts it was shown in [20] that Devaney chaos implies distributional chaos. We conjecture that using the characterization of distributional chaos given in [9], one may be able to prove a large class of Devaney chaotic operators, which include weighted shifts as well as Rolewicz-type operators defined in this article, are distributionally chaotic. then T = Pt 1. Notations and preliminaries In what follows X is a separable infinite dimensional Banach space. Let T : X → X be a linear bounded operator. For x ∈ X the orbit of x under T is Orb(T, x) = {x, T (x), T 2(x), ...} where T n = T ◦ T ◦ ... ◦ T is the nth iterate of T obtained by composing T with itself n times. A point x ∈ X is a periodic point of T if T n(x) = x for some n ≥ 1. Also T is transitive if for any two non-empty open sets U, V in X, there exists an integer n ≥ 1 such that T n(U ) ∩ V 6= ∅. Devaney's Definition of Chaos ([12]) We say that a linear bounded op- erator T : X → X is chaotic if 4 (i) T is transitive; (ii) the periodic points of T are dense in X. It is well known that, in a complete metric space with no isolated points, being transitive is equivalent (via the Baire Category Theorem) to having a point with dense orbit, which it is equivalent to having a dense Gδ set of points each of which has a dense orbit. We refer the reader to texts [4] and [17] for further information on linear chaos. From now on we will consider linear bounded operators in lp, for some fixed i=1 xip) p where 1 ≤ p < ∞. If ~x = (x1, x2, . . . ) ∈ lp we denote by k~xkp := (P∞ and by k~xk∞ = sup{xi : i = 1, 2, . . . }, respectively its lp and sup norm. 1 p We start with some preliminary facts and notations. Let N = {1, 2, . . . }. We say that a map f : N → N is increasing if f (i) < f (j) whenever i < j. Let f : N → N and g : N → N be given. We say that f and g are almost disjoint if the set {k : f (k) = g(k)} is finite and f (k) 6= g(k′) if k 6= k′. For a given increasing function f : N → N, denote by Tf : lp → lp the linear bounded operator defined by Tf (x1, x2, . . . ) := (xf (1), xf (2), . . . ). Let f1, f2, . . . , ft be increasing functions and let c1, c2, . . . , ct ∈ R. Then for each σ ∈ {1, 2, . . . , t}n we define and fσ := fσ(n) ◦ fσ(n−1) ◦ · · · ◦ fσ(1) cσ := cσ(n) · cσ(n−1) · · · · · cσ(1). If σ, τ are in {1, 2, . . . , t}n and {1, 2, . . . , t}m, respectively, then στ is just the concatenation of σ followed by τ , and, in general, if S ⊆ {1, 2, . . . , t}n then σS = {στ : τ ∈ S}. We use σ to denote the length of σ. In particular fσfτ = fστ and cσcτ = cστ . For notational convenience we let f∅ = id and c∅ = 1. If S ⊂ {1, 2, . . . , t}r, then we let c(S) := Pσ∈S cσ. Now we define a collection of equivalence relations. Fix r, i ∈ N. Let σ, τ ∈ {1, 2, . . . , t}r. We say that σ ∼i τ if and only if fσ(i) = fτ (i). We note that ∼i is an equivalence relation on {1, 2, · · · , t}r. If σ ∈ {1, 2, . . . , t}r, then [σ]i = {τ ∈ {1, 2, . . . , t}r : σ ∼i τ }. We note that in general there is no ambiguity in writing [σ]i as r = σ. We let [σ]i the number of elements of [σ]i and A(r, i) = {[σ]i : σ ∈ {1, . . . , t}r}, 5 and ♯(r, i) the cardinality of A(r, i). We say that the constants c1, c2, . . . , ct ∈ R satisfy the non-zero condition at level m, m ∈ N, if the following condition holds: c([σ]i) 6= 0, ∀σ ∈ {1, 2, . . . , t}≤m, and 1 ≤ i ≤ m. We note that, in particular, this condition implies that all of c1, . . . , ct are non- zero and hence cσ 6= 0 for all σ ∈ {1, . . . , t}r, for all r ∈ N. Moreover, when this condition is satisfied we let γ := γ(c1, . . . , ct) := min{1, c([σ]i) : σ ∈ {1, . . . , t}≤m, 1 ≤ i ≤ m} > 0. Below we collect some basic propositions which we will need for our main results. Throughout the rest of this section we fix a sequence f1, . . . , ft of increasing functions which are pairwise almost disjoint. We also fix m ∈ N such that fi(k) 6= fj(k) for all 1 ≤ i < j ≤ t and k ≥ m. And c1, . . . , ct are some fixed constants in R. Proposition 1. Let σ, τ ∈ {1, 2, . . . , t}r, k, k′ ∈ N such that fσ(k) = fτ (k′). Then k = k′. Proof. We start with the case r = 1. Let σ = i and τ = j. If i = j, then as fi is increasing, we have k = k′. If i 6= j, then k = k′ follows by the fact that fi and fj are almost disjoint. Suppose the proposition holds for 1, 2, ..., r. Let σ, τ ∈ {1, 2, . . . , t}r+1, σ = iσ′, τ = jτ ′, with σ′, τ ′ ∈ {1, 2, . . . , t}r, and assume fσ(k) = fτ (k′). So fi(fσ′ (k)) = fσ(k) = fτ (k′) = fj(fτ ′(k′)). By step 1 of the induction we have fσ′ (k) = fτ ′(k′). By step r of the induction, we have k = k′. Proposition 2. Suppose σ, τ ∈ {1, 2, · · · , t}r and k ≥ m. If fσ(k) = fτ (k), then σ = τ . Proof. This simply follows from the fact that f1(k), . . . , ft(k) are disjoint for input values k grater than m. Proposition 3. Let σ ∈ {1, 2, . . . , t}r, r ≥ m, i ∈ N, and σ = σ1σ2 where σ2 = m. Then, [σ]i = σ1[σ2]i. Proof. Let τ ∈ [σ]i with τ = τ1τ2, and τ2 = m. As fσ1(fσ2 (i)) = fσ(i) = fτ (i) = fτ1(fτ2(i)) and σ1 = τ1, by Proposition 1, we have that fσ2(i) = fτ2(i). This means that σ2 ∼i τ2. As fσ2 (i) ≥ m by Proposition 2, we have that σ1 = τ1. Hence, we have shown that τ = σ1τ2 where τ2 ∈ [σ2], completing the proof of the proposition. Proposition 4. Let s > r ≥ m and i ∈ N. Then, ♯(s, i) = ts−r♯(r, i). Proof. This follows directly from Proposition 3. Proposition 5. Let σ ∈ {1, 2, . . . , t}r, r ≥ m. Let σ = σ1σ2, where σ2 = m. Then c([σ]i) = cσ1 c([σ2]i). 6 Proof. This simply follows from Proposition 3. Proposition 6. Assume that the coefficients c1, c2, . . . , ct satisfy the non-zero condition at level m. Then, for all σ ∈ {1, 2, . . . , t}r, with r ≥ m, and i ∈ N we have that where γ = γ(c1, . . . , ct). In particular, c([σ]i) > 0. c([σ]i) ≥ γr Proof. Let σ ∈ {1, 2, . . . , t}r, with r ≥ m, and i ≥ 1. Let us first consider the case that i ≥ m. Then, by Proposition 2 we have that [σ]i = {σ}. Hence c([σ]i) = cσ ≥ γr. Now assume 1 ≤ i ≤ m. Write σ = σ1σ2 where σ2 = m. Then, by Proposition 5, we have that c([σ]i) = cσ1 c([σ2]i). Hence, c([σ]i) = cσ1 c([σ2]i) ≥ γr−m · γ = γr−m+1 ≥ γr. 2. Main results We would like to prove the following result. Theorem 7. Let f1, f2, · · · , ft be increasing functions which are pairwise al- most disjoint. Let m ∈ N be such that fi(k) 6= fj(k) for all 1 ≤ i < j ≤ t and k ≥ m. Assume that the coefficients c1, c2, · · · ct ∈ R satisfy the non-zero i=1 ciTfi . Then, for sufficiently large λ, the condition at level m and let T = Pt operator λT is chaotic. Proof. Let γ = γ(c1, . . . , ct). Let λ be large enough so that 4t2 γ < λ. (1) Let us first prove that λT is transitive. To this end, let ~x = (x1, x2, . . . , xj , 0, 0, 0, . . . ) and ~y = (y1, y2, . . . , yj, 0, 0, 0, . . . ) be in lp and let ε > 0 be given. It is enough to show that there exists ~w = (w1, w2, . . . ) ∈ lp such that k~x − ~wkp p < ε and k~y − λnT n( ~w)kp p < ε for some n. Without loss of generality we may assume that j > m. Let n be large enough so that n > j and so that j · kykp ∞ ∞ Xl=1 1 4ln < ε. (2) Now we define ~w = (w1, w2, . . . ) ∈ lp. Let wk = xk, for 1 ≤ k ≤ j. For each l ≥ 1, σ ∈ {1, 2, . . . , t}ln and 1 ≤ k ≤ j, we define wfσ (k) = ykλ−ln c([σ]k)♯(ln, k) . 7 (3) Let us observe that c([σ]k) 6= 0, by Proposition 6. We let other wi's be zero. Let us first show that ~w is well defined. Clearly, for l ≥ 1, σ ∈ {1, 2, . . . , t}ln, 1 ≤ k ≤ j, we have that fσ(k) ≥ ln ≥ n > j. Hence, the definition of wfσ (k) does not interfere with the definition of {w1, w2, . . . , wj}. Suppose fσ(k) = fσ′ (k′), where σ ∈ {1, 2, . . . , t}ln, σ′ ∈ {1, 2, . . . , t}l′n and 1 ≤ k, k′ ≤ j, and l, l′ ≥ 1. We need to verify that wfσ (k) = wfσ′ (k′). We first verify that l = l′. To obtain a contradiction, assume that l 6= l′. Without loss of generality assume that l < l′. Then, we write σ′ = σ′ 2 ≥ n and that fσ(k) = fσ′ (k′) = fσ′ 2(k′), 1 (fσ′ 2 (k′) > j and this by Proposition 1. However, as n > j we have that k = fσ′ contradicts that 1 ≤ k ≤ j. Hence l = l′. Now we have that σ, σ′ ∈ {1, 2, . . . , t}ln and since fσ(k) = fσ′(k′), by Proposition 1 we have that k = k′. Hence σ ∼k σ′. Therefore, 1σ′ 2 (k′)). As σ = σ′ 1, we have that k = fσ′ 2, where σ′ 1 = σ. Note that σ′ wfσ (k) = ykλ−ln c([σ]k)♯(ln, k) = yk′ λ−ln c([σ′]k′ )♯(l′n, k′) = wfσ′ (k′). Therefore ~w is well defined. Moreover, k~x − ~wkp p = ∞ j Xl=1 Xk=1 X[σ]k∈A(ln,k) wfσ (k)p = ≤ = ∞ j Xl=1 Xk=1 X[σ]k∈A(ln,k) ∞ j Xl=1 Xk=1 X[σ]k∈A(ln,k) ykpλ−pln c([σ]k)p♯(ln, k)p kykp ∞λ−pln (γln)p♯(ln, k)p ∞ j Xl=1 Xk=1 kykp ∞λ−pln♯(ln, k) γlnp♯(ln, k)p ≤ jkykp ∞ ≤ jkykp ∞ ∞ Xl=1 Xl=1 ∞ 1 (γλ)pln 1 4pln ≤ jkykp ∞ ∞ Xl=1 1 4ln < ε. (4) (5) (6) (7) (8) (9) Let us give explanations for some of the equalities and inequalities above. Let us first justify Equality (4). Note that for all i > j where wi 6= 0 we have that i = fσ(k) for some σ ∈ {1, . . . , t}nl, l ≥ 1, and 1 ≤ k ≤ j. Moreover, if σ ∼k τ implies that fσ(k) = fτ (k). Hence, we only need to sum over one element from each equivalence class in A(nl, k). Equality (5) is simply inserting the definition of wfσ (k). Inequality (6) follows from Proposition 6. Equality (7) follows from the definition of ♯(ln, k). Inequality (8) is simply algebra. Inequality (9) follows from the manner in which λ and n were chosen see (1) and (2) respectively. 8 Let us next show that kλnT n( ~w) − ~ykp p < ε. We first note that for all k we have that T n( ~w)(k) = Xσ∈{1,2,...,t}n cσwfσ (k). In the case that 1 ≤ k ≤ j, we have that T n( ~w)(k) = Xσ∈{1,2,...,t}n = Xσ∈{1,2,...,t}n cσwfσ (k) cσ ykλ−n c([σ]k)♯(n, k) = = = = = cσ c([σ]k) ykλ−n ykλ−n ykλ−n ♯(n, k) Xσ∈{1,2,...,t}n ♯(n, k) X[σ]k∈A(n,k) Xτ ∈[σ]k ♯(n, k) X[σ]k∈A(n,k) Xτ ∈[σ]k ♯(n, k) X[σ]k∈A(n,k) ♯(n, k) X[σ]k∈A(n,k) ykλ−n ykλ−n cτ c([τ ]k) cτ c([σ]k) c([σ]k) Xτ ∈[σ]k 1 1 (10) cτ c([σ]k) · c([σ]k) = ykλ−n ♯(n, k) X[σ]k∈A(n,k) 1 = ykλ−n. (11) We note that Equality (10) follows from the fact that c([σ]k) = c([τ ]k) whenever σ ∼k τ . Equality (11) follows from the definitions of c([σ]k) and A(n, k). Hence we have just shown that for 1 ≤ k ≤ j, we have λnT n( ~w)(k) = yk. Let us now consider the case k > j. Let us compute T n( ~w)(k), for k > j. In order to do that, we have to compute wfσ (k) for σ ∈ {1, 2, . . . , t}n. If wfσ (k) = 0, ∀σ ∈ {1, 2, · · · , t}n, then T n( ~w)(k) = 0. In the case in which there exists a σ ∈ {1, 2, · · · , t}n such that wfσ (k) 6= 0, we assign in a unique way to k a triple (lk, τk, ik) such that lk ≥ 1, τk ∈ {1, 2, · · · , t}lkn, 1 ≤ ik ≤ j and k = fτk (ik). Indeed as wfσ (k) 6= 0, we have that fσ(k) = fτ (ik), for some 1 ≤ l ≥ 1. Now, let us write τ = τ ′τk, where τ ′ = n ik ≤ j, and τk = (l − 1)n. Then, fσ(k) = fτ ′(fτk (ik)). Since σ = τ ′ we have that k = fτk(ik) by Proposition 1. As k > j and ik ≤ j, we have also that τk 6= ∅. Hence l − 1 ≥ 1. Let lk = l − 1, therefore (lk, τk, ik) is the desired triple. τ ∈ {1, 2, . . . , t}ln, 9 Now, T n( ~w)(k)p = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xσ∈{1,2,...,t}n ≤ Xσ∈{1,2,...,t}n(cid:12)(cid:12)(cid:12) = Xσ∈{1,2,...,t}n(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) = Xσ∈{1,2,...,t}n = Xσ∈{1,2,...,t}n ≤ Xσ∈{1,2,...,t}n p p p cσwfσ (k)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) cσwfσ (fτk (ik))(cid:12)(cid:12)(cid:12) cσwfστk (ik)(cid:12)(cid:12)(cid:12) c([στk]ik )♯((lk + 1)n; ik)(cid:12)(cid:12)(cid:12)(cid:12) λ−(lk+1)nyik λ−lk npλ−npyik p cσ c([τk]ik )p♯((lk + 1)n; ik)p λ−lk npλ−npyik p γlknp♯((lk + 1)n; ik)p p (12) (13) ≤ tn λnp kykp ∞ (λγ)lk np In above, Equality (12) follows by Proposition 5 and Inequality (13) follows by Proposition 6. Now let us make the following observation. Suppose k 6= k′ > j, T n( ~w)(k) 6= 0 and T n( ~w)(k′) 6= 0. Then, k → (lk, τk, ik) and k′ → (lk′ , τk′ , ik′) are defined and, moreover, (lk, τk, ik) 6= (lk′ , τk′ , ik′ ). Hence, kλnT n( ~w) − ~ykp p = ∞ Xk=j+1 λnT n( ~w)(k)p = λnp ∞ Xk=j+1 T n( ~w)(k)p ∞ j Xl=1 Xτ ∈{1,2,...,t}ln Xi=1 tn λnp kykp ∞ (λγ)lnp ≤ λnp = ∞ Xl=1 tln · j · ∞ ≤ jkykp ∞ ≤ jkykp ∞ ≤ jkykp ∞ ∞ Xl=1 Xl=1 Xl=1 ∞ tnkykp ∞ (λγ)pln t(l+1)n (λγ)pln t2ln (λγ)pln λγ(cid:19)pln (cid:18) t2 < jkykp ∞ ∞ Xl=1 4(cid:19)pln (cid:18) 1 < ε. (14) In above, Inequality (14) follows by (1) and (2). 10 Hence k~y − λnT n( ~w)kp p < ε and we have shown that λT is transitive. Let us now prove that the periodic points of λT are dense in ℓp. Indeed, ~x p < ε. Hence, it will suffice to show that was chosen arbitrarily and k~x − ~wkp λnT n( ~w) = ~w. As we have already observed, λnT n( ~w)(k) = yk, ∀ 1 ≤ k ≤ j. Let us assume that k > j. Let A = {fσ(i) : σ ∈ {1, 2, . . . , t}ln, 1 ≤ i ≤ j, l ≥ 1}. Note that the elements of A are all greater than j. Let B = N \ A ∩ [j + 1, ∞). Let us first consider the case k ∈ A. Then k = fσ(i) for same σ ∈ {1, 2, . . . , t}ln, 1 ≤ i ≤ j and l ≥ 1. By definition we have that, λnT n( ~w)(k) = λn Xτ ∈{1,2,...,t}n = λn Xτ ∈{1,2,...,t}n = λn Xτ ∈{1,2,...,t}n = λn Xτ ∈{1,2,...,t}n = Xτ ∈{1,2,...,t}n cτ wfτ (k) cτ wfτ (fσ (i)) cτ wfτσ (i) cτ λ−(l+1)nyi c([τ σ]i)♯((l + 1)n, i) λ−lnyi c([σ]i)♯((l + 1)n, i) = tn λ−lnyi c([σ]i)tn♯(ln, i) λ−lnyi = c([σ]i)♯(ln, i) (15) (16) = wk. We note that Equality (15) follows from Proposition 5 and Equality (16) follows from Proposition 4. Now let us assume that k ∈ B. We claim that for each σ ∈ {1, 2, . . . , t}n we have that fσ(k) /∈ A. To obtain a contradiction, assume fσ(k) ∈ A and that there exist l, τ ∈ {1, 2, . . . , t}ln and 1 ≤ i ≤ j such that fσ(k) = fτ (i). Let τ = τ1τ2, τ1 = n, then fσ(k) = fτ1 (fτ2(i)) . By the Proposition 1 we have that k = fτ2(i). Moreover, as k > j, τ2 6= ∅. Hence k ∈ A and this is a contradiction as k ∈ B. Therefore, for all k ∈ B we have that fσ(k) ∈ B. Recall that the values of ~w at each coordinate of B is zero. Hence, when k ∈ B we have that λnT n( ~w)(k) = λn Xσ∈{1,2,...,t}n cσwfσ (k) = 0 = wk, completing the proof of the theorem. Remark 8. By the proof of the above Theorem it follows at once that if f : N → N is an increasing function, then for each λ > 1 the operator λTf is chaotic. 11 Remark 9. We observe that if the coefficients c1, c2, . . . , ct do not satisfy the non-zero condition at level m, Theorem 7 does not hold. Indeed, let us con- sider the following increasing almost disjoint functions f1(n) = 2n, ∀n ≥ 1 and f2(1) = 2 and f2(n) = 2n − 1, ∀n > 1. The operator T = 2Tf1 − 2Tf2 has the property that T (~x) is always zero in the first coordinate. Hence, λT is not chaotic for any λ. Corollary 10. Suppose that f1, f2, . . . , ft are increasing function with disjoint ranges. Let c1, c2, . . . , ct ∈ R \ {0}, then for sufficiently large λ, the operator i=1 ciTfi is chaotic. λPt Proof. We note that c1, c2, . . . , ct satisfies the non-zero condition at the level m = 1. The claim follows from Theorem 7. Corollary 11. There exists an infinite family T of chaotic operators such that i=1 ciTfi is for all T1, T2, . . . , Tt ∈ T and c1, c2, . . . , ct ∈ R the operator λPt chaotic for sufficiently large λ, provided that the following sum t Xi=1 ciTfi is not the zero operator. Proof. Let f1, f2, . . . , fi, . . . be increasing functions with disjoint ranges. Let T = {2Tfi : i ≥ 1}. Then T is the desired family. Proposition 12. There exists a family F of cardinality continuum consisting of increasing functions such that all distinct f, g ∈ F are pairwise almost disjoint. Proof. For each c > 1 , let gc : N → N be defined by gc(k) = ⌈ck⌉ where ⌈x⌉ denotes the ceiling function. Trivially gc is increasing as c > 1. Moreover, if c, d > 1 and c 6= d, then {k : gc(k) = gd(k)} is finite. Let α : N2 → N be a bijection such that if (i, j), (i′, j′) ∈ N2, with i < i′ and j < j′, then α(i, j) < α(i′, j′). For each c > 1, define fc : N → N by fc(k) = α(k, gc(k)). Let F = {fc : c > 1}. First of all, let us note that fc is an increasing map. Indeed, let k ∈ N. Then, fc(k + 1) = α(k + 1, gc(k + 1)) > α(k, gc(k)) = fc(k). Let c, d ≥ 1, with c 6= d. As {k : gc(k) = gd(k)} is finite and α is 1-1, we have that {k : fc(k) = fd(k)} is finite. Let k 6= k′, then (k, gc(k)) 6= (k′, gd(k′)). As α is 1-1 we have that fc(k) = α(k, gc(k)) 6= α(k′, gd(k′)) = fd(k′). Hence F has the desired proprieties. In the following, we use Ln to denote the n-dimensional Lebesgue measure in Rn. Lemma 13. Let P (x1, x2, . . . , xt) be a non-zero polynomial in variables x1, . . . , xt. Then, the set {(b1, b2, . . . , bt) ∈ Rt : P (b1, b2, . . . , bt) = 0} has t-dimensional Lebesgue measure zero. 12 Proof. We proceed by induction. The lemma is clear for t = 1 since non-zero polynomials have only finitely many roots. Suppose that the Lemma is true for t = 1, . . . , n. Now let us consider stage t = n + 1. Let P (x1, x2, . . . , xn+1) be a non-zero polynomial in the variables x1, . . . , xn+1. Fix (r1, . . . , rn+1) ∈ Rn+1 such that P (r1, . . . , rn+1) 6= 0. Consider the polynomial P (x1, x2, . . . , xn, rn+1) in variables x1, . . . , xn. Then this is a non-zero polynomial in n variables and, by induction, we have that A = {(b1, b2, . . . , bn) ∈ Rn : P (b1, b2, . . . , bn, rn+1) = 0} has n-dimensional Lebesgue measure zero. Now, let B = {(b1, b2, . . . , bn) ∈ Rn : P (b1, b2, . . . , bn, rn+1) 6= 0}. For each (b1, b2, . . . , bn) ∈ B, consider the polynomial of one variable P (b1, b2, . . . , bn, xn+1) in variable xn+1. Then, this is a non-zero polynomial in one variable and, hence, the set Sb1,b2,...,bn = {y ∈ R : P (b1, b2, . . . , bn, y) = 0} is a finite set. Let us now make an observation. If P (b1, . . . , bn+1) = 0 and P (b1, . . . , bn, rn+1) = 0, then (b1, . . . , bn) ∈ A and hence (b1, . . . , bn+1) ⊆ A × R. In the case, that P (b1, . . . , bn, rn+1) 6= 0, we have that (b1, . . . , bn) ∈ B and bn+1 ∈ Sb1,b2,...,bn . Putting all this together, we have that {(b1, . . . , bn+1) ∈ Rn+1 : P (b1, . . . , bn+1) = 0} ⊆ (A × R) ∪ {(b1, . . . , bn, y) ∈ Rn+1 : (b1, . . . , bn) ∈ B, y ∈ Sb1,...,bn}. As Ln(A) = 0, we have that By Fubini's Theorem, Ln+1(A × R) = 0. Ln+1({(b1, b2, . . . , bn, y) : (b1, . . . , bn) ∈ B, y ∈ Sb1,x2,...,bn}) = 0. Hence, the set in question at the (n + 1)st step of induction has (n + 1)- dimensional Lebesgue measure zero. Corollary 14. There exists a family T of cardinality continuum of chaotic op- erators such that for almost all (c1, c2, . . . , ct) ∈ Rt and T1, T2, . . . , Tt ∈ T the i=1 ciTfi is chaotic for sufficiently large λ. operator λPt Proof. Let T = {2Tf : f ∈ F}, where F is the family defined in the Propo- sition 12. Fix distinct elements in F , let say f1, f2, . . . , ft. Choose m accord- ing to the definition of almost disjoint functions and fix σ ∈ {1, 2, . . . , t}≤m and 1 ≤ i ≤ j. Note that c([σ]i) is a non-zero polynomial in c1, c2, . . . , ct. By Lemma 13, the t- dimensional Lebesgue measure of {(c1, c2, . . . , ct) : c([σ]i) = 0, for some σ ∈ {1, 2, . . . , t}≤m and1 ≤ i ≤ m} is equal to zero. Hence by the Theorem 7, for sufficiently large λ and almost all (c1, c2, . . . , ct) ∈ Rt, the operator λPt i=1 2ciTfi is chaotic. 13 Remark 15. We observe that, since the vector space of finitely non-zero se- quences is a dense subspace of c0 (as well as of lp, 1 ≤ p < ∞) and since lp ⊂ c0, then Theorem 7 and the other results of this section, hold also if the Banach space X is c0. References References [1] S. I. Ansari, Existence of hypercyclic operators on topological vector spaces, J. Funct. Anal. 148 (1997), no. 2, 384–390. [2] R. M. Aron, V. I. Gurariy, J. B. Seoane-Sep´ulveda, Lineability and space- ability of sets of functions on R, Proc. Amer. Math. Soc. 133 (2005), 795- 803. [3] R. M. Aron, D. P´erez-Garc´a, J. B. Seoane-Sep´ulveda, Algebrability of the set of non- convergent Fourier series. Studia Math. 175 (2006), no. 1, 83–90 [4] F. Bayart, E. Matheron, Dynamics of linear operators, Cambridge Tracts in Mathematics, 179. Cambridge University Press, Cambridge, 2009. xiv+337 pp. ISBN: 978-0-521-51496-5. [5] T. Berm`udez, A. Bonilla, F. Mart`ınez-Gim`enez, A. Peris, Li-Yorke and distributionally chaotic operators, J. Math. Anal. Appl. 373 (2011), no. 1, 83–93. [6] L. Bernal-Gonz`alez, On hypercyclic operators on Banach spaces Proc. Amer. Math. Soc. 127 (1999), no. 4, 1003–1010. [7] L. Bernal-Gonz`alez, M. Ord`onez Cabrera, Lineability criteria, with appli- cations. J. Funct. Anal. 266 (2014), no. 6, 3997–4025 [8] N. C. Bernardes, JR, A. Bonilla, V. Muller, A. Peris Li-Yorke chaos in linear dynamics, Ergod. Th. & Dyn. Sys.July (2014), 1–23 doi:10.1017/etds.2014.20 [9] N. C. Bernardes, JR, A. Bonilla, V. Muller, A. Peris, Distributional chaos for linear operators, J. Funct. Anal. 265 (2013), 2143-2163. [10] J. Bonet, F. Mart`ınez-Gim`enez, A. Peris, A Banach space which admits no chaotic operator Bull. London Math. Soc. 33 (2001), no. 2, 196–198. [11] B. Bongiorno, D. Darji, L. Di Piazza, Lineability of non-differentiable Pettis primitives, to appear on Monatsh. Math. doi: 10.1007/s00605-014-0703-6 [12] R. L. Devaney, An introduction to chaotic dynamical systems, Second edi- tion. Addison-Wesley Studies in Nonlinearity. Addison-Wesley Publishing Company, Advanced Book Program, Redwood City, CA, 1989. xviii+336 pp. ISBN: 0-201-13046-7 14 [13] P. H. Enflo, V. I. Gurariy, J. B. Seoane-Sep´ulveda, Some results and open questions on spaceability in function space, Trans. Amer. Math. Soc. 366 no 2 (2014), 611-625. [14] M. Fabian, P. Habala, P. H`jek, V. Montesinos Santaluca, J. Pelant, V. Zizler, Functional analysis and infinite-dimensional geometry, CMS Books in Mathematics/Ouvrages de Mathmatiques de la SMC, 8. Springer-Verlag, New York, 2001. x+451 pp. ISBN: 0-387-95219-5. [15] W. T. Gowers, B. Maurey, The unconditional basic sequence problem, J. Amer. Math. Soc. 6 (1993), no. 4, 851–874. [16] S. Grivaux, Sums of hypercyclic operators, J. Funct. Anal. 202 (2003), no. 2, 486–503. [17] K. G. Grosse-Erdmann, A. Peris Manguillot, Linear chaos, Universitext. Springer, London, 2011. [18] V. I. Gurariy, Linear spaces composed of nondifferentiable functions, C. R. Acad. Bulgare Sci. 44, (1991) no 5, 13-16. [19] V. I. Gurariy, L. Quarta. On lineability of sets of continuous functions, J. Math. Anal. Appl. 294 (2004), 62-72. [20] F. Mart`ınez-Gim`enez, P. Oprocha, A. Peris, Distributional chaos for back- ward shifts, J. Math. Anal. Appl. 351 (2009), no. 2, 607-615. [21] S. Rolewicz, On orbits of elements, Studia Math. 32 (1969), 17-22. 15
1202.0965	1	1202	2012-02-05T14:21:16	A Proof of Bobkov's Spectral Bound For Convex Domains via Gaussian Fitting and Free Energy Estimation	[ "math.FA", "math-ph", "math-ph" ]	We obtain a new proof of Bobkov's lower bound on the first positive eigenvalue of the (negative) Neumann Laplacian (or equivalently, the Cheeger constant) on a bounded convex domain $K$ in Euclidean space. Our proof avoids employing the localization method or any of its geometric extensions. Instead, we deduce the lower bound by invoking a spectral transference principle for log-concave measures, comparing the uniform measure on $K$ with an appropriately scaled Gaussian measure which is conditioned on $K$. The crux of the argument is to establish a good overlap between these two measures (in say the relative-entropy or total-variation distances), which boils down to obtaining sharp lower bounds on the free energy of the conditioned Gaussian measure.	math.FA	math	A Proof of Bobkov's Spectral Bound For Convex Domains via Gaussian Fitting and Free Energy Estimation Emanuel Milman1 Abstract We obtain a new proof of Bobkov's lower bound on the first positive eigenvalue of the (negative) Neumann Laplacian (or equivalently, the Cheeger constant) on a bounded convex domain K in Euclidean space. Our proof avoids employing the localization method or any of its geometric extensions. Instead, we deduce the lower bound by invoking a spectral transference principle for log-concave measures, comparing the uniform measure on K with an appropriately scaled Gaussian mea- sure which is conditioned on K. The crux of the argument is to establish a good overlap between these two measures (in say the relative-entropy or total-variation distances), which boils down to obtaining sharp lower bounds on the free energy of the conditioned Gaussian measure. 1 Introduction The following theorem was proved by Sergey Bobkov in [3] (we use the formulation from [37], which is formally stronger but ultimately equivalent in the cases of interest): Theorem 1.1 (Bobkov). Let K denote a convex bounded domain in Euclidean space (Rn,·). Let X denote a random vector uniformly distributed in K (with respect to normalized Lebesgue measure). Then: DChe(K) ≥ sup x0∈Rn for some universal constant c > 0. c pE(X − x0)S(X − x0) , (1.1) Let us explain the notation used above. We denote the expectation of a random variable Y by E(Y ), and set S(Y ) :=pE((Y − E(Y ))2) to denote the square root of the variance. We use DChe(Ω) to denote the Cheeger constant of the domain Ω ⊂ (Rn,·), defined as: DChe(Ω) := inf A⊂Ω , Hn−1(∂A ∩ Ω) min(Hn(A),Hn(Ω \ A)) 1Department of Mathematics, Technion - Israel Institute of Technology, Haifa 32000, Israel. Supported by ISF, BSF, GIF and the Taub Foundation (Landau Fellow). Email: [email protected]. 1 where the infimum ranges over all Borel subsets A ⊂ Ω, and Hk denotes the k-dimensional Hausdorff measure. When K is a convex domain, it is known by results of Maz'ya [34, 35] and Cheeger [17] on one hand, and Ledoux [31] on the other, that DChe(K) ≃pλ1(K), where λ1(Ω) denotes the first positive eigenvalue of the (negative) Laplacian −∆ on Ω with Neumann boundary conditions. Here and elsewhere A ≃ B signifies that there exist universal numeric constants C1, C2 > 0 so that C1A ≤ B ≤ C2A. Bobkov's Theorem may therefore be interpreted as a spectral bound for general convex domains. Of course, classical lower bounds on λ1(K) are known for convex domains K, and we mention the sharp lower bound in Euclidean space due to Payne and Weinberger [43]: λ1(K) ≥ π2 diam(K)2 . (1.2) This was extended by Li and Yau [32] and Zhong and Yang [45] to locally convex do- mains with smooth boundary on Riemannian manifolds with non-negative Ricci curva- ture. Back to the Euclidean setting, the bound (1.2) was extended (using the equivalent DChe(K) parameter) by Kannan, Lov´asz and Simonovits [27] to the bound: DChe(K) ≥ sup x0∈Rn log 2 E(X − x0) , (1.3) requiring control over the average distance to the centroid x0 rather than control over the diameter diam(K). To be completely accurate, the bound obtained in [27] was for a slightly different isoperimetric parameter, which is equivalent to within a factor of 2 to the DChe parameter; however, a simple modification of the argument (as in [6, Section 9]) yields the asserted (1.3). The latter bound was extended to the Riemannian setting (with a different numerical constant) in [40]. By testing very elongated n-dimensional cylinders or cones essentially degenerating to one-dimensional densities (and taking the limit as n → ∞ in the latter case), it is easy to verify that π2 and log 2 are the best possible (dimension independent) constants one can use in (1.2) and (1.3), respectively. However, assuming that such "biased" or "degenerate" domains are prohibited (for instance, a natural non-degeneracy condition is the isotropic condition, requiring that the variance of all unit linear functionals is equal to 1), Bobkov's bound is in fact the best general bound known to date. Indeed, up to a numeric constant it is certainly better than (1.3), since S(X − x0)2 ≤ E(X − x02) ≤ C(EX−x0)2 for some universal constant C > 1, where the last Khinchine-type inequal- ity is a well known consequence of the convexity of K (as follows from Borell's Lemma [9], see [41]). And indeed, Bobkov's bound was used in [25] to deduce the currently best-known estimate on the Cheeger constant for general isotropic convex domains K in Rn, namely DChe(K) ≥ cn−5/12; a conjecture of Kannan -- Lov´asz -- Simonovits [27] asserts that the bound should be DChe(K) ≥ c, for some dimension independent constant c > 0. Bobkov's proof in [3] is based on a localization method, having its origins in the work of Payne and Weinberger [43], rediscovered by Gromov and V. Milman [23], and 2 systematically developed by Kannan -- Lov´asz -- Simonovits [33, 27]. A more geometric proof of Bobkov's theorem was given in [37, Theorem 5.15], which was based on a general bound on DChe(K) obtained by Kannan -- Lov´asz -- Simonovits [27] using again localization. The latter general bound was recently given a completely geometric proof in [40], which generalizes to Riemannian manifolds with non-negative Ricci curvature. In this work, we propose yet another alternative proof of Bobkov's Theorem 1.1 in the Euclidean setting. However, we believe that the method of proof is of independent interest, and leads to interesting questions on the distribution of mass on convex sets. Furthermore, amongst the known proofs of Theorem 1.1, the approach described in this work is the only one which avoids directly employing the localization method (or its isoperimetric geometric extension, developed in [40]). Our method is based on a comparison between the uniform probability measure on K, denoted µK, and a suitably scaled Gaussian measure which we condition on K: γw K := exp(ZK (w)) exp(− w 2 x2)dµK (x) , where ZK(w) ≥ 0 is a normalization term ensuring that γw K is a probability measure. When K has unit volume, this measure may be thought of as the Gibbs measure at inverse-temperature w > 0 associated with the Hamiltonian: H(x) :=(x2 2 +∞ otherwise x ∈ K , and so exp(−ZK (w)) represents the partition function and ZK(w)/w represents the system's (Helmholtz) free energy. We have learned from Dario Cordero-Erausquin that the idea of adding wH(x) to a given potential and optimizing on w > 0 may be traced back to the work of Hormander (e.g. [26, Theorem 2.2.3]), who derived a generalized version of the Payne -- Weinberger estimate (1.2) for pseudo-convex domains in Cn. Clearly, it is enough to prove the desired bound (1.1) for x0 = 0, since DChe(Ω) is invariant under translation of the domain Ω. We consequently denote E = E0 = E(X) and S = S0 = S(X). It turns out that when w ≤ c ES , for an appropriately chosen small constant c > 0, µK and γw K overlap rather well in the total-variation or relative entropy senses, and general transference principles from [37, 39] imply that µK and γw K have comparable Cheeger constants (or equivalently, spectral-gaps). Since γw K is well-known to inherit all the isoperimetric and spectral properties of the (non-conditioned) Gaussian measure γw := γw [1, 2, 15] or see below), a delicate estimation of the overlap yields the desired bound (1.1). Rn (e.g. Intuitively it is clear that w should be of the order of 1 ES to get a good overlap. Indeed, since µK is mostly concentrated (by Chebyshev's inequality) in the annulus A := K ∩ (B(E + 2S) \ B(E − 2S)) (where B(R) denotes the Euclidean ball of radius R), we would like the density of γw K to be almost uniform in A, yielding the requirement that w((E + 2S)2 − (E − 2S)2) ≃ 1, i.e. that w ≃ 1 ES . However, we also need to control ZK(w) to push this argument through, which turns out to be a rather delicate task. 3 This constitutes the main part of this work, and leads to questions on the distribution of X in K which may be of independent interest. In particular, the crux of the argument relies on controlling from below the free energy ZK(w)/w for w ≤ c ES . The rest of this work is organized as follows. We recall some preliminaries in Section 2, and provide the full justification to the intuitive claims made above in Section 3. We conclude with Section 4, where we demonstrate that our lower bound for the free energy is in fact sharp. Unless otherwise stated, all constants c, c′, c′′, C, C′ etc. appearing in this work are universal numeric constants, whose values may change from one occurrence to the next. Acknowledgement. I thank Dario Cordero-Erausquin for his interest, discussions and references, and for making me realize that there is something to prove. I also thank the organizers of the 2011 S´eminaire de Math´ematiques Sup´erieures for the invitation to give a mini-course and to present this work. 2 Preliminaries 2.1 Isoperimetry, Spectral-Gap and Concentration We start by extending our definitions of DChe and λ1 to a general metric-measure space setting. A standard text-book reference for most of the notions we employ below is the excellent book by Ledoux [30] (see also [39] for a general overview). Let (Ω, d, µ) denote a measure-metric space, meaning that (Ω, d) is a separable metric space and µ is a Borel probability measure on (Ω, d). There are various known ways of measuring the interaction between the metric d and the measure µ. One of the strongest forms of measuring this interplay is given by isoperimetric inequalities. Recall that Minkowski's (exterior) boundary measure of a Borel set A ⊂ Ω, denoted µ+(A), is defined as µ+(A) := lim inf ε→0 := {x ∈ Ω;∃y ∈ A d(x, y) < ε} denotes the ε-neighborhood of A in (Ω, d). The isoperimetric profile I = I(Ω,d,µ) is then defined as the function I : [0, 1] → R+ given by I(v) = inf {µ+(A); µ(A) = v}. An isoperimetric inequality measures the relation between the boundary measure and the measure of a set, by providing a lower bound on I(Ω,d,µ). In this work, we will only be interested in the Cheeger constant DChe of the space (Ω, d, µ), defined as: ε = AΩ,d , where Ad ε µ(Ad ε )−µ(A) ε DChe(Ω, d, µ) := inf v∈[0,1] I(Ω,d,µ)(v) min(v, 1 − v) = inf A⊂Ω µ+(A) min(µ(A), 1 − µ(A)) , measuring a certain linear isoperimetric property of the space. When d and (or) µ are implied from the context, we simply write DChe(Ω) or DChe(µ). We will mostly work in Euclidean space (Rn, · ), and so given a bounded domain Ω ⊂ Rn, we denote DChe(Ω) = DChe(Ω, · , µΩ), where µΩ denotes the uniform probability measure on Ω, and given a Borel probability measure ν, denote DChe(ν) = DChe(Rn,·, ν). Note that in the former case, when A ⊂ Ω has smooth boundary, then µ+ Ω(A) = Hn−1(∂A∩Ω)/Hn(Ω). For the standard Gaussian measure γ on (Rn,·), a classical result of Sudakov -- Tsirelson 4 −∞ spaces). exp(−t2/2) and Φ(t) := R t [44] and independently Borell [10], asserts that Iγ = I(Rn,·,γ) = ϕ ◦ Φ−1, where ϕ(t) := 1√2π ϕ(s)ds, corresponding to the fact that half-planes are isoperimetric minimizers of Gaussian boundary measure. It is elementary to verify that Iγ is a concave function on [0, 1] vanishing at the endpoints and symmetric about the point 1/2, and so consequently DChe(γ) = 2Iγ(1/2) =p2/π, and by scaling DChe(γw) = p2/π√w (see [37, 38] for the concavity of the isoperimetric profile in more general Another way of measuring the interaction between metric and measure is given by Sobolev inequalities. In this work, we will only be interested in the Poincar´e inequality. Let F = F(Ω, d) denote the space of functions which are Lipschitz on every ball in (Ω, d). Given f ∈ F, define ∇f as the following Borel function: ∇f (x) := lim sup d(y,x)→0+ f (y) − f (x) d(x, y) . (and we define it as 0 if x is an isolated point - see [7, pp. 184,189] for more details). In the smooth Euclidean setting, ∇f of course coincides with the Euclidean length of the gradient of f . We say that (Ω, d, µ) satisfies a Poincar´e inequality if: ∃λ1 > 0 ∀f ∈ F Z ∇f2dµ ≥ λ1 Z f 2dµ −(cid:18)Z f dµ(cid:19)2! . pλ1(Ω, d, µ) ≥ 1 The best possible constant λ1 above, called the Poincar´e or spectral-gap constant, is denoted λ1(Ω, d, µ). As usual, we will use λ1(Ω) or λ1(µ) when the space is implied from the context. Note that when Ω is a smooth domain in Euclidean space, λ1(Ω) coincides with the first positive eigenvalue of the (negative) Laplacian −∆ on Ω with Neumann boundary conditions. For instance, it is well known (e.g. [30]) for the standard Gaussian measure γ on (Rn, · ) that λ1(γ) = 1, and hence by scaling λ1(γw) = w. As already alluded to in the Introduction, there is an intimate relation between DChe and λ1. It was shown by Cheeger [17] and independently by Maz'ya [34, 35] that 2 DChe(Ω, d, µ) (their proof extends from the Euclidean or Riemannian setting to the general metric-measure space one, cf. [37]). The converse inequality is in general false, due to the possible existence of narrow "necks" in the space (see e.g. [37]). However, it was shown by Buser [14] for the case of a closed Riemannian manifold having non-negative Ricci curvature, and extended by Ledoux [31] to the case of a manifold-with- density having non-negative generalized Ricci curvature, that the converse inequality also holds up to a universal numeric constant. We do not provide unnecessary definitions here and only mention that in particular, in the case of Euclidean space (Rn, · ) endowed with a log-concave probability measure µ, it follows thatpλ1(µ) ≃ DChe(µ). Recall that a measure µ on Rn is called log-concave if µ = exp(−V (x))dx with V : Rn → R ∪{+∞} convex (cf. Borell [9]). Note that all of the probability measures mentioned in the Introduction µK, γw K and γw are log-concave. A third way of measuring the interaction between metric and measure is given by concentration inequalities. The concentration profile of our space K = K(Ω,d,µ) : R+ → 5 [0, 1/2] is defined as the pointwise minimal function so that 1 − µ(Ad r ) ≤ K(r) for all Borel sets A ⊂ Ω with µ(A) ≥ 1/2 (r ≥ 0). Clearly K is a non-increasing function. Concentration inequalities measure how tightly the measure µ is concentrated around sets having measure 1/2 as a function of the distance r away from these sets, by providing an upper bound on K(r) which decays to 0 as r → ∞. To measure this decay, we denote by DExpp = DExpp(Ω, d, µ) ≥ 0 the best constant in the following inequality (p > 0): ∀r ≥ 0 K(r) ≤ exp(−(DExppr)p) . When DExp1 > 0 (DExp2 > 0) we will say that our space has exponential (Gaussian) concentration (respectively). Indeed, it is well-known (see e.g. [30]) that the standard Gaussian measure γ on (Rn, · ) satisfies DExp2(γ) = 1/√2 , and hence by scaling DExp2(γw) = pw/2. Furthermore, it was shown by M. Gromov and V. Milman [22] that DExp1 ≥ c√λ1 for some universal numeric constant c > 0, i.e. that having positive spectral-gap implies exponential concentration (their proof applies in a general metric- measure space setting). The converse inequality is again in general false (consider e.g. a measure with disconnected support). However, for the class of log-concave measures in Euclidean space, it was shown in our previous work [37] that in fact DExp1 ≥ c′DChe for some universal constant c′ > 0, which implies together with the above mentioned results that for this class DChe ≃ √λ1 ≃ DExp1 (see [37] for a much stronger result in a more general setting). 2.2 Stability of DChe in the class of log-concave measures Next, we will require two notions of proximity between two Borel probability measures µ1 and µ2 on (Ω, d). The first is given by the total-variation distance, denoted dT V (µ1, µ2) and defined as: dT V (µ1, µ2) := sup A⊂Ωµ1(A) − µ2(A) . The second is given by the relative entropy (or Kullback -- Leibler divergence) of µ2 with respect to µ1, denoted H(µ2µ1) and defined as: H(µ2µ1) := Entµ1(cid:18)dµ2 dµ1(cid:19) =Z log(cid:18) dµ2 dµ1(cid:19) dµ2 , The key ingredient in our proof of Bobkov's Theorem 1.1 is the following theorem, if µ2 ≪ µ1, and ∞ otherwise. compiled from our previous results from [37, 39]: Theorem 2.1 ([37, 39]). Let µ1, µ2 denote two log-concave probability measures on (Rn,· ). • There exists a universal constant c > 0 so that for any ε > 0: dT V (µ1, µ2) ≤ 1 − ε ⇒ DChe(µ2) ≥ c ε2 log(1/ε) DChe(µ1) . 6 • There exists a universal constant c > 0 so that for any p ≥ 1: H(µ2µ1) ≤ L ⇒ DChe(µ2) ≥ Since DExp1(µ1) ≥ c′DChe(µ1), in particular: c 1 + L1/p DExpp(µ1) . H(µ2µ1) ≤ L ⇒ DChe(µ2) ≥ cc′ 1 + L DChe(µ1) . 2.3 Conditioning the Gaussian measure onto K Recall that given w > 0, we define γw K, namely: K by conditioning the Gaussian measure γw onto γw K := exp(ZK (w)) exp(cid:16)− w 2 x2(cid:17) dµK (x) , (2.1) where µK denotes the uniform (probability) measure on K, and ZK (w) ≥ 0 is a nor- malization term ensuring that γw K is a probability measure. The following estimates are well-known: Theorem 2.2. ∀w > 0 DChe(γw K ) ≥r 2 π √w , λ1(γw K ) ≥ w , DExp2(γw K ) ≥pw/2 . For the sake of completeness, we present (several variants of) the proof. Proof. One way to obtain all these estimates simultaneously and with the optimal constants is by applying Caffarelli's Contraction Theorem. It was shown in [15] that given a source probability measure µ1 and target probability measure µ2 on Rn, hav- ing the form µ1 = γw and µ2 = f γw with f log-concave, there exists a Borel map T : (Rn,·) → (Rn,·) pushing forward µ1 onto µ2 which contracts Euclidean distance: T (x) − T (y) ≤ x − y for all x, y ∈ Rn. In fact, T is the Brenier optimal-transport map [13], minimizing the L2-average transport cost R T (x) − x2dµ1(x) among all maps pushing forward µ1 onto µ2 (see also [28] for a different construction in a more general setup). We note that Caffarelli's result applies regardless of the smoothness of f , since we can always approximate f by smooth log-concave functions {fn} so that {fnµ1} con- verge in total-variation to f µ1, and the corresponding contractions {Tn} pushing forward µ1 onto {fnµ1} will necessarily converge (by the Arzel`a -- Ascoli Theorem) to our desired contraction T (see [28, Lemma 3.3] for the precise argument). Note that the convexity of K ensures that f0 := dγw dγw = cK,w1K is log-concave, and so by Caffarelli's Contraction Theorem there exists a contraction pushing forward γw onto γw K . K It is easy to verify that all of forms of interaction between metric and measure described in Subsection 2.1 can only improve under contracting maps. Namely, if T : (Ω1, d1, µ1) → (Ω2, d2, µ2) is a Borel map pushing forward µ1 onto µ2 and contract- ing distances d2(T (x), T (y)) ≤ d1(x, y) for all x, y ∈ Ω1, then I(Ω2,d2,µ2) ≥ I(Ω1,d1,µ2), 7 ∇(f ◦ T ) (x) ≤ ∇f (T x) and K(Ω2,d2,µ2) ≤ K(Ω1,d1,µ2). In particular, it immediately follows that: DChe(γw K ) ≥ DChe(γw) , λ1(γw K ) ≥ λ1(γw) , DExp2(γw K ) ≥ DExp2(γw) , and the asserted bounds follow from the well-known bounds for the Gaussian measure γw mentioned in Subsection 2.1. Another possible way to derive the above mentioned bounds (with perhaps inferior numerical constants) is as follows. The Brascamp -- Lieb inequality [12] states that if µ = exp(−V (x))dx is a log-concave measure on Rn with V ∈ C 2(Rn) having positive definite Hessian Hess V > 0, then for any f ∈ C 1(Rn): Z (cid:10)(Hess V )−1∇f,∇f(cid:11) dµ ≥Z f 2dµ −(cid:18)Z f dµ(cid:19)2 . In particular, it follows that if Hess V ≥ λId with λ > 0 then λ1(µ) ≥ λ (the latter statement may also be attributed to Lichnerowicz, at least in an analogous Riemannian [20, Theorem 4.70]). Moreover, if Hess V ≥ λId with λ > 0, then by the setting, cf. Bakry -- ´Emery criterion [1], (Rn, · , µ) satisfies in fact a stronger log-Sobolev inequality, and the Herbst argument implies that DExp2(µ) ≥ c√λ for some universal constant c > 0 (the interested reader may consult [30, 39] for missing definitions and arguments). The inequality DChe(µ) ≥ c′√λ may be deduced by using that DChe(µ) ≃ pλ1(µ) for log- concave measures µ; alternatively, one may directly deduce that DChe(µ) ≥ DChe(γλ) by applying the isoperimetric inequality of Bakry -- Ledoux [2]. Finally, an approximation argument as the one described above ensures that all isoperimetric, Sobolev and concen- tration estimates pass onto γw K = f0γw (see [37, 38, 36] for more technical approximation results, which in fact apply in the more challenging Riemannian setting). 3 Proof of Theorem 1.1 In this section we provide a proof of Bobkov's Theorem 1.1 via Gaussian Fitting and estimation of the free energy ZK (w)/w. We denote E2 =pE(X2). Clearly E ≤ E2 by Jensen's inequality, but as mentioned in the Introduction, it is also well-known that the convexity of K ensures that E2 ≤ CE for some universal numeric constant C > 1. Together with the obvious reduction to the case x0 = 0 described in the Introduction, the proof of Theorem 1.1 thus reduces to proving the following equivalent bound: DChe(K) ≥ c √E2S . (3.1) 3.1 Reduction to Estimation of the Free Energy Note that the normalization term ZK(w) ≥ 0 appearing in (2.1) is given by: ZK(w) := − logZ exp(− w 2 x2)dµK (x) . 8 We summarize several useful properties of ZK(w) below: Lemma 3.1. • The function [0,∞) ∋ w 7→ ZK (w) is concave increasing. • ZK (0) = 0 and d • ZK (w) = n 2 log(w/(2π)) − log γ(√wK), where γ denotes the standard Gaussian dw ZK (0) = 1 2 E2 2 . measure on Rn. • (0,∞) ∋ w 7→ ZK(w)/w decreases continuously from 1 2 E2 2 to 0. Proof. The first assertion follows immediately by direct differentiation of by employing Holder's inequality. The second and third ones are immediate to verify as well. The fourth one is an immediate consequence of the first three. Remark 3.2. Some properties and extremal characterizations of γ(K) for symmetric convex sets K in Rn have been studied by Cordero-Erausquin, Fradelizi and Maurey [18] and by Bobkov [4]. In view of the third assertion of the above lemma, it is immediate to translate these properties into statements about ZK (w). However, we have not found any concrete applications of these connections, and so we only vaguely mention these in passing, and refer the interested reader to the above mentioned references. To transfer the lower bound on DChe(γw K ) given in Theorem 2.2, namely DChe(γw K ) ≥ c√w, onto a lower bound on DChe(µK), we would like to invoke one of the transference principles given by Theorem 2.1. To this end, we must control the total-variation distance dT V (µK , γw K ). The calculation is a little more concise for the latter option, so we proceed with the task of bounding the relative entropy from above. K) or relative entropy H(µKγw Lemma 3.3. • H(µKγw K ) = Z (cid:16) w = E2 1 2 2 x2 − ZK(w)(cid:17) dµK(x) 2 w − ZK(w) = w(cid:18) 1 E2 2 − K) is convex and increases from 0 to ∞. w (cid:19) . ZK (w) 2 (3.2) • The function [0,∞) ∋ w 7→ H(µKγw Proof. The first assertion follows by definition. The second assertion is a direct conse- quence of the first one and the concavity of ZK(w). 9 Since the measure γw K) ≥ pw/2, the transference principle for the relative entropy given by Theorem 2.1 guaran- K has Gaussian concentration by Theorem 2.2, namely DExp2(γ2 tees that if: H(µKγw K ) ≤ wLw , then for some universal constant c′ > 0: DChe(µK) ≥ c′ DExp2(γw 1 + √wLw ≥ K ) c′/√2 1/√w + √Lw . (3.3) Optimizing on w > 0 in (3.3), we see that we would like to choose w0 > 0 as large as possible so that H(µKγw0 Recalling (3.2), this boils down to choosing w0 > 0 as large as possible so that: K ) ≤ w0Lw0 ≃ 1, resulting in the bound DChe(µK) ≥ c′′√w0. ZK (w0) w0 ≥ 1 2 E2 2 − C w0 , (3.4) for some numeric constant C > 0. We will show that (3.4) is indeed satisfied with w0 = c/E2S, for some appropriately small numeric constant c > 0, from whence Bobkov's bound (3.1) follows. Theorem 3.4 (Free Energy Lower Bound). There exists universal constants c, C > 0 so that for any convex bounded domain K in (Rn, · ), we have the following estimate for the free energy: w ≤ c E2S ⇒ ZK (w) w ≥ 1 2 E2 2 − CE2S . (3.5) Before proceeding with the proof of Theorem 3.4, we summarize below its conse- quences as described in the preceding discussion: Corollary 3.5. There exist two numeric constants c′, c′′ > 0 so that for any convex bounded domain K in (Rn, · ): • Setting w0 = c′/(E2S), we have H(µKγw0 • Bobkov's bound: DChe(K) ≥ c′′/√E2S. K ) ≤ 1/2 and dT V (µK, γw0 K ) ≤ 1/2. Proof. Theorem 3.4 and (3.2) guarantee that setting c′ = min(c, 1/(2C)), where c, C > 0 K ) follows. The bound on dT V (µK , γw0 are given in Theorem 3.4, the bound on H(µKγw0 K ) follows by the Csisz´ar -- Kullback -- Pinsker inequality (e.g. [30]), stating that for any two Borel probability measures µ1, µ2: Bobkov's lower bound on DChe(K) follows by the discussion preceding the statement of Theorem 3.4, employing the transference principle given by Theorem 2.1 and the well-known bound DChe(γw dT V (µ1, µ2) ≤pH(µ2µ1)/2 . K ) ≥ c′′′√w stated in Theorem 2.2. 10 3.2 Main Calculation For the proof of Theorem 3.4, we require the following: Lemma 3.6. There exists a universal constant c1 := log(3)/4 > 0 so that for any convex bounded domain K in (Rn, · ): µK {x ≤ r} ≤ exp(cid:18)−c1 S (cid:19) ∀r ∈ [0, E2 − 3S] . E2 − r Proof. Note that by the Brunn -- Minkowski inequality (e.g. [21]), the function r 7→ log µK {x ≤ r} is concave. Indeed, denoting K∞ := R×K and L∞ = {(r, x) ∈ R × Rn ; r ≥ 0 , x ≤ r}, then µK {x ≤ r} = Vol(K+∩Hr)/Vol(K), where K+ is the convex set K∞∩L∞ and Hr is the hyperplane {r} × Rn in Rn+1. Consequently, the function r 7→ (µK {x ≤ r})1/n is concave, and in particular, the asserted log-concavity of r 7→ µK {x ≤ r} follows. Next, note that by Chebyshev's inequality: µK {x ≤ E − 2S} ≤ 1 4 , µK {x ≤ E + 2S} ≥ 3 4 . The log-concavity of r 7→ µK {x ≤ r} consequently implies: t ln(3) µK {x ≤ E − 2S − t} ≤ 1 4 exp(cid:18)− 4 S(cid:19) ∀t ≥ 0 , or equivalently: µK {x ≤ r} ≤ 1 4 exp(cid:18)− ln(3) 4 (cid:18) E − r S − 2(cid:19)(cid:19) ∀r ≤ E − 2S . Observing that E2 2 = E2 + S2 and hence E2 ≤ E + S, it follows that: µK {x ≤ r} ≤ 1 4 exp(cid:18)3 ln(3) 4 (cid:19) exp(cid:18)− ln(3) 4 S (cid:19) ∀r ≤ E2 − 3S , E2 − r and the assertion of the lemma follows by direct numerical inspection. Proof of Theorem 3.4. First, we may assume that E2 ≥ 3S, since otherwise the assertion follows trivially with C = 3/2. Next, since ZK(w)/w is decreasing, it is enough to prove the assertion for w0 = c1/(αE2S), where c1 is the constant from the above Lemma and α := 2. Integrating by parts, evaluating separately the integral over the ranges [0, E2 − 3S] and [E2 − 3S,∞), and using the Lemma, we have: Z exp(− ≤ exp(cid:16)− w0 2 w0 2 x2)dµK (x) =Z ∞ (E2 − 3S)2(cid:17) +Z E2−3S 0 0 w0r exp(− w0 2 w0r exp(− r2)µK {x ≤ r} dr r2) exp(cid:18)−c1 w0 2 S (cid:19) dr . E2 − r 11 To evaluate the second term above, denoted T2, we use the change of variables r = c1 w0S − t = αE2 − t: T2 = exp(cid:18)−c1 ≤ exp(cid:18)−c1 2(cid:18)√w0r − (αw0E2 − w0t) exp(− 2w0S2(cid:19)Z E2−3S 2w0S2(cid:19)Z ∞ c1√w0S(cid:19)2! dr w0r exp − t2)dt . E2 S E2 S w0 2 c2 1 c2 1 + + 1 0 E2(α−1)+3S Since α ≥ 1, we estimate the latter expression by: ≤ exp(cid:18)−c1 ≤ exp(cid:18)−c1 E2 S E2 S + + c2 1 αE2 E2(α − 1) + 3S − 1(cid:19)Z ∞ 2w0S2(cid:19)(cid:18) 2w0S2(cid:19) 1 α − 1 exp(cid:16)− w0 2 c2 1 (E2(α − 1) + 3S)2(cid:17) . E2(α−1)+3S w0t exp(− w0 2 t2)dt Recalling that w0 = c1/(αE2S), this is the same as: = exp(cid:18)c1 E2 S (cid:16) α 2 − 1(cid:17)(cid:19) 1 α − 1 w0 2 exp(cid:16)− (E2(α − 1) + 3S)2(cid:17) . It is easy to check that, in fact, for any α > 0: exp(cid:18)c1 E2 S (cid:16) α 2 − 1(cid:17)(cid:19) exp(cid:16)− w0 2 (E2(α − 1))2(cid:17) = exp(− w0 2 E2 2) , but this is most apparent for α = 2 (which satisfies the requirement α ≥ 1 used above). Consequently: T2 ≤ 1 α − 1 w0 2 exp(cid:16)− E2 2(cid:17) ≤ 1 α − 1 w0 2 exp(cid:16)− (E2 − 3S)2(cid:17) , and we conclude that for w0 = c1/(2E2S): Z exp(− w0 2 x2)dµK (x) ≤ 2 exp(cid:16)− w0 2 (E2 − 3S)2(cid:17) . It follows that: ZK(w0) w0 = − 1 2 ≥ log(cid:18)Z exp(− 1 w0 (E2 − 3S)2 − w0 2 x2)dµK (x)(cid:19) E2S ≥ E2 1 2 2 log(2) c1 2 −(cid:18)3 + log(4) c1 (cid:19) E2S . The assertion now follows from the monotonicity of w 7→ ZK(w)/w with c = c1/2 and C = 3 + log(4) . c1 12 Remark 3.7. To appreciate the delicate nature of the bound (3.5) obtained in Theorem 3.4, note that the requirement that α > 1 was crucially used in the proof. In particu- lar, we do not know how to extend (3.5) to the following statement, which is perhaps suggested by the heuristic argument outlined in the Introduction: c E2S ⇒ ZK(w) w ≥ 1 2 ∀c > 0 ∃C > 0 such that w ≤ E2 2 − CE2S . Furthermore, note that our proof above crucially relied on the estimateR ∞a exp(− t2 2 ) valid for positive a > 0, and that the slightly rougher boundR ∞a exp(− t2 a exp(− a2 C exp(− a2 w0 ≥ 1 2 )dt ≤ 2 )dt ≤ 2 ) would have incurred an extraneous logarithmic term in the final result: 2 E2 2 − CE2S log(E2/S). 1 ZK (w0) Remark 3.8. A more delicate analysis of the proof above (e.g. using α = 5) reveals that we may in fact assert the existence of a constant c > 0 so that: w ≤ c E2S ⇒ ZK(w) w ≥ 1 2 (E2 − 3S)2 . 4 Sharpness of Free Energy Estimate Before concluding, we observe that the lower bound on the free energy we obtained in Theorem 3.4 is in fact sharp in the following sense: Proposition 4.1 (Free Energy Upper Bound). There exist two universal constants c, C > 0 so that for any convex bounded domain K in Rn: w ≥ C E2S ⇒ ZK (w) w ≤ 1 2 E2 2 − cE2S . The proof is immediate given the following lemma, asserting a reverse Chebyshev- type inequality for log-concave measures, which may be of independent (technical) in- terest: Lemma 4.2. Given a log-concave probability measure µ on Euclidean space (Rn, · ), let X be a random vector distributed according to µ and set E = E(X) and S = S(X). There exists a universal constant c0 > 0 so that: P(X ≤ E − c0S) ≥ c0 . (4.1) Proof of Proposition 4.1. Since µK is log-concave and E ≤ E2, we know by the previous lemma that: Consequently, we may estimate: µK {x ≤ E2 − c0S} ≥ c0 . Z exp(− w 2 x2)dµK (x) ≥ c0 exp(− w 2 (E2 − c0S)2) , 13 and hence: ZK (w) w ≤ 1 2 (E2 − c0S)2 + w It follows that if C > 0 is chosen large enough, then: log(1/c0) ∀w > 0 . where: w ≥ C E2S ⇒ ZK (w) w ≤ 1 2 E2 2 − cE2S , c := 2c0 − c2 0 − log(1/c0) C > 0 . It remains to establish Lemma 4.2. Surprisingly, we have not found a reference for it in the literature, perhaps due to the fact that the inequality (4.1) goes in the opposite direction to the standard large-deviation or small-ball estimates for X, and so for completeness we provide a proof. It is clear for experts that the statement of the lemma basically reduces to the one-dimensional case using the localization technique, and so our task is ultimately one-dimensional. However, to make the argument as short as possible, we prefer to rely on the available tools in the literature, and do not make any attempts to obtain good numerical constants. We therefore turn to the known reverse-Holder Khinchine-type inequalities between positive and negative moments of f (X) = X2−E2. Such inequalities between positive moments of semi-norms are classical and readily follow by Borell's Lemma [9] (see e.g. [41]); these have been extended to p-th order moments for p = 0 and p ∈ (−1, 0) in [29, 24]. However, we require such moment inequalities for the more general function f , and thus turn to the reverse-Holder inequalities initiated by Bourgain in [11] for polynomials, and extended using the localization technique by Bobkov[5], Carbery and Wright [16] and Nazarov, Sodin and Volberg [42]. For more on this topic, we refer to the excellent survey paper of Fradelizi [19], which also extends these results even further. Theorem 4.3 ([42, 19, 8]). Let P denote a degree d polynomial on Rn. Then for any −1/d < q ≤ p < ∞ and random-vector X distributed according to a log-concave measure on Rn: for some (explicit) constant C(p, q, d) > 0 depending solely on its arguments. kP (X)kp ≤ C(p, q, d)kP (X)kq , Here and below we use kY kp to denote E(Y p) p = 0, which in any case we will not require). 1 p (with the usual interpretation when Proof of Lemma 4.2. By the union bound and Markov-Chebyshev inequality: P(X − E ≤ c1S) ≤ P(X ≥ 8E) + P(X − E ≤ c1S ∧ X ≤ 8E) ≤ 1/8 + P((cid:12)(cid:12)X2 − E2(cid:12)(cid:12) ≤ 9c1SE) . 14 The function f (x) = x2 − E2 is a degree two polynomial, and so applying Theorem 4.3 with q = −1/4 and p = 2, preceded by Chebyshev's inequality, we have: P((cid:12)(cid:12)X2 − E2(cid:12)(cid:12) ≤ 9c1SE) ≤ (9c1SE)1/4 kf (X)k−1/4 Since: −1/4 ≤ C′(9c1SE)1/4 kf (X)k−1/4 2 . kf (X)k2 ≥ E kX − Ek2 = ES , we see that choosing the constant c1 > 0 small enough, we can ensure that: P(X − E ≤ c1S) ≤ 1/4 . Now set ε := P(X ≤ E − c1S), and observe that: 0 = E(X − E) ≥ (P(X ≥ E + c1S) − P (X ≤ E + c1S)) c1S −Z E P(X ≤ E − (c1 + t)S)dt(cid:19) . ≥ S(cid:18)(1/2 − 2ε)c1 −Z ∞ c1S 0 P(X ≤ E − r)dr (4.2) Since P(X ≤ E + 2S) ≥ 3/4 by Chebyshev's inequality, we may argue exactly as in the proof of Lemma 3.6 that the log-concavity of the function r 7→ P(X ≤ r) ensures that: P(X ≤ E − (c1 + t)S) ≤ ε exp(cid:18)− log 3/4 − log ε 2 + c1 t(cid:19) ∀t ≥ 0 . Plugging this into (4.2), we obtain: 0 ≥ (1/2 − 2ε)c1 − ε 2 + c1 log 3/4 − log ε . However, this is clearly impossible if ε ≥ 0 is too small, and so ε must be bounded from below by some universal constant c2 > 0. Setting c0 = min(c1, c2), the asserted claim follows. References [1] D. Bakry and M. ´Emery. Diffusions hypercontractives. In S´eminaire de probabilit´es, XIX, 1983/84, volume 1123 of Lecture Notes in Math., pages 177 -- 206. Springer, Berlin, 1985. [2] D. Bakry and M. Ledoux. L´evy-Gromov's isoperimetric inequality for an infinite- dimensional diffusion generator. Invent. Math., 123(2):259 -- 281, 1996. [3] S. Bobkov. On isoperimetric constants for log-concave probability distributions. In Geometric aspects of functional analysis, Israel Seminar 2004-2005, volume 1910 of Lecture Notes in Math., pages 81 -- 88. Springer, Berlin, 2007. [4] S. Bobkov. On Milman's ellipsoids and M-position of convex bodies. In C. Houdr´e, M. Ledoux, E. Milman, and M. Milman, editors, Concentration, Functional In- equalities and Isoperimetry, volume 545 of Contemporary Mathematics, pages 23 -- 34. Amer. Math. Soc., 2011. 15 [5] S. G. Bobkov. Remarks on the growth of Lp-norms of polynomials. In Geometric aspects of functional analysis, volume 1745 of Lecture Notes in Math., pages 27 -- 35. Springer, Berlin, 2000. [6] S. G. Bobkov. Large deviations and isoperimetry over convex probability measures with heavy tails. Electron. J. Probab., 12:1072 -- 1100 (electronic), 2007. [7] S. G. Bobkov and C. Houdr´e. Isoperimetric constants for product probability mea- sures. Ann. Probab., 25(1):184 -- 205, 1997. [8] S. G. Bobkov and F. L. Nazarov. Sharp dilation-type inequalities with fixed pa- rameter of convexity. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI), 351(Veroyatnost i Statistika. 12):54 -- 78, 299, 2007. [9] Ch. Borell. Convex measures on locally convex spaces. Ark. Mat., 12:239 -- 252, 1974. [10] Ch. Borell. The Brunn -- Minkowski inequality in Gauss spaces. Invent. Math., 30:207 -- 216, 1975. [11] J. Bourgain. On the distribution of polynomials on high dimensional convex sets. In Geometric Aspects of Functional Analysis, volume 1469 of Lecture Notes in Mathe- matics, pages 127 -- 137. Springer-Verlag, 1991. [12] H. J. Brascamp and E. H. Lieb. On extensions of the Brunn-Minkowski and Pr´ekopa- Leindler theorems, including inequalities for log concave functions, and with an application to the diffusion equation. J. Func. Anal., 22(4):366 -- 389, 1976. [13] Y. Brenier. Polar factorization and monotone rearrangement of vector-valued func- tions. Comm. Pure Appl. Math., 44(4):375 -- 417, 1991. [14] P. Buser. A note on the isoperimetric constant. Ann. Sci. ´Ecole Norm. Sup. (4), 15(2):213 -- 230, 1982. [15] L. A. Caffarelli. Monotonicity properties of optimal transportation and the FKG and related inequalities. Comm. Math. Phys., 214(3):547 -- 563, 2000. [16] A. Carbery and J. Wright. Distributional and Lq norm inequalities for polynomials over convex bodies in Rn. Math. Res. Lett., 8(3):233 -- 248, 2001. [17] J. Cheeger. A lower bound for the smallest eigenvalue of the Laplacian. In Problems in analysis (Papers dedicated to Salomon Bochner, 1969), pages 195 -- 199. Princeton Univ. Press, Princeton, N. J., 1970. [18] D. Cordero-Erausquin, M. Fradelizi, and B. Maurey. The (B) conjecture for the Gaussian measure of dilates of symmetric convex sets and related problems. J. Funct. Anal., 214(2):410 -- 427, 2004. [19] M. Fradelizi. Concentration inequalities for s-concave measures of dilations of Borel sets and applications. Electron. J. Probab., 14:no. 71, 2068 -- 2090, 2009. [20] S. Gallot, D. Hulin, and J. Lafontaine. Riemannian geometry. Universitext. Springer-Verlag, Berlin, 1987. [21] R. J. Gardner. The Brunn-Minkowski inequality. Bull. Amer. Math. Soc. (N.S.), 39(3):355 -- 405, 2002. [22] M. Gromov and V. D. Milman. A topological application of the isoperimetric in- equality. Amer. J. Math., 105(4):843 -- 854, 1983. [23] M. Gromov and V. D. Milman. Generalization of the spherical isoperimetric inequal- ity to uniformly convex Banach spaces. Compositio Math., 62(3):263 -- 282, 1987. [24] O. Gu´edon. Kahane-khinchine type inequalities for negative exponent. Mathe- matika, 46:165 -- 173, 1999. [25] O. Gu´edon and E. Milman. Interpolating thin-shell and sharp large-deviation es- timates for isotropic log-concave measures. Geom. Func. Anal.,, 21(5):1043 -- 1068, 2011. [26] L. Hormander. L2 estimates and existence theorems for the ¯∂ operator. Acta Math., 113:89 -- 152, 1965. [27] R. Kannan, L. Lov´asz, and M. Simonovits. Isoperimetric problems for convex bodies and a localization lemma. Discrete Comput. Geom., 13(3-4):541 -- 559, 1995. 16 [28] Y.-H. Kim and E. Milman. A generalization of caffarelli's contraction theorem via (reverse) heat flow. to appear in Math. Annal., arxiv.org/abs/1002.0373, 2010. [29] R. Lata la. On the equivalence between geometric and arithmetic means for log- concave measures. In Convex geometric analysis (Berkeley, CA, 1996), volume 34 of Math. Sci. Res. Inst. Publ., pages 123 -- 127. Cambridge Univ. Press, Cambridge, 1999. [30] M. Ledoux. The concentration of measure phenomenon, volume 89 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2001. [31] M. Ledoux. Spectral gap, logarithmic Sobolev constant, and geometric bounds. In Surveys in differential geometry. Vol. IX, pages 219 -- 240. Int. Press, Somerville, MA, 2004. [32] P. Li and S.T. Yau. Estimates of eigenvalues of a compact Riemannian manifold. In Geometry of the Laplace operator (Proc. Sympos. Pure Math., Univ. Hawaii, Honolulu, Hawaii, 1979), Proc. Sympos. Pure Math., XXXVI, pages 205 -- 239. Amer. Math. Soc., Providence, R.I., 1980. [33] L. Lov´asz and M. Simonovits. Random walks in a convex body and an improved volume algorithm. Random Structures Algorithms, 4(4):359 -- 412, 1993. [34] V. G. Maz′ja. Classes of domains and imbedding theorems for function spaces. Dokl. Acad. Nauk SSSR, 3:527 -- 530, 1960. Engl. transl. Soviet Math. Dokl., 1 (1961) 882 -- 885. [35] V. G. Maz′ja. The negative spectrum of the higher-dimensional Schrodinger opera- tor. Dokl. Akad. Nauk SSSR, 144:721 -- 722, 1962. Engl. transl. Soviet Math. Dokl., 3 (1962) 808 -- 810. [36] E. Milman. On the role of convexity in functional and isoperimetric inequalities. Proc. London Math. Soc., 99(3):32 -- 66, 2009. [37] E. Milman. On the role of convexity in isoperimetry, spectral-gap and concentration. Invent. Math., 177(1):1 -- 43, 2009. [38] E. Milman. Isoperimetric and concentration inequalities - equivalence under curva- ture lower bound. Duke Math. J., 154(2):207 -- 239, 2010. [39] E. Milman. Properties of isoperimetric, inequalities via concentration. arxiv.org/abs/0909.0207, 2010. to appear functional and transport-entropy in Prob. Theor. Rel. Fields, [40] E. Milman. Isoperimetric bounds on convex manifolds. In C. Houdr´e, M. Ledoux, E. Milman, and M. Milman, editors, Concentration, Functional Inequalities and Isoperimetry, volume 545 of Contemporary Mathematics, pages 195 -- 208. Amer. Math. Soc., 2011. [41] V. D. Milman and G. Schechtman. Asymptotic theory of finite-dimensional normed spaces, volume 1200 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. With an appendix by M. Gromov. [42] F. Nazarov, M. Sodin, and A. Vol′berg. The geometric Kannan-Lov´asz-Simonovits lemma, dimension-free estimates for the distribution of the values of polynomials, and the distribution of the zeros of random analytic functions. Algebra i Analiz, 14(2):214 -- 234, 2002. [43] L. E. Payne and H. F. Weinberger. An optimal Poincar´e inequality for convex domains. Arch. Rational Mech. Anal., 5:286 -- 292, 1960. [44] V. N. Sudakov and B. S. Cirel′son [Tsirelson]. Extremal properties of half-spaces for spherically invariant measures. Zap. Naucn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI), 41:14 -- 24, 165, 1974. Problems in the theory of probability distri- butions, II. [45] J. Q. Zhong and H. C. Yang. On the estimate of the first eigenvalue of a compact Riemannian manifold. Sci. Sinica Ser. A, 27(12):1265 -- 1273, 1984. 17
1212.1099	2	1212	2013-08-01T10:55:29	Measures and Dirichlet forms under the Gelfand transform	[ "math.FA", "math.GN", "math.PR" ]	Using the standard tools of Daniell-Stone integrals, Stone-\v{C}ech compactification and Gelfand transform, we discuss how any Dirichlet form defined on a measurable space can be transformed into a regular Dirichlet form on a locally compact space. This implies existence, on the Stone-\v{C}ech compactification, of the associated Hunt process. As an application, we show that for any separable resistance form in the sense of Kigami there exists an associated Markov process.	math.FA	math	MEASURES AND DIRICHLET FORMS UNDER THE GELFAND TRANSFORM MICHAEL HINZ1 , 2, DANIEL KELLEHER2, AND ALEXANDER TEPLYAEV2 Dedicated with deep respect to Professor Ildar Ibragimov on the occasion of his eightieth birthday. Abstract. Using the standard tools of Daniell-Stone integrals, Stone- Cech compactification and Gelfand transform, we discuss how any Dirichlet form defined on a measurable space can be trans- formed into a regular Dirichlet form on a locally compact space. This implies existence, on the Stone- Cech compactification, of the associated Hunt process. As an application, we show that for any separable resistance form in the sense of Kigami there exists an associated Markov process. Contents Introduction 1. 1.1. Acknowledgements 2. Gelfand theory and the Daniell-Stone Theorem 3. Multiplicative Stonean vector lattices 4. Positive linear functionals and measures 5. Dirichlet forms under the Gelfand map 6. Beurling-Deny decomposition and energy measures 7. Separation of points and separable resistance forms References 1 5 5 6 9 11 12 15 18 1. Introduction The main object of our study is a Dirichlet forms (E, F ) on the L2-space over a measure space (X, X , µ). The notion of the Dirichlet Date: September 16, 2018. 2010 Mathematics Subject Classification. 60J25, 60J35, 28A80, 81Q35. Key words and phrases. Regular symmetric Dirichlet form, C* algebra, Daniell- Stone integral, Stone-Cech compactification, Gelfand transform, fractals. 1Research supported in part by the Alexander von Humboldt Foundation Feodor (Lynen Research Fellowship Program). 2Research supported in part by NSF grant DMS-0505622. 1 2 HINZ, KELLEHER, AND TEPLYAEV form means that E is a closed nonnegative (bilinear) quadratic form on L2(X, X , µ) with a dense domain F ⊂ L2(X, X , µ). Moreover (E, F ) has what is called Markov (or positivity preserving, or normal contrac- tion property): if u ∈ F then ¯u = min(u, 1) ∈ F and E(¯u, ¯u) 6 E(u, u). By the combination of the standard theories of quadratic forms on Hilbert spaces, the spectral theory of self-adjoint operators and the Hille-Yosida theorem, there exists an associated self-adjoint operator (non-negative or non-positive, depending on the analytic or probabilis- tic conventions), which generates a positivity preserving contraction semigroup on L2(X, X , µ). This is equivalent to having a semigroup of transition probability kernels which, by the Kolmogorov's general theory of random process, is equivalent to the existence of a symmet- ric Markov process (in the usual way one may have to allow for the extinction of the process, or to augment the state space X with a "cemetery" point). This set up has generated an abundance of strong and well-known results, see e.g. [9, 13, 14, 18, 29, 30, 31], and re- cently was extensively used in analysis and probability on fractals, see [24, 32, 27]. However most of the basic results in the theory of Dirichlet forms and Markov processes rely on a set up where X is assumed to be a topological space. Examples include the classical Beurling-Deny decomposition for regular Dirichlet forms, the existence of energy mea- sures in the sense of Fukushima [18] and LeJan [28] or the existence of an associated Hunt process. To discuss them most references require X to be locally compact. Of course it is desirable to have versions of these theorems in more general situations (for instance for quasi- regular Dirichlet forms on Souslin spaces), and therefore a reduction of topological assumptions was one of the various directions into which the standard theory for regular Dirichlet forms has been extended. One of the typical strategies is to embed the possibly non-locally compact state space X into a larger but (locally) compact space and to transfer the Dirichlet form to this new space, where the standard theory for the locally compact case applies. See for instance [1, 2, 3, 10, 16, 29] for some applications of such compactification methods. In [1] this idea was used to prove a Beurling-Deny type theorem for quasi-regular Dirichlet forms on Hausdorff spaces X that are such that each compact is metrizable and its Borel σ-algebra is countably generated. Histor- ically the representation theoretic point of view upon Dirichlet forms already dates back to the work of Beurling and Deny, [6, 7], and was DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 3 later taken up by Fukushima in connection with regularization tech- niques, see [19] and [18, Appendix A.4]. A study of Banach algebras naturally induced by Dirichlet forms was carried out by Cipriani, [11]. The main ideas of the present note are not new but versions of these ideas. In contrast to the mentioned references we do not assume the given state space X to carry any topology (except for Section 7), and one item we would like to highlight in this context is the Daniell-Stone representation theorem, see e.g. [15]. Given a multiplicative Stonean vector lattice B of bounded real-valued functions on a set X we use the connection between the Daniell-Stone theorem and Gelfand's represen- tation theorem for C ∗-algebras to establish an injection of a suitable class of measures on X into the space of nonnegative Radon measures on the spectrum ∆ of the complex uniform closure of B. We apply this idea to show that for any given Dirichlet form over a measurable space there is a corresponding uniquely determined regular Dirichlet form on a larger and locally compact state space. We consider the algebra B(E) of bounded measurable functions on (X, X ) that are µ- square integrable and have finite energy. The uni- form closure of its complexification is a C ∗-algebra, and its spectrum If B(E) vanishes nowhere, ∆ is a locally compact Hausdorff space. then ∆ (roughly speaking) contains X as a dense subset, and there is a Radon measure µ on ∆ which is uniquely determined by µ in a way that makes the restriction of the Gelfand transform f 7→ f to B(E) an L2-isometry, i.e. (1) = kf kL2(X,µ) , f ∈ B(E). f(cid:13)(cid:13)(cid:13)L2(∆,µ) (cid:13)(cid:13)(cid:13) This allows to define a symmetric bilinear form by E( f , g) := E(f, g), f, g ∈ B(E). Our main result, Theorem 5.1, says that E, together with the image B(E) of B(E) under the Gelfand map, is closable, and its closure ( E, F) in L2(∆, µ) is a symmetric regular Dirichlet form. In other words, we can find a locally compact Hausdorff space ∆ which 'contains' the state space X, and a regular Dirichlet form ( E, F ) that is the image of (E, F ). For this Dirichlet form we can now apply the standard theory [18] and for instance obtain a Beurling-Deny representation and the existence of energy measures. We would like to point out that in [2] the embedding of a Souslin standard Borel space into the Gelfand spectrum of a countably generated and point separating algebra of continuous functions had been used to construct a symmetric Hunt process associated with the given Dirichlet form. 4 HINZ, KELLEHER, AND TEPLYAEV In contrast to references like [1, 2] it may not be possible to pull these results back to the Dirichlet form (E, F ) on the original state space X. For instance, the energy measure of ( E, F) on ∆ may be such that the image of X under the embedding into ∆ is of zero energy measure, see Example 6.1. This is reminiscent of the situation in infi- nite dimensional analysis where the Cameron-Martin space typically is a null set, cf. Remark 6.2 and such references as [20, 21, 22, 33]. The study of the Dirichlet form ( E, F) on the spectrum ∆ may be a natural way to enlarge the space to support energy measures. Under additional topological assumptions we can recover results similar to those in [1, 2]. Before we turn to Dirichlet forms we discuss how to naturally relate suitable measures µ on X to Radon measures µ on ∆. This corre- spondence relies on a connection between the Daniell-Stone theorem and the Gelfand transform. Although this idea is not new, see for in- stance [17], it does not seem to be all too widely used. We consider a multiplicative vector lattice B of bounded real-valued functions on X. The uniform closure A(B) of its complexification is a commutative C ∗-algebra. If µ is uniquely associated with a positive linear functional on B then we may use positivity arguments to obtain a uniquely as- sociated positive linear functional on the space Cc(∆, R) of real-valued compactly supported functions on the spectrum of A(B). By the Riesz representation theorem this functional can be represented by integra- tion with respect to some uniquely determined Radon measure µ on ∆. Proceeding this way we obtain an injective mapping from a cone of nonnegative measures on X into the cone of nonnegative Radon mea- sures on ∆. The isomorphism property of the Gelfand transform finally yields the L2-isometry (1). The paper is organized as follows. For convenience, we recall some preliminaries concerning Gelfand theory and the Daniell-Stone theo- rem in the section 2. In Section 3 we investigate the connection for multiplicative Stonean vector lattices of bounded real-valued functions and establish some lemmas on positivity, support properties and dense- ness. The main result of Section 4 is Theorem 4.1, which states the correspondence between measures on X and ∆. As a consequence we also obtain the L2-isometry (1). In Section 5 we apply these results to Dirichlet forms to obtain the closability of ( E, B(E)) in L2(∆, µ) and the regularity of its closure ( E, F ), Theorem 5.1. Consequences include the Beurling-Deny representation and the existence of Radon energy measures for the transferred Dirichlet form ( E, F) on ∆, sketched in Section 6. DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 5 We write C0(∆) to denote the space of continuous functions on ∆ that vanish at infinity and Cc(∆) to denote its subspace of functions with compact support. For their subspaces of real-valued functions we write C0(∆, R) and Cc(∆, R), respectively, and we will do similarly for other function spaces. If the index set of a sequence is not specified, it is the set of natural numbers, and if corresponding limits are taken, they are taken with the index going to infinity. 1.1. Acknowledgements. Helpful discussions with Mikhail Gordin, Masha Gordina, Naotaka Kajino, Jun Kigami and Takashi Kumagai are gratefully acknowledged. 2. Gelfand theory and the Daniell-Stone Theorem For multiplicative vector lattices of bounded real valued functions the theorem of Daniell-Stone can be connected to Gelfand's representation theorem for commutative C ∗-algebras. In this section we briefly recall these two concepts. We start with remarks on commutative Gelfand theory, cf. [5, 8, 23]. Let A be a commutative C ∗-algebra of bounded functions a : X → C, with the supremum norm k·k and with the algebra operations defined pointwise and the involution ∗ defined by complex conjugation a∗ := a. By ∆(A) we denote the spectrum (Gelfand space) of A, the space of continuous, complex-valued, multiplicative functionals on A. Equipped with the Gelfand topology the spectrum ∆(A) becomes a regular locally compact Hausdorff space, cf. If A contains the constant function 1 then ∆(A) is compact. The space ∆(A) is second countable if and only if the C ∗-algebra A is separable, and this in turn is equivalent to A being countably generated. For any a ∈ A the Gelfand transform a : ∆(A) → C of a is defined by a(ϕ) := ϕ(a), and by the Gelfand representation theorem the Gelfand map a 7→ a is seen to be an isometric ∗-isomorphism from the Banach algebra A onto the algebra C0(∆(A)) of continuous functions on ∆(A) vanishing at infinity. If the algebra A vanishes nowhere on X, that is, if for any x ∈ X there exists some a ∈ A such that a(x) 6= 0, then X may be identified with a subset of ∆(A) by the map ι : X → ∆(X), where [23]. (2) ι(x)(a) := a(x) , a ∈ A, for any x ∈ X. Note that multiplication in C0(∆(A)) is given pointwise, and ι(x)(a1a2) = (a1a2)(x) = a1(x)a2(x) = ι(x)(a1)ι(x)(a2) 6 HINZ, KELLEHER, AND TEPLYAEV for any x ∈ X and a1, a2 ∈ A. Thus, we observe the set-theoretic inclusion ι(X) ⊂ ∆(A). The set ι(X) is dense in ∆(A). For if not, we could find a nonzero function f ∈ C0(∆(A)) such that f (ι(x)) = 0 for all x ∈ X. Then, however, some nonzero a ∈ A would have to exist with a = f ∈ C0(∆(A)), hence (3) a(ι(x)) = ι(x)(a) = a(x) would have to be zero for all x ∈ X and consequently a ≡ 0 in A, a contradiction. The second tool we would like to sketch is the Daniell-Stone Theo- rem. Let X 6= 0 and let L be a real vector lattice of functions on X, i.e. a vector space of functions f : X → R that is closed under minimum and maximum operations f ∧ g = min(f, g) and f ∨ g = max(f, g). We assume that L possesses the Stone property: for any f ∈ L, f ∧ 1 ∈ L. By σ(L) we denote the σ-ring of subsets of X generated by L and by M+(σ(L)), the cone of (nonnegative) measures on σ(L). A positive linear functional I : L → R is called a Daniell integral on L if for any sequence (fn)n ⊂ L of nonnegative functions decreasing to zero pointwise at all x ∈ X also the sequence of integrals (I(fn))n decreases to zero. The Daniell-Stone Theorem says that for any Daniell integral I on L there exists a uniquely determined measure µ ∈ M+(σ(L)) on σ(L) such that (4) I(f ) =ZX f dµ , f ∈ L. See for instance [15]. We use the notation D(L) :=(cid:8)µ ∈ M+(σ(L)) : all functions from L are µ-integrable (cid:9) . If I is a Daniell integral on L then the measure µ uniquely associated with I by (4) is a member of D(L). Conversely any µ ∈ D(L) defines a Daniell integral on L by (4). Note that if L contains a strictly positive function, then all measures in D(L) are σ-finite, and if it contains the constant function 1, then all measures in D(L) are finite. 3. Multiplicative Stonean vector lattices We are interested in special cases to which both theories apply. Let B be a real multiplicative vector lattice of bounded functions on X 6= ∅ that has the Stone property. By B + iB we denote its complexification, that is the complex vector space of functions f1 + if2 with f1, f2 ∈ B. The vector space operations and the complex conjugation are defined pointwise. We endow B + iB with the supremum norm k·k and denote DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 7 its closure by A(B), clearly a Banach space. Pointwise multiplication turns A(B) into a commutative Banach algebra, and with the invo- lution ∗ defined by complex conjugation it becomes a commutative C ∗-algebra. Under the Gelfand transform f 7→ f the C ∗-algebra A(B) is isometrically ∗-isomorphic to C0(∆(A(B))). To shorten notation we will write ∆ to abbreviate ∆(A(B)). From now on we will assume the following. Assumption 3.1. The space B vanishes nowhere. Under this assumption the set ι(X), where ι is defined as in (2) with A = A(B), is a dense subset of ∆, and according to (3) we have f (ι(x)) = f (x) for any f ∈ B and x ∈ X. To discuss nonnegativity issues let A(B)+ and C0(∆)+ denote the cones of real-valued nonnegative functions in A(B) and C0(∆), respec- tively. For a real-valued function f we write f + = max(f, 0) and f − = max(−f, 0). If f is a member of B then so are f + and f −. Lemma 3.1. A function f ∈ A(B) is real-valued if and only if f ∈ C0(∆) is. Moreover, we have f ∈ A(B)+ if and only if f ∈ C0(∆)+. This lemma is a consequence of (3) together with the denseness of ι(X) in ∆. Lemma 3.2. For any real-valued f ∈ A(B) we have (f +)∧ = f + and (f −)∧ = f −. Proof. For any x ∈ X we have (f +)∧(ι(x)) = f +(x) by (3). If f (x) ≥ 0 then f +(x) = f (x) = f (ι(x)) = f +(ι(x)). If f (x) < 0 then f (ι(x)) < 0 and f +(x) = 0. Consequently (f +)∧(ι(x)) = f +(ι(x)) for all x ∈ X, and by linearity also (f −)∧(ι(x)) = f −(ι(x)). By continuity and the denseness of ι(X) in ∆ the lemma follows. (cid:3) The members of A(B)+ are all monotone limits of nonnegative func- tions from B. For this statement Assumption 3.1 is not needed. Lemma 3.3. For any function f ∈ A(B)+ there exists a monotoni- cally increasing sequence (fn)n of nonnegative functions fn ∈ B that converges to f pointwise. Proof. By the lattice property in B, we can see that there is a se- quence (gn)n of nonnegative functions gn ∈ B converging uniformly to f . We may assume that the nonnegative numbers δn := supX gn−gn+1 are such that Pn δn < ∞ (otherwise pass to a subsequence). Setting fn := gn − gn ∧ (P∞ k=n δk) we obtain a sequence (fn)n with the desired (cid:3) properties. 8 HINZ, KELLEHER, AND TEPLYAEV We discuss compactly supported functions. If B contains the con- stant functions, then ∆ is compact, hence every function in B has compact support. To formulate a result for the general case, set Bc := {ϕ ∈ B : ϕ ∈ Cc(∆)} . Clearly Bc is again a multiplicative vector lattice having the Stone property. Lemma 3.4. The space Bc is uniformly dense in B. To prove Lemma 3.4 we use a property of upper level sets. Given ϕ ∈ B and k ∈ N \ {0} set Nk(f ) :=(cid:26)x ∈ X : f (x) ≥ 1 k(cid:27) . Lemma 3.5. For any f ∈ B and any k the closure of the set ι(Nk(f )) is compact in ∆. Proof. We have f ∈ B and, according to Lemma 3.2, f ∧ = f . Consequently we may assume f ≥ 0. Since f ∈ C0(∆), the closed set Lk(f ) :=(cid:26)y ∈ ∆ : f (y) ≥ 1 k(cid:27) is contained in a compact set and therefore compact itself. On the other hand ι(Nk(f )) ⊂ Lk(f ), what implies that ι(Nk(f )) is a closed subset of Lk(f ), hence compact. (cid:3) We verify Lemma 3.4. Proof. It suffices to show that nonnegative functions can be approxi- mated. Given f ∈ B with f ≥ 0 consider the functions ϕk := f − f ∧ 1 k . Obviously the sequence (ϕk)k uniformly converges to f , and for fixed k the set N k := {x ∈ X : ϕk(x) > 0} is a subset of Nk(f ). On the other hand, we have {y ∈ ∆ : ϕk(y) > 0} ⊂ ι(N k). For if there were some y ∈ ∆ with ϕk(y) > 0 having an open neigh- borhood Uy such that ϕk(x) = 0 for all x ∈ X with ι(x) ∈ Uy, then we would have ϕk(z) = 0 for all z ∈ Uy by the density of ι(X) in ∆, a contradiction. It also follows that supp ϕk ⊂ ι(N k) ⊂ Nk(f ), DIRICHLET FORMS UNDER THE GELFAND TRANSFORM and Lemma 3.5 implies that supp ϕk is compact. 9 (cid:3) 4. Positive linear functionals and measures In this section we establish a correspondence between suitable mea- sures µ on X and Radon measures µ on ∆ and list some consequences. As before we assume that B is a Stonean multiplicative vector lattice of bounded real-valued functions on X. Let I : B → R be a positive linear functional. Given a function f ∈ A(B)+ and an increasing sequence (fn)n ⊂ B of nonnegative function as in Lemma 3.3, we set (5) I(f ) := sup n I(fn). The lattice property of B guarantees that (5) provides a well-defined positive linear (i.e. positively homogeneous and additive) functional I : A(B)+ → [0, +∞]. In what follows let Assumption 3.1 be satisfied. In view of Lemma 3.1 we can then define a bounded positive linear functional I : C0(∆)+ → [0, +∞] by I( f ) := I(f ), f ∈ C0(∆)+, (6) and according to Lemma 3.2 we may set I(f ) := I(f +) − I(f −) and I( f ) := I(f ) to extend (6) to all f ∈ B. Let M+(∆) denote the cone of nonnegative Radon measures on ∆. The Riesz representation theorem ensures the existence of a uniquely determined µ ∈ M+(∆) such that for any f ∈ Cc(∆) we have (7) I( f ) =Z∆ f dµ. Remark 4.1. 2 Recall that to prove the existence part of the Riesz representation theorem one usually sets µ(K) := infn I( f ) : f ∈ Cc(∆, R), and f ≥ 1Ko for compact K ⊂ ∆ and defines the µ-measure of an arbitrary Borel set by inner approximation by compacts. It is therefore sufficient to know the functional I on the cone Cc(∆)+. Now assume that I : B → R is a Daniell integral and µ ∈ D(B) is the unique measure on σ(B) associated with I as in (4). In this case definition (6) yields (8) ZX f dµ =Z∆ f dµ, f ∈ B. 10 HINZ, KELLEHER, AND TEPLYAEV The map µ 7→ µ is positive and linear (i.e. additive and positively homogeneous). By (8) and the uniqueness part of the Daniell-Stone Theorem we obtain the following result. Theorem 4.1. The map µ 7→ µ is an injection of D(B) into M+(∆). We may also consider equivalence classes of functions. Lemma 4.1. Let µ ∈ D(B) and f ∈ B. Then f = 0 µ-a.e. on X if and only if f = 0 µ-a.e. on ∆. Proof. Let f = 0 µ-a.e. on X. Then also f + and f − vanish µ-a.e. on f +dµ = 0, hence f + = 0 µ-a.e. The same is true for f − and consequently f = 0 µ-a.e. The converse implication follows in a similar manner. (cid:3) X. By Lemma 3.2 and (8) therefore R∆ Therefore the Gelfand map induces a well-defined map from the space of µ-equivalence classes of functions from B into the space of µ-equivalence classes of functions on ∆. We denote it again by f 7→ f . We investigate corresponding L2-spaces. Lemma 4.2. Let µ ∈ D(B). For f ∈ B we have . kf kL2(X,µ) =(cid:13)(cid:13)(cid:13) f 2dµ =Z∆ f(cid:13)(cid:13)(cid:13)L2(∆,µ) (f 2)∧dµ =Z∆ Proof. Being an algebra homomorphism, the Gelfand map satisfies ( f )2 = (f 2)∧ for any f ∈ A(B). For f ∈ B the identity (8) then yields ZX f 2dµ. (cid:3) The following fact will be used in the next section. Lemma 4.3. For any µ ∈ D(B) the image B of B is dense in L2(∆, µ, R). Proof. Since C0(∆, R) is a dense subspace of L2(∆, µ, R), it suffices to show that any f ∈ C0(∆, R) can be approximated in L2(∆, µ, R) by functions from B. However, as C0(∆) is isometrically isomorphic to the uniform closure A(B) of the complexification of B, there is a sequence (fn)n ⊂ B such that ( fn)n approximates f uniformly. Given ε > 0 we can find a compact set Kε ⊂ ∆ such that µ(∆ \ Kε) < ε. Then obviously lim n ZKε fn − f 2dµ = 0 DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 11 and Z∆\Kε fn − f 2dµ ≤ ε(kf k + sup n kfnk). (cid:3) 5. Dirichlet forms under the Gelfand map We use the setup of the previous section to transfer from a Dirich- let form on a measure space to a regular Dirichlet form on a locally compact second countable Hausdorff space. Let (X, X , µ) be a measure space and (E, F ) a Dirichlet form on L2(X, µ, R), see for example [9, Chapter I]. We will frequently use the shorthand notation E(f ) := E(f, f ) and do similarly for other bilin- ear expressions. The space of bounded measurable functions on X is denoted by bX . Set (9) B(E) := {f ∈ bX : the µ-equivalence class of f is in F ∩ L1(X, µ, R)} . The Cauchy-Schwarz inequality and the Markov property of (E, F ) imply that B(E) is a multiplicative vector lattice that has the Stone property. In addition we assume the following: Assumption 5.1. The space B(E) vanishes nowhere on X and is sepa- rable with respect to the supremum norm. Let ∆ be the spectrum of the uniform closure A(B(E)) of the com- plexification of B. For f, g ∈ B(E) we set (10) E( f , g) := E(f, g). Obviously E is a nonnegative definite symmetric bilinear form on the dense subspace B(E) =n f ∈ C0(∆, R) : f ∈ B(E)o of L2(∆, µ, R). It enjoys the Markov property. regular symmetric Dirichlet form on L2(∆, µ, R). In fact, it defines a Theorem 5.1. The form ( E, B(E)) is closable on L2(∆, µ, R). Its clo- sure ( E, F) defines a symmetric regular Dirichlet form. Proof. Let ( fn)n be a sequence of functions from B(E) that is E-Cauchy and tends to zero in L2(∆, µ, R). Then by (10) the sequence (fn)n of preimages fn ∈ B(E) of the functions fn under the Gelfand map is 12 HINZ, KELLEHER, AND TEPLYAEV E-Cauchy, and by Lemma 4.2 it tends to zero in L2(X, µ). From the closability of (E, F ) together with (10) it then follows that lim n E( fn) = lim n E(fn) = 0. Therefore ( E, B(E)) is closable. According to Lemma 3.4 the set Bc(E) :=n f ∈ Cc(∆) : f ∈ B(E)o is uniformly dense in B(E), hence also in C0(∆). On the other hand, given f ∈ B(E), the functions ϕk := f − (f ∨ (− 1 k )) ∧ 1 k converge to f in E1-norm, see for instance [18, Theorem 1.4.2]. Con- sequently Bc(E) is a core for ( E, F). Note that as a consequence of Assumption 5.1 the Gelfand spectrum ∆ of A(B(E)) is second count- able. (cid:3) To the symmetric regular Dirichlet form ( E, F) on L2(∆, µ, R) we refer as the transferred Dirichlet form. 6. Beurling-Deny decomposition and energy measures We record some consequences of the existing theory for Dirichlet forms on locally compact spaces when applied to ( E, F). As before let (X, X , µ) be a measure space and (E, F ) a symmetric Dirichlet form on L2(X, µ) such that Assumption 5.1 is satisfied. The first theorem is the Beurling-Deny representation. Theorem 6.1. The transferred Dirichlet form ( E, F) on L2(∆, µ, R) admits the decomposition E( f , g) = E c( f , g) +Z Z∆×∆ ( f (x) − f (y))(g(x) − g(y)) J(dx, dy) f (x)g(x)κ(dx) +Z∆ for any f , g ∈ B(E), where E c is a symmetric nonnegative definite bilin- ear form on B(E) that is strongly local, J is a symmetric nonnegative Radon measure on ∆ × ∆ \ {(x, x) : x ∈ ∆}, and k is a nonnegative Radon measure on ∆. The normal contraction operates on E c, and the triple ( E c, J, k) is uniquely determined. For a proof see for instance [4] or [18]. DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 13 Remark 6.1. Note that these proofs require the local compactness but not the second countability of ∆. However, if B(E) has a countable subset from which any element in B(E) can be produced by linear operations, multiplication, truncation by 1 and taking uniform limits, then ∆ is second countable and by Urysohn's theorem there exists a metric turning ∆ into a locally compact separable metric space. Another result is the existence of energy measures for the transferred Dirichlet form ( E, F), which is an immediate consequence its regularity, [18, 28]. Theorem 6.2. For any f ∈ B(E) there exists a uniquely determined finite nonnegative Radon measure Γ( f ) on ∆ such that 2Z∆ for any ϕ ∈ B(E). ϕdΓ( f ) = 2 E( ϕ f , f ) − E( ϕ, f 2) If the original Dirichlet form (E, F ) itself admits energy measures, that is if for any f ∈ B(E) there exists some nonnegative measure Γ(f ) such that (11) ϕdΓ(f ) = 2E(ϕf, f ) − E(ϕ, f 2), ϕ ∈ B(E), 2ZX then the energy measures Γ( f ) are consistent with these original ones. Theorem 6.3. Assume that (E, F ) admits energy measures (11). Then for any f ∈ B(E) we have (Γ(f ))∧ = Γ( f ). Proof. For any ϕ ∈ C0(∆) we have Z∆ ϕd(Γ(f ))∧ =ZX ϕdΓ(f ) = 2E(f ϕ, f ) − E(ϕ, f 2) = 2 E((f ϕ)∧, f ) − E( ϕ, (f 2)∧) = 2 E( f ϕ, f ) − E( ϕ, f 2) ϕdΓ( f ). =Z∆ (cid:3) Theorem 6.2 is significant, because as the following examples show, the original Dirichlet form (E, F ) itself may not admit energy measures. 14 HINZ, KELLEHER, AND TEPLYAEV Examples 6.1. Consider the classical Dirichlet integral on the unit in- terval [0, 1], given by E0(g) :=Z 1 0 g′(x)2dx for any function g from F0 := {g ∈ C([0, 1]) : E(g) < ∞} . The form (E0, F0) is a resistance form on [0, 1] in the sense of Kigami [25, 26]. We consider the countable state space X = Q ∩ [0, 1]. Set F0X := {f : X → R : there exists some g ∈ F0 such that f = gX} and E(f ) := E0(g), f ∈ F0X. Here gX denotes the pointwise restriction of the continuous function g to X. By continuity and the density of X in [0, 1] each f ∈ F0X is the restriction of exactly one function g ∈ F0. Now let δq denote the normed Dirac point measure at a given point q and let {qn}∞ n=1 be an enumeration of X. Then µ := ∞ Xn=1 2−nδqn is a probability measure. The form (E, F0X) is closable in L2(X, µ), see for instance [26, Lemma 9.2 and Theorem 9.4], and its closure (E, F ) is a Dirichlet form. For a function f ∈ F0X with E(f ) > 0 (such as for instance the restriction to X of a nonconstant linear function) and g ∈ F0 is such that f = gX we have (12) 2E(ϕf, f ) − E(ϕ, f 2) = 2Z 1 0 ψ(x)g′(x)2dx, for all ϕ ∈ F0X with ϕ = ψX, ψ ∈ F0. On the other hand approxima- tion by piecewise linear functions shows that F0 is dense in C([0, 1]), and consequently any bounded Borel function on [0, 1] can be approx- imated pointwise by a uniformly bounded sequence of functions from F0. Let (ψn)n ⊂ F0 be a uniformly bounded sequence of functions that approximate 1X pointwise. If for some f as above (E, F ) would admit energy measures as in (11) then we would obtain ZX ψnX dΓ(f ) =Z∆ (ψnX)∧ dΓ(f ) =Z 1 0 ψn(x)g′(x)2dx, DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 15 and by bounded convergence E(f ) = Γ(f )(X) =ZX g′(x)2dx = 0, because the restriction of g′(x)2dx to X is the zero measure. This contradicts E(f ) > 0. Remark 6.2. In some sense the situation of Example 6.1 displays a similar feature as we encounter it for Dirichlet forms on infinite dimen- sional spaces. For instance, let (E, H, µ) be an abstract Wiener space, cf. [9, 20, 21, 29, 33], let F C ∞ b := {f (l1, ..., ln) : n ∈ N, f ∈ C ∞ b (Rn), l1, ..., ln ∈ E′} , h∇u(z), hiH := for any u ∈ F C ∞ b and ∂u ∂h (z), h ∈ H, k∇uk2 H dµ. E(u) :=ZE Then (E, F C ∞ b ) is closable on L2(E, µ) and its closure (E, F ) is a Dirich- let form. Its energy measure is given by k∇uk2 H dµ on E. However, as the Gaussian measure µ is quasi-invariant under translations by ele- ments of the (infinite dimensional) generalized Cameron-Martin space H, the space H has zero Gaussian measure, hence zero energy measure. In other words, the space H is too small to carry a nontrivial energy measure, but on the larger space E the energy measures generally are nontrivial. 7. Separation of points and separable resistance forms In addition to Assumption 3.1 respectively 5.1 we now assume the following. Assumption 7.1. The space B separates points, that is for each x, y ∈ X, there are f ∈ B such that f (x) 6= f (y). An immediate consequence of this assumption is that ι : X → ∆ is injective, so X is embedded in ∆ as ι(X). Thus we will use X and ι(X) interchangeably. We further assume that ι(X) is a Borel set with respect to the Gelfand topology in ∆, although this assumption is technical and often can be weakened or eliminated, depending on the situation. Remark 7.1. (i) Assumption 7.1 leads to a situation similar to the one in [2, Section 2]. See also the references cited there. 16 HINZ, KELLEHER, AND TEPLYAEV (ii) If B does not separate points, one can define an an equivalence relation ∼ on X by x ∼ y if f (x) = f (y) for all f ∈ B. Then all functions in B naturally define functions on the quotient space X = X/ ∼, and functions in B separates equivalent classes. In this case, ι : X → ∆, defined by ι([x]) = ι(x) is an embedding with ι( X) = ι(X). In light of 4.1, any µ ∈ D(B) can be extended to a positive measure on ∆. By Assumption 7.1 we may consider the measure of X in ∆. In particular, for any σ-finite µ ∈ D(B), we can extend µ to ∆ either by considering µ, or by ν(A) = µ(A ∩ X). However, by equation (8) and the Riesz representation theorem, µ and ν coincide. The fact that X is a set of full measure µ allows us a technique for extending results for Dirichlet forms on locally compact spaces to a more general class of spaces. The following result is a version of [2, Theorem 2.7]. Proposition 7.1. Since E is a regular Dirichlet form on ∆, there is a µ-symmetric Hunt process on ∆ with Dirichlet form E. Since A(B(E)) separates points, X is naturally identified as a subset of ∆ with full µ-measure. By [18, Lemma 4.1.1], this implies that the process on ∆ is contained in X with probability 1, thus can be thought of as a process on X. Note that we do not claim that this process is a Hunt process on X because we do not consider X as a topological space. However the random process is well defined, which is useful in some applications such as the following. In what follows we will consider a special class of Dirichlet forms, the resistance forms of Kigami [24, 25, 26], for which points have positive capacity. For simplicity we define these forms in the separable case, which can be essentially reduced to a form on a countable set. Definition 7.1. A pair (E, F ) is called a resistance form on a countable set V∗ if it satisfies: (RF1) F is a linear subspace of the functions V∗ → R that contains the constants, E is a nonnegative symmetric quadratic form on F , and E(u, u) = 0 if and only if u is constant. (RF2) The quotient of F by constant functions is Hilbert space with the norm E(u, u)1/2. DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 17 (RF3) If v is a function on a finite set V ⊂ V∗ then there is u ∈ F (RF4) For any x, y ∈ V∗ the effective resistance between x and y is with u(cid:12)(cid:12)V = v. R(x, y) = sup((cid:0)u(x) − u(y)(cid:1)2 E(u, u) : u ∈ F , E(u, u) > 0) < ∞. (RF5) (Markov Property.) If u ∈ F then ¯u(x) = max(0, min(1, u(x))) ∈ F and E(¯u, ¯u) 6 E(u, u). The resistance forms on countable sets are determined by a sequence of traces on finite subsets, as in the following two propositions. Proposition 7.2 ([24, 25, 26]). Resistance forms have the following properties. (i) R(x, y) is a metric on V∗. Functions in F are R-continuous, thus have unique R-continuous extension to the R-completion XR of V∗. (ii) If U ⊂ V∗ is finite then a Dirichlet form EU on U may be defined by EU (f, f ) = inf{E(g, g) : g ∈ F , g(cid:12)(cid:12)U = f } in which the infimum is achieved at a unique g. The form EU is called the trace of E on U, denoted EU = TraceU(E). If U1 ⊂ U2 then EU1 = TraceU1(EU2). that Vn ⊂ Vn+1 and S∞ Proposition 7.3 ([24, 25, 26]). Suppose Vn ⊂ V∗ are finite sets such n=0 Vn is R-dense in V∗. Then EVn(f, f ) is non- decreasing and E(f, f ) = limn→∞ EVn(f, f ) for any f ∈ F . Hence E is uniquely defined by the sequence of finite dimensional traces EVn on Vn. Conversely, suppose Vn is an increasing sequence of finite sets each supporting a resistance form EVn, and the sequence is compatible in that each EVn is the trace of EVn+1 on Vn. Then there is a resistance form E n=0 Vn such that E(f, f ) = limn→∞ EVn(f, f ) for any f ∈ F . The following theorem follows easily from the analysis presented above. See [Chapter 5][26] for discussion why the effective resistance metric is not suitable to define topology to produce a regular Dirichlet form. on V∗ =S∞ Theorem 7.1. There exists a finite measure µ on X = V∗ such that: (i) any point of X has positive measure and any function of finite energy is in L2(X, µ); (ii) E is a Dirichlet form on L2(X, µ); (iii) the embedding of X into the Gelfand spectrum ∆ yields a regular Dirichlet form on L2(∆, µ). 18 HINZ, KELLEHER, AND TEPLYAEV Moreover, one can see that for any other finite measure µ on V∗ one can obtain a regular Dirichlet form on L2(∆, µ) by modifying the domain. However the case of infinite measures is more delicate. For instance, in [Chapter 5][26] one can see that choosing the counting measure on V∗ may not produce a regular Dirichlet form, even though the space X is compact in the topology induced by the set of functions of finite energy (but is not locally compact in the topology induced by the effective resistance metric). References [1] S. Albeverio, Z.-M. Ma, M. Rockner, A Beurling-Deny type structure theorem for Dirichlet forms on general state spaces. -- in: Ideas and Methods in Math- ematical Analysis, Stochastics and Applications, 115-123 Ed. S. Albeverio, J.E. Fenstad, H. Holden, T. Lindstrøm, Cambridge Univ. Press, Cambridge 1992. [2] S. Albeverio, M. Rockner, Classical Dirichlet forms on topological vector spaces - construction of an associated diffusion process. -- Probab. Th. Rel. Fields 83 (1989), 405-434. [3] S. Albeverio, M. Rockner, Classical Dirichlet forms on topological vector spaces - closability and a Cameron-Martin formula. -- J. Funct. Anal. 88 (1990), 395- 436. [4] G. Allain, Sur la repr´esentation des formes de Dirichlet. -- Ann. Inst. Fourier 25 (1975), 1-10. [5] W. Arveson, An Invitation to C ∗-Algebras. Springer Graduate Texts in Math. 39, Springer, New York, 1976. [6] A. Beurling, J. Deny, Espaces de Dirichlet I, le cas ´el´ementaire. -- Acta Math. 99 (1958), 203-224. [7] A. Beurling, J. Deny, Dirichlet spaces -- Proc. Nat. Acad. Sci. U.S.A. 45 (1959), 208-215. [8] B. Blackadar, Operator Algebras: Theory of C ∗-Algebras and von Neumann Algebras. Encyclopedia of Math. Sciences 122, Springer, New York, 2006. [9] N. Bouleau, F. Hirsch, Dirichlet Forms and Analysis on Wiener Space. de- Gruyter Studies in Math. 14, deGruyter, Berlin, 1991. [10] Z.-Q. Chen, M. Fukushima, Symmetric Markov Processes, Time Change, and Boundary Theory, Princeton Univ. Press, Princeton, 2012. [11] F. Cipriani, Dirichlet forms as Banach algebras and applications, Pacific J. Math. 223 (2) (2006) -- 229-249. [12] J. Deny, M´ethodes Hilbertiennes et th´eorie du potentiel. CIME, Rome, 1970. [13] E.B. Dynkin, Foundations of the Theory of Markov Processes. Gosudarstv. Izdat. Fiz.-Mat. Lit., Moscow 1959, English translation Theory of Markov pro- cesses. Prentice-Halland Pergamon Press, 1965, Dover Pub., 2006. [14] E.B. Dynkin, Markov Processes. Gosudarstv. Izdat. Fiz.-Mat. Lit., Moscow 1963, English translation Academic Press and Springer-Verlag, 1965. [15] R.M. Dudley, Real Analysis and Probability. Cambridge studies in adv. math. 74, Cambridgr Univ. Press, Cambridge, 2002. [16] P.J. Fitzsimmons, Markov processes and nonsymmetri Dirichlet forms without regularity. -- J. Funct. Anal. 85 (1989), 287-306. DIRICHLET FORMS UNDER THE GELFAND TRANSFORM 19 [17] B. Fuchssteiner, When does the Riesz representation theorem hold ? -- Arch. Math. 28 (1977), 173-181. [18] M. Fukushima, Y. Oshima and M. Takeda, Dirichlet Forms and Symmetric Markov Processes. deGruyter, Berlin, New York, 1994. [19] M. Fukushima, Regular representations of Dirichlet spaces -- Trans. Amer. Math. Soc. 155 (2) (1971), 455-472. [20] L. Gross Potential theory on Hilbert space. -- J. Functional Analysis 1 (1967) 123 -- 181. [21] L. Gross Abstract Wiener spaces. -- 1967 Proc. Fifth Berkeley Sympos. Math. Statist. and Probability (Berkeley, Calif., 1965/66), Vol. II: Contributions to Probability Theory, Part 1 pp. 31 -- 42 Univ. California Press, Berkeley, Calif. [22] I.A. Ibragimov, Y.A. Rozanov, Gaussian Random Processes. Nauka, Moscow, 1970 and Springer, New York - Berlin, 1978. [23] E. Kaniuth, A Course in Commutative Banach Algebras. Springer, New York, 2009. [24] J. Kigami, Analysis on Fractals. Cambridge Tracts in Mathematics 143, Cam- bridge University Press, 2001. [25] J. Kigami, Harmonic analysis for resistance forms. -- J. Funct. Anal. 204 (2003), 525 -- 544. [26] J. Kigami, Resistance forms, quasisymmetric maps and heat kernel estimates. Mem. Amer. Math. Soc. 216 (2012). [27] A. A. Kirillov, A tale of two fractals. (in Russian) MCMNO 2010. [28] Y. LeJan, Mesures associ´ees `a une forme de Dirichlet. Applications. -- Bull. Soc. Math. France 106 (1978), 61-112. [29] Z.-M. Ma, M. Rockner, Introduction to the Theory of Non-Symmetric Dirichlet Forms, Universitext, Springer, Berlin, 1992. [30] M. Reed, B. Simon, Methods of Modern Mathematical Physics. I. Functional Analysis. Academic Press, 1980. [31] L.C.G. Rogers, D. Williams, Diffusions, Markov Processes, and Martingales. Volume one: Foundations, 2nd ed. Wiley, 1994. [32] R. S. Strichartz, Differential Equations on Fractals: A Tutorial. Princeton University Press, 2006. [33] V. N. Sudakov, Geometric problems of the theory of infinite-dimensional prob- ability distributions. Trudy Mat. Inst. Steklov. 141 (1976), 191 pp. Proc. Steklov Inst. Math., Springer (1979), 178pp. Mathematisches Institut, Friedrich-Schiller-Universitat Jena, Ernst- Abbe-Platz 2, 07737, Germany and Department of Mathematics, Uni- versity of Connecticut, Storrs, CT 06269-3009 USA E-mail address: [email protected] and [email protected] Department of Mathematics, University of Connecticut, Storrs, CT 06269-3009 USA E-mail address: [email protected] Department of Mathematics, University of Connecticut, Storrs, CT 06269-3009 USA E-mail address: [email protected]
1708.09640	4	1708	2018-08-30T16:45:25	Certain Liouville properties of eigenfunctions of elliptic operators	[ "math.FA", "math.AP", "math.PR" ]	We present certain Liouville properties of eigenfunctions of second-order elliptic operators with real coefficients, via an approach that is based on stochastic representations of positive solutions, and criticality theory of second-order elliptic operators. These extend results of Y. Pinchover to the case of nonsymmetric operators of Schr\"odinger type. In particular, we provide an answer to an open problem posed by Pinchover in [Comm. Math. Phys. 272 (2007), no. 1, 75-84, Problem 5]. In addition, we prove a lower bound on the decay of positive supersolutions of general second-order elliptic operators in any dimension, and discuss its implications to the Landis conjecture.	math.FA	math	CERTAIN LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Abstract. We present certain Liouville properties of eigenfunctions of second- order elliptic operators with real coefficients, via an approach that is based on stochastic representations of positive solutions, and criticality theory of second- order elliptic operators. These extend results of Y. Pinchover to the case of nonsymmetric operators of Schrodinger type. In particular, we provide an answer to an open problem posed by Pinchover in [Comm. Math. Phys. 272 (2007), no. 1, 75 -- 84, Problem 5]. In addition, we prove a lower bound on the decay of positive supersolutions of general second-order elliptic operators in any dimension, and discuss its implications to the Landis conjecture. Contents Introduction 1. Notation 2. Preliminaries and main results 3. Proofs of Theorems 2.1 to 2.4 4. A lower bound on the decay of eigenfunctions Acknowledgements References 1 3 3 12 23 31 31 1. Introduction The main objective of this paper is to establish Liouville properties of eigenfunc- tions of second-order elliptic operators. These type of results came to prominence after the paper of Pinchover [38] where he proved a very interesting property which can be stated as follows. Let D be a domain in Rd and let Pi = −div(Au) − Viu, i = 1, 2, be two nonnegative Schrodinger operators, with Vi ∈ Lp loc(D) for p > d/2, and A locally non-degenerate in D. Suppose that P1 is critical in D with ground state Ψ∗1, that the generalized principal eigenvalue of P2 is nonnegative, and that there exists a subsolution Ψ, with Ψ+ 6= 0, to P2u = 0 in D satisfying Ψ+ ≤ CΨ∗1 for some constant C. Then P2 is also critical with ground state Ψ. In partic- ular, the principal eigenvalue of P2 equals 0, and Ψ > 0. In the same paper, Pinchover proposed two problems on the generalization of this result for (a) gen- eral non-symmetric second order elliptic operators and (b) quasilinear operators of p-Laplacian type. Later, a similar result for p-Laplacian operators was proved by 2000 Mathematics Subject Classification. Primary 35J15; Secondary 35A02, 35B40, 35B60. Key words and phrases. Principal eigenvalue, Landis' conjecture, decay of eigenfunctions, Li- ouville property. 1 2 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Pinchover, Tertikas and Tintarev in [39]. However, the problem concerning general second-order elliptic operators remains open so far. The main goal of this paper is to address this problem for a large class of second-order elliptic operators. Pinchover's approach was variational. He first established the existence of a null sequence for the quadratic form associated with Pi, and then using criticality theory, together with a bound on the positive part of the subsolution, he obtained the above mentioned Liouville-type result. Unfortunately, for general (nonsymmetric) operators the existence of such a null sequence is not possible, despite the fact that criticality theory is well developed for general operators. Moreover, in [38, Remark 4.1] Pinchover discussed the difficulty in obtaining the above Liouville-type results for general (nonsymmetric) second-order elliptic operators. In this paper we show that the above Liouville-type result holds for a fairly general class of second- order elliptic operators and potentials. Our approach differs significantly from variational arguments, and relies on stochastic representations of positive solutions studied in [5], and criticality theory of second-order elliptic operators. This allows us to bypass the use of a null sequence. For D = Rd, it is known that the criticality of the operator is equivalent to the recurrence of the twisted process [5, 40]. An interesting observation in this paper is that criticality is also equivalent to the strict right monotonicity of the (generalized) principal eigenvalue. Let L be a second-order elliptic operator and λ∗(V ) denote the principal eigenvalue of the operator L + V , with potential V . We say that λ∗(V ) is strictly right monotone at V , or strictly monotone at V on the right, if λ∗(V ) < λ∗(V + h) for any non-zero, nonnegative continuous function h that vanishes at infinity. This equivalence is established in Theorem 2.1. We also show that given two potential functions Vi ∈ L∞loc(Rd), i = 1, 2, if V1 − V2 has a fixed sign outside some compact subset of Rd, then a result analogous to the one described in the preceding paragraph holds (see Theorems 2.2 and 2.3). In particular, if P1 = L1 + V1 is a small perturbation of P2 = L2 + V2, then, under suitable assumptions, the criticality of P1 implies that of P2. To further strengthen these results, we study the strict monotonicity of the principal eigenvalue, by which we mean that the principal eigenvalue is strictly left and right monotone at V (i.e., λ∗(V − h) < λ∗(V ) < λ∗(V + h)). It is shown in Theorem 2.4 that if the principal eigenvalue corresponding to P1 is strictly monotone, L1 = L2 outside a compact subset of Rd, and V1− V2 vanishes at infinity, then under analogous hypotheses, the principal eigenvalue of P2 is also strictly monotone. These results can be further improved to Vi ∈ L∞loc(Rd), provided we impose a 'stability' assumption on L. See Theorem 3.1 for more details. The second part of this paper deals with the lower bound on the decay of pos- itive supersolutions of general second-order elliptic operators in any dimension. The results obtained here extend those of Agmon [2], Carmona [16], Carmona and Simon [17]. Our proof is based on the stochastic representation of positive solu- tions (see Theorem 4.1). As a consequence, we prove the Landis conjecture for a large class of potentials. Landis' conjecture [28] can be loosely stated as follows: for a bounded potential V , if a solution u of ∆u + V u = 0 satisfies the estimate u(x) ≤ C exp(−cx1+), for some positive constants C and c, then u is identically 0. For a precise statement, we refer the reader to Section 4. This conjecture is open when u and V are real valued, while a counterexample was constructed by Meshkov [30] for complex-valued u and V . This conjecture was revisited recently LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 3 in [18, 26], for dimension 2 and for V ≤ 0. Note that this conjecture trivially holds for V ≤ 0 due to the strong maximum principle. The main contribution in [18, 26] is the lower bound on the decay rate of solutions that may not vanish at infinity. In this direction, Kenig has conjectured in [27, Question 1] a lower bound on the decay of the eigenfunctions of the Schrodinger's equation. In Theorem 4.3 we validate the Landis conjecture for a large class of potentials and in any dimension d ≥ 2. This class of potentials includes compactly supported functions. We also wish to bring to the attention of the reader a recent study of Liouville properties for nonlinear operators [9]. The paper is organized as follows. In Section 2 we briefly review some basic results from the criticality theory of second-order elliptic operators and state our main results. Section 3 is devoted to the proofs of Theorems 2.1 and 2.2 to 2.4. In Section 4 we establish a lower bound on the decay of positive supersolutions (Theorem 4.1), and discuss its implication to Landis conjecture. Notation. The open ball of radius r around a point x ∈ Rd is denoted by Br(x), and Br stands for Br(0). By C0(Rd) (B0(Rd)) we denote the collection of all real valued continuous (Borel measurable) functions on Rd that vanish at infinity. By k·k∞ we denote the L∞ norm. Also κ1, κ2, . . . are used as generic constants whose values might vary from place to place. 2. Preliminaries and main results In this section we introduce our assumptions and state our main results. The conditions (A1) -- (A3) on the coefficients of the operator that follow are used in most of the results of the paper, so we assume that they are in effect throughout unless otherwise mentioned. A notable exception to this is Theorem 2.2, where only (A3) is assumed. (A1) Local Lipschitz continuity: The function a = (cid:2)aij(cid:3) : Rd → Sd×d + , where Sd×d denotes the set of real, symmetric positive definite matrices, is locally Lipschitz in x with a Lipschitz constant CR > 0 depending on R > 0. In other words, we have + ka(x) − a(y)k ≤ CR x − y ∀ x, y ∈ BR , where kak2 := trace(cid:0)aaT(cid:1). The drift function b : Rd → R is a locally bounded Borel measurable function. (A2) Affine growth condition: b and a satisfy a global growth condition of the form ∀ x ∈ Rd, for some constant C0 > 0. (A3) Nondegeneracy: For each R > 0, it holds that hb(x), xi+ + ka(x)k ≤ C0(cid:0)1 + x2(cid:1) dXi,j=1 aij (x)ξiξj ≥ C−1 R ξ2 ∀ x ∈ BR , and for all ξ = (ξ1, . . . , ξd)T ∈ Rd. We define σ(x) = √2a1/2(x). Then under (A1) and (A3), σ is also locally Lip- schitz and has at most linear growth. We say that a is uniformly elliptic if (A3) holds for a positive C = CR which is independent of R. Xt = X0 +Z t 0 b(Xs) ds +Z t 0 σ(Xs) dWs , t ≥ 0, a.s. , 4 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Consider the Ito stochastic differential equation (SDE) given by (2.1) dXs = b(Xs) ds + σ(Xs) dWs , where W is a standard d-dimensional Wiener process defined on some complete, filtered probability space (Ω, F,{Ft}, P). By a strong solution of (2.1) we mean an Ft -- adapted process Xt which satisfies where third term on the right hand side is an Ito stochastic integral. It is well known that given a complete, filtered probability space (Ω, F,{Ft}, P) with a Wiener process W , there exists a unique strong solution of (2.1) [22, Theorem 2.8]. The process X is also strong Markov, and we denote its transition kernel by P t(x,· ). It also follows from the work in [15] that the transition probabilities of X have densities which are locally Holder continuous. The extended generator L is given by Lf (x) = aij(x) ∂ij f (x) + bi(x) ∂if (x) , (2.2) for f ∈ C2(Rd). The operator L is the generator of a strongly-continuous semigroup on Cb(Rd), which is strong Feller. We let Px denote the probability measure, and Ex the expectation operator on the canonical space of the process conditioned on X0 = x. The closure, boundary, and the complement of a set A ⊂ Rd are denoted by ¯A, ∂A, and Ac, respectively. We write A ⋐ B to indicate that ¯A ⊂ B. By τ(D) we denote the first exit time of the process X from a domain D ⊂ Rd, i.e., τ(D) := inf {t : Xt /∈ D} . The process X is said to be recurrent if for any bounded domain D we have Px(τ(Dc) < ∞) = 1 for all x ∈ ¯Dc. Otherwise the process is called transient. A recurrent process is said to be positive recurrent if Ex[τ(Dc)] < ∞ for all x ∈ ¯Dc. It is known that for a non-degenerate diffusion the property of recurrence (or positive recurrence) is independent of domain D and x, i.e., if it holds for some domain D and some x ∈ ¯Dc, then it also holds for every bounded domain D, and all x ∈ ¯Dc [6, Theorem 2.6.12 and Theorem 2.6.10]. By τr (τr) we denote the first hitting (exit) time of the ball Br of radius r around 0, i.e., τr = τ(Bc r) and τr = τ(Br). In order to state the results in this paper, we review some basic definitions from criticality theory which have been introduced by various authors [1, 31, 32, 42], and have been further developed by Y. Pinchover (see [35 -- 37] and references therein). The reader should keep in mind that although the convention in criticality theory is to consider the eigenvalues of the operator −L, we find it more convenient to work with the eigenvalues of L. Definition 2.1. Throughout the paper, D ⊂ Rd denotes a domain, and D := {Dj}∞j=1 a sequence of bounded subdomains with smooth boundaries, such that ¯Dj ⊂ Dj+1, and D = ∪∞j=1Dj. We denote the cone of all positive solutions of the equation Lu = 0 in D by CL(D). We always assume that solutions u are in W2,d loc(D), i.e., u is a strong solution, so that Lu is defined pointwise almost everywhere. Given a potential V ∈ L∞loc(D), we introduce the operator LV := L + V . LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 5 We say that −LV is nonnegative in D (and denote it by −LV ≥ 0 in D), if CLV (D) 6= ∅. The generalized principal eigenvalue of the operator LV is defined by λ∗(L, V ) := inf(cid:8)λ ∈ R : CLV −λ(D) 6= ∅(cid:9) . Note that −LV is nonnegative in D if and only if λ∗(L, V ) ≤ 0. It is clear that −L is always nonnegative, since 1 ∈ CL(D), where 1 is the constant function on D having value 1 at every x ∈ D. In the sequel we shall use the notation λ∗(V ) instead of λ∗(L, V ), whenever this is not ambiguous. In most of the paper we deal with the case D = Rd. An exception to this is Theorem 2.2, where we address the question of Pinchover [38, Problem 5] for general domains D. Let us now recall the definitions of critical and subcritical operators and the ground state. Definition 2.2 (Minimal growth at infinity). A positive function u ∈ W2,d satisfying loc(D) LV u = 0 a.e. in D , is said to be a solution of minimal growth at infinity, if for any compact K ⊂ D and any positive function v ∈ W2,d loc (D \ K) which satisfies LV v ≤ 0 a.e. in D \ K, there exist Di ∈ D, with K ⊂ Di, and a constant κ > 0 such that κu ≤ v in ¯Dc i ∩ D. A positive solution u ∈ CLV (D) which has minimal growth at infinity in D is called the (Agmon) ground state of LV in D. Remark 2.1. Definition 2.2 is equivalent to what is generally used in criticality theory. In criticality theory for an operator P , a function u ∈ CP (D) is said to have a minimal growth at infinity in D, if for any K ⋐ D, with a smooth boundary, and any positive supersolution v ∈ W2,d loc(D \ K) of P v = 0 in D \ K such that v ∈ C((D \ K) ∪ ∂K), and u ≤ v on ∂K, it holds that u ≤ v in D \ K. It is easy to see that this definition implies minimal growth at infinity according to Definition 2.2 for P = LV . To see the converse direction, define κ0 = inf D∩K c v u . Since u ≤ v on ∂K, we must have κ0 ≤ 1 by continuity. We claim that κ0 = 1. Arguing by contradiction, suppose that κ0 < 1. Then v − κ0u must be positive on D\K by the strong maximum principle. Since P (v−κ0u) ≤ 0, then by Definition 2.2 there exist κ ∈ (0, 1− κ0) and Di ∈ D, such that κu ≤ v − κ0u in ¯Dc i ∩ D. Without loss of generality, suppose that Di ⊃ K. Applying the strong maximum principle in Di ∩ K c to LΦ − V −Φ ≤ 0 , Φ = v − (κ0 + κ)u , we have κu ≤ v − κ0u in ¯Di ∩ K c, and therefore (κ0 + κ)u ≤ v in D ∩ K c. But this contradicts the definition of κ0. Hence κ0 = 1. Definition 2.3. The operator LV is said to be critical in D, if LV admits a ground state in D. The operator LV is called subcritical in D, if −LV ≥ 0 in D, but LV does not admit a ground state solution. Example 2.1. Let L = ∆ in Rd, d ≥ 1. It is well known that λ∗(L, 0) = 0. Moreover, L is critical if and only if d ≤ 2. LV := ∆ + (d − 2)2 4 1 x2 . 6 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Example 2.2. Let D = Rd \ {0}, d ≥ 3, and consider the Hardy operator 2−d 2 . Then it is well known that LV is critical, and the corresponding ground state is x Remark 2.2. For D = Rd one can also define the (generalized) principal eigenvalue in the sense of Berestycki and Rossi [14] (see also [33]) by λ∗(L, V ) := inf(cid:8)λ ∈ R : ∃ ϕ ∈ W2,d loc (Rd), ϕ > 0, LV ϕ − λϕ ≤ 0 a.e. in Rd(cid:9) . For V ∈ L∞loc(Rd), it is known from [14, Theorem 1.4] that there exists a (gen- eralized) positive eigenfunction corresponding to λ∗(L, V ), whenever this is finite. Thus λ∗(L, V ) = λ∗(L, V ). Remark 2.3. Let P = LV in D. It is well known that the operator P is critical in D, if and only if the equation P u = 0 in D has a unique (up to a multiplicative constant) positive supersolution (see [35,36]). In particular, P is critical in D if and only if P does not admit a positive Green's function in D. However, there exists a sign-changing Green's function for a P which is critical in D (see [20]). In addition, in the critical case, we have dim CP (D) = 1, and the unique positive solution (up to a multiplicative positive constant) is a ground state of P in D. On the other hand, P is subcritical in D if and only if P admits a unique positive minimal Green's function GDP (x, y) in D. Moreover, for any fixed y ∈ D, the function GDP (·, y) is a positive solution of minimal growth in a neighborhood of infinity in D, i.e., in D \ K for some compact set K (see [19]). For an eigenpair (Ψ, λ) of LV in Rd, i.e., a solution of LV Ψ = λΨ, Ψ > 0 in Rd , the twisted process corresponding to (Ψ, λ) is defined by the SDE dYs = b(Ys) ds + 2a(Ys)∇ψ(Ys) ds + σ(Ys) dWs , (2.3) with ψ = log Ψ. The process Y also goes by the name of Doob's h-transformation in the literature. Since ψ ∈ W2,p loc (Rd), p > d, it follows that ψ is locally bounded (in fact, it is locally Holder continuous), and therefore (2.3) has a unique strong solution up to its explosion time. In what follows, we use the notation (Ψ∗, λ∗(V )) to denote a principal eigenpair. Let us introduce one more definition which is related to the criticality of an operator. By C+ 0 (Rd) we denote the collection of all nonnegative, non-zero, real valued continuous functions on Rd that vanish at infinity. We fix L, and dropping the dependence on L in the notation, as mentioned earlier, we let λ∗(V ) denote the principal eigenvalue of L + V . Definition 2.4. λ∗(V ) is said to be strictly monotone at V if for all h ∈ C+ 0 (Rd) we have λ∗(V − h) < λ∗(V ) < λ∗(V + h). Also, λ∗(V ) is said to be strictly monotone at V on the right if for all h ∈ C+ 0 (Rd) we have λ∗(V ) < λ∗(V + h). It is known from [5, Theorem 2.2] that if λ∗(V −h) < λ∗(V ) for some h ∈ C+ then λ∗(V − h) < λ∗(V ) < λ∗(V + h) for all h ∈ C+ using the fact that V 7→ λ∗(V ) is a convex function (see for instance, [14]). 0 (Rd), 0 (Rd). This assertion also follows LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 7 Example 2.3. For L = ∆ in R2 and V = 0, it is known that λ∗(V ) strictly monotone at V on the right, but not strictly monotone at V . Throughout the paper, with the exception of Theorem 2.2, we consider potential functions V that are Borel measurable and bounded from below. We also assume that λ∗(V ) is finite. Let us begin with the following equivalence between the strict right monotonicity of the principal eigenvalue and the criticality of the operator [5]. See also [40, Theorems 4.3.3 and 7.3.6] for similar results concerning operators with regular coefficients. Theorem 2.1. Let D = Rd. The following are equivalent. (a) A function Ψ ∈ W2,d (b) The twisted process corresponding to the eigenpair (Ψ, λ) is recurrent. (c) λ∗(V ) is strictly monotone at V on the right. (d) For any r > 0, the eigenpair (Ψ, λ) satisfies loc(D) is a ground state for LV − λ, with λ ∈ R. (2.4) Ψ(x) = ExheR τr 0 (V (Xs)−λ) ds Ψ(X τr )1{ τr<∞}i , x ∈ Bc r , where, as defined earlier, τr denotes the first hitting time to the ball Br. We often exploit the above equivalence between strict monotonicity and critical- ity. To state our next result we need some additional notation. Let Lkf = aij k (x) ∂ij f (x) + bi k(x) ∂if (x) , k = 1, 2 . We assume that (ak, bk), k = 1, 2, satisfies (A1) -- (A3). We say that L1 is a small perturbation [35] of L2 if ka1(x)− a2(x)k +b1(x)− b2(x) = 0 outside some compact set. The first main result of this section is the following theorem which gives a partial answer (see also Theorem 2.3) to the open question posed by Y. Pinchover in [38, Problem 5]. Simplifying the notation, in the sequel we sometimes denote by λ∗k (instead of λ∗(Lk, Vk)) the principal eigenvalue of the operator Lk + Vk, k = 1, 2. Theorem 2.2. Let D be a domain in Rd, d ≥ 1. Consider two Schrodinger oper- ators defined on D of the form Pk := Lk + Vk , k = 1, 2, where ak, k = 1, 2, are continuous and satisfy (A3), bk, Vk ∈ L∞loc(D), and V2 ≥ V1 outside a compact set in D. In addition, assume that L1 is a small perturbation of L2 in D, and (1) The operator P1 − λ∗1 is critical in D. Denote by Ψ∗1 its ground state. (2) λ∗2 ≤ λ∗1 and there exists Ψ ∈ W2,d loc(D), with Ψ+ 6= 0, satisfying L2Ψ + V2Ψ ≥ λ∗1Ψ , (2.5) and (2.6) Ψ+(x) ≤ C Ψ∗1(x) for all x ∈ D , for some constant C > 0. Then the operator P2 − λ∗2 is critical in D, λ∗1 = λ∗2, and Ψ is its ground state. Remark 2.4. One can not expect any pair V1, V2 to satisfy the hypotheses of Theorem 2.2, even if we restrict V1 and V2 to have compact support, and consider the same second-order operator L = L1 = L2. To see this, let us take V2 (cid:8) V1, both of them compactly supported, and suppose that L + V1 − λ∗1 is critical. Then 8 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY it is not possible to have λ∗1 = λ∗2 and Ψ∗2 ≤ CΨ∗1, for some constant C. Indeed, if L denotes the generator of the twisted process corresponding to the eigenpair (Ψ∗1, λ∗1), and λ∗1 = λ∗2, then for Φ = Ψ∗ we have 2 Ψ∗ 1 LΦ = (V1 − V2)Φ ≥ 0 . Since the twisted process Ys corresponding to (Ψ∗1, λ∗1) is recurrent by Theorem 2.1, and Φ(Ys) is a bounded submartingale, Φ must be constant. This implies (V1 − V2)Ψ∗1 = 0, which is a contradiction. More precisely, one can find a relation between V1 and V2 as follows. Suppose D = Rd and the operators L + Vi, Vi ∈ L∞loc(Rd), are critical in Rd with principal eigenfunctions Ψ∗i , i = 1, 2. Then by Theorem 2.1 we know that for any r > 0 we have (2.7) Ψ∗i (x) = ExheR τr 0 Vi(Xs) ds Ψ∗i (X τr )1{ τr <∞}i , x ∈ Bc r . Now if (2.6) holds, i.e., Ψ∗2 ≤ CΨ∗1 in Rd, then by (2.7), for every r > 0 we can find a constant Cr such that ExheR τr 0 V2(Xs) ds 1{ τr<∞}i ≤ Cr ExheR τr 0 V1(Xs) ds 1{ τr <∞}i , x ∈ Bc r . This in particular, provides a necessary condition on the potentials for Liouville type theorems like Theorem 2.2 to hold. For the rest of the results in this section we let D = Rd. Theorem 2.3. Consider two Schrodinger operators defined on Rd of the form Pk := Lk + Vk , k = 1, 2, whose coefficients satisfy (A1) -- (A3), and Vk ∈ L∞loc(Rd). Let V (x) := max{V1(x), V2(x)} . Suppose that there exists a positive Φ ∈ W2,d (2.8) L1 = L2, L2 Φ + V Φ ≤ λ∗1 Φ in K c . loc (Rd) and a compact set K such that In addition, assume that (1) The operator P1 − λ∗1 is critical in Rd. Denote by Ψ∗1 its ground state. (2) λ∗2 ≤ λ∗1, and there exists subsolution Ψ ∈ W2,d loc (Rd), which may be sign- changing but Ψ+ 6= 0, that satisfies (2.9) L2Ψ + V2Ψ ≥ λ∗1Ψ , and for some constant C > 0, Ψ+(x) ≤ C Ψ∗1(x) for all x ∈ Rd . Then the operator P2 − λ∗2 is critical in Rd, λ∗1 = λ∗2, and Ψ is its ground state. It should be noted that the second display in (2.8) is an assumption on the operators; compare to [38, Theorem 1.7]. However, there is a large family of elliptic operators for which (2.8) holds, as the following examples show. LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 9 Example 2.4. Note that (2.8) is satisfied if V1−V2 has a fixed sign outside a compact set K. If V = V1 in K c we can choose Φ = Ψ∗1. On the other hand, if V = V2 in K c, we know from [14, Theorem 1.4] that there exists a positive Φ1 satisfying L2Φ1 + V2Φ1 = λ∗1Φ1 in Rd . Hence we can take Φ = Φ1 in K c. Example 2.5. Let us now give an example where the sign of V1 − V2 may not be fixed outside some compact set. Consider P1 = ∆ + V1 in Rd, d ≥ 3, such that V1 has compact support and P1 is critical in Rd with λ∗ = 0. Now let L2 := ∆ + bi∂i , where the vector field b has compact support in Rd. Let a nonnegative W be a small perturbation (see [36, Definition 2.1, Example 2.2]) with respect to the operator −∆. Then there exists a positive ε such that −∆Ψ− ε W Ψ = 0 has a positive solution Ψ in Bc r, r > 0 [36, Lemma 2.4]. Therefore, if we choose a potential V2 which decays faster than W at infinity, i.e., for every δ > 0 there exists a compact Kδ such that V2(x) ≤ δ W (x) for x ∈ K c L2Ψ + V Ψ ≤ ∆Ψ + V2Ψ ≤ 0 Therefore (2.8) holds. There are several choices for a small perturbation W (see [36, Example 2.2]). For instance, we could take any nonnegative W which is locally Holder continuous and satisfies δ , it is easy to see that, by choosing δ < ε, we have on the complement of a compact set in Rd . ϕ(r) dr < ∞ , r0 > 0 . r0 1 r Z ∞ in Rd , d ≥ 3 , (1 + x)2 W (x) ≤ ϕ(x) ∀ x ∈ Rd , and Example 2.6. We define for k = 1, 2, Pk := ∆ + bi k∂i + Vk with the vector fields bk smooth, and satisfying C bk(x) ≤ (1 + x)1+ε , for some constants C > 0 and ε > 0. It is known that the operator Lk := ∆ + bi k∂i is subcritical; this follows from the fact that 1 is a positive solution, together with the above decay estimate on bk. Hence there exists a minimal Green's function for Lk. Also, the Green functions G−∆ and G−Lk are comparable [3, Theorem 1]. Therefore, a small perturbation of ∆ is also a small perturbation of Lk [36]. Let b1(x) − b2(x) = 0 outside a compact set. As earlier, we suppose that P1 is critical and λ∗1 = 0. Assume that V = max{V1, V2} decays faster than (1 + x)−2−ǫ at infinity. In particular, we may choose V2(x) = (1 + x)−2−ε. Then V satisfies the estimate above, and hence, as before, there exists a positive supersolution Ψ to L2Ψ + V Ψ ≤ 0 on the complement of some compact set in Rd. Thus (2.8) holds. Remark 2.5. Recall that the criticality of LV − λ∗ is equivalent to the strict mono- tonicity of λ∗ at V on the right by Theorem 2.1. However, strict right monotonicity does not necessarily imply strict monotonicity of λ∗. Later, in Theorem 2.4, we show that if λ∗ is strictly monotone at V , then we do not require (2.8). Also ob- serve that if V ∈ B0(Rd) and λ∗ is not strictly monotone at V , then λ∗(V ) ≤ 0. Indeed, since λ∗ is not strictly monotone at V and λ∗(V ) ≥ λ∗(−V −) it is obvious 10 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY In addition, the following hold. that λ∗(V ) ≤ 0. If X is not positive recurrent and λ∗(0) = 0, then λ∗(V ) = 0, otherwise λ∗(−V −) ≤ λ∗(V ) < 0 = λ∗(0). How- ever, this implies that X is geometrically ergodic [5, Theorem 2.7], and therefore, positive recurrent. If a is bounded and uniformly elliptic, and x−1hb(x), xi → 0 as x → ∞, then λ∗(V ) = 0 for any Lipschitz V ∈ C0(Rd), since by [25, Proposi- tion 6.2] λ∗(V ) ≥ 0. Therefore assuming that λ∗1 = 0 in Examples 2.5 and 2.6 is not very restrictive. In [12], Berestycki, Caffarelli and Nirenberg asked the following question. Is it true that if there exists a bounded, sign-changing solution Ψ to ∆Ψ+V Ψ = 0 in Rd, for some locally bounded potential V, then necessarily λ∗(V ) > 0? This question has been resolved in [10, 12, 21], and the answer is "yes" if and only if d = 1, 2. Applying Theorems 2.1 and 2.3 we can extend the sufficiency part of this answer to a more general class of elliptic operators. Noting that the Brownian motion is recurrent for d = 1, 2, and transient for higher dimensions, we focus on elliptic operators L satisfying (A1) -- (A3) which are generators of a recurrent process. Using Theorems 2.1 and 2.3 we obtain the following two corollaries. Corollary 2.1. Suppose the solution of (2.1) is recurrent, and V is a locally bounded function which does not change sign outside some compact set in Rd. Then the existence of a bounded, sign-changing solution Ψ ∈ W2,d loc (Rd) to LΨ + V Ψ = 0 implies that λ∗(V ) > 0. Proof. Since (1, 0) is an eigenpair of L and the corresponding twisted process is given by X, it follows by Theorem 2.1 that L is a critical operator with principal eigenvalue 0. Moreover, CL = {c1 : c ∈ (0,∞)}. We apply Theorem 2.3, with L1 = L2 = L, V1 = 0, λ∗1 = 0, V2 = V , and Ψ∗1 = c1. Suppose λ∗(V ) ≤ 0. If V is positive outside a compact set, then (2.8) holds with Φ = Ψ∗2, the principal eigenfunction corresponding to λ∗(V ). On the other hand, if V is negative outside a compact set, then (2.8) holds for some positive Φ by [14, Theorem 1.4] (see also Example 2.4). It then follows from Theorem 2.3 that Ψ is a ground state, and therefore cannot be sign-changing. This contradicts the hypothesis that λ∗(V ) ≤ 0, and completes the proof. (cid:3) Corollary 2.2. Let the process (2.1) be recurrent and V ≤ 0 be a bounded function. Then there does not exist any nonconstant bounded solution u to Lu + V u = 0 . (2.10) Proof. Since λ∗(V ) ≤ 0, Corollary 2.1 implies that any bounded solution u to (2.10) cannot be sign-changing. So without loss of generality we assume that 0 ≤ u < C. Then C − u is a positive supersolution of Lu = 0. Since L is critical by hypothesis, it has a unique supersolution (up to a multiplicative constant). Hence u must be constant. (cid:3) The conclusion of Corollary 2.2 might not hold if V 6≤ 0. For instance, in dimension d = 2 we know that the standard Wiener process is recurrent. But u(x, y) = sin(x) sin(y) satisfies ∆u + 2u = 0. Corollary 2.2 is also comparable to [38, Theorem 1.7]. Note that for V = 0 the operator in (2.10) is critical in the sense of Pinchover (see Theorem 2.1 above). Therefore Corollary 2.2 provides a Liouville property for the perturbed operator. LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 11 As shown in Theorem 2.1, criticality is equivalent to the strict right monotonic- ity of the principal eigenvalue λ∗. However, if we assume strict monotonicity of λ∗(L1, V1) at V1, then Theorem 2.3 holds for a bigger class of potentials without assuming (2.8). This is the subject of our next result. Also note that the the theorem which follows provides sufficient conditions for strict monotonicity of the principal eigenvalue of the perturbed problem. Theorem 2.4. Let L1 be a small perturbation of L2, Vi ∈ L∞loc(Rd), i = 1, 2, and V1 − V2 ∈ B0(Rd). Let λ∗i denote the principal eigenvalue of Li + Vi, i = 1, 2, and suppose that λ∗1 is strictly monotone at V1. Suppose also that λ∗2 ≤ λ∗1, and that there exists Ψ ∈ W2,d loc (Rd), which may be sign-changing but Ψ+ 6= 0, that satisfies (2.11) L2Ψ + V2Ψ ≥ λ∗1Ψ , and (2.12) for some constant C > 0. Then λ∗2 is strictly monotone at V2, and Ψ = Ψ∗2 (up to a multiplicative constant ), where Ψ∗2 is the principal eigenfunction of L2 + V2. Ψ+(x) ≤ C Ψ∗1(x) for all x ∈ Rd , Remark 2.6. Strict monotonicity sometimes implies an interesting spectral prop- erty. To explain this, we restrict ourselves to symmetric operators. In particular, we consider a second-order elliptic operator in D in divergence form given by Lu = div(cid:0)A∇u(cid:1) , where A : Rd → Sd×d + is locally non-degenerate. The assumptions on the coefficients are the same as before. Let dν = ρ(x)dx, where ρ(x) is a positive measurable function on D. The operator L is self-adjoint in the space L2(D, dν) (in the sense of the Friedrichs extension). Let V ∈ L∞(D), and σ(LV ) denote the L2(D, dν)-spectrum of the Friedrichs extension of LV , which is also denoted as LV , abusing the notation in the interest of simplicity. We next show that if λ∗(V ) is strictly monotone at V , then it must be an isolated eigenvalue in σ(LV ). Indeed, by Persson's formula (see [34] or [19, Proposition 4.2]) the supremum of the essential spectrum σess(LV ) is given by λ∞(V ) := inf {λ : ∃ K ⋐ D , CLV −λ(D \ K) 6= ∅} . In addition, λ∗(V ) is the supremum of σ(LV ). It is clear that λ∞(V ) ≤ λ∗(V ). We claim that λ∞(V ) < λ∗(V ). Arguing by contradiction, let us assume λ∞(V ) = λ∗(V ). It is known that σess(LV ) = σess(LV − h) (2.13) for any h ∈ C0(D). Using (2.13), we have λ∞(V ) = λ∞(V − h) for all h ∈ C+ By hypothesis, λ∗(V ) is strictly monotone at V and therefore, we have λ∗(V − h) < λ∗(V ) = λ∞(V ) = λ∞(V − h) ≤ λ∗(V − h) . 0 (D). Thus we arrive at a contradiction, which implies that λ∞(V ) < λ∗(V ) (for a more general related result see [5, Theorem 2.5]). Since the Friedrichs extension is a self-adjoint operator, its spectrum can be written as σ(LV ) = σess(LV ) ∪ σdis(LV ), with σess(LV ) ∩ σdis(LV ) = ∅, where σdis(LV ) is the discrete spectrum. On the other hand, [11, Theorem 1.1] shows that λ∗(V ) ∈ σ(LV ). Therefore, λ∗(V ) is an isolated eigenvalue in σ(LV ). 12 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Remark 2.7. Let Pi = L + Vi, i = 1, 2, be self-adjoint operators, and λ∞,i denote the supremum of the essential spectrum of Pi. If V1 − V2 ∈ C0(Rd), then it is known that σess(P1) = σess(P2), which in turn implies that λ∞,1 = λ∞,2. Suppose that the hypotheses of Theorem 2.4 hold. Then using Theorem 2.4 and Remark 2.6 we deduce that λ∗2 > λ∗ ∞,2, and that the corresponding operator P2 − λ∗2 is critical. In particular, Theorem 2.4 provides a necessary condition for the spectral gap of the operator P2. Example 2.7. Let D = Rd \ {0}, where d ≥ 3, and consider the Hardy operator LV := ∆ + (d − 2)2 4 1 x2 . Then it is well known (see [19]) that for this operator we have λ∗(V ) = λ∞(V ). Hence λ∗(V ) cannot be strictly monotone at V , although it is strictly right mono- tone. There is a large class of operators for which the strict monotonicity property holds. The following example suggests that the assumptions in Theorems 2.2 and 2.4 hold for a large class of operators. Example 2.8. Suppose that the solution of (2.1) is recurrent. Consider two functions Vi ∈ C+ 0 (Rd), i = 1, 2, which are compactly supported. Then as shown in [5, β := λ∗(β Vi) is strictly monotone in [0,∞), and Theorem 2.7], the map β 7→ Λi 0 = 0, for i = 1, 2. Since β 7→ Λi Λi β is an increasing, convex function [14], we have limβ→∞ Λi β = ∞. Therefore, for any β1 > 0, we can find β2 > 0 such that . Thus by defining Vi := βi Vi, i = 1, 2, we note that λ∗1 = λ∗(V1) = Λ1 β1 λ∗(V2) = λ∗2, and Vi has compact support. On the other hand, L + Vi− λ∗i is critical by Theorem 2.1. In fact, the corresponding twisted processes are geometrically ergodic by [5, Theorem 2.7]. Thus, if Ψ∗i , i = 1, 2, are the principal eigenfunctions, then they have a stochastic representation by Theorem 2.1. Hence, if we choose r large enough such that support(Vi) ⊂ Br, we have = Λ2 β2 Ψ∗i (x) = Exhe−λ∗ i τr Ψ∗i (X τr )1{ τr <∞}i, x ∈ Bc r . Since λ∗1 = λ∗2, it is easy to see from the above that Ψ∗2 ≤ C Ψ∗1 for some C > 0. 3. Proofs of Theorems 2.1 to 2.4 loc (Rd), we have Before we proceed with the proofs of the results in Section 2, let us recall the Ito -- Krylov formula [29, p. 122] for generalized derivatives. Let D be a bounded domain in Rd with smooth boundary and V ∈ L∞loc(Rd). Let τ = τ(D). Then for any ϕ ∈ W2,d (3.1) ExheR T ∧τ V (Xs) ds ϕ(XT∧τ)i − ϕ(x) = Ex(cid:20)Z T∧τ 0 V (Xs) ds LV ϕ(Xt) dt(cid:21) for all x ∈ D and T > 0. We start with the proof of Theorem 2.1. Proof of Theorem 2.1. The equivalence between (b), (c) and (d) is established in [5]. Since the twisted process corresponding to an eigenpair (Ψ, λ) with λ > λ∗(V ) is transient by [5, Theorem 2.1 (c)], part (b) together with [5, Corollary 2.1] imply that λ = λ∗(V ). eR t 0 0 LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 13 Let us show that (b) ⇒ (a). Suppose that v ∈ W2,d loc (Rd) is a positive function which satisfies Lv + (V − λ)v ≤ 0 a.e. in Bc r1, with r1 > 0. Recall that τR denotes the first exit time from the ball BR. Then by the Ito -- Krylov formula in (3.1) we have 0 v(x) ≥ ExheR τr ∧τR ∧T (V (Xs)−λ) ds v(X τr∧τR∧T )i 0 (V (Xs)−λ) ds v(X τr ) 1{ τr <τR∧T}i , 0 (V (Xs)−λ) ds v(X τr )1{ τr <∞}i, ≥ ExheR τr v(x) ≥ ExheR τr Hence (a) follows by applying (2.4). x ∈ Bc r ∩ BR , r > r1 . x ∈ Bc r , r > r1 . Now letting first T → ∞, and then R → ∞, and using Fatou's lemma we obtain Next we show that (a) ⇒ (b). By Corollary 3.2, which appears later in this section, there exists a ball B, a constant δ ≥ 0, and a positive solution Ψ∗ ∈ W2,d loc(Rd) to LΨ∗ + (V + δ1B − λ)Ψ∗ = 0, such that λ = λ∗(V + δ1B), and (3.2) Ψ∗(x) = ExheR τr 0 (V (Xs)−λ) ds Ψ∗(X τr )1{ τr <∞}i Ψ∗ for Br ⊃ B. Let κ0 := inf Rd Ψ . By the assumption of minimal growth, we have κ0 > 0. Note that the value κ0 is attained. If not, then the function Φ = Ψ∗ − κ0Ψ is positive and satisfies LΦ + (V − λ∗)Φ ≤ 0. Consequently, the minimal growth of Ψ would imply that inf Rd(cid:16) Ψ∗ Ψ − κ0(cid:17) > 0, thus contradicting the definition of κ0. Therefore, defining Φ = Ψ∗−κ0Ψ, it is easy to see that LΦ−(V +δ1B−λ∗)−Φ ≤ 0, and Φ attains its minimum value 0 in Rd, which implies by the strong maximum principle that κ0Ψ∗ = Ψ. This of course implies that δ = 0. Hence λ = λ∗(V ), and, in turn, (3.2) implies that the twisted process corresponding to the eigenpair (Ψ, λ) is recurrent. This completes the proof. (cid:3) We continue with the proof of Theorem 2.2. Proof of Theorem 2.2. Let K ⋐ D be a compact set such that V2−V1 ≥ 0 and L1 = L2 in K c. Since λ∗2 ≤ λ∗1, using Harnack's inequality, it follows that there exists a positive generalized eigenfunction Ψ∗2 corresponding to the generalized eigenvalue λ∗2, i.e., Thus we have P2Ψ∗2 = λ∗2 Ψ∗2 in D . L1Ψ∗2 + (V1 − λ∗1)Ψ∗2 ≤ L2Ψ∗2 + (V2 − λ∗2)Ψ∗2 = 0 (3.3) By the minimal growth property of Ψ∗1 and (3.3), we can find a positive constant κ and a set Di ∈ D, with Di ⊃ K, such that κΨ∗1 ≤ Ψ∗2 for all x ∈ D \ Di. Let in K c . κ = sup D Ψ Ψ∗2 = sup D Ψ+ Ψ∗2 . Then, using (2.6) and the bound κΨ∗1 ≤ Ψ∗2, we conclude that κ ∈ (0,∞). Let us now define Φ(x) := κΨ∗2(x) − Ψ(x) in D. We claim that there exists x0 ∈ D such that Φ(x0) = 0. If not, then Φ(x) > 0 in D. Then in K c we have L1Φ + (V1 − λ∗1)Φ = L2Φ + (V1 − λ∗1)Φ 14 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY ≤ L2Φ + (V2 − λ∗1)Φ = (cid:0)L2 + V2 − λ∗1(cid:1)(κΨ⋆ ≤ (cid:0)L2 + V2 − λ∗1(cid:1)κΨ∗2 = (λ∗2 − λ∗1)κ Ψ∗2 ≤ 0 . 2 − Ψ) By the minimality of the growth of the ground state Ψ∗1, there exist a positive constant κ1 and a compact set K2 ⊃ K such that κ1Ψ∗1 ≤ Φ in K c 2. Next, using (2.6), we obtain κΨ∗2(x) − Ψ(x) ≥ κ1 C Ψ(x) ⇒ Ψ(x) Ψ∗2(x) ≤ κ 1 + κ1/C < κ ∀ x ∈ K c 2 . Thus the value κ is attained for some x0 ∈ K2. This shows that Φ(x0) = 0 at some x0 ∈ D. On the other hand, Φ is nonnegative, and it satisfies L2Φ + (V2 − λ∗1)Φ ≤ (λ∗2 − λ∗1)κΨ∗2 ≤ 0 in D , which in turn implies that L2Φ − (V2 − λ∗1)−Φ ≤ −(V2 − λ∗1)+Φ ≤ 0 in D . Thus by strong maximum principle we must have Φ ≡ 0 in D. This shows that κΨ∗2 = Ψ, which implies by (2.5) that λ∗2 = λ∗1. To complete the proof it remains to show that Ψ∗2 is a ground state of L2+V2−λ∗2. loc ( K c) be a positive supersolution of Consider a compact set K, and let v ∈ W2,d L2 + V2 − λ∗2, i.e., By hypothesis, we have L2v + (V2 − λ∗2)v ≤ 0 in K c . L1v + (V1 − λ∗1)v ≤ L2v + (V2 − λ∗2)v ≤ 0 on K c ∩ K c . Since Ψ∗1 has minimal growth at infinity, we can find a constant κ2 and a compact set K2 satisfying κ2Ψ∗1 ≤ v in K c C Ψ∗2 ≤ v in K c 2. Therefore Ψ∗2 also has minimal growth at infinity, and hence is a ground state. This completes the proof. (cid:3) 2. Combining this with (2.6) we have κ2 As an immediate corollary to Theorem 2.2, we have the following generalization of the result in [38, Corollary 1.8] Corollary 3.1. Let P1 and P2 be as in Theorem 2.2. Suppose that any Ψ which satisfies (2.5) and (2.6) cannot be a solution of (L2 + V2 − λ∗1)Ψ = 0 unless it is sign-changing. Then λ∗2 > λ∗1. To prove Theorems 2.3 and 2.4 we need several lemmas which are stated next. Lemma 3.1. Suppose that λ∗(V ) is not strictly right monotone at V . Then for any ball B there exists a constant δ > 0 such that λ∗(V ) = λ∗(V + δ1B), and λ∗ is strictly right monotone at V + δ1B. Proof. Let Fα(x) := V (x) − λ∗(V ) − α, for α > 0. It is evident that the Dirichlet eigenvalue of −L − Fα on every ball Bn is positive. Thus by Proposition 6.2 and Theorem 6.1 in [13], for any n ∈ N, the Dirichlet problem (3.4) Lϕα,n(x) + Fα(x) ϕα,n(x) = −1B(x) a.e. x ∈ Bn , ϕα,n = 0 on ∂Bn , LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 15 has a unique solution ϕα,n ∈ W2,p loc (Bn) ∩ C( ¯Bn), for any p ≥ 1. In addition, by the refined maximum principle in [13, Theorem 1.1] ϕα,n is nonnegative. It is clear that ϕα,n cannot be identically equal to 0. Thus if we write (3.4) as Lϕα,n − F −α ϕα,n = −F + α ϕα,n − 1B , it follows by the strong maximum principle that ϕα,n > 0 in Bn. By the Ito -- Krylov formula in (3.1), since ϕα,n = 0 on ∂Bn, we obtain from (3.4) that (3.5) ϕα,n(x) = ExheR T 0 Fα(Xs) ds ϕα,n(XT ) 1{T≤τn}i + Ex(cid:20)Z T∧τn 0 eR t 0 Fα(Xs) ds 1B(Xt) dt(cid:21) for all (T, x) ∈ R+ × Bn. Now fix α > 0. Let Ψ be a positive principal eigenfunction of L + V constructed canonically from Dirichlet eigensolutions. We can scale Ψ so that Ψ ≥ 1 on B. Let Ψα = α−1Ψ. Then LΨα(x) + Fα(x) Ψα(x) ≤ −1B(x) Using the Ito -- Krylov formula and Fatou's lemma, we obtain Ψα(x) ≥ ExheR t 0 Fα(Xs) ds Ψα(Xt)i + Ex(cid:20)Z t 0 a.e. x ∈ Rd . 0 Fα(Xs) ds 1B(Xt) dt(cid:21) , eR t for any finite stopping time t, and any α > 0. Also, Ψ being an eigenfunction, we have (3.6) (3.7) 0 Ψα(x) ≥ ExheR T ∧τn ≥ α−1(cid:16)inf F0(Xt) dt Ψα(XT ) 1{T≤τn}i Ψ(cid:17) ExheR T ∧τn F0(Xt) dt 1{T≤τn}i . Bn 0 Thus by (3.7) we have ExheR T 0 Fα(Xs) ds ϕα,n(XT ) 1{T≤τn}i ≤ e−αT(cid:18)sup Bn ϕα,n(cid:19) ExheR T 0 F0(Xs) ds 1{T≤τn}i , and the right hand side tends to 0 as T → ∞. Taking limits in (3.5) as T → ∞, using monotone convergence for the second integral, we obtain ϕα,n(x) = Ex(cid:20)Z τn 0 eR t 0 Fα(Xs) ds 1B(Xt) dt(cid:21) , which implies by (3.6) that ϕα,n ≤ Ψα for all n ∈ N. It therefore follows by the a priori estimates that {ϕα,n} is relatively weakly compact in W2,p(Bn), for any p ≥ 1 and n ∈ N, and thus ϕα,n converges uniformly on compact sets along some sequence n → ∞ to a nonnegative Φα ∈ W2,p loc(Rd), for any p ≥ 1, which solves LΦα(x) + Fα(x) Φα(x) = −1B(x) a.e. x ∈ Rd . It is clear by the strong maximum principle that Φα > 0. Since, as we have already shown, ϕα,n ≤ Ψα for all n ∈ N, it follows that Φα ≤ Ψα. Using (3.6) with t = T 16 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY and a slightly smaller α, then by (3.5) and dominated convergence, we obtain (3.8) Φα(x) = ExheR T 0 Fα(Xs) ds Φα(XT )i + Ex(cid:20)Z T 0 eR t 0 Fα(Xs) ds 1B(Xt) dt(cid:21) for all T > 0 and x ∈ Rd. Since (3.5) also holds with T replaced by T ∧ τr, then again dominating this by (3.6) with t = τr and choosing a slightly smaller α, we similarly obtain (3.9) Φα(x) = ExheR τr 0 Fα(Xs) dt Φα(X τr )1{ τr <∞}i + Ex(cid:20)Z τr 0 F0(Xs) ds Ψα(XT )i 0 Fα(Xs) ds Φα(XT )i ≤ e−αTheR T r and r > 0. Using the bound Φα ≤ Ψα we have ExheR T eR t 0 0 Fα(Xs) ds 1B(Xt) dt(cid:21) for all x ∈ Bc Thus by (3.8), we obtain 0 . ≤ e−αT Ψα(x) −−−−→T→∞ 0 Fα(Xs) ds 1B(Xt) dt(cid:21) . eR t Φα(x) = Ex(cid:20)Z ∞ 0 Since λ∗ is not strictly right monotone at V , the twisted process is transient, and by [5, Lemma 2.7] we have E0(cid:20)Z ∞ 0 eR t 0 F0(Xs) ds 1B(Xt) dt(cid:21) < ∞ . It follows that Φα(0) is bounded uniformly over α ∈ (0, 1), and therefore is uniformly locally bounded by the superharmonic Harnack inequality [7]. Thus letting α ց 0, we obtain a positive Φ as a limit of Φα, which solves LΦ(x) + F0(x) Φ(x) = −1B(x) a.e. x ∈ Rd . Write this as a.e. x ∈ Rd . On the other hand, taking limits in (3.9) as α ց 0, choosing r > 0 such that Br ⊃ B, we obtain LΦ(x) +(cid:0)V (x) + Φ−1(x)1B(x)(cid:1) Φ(x) = λ∗(V )Φ(x) 0 F0(Xs) dt Φ(X τr )1{ τr <∞}i . Φ(x) = ExheR τr This shows that Φ has a stochastic representation, which implies that λ∗ is strictly monotone at V + Φ−11B on the right. Then the monotonicity property of δ 7→ λ∗(V + δ1B) implies that limδ→∞ λ∗(V + δ1B) > λ∗(V ). So we define δ0 = inf{δ > 0 : λ∗(V + δ1B) > λ∗(V )}, then λ∗ is strictly monotone at V + δ01B on the right by [5, Corollary 2.4]. (cid:3) Corollary 3.2. For any λ > λ∗(V ) and ball B, there exists a constant δ such that λ = λ∗(V + δ1B) and λ∗ is strictly right monotone at V + δ1B. LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 17 Proof. By Lemma 3.1 there exists δ∗ ≥ 0 such that λ∗(V + δ∗1B) = λ∗(V ), and λ∗ is strictly right monotone at V + δ∗1B. Recall that the map δ 7→ λ∗(V + δ1B) is non-decreasing and convex. Since λ∗(V + δ1B) > λ∗(V ) for any δ > δ∗ by strict right monotonicity, it follows that this map is strictly increasing and convex on [δ∗,∞), and hence its image is [λ∗(V ),∞). (cid:3) Using Lemma 3.1 we can show the following. Lemma 3.2. Let D = Rd. Assume the hypotheses of Theorem 2.2 and (A1) -- (A2), and in addition, suppose that V2 − V1 ∈ B0(Rd), and L1 = L2 outside a compact set K. Then λ∗1 = λ∗2. Proof. Suppose that λ∗2 < λ∗1. By Lemma 3.1 there exists δ ≥ 0, such that λ∗2(V2 + δ1B) is strictly monotone at V2 + δ1B on the right, and λ∗2(V2 + δ1B) = λ∗2. Let Φδ denote the ground state corresponding to λ∗2(V2 + δ1B). Then L1Φδ + (V1 − λ∗1)Φδ = (λ∗2 − V2 − δ1B + V1 − λ∗1)Φδ outside the compact set K. Hence by the minimal growth property of Ψ∗1 we have Ψ∗1 ≤ κ1Φδ. Note that the choice of B is arbitrary. This means we can select B so that Ψ > 0 on B. Therefore Moreover, Ψ Φδ ≤ Ψ+ Φδ L2Ψ + (V2 + δ1B)Ψ ≥ λ∗1Ψ. is bounded above by (2.6). By L we denote the generator of the twisted process (2.3) corresponding to (Φδ, λ∗(V2 + δ1B)) and L2. Therefore Lf = L2f + 2ha2(x)∇ϕδ,∇fi , for f ∈ C2(Rd) , where ϕδ = log Φδ. Since the twisted process (2.3) corresponding to (Φδ, λ∗(V2 + δ1B)) is recurrent by Theorem 2.1, it exists for all time. Moreover, we note that Φδ we obtain from (2.5) that for bΦ = Ψ Now since bΦ is bounded above, by applying the Ito -- Krylov formula to the above equation, we obtain LbΦ − (λ∗1 − λ∗2)bΦ ≥ 0 . 2 )TbΦ+( YT )(cid:3) ≤ kbΦ+k∞e−(λ∗ Letting T → ∞ in this inequality, it follows that Φ(x) ≤ 0 for all x, which contra- dicts the fact that Ψ+ 6= 0. Hence we have λ∗1 = λ∗2. (cid:3) bΦ(x) ≤ Ex(cid:2)e−(λ∗ ∀ T > 0 . 1−λ∗ 2 )T , 1−λ∗ Note that the generators L1 and L2 agree outside the compact set K. Therefore, the processes associated to these generators must agree up to the hitting time τ(K). Lemma 3.3. Let the assumptions of Theorem 2.3 hold, and r > 0 be large enough so that K ⊂ Br. Then we have (3.10) 0 (V2(Xs)−λ∗ Ψ(x) ≤ ExheR τr 1) ds Ψ+(X τr )1{ τr <∞}i . Proof. Choose R > r and x ∈ BR \ Br. Applying the Ito -- Krylov formula to (2.11) we obtain (3.11) Ψ(x) ≤ ExheR τr 0 (V2(Xs)−λ∗ + ExheR τR 0 1) ds Ψ(X τr )1{ τr<τR∧T}i (V2(Xs)−λ∗ 1) ds Ψ(XτR)1{τR< τr∧T}i 18 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY ≤ ExheR τr 0 (V2(Xs)−λ∗ 0 + ExheR τR 1) ds Ψ(XT )1{T≤ τr∧τR}i 0 (V2(Xs)−λ∗ + ExheR T 1) ds Ψ+(X τr )1{ τr <τR∧T}i (V2(Xs)−λ∗ 1) ds Ψ+(XτR)1{τR< τr∧T}i {z } 0 (V2(Xs)−λ∗ I1 1) ds Ψ(XT )1{T≤ τr∧τR}i } {z I2 . + ExheR T We first show that I2 tends 0 as T → ∞. By (ΨR, λR) we denote the principal eigenpair of L2 + V2 in BR with Dirichlet boundary condition. It is known that λR is strictly increasing to λ∗2 as R → ∞. An application of the Ito -- Krylov formula shows that (3.12) ΨR+1(x) = ExheR τr 0 (V2(Xs)−λR+1) ds ΨR+1(X τr )1{ τr<τR∧T}i + ExheR τR (V2(Xs)−λR+1) ds ΨR+1(X τR )1{τR< τr∧T}i + ExheR T 0 (V2(Xs)−λR+1) ds ΨR+1(XT )1{T≤ τr∧τR}i 0 for x ∈ BR \ Br. Since λR < λ∗2 ≤ λ∗1 and ΨR+1 > 0 in BR+1, we deduce that I2 = ExheR T 1 0 (V2(Xs)−λ∗ ≤ ≤ minBR ΨR+1 1)T e(λR+1−λ∗ minBR ΨR+1(cid:16)max max 1) ds Ψ(XT )1{T≤ τr∧τR}i BR Ψ ExheR T BR Ψ(cid:17) ΨR+1(x) → 0, 0 (V2(Xs)−λ∗ where in the last inequality we used (3.12). 1) ds ΨR+1(XT )1{T≤ τr∧τR}i as T → ∞, Therefore letting T → ∞ in (3.11) and using the monotone convergence theorem, we obtain (3.13) Ψ(x) ≤ ExheR τr 0 (V2(Xs)−λ∗ 1) ds Ψ+(X τr )1{ τr <τR}i 1) ds Ψ+(XτR)1{τR< τr}i + ExheR τR {z } (V2(Xs)−λ∗ I3 0 . We next show that lim sup I3 ≤ 0 as R → ∞. Recall that P1 − λ∗1 is critical and therefore, by Theorem 2.1, we have Ψ∗1(x) = ExheR τr (3.14) Since Ψ+ ≤ CΨ∗1 by (2.12), we see using (3.14) that (3.15) (V2(Xs)−λ∗ 0 (V1(Xs)−λ∗ 1) ds Ψ∗1(X τr )1{ τr <∞}i, 1) ds Ψ∗1(XτR)1{τR< τr}i x ∈ Bc r . 0 I3 ≤ C ExheR τR = C Ex(cid:20)eR τR 0 (V2(Xs)−λ∗ 1) ds 1{τR< τr} EXτRheR τr 0 (V1(Xs)−λ∗ 1) ds Ψ∗1(X τr )1{ τr <∞}i(cid:21) LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 19 where in the third line we used strong Markov property. On the other hand, using (2.8) we note that ≤ C ExheR τr ExheR τr 0 ( V (Xs)−λ∗ 0 ( V (Xs)−λ∗ 1) ds 1{τR< τr <∞}Ψ∗1(X τr )i , 1) ds 1{ τr<∞}i < ∞, for x > r . Therefore, since τR → ∞ a.s. as R → ∞, applying the dominated convergence theorem to (3.15) we have (3.16) lim sup R→∞ I3 ≤ 0 . Hence, (3.10) follows from (3.13) and (3.16) by applying the monotone convergence theorem. (cid:3) We are now ready to present the proofs of Theorems 2.3 and 2.4. Proof of Theorem 2.3. Without loss of generality, we may assume that the compact K is large enough so that there exists a ball B ⊂ K satisfying Ψ > 0 in B. Using Lemma 3.1, we deduce that there exists δ ≥ 0 such that λ∗(V2 + δ1B) = λ∗2, and λ∗ is strictly monotone at V2 + δ1B on the right. Let Φδ be the ground state of the operator L + V2 + δ1B − λ∗2. Then, for any r > 0, we have from Theorem 2.1 that (3.17) Fix r > 0 large enough so that K ⊂ Br. Since λ∗2 ≤ λ∗1 we obtain from Lemma 3.3 . Let L be the generator of the twisted process corresponding to (Φδ, λ∗2) and L2. Since L2Ψ + (V2 + δ1B − λ∗1)Ψ ≥ 0, we have and (3.17) that Ψ ≤ κ1Φδ. Define bΦ = Ψ 2) ds Φδ(X τr )1{ τr <∞}i, Φδ(x) = ExheR τr 0 (V2(Xs)+δ1B(Xs)−λ∗ x ∈ Bc r . Φδ (3.18) LbΦ + (λ∗2 − λ∗1)bΦ ≥ 0 . Thus repeating the arguments in the proof of Lemma 3.2, we obtain λ∗1 = λ∗2. But it then follows from (3.18) that {bΦ(Ys)} is a submartingale which is bounded above. This of course, implies that bΦ(Ys) converges almost surely as s → ∞. Since Ys is recurrent, bΦ has to be constant, implying that Ψ = κ2Φδ for some positive κ2 > 0. Using (2.9), we obtain δ = 0, and this completes the proof. Proof of Theorem 2.4. Let K be a compact set such that L1 ≡ L2 in K c. Let h ∈ C+ 0 (Rd) be a function with compact support. Then we know that (cid:3) β 7→ Λβ = λ∗(V1 + βh) is an increasing, convex function [14, Proposition 2.3]. In addition, it is strictly monotone at β = 0. Let βc := inf {β ∈ R : Λβ > Λ−∞}. It is then clear that βc < 0, and hence it follows from [5, Theorem 2.7] that for some β < 0, close to 0, the twisted process corresponding to the eigenpair (Ψβ, Λβ) and L1 is recurrent (in fact, geometrically ergodic), and Λβ < λ∗1 = Λ0. We also have (3.19) L1Ψβ + (V1 + βh)Ψβ = ΛβΨβ . Moreover, by Theorem 2.1, Ψβ has a stochastic representation, i.e., (3.20) Ψβ(x) = ExheR τr 0 (V1(Xs)−Λβ ) ds Ψβ(X τr )1{ τr<∞}i. 20 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY In (3.20) we use a radius r large enough so that the support of h and the set K lie in Br. Also by Lemma 3.2 we have λ∗1 = λ∗2. Let δ = 1 2 (λ∗1 − Λβ) > 0. It is clear that we can choose r large enough so that For such a choice of r, we note from (3.20) that V2(x) − λ∗2 + δ = V2(x) − λ∗1 + δ < V1(x) − Λβ 2+δ) ds 1{ τr<∞}i < ∞, ExheR τr 0 (V2(Xs)−λ∗ (3.21) for all x ≥ r . x ≥ r. Using (3.21) and the arguments in [5, Theorem 2.2] (see for instance, (2.30) in [5]) it is easy to show that λ∗2 is strictly monotone at V2. Therefore, in order to complete the proof, it remains to show that Ψ is a positive multiple of Ψ∗2. Since V1 + βh − Λβ ≥ V1 − λ∗1 outside some compact set K0, we obtain from (3.19) that L1Ψβ + (V1 − λ∗1)Ψβ ≤ 0 ∀ x ∈ K c 0 . Therefore, by the minimal growth at infinity of Ψ∗1, we can find a constant κβ satisfying Ψ∗1 ≤ κβΨβ in Rd. Combining this with (2.12), we have Ψ+ ≤ CκβΨβ. As earlier, we fix r large enough so that V2(x) − λ∗2 < V1(x) − Λβ, L1 ≡ L2, and h(x) = 0 for x ≥ r. We apply the Ito -- Krylov formula to (2.11) to obtain By the choice of r, we can estimate the second term as follows Ψ(x) ≤ ExheR τr 0 (V2(Xs)−λ∗ (3.22) ExheR τR 0 (V2(Xs)−λ∗ 0 2) ds Ψ(X τr )1{ τr <τR}i + ExheR τR 2) ds Ψ+(X τr )1{ τr>τR}i ≤ κ2 ExheR τR (V2(Xs)−λ∗ 2) ds Ψ(XτR)1{ τr >τR}i . The right hand side of (3.22) tends to 0, as R → ∞, by (3.20). Hence letting R → ∞, we obtain 0 (V1(Xs)−Λβ ) ds Ψβ(XτR)1{ τr >τR}i . 2) ds Ψ(X τr )1{ τr<∞}i . Ψ(x) ≤ ExheR τr 0 (V2(Xs)−λ∗ Since Ψ∗2 also has a stochastic representation by Theorem 2.1, this implies that Ψ ≤ κ1Ψ∗2 for some κ1 > 0. With Φ = Ψ we have Ψ∗ 2 LΦ ≥ 0 , where L is the generator of twisted process Y corresponding to (Ψ∗2, λ∗2) and L2. Thus, {Φ(Ys)} is a submartingale which is bounded from above. Since the twisted process Y is recurrent by Theorem 2.1, Φ must be constant. Since Ψ+ 6= 0, this implies that Ψ is a positive function, which means of course, that it is a positive multiple of Ψ∗2. (cid:3) One interesting by-product of the proof of Theorem 2.4 is the corollary that follows. This result however might be known, but we could not locate it in the literature. LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 21 Corollary 3.3. Let L be the operator in (2.2), and λ∗ be the principal eigenvalue of L+V , where V is a locally bounded function. In addition, suppose that L+V −λ∗ is critical, and let Ψ∗1 denote the ground state. Then, there does not exist any non-zero solution Ψ ∈ W2,d loc(Rd) of LΨ + V Ψ = λΨ, for λ > λ∗, with Ψ ≤ κΨ∗1. We can improve the above results to a larger class of potentials if we impose a 'stability' condition of the underlying dynamics X. Let us assume the following (H) There exists a lower-semicontinuous, inf-compact function ℓ : Rd → [0,∞) such that 0 ℓ(Xs) dsi < ∞ for all x ∈ Rd . By o(ℓ) we denote the collection of functions f : Rd → R satisfying ExheR T 1 T lim sup T→∞ f (x) ℓ(x) = 0 . lim sup x→∞ We say that the elliptic operator L satisfies (H) if the process X with extended generator L satisfies (H). It is easy to see that under hypothesis (H), the process is recurrent. Therefore, if (H) holds for L1, it follows from [4, Lemma 2.3] that λ∗1(ℓ) is finite. Moreover, there exists a positive eigenfunction ϕ1 ∈ W2,p loc (Rd), p > 1, with inf Rd ϕ1 > 0, that satisfies L1ϕ1 + (ℓ − λ∗1(ℓ))ϕ1 = 0, in Rd . If L2 is a small perturbation of L1, then L2 also satisfies (H). To see this, consider a ball B ⊂ Rd such that L1 = L2 in Bc. Let χ : Rd → [0, 1] be a smooth function that vanishes in B and equals 1 outside a ball Br ⊃ ¯B. Define ϕ2 = (1 − χ) + χϕ1. Note that ϕ2 = 1 in B, and ϕ2 ≥ 1 ∧ inf Rd ϕ1 > 0 on Rd. Then, for some positive constants κ1 and κ2, we have (3.23) L2ϕ2 = L2(1 − χ) + χL2ϕ1 + ϕ1L2χ + 2ha2∇χ,∇ϕ1i = L1(1 − χ) + χL1ϕ1 + ϕ1L1χ + 2ha1∇χ,∇ϕ1i = L1(1 − χ) + χ(λ∗1(ℓ) − ℓ)ϕ1 + ϕ1L1χ + 2ha1∇χ,∇ϕ1i ≤ (κ1 − ℓ) ϕ1 ≤ (κ2 − ℓ) ϕ2 on Rd . In (3.23), the first inequality arises from the fact that inf Rd ϕ1 > 0, while in the second inequality we use the fact that ϕ1 = ϕ2 on Bc r, and inf Rd ϕ2 > 0. Equation (3.23) of course implies that 1 T lim sup T→∞ ExheR T 0 ℓ(X 2 s ) dsi < κ2 for all x ∈ Rd , where X 2 denotes the diffusion process with generator L2. We have the following result. Theorem 3.1. Let all the assumptions of Theorem 2.4 hold, except we replace V1 − V2 ∈ B0(Rd) with Vi ∈ o(ℓ). Moreover, assume that (H) holds for L1. Then the conclusion of Theorem 2.4 also holds. 22 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY (3.24) Proof. By [5, Theorem 3.2] we know that λ∗ is strictly monotone at both V1 and V2. Therefore, in order to complete the proof, we only need to show that λ∗1 = λ∗2 and Ψ∗2 = Ψ. Since ℓ is inf-compact, (H) implies that the processes X i, i = 1, 2, in Rd, i = 1, 2 , loc(Rd), p ≥ 1, such that are recurrent. Moreover, there exists a positive bV i ∈ W2,p LibV i + ℓbV i = λ∗i (ℓ)bV i and inf Rd bV i > 0. Let Br be a ball such that (cid:12)(cid:12)Vi(x) − max{λ∗1(ℓ), λ∗2(ℓ)}(cid:12)(cid:12) ≤ θ(cid:0)ℓ(x) − max{λ∗1(ℓ), λ∗2(ℓ)}(cid:1) ∀ x ∈ Bc and L1 = L2, i = 1, 2, on Bc r, for some constant θ ∈ (0, 1). Recall that τr denotes the first hitting time to Br. Since both processes agree outside Br, in what follows we use X to denote any one of these processes. Then applying the Ito -- Krylov formula to (3.24), followed by Fatou's lemma, we obtain (3.25) r , (3.26) ExheR τr 0 (ℓ(Xs)−λ∗ i (ℓ)) dsbV i(X τr )i ≤ bV i(x) for x ∈ Bc r . We can choose Br large enough so that Ψ+ 6= 0 in Br. Let B ⋐ Br be such that Ψ > 0 in B. By Lemma 3.1 we can find δ ≥ 0 such that λ∗ is strictly monotone on the right at V2 + δ1B and λ∗2 = λ∗(V2 + δ1B). Let Ψδ be the corresponding principal eigenfunction. By Theorem 2.1 we have (3.27) 0 (V2(Xs)−λ∗ Since L1 + V1 − λ∗1 is critical by hypothesis, we have (3.28) 0 (V1(Xs)−λ∗ ExheR τr ExheR τr 2) ds Ψδ(X τr )i = Ψδ(x) 1) ds Ψ∗1(X τr )i = Ψ∗1(x) for x ∈ Bc r . for x ∈ Bc r . It follows by (3.25), (3.26), and (3.28) that Ψ∗1(x) ≤ κ(bV 1(x))θ in Rd, for some We claim that constant κ. To prove the claim we define Γ(R, m) = {x ∈ ∂Br : Ψ∗1(x) ≥ m} for m ≥ 1. Then 1) ds Ψ∗1(XτR )1{τR< τr}i −−−−→R→∞ 0 . (3.29) ExheR τR 0 (V2(Xs)−λ∗ 0 0 0 (V2(Xs)−λ∗ 0 θ(ℓ(Xs)−λ∗ (V2(Xs)−λ∗ ExheR τR 1) ds Ψ∗1(XτR)1{τR< τr}i ≤ m ExheR τR + ExheR τR ≤ m ExheR τR + κ1m1−1/θ ExheR τR ≤ m ExheR τR Ψ(x) ≤ ExheR τr ∧τR∧T (V2(Xs)−λ∗ θ(ℓ(Xs)−λ∗ θ(ℓ(Xs)−λ∗ 0 0 0 (ℓ(Xs)−λ∗ 1(ℓ)) ds 1{τR< τr}i 1) ds Ψ∗1(XτR )1{x∈Γ(R,m)}1{τR< τr}i 1(ℓ)) ds 1{τR< τr}i 1(ℓ)) dsbV 1(XτR )1{τR< τr}i 1(ℓ)) ds 1{τR< τr}i + κ1m1−1/θbV 1(x) . 1) ds Ψ(X τr∧τR∧T )i , T > 0 . Then (3.29) follows by first letting R → ∞, and then m → ∞. Applying the Ito -- Krylov formula (3.1) to (2.11) we obtain (3.30) LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 23 Since V2 − λ∗2 ≤ ℓ − λ∗1(ℓ) in Bc r, and for every fixed R > r, we have (ℓ(Xs)−λ∗ 0 ExheR τr ∧τR ExheR T 0 (V2(Xs)−λ∗ r < x < R , 1(ℓ)) dsi < ∞ , 1) ds Ψ(XT )1{T≤ τr∧τR}i −−−−→T→∞ 1) ds Ψ(X τr )i . 0 (V2(Xs)−λ∗ 0 . Ψ(x) ≤ ExheR τr Hence, first letting T → ∞, and then R → ∞ in (3.30), and using (2.12) and (3.29), we obtain Combining this with (3.27) we have Ψ ≤ C1Ψδ. Now mimicking the arguments in the last part of the proof of Theorem 2.3, we obtain λ∗1 = λ∗2, and Ψ = Ψδ with δ = 0. (cid:3) We next exhibit a family of operators for which (H) holds. Example 3.1. Let δ1I ≤ a(x) ≤ δ2I, for δ1, δ2 > 0 and x ∈ Rd. Also b(x) = b1(x) + b2(x) where b2 ∈ L∞(Rd), and hb1(x), xi ≤ −κxα on the complement of a compact set in Rd , for some constant κ > 0, and some α ∈ (1, 2]. Let ζ be a positive, twice differentiable function in Rd such that ζ(x) = exp(θxα) for x ≥ 1. If we choose θ ∈ (0, 1) small enough, then it is routine to check that there exists R0 > 0 such that Lζ(x) ≤ − κθ 2 x2α−2ζ(x) for x ≥ R0 . The above inequality is known as a (geometric) Foster -- Lyapunov stability condition and ζ is generally referred to as a Lyapunov function. Therefore, if we choose a function ℓ which coincides with κθ 2 x2α−2 outside a compact set, then using the above inequality and Ito's formula one can easily verify that (H) holds. 4. A lower bound on the decay of eigenfunctions The main goal of this section is to exhibit a sharp lower bound on the decay of supersolutions, and also to use this estimate to prove several results for positive solutions. Lemma 4.1. Suppose that there exist positive constants M and η0, and some β ∈ [0, 2] such that (4.1) for all x outside some compact set in Rd. Let α ≥ β and K, γ be any positive constants satisfying hξ, a(x)ξi ≥ η0ξ2 ∀ ξ ∈ Rd and (cid:12)(cid:12)hb(x), xi(cid:12)(cid:12) ≤ Mxβ , +s M 2 M 2η0 4η2 0 + (4.2) Kα > γ η0 , and lim x→∞ 1 xα dXi=1 aii(x) = 0 , and define V(x) := exp(−Kxα). Then there exists r0 > 0 such that for every r ≥ r0 we have (4.3) 0 Xs2α−2 ds V(X τr )1{ τr <∞}i ≥ V(x) Exhe−γ R τr for x ≥ r. 24 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY Proof. By (4.2) we see that (4.4) We have η0(Kα)2 − M (Kα) − γ > 0 . ∂ ∂xi V(x) = −Kαxα−2xiV ∂2 ∂xi∂xj V(x) = (Kα)2x2α−4xixjV(x) − Kα(α − 2)xα−4xixjV(x) − Kαxα−2V(x)δij for 1 ≤ i, j ≤ d, and x ≥ 1. This implies that LV(x) = Kαx2α−2(cid:18)Kα − α − 2 xα (cid:19)V(x) x2 hx, axi dXi=1 − Kαxα−2V(x) ≥ Kαx2α−2 Kαη0 − M − η0(α − 2) xα − aii(x) + hb(x),∇V(x)i aii(x)!V(x) , dXi=1 1 xα V(x) ≤ Exhe−γ R τr ≤ Exhe−γ R τr which combined with (4.2) and (4.4) shows that there exists r0 ≥ 1, such that (4.5) Let R > r ≥ r0. Applying the Ito -- Krylov formula to (4.5), we obtain (4.6) for x ≥ r0 . LV(x) ≥ γx2α−2V(x) 0 Xs2α−2 ds V(X τr )1{ τr<τR}i 0 Xs2α−2 ds V(X τr )1{ τr<∞}i + Exhe−γ R τR + Exhe−γ R τR Xs2α−2 ds V(XτR)1{τR< τr}i Xs2α−2 ds V(XτR)1{τR< τr}i . Xs2α−2 ds V(XτR)1{τR< τr}i ≤ ExhV(XτR )1{τR< τr}i ≤ e−KRα for r ≤ x ≤ R. On the other hand, Exhe−γ R τR 0 as R → ∞. Thus by letting R → ∞ in (4.6), we obtain (4.3). 0 0 → 0 , (cid:3) The above result should be compared with Carmona [16], Carmona and Simon [17], where a weaker lower bound was obtained for L´evy processes. In these papers, the stationarity and independent increment property of the underlying process is used, and also the proof is much more complicated. For instance, see [16, Propo- sition 4.1] when X is a Brownian motion. We next use Lemma 4.1 to provide a quantitative estimate on the decay of positive supersolutions in the outer domain. Theorem 4.1. Assume (4.1), and let γ, α, and V be as in Lemma 4.1. Let K ⊂ Rd be a compact set, and suppose u ∈ W2,d loc(Kc) is a nontrivial nonnegative function such that (4.7) Lu + V u = 0 in Kc , LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 25 where V is locally bounded, and V (x) ≥ −γx2α−2 for all x sufficiently large. Then there exists a positive constant C, not depending on u, provided we fix u(x0) = 1 at some x0 ∈ Kc, and r > 0 such that (4.8) u(x) ≥ C V(x) for x > r . Proof. Let r0 be as in Lemma 4.1. By the hypotheses of the theorem we may choose r > r0 and sufficiently large, so that applying the Ito -- Krylov formula to (4.7) we have (4.9) u(x) ≥ ExheR τr ≥ Exhe−γ R τr 0 V (Xs) ds u(X τr )1{ τr <τR}i + ExheR τR 0 Xs2α−2 ds u(X τr )1{ τr <τR}i 0 V (Xs) ds u(XτR)1{τR< τr}i for x > r . By the Harnack inequality we have minz=r u(z) ≥ κ for some positive constant κ which does not depend on u. We let R → ∞ in (4.9) and apply Fatou's lemma to obtain u(x) ≥ Exhe−γ R τr 0 Xs2α−2 ds u(X τr )1{ τr <∞}i u(z)(cid:19) eKrα Exhe−γ R τr 0 Xs2α−2 ds V(X τr )1{ τr <∞}i ≥ (cid:18) min z=r ≥ κ eKrα for x > r , V(x) by (4.3). Thus (4.8) follows. Remark 4.1. If u ∈ W2,d loc (Kc) is a nontrivial nonnegative supersolution of (4.7) then it is necessarily positive on Kc by the strong maximum principle. Thus (4.8) is valid for nonnegative supersolutions; however the constant C depends, in general, on u. (cid:3) As an immediate corollary to Lemma 4.1 and Theorem 4.1 we have the following. Corollary 4.1. Let u ∈ W2,d loc (Rd) be a nontrivial nonnegative solution of trace(a∇2u) + hb,∇ui + V u = 0 in Kc . Here, we assume that supKcb(x) ≤ M , supKcV (x) ≤ γ, that a is bounded, and hξ, a(x)ξi ≥ η0ξ2 for all ξ ∈ Rd. Then for every ε′ > 0 there exist positive constants Cε′ and Rε′ such that u(x) ≥ Cε′ exp −(cid:18) M 2η0 +s M 2 4η2 0 + γ η0 + ε′(cid:19)x! , x ≥ Rε′ . Proof. Let K = M 2η0 from Theorem 4.1. +q M 2 4η2 0 + γ η0 + ε′, α = 1, and β = 0. Then the result follows (cid:3) Let us now discuss some important aspects of Theorem 4.1 and Corollary 4.1. When a = I, b = 0, and V is the potential function for the two body problem, a similar lower bound was obtained by Agmon [2]. In the context of Corollary 4.1, a lower bound was also obtained by Kenig, Silvestre and Wang [26, Theorem 1.5] for solutions which can be sign-changing; however it is assumed in [26] that V ≤ 0, a is the identity matrix, and d = 2. In contrast, Corollary 4.1 does not require 26 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY these assumptions, but applies only to nonnegative solutions u. Note that the lower bound in [26, Theorem 1.5] is of the form e−CR(log R)2 in the radial direction R, whereas the lower bound in Corollary 4.1 is of the form e−CR, and hence it is tighter. When b = 0, this bound is also sharper than the one conjectured by Kenig in [27, Question 1]. In fact, this bound is optimal in some sense. To see this, take u(x) = e−x in Rd. Then ∆u + V u = 0 in {x > d} where V (x) = −1 + d−1 . Since x we can take ε′ arbitrarily small, the bound in Corollary 4.1 is very sharp. We apply Corollary 4.1 to semi-linear or quasi-linear operators to find a lower bound on the decay of solutions. Corollary 4.2. Grant the hypotheses in Lemma 4.1. (a) Let f : (0,∞) → (0,∞) be a continuous function such that 1 s f (s) < +∞ , lim sup sց0 loc (Rd) be a bounded, positive solution of Lu = f (u). Then there and u ∈ W2,d exist constants γ > 0 and Cγ, depending on kuk∞, such that (4.10) u(x) ≥ Cγe−γx for all x sufficiently large. (b) Let U1, U2 be two compact metric spaces, and V, b : Rd × U1 × U2 → Rd be two continuous functions with kV k∞ < ∞, and b(·, v1, v2) satisfying (A2 ) uniformly in (v1, v2) ∈ U1 × U2. If u ∈ W2,d loc(Rd) is a positive solution of min v1∈U1 max v2∈U2(cid:2)aij ∂iju + bi(x, v1, v2)∂iu + V (x, v1, v2)u(cid:3) = 0 , then it satisfies (4.10). Proof. For part (a), note that since u is bounded, we have f (u) ≤ Cu for some constant C (depending on kuk∞). Thus we obtain Lu ≤ C u , and the proof follows from Corollary 4.1. For part (b), observe that we can find measurable selectors v∗i : Rd → Ui satisfying aij ∂iju + bi(cid:0)x, v∗1(x), v∗2 (x)(cid:1)∂iu + V(cid:0)x, v∗1 (x), v∗2 (x)(cid:1)u = 0 . The rest follows as before using Corollary 4.1. (cid:3) In the rest of this section we discuss some connections of Theorem 4.1 and Corol- lary 4.1 with the Landis conjecture, and provide a partial answer to this con- jecture. In 1960s, E. M. Landis conjectured (see [28]) that if u is a solution to ∆u + V u = 0, with kV k∞ ≤ q2, and there exist positive constants ε and Cε such that u(x) ≤ Cε e−(q+ε)x, then u ≡ 0. He also proposed a weaker version of this conjecture which states that if u(x) ≤ Cke−kx for any positive k, and some constant Ck, then u ≡ 0. This conjecture was disproved by Meshkov in [30] who constructed a non-zero solution to ∆u + V u = 0 which satisfies u(x) ≤ Ce−cx for some positive constants c and C. It is also shown in [30] that if for any k > 0, there exists a constant Ck satisfying u(x) ≤ Cke−kx , then u is identically 0. The counterexample by Meshkov has V and u complex valued. Therefore the Lan- dis conjecture remains open for real valued solutions and potentials. It is interesting to note that the Landis conjecture concerns the unique continuation property of u at infinity. In practice, Carleman type estimates are commonly used to treat such 4/3 4/3 LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 27 problems, but since a Carleman estimate does not distinguish between real and complex valued functions, it is hard to improve the results of Meshkov using such estimates. Landis' conjecture was recently revisited by Kenig, Silvestre and Wang [26], and Davey, Kenig and Wang [18] for V ≤ 0 and d = 2. Note that if V ≤ 0 then Landis' conjecture follows from the strong maximum principle. The key con- tribution of [18, 26] is a lower bound on the decay of u. On the other hand, if we assume u to be nonnegative, then the Landis conjecture follows from Corollary 4.1. In Theorem 4.2 below we show that Landis' conjecture holds under the assumption that λ∗(V ) ≤ 0. It should be observed that λ∗(V ) ≤ 0 does not necessarily imply that V ≤ 0. Theorem 4.2. Let Lu + V u = 0 and suppose that the following hold. (i) a is bounded and uniformly elliptic with ellipticity constant η0, kbk∞ ≤ M , (ii) For some positive constants ε and Cε, we have kV k∞ ≤ γ, and λ∗(V ) ≤ 0. u(x) ≤ Cε exp −(cid:18) M 2η0 +s M 2 4η2 0 + γ η0 + ε(cid:19)x! , ∀ x ∈ Rd . Then u ≡ 0. Proof. Let Ψ be a positive function in W2,d loc (Rd) satisfying (4.11) LΨ + V Ψ = 0 . Existence of such Ψ follows, for example, from [14, Theorem 1.4] and the fact that λ∗(V ) ≤ 0. By L we denote the twisted process corresponding to the eigenpair (Ψ, 0) i.e., where ψ = log Ψ. Let Φ = u Ψ . Then it is easy to check from (4.11) that Lf = Lf + 2ha∇ψ,∇fi , f ∈ C2(Rd) , (4.12) LΦ = 0 . On the other hand, by Corollary 4.1 we have Ψ(x) ≥ Cε′ exp −(cid:18) M 2η0 +s M 2 4η2 0 + γ η0 + ε′(cid:19)x! ∀ x > Rε′ . for ε′ < ε, and constants Cε′ and Rε′ . This of course, implies that Φ(x) → 0 as x → ∞. Therefore, applying the strong maximum principle to (4.12), we deduce that Φ ≡ 0, which in turn implies that u ≡ 0. This completes the proof. Remark 4.2. Recall that W denotes the standard Brownian motion. Let V ∈ C0(Rd) be such that (cid:3) (4.13) ExheR ∞ 0 1 2 V +(Ws) dsi < ∞ , 1 ∀ x ∈ Rd . It is then known that v(x) := ExheR ∞ 0 2 V +(Ws) dsi is a positive solution to ∆v + V +v = 0 . This of course implies that ∆v + V v ≤ 0, and therefore, λ∗(V ) ≤ 0. For d ≥ 3, if we have 2 (d − 2)ωdZRd sup x∈Rd V +(y) x − yd−2 < 1 , 28 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY with ωd denoting the surface measure of the unit sphere in Rd, then V also satisfies (4.13) by Khasminskii's lemma [43, Lemma B.1.2]. Recall that, as shown in Remark 2.5, if V ∈ C0(Rd) and λ∗ is not strictly mono- tone at V , then λ∗(V ) ≤ 0. This allows us to apply Theorem 4.2, to conclude that Landis' conjecture holds if λ∗(L, V ) is not strictly monotone at V ∈ C0(Rd). This of course applies to the general class of operators L satisfying (A1) -- (A3). Though we have not been able to prove the Landis conjecture in its full gen- erality, we can validate this conjecture for a large class of potentials, including compactly supported potentials. This can be done with the help of the Radon transformation and a support theorem from Helgason [23]. See also [8] which uses a similar approach, albeit for homogeneous elliptic equations. Theorem 4.3. Suppose ∆u + hb,∇ui + V u = 0 in Bc, where B is a bounded ball, and the following hold. (i) kbk∞ ≤ M, kV k∞ ≤ γ. There exists a ball Br, r > 0, with B ⊂ Br, such that b and V are constant in Bc r. (ii) For some positive ε, Cε, we have u(x) ≤ Cε exp −(cid:18) M 2 +r M 2 4 + γ + ε(cid:19)x! , ∀ x ∈ Bc . Then u ≡ 0. Proof. Without loss of generality we may assume 0 ∈ B, and b(x) = b0, V (x) = k for all x ∈ Bc r. Also by standard regularity theory of elliptic PDE we may assume that u is smooth in Bc 1. Let (ω, p) ∈ Sd × R where Sd is the (d − 1)-dimensional unit sphere in Rd. We note that any hyperplane Rd can be identified by (ω, p) up to the equality (−ω,−p) = (ω, p). Let ξ be a hyperplane in Rd, i.e., for some (ω, p) we have ξ = {x ∈ Rd : hx, ωi = p}. The Radon transformation of u is defined as u(ξ) := Zξ u(y) S(dy) , where S(dy) is the surface measure on ξ . We claim that if the hyperplane does not intersect B1, then we have (4.14) u(ξ) = 0 . If (4.14) is true, then since u decays exponentially fast to 0 at infinity by (ii), then the support theorem [23, Theorem 1.2.6 and Corollary 1.2.8] implies that u ≡ 0 in Bc r. This in turn, implies that u ≡ 0 by the unique continuation property of Hormander [24, Theorem 2.4]. In order to complete the proof we need to prove (4.14). Note that if we define v(x) = u(Mx), where M is any rotation matrix, then T ∆v + hM b0,∇vi + kv = 0 , x ∈ Bc 1 . Also a rotation does not change the norm of b. Therefore, without loss of generality, we may assume that ξ = ξ(κ0) := {x ∈ Rd : x1 = κ0, κ0 > 0}. Define for s ≥ κ0, u(s, ¯x) d¯x, where ¯x = (x2, . . . , xd) ∈ Rd−1 . u(y) S(dy) =ZRd−1 w(s) := Zξ(s) LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 29 Note that w is smooth in [κ0,∞) due to the smoothness of u and its decay at infinity. Moreover, (4.15) d2u ds2 (s, ¯x) d¯x d2w(s) ds2 = ZRd−1 = −ZRd−1 bi 0∂iu(s, ¯x) d¯x ∂iiu(s, ¯x) d¯x −ZRd−1 dXi=2 b1 0∂1u(s, ¯x) d¯x − kZRd−1 dXi=1 − kZRd−1 u(s, ¯x) d¯x u(s, ¯x) d¯x = −ZRd−1 = −b1 0 dw(s) ds − kw(s) , where in the second equality we use the equation satisfied by u, and in the third equality we use the fundamental theorem of calculus. Thus we obtain from (4.15) a second-order ODE with constant coefficients, given by (4.16) d2w(s) ds2 + b1 0 dw(s) ds + kw(s) = 0 in [κ0,∞) . We solve this ODE explicitly, and using the decay property of u in (ii) we show that w(s) = 0 in [κ0,∞). In particular, w(κ0) = 0 which proves (4.14). Denote by κ1 = M 4 + γ. We first show that for ε′ < ε there exists a positive constant Cε′ such that 2 +q M 2 (4.17) for s ∈ [κ0,∞) . By (ii) and a choice of s0, satisfying √s0(s0 − 1) > 2, we get w(s) ≤ Cε′ e−(κ1+ε′)s , w(s) ≤ ZRd−1u(s, ¯x) d¯x e−κεs√1+r2 rd−2dr e−κεs√1+¯x2 d¯x e−κεx d¯x [for κε = κ1 + ε] ≤ CεZRd−1 = Cεsd−1ZRd−1 = κ2sd−1Z ∞ = κ2sd−1(cid:20)Z s0 e−κεs√1+r2 rd−2dr +Z ∞ ≤ κ2sd−1(cid:20)sd−1 e−κεs +Z ∞ ≤ κ3sd−1e−κεs(cid:20)1 +Z ∞ s0 s0 e−κεs(1+√r) rd−2dr(cid:21) e−κεκ0√r rd−2dr(cid:21) , s0 0 0 0 e−κεs√1+r2 rd−2dr(cid:21) 30 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY where in the third inequality we have used the fact that √1 + r2 > 1 + √r for r ≥ s0. Equation (4.17) easily follows from the above estimate. To solve (4.16) we find the roots of the characteristic polynomial of the ODE which are given by r1 = − 1 2 b1 0 + 0)2 − 4k , and r2 = − 1 2 b1 0 − 1 2q(b1 1 2q(b1 0)2 − 4k . The solution of (4.16) can be written as w(s) = c1er1s + c2er2s , 0)2 − 4k < 0, then where the constants c1 and c2 are uniquely determined. Now if (b1 the roots are complex and the decay of w is of order e− 1 0s which is larger than the RHS of (4.17). Therefore, we must have c1 = c2 = 0. On the other hand, if (b1 Hence we must have w(s) = 0 in [κ0,∞). This completes the proof. 0)2 − 4k(cid:12)(cid:12) ≤ κ1, we conclude that c1 = c2 = 0. 0)2−4k ≥ 0, and since(cid:12)(cid:12)− 1 As a concluding remark, we show that the Landis conjecture is true for solutions that satisfy a reverse Poincar´e inequality. Specifically, consider an operator L + V with bounded coefficients and a the identity matrix. We say a solution u to Lu + V u = 0 satisfies (G) if the following holds: 2p(b1 0± 1 2 b1 2 b1 (cid:3) (G) There exist positive constants r and C, independent of x, such that ZBr (x)∇u(y)2 dy ≤ CZBr (x)u(y)2 dy, ∀ x ≫ 1 . Remark 4.3. Note that ifRBr (x)∇u(y)2 dy = 0 for some x ∈ Rd, then it is shown by Hormander [24, Theorem 2.4] that u ≡ 0. The following (weaker) Landis' conjecture is true for solutions satisfying (G). Suppose that ∆u + hb,∇ui + V u = 0 in Bc, with B a bounded ball, u ∈ C2(Bc), and the following hold. • kbk∞ ≤ M, kV k∞ ≤ q2, and u satisfies (G). Then for κ =(cid:2)2(M√C + q2 + C)(cid:3)1/2 + 1, we have u2(y) dy ≥ Cκ exp(−κx) (4.18) ZBr (x) for all x ≫ 1 . In particular, if for every k ∈ N, we have u(x) ≤ Cke−kx for some constant Ck, then u ≡ 0. In order to prove this claim, we define v(x) := ZBr (x) u2(y) dy . Since u ∈ C2(Bc), we have v ∈ C2(Bc calculation shows that 1) for some large ball B1 ⋑ B. A straightforward u∆u dy + 2ZBr(x)∇u2 dy uhb,∇ui dy − 2ZBr(x) ∆v(x) = 2ZBr(x) = −2ZBr (x) ≤ 2M(cid:18)ZBr(x)u2 dy(cid:19)1/2(cid:18)ZBr (x)∇u2 dy(cid:19)1/2 V u2 dy + 2ZBr (x)∇u2 dy + 2q2ZBr (x) u2 dy LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 31 ≤ (2M√C + 2q2 + 2C)v(x) . Therefore (4.18) follows from Corollary 4.1. + 2ZBr (x)∇u2 dy Acknowledgements We wish to convey our sincere thanks to Yehuda Pinchover for his helpful com- ments and remarks throughout this project. The research of Ari Arapostathis was supported in part by the Army Research Office through grant W911NF-17-1-001, in part by the National Science Foundation through grant DMS-1715210, and in part by Office of Naval Research through grant N00014-16-1-2956. The research of Anup Biswas was supported in part by an INSPIRE faculty fellowship (IFA13/MA-32) and DST-SERB grant EMR/2016/004810. Debdip Ganguly was supported in part at the Technion by a fellowship of the Israel Council for Higher Education and the Israel Science Foundation (Grant No. 970/15) founded by the Israel Academy of Sciences and Humanities. References [1] S. Agmon, On positivity and decay of solutions of second order elliptic equations on Rie- mannian manifolds, Methods of functional analysis and theory of elliptic equations (Naples, 1982), 1983, pp. 19 -- 52. MR819005 [2] , Bounds on exponential decay of eigenfunctions of Schrodinger operators, Schrodinger operators (Como, 1984), 1985, pp. 1 -- 38. MR824986 [3] A. Ancona, First eigenvalues and comparison of Green's functions for elliptic operators on manifolds or domains, J. Anal. Math. 72 (1997), 45 -- 92. MR1482989 [4] A. Arapostathis and A. Biswas, Infinite horizon risk-sensitive control of diffusions without any blanket stability assumptions, Stochastic Processes and their Applications 128 (2018), no. 5, 1485 -- 1524. MR3780687 [5] A. Arapostathis, A. Biswas, and S. Saha, Strict monotonicity of principal eigenvalues of elliptic operators in Rd and risk-sensitive control, J. Math. Pure. Appl. (2018), to appear, available at https://arxiv.org/abs/1704.02571. [6] A. Arapostathis, V. S. Borkar, and M. K. Ghosh, Ergodic control of diffusion processes, Encyclopedia of Mathematics and its Applications, vol. 143, Cambridge University Press, Cambridge, 2012. MR2884272 [7] A. Arapostathis, M. K. Ghosh, and S. I. Marcus, Harnack's inequality for cooperative weakly coupled elliptic systems, Comm. Partial Differential Equations 24 (1999), no. 9-10, 1555 -- 1571. MR1708101 [8] A. T. Astakhov and V. Z. Meshkov, On the possible decay at infinity of solutions of homo- geneous elliptic equations, Differ. Equ. 50 (2014), no. 11, 1548 -- 1550. Translation of Differ. Uravn. 50 (2014), no. 11, 1548 -- 1550. MR3369163 [9] M. Bardi and A. Cesaroni, Liouville properties and critical value of fully nonlinear elliptic operators, J. Differential Equations 261 (2016), no. 7, 3775 -- 3799. MR3532054 [10] M. T. Barlow, On the Liouville property for divergence form operators, Canad. J. Math. 50 (1998), no. 3, 487 -- 496. MR1629807 [11] S. Beckus and Y Pinchover, Shnol-type theorem for the Agmon ground state, ArXiv e-prints 1706.04869 (2017), available at https://arxiv.org/abs/1706.04869. [12] H. Berestycki, L. Caffarelli, and L. Nirenberg, Further qualitative properties for elliptic equa- tions in unbounded domains, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 25 (1997), no. 1-2, 69 -- 94 (1998). Dedicated to Ennio De Giorgi. MR1655510 [13] H. Berestycki, L. Nirenberg, and S. R. S. Varadhan, The principal eigenvalue and maximum principle for second-order elliptic operators in general domains, Comm. Pure Appl. Math. 47 (1994), no. 1, 47 -- 92. MR1258192 32 ARI ARAPOSTATHIS, ANUP BISWAS, AND DEBDIP GANGULY [14] H. Berestycki and L. Rossi, Generalizations and properties of the principal eigenvalue of elliptic operators in unbounded domains, Comm. Pure Appl. Math. 68 (2015), no. 6, 1014 -- 1065. MR3340379 [15] V. I. Bogachev, N. V. Krylov, and M. Rockner, On regularity of transition probabilities and invariant measures of singular diffusions under minimal conditions, Comm. Partial Differ- ential Equations 26 (2001), no. 11-12, 2037 -- 2080. MR1876411 [16] R. Carmona, Pointwise bounds for Schrodinger eigenstates, Comm. Math. Phys. 62 (1978), no. 2, 97 -- 106. MR505706 [17] R. Carmona and B. Simon, Pointwise bounds on eigenfunctions and wave packets in N -body quantum systems. V. Lower bounds and path integrals, Comm. Math. Phys. 80 (1981), no. 1, 59 -- 98. MR623152 [18] B. Davey, C. Kenig, and J.-N. Wang, The Landis conjecture for variable coefficient second- order elliptic PDEs, Trans. Amer. Math. Soc. 369 (2017), no. 11, 8209 -- 8237. MR3695859 [19] B. Devyver, M. Fraas, and Y. Pinchover, Optimal Hardy weight for second-order elliptic operator: an answer to a problem of Agmon, J. Funct. Anal. 266 (2014), no. 7, 4422 -- 4489. MR3170212 [20] D. Ganguly and Y. Pinchover, On Green functions of second-order elliptic operators on Riemannian manifolds: the critical case, J. Funct. Anal. 274 (2018), no. 9, 2700 -- 2724. MR3771841 [21] N. Ghoussoub and C. Gui, On a conjecture of De Giorgi and some related problems, Math. Ann. 311 (1998), no. 3, 481 -- 491. MR1637919 [22] I. Gyongy and N. Krylov, Existence of strong solutions for Ito's stochastic equations via approximations, Probab. Theory Related Fields 105 (1996), no. 2, 143 -- 158. MR1392450 [23] S. Helgason, The Radon transform, Second, Progress in Mathematics, vol. 5, Birkhauser Boston, Inc., Boston, MA, 1999. MR1723736 [24] L. Hormander, Uniqueness theorems for second order elliptic differential equations, Comm. Partial Differential Equations 8 (1983), no. 1, 21 -- 64. MR686819 [25] N. Ichihara, The generalized principal eigenvalue for Hamilton-Jacobi-Bellman equations of ergodic type, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire 32 (2015), no. 3, 623 -- 650. MR3353703 [26] C. Kenig, L. Silvestre, and J.-N. Wang, On Landis' conjecture in the plane, Comm. Partial Differential Equations 40 (2015), no. 4, 766 -- 789. MR3299355 [27] C. E. Kenig, Some recent quantitative unique continuation theorems, S´eminaire: ´equations aux D´eriv´ees Partielles. 2005 -- 2006, 2006, pp. Exp. No. XX, 12. MR2276085 [28] V. A. Kondrat′ev and E. M. Landis, Qualitative properties of the solutions of a second-order nonlinear equation, Mat. Sb. (N.S.) 135(177) (1988), no. 3, 346 -- 360, 415. MR937645 [29] N. V. Krylov, Controlled diffusion processes, Applications of Mathematics, vol. 14, Springer- Verlag, New York-Berlin, 1980. Translated from the Russian by A. B. Aries. MR601776 [30] V. Z. Meshkov, On the possible rate of decrease at infinity of the solutions of second-order partial differential equations, Mat. Sb. 182 (1991), no. 3, 364 -- 383. MR1110071 [31] M. Murata, Structure of positive solutions to (−∆ + V )u = 0 in Rn, Duke Math. J. 53 (1986), no. 4, 869 -- 943. MR874676 [32] , Semismall perturbations in the Martin theory for elliptic equations, Israel J. Math. 102 (1997), 29 -- 60. MR1489100 [33] R. D. Nussbaum and Y. Pinchover, On variational principles for the generalized principal eigenvalue of second order elliptic operators and some applications, J. Anal. Math. 59 (1992), 161 -- 177. Festschrift on the occasion of the 70th birthday of Shmuel Agmon. MR1226957 [34] A. Persson, Bounds for the discrete part of the spectrum of a semi-bounded Schrodinger operator, Math. Scand. 8 (1960), 143 -- 153. MR0133586 [35] Y. Pinchover, On positive solutions of second-order elliptic equations, stability results, and classification, Duke Math. J. 57 (1988), no. 3, 955 -- 980. MR975130 [36] [37] [38] , Criticality and ground states for second-order elliptic equations, J. Differential Equa- tions 80 (1989), no. 2, 237 -- 250. MR1011149 , On criticality and ground states of second order elliptic equations. II, J. Differential Equations 87 (1990), no. 2, 353 -- 364. MR1072906 , A Liouville-type theorem for Schrodinger operators, Comm. Math. Phys. 272 (2007), no. 1, 75 -- 84. MR2291802 LIOUVILLE PROPERTIES OF EIGENFUNCTIONS OF ELLIPTIC OPERATORS 33 [39] Y. Pinchover, A. Tertikas, and K. Tintarev, A Liouville-type theorem for the p-Laplacian with potential term, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire 25 (2008), no. 2, 357 -- 368. MR2400106 [40] R. G. Pinsky, Positive harmonic functions and diffusion, Cambridge Studies in Advanced Mathematics, vol. 45, Cambridge University Press, Cambridge, 1995. MR1326606 [41] L. Rossi, The Landis conjecture with sharp rate of decay, ArXiv e-prints 1807.00341 (2018). [42] B. Simon, Large time behavior of the Lp norm of Schrodinger semigroups, J. Funct. Anal. 40 (1981), no. 1, 66 -- 83. MR607592 [43] , Schrodinger semigroups, Bull. Amer. Math. Soc. (N.S.) 7 (1982), no. 3, 447 -- 526. MR670130 Department of ECE, The University of Texas at Austin, 2501 Speedway, EER 7.824, Austin, TX 78712, USA E-mail address: [email protected] Department of Mathematics, Indian Institute of Science Education and Research, Dr. Homi Bhabha Road, Pune 411008, India E-mail address: [email protected] Department of Mathematics, Indian Institute of Science Education and Research, Dr. Homi Bhabha Road, Pashan, Pune 411008, India E-mail address: [email protected]
1908.09012	1	1908	2019-08-23T19:41:18	The H\'ajek-R\'enyi-Chow maximal inequality and a strong law of large numbers in Riesz spaces	[ "math.FA", "math.PR", "math.ST", "math.ST" ]	In this paper we generalize the H\'ajek-R\'enyi-Chow maximal inequality for submartingales to $L^p$ type Riesz spaces with conditional expectation operators. As applications we obtain a submartingale convergence theorem and a strong law of large numbers in Riesz spaces. Along the way we develop a Riesz space variant of the Clarkson's inequality for $1\le p\le 2$.	math.FA	math	9 The H´ajek-R´enyi-Chow maximal inequality and a strong 1 0 2 law of large numbers in Riesz spaces ∗ . A F h t a m [ 1 v 2 1 0 9 0 . 8 0 9 1 : v i X r a g u A 3 2 ] Wen-Chi Kuo†, David F. Rodda‡ & Bruce A. Watson § School of Mathematics University of the Witwatersrand Private Bag 3, P O WITS 2050, South Africa August 27, 2019 Abstract In this paper we generalize the H´ajek-R´enyi-Chow maximal inequality for submartingales to Lp type Riesz spaces with conditional expectation operators. As applications we obtain a submartingale convergence theorem and a strong law of large numbers in Riesz spaces. Along the way we develop a Riesz space variant of the Clarkson's inequality for 1 ≤ p ≤ 2. 1 Introduction In a Dedekind complete Riesz space E with weak order unit, say e, we say that T is a con- ditional expectation on E if T is a positive order continuous linear projection on E which maps weak order units to weak order units and has range, R(T ), a Dedekind complete Riesz subspace of E, see [13] for more details. It should be noted that the only conditional expecta- tion operator which is also a band projection is the identity map. A conditional expectation operator T is said to be strictly positive if T f = 0 implies that f = 0. Every Archimedean Riesz space E can be extended uniquely (up to Riesz isomorphism) to a universally complete space Eu, see [24]. It was shown in [13] that the domain of a strictly positive conditional ∗Keywords: Riesz spaces; vector lattices; maximal inequality; Clarkson's inequality; submartingale con- vergence; strong law of large numbers. Mathematics subject classification (2010): 46B40; 60F15; 60F25. †Supported in part by National Research Foundation of South Africa grant number CSUR160503163733. ‡Supported in part by National Research Foundation of South Africa grant number 110943. §Supported in part by the Centre for Applicable Analysis and Number Theory and by National Research Foundation of South Africa grant IFR170214222646 with grant no. 109289. 1 expectation operator T can be extended to its natural domain L1(T ) in Eu. In particular L1(T ) = dom(T ) − dom(T ) where f ∈ dom(T ) if f ∈ Eu +, the positive cone of Eu, and there is an upwards directed net fα in E+ with the net T fα order bounded in Eu and in this case the value assigned to T f is the order limit in Eu of the net T fα. Given that the above extensions can be made we will assume throughout that T is a conditional expecta- tion operator acting on L1(T ). The space Eu is an f -algebra with multiplication defined so that the chosen weak order unit, e, is the algebraic unit. Further, it was shown in [16] that R(T ) = {T f f ∈ L1(T )} is an f -algebra and L1(T ) is an R(T )-module with R(T )- valued norm kf k1,T := T f . The space L2(T ) was introducted in [19] and generalized to Lp(T ) = {x ∈ L1(T ) : xp ∈ L1(T )}, 1 < p < ∞, in [3] where functional calculus was used to define f (x) = xp for x ∈ Eu +. Much of the mathematical machinery needed to work in Lp(T ), 1 < p < ∞, was developed in [10] even though such spaces were not considered there. Again these spaces are R(T )-modules with associated R(T )-valued norms kf kp,T := (T f p)1/p. We note that in [16] the R(T )-module L∞(T ) = {x ∈ L1(T ) : x ≤ y for some y ∈ R(T )} was considered, with R(T )-valued norm kf k∞,T := inf{y ∈ R(T )+ : f ≤ y}. Here we have that L∞(T ) is an f -algebra order dense in L1(T ) and having L∞(T ) ⊂ Lp(T ) ⊂ L1(T ) for all 1 < p < ∞. Regarding Lp type spaces we also note the work of Boccuto, Candeloro and Sambucini in [1]. In Section 3, we generalize Clarkson's inequality to Lp(T ), 1 ≤ p ≤ 2. A filtration on a Dedekind complete Riesz space E with weak order unit is a family of con- ditional expectation operators (Ti)i∈N defined on E having TiTj = TjTi = Ti for all i < j. We say that a sequence of elements (fi)i∈N in a Riesz space is adapted to a filtration (Ti)i∈N if fi ∈ R(Ti) for all i ∈ N. A sequence (fi)i∈N is said to be predictable if fi ∈ R(Ti−1) for each i ∈ N. A Riesz space (sub, super) martingale is a double sequence (fi, Ti)i∈N with (fi)i∈N adapted to the filtration (Ti)i∈N and Tifj(≥, ≤) = fi for i < j. The fundamentals of such processes can be found in [12, 13, 14] as well as their continuous time versions in [8, 9]. In Section 4, we give the H´ajek-R´enyi-Chow maximal inequality for Riesz space sub- martingales, see [4, Theorem 1] and [6, Proposition (6.1.4)] for measure theoretic versions. The H´ajek-R´enyi-Chow maximal inequality for submartingales has as a special case Doob's maximal inequality. We note that maximal inequalities have been obtained for Riesz space positive supermartingales in [11, Lemma 3.1] and for Riesz space quasi-martingales in [22, Theorem 6.2.10]. The H´ajek-R´enyi-Chow maximal inequality for submartingales is applied, in Theorem 5.1, to non-negative submartingales to obtain weighted convergence, via a proof which does not use of upcrossing. We note that this theorem can be deduced directly from [14, Theorem 3.5], which is, however, based on the Riesz space upcrossing theorem. For E a Dedekind complete Riesz space with weak order unit, e, and (Bn) an increasing sequence of bands in E, with associated band projections (Pn), it was proved in [21] that xn/bn → 0, in order, as n → ∞ if xn ∈ Bn with Pnxn+1 = xn and xn+1 − xn ≤ cne, for all n ∈ N. Here cn > 0 and 0 < bn ↑ ∞, for all n ∈ N, with → 0 as n → ∞. In Section 5, we conclude by giving Chow's strong law of large numbers in Lp(T ), 1 < p < ∞, see [4, 5] and 1 bn n Xi=1 c2 i!1/2 2 [6, Theorems 6.1.8 and 6.1.9] for measure versions. We note that, for Riesz space processes, a strong law of large numbers for ergodic processes was given in [15], a weak law of large number for mixingales in [17] and Bernoulli's law of large numbers in [18]. 2 Weighted Ces`aro means In this section we give a version of Kronecker's Lemma for weighted Ces`aro means in an Archimedean Riesz space. Lemma 2.1 Let E be an Archimedean Riesz space and (sn) be a sequence in E+ order con- vergent to 0. If bn is a non-decreasing sequence of non-negative real numbers divergent to +∞, then 1 i=1 (bi+1 − bi)si converges to zero in order as n → ∞. bn Pn−1 Proof: By the order convergence of (sn) to 0, there is sequence (vn) in E such that sn ≤ vn ↓ 0, for n ∈ N. As 0 ≤ 1 bn (bi+1 − bi)si ≤ 1 bn n−1 Xi=1 n−1 Xi=1 (bi+1 − bi)vi =: zn, it suffices to show that zn → 0 in order. For n ∈ N, let Nn := max{j ∈ N j2bj ≤ bn}, then (Nn) is a non-decreasing sequence in N with Nn → ∞ as n → ∞ and Nn < n for n ≥ 2. Now, for n ≥ 2, zn = 1 bn bnvn−1 − b1v1 + ≤ vn−1 + ≤ vn−1 + 1 bn Nn Xi=2 NnbNn bn bNnv1 + bn−1 bn Nn n−1 bi(vi−1 − vi) + Xi=2 Xi=Nn+1 n−1 Xi=Nn+1 bn−1(vi−1 − vi)! v1 + vNn ≤ vn−1 + 1 Nn v1 + vNn ↓ 0. bi(vi−1 − vi)! In [7, lemma 3.14], this result was proved for the case of bn = n, n ∈ N. Lemma 2.2 (Kronecker's Lemma) Let (xn) be a summable sequence of elements in an Archimedean Riesz space E. Let (bn)n∈N be a non-decreasing sequence of non-negative real numbers divergent to +∞. Then 1 bn n bixi → 0 in order as n → ∞. Xi=1 3 ∞ 1 Xi=n+1 bn (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) bixi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 bn n Xi=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) which converges to zero in order by Lemma 2.1. Proof: Let sn := xi, n = 0, 1, 2, . . ., then sn → 0 in order and = bnsn − b1s0 − n−1 Xi=1 (bi+1 − bi)si(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ sn + b1 bn s0 + 1 bn n−1 (bi+1 − bi)si Xi=1 3 Inequalities The inequalities presented in this section form the foundation on which much of the rest of this paper is based. Taking the product Riesz space K = [L1(T )]n with componentwise ordering and defining F(xi)n i=1 we have that F is a conditional expectation operator on K. Hence from [10, Corollary 6.4] or [3, Theorem 3.7] we have the following theorem. i=1 = ( 1 j=1 xj)n nPn Theorem 3.1 (Holder's inequality for sums) Let T be a conditional expectation with nat- ural domain L1(T ) and 1 ≤ p, q ≤ ∞ with 1 q = 1 and p = 1 for q = ∞. Let n ∈ N. If xi ∈ Lp(T ) and yi ∈ Lq(T ) for all i ∈ {1, 2, ..., n}, then xiyi ∈ L1(T ) for each i, and p + 1 n Xi=1 T xiyi ≤ n Xi=1 T xip! 1 p n Xi=1 1 q . T yiq! From [3, page 809] we have that x + yp + x − yp ≤ 2p(xp + yp) for 1 < p < ∞ with x, y ∈ Eu. This inequality, however, is inadequate for our purposes and we require a refined version, i.e. the Clarkson's inequalities for 1 < p < 2. To this end we follow the approach of Ramaswamy [20]. Theorem 3.2 (Clarkson's Inequality) Let E be a Dedekind complete Riesz space with weak order unit, say e which we take as the multiplicative unit in the f -algebra Eu. For x, y ∈ Eu and 1 ≤ p ≤ 2 we have x + yp + x − yp ≤ 2(xp + yp). (3.1) 4 Proof: For p = 1 the result follows from the triangle inequality while for p = 2 the result follows from f 2 = f 2, so we now consider only 1 < p < 2. Taking g(X) = X 2/p and F(X, Y ) = ( 1 2(X +Y )) in Jensen's inequality of [10] on the Riesz space F := Eu ×Eu with componentwise ordering, we have that F(g(ap, bp)) ≥ g(F(ap, bp)) so a2 + b2 ≥ 2(p−2)/p(ap + bp)2/p and 2(2−p)/2(a2 + b2)p/2 ≥ ap + bp. Setting a = x + y and b = x − y we have 2(X +Y ), 1 x + yp + x − yp ≤ 2(2−p)/2((x + y)2 + (x − y)2)p/2 = 2(x2 + y2)p/2. (3.2) We now apply the ∞ case of Holder's inequality of [10] on the space F with conditional expectation F as above to get F((xp, yp)(x2−p, y2−p)) ≤ F((xp, yp))(x2−p ∨ y2−p, x2−p ∨ y2−p). (3.3) Here x2−p ∨ y2−p = (xp ∨ yp)(2−p)/p ≤ (xp + yp)(2−p)/p by the commutation of multipli- cation and band projections in the f -algebra Eu. Hence from (3.3) we get 1 2 (x2 + y2) ≤ 1 2 (xp + yp)(xp + yp)(2−p)/p = 1 2 (xp + yp)2/p, which when combined with (3.2) gives (3.1). The strong law of large numbers for p > 2 will make use of Riesz space versions of Burkholder's inequality, [2, Theorem 16], which we give here for completeness. Theorem 3.3 (Burkholder's inequality) For 1 < p < ∞, there are constants cp, Cp > 0 such that 1 2 CpT Xnp ≤ T S n p ≤ cpT Xnp, for each (Xn, Tn)n∈N a martingale in Lp(T ) compatible with T , i.e. T Tn = T = TnT , for all n ∈ N. Here Sn := Xi=1 (Xi − Xi−1)2 and X0 := 0. 4 H´ajek-R´enyi-Chow maximal inequality We now recall some well known results regarding band projections on a Dedekind complete Riesz space, E, with a weak order unit, say e. If g ∈ E+ we denote the band projection onto the band generated by g by Pg. In this setting every band is a principal band and if B is a band in E with band projection Q onto E then a generator of the band is Qe. Moreover for f ∈ E+ we have Pgf = sup (f ∧ (ng)), see [23, Theorem 11.5]. Further if (fn) is a sequence in n∈N E+ then ∞ _n=1 Pfn = PW∞ n=1 fn 5 (4.1) since n=1 fne = PW∞ ∞ _m=1 e ∧ m ∞ _n=1 fn!! = _m,n=1 ∞ ∞ (e ∧ mfn) = Pfne. _n=1 We note however that for the case of infima only the following inequality can be assured For reference we note that if (gn) is a sequence in E then 0 ≤ PV∞ n=1 g− n ∞ ^n=1 Pfn ≥ PV∞ n=1 fn. (4.2) g+ m ≤ Pg− m g+ m = 0 giving (I − PV∞ n=1 g− n ) Hence )g+ m = n=1 g− n g+ m. ∞ _m=1 ∞ _m=1 PW∞ g+ m = n=1 g+ n ∞ (I − PV∞ _m=1 ≤ I − PV∞ n=1 g− n . (4.3) Using telescoping series we generalise [6, Lemma (6.1.1)] to vector lattices. Lemma 4.1 Let E be a Dedekind complete Riesz space weak order unit e. Let (Xi) ⊂ E be a sequence in E and g ∈ E. Let Pi := P(g−Xi)+, i ∈ N, be the band projection onto the band generated by (g − Xi)+, then n−1 (I − Qn)g ≤ X1 + [Qi(Xi+1 − Xi)] − QnXn, (4.4) Xi=1 j=1 Xj)+, n ∈ N. where Qn := Pj = P n Yj=1 (g−Wn Proof: Let Q0 := I. From the definition of Pi we have Pj(g − Xj) = (g − Xj)+ and thus (I − Pj)(g − Xj) = −(g − Xj)− ≤ 0. However Qj−1 − Qj = Qj−1(I − Pj), so applying Qj−1 to both sides of (I − Pj)(g − Xj) ≤ 0, gives (Qj−1 − Qj)(g − Xj) ≤ 0. Hence (Qj−1 − Qj)g ≤ (Qj−1 − Qj)Xj, which when summed over j = 1, . . . , n gives (4.4). If (fi, Ti) is a submartingale in the Riesz space E then so is (f + that as Tj a positive operator and f + for i ≤ j. Hence Tif + j ≥ fj so Tif + for i ≤ j. j ≥ 0 ∨ fi = f + i j ≥ Tifj ≥ fi and as f + i , Ti). To see this we observe j ≥ 0, j ≥ 0 so Tif + 6 Theorem 4.2 (H´ajek-R´enyi-Chow maximal inequality) Let (Yi, Ti)i∈N be a submartin- gale in L1(T ). For (ai)i∈N a non-decreasing sequence of positive real numbers and g ∈ R(T1)+ we have T1(I − Un)g ≤ Y + 1 a1 + where Un :=Qn i=1 P Yi ai(cid:17)+ = P (cid:16)g− (cid:16)g−Wn i=1 Yi ai(cid:17)+. n−1 Xi=1 T1(cid:20)Y + i+1 − Y + i ai+1 (cid:21) (4.5) Proof: Let Q = Pg be the band projection onto the band generated by g. Now as g ∈ R(T1)+ it follows that Q and T1 commute, see [13, Theorem 3.2]. As (Y + , Ti) is a submartingale, for i i ≤ j, TiY + , hence i = TiY + j ≥ Y + i Ti(Y + j+1 − Y + j ) ≥ 0, and thus (I − Q)T1(I − Un)g = T1(I − Un)(I − Q)g = 0 ≤ (I − Q) Y + 1 a1 + n−1 Xi=1 T1(cid:20) Y + i+1 − Y + i ai+1 Letting Xi = Y + i /ai, i ∈ N, in Lemma 4.1 we have, for n ∈ N, (I − Qn)g ≤ Y + 1 a1 + n−1 Xi=1 Qi(cid:18) Y + i+1 ai+1 − Y + i ai (cid:19) − Qn Y + n an ≤ Y + 1 a1 + n−1 Xi=1 Qi(cid:18)Y + i+1 − Y + i ai+1 (4.6) (cid:21)! . (4.7) (cid:19) (4.8) where Qi = P Ti(Y + i+1 − Y + j=1 Xj)+. Here 0 ≤ Qi ≤ I and Ti(Y + i ). Hence, as T1 = T1Ti, from (4.6) and (4.8), (g−Wi i+1 − Y + i ) ≥ 0 so QiTi(Y + i+1 − Y + i ) ≤ T1(I − Qn)g ≤ Y + 1 a1 + n−1 Xi=1 Since g ≥ 0, we have T1(cid:18)Y + i+1 − Y + i ai+1 (cid:19) . (4.9) Yi g ∧(cid:18)(cid:18)g − ai(cid:19) ∨ 0(cid:19) =(cid:18)g ∧(cid:18)g − ai(cid:17)+ ai (cid:17)+ =(cid:16)g − Y + giving g ∧(cid:16)g − Yi QT1(I − Un)g ≤ Q Y + that T1Q = QT1 we have 1 a1 i Combining (4.7) and (4.10) gives (4.5). Yi ai(cid:19)(cid:19) ∨ (g ∧ 0) =(cid:18)g −(cid:18)0 ∨ Yi ai(cid:19)(cid:19) ∨ 0, , thus QUn = QQn. Now, applying Q to (4.9) and noting T1(cid:18)Y + i+1 − Y + i ai+1 (cid:19)! . (4.10) + n−1 Xi=1 7 5 Submartingale convergence As an application of the H´ajek-R´enyi-Chow Maximal Inequality we give a weighted conver- gence theorem for submartingales, with a proof that is independent of upcrossing. Theorem 5.1 (Submartingale convergence) Let p ≥ 1 and (Xi, Ti)i∈N be a non-negative submartingale in Lp(T ). Let (ai)i∈N be a positive, non-decreasing, sequence of real numbers diverging to ∞. If ∞ Xi=1 T1(cid:18)X p i+1 − X p ap i+1 i (cid:19) (5.1) converges in order, then Xn an tends to zero in order as n tends to ∞. i )/ap Proof: By [10, Corollary 4.5], (X p X p i+1 has T1Zi ≥ 0 and by assumption P∞ m m+1 T1X p ap m+1 = 1 ap m+1 T1 X p 1 + (X p i+1 − X p Xi=1 in order as m → ∞. n, Tn) is a non-negative submartingale so Zi := (X p i+1 − i=1 T1Zi is order convergent, so by Lemma 2.2 i )! = X p 1 ap m+1 + 1 ap m+1 m Xi=1 ap i+1T1Zi → 0 (5.2) By (4.3) and Theorem 4.2 applied to (X p have i )n i=m, with g = te, t ∈ R where t > 0, for n > m, we T1P Wn i=m(cid:18) X a p i p i −te(cid:19)+!te ≤ T1 I − P i=m(cid:18) X Vn a p i p i −te(cid:19)−! te ≤ X p m ap m + n−1 Xi=m T1Zi. (5.3) Applying T1 to (5.3) and taking the order limit as n → ∞, by (4.1) we have 0 ≤ tT1 ∞ _i=m P p i p i (cid:18) X a −te(cid:19)+! e ≤ T1(cid:20) X p m ap m(cid:21) + T1Zi. ∞ Xi=m Taking the order limit as m → ∞ of (5.4), by (5.2), we have 0 ≤ tT1 lim m→∞ ∞ _i=m −te(cid:19)+e = 0 and by the strict positivity of T1,Vm∈NW∞ (cid:18) X 0. Now, by (4.1) and (4.2), Hence T1Vm∈NW∞ −te(cid:19)+! e ≤ lim m(cid:21) + lim T1(cid:20) X p m ap Xi=m i=m P (cid:18) X m→∞ m→∞ p i p i p i p i P ∞ a a T1Zi = 0. i=m P (5.4) (5.5) −te(cid:19)+e = p i p i (cid:18) X a 0 ≤ P lim supi→∞(cid:18) X a p i p i −te(cid:19)+e ≤ ^m∈N P p i p i a i=m(cid:18) X W∞ 8 ∞ −te(cid:19)+e = ^m∈N _i=m P p i p i (cid:18) X a −te(cid:19)+e = 0 and so 0 ≤ lim inf i→∞ X p i ap i ≤ lim sup i→∞ X p i ap i ≤ te for all t > 0. Thus X p i ap i → 0 in order as i → ∞. Remark Since (X p n, Tn) is a non-negative submartingale, by [10, Corollary 4.5], taking Yj+1 = i=1 Zi in the above theorem, we have that (Yj, Tj) is a T1-bounded submartingale and Theorem 5.1 follows directly from [14, Theorem 3.5]. Pj 6 Chow's strong laws of large numbers We recall that (Yi, Ti) is a martingale difference sequence if (Ti) is a filtration, (Yi) is adapted to (Ti) and Ti−1Yi = 0 for i ≥ 2. In Theorem 6.1, for 1 ≤ p ≤ 2, and Corollary 6.2 and Theorem 6.3, for p > 2, Chow's strong law of large numbers is extended to martingale difference sequences in Riesz spaces. Theorem 6.1 Let 1 ≤ p ≤ 2, and (Yn, Tn)n∈N be a martingale difference sequence in Lp(T ). Let (ai)i∈N be a positive, non-decreasing sequence of real numbers divergent to infinity with ∞ Xi=1 T1(cid:18)Yip i (cid:19) ap (6.1) Yi → 0, in order, as n tends to infinity. order convergent, then 1 an n Xi=1 n Proof: Let Xn = Yi then Xi + Yi+1 = Xi+1 and Xi − Yi+1 = 2Xi − Xi+1 so Theorem 3.2 can be applied to give Xi=1 Xi+1p + 2Xi − Xi+1p ≤ 2(Xip + Yi+1p). (6.2) Now as (Xn, Tn) is a martingale, so Ti(2Xi − Xi+1) = Xi and by functional calculus, see [10], Xip ∈ R(Ti) giving TiXip = Xip, hence TiXip = Ti(2Xi − Xi+1)p ≤ Ti2Xi − Xi+1p (6.3) where the final inequality follows from Jensen's inequality, [10, Theorem 4.4]. Combining (6.2) and (6.3) we have TiXi+1p − TiXip ≤ 2TiYi+1p. (6.4) 9 By [10, Corollary 4.5], (Xi, Ti) and (Xip, Ti) are submartingales so 0 ≤ TiXi+1p − TiXip ap i+1 ≤ 2 TiYi+1p ap i+1 (6.5) which, with (6.1), yields that follows from Theorem 5.1. TiXi+1p − TiXip ap i+1 ∞ Xi=1 is order convergent. The theorem now We can now bootstrap on Theorem 6.1 to obtain a strong law for p > 2. Corollary 6.2 Let p > 2 and (Yn, Tn)n∈N be a martingale difference sequence in Lp(T1). Let (ai)i∈N be a positive, non-decreasing sequence of real numbers with and ∞ Xi=1 T1(cid:18)Yip aγ i (cid:19) order convergent, where p ≥ γ + ( p 2 − 1)k, then as n tends to infinity. ∞ Xi=1 1 an 1 ak i n Xi=1 convergent in R, Yi → 0, in order, Proof: From Holder's inequality, Theorem 3.1, for n > m we have T1 Yi2 a2 i n Xi=m ≤ n Xi=m T1 Yip aγ i ! 2 p n Xi=m 2 p i!1− e aδ where δ = p−γ −1 ≥ k ensuring that p 2 1 aδ i ∞ Xi=1 converges. Hence from Theorem 6.1 with p = 2, Yi → 0, in order, as n tends to infinity. From Corollary 6.2, if p > 2 and (Yn, Tn)n∈N is a martingale difference sequence in Lp(T1) with T1(cid:18) Yip i1+ p 2 −δ(cid:19) order convergent for some δ > 0 then 1 n n Xi=1 infinity. For this special case, of ai = i, a more precise result can be given, as per [5, 4]. Yi → 0, in order, as n tends to 1 an n Xi=1 ∞ Xi=1 Theorem 6.3 Let p > 2 be a fixed number and let (Yn, Tn)n∈N be a martingale difference sequence in Lp(T1). If ∞ Xi=1 T1(cid:18)Yip i1+ p 2(cid:19) converges in order then 1 n 10 n Yi → 0 in order as n → ∞. Xi=1 Proof: Let Xn =: as n → ∞ of n Xi=1 Yi for n ∈ N, then, from Theorem 5.1, it suffices to prove the convergence Zn = n Xi=2 T1(cid:18)Xip − Xi−1p ip (cid:19) = n−1 Xi=2 (cid:18) 1 ip − 1 (i + 1)p(cid:19) T1(Xip) + T1(Xnp) np − T1(X1p) 2p . Since each term in the above summations is non-negative we need only show the boundedness of Zn, n ∈ N. From Burkholder inequality, Theorem 3.3, there is Cp > 0 so that CpT1Xnp ≤ T1 n Xi=1 Yi2!p/2 , for all n ∈ N. Applying Jensen's inequality of [10] we have T1 n Xi=1 Yi2!p/2 ≤ n Xi=1 T1Yi2!p/2 . Now Holder inequality, Theorem 3.1, gives T1Yi2!p/2 n Xi=1 p 2 −1 ≤ n n Xi=1 T1Yip. Combining (6.6), (6.7) and (6.8) gives T1Xnp np ≤ 1 p 2 +1 Cpn T1Yip. n Xi=1 (6.6) (6.7) (6.8) (6.9) From Kronecker's Lemma, Theorem 2.2, we have that 1 p 2 +1 n T1Yip → 0 in order as n → ∞. Thus the left hand side of (6.9) is order bounded by say h ∈ L1(T1) and n Xi=1 n−1 Xi=2 (cid:18) 1 ip − 1 (i + 1)p(cid:19) T1(Xip) ≤ p 1 ip/2 h, ∞ Xi=1 giving ∞ 1 ip/2 h + h − T1(X1p) 2p . Zn ≤ p Xi=1 Here we have used that n−p − (n + 1)−p ≤ pn−p−1, n ∈ N. 11 References [1] A. Boccuto, D. Candeloro, A. Sambucini, Lp spaces in vector lattices and appli- cations, Mathematica Slovaca, 67 (2017), 1409-1426. [2] Y. Azouzi, K. Ramdane, Burkholder Inequalities in Riesz spaces, Indag. Math., 28 (2017), 1076-1094. [3] Y. Azouzi, M. Trabelsi, Lp-spaces with respect to conditional expectation on Riesz spaces, J. Math. Anal. Appl., 447 (2017), 798-816. [4] Y. S. Chow, A martingale inequality and the law of large numbers, Proc. Amer. Math. Soc., 11 (1960), 107-111. [5] Y. S. Chow, On a strong law of large numbers, Ann. Math. Statist., 38 (1967), 610-611. [6] G. A. Edgar, L. Sucheston, Stopping times and directed processes, Cambridge Univ. Press, 1992. [7] N. Gao, V. G. Troitsky, and F. Xanthos, UO-Convergence and its applications to Ces`aro means in Banach lattices, Israel Journal of Mathematics, 220 (2017), 649-689. [8] J. J. Grobler, Continuous stochastic processes in Riesz spaces: the Doob-Meyer de- composition, Positivity, 14 (2010) 731-751. [9] J. J. Grobler, Doob's optional sampling theorem in Riesz spaces, Positivity, 15 (2011) 617-637. [10] J. J. Grobler, Jensen's and martingale inequalities in Riesz spaces, Indag. Math., N.S., 25 (2014), 275-295. [11] J. J. Grobler, C. Labuschagne, V. Marraffa, Quadratic variation of martingales in Riesz spaces, J. Math. Anal. Appl., 410 (2014), 418-426. [12] W.-C. Kuo, C. C. A. Labuschagne, B. A. Watson, Discrete time stochastic pro- cesses on Riesz spaces, Indag. Math., N.S., 15 (2004), 435-451. [13] W.-C. Kuo, C. C. A. Labuschagne, B. A. Watson, Conditional expectations on Riesz spaces, J. Math. Anal. Appl., 303 (2005), 509-521. [14] W.-C. Kuo, C. C. A. Labuschagne, B. A. Watson, Convergence of Riesz Space Martingales, Indag. Math., 17 (2006), 271-283. [15] W.-C. Kuo, C. C. A. Labuschagne, B. A. Watson, Ergodic Theory and the Law of Large Numbers on Riesz Spaces, J. Math. Anal. Appl., 325 (2007), 422-437. 12 [16] W.-C Kuo, M.J. Rogans, B. A. Watson, Mixing Inequalities in Riesz spaces, J. Math. Anal. Appl., 456 (2017), 992-1004. [17] W.-C Kuo, J. J. Vardy, B. A. Watson, Mixingales on Riesz spaces, J. Math. Anal. Appl., 402 (2013), 731-738. [18] W.-C Kuo, J. J. Vardy, B. A. Watson, Bernoulli processes in Riesz spaces, Ordered Structures and Applications: Positivity VII Trends in Mathematics, 263-274. [19] C. C. A. Labuschagne, B. A. Watson, Discrete stochastic integrals in Riesz Spaces, Positivity, 14 (2010), 859-875. [20] S. Ramaswamy, A simple proof of Clarkson's inequality, Proc. Amer. Math. Soc., 68 (1978), 249-250. [21] G. Stoica, Limit laws for martingales in vector lattices, J. Math. Anal. Appl., 476 (2019), 715-719. [22] J. J. Vardy, Markov Processes and Martingale Generalisations on Riesz Spaces, PhD Thesis, University of the Witwatersrand, 2013. [23] A. C. Zaanen, Introduction to operator theory in Riesz spaces, Springer-Verlag, Berlin and Heidelberg, 1991. [24] A. C. Zaanen, Riesz Spaces II, North-Holland, Amsterdam, New York, 1983. 13
1303.6078	1	1303	2013-03-25T10:45:42	The Bishop-Phelps-Bollob\'{a}s theorem for operators on $L_1(\mu)$	[ "math.FA" ]	In this paper we show that the Bishop-Phelps-Bollob\'as theorem holds for $\mathcal{L}(L_1(\mu), L_1(\nu))$ for all measures $\mu$ and $\nu$ and also holds for $\mathcal{L}(L_1(\mu),L_\infty(\nu))$ for every arbitrary measure $\mu$ and every localizable measure $\nu$. Finally, we show that the Bishop-Phelps-Bollob\'as theorem holds for two classes of bounded linear operators from a real $L_1(\mu)$ into a real $C(K)$ if $\mu$ is a finite measure and $K$ is a compact Hausdorff space. In particular, one of the classes includes all Bochner representable operators and all weakly compact operators.	math.FA	math	THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON L1(µ) YUN SUNG CHOI, SUN KWANG KIM, HAN JU LEE, AND MIGUEL MART´IN Dedicated to the memory of Robert R. Phelps Abstract. In this paper we show that the Bishop-Phelps-Bollob´as theorem holds for L(L1(µ), L1(ν)) for all measures µ and ν and also holds for L(L1(µ), L∞(ν)) for every arbitrary measure µ and every localizable measure ν. Finally, we show that the Bishop-Phelps-Bollob´as theorem holds for two classes of bounded linear operators from a real L1(µ) into a real C(K) if µ is a finite measure and K is a compact Hausdorff space. In particular, one of the classes includes all Bochner representable operators and all weakly compact operators. 1. Introduction The celebrated Bishop-Phelps theorem of 1961 [12] states that for a Banach space X, every element in its dual space X∗ can be approximated by ones that attain their norms. Since then, there has been an extensive research to extend this result to bounded linear operators between Banach spaces [14, 27, 33, 34, 36] and non-linear mappings [2, 7, 11, 16, 17, 30]. On the other hand, Bollob´as [13], motivated by problems arising in the theory of numerical ranges, sharpened the Bishop-Phelps theorem in 1970, and got what is nowadays called the Bishop-Phelps-Bollob´as theorem. Previous to presenting this result, let us introduce some notations. Given a (real or complex) Banach space X, we write BX for the unit ball, SX for its unit sphere, and X∗ for the topological dual space of X. If Y is another Banach space, we write L(X, Y ) to denote the space of all bounded linear operators from X into Y . Theorem 1.1 (Bishop-Phelps-Bollob´as theorem). Let X be a Banach space. If x ∈ SX and x∗ ∈ SX∗ satisfy x∗(x) − 1 < ε2/4, then there exist y ∈ SX and y∗ ∈ SX∗ such that y∗(y) = 1, (cid:107)x∗ − y∗(cid:107) < ε and (cid:107)x − y(cid:107) < ε. In 2008, Acosta, Aron, Garc´ıa and Maestre [3] introduced the Bishop-Phelps-Bollob´as property to study extensions of the theorem above to operators between Banach spaces. Definition 1.2. Let X and Y be Banach spaces. The pair (X, Y ) is said to have the Bishop-Phelps- Bollob´as property (BPBp) if for every 0 < ε < 1, there is η(ε) > 0 such that for every T ∈ L(X, Y ) with (cid:107)T(cid:107) = 1 and x0 ∈ SX satisfying (cid:107)T (x0)(cid:107) > 1 − η(ε), there exist y0 ∈ SX and S ∈ L(X, Y ) with (cid:107)S(cid:107) = 1 satisfying the following conditions: In this case, we also say that the Bishop-Phelps-Bollob´as theorem holds for L(X, Y ). (cid:107)Sy0(cid:107) = 1, (cid:107)y0 − x0(cid:107) < ε, and (cid:107)S − T(cid:107) < ε. Date: March 25th, 2013. 2010 Mathematics Subject Classification. Primary 46B20; Secondary 46B04, 46B22. Key words and phrases. Banach space, approximation, norm-attaining operators, Bishop-Phelps-Bollob´as theorem. First named author partially supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (No. 2010-0008543), and also by Priority Research Centers Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (MEST) (No. 2012047640). Third named author partially supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology (2012R1A1A1006869). Fourth named author partially supported by Spanish MINECO and FEDER project no. MTM2012-31755, Junta de Andaluc´ıa and FEDER grants FQM-185 and P09-FQM-4911, and by CEI-Granada 2009. 1 2 CHOI, KIM, LEE, AND MART´IN This property has been studied by many authors. See for instance [5, 8, 10, 15, 18, 19, 28, 29]. Observe that the BPBp of a pair (X, Y ) implies obviously that the set of norm attaining operators is dense in L(X, Y ). However, its converse is false, as shown by the pair (X, Y ) where X is the 2-dimensional L1- space and Y is a strictly, but not uniformly convex space (see [3] or [10]). Let us also comment that the Bishop-Phelps-Bollob´as theorem states that the pair (X, K) has the Bishop-Phelps-Bollob´as property for every Banach space X (K is the base scalar field R or C). In this paper we first deal with the problem of when the pair (Lp(µ), Lq(ν)) has the BPBp. Let us start with a presentation of both already known results and our new results. Iwanik [25] showed in 1979 that the set of norm-attaining operators from L1(µ) to L1(ν) is dense in the space L(L1(µ), L1(ν)) for arbitrary measures µ and ν. Our first main result in this paper is that the pair L(L1(µ), L1(ν)) has the BPBp. This is the content of section 3. On the other hand, Aron et al. [9] showed that if µ is a σ-finite measure, then the pair (L1(µ), L∞[0, 1]) has the BPBp, improving a result of Finet and Pay´a [24] about the denseness of norm-attaining operators. We generalize this result in section 4 showing that (L1(µ), L∞(ν)) has the BPBp for every measure µ and every localizable measure ν. This is also a strengthening of a result of Pay´a and Saleh [35] which stated only the denseness of norm-attaining operators. One of the tools used to prove the results above is the fact that one can reduce the proofs to some particular measures. We develop this idea in section 2, where, as its first easy application, we extend to arbitrary measures µ the result in [18] that (L1(µ), Lp(ν)) has the BPBp for σ-finite measures µ. The following result summarizes all what is known about the BPBp for the pair (Lp(µ), Lq(ν)). Corollary 1.3. The pair (Lp(µ), Lq(ν)) has the BPBp (1) for all measures µ and ν if p = 1 and 1 (cid:54) q < ∞. (2) for any measure µ and any localizable measure ν if p = 1, q = ∞. (3) for all measures µ and ν if 1 < p < ∞ and 1 (cid:54) q (cid:54) ∞. (4) for all measures µ and ν if p = ∞, q = ∞, in the real case. (1) and (2) follows from the results of this paper (Corollary 2.3, Theorem 3.1 and Theorem 4.1). Since Lp(µ) is uniformly convex when 1 < p < ∞, (3) follows from [5, 29] in the σ-finite case, generalized here to arbitrary measures µ (Corollary 2.3). Finally, (4) follows from [4], because every L∞ space is isometrically isomorphic to a C(K) space. As far as we know, the cases (L∞(µ), Lq(ν)) for 1 (cid:54) q < ∞ and the complex case of (4) remain open. Let µ be a finite measure. Since any L∞ space is isometrically isomorphic to C(K) for some compact Hausdorff space K, it is natural to ask when (L1(µ), C(K)) has the BPBp. Schachermayer [37] showed that the set of all norm-attaining operators is not dense in L(L1[0, 1], C[0, 1]). Hence, (L1[0, 1], C[0, 1]) cannot have the BPBp. On the other hand, Johnson and Wolfe [27] proved that if X is a Banach space and if either Y or Y ∗ is a L1(µ) space, then every compact operator from X into Y can be approximated by norm-attaining finite-rank operators. They also showed that every weakly compact operator from L1(µ) into C(K) can be approximated by norm-attaining weakly compact ones. In this direction, Acosta et al. have shown that (L1(µ), Y ) has the BPBp for representable operators (in particular, for weakly compact operators) if ((cid:96)1, Y ) has the BPBp, and this is the case of Y = C(K) [6]. On the other hand, Iwanik [26] studied two classes of bounded linear operators from a real L1(µ) space to a real C(K) space such that every element of each class can be approximated by norm-attaining elements, and showed that one of the classes strictly contains all Bochner representable operators and all weakly compact operators. In section 5, we deal with Bishop-Phelps-Bollob´as versions of these Iwanik's results. In particular, we show that for every 0 < ε < 1, there is η(ε) > 0 such that if T ∈ L(L1(µ), C(K)) with (cid:107)T(cid:107) = 1 is Bochner representable (resp. weakly compact) and f0 ∈ SL1(µ) satisfy (cid:107)T f0(cid:107) > 1 − η(ε), then there is a Bochner representable (resp. weakly compact) operator S ∈ L(L1(µ), C(K)) and f ∈ SL1(µ) such that (cid:107)Sf(cid:107) = (cid:107)S(cid:107) = 1, (cid:107)S − T(cid:107) < ε and (cid:107)f − f0(cid:107) < ε. THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 3 Let us finally comment that the proofs presented in sections 3 and 4 are written for the complex case. Their corresponding proofs for the real case are easily obtained, even easier, from the ones presented there. 2. Some preliminary results of characteristic functions of measurable subsets with finite positive measure. Let E = (cid:83) We start with some terminologies and known facts about L1(µ). Suppose that (Ω, Σ, µ) is an arbitrary measure space and put X = L1(µ). Suppose G is a countable subset of X. Since the closed linear span [G] of G is separable, we may assume that [G] is the closed linear span of a countable set {χEn} n En and Z = {f χE : f ∈ X}. Then, Z = L1(µE), where µE is restriction of the measure µ to the σ- algebra ΣE = {E ∩ A : A ∈ Σ}. Since µE is σ-finite, Z is isometrically (lattice) isomorphic to L1(m) for some positive finite Borel regular measure m defined on a compact Hausdorff space by the Kakutani representation theorem (see [32] for a reference). This space Z is called the band generated by G, and the canonical band projection P : X −→ Z, defined by P (f ) := f χE for f ∈ X, satisfies (cid:107)f(cid:107) = (cid:107)P f(cid:107) + (cid:107)(Id−P )f(cid:107) for all f ∈ X. For more details, we refer the reader to the classical books [32, 38]. and define, for 0 < ε < 1, Next, we present the following equivalent formulation of the BPBp from [10] which helps to better understand the property and will be useful for our preliminary results. Given a pair (X, Y ) of Banach spaces, let Π(X, Y ) = {(x, T ) ∈ X × L(X, Y ) : (cid:107)T(cid:107) = (cid:107)x(cid:107) = (cid:107)T x(cid:107) = 1} η(X, Y )(ε) = inf(cid:8)1 − (cid:107)T x(cid:107) : x ∈ SX , T ∈ L(X, Y ), (cid:107)T(cid:107) = 1, dist(cid:0)(x, T ), Π(X, Y )(cid:1) (cid:62) ε(cid:9), where dist(cid:0)(x, T ), Π(X, Y )(cid:1) = inf(cid:8)max{(cid:107)x − y(cid:107),(cid:107)T − S(cid:107)} : (y, S) ∈ Π(X, Y )(cid:9). Equivalently, for every ε ∈ (0, 1), η(X, Y )(ε) is the supremum of those ξ (cid:62) 0 such that whenever T ∈ L(X, Y ) with (cid:107)T(cid:107) = 1 and x ∈ SX satisfy (cid:107)T x(cid:107) (cid:62) 1 − ξ, then there exists (y, S) ∈ Π(X, Y ) with (cid:107)T − S(cid:107) (cid:54) ε and (cid:107)x − y(cid:107) (cid:54) ε. It is clear that (X, Y ) has the BPBp if and only if η(X, Y )(ε) > 0 for all 0 < ε < 1. Our first preliminary result deals with operators acting on an L1(µ) space and shows that the proof of some results can be reduced to the case when µ is a positive finite Borel regular measure defined on a compact Hausdorff space. Proposition 2.1. Let Y be a Banach space. Suppose that there is a function η : (0, 1) −→ (0,∞) such that η(cid:0)L1(m), Y(cid:1)(ε) (cid:62) η(ε) > 0 measure µ, the pair (L1(µ), Y ) has the BPBp with η(cid:0)L1(µ), Y(cid:1) (cid:62) η. η(cid:0)L1(m1), L1(m2)(cid:1)(ε) (cid:62) η(ε) > 0 get that (L1(µ), L1(ν)) has the BPBp with η(cid:0)L1(µ), L1(ν)(cid:1) (cid:62) η. (0 < ε < 1) (0 < ε < 1) Moreover, if Y = L1(ν) for an arbitrary measure ν, then it is enough to show that for every positive finite Borel regular measure m defined on a compact Hausdorff space. Then, for every for all positive finite Borel regular measures m1 and m2 defined on Hausdorff compact spaces in order to Proof. Let 0 < ε < 1. Suppose that T ∈ L(L1(µ), Y ) is a norm-one operator and f0 ∈ SX satisfy that n=1 be a sequence in X such that (cid:107)fn(cid:107) (cid:54) 1 for all n and limn→∞ (cid:107)T fn(cid:107) = (cid:107)T f0(cid:107) > 1 − η(ε). Let {fn)∞ (cid:107)T(cid:107) = 1. The band X1 generated by {fn : n (cid:62) 0} is isometric to L1(J, m) for a finite positive Borel regular measure m defined on a compact Hausdorff space J by the Kakutani representation theorem. Let T1 be the restriction of T to X1. Then (cid:107)T1(cid:107) = 1 and (cid:107)T1f0(cid:107) > 1 − η(ε). By the assumption, there exist a norm-one operator S1 : X1 −→ Y and g ∈ SX1 such that (cid:107)S1g(cid:107) = 1, (cid:107)T1 − S1(cid:107) < ε and (cid:107)f − g(cid:107) < ε. Let P denote the canonical band projection from L1(µ) onto X1. Then S := S1P + T (Id−P ) is a norm- one operator from L1(µ) to Y , g can be viewed as a norm-one element in SL1(µ) (just extending by 0), (cid:107)Sg(cid:107) = 1, (cid:107)S − T(cid:107) < ε and (cid:107)f − g(cid:107) < ε. This completes the proof of the first part of the proposition. 4 CHOI, KIM, LEE, AND MART´IN In the case when Y = L1(ν), we observe that the image T (X1) is also contained in a band Y1 of L1(ν) which, again, is isometric to L1(m2) for a finite positive Borel regular measure m2 on a compact Hausdorff space J2. Now, we work with the restriction of T to X1 with values in Y1, we follow the proof of the first part and finally we consider the operators S there as an operator with values in L1(ν) (just (cid:3) composing with the formal inclusion of Y1 into L1(ν)). Since for every positive finite Borel regular measure m defined on a compact Hausdorff space, L1(m) is isometric to L1(µ) for a probability measure µ, we get the following. Corollary 2.2. Let Y be a Banach space. Suppose that there is a strictly positive function η : (0, 1) −→ (0,∞) such that η(cid:0)L1(µ1), Y(cid:1) (cid:62) η for every probability measure µ1. Then (L1(µ), Y ) has the BPBp for every measure µ, with η(cid:0)L1(µ), Y(cid:1) (cid:62) η. Let us present the first application of the above results. For a σ-finite measure µ1, it is shown in [18] that (L1(µ1), Y ) has the BPBp if Y has the Radon-Nikod´ym property and ((cid:96)1, Y ) has the BPBp. By following the proof of [18, Theorem 2.2], we conclude that there is a strictly positive function ηY : (0, 1) −→ (0,∞) such that η(L1(µ1), Y ) (cid:62) ηY for every probability measure µ1. Therefore, the corollary above provides the same result without the assumption of σ-finiteness. We also recall that Lq(ν) is uniformly convex for all 1 < q < ∞ and for all measures ν, so it has the Radon-Nikod´ym property and ((cid:96)1, Lq(ν)) has the BPBp [3]. Hence we get the following. Corollary 2.3. Let µ be an arbitrary measure. If Y is a Banach space with the Radon-Nikod´ym property and such that ((cid:96)1, Y ) has the BPBp, then the pair (L1(µ), Y ) has the BPBp. In particular, (L1(µ), Lq(ν)) has the BPBp for all 1 < q < ∞ and all arbitrary measures ν. (cid:104)(cid:76) We now deal with operators with values on an (cid:96)∞-sum of Banach spaces, presenting the following result from [10] which we will use in section 4. Given a family {Yj : j ∈ J} of Banach spaces, we denote by Proposition 2.4 ([10]). Let X be a Banach space and let {Yj : j ∈ J} be a family of Banach spaces η(X, Yj)(ε) > 0 for all 0 < ε < 1, then (X, Y ) has and let Y = the (cid:96)∞-sum of the family. j∈J Yj (cid:105) denote their (cid:96)∞-sum. If inf j∈J j∈J Yj (cid:96)∞ (cid:105) (cid:104)(cid:76) (cid:96)∞ the BPBp with We will use this result for operators with values in L∞(ν). To present the result, we fist recall that given a localizable measure ν, we have the following representation (1) η(X, Y ) = inf j∈J η(X, Yj). L∞(ν) ≡(cid:104)(cid:77) (cid:105) Yj j∈J , (cid:96)∞ where each space Yj is either 1-dimensional or of the form L∞([0, 1]Λ) for some finite or infinite set Λ and [0, 1]Λ is endowed with the product measure of the Lebesgue measures. We refer to [32] for its background. With this in mind, the following corollary follows from the proposition above. Corollary 2.5. Let X be a Banach space. Suppose that there is a strictly positive function η : (0, 1) −→ (0,∞) such that for every finite or infinite set Λ. Then the pair (X, L∞(ν)) has the BPBp for every localizable measure ν with η(cid:0)X, L∞([0, 1]Λ)(cid:1)(ε) (cid:62) η(ε) η(cid:0)X, L∞(ν)(cid:1)(ε) (cid:62) min{η(ε), ε2/2} (0 < ε < 1) (0 < ε < 1). The proof is just an application of Proposition 2.4, the representation formula given in (1) and the Bishop-Phelps-Bollob´as theorem (Theorem 1.1). THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 5 Let us comment that the analogue of Proposition 2.4 is false for (cid:96)1-sums in the domain space (see [10]), so Proposition 2.1 cannot be derived directly from the decomposition of L1(µ) spaces analogous to (1). Before finishing this section, we state the following lemma of [3] which we will frequently use afterwards. Lemma 2.6 ([3, Lemma 3.3]). Let {cn} be a sequence of complex numbers with cn (cid:54) 1 for every n, and n=1 αncn > 1 − η. Then for every 0 < r < 1, the let η > 0 be such that for a convex series(cid:80) αn, Re(cid:80)∞ set A := {i ∈ N : Re ci > r}, satisfies the estimate(cid:88) αi (cid:62) 1 − η 1 − r . i∈A 3. The Bishop-Phelps-Bollob´as property of (L1(µ), L1(ν)) Our goal in this section is to prove the following result. Theorem 3.1. Let µ and ν be arbitrary measures. Then the pair (L1(µ), L1(ν)) has the BPBp. Moreover, there exists a strictly positive function η : (0, 1) −→ (0,∞) such that η(cid:0)L1(µ), L1(ν)(cid:1)(ε) (cid:62) η(ε) (0 < ε < 1). By Proposition 2.1, it is enough to get the result for finite regular positive Borel measures defined on compact Hausdorff spaces. Therefore, Theorem 3.1 follows directly from the next result. Theorem 3.2. Let m1 and m2 be finite regular positive Borel measures on compact Hausdorff spaces J1 and J2, respectively. Let 0 < ε < 1 and suppose that T ∈ L(L1(m1), L1(m2)) with (cid:107)T(cid:107) = 1 and f0 ∈ SL1(m1) satisfy (cid:107)T f0(cid:107) > 1 − ε18 53227 . Then there are S ∈ SL(L1(m1),L1(m2)) and g ∈ SL1(m1) such that (cid:107)Sg(cid:107) = 1, √ (cid:107)T − S(cid:107) < 4 (cid:107)f − g(cid:107) < 4ε and ε. Prior to presenting the proof of this theorem, we have to recall the following representation result for operators from L1(m1) into L1(m2). As we announced in the introduction, we deal with only complex spaces, being the real case easily deductible, even easier, from the proof of the complex case. Let m1 and m2 be finite regular positive Borel measures on compact Hausdorff spaces J1 and J2, respectively. For a complex-valued Borel measure µ on the product space J1 × J2, we define their marginal measures µi on Ji (i = 1, 2) as follows: µ1(A) = µ(A × J2) µ2(B) = µ(J1 × B), and where A and B are Borel measurable subsets of J1 and J2, respectively. Let M (m1, m2) be the complex Banach lattice consisting of all complex-valued Borel measures µ on the product space J1 × J2 such that each µi is absolutely continuous with respect to mi for i = 1, 2 with the norm It is clear that to each µ ∈ M (m1, m2) there corresponds a unique bounded linear operator Tµ ∈ L(L1(m1), L1(m2)) defined by (cid:104)Tµ(f ), g(cid:105) = f (x)g(y) dµ(x, y), (cid:13)(cid:13)(cid:13)(cid:13)∞ . (cid:13)(cid:13)(cid:13)(cid:13) dµ1 (cid:90) dm1 J1×J2 (cid:13)(cid:13)(cid:13)(cid:13) dµ1 dm1 (cid:13)(cid:13)(cid:13)(cid:13)∞ . (cid:107)Tµ(cid:107) = where f ∈ L1(m1) and g ∈ L∞(m2). Iwanik [25] showed that the mapping µ (cid:55)−→ Tµ is a surjective lattice isomorphism and Even though he showed this for the real case, it can be easily generalized to the complex case. For details, see [25, Theorem 1] and [38, IV Theorem 1.5 (ii), Corollary 2]. Since the proof of Theorem 3.2 is complicated, we divide it into the following two lemmas. 6 CHOI, KIM, LEE, AND MART´IN Lemma 3.3. Let 0 < ε < 1. Suppose that Tµ is an element of L(L1(m1), L1(m2)) with (cid:107)Tµ(cid:107) = 1 for some µ ∈ M (m1, m2) and that f0 ∈ SL1(m1) is a nonnegative simple function such that (cid:107)Tµf0(cid:107) > 1 − ε3 26 . Then there are a norm-one bounded linear operator Tν for some ν ∈ M (m1, m2) and a nonnegative simple function f1 in SL1(m1) such that (cid:107)Tµ − Tν(cid:107) < ε, (cid:107)f1 − f0(cid:107) < 3ε and we have, for all x ∈ supp(f1), Proof. Let f0 =(cid:80)n m1(Bj) > 0 for all 1 (cid:54) j (cid:54) n, and(cid:80)n j=1 αj χBj m1(D) > 0. Since (cid:107)Tµf0(cid:107) > 1 − ε3 m1(Bj ) , where {Bj}n (x) = 1. dν1 dm1 j=1 are mutually disjoint Borel subsets of J1, αj (cid:62) 0 and 8}. It is clear that (x) > 1 − ε j=1 αj = 1. Let D = {x ∈ J1 : dµ1 dm1 26 , there is g0 ∈ SL∞(m2) such that Re (cid:104)Tµf0, g0(cid:105) > 1 − ε3 26 . (cid:90) j ∈ {1, . . . , n} : 1 m1(Bj) Bj dµ1 dm1 (x) dm1(x) > 1 − ε2 26 (cid:41) . (cid:40) J = Let Then we have Indeed, since αj (cid:62) 1 − ε > 0. (cid:88) (cid:90) j∈J 1 − ε3 26 < Re (cid:104)Tµf0, g0(cid:105) = Re (cid:90) J1×J2 f0(x) dµ(x, y) = J1×J2 (cid:90) n(cid:88) (cid:54) 1 (cid:90) J1 n(cid:88) f0(x)g0(y) dµ(x, y) f0(x) dµ1(x) (cid:90) we have(cid:80) = dµ1 dm1 j∈J αj (cid:62) 1 − ε > 0 by Lemma 2.6. Note also that for each j ∈ J, dµ1(x) = m1(Bj) m1(Bj) αj αj j=1 j=1 Bj Bj 1 (x) dm1(x), 1 − ε2 26 < dµ1 dm1 1 m1(Bj) (cid:90) (cid:90) Bj∩D (cid:54) m1(Bj ∩ D) m1(Bj) = Bj 1 + m1(Bj) = 1 − ε 8 m1(Bj \ D) m1(Bj) . (cid:90) 1 dµ1 dm1 (x) dm1(x) m1(Bj) Bj\D (x) dm1(x) (cid:16) dµ1 dm1 1 − ε 8 (x) dm1(x) + (cid:17) m1(Bj \ D) m1(Bj) Hence we deduce that, for all j ∈ J, Let Bj = Bj ∩ D and βj = αj(cid:80) j∈J αj . m1(Bj \ D) m1(Bj) (cid:54) ε 8 for all j ∈ J and define χ Bj (cid:88) f1 = βj . m1( Bj) j∈J THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 7 (cid:88) βj χ Bj m1( Bj) αj χBj m1(Bj) j∈{1,...,n}\J (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + It is clear that f1 is a nonnegative element in SL1(m1) and (cid:107)f0 − f1(cid:107) (cid:54) (cid:54) < (cid:54) j∈J j∈J (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:88) j∈J j∈J j∈J αj αj αj αj αj −(cid:88) j∈J m1(Bj) χBj χBj m1(Bj) (cid:32) (cid:32) (cid:18) χBj χBj m1(Bj) m1(Bj) m1(Bj \ D) m1(Bj) − χ Bj m1( Bj) − χ Bj m1( Bj) − χ Bj m1(Bj) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:88) (cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:19)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:88) j∈J j∈J j∈J (cid:54) 2 + 2ε (cid:54) 3ε. (αj − βj) χ Bj m1( Bj) αj j∈{1,...,n}\J αj − βj + ε (cid:32) χ Bj αj m1(Bj) − χ Bj m1( Bj) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 −(cid:88) j∈J (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + ε αj (cid:88) (cid:33)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + Define where B =(cid:83) x ∈ J1, j∈J (cid:88) j∈J χ Bj (x) (cid:18) dµ1 dm1 dν(x, y) = (cid:19)−1 (x) dµ(x, y) + χ(J1\ B) dµ(x, y), Bj. It is clear that dν1 (x) = 1 on B and dν1 (x) (cid:54) 1 elsewhere. Note also that for all dν − µ1 dm1 (x) = dm1 (cid:88) j∈J dm1 (cid:32)(cid:18) dµ1 (cid:33) (cid:19)−1 − 1 χ Bj (x) (x) dm1 dµ1 dm1 (cid:54) 1 1 − ε/8 − 1 = ε 8 − ε < ε. Hence Tν is a norm-one operator such that (cid:107)Tµ − Tν(cid:107) < ε, (cid:107)f1 − f0(cid:107) < 3ε and dν1 (x) = 1 for all x ∈ supp(f1). (cid:3) Lemma 3.4. Let 0 < ε < 1. Suppose that Tν is a norm-one operator in L(L1(m1), L1(m2)) and that f is a nonnegative norm-one simple function in SL1(m1) satisfying (cid:107)Tνf(cid:107) > 1 − ε6 (x) = 1 for all x in the support of f . Then there are a nonnegative simple function f in SL1(m1) and a norm-one operator Tν in L(L1(m1), L1(m2)) such that 27 and dν1 dm1 dm1 √ (cid:107)Tν − Tν(cid:107) < 3 ε and (cid:107)f − f(cid:107) < 3ε. f(cid:107) = 1, Proof. Let f = (cid:80)n m1(Bj) > 0 for all 1 (cid:54) j (cid:54) n, and(cid:80)n (cid:107)Tν m1(Bj ) , where {Bj}n n(cid:88) j=1 βj χBj 1 − ε6 27 < Re (cid:104)Tνf, g(cid:105) = (cid:26) Let J = j=1 are mutually disjoint Borel subsets of J1, βj (cid:62) 0 and j=1 βj = 1. Since (cid:107)Tνf(cid:107) > 1 − ε6 27 , there is g ∈ SL∞(m2) such that χBj (x) (cid:27) m1(Bj) (g(y)) dν(x, y). J1×J2 βj Re (cid:90) j=1 . From Lemma 2.6 it follows that (cid:90) j ∈ {1, . . . , n} : Re χBj (x) m1(Bj) g(y) dν(x, y) > 1 − ε3 26 J1×J2 (cid:88) j∈J βj > 1 − ε3 2 . 8 Let f1 =(cid:80) j∈J βj χBj m1(Bj ) , where βj = βj/((cid:80) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:26) (cid:107)f1 − f(cid:107) (cid:54) j∈J CHOI, KIM, LEE, AND MART´IN j∈J βj) for all j ∈ J. Then ( βj − βj) χBj m1(Bj) βj (cid:54) ε3 < ε. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:88) j∈J (cid:27) . √ ε 23/2 C = (x, y) : g(y)h(x, y) − 1 < Note that there is a Borel measurable function h on J1 × J2 such that dν(x, y) = h(x, y) dν(x, y) and h(x, y) = 1 for all (x, y) ∈ J1 × J2. Let Define two measures νf and νc as follows: νf (A) = ν(A \ C) and νc(A) = ν(A ∩ C) for every Borel subset A of J1 × J2. It is clear that dνf = ¯hdνf , dνc = ¯hdνc, and dν = dνf + dνc. dν = dνf + dνc, (x) = 1 for all x ∈(cid:83)n Since dν1 dm1 for all x ∈ B =(cid:83)n 1 − ε We claim that 24 . So we have j=1 Bj, we have 1 = (x) = (x) + (x) dν1 dm1 dνf1 dm1 dνc1 dm1 j=1 Bj, and we deduce that ν1(Bj) = m1(Bj) for all 1 (cid:54) j (cid:54) n. νf1(Bj ) m1(Bj ) 22 for all j ∈ J. Indeed, if g(y)h(x, y) − 1 (cid:62) √ (cid:54) ε2 ε 23/2 , then Re (g(y)h(x, y)) (cid:54) (cid:90) 1 − ε3 26 (cid:54) = = m1(Bj) m1(Bj) 1 1 1 (cid:90) (cid:90) Re (cid:90) J1×J2 J1×J2 J1×J2 χBj (x)g(y) dν(x, y) χBj (x) Re(cid:0)g(y)h(x, y)(cid:1) dν(x, y) χBj (x) Re(cid:0)g(y)h(x, y)(cid:1) dνf(x, y) χBj (x) Re(cid:0)g(y)h(x, y)(cid:1) dνc(x, y) (cid:17) J1×J2 24 )νf1(Bj) + νc1(Bj) (1 − ε νf1(Bj) m1(Bj) (cid:16) . m1(Bj) 1 + m1(Bj) 1 (cid:54) m1(Bj) = 1 − ε 24 This proves our claim. We also claim that for each j ∈ J, there exists a Borel subset Bj of Bj such that (cid:17) (cid:16) 1 − ε 2 m1(Bj) (cid:54) m1( Bj) (cid:54) m1(Bj) and for all x ∈ Bj. Indeed, set Bj = Bj ∩(cid:110) (cid:90) x ∈ J1 : dνf1 (cid:90) dm1 dνf1 dm1 dm1(x) (cid:54) ε 2 Bj\ Bj dνf1 dm1 Bj (x) (cid:54) ε 2 (x) (cid:54) ε 2 (cid:111) . Then (x) dm1(x) = ν1 f(Bj) (cid:54) ε2 22 m1(Bj). This shows that m1(Bj \ Bj) (cid:54) ε 2 m1(Bj). This proves our second claim. THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) g(y) if g(y) (cid:54)= 0 and g(y) = 1 if g(y) = 0, and we write f =(cid:80) j∈J 9 βj χ Bj m1( Bj ) . Now, we define g by g(y) = g(y) Finally, we define the measure (cid:88) j∈J where B =(cid:83) j∈J dν(x, y) = χ Bj (x)g(y)h(x, y)dνc(x, y) (x) + χJ1\ B(x)dν(x, y), Bj. It is easy to see that dν1 (x) = 1 on B and dν1 (x) (cid:54) 1 elsewhere. Note that (cid:34) dm1 d(ν − ν)(x, y) = j∈J (cid:88) −(cid:88) (cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)g(y)h(x, y) − 1 ε 23/2 If (x, y) ∈ C, then g(y) (cid:62) 1 − √ Hence, for all (x, y) ∈ C we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)g(y)h(x, y) (cid:18) dνc1 dm1 (cid:19)−1 − 1 (x) So, we have for all x ∈ J1, dν − ν1 dm1 j∈J (x) (cid:54)(cid:88) (cid:88) (cid:54)(cid:88) j∈J + χ Bj (x) χ Bj (x) (x) χ Bj 2 j∈J √ √ (cid:54) 2 ε + ε < 3 √ This gives that (cid:107)Tν − Tν(cid:107) < 3 χ Bj (cid:42) (cid:43) ε. Tν m1( Bj) , g = (cid:18) dνc1 (cid:18) dνc1 dm1 dm1 (cid:19)−1 (cid:35) (cid:19)−1 − 1 dm1 (x) χ Bj (x) g(y)h(x, y) dνc(x, y) χ Bj (x)dνf (x, y). j∈J (cid:62) 1 − 1 (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) g(y) 23/2 and g(y) h(x, y) − 1 (cid:54) g(y)h(x, y) − 1 + g(y)h(x, y) − 1 g(y) (cid:54) 2 g(y) (cid:12)(cid:12)1 − g(y)(cid:12)(cid:12) g(y) √ (cid:54) 2 ε 23/2 23/2 23/2 − 1 √ ε. (cid:54) 2 (cid:19)−1 − 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (x) ε + + (x) (x) √ dm1 dm1 (cid:54) 2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:54)(cid:12)(cid:12)(cid:12)g(y)h(x, y) − 1 (cid:12)(cid:12)(cid:12) (cid:18) dνc1 (cid:19)−1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:18) dνc1 (cid:19)−1 (cid:18) dνc1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:34) (cid:18) dνc1 (cid:18) dνc1 (cid:19)−1 (cid:19)(cid:19) ε dνf1 dm1 √ (cid:18) (cid:18) dm1 dm1 dm1 (cid:88) dm1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:18) dνc1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . (cid:19)−1 − 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:35) (cid:19)−1 − 1 (cid:18) dνf1 √ (x) (x) (x) (x) + 2 ε + 1 − dνc1 dm1 (x) + χ Bj (x) j∈J (x) dm1 (cid:19) dνc1 dm1 (x) ε. Note also that, for all j ∈ J, (cid:90) (cid:90) (cid:90) (cid:90) = = = g(y) dν(x, y) (cid:18) dνc1 (cid:19)−1 dm1 J1×J2 (x) χ Bj m1( Bj) χ Bj (x) m1( Bj) J1×J2 χ Bj (x) m1( Bj) χ Bj (x) m1( Bj) J1 J1 h(x, y) (cid:18) dνc1 (x) dm1 dm1(x) = 1. (cid:19)−1 (x) dνc(x, y) dνc1(x) 10 Hence we get (cid:68) (cid:69) CHOI, KIM, LEE, AND MART´IN = 1, which implies that (cid:107)Tν f , g Tν (cid:107) f − f(cid:107) (cid:54) (cid:107) f − f1(cid:107) + (cid:107)f1 − f(cid:107) f(cid:107) = (cid:107)Tν(cid:107) = 1. Finally, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + ε (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) χBj −(cid:88) βj χ Bj m1( Bj) (cid:32)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) χ Bj βj βj j∈J − χBj χBj m1(Bj) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + m1( Bj) m1(Bj \ Bj) m1( Bj) + ε m1( Bj) m1( Bj) = j∈J (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:54)(cid:88) (cid:88) (cid:88) j∈J j∈J = 2 (cid:54) 2 j∈J βj βj (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:33) + ε − χBj m1(Bj) ε 2 m1(Bj) m1( Bj) + ε (cid:54) ε 1 − ε/2 + ε < 3ε. (cid:3) We are now ready to prove Theorem 3.2. Proof of Theorem 3.2. Let 0 < ε < 1. Suppose that T is a norm-one element in L(L1(m1), L1(m2)) and there is f ∈ SL1(m1) such that (cid:107)T f(cid:107) > 1− ε18 53227 . Then there is an isometric isomorphism ψ from L1(m1) onto itself such that ψ(f ) = f. Using T ◦ ψ−1 instead of T , we may assume that f is nonnegative. Since simple functions are dense in L1(m1), we can choose a nonnegative simple function f0 ∈ SL1(m1) arbitrarily close to f so that (cid:107)T f0(cid:107) > 1 − ε18 53227 = 1 − ε3 26 , 1 5·27 . By Lemma 3.3, there exist a norm-one bounded linear operator Tν for some ν ∈ where ε1 = ε6 M (m1, m2) and a nonnegative simple function f1 in SL1(M1) such that (cid:107)T − Tν(cid:107) < ε1, (cid:107)f1 − f(cid:107) < 3ε1 and dν1 dm1 (x) = 1 for all x ∈ supp(f1). Then (cid:107)Tνf1(cid:107) (cid:62) (cid:107)T f(cid:107) − (cid:107)T f − Tνf(cid:107) − (cid:107)Tν(f − f1)(cid:107) (cid:62) 1 − ε3 1 Now, by Lemma 3.4, there exist a nonnegative simple function f and an operator Tν in L(L1(m1), L1(m2)) such that (cid:107)Tν ε and (cid:107)f − f(cid:107) < 4ε, which complete the proof. (cid:3) f(cid:107) = (cid:107)Tν(cid:107) = 1, (cid:107)Tν − Tν(cid:107) (cid:54) 3 26 − ε1 − 3ε1 (cid:62) 1 − 5ε1 = 1 − ε6 27 . √ ε and (cid:107)f1 − f(cid:107) (cid:54) 3ε. Therefore, (cid:107)T − Tν(cid:107) < 4 √ 4. The Bishop-Phelps-Bollob´as property of (L1(µ), L∞(ν)) Our aim now is to show that (L1(µ), L∞(ν)) has the BPBp for any measure µ and any localizable measure ν. Theorem 4.1. Let µ be an arbitrary measure and let ν be a localizable measure. Then the pair (L1(µ), L∞(ν)) has the BPBp. Moreover, η(cid:0)L1(µ), L∞(ν)(cid:1)(ε) (cid:62)(cid:16) ε (cid:17)8 10 (0 < ε < 1). By Corollaries 2.2 and 2.5, it is enough to prove the result in the case where µ is σ-finite and ν is the product measure on [0, 1]Λ. Therefore, we just need to prove the following result. Theorem 4.2. Assume µ is a σ-finite measure and ν is the product measure of Lebesgue measures on [0, 1]Λ. Let 0 < ε < 1/3, let T : L1(µ) −→ L∞(ν) be a bounded linear operator of norm one and let f0 ∈ SL1(µ) satisfy (cid:107)T (f0)(cid:107)∞ > 1 − ε8. Then there exist S ∈ L(L1(µ), L∞(ν)) with (cid:107)S(cid:107) = 1 and g0 ∈ SL1(µ) such that (cid:107)S(g0)(cid:107)∞ = 1, (cid:107)T − S(cid:107) < 2ε and (cid:107)f0 − g0(cid:107)1 < 10ε. THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 11 Recall that the particular case where Λ reduces to one point was established in [9]. Actually, our proof is based on the argument given there. Prior to giving the proof of Theorem 4.2, we present the following representation result for operators from L1(µ) into L∞(cid:0)[0, 1]Λ(cid:1) and one lemma. measure such that Πn+1 is a refinement of Πn for each n, and the σ-algebra generated by(cid:83)∞ Let (Ω, Σ, µ) be a σ-finite measure space and let K = [0, 1]Λ be the product space equipped with the product measure ν of the Lebesgue measures. Let J be a countable subset of Λ and let πJ be the natural projection from K onto [0, 1]J . Fix a sequence (Πn) of finite partitions of [0, 1]J into sets of positive n=1 Πn is the Borel σ-algebra of [0, 1]J . For each y ∈ K and n ∈ N, let B(n, πJ (y)) be the set in Πn containing πJ (y). Then, given a Borel set F of the form F0 × [0, 1]Λ\J with F0 ⊂ [0, 1]J , define (cid:26) (cid:26) y ∈ K : δ(F ) = ν(π−1 It is easy to check that δ(F ) = δJ (F0) × [0, 1]Λ\J , where ν(π−1 lim n→∞ δJ (F0) = y ∈ [0, 1]J : lim n→∞ ν(F ∩ π−1 J (B(n, πJ (y))) J (B(n, πJ (y))) = 1 J (F0 ∩ B(n, y))) ν(π−1 J (B(n, y))) = 1 . Using the martingale almost everywhere convergence theorem [22], we have ν(F ∆δ(F )) = 0 . (cid:27) (cid:27) where F ∆δ(F ) denotes the symmetric difference of the sets F and δ(F ). On the other hand, it is well-known that the space L(L1(µ), L∞(ν)) is isometrically isomorphic to the space L∞(µ ⊗ ν), where µ ⊗ ν denotes the product measure on Ω × K. More precisely, the operator(cid:98)h corresponding to h ∈ L∞(µ ⊗ ν) is given by (cid:98)h(f )(t) = (cid:90) Ω h(ω, t)f (ω) dµ(ω) for ν-almost every t ∈ K. For a reference, see [20]. Lemma 4.3. Let M be a measurable subset of Ω × K with positive measure, 0 < ε < 1, and let f0 be a simple function. If (cid:107)(cid:100)χM (f0)(cid:107)∞ > 1 − ε, then there exists a simple function g0 ∈ SL1(µ) such that (cid:107)[(cid:100)χM +(cid:98)ϕ](g0)(cid:107)∞ = 1 and (cid:107)f0 − g0(cid:107)1 < 4 √ ε for every simple function ϕ in L∞(µ ⊗ ν) with (cid:107)ϕ(cid:107)∞ (cid:54) 1 and vanishing on M . Proof. Write f0 =(cid:80)m measure, Ak ∩ Al = ∅ for k (cid:54)= l, and αj is a positive real number for every j = 1, . . . , m with(cid:80)m Since (cid:107)(cid:100)χM (f0)(cid:107)∞ > 1 − ε, there is a measurable subset B of K such that 0 < ν(B) and µ(Aj ) ∈ SL1(µ), where each Aj is a measurable subset of Ω with finite positive j=1 αj = 1. j=1 αj χAj (cid:28)(cid:100)χM (f0), (cid:29) χB ν(B) > 1 − ε. We may assume that there is a countable subset J of Λ such that M = M0×[0, 1]Λ\J and B = B0×[0, 1]Λ\J for some measurable subsets M0 ⊂ Ω × [0, 1]J and B0 ⊂ [0, 1]J . For each j ∈ {1, . . . , m}, we write Mj = M (cid:84) (Aj × B) = (M0 ∩ (Aj × B0)) × [0, 1]Λ\J and define Hj = {(x, y) : x ∈ Aj, y ∈ δ(cid:0)(Mj)x (cid:1)}. As in the proof of [35, Proposition 5], the Hj's are disjoint measurable subsets of Ω × K. We note that for each j ∈ {1, . . . , m}, we have Hj ⊂ Aj × δ(B) and (µ ⊗ ν)(Mj∆Hj) = 0. 12 CHOI, KIM, LEE, AND MART´IN Now, by Fubini theorem, we have that 1 − ε < (cid:104)(cid:98)χM (f0), χB ν(B) (cid:105) (cid:90) (cid:90) j=1 m(cid:88) m(cid:88) m(cid:88) j=1 j=1 = = = = αj µ(Aj)ν(B) Ω×K χMj (x, y) d(µ ⊗ ν) χHj (x, y) d(µ ⊗ ν) µ(Aj)ν(B) αj ν(B) δ(B) αj (cid:90) (cid:90) Ω×K µ(H y j ) µ(Aj) m(cid:88) dν(y) 1 ν(δ(B)) δ(B) j=1 αj µ(H y j ) µ(Aj) dν(y). So, there exists y0 ∈ δ(B) such that (cid:110) Let J = by Lemma 2.6, we also have αJ :=(cid:80) j ∈ {1, . . . , m} : µ(H y0 j ) m(cid:88) (cid:111) µ(Aj ) > 1 − √ j=1 ε αj µ(H y0 j ) µ(Aj) > 1 − ε. j∈J αj > 1 − √ (cid:88) g0 = βj j∈J ε. Define j χH y0 µ(H y0 j ) , . For each j ∈ J, we have that µ(Aj \ H y0 j ) < √ εµ(Aj) and, −(cid:88) −(cid:88) j∈J j∈J αj χAj µ(Aj) βj χAj µ(Aj) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + 2 √ ε √ ε βj βj βj αj j∈J j∈J (cid:54) (cid:54) j∈J j∈J √ ε j j j χAj µ(Aj) χAj µ(Aj) χH y0 µ(H y0 j ) χH y0 µ(H y0 j ) (cid:107)g0 − f0(cid:107) < where βj = αj/αJ . Then (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) −(cid:88) −(cid:88) −(cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) We claim that (cid:100)χM + (cid:98)ϕ attains its norm at g0. Let Bn = π−1 (cid:1) ν(cid:0)(Mj)x ∩ Bn ν(cid:0)(Mj)x ∩ Bn (cid:90) j we have (x, y0) ∈ Hj, which implies that χH y0 µ(H y0 j ) j∈J µ(Aj \ H y0 j ) µ(Aj) χH y0 µ(Aj) √ j∈J √ ε (cid:54) 4 every x ∈ H y0 lim n→∞ ν(Bn) (cid:54) 2 = 1. j∈J j∈J j ε + 2 βj βj 1 βj χAj µ(Aj) βj χH y0 µ(Aj) j (cid:1) 1 = lim n→∞ = lim n→∞ H y0 j µ(H y0 j ) (µ ⊗ ν)(Mj ∩ (H y0 j )ν(Bn) µ(H y0 ν(Bn) j × Bn)) . dµ(x) It follows from the Lebesgue dominated convergence theorem and Fubini theorem that, for each j ∈ J, J (B(n, πJ (y0))) for each n. Note that for THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 13 On the other hand, since the simple function ϕ is assumed to vanish on M and (cid:107)ϕ(cid:107)∞ (cid:54) 1, we have (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:10)(cid:98)ϕ(cid:0) χH y0 µ(H y0 j ) j (cid:1), (cid:11)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) χBn ν(Bn) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ d(µ ⊗ ν) 1 µ(H y0 (cid:90) j × Bn) \ Mj) j )ν(Bn) j ×Bn H y0 j )ν(Bn) (cid:54) (µ ⊗ ν)((H y0 µ(H y0 = 1 − (µ ⊗ ν)(Mj ∩ (H y0 µ(H y0 j )ν(Bn) j × Bn)) (cid:17) χH y0 µ(H y0 j ) j × Bn)) , j −→ 0, χBn ν(Bn) (cid:43)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:16)(cid:88) j∈J βj µ(H y0 (µ ⊗ ν)(M ∩ (H y0 j )ν(Bn) (µ ⊗ ν)(Mj ∩ (H y0 j )ν(Bn) µ(H y0 j × Bn)) as n → ∞. Therefore, 1 (cid:62)(cid:13)(cid:13)[(cid:100)χM +(cid:98)ϕ](g0)(cid:13)(cid:13)∞ (cid:62) lim (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:42) ((cid:100)χM +(cid:98)ϕ) (cid:88) (cid:88) which shows that(cid:100)χM +(cid:98)ϕ attains its norm at g0. (cid:62) lim n→∞ = lim n→∞ n→∞ j∈J j∈J βj βj = 1, (cid:3) We are now ready to present the proof of the main result in this section. Proof of Theorem 4.2. Since the set of all simple functions is dense in L1(µ), we may assume m(cid:88) f0 = αj χAj µ(Aj) ∈ SL1(µ), j=1 αj is a nonzero complex number with (cid:80)m where each Aj is a measurable subset of Ω with finite positive measure, Ak ∩ Al = ∅ for k (cid:54)= l, and every j=1 αj = 1. We may also assume that 0 < αj (cid:54) 1 for every j = 1, . . . , m. Indeed, there exists an isometric isomorphism Ψ : L1(µ) −→ L1(µ) such that Ψ(f0) = f0. Let h be the element in L∞(Ω× K, µ⊗ ν) with (cid:107)h(cid:107)∞ = 1 corresponding to T , that is, T =(cid:98)h. We may Hence we may replace T and f0 by T ◦ Ψ−1 and Ψ(f0), respectively. such that (cid:107)h − h0(cid:107)∞ < (cid:107)T (f0)(cid:107)∞ − (1 − ε8), hence (cid:107)(cid:98)h0(f0)(cid:107)∞ > 1 − ε8. We can write h0 =(cid:80)p l=1 clχDl , where each Dl is a measurable subset of Ω × K with positive measure, Dk ∩ Dl = ∅ for k (cid:54)= l, cl (cid:54) 1 for every l = 1, . . . , p, and cl0 = 1 for some 1 (cid:54) l0 (cid:54) p. h0 ∈ L∞(Ω × K, µ ⊗ ν), find a simple function (cid:107)h0(cid:107)∞ = 1 Let B be a Lebesgue measurable subset of K with 0 < ν(B) < ∞ such that Choose θ ∈ R so that χB ν(B) (cid:12)(cid:12)(cid:12)(cid:12)(cid:28)(cid:98)h0(f0), (cid:29)(cid:12)(cid:12)(cid:12)(cid:12) > 1 − ε8. (cid:105)(cid:12)(cid:12) 1 − ε8 < (cid:12)(cid:12)(cid:104)(cid:98)h0(f0), = eiθ(cid:104)(cid:98)h0(f0), (cid:28)(cid:98)h0( m(cid:88) χB ν(B) χB ν(B) αj eiθ = (cid:105) ), χAj µ(Aj) j=1 (cid:29) . χB ν(B) αj > 1 − ε8 1 − (1 − ε4) (cid:88) j∈J (cid:88) j∈J f1 = = 1 − ε4. . µ(Aj) + αJ (cid:19) χAj (cid:16) 1 (cid:18) αj (cid:13)(cid:13)(cid:13)1 (cid:28)(cid:98)h0(f1), (cid:20) αJ αj Re eiθ (cid:17)(cid:13)(cid:13)(cid:13)(cid:88) αj (cid:13)(cid:13)(cid:13)1 χAj µ(Aj) − 1 (cid:29)(cid:21) (cid:28)(cid:98)h0( χB ν(B) (cid:29)(cid:21) ), χB ν(B) χAj µ(Aj) j /∈J αj + (1 − αJ ) = 2(1 − αJ ) < 2ε4. j∈J αj (cid:107)f0 − f1(cid:107)1 (cid:54) (cid:13)(cid:13)(cid:13)(cid:88) (cid:88) (cid:12)(cid:12)(cid:12)(cid:12) (cid:62) Re (cid:12)(cid:12)(cid:12)(cid:12)(cid:104)(cid:98)h0(f1), χB ν(B) j /∈J = (cid:105) χAj µ(Aj) eiθ j∈J (cid:20) (cid:88) (cid:88) (cid:111) (cid:28)(cid:98)h0( = 1 αJ (cid:20) (cid:110) 1 αJ l ∈ {1, . . . , p} : Re (eiθ cl) > 1 − ε2 > 2 αj(1 − ε4) = 1 − ε4. j∈J . On the other hand, for each j ∈ J, we have (cid:29)(cid:21) and Let L = 14 Set (cid:26) J = j ∈ {1, . . . , m} : Re(cid:2) eiθ(cid:104)(cid:98)h0( CHOI, KIM, LEE, AND MART´IN (cid:105)(cid:3) > 1 − ε4 (cid:27) . χAj µ(Aj) ), χB ν(B) By Lemma 2.6, we have αJ = We define We can see that (cid:107)f1(cid:107)1 = 1, 1 − ε4 < Re ), χB ν(B) χAj µ(Aj) (µ ⊗ ν)(Dl ∩ (Aj × B)) Re (eiθ cl) = eiθ p(cid:88) (cid:54) (cid:88) (cid:88) l=1 l∈{1,...,p}\L + l∈L (cid:54) 1 − ε2 2 µ(Aj)ν(B) (µ ⊗ ν)(Dl ∩ (Aj × B)) µ(Aj)ν(B) ) (1 − ε2 2 (µ ⊗ ν)(Dl ∩ (Aj × B)) (cid:88) µ(Aj)ν(B) (µ ⊗ ν)(Dl ∩ (Aj × B)) . l∈{1,...,p}\L µ(Aj)ν(B) This implies that for each j ∈ J (cid:88) p(cid:88) Since l∈{1,...,p}\L for every j ∈ J we have that(cid:88) (µ ⊗ ν)(Dl ∩ (Aj × B)) µ(Aj)ν(B) (cid:54) 2ε2. (µ ⊗ ν)(Dl ∩ (Aj × B)) > 1 − ε4, l=1 µ(Aj)ν(B) (µ ⊗ ν)(Dl ∩ (Aj × B)) l∈L µ(Aj)ν(B) (cid:62) (1 − ε4 − 2ε2) (cid:62) 1 − 3ε2. Set D =(cid:83) THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 15 l∈L Dl. Then we can see (cid:105) = (cid:104)(cid:98)χD(f1), χB ν(B) (cid:88) (cid:0) αj (cid:1) ·(cid:88) µ ⊗ ν(Dl ∩ (Aj × B)) (cid:62) 1 − 3ε2. By Lemma 4.3, there is g0 ∈ SL1(µ) such that (cid:107)((cid:98)χD + (cid:98)ϕ)(g0)(cid:107)∞ = 1 and (cid:107)f1 − g0(cid:107) < 4 √ every simple function ϕ in L∞(µ ⊗ ν) vanishing on D with (cid:107)ϕ(cid:107)∞ (cid:54) 1. Therefore, we have j∈J l∈L µ(Aj)ν(B) αJ 3ε2 < 8ε for (cid:107)f0 − g0(cid:107)1 (cid:54) (cid:107)f0 − f1(cid:107)1 + (cid:107)f1 − g0(cid:107)1 (cid:54) 2ε4 + 8ε < 10ε. (cid:88) l /∈L Define h1 = e−iθ χD + cl χDl ∈ L∞(µ ⊗ ν). Let S be the operator in L(L1(µ), L∞(m)) corresponding to h1. Then we get (cid:107)S(g0)(cid:107)∞ = (cid:107)(cid:99)h1(g0)(cid:107)∞ = 1 and (cid:107)h0 − h1(cid:107)∞ = max l∈L cl − e−iθ = max l∈L eiθ cl − 1. As Re (eiθ cl) > 1 − ε2 2 for every l ∈ L, we have that (cid:0) Im (eiθ cl)(cid:1)2 (cid:54) 1 −(cid:0) Re (eiθ cl)(cid:1)2 (cid:113)(cid:0)1 − Re (eiθ cl)(cid:1)2 < (cid:112)ε4/4 + (ε2 − ε4/4) = ε, eiθ cl − 1 = < 1 − (1 − ε2 2 +(cid:0) Im (eiθ cl)(cid:1)2 )2 = ε2 − ε4 4 . Since we conclude that and (cid:107)h0 − h1(cid:107)∞ < ε (cid:107)T − S(cid:107)∞ (cid:54) (cid:107)h − h0(cid:107)∞ + (cid:107)h0 − h1(cid:107)∞ < ε8 + ε < 2ε. (cid:3) 5. The Bishop-Phelps-Bollob´as Property for some operators from L1(µ) into C(K) Throughout this section, we consider only a finite measure µ on a measurable space (Ω, Σ) and real Banach spaces L1(µ) and C(K). Our aim is to obtain the Bishop-Phelps-Bollob´as property for some classes of operators from L1(µ) to C(K), sharpening the results about denseness of norm-attaining operators given by Iwanik in 1982 [26]. We use the following standard representation of operators into C(K) [23, Theorem 1 in p. 490]. Lemma 5.1. Given a bounded linear operator T : X −→ C(K), define F : K −→ X∗ by F (s) = T ∗(δs), where δs is the point measure at s ∈ K. Then, for x ∈ X, the relation T x(s) = (cid:104)x, F (s)(cid:105) defines an isometric isomorphism of L(X, C(K)) onto the space of weak∗ continuous functions from K to X∗ with the supremum norm. Moreover, compact operators correspond to norm continuous functions. Iwanik [26] considered operators T ∈ L(L1(µ), C(K)) satisfying one of the following conditions: (1) The map s (cid:55)−→ T ∗δs is continuous in measure. (2) There exists a co-meager set G ⊂ K such that {T ∗δs : s ∈ G} is norm separable in L∞(µ). We recall that a subset A is said to be a co-meager subset of K if the set K \ A is meager, that is, of first category. 16 CHOI, KIM, LEE, AND MART´IN Theorem 5.2. Let 0 < ε < 1. Suppose that T ∈ L(L1(µ), C(K)) (real case) has norm one and satisfies condition (1). If (cid:107)T f(cid:107) > 1 − ε2 6 for some f ∈ SL1(µ), then there exist S ∈ L(L1(µ), C(K)) with (cid:107)S(cid:107) = 1 and g ∈ SL1(µ) such that (cid:107)Sg(cid:107) = 1, (cid:107)S − T(cid:107) < ε, and (cid:107)f − g(cid:107) < ε. Moreover, S also satisfies condition (1). Proof. Without loss of generality, we assume that there exists s0 ∈ K such that Consider the function G : L∞(µ) −→ L∞(µ) given by (cid:16) G(h) = h ∧ (1 − ε/3) . T f (s0) > 1 − ε2 6 (cid:17) ∨ (−1 + ε/3) (cid:0)h ∈ L∞(µ)(cid:1). Since the lattice operation G is continuous in the L∞ norm and T satisfies condition (1), we can see that the mapping s (cid:55)−→ GT ∗δs is continuous in measure, hence weak∗-continuous. Let ¯S be the element of L(L1(µ), C(K)) represented by the function F (s) := GT ∗δs. Then Let C = (cid:107) ¯S − T(cid:107) = sup s∈K (cid:107)F (s) − T ∗δs(cid:107) (cid:54) ε 3 (cid:110) ω ∈ Ω : sign(cid:0)f (ω)(cid:1)T ∗δs(ω) > 1 − ε . (cid:111) 3 and define S = S/(cid:107)S(cid:107) and g = fC/(cid:107)fC(cid:107), where fC is the restriction of f to the subset C. It is easy to see that S satisfies condition (1) and (cid:107)S − T(cid:107) (cid:54) (cid:107)S − S(cid:107) + (cid:107)S − T(cid:107) = (cid:107) ¯S(cid:107) − 1 + (cid:107) ¯S − T(cid:107) (cid:54) 2(cid:107) ¯S − T(cid:107) < ε. sign(cid:0)f (ω)(cid:1)T ∗δs0 (ω)f (ω) dµ (cid:107)g − f(cid:107) (cid:54) (cid:107)g − fC(cid:107) + (cid:107)fC − f(cid:107) = 2(1 − (cid:107)fC(cid:107)) On the other hand, we see that Sg(s0) = (cid:104)S∗δs0, g(cid:105) = 1 because S∗δs0(ω) = sign(cid:0)f (x)(cid:1) = sign(cid:0)g(ω)(cid:1) for f (x)dµ < ε = 2 Ω\C every ω ∈ C. This completes the proof. (cid:3) We do not know, and it is clearly of interest, for which topological compact Hausdorff spaces K all operators in L(L1(µ), C(K)) satisfy condition (1). We recall that a bounded linear operator T from L1(µ) into a Banach space X is said to be Bochner representable if there is a bounded strongly measurable function g : Ω −→ X such that (cid:90) T f = f (ω)g(ω) dµ(ω) (cid:0)f ∈ L1(µ)(cid:1). Moreover, we get 1 − ε2 6 which implies that Therefore, T ∗δs0(ω)f (ω) dµ (cid:90) Ω\C Ω = (cid:90) < T f (s0) = (cid:104)T ∗δs0, f(cid:105) = (cid:90) sign(cid:0)f (ω)(cid:1)T ∗δs0(ω)f (ω) dµ + (cid:90) f (x) dµ + (1 − ε 3 f (x)dµ, Ω\C (cid:54) C C ) = 1 − ε 3 Ω\C f (ω) dµ f (x) dµ < ε 2 . Ω\C (cid:90) (cid:90) (cid:90) (cid:90) THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 17 The Dunford-Pettis-Phillips Theorem [21, Theorem 12 p. 75] says that T ∈ L(L1(µ), X) is weakly compact if and only if T is Bochner representable by a function g which has an essentially relatively weakly compact range. Iwanik [26] showed that every Bochner representable operator from L1(µ) into C(K) satisfies condition (1). Moreover, we get the following. Corollary 5.3. Let 0 < ε < 1. Suppose that T ∈ L(L1(µ), C(K)) (real case) has norm-one and it is Bochner representable (resp. weakly compact). If (cid:107)T f(cid:107) > 1 − ε2 6 for some f ∈ SL1(µ), then there exist a Bochner representable (resp. weakly compact) operator S ∈ L(L1(µ), C(K)) with (cid:107)S(cid:107) = 1 and g ∈ SL1(µ) such that (cid:107)Sg(cid:107) = 1, (cid:107)S − T(cid:107) < ε, and (cid:107)f − g(cid:107) < ε. Proof. By Theorem 5.2, it is enough to show that if T is a Bochner representable operator from L1(µ) into C(K), then F (s) = T ∗δs is continuous in measure and that the operator S defined in the proof is Bochner representable. Let g : Ω −→ C(K) be a bounded strongly measurable function which represents T . It is easy to check that F (s) = g(·)(s) for all s ∈ K. Since the range of g is separable, the range of T is separable and contained in a separable sub-algebra A of C(K) with unit. By the Gelfand representation theorem, A is isometrically isomorphic to C( ¯K) for some compact metrizable space ¯K. So, we may assume that K is metrizable. To show that the mapping F (s) = T ∗δs = g(ω)(s) is continuous in measure, assume that a sequence (sn) converges to s in K. Then for all ω ∈ Ω, By the dominated convergence theorem, we have that lim n→∞g(ω)(sn) − g(ω)(s) = 0. (cid:90) (cid:90) lim n→∞ sup f∈SL∞ (µ) f (ω)(g(ω)(sn) − g(ω)(s)) dµ(ω) (cid:54) lim n→∞ g(ω)(sn) − g(ω)(s) dµ(ω) = 0. Hence the sequence (g(·)(sn))n converges to g(·)(s) in measure. That is, (F (sn))n converges to F (s). We note that the operator ¯S in the proof of Theorem 5.2 is determined by GT ∗δs = G(g(·)(s)). Since the mapping s (cid:55)−→ G(g(·)(s))(ω) = (g(ω)(s) ∧ (1 − ε/3)) ∨ (−1 + ε/3)) is continuous for each ω ∈ Ω, the operator ¯S is Bochner representable by this mapping. Finally, if T is (cid:3) weakly compact, then the proof is done by the Dunford-Pettis-Phillips theorem. As observed in [26], the operator T : L1[0, 1] −→ C[0, 1] determined by T ∗δs = χ[0,s] is not Bochner representable, but satisfies condition (1). For condition (2), we have the following result. Theorem 5.4. Let 0 < ε < 1. Suppose that T ∈ L(L1(µ), C(K)) (real case) has norm-one and satisfies 4 for some f ∈ SL1(µ), then there exist S ∈ L(L1(µ), C(K)) with (cid:107)S(cid:107) = 1 condition (2). If (cid:107)T f(cid:107) > 1 − ε2 and g ∈ SL1(µ) such that (cid:107)Sg(cid:107) = 1, (cid:107)S − T(cid:107) < ε, and (cid:107)f − g(cid:107) < ε. Moreover, S also satisfies condition (2). Proof. By using a suitable isometric isomorphism, we may first assume that f is nonnegative. Let G be the co-meager set in the condition (2) and (T ∗δsk )k be a sequence which is (cid:107) · (cid:107)∞-dense in the closure of {T ∗δs : s ∈ G} ⊂ L∞(µ). Observe that the sets {ω ∈ Ω : a < T ∗δsk (ω) < b} where a, b ∈ Q and k (cid:62) 1, form a countable family {Ai}i of measurable subsets of Ω. We define, for each i, the functions ui(s) = ess. inf{T ∗δs(ω) : ω ∈ Ai} and vi(s) = ess. sup{T ∗δs(ω) : ω ∈ Ai}. 18 CHOI, KIM, LEE, AND MART´IN Let Ui and Vi be the set of all continuity points of ui and vi for all i, respectively. Let F be the intersection of all subsets Ui's and Vi's. We claim that the functions ui's are upper semi-continuous and the functions vi's are lower semi-continuous. Indeed, recall that vi(s) = inf λ ∈ R : µ{ω ∈ Ai : T ∗δs(ω) > λ} = 0 (cid:110) (cid:111) , where inf ∅ = ∞ and inf R = −∞. To show that the set {s : λ < vi(s)} is open in K for all λ ∈ R, suppose that vi(s0) > λ0 for some s0 ∈ K and λ0 ∈ R. It suffice to prove that there is an open neighborhood V of s0 such that V ⊂ {s : vi(s) > λ0}. We note that µ{ω ∈ Ai : T ∗δs0 (ω) > λ0} > 0 and there exists λ1 > λ0 such that µ{ω ∈ Ai : T ∗δs0(ω) > λ1} > 0. (cid:90) Let E = {ω ∈ Ai : T ∗δs0(ω) > λ1}. Then T ∗δs0(ω)dµ(ω) > λ1 > λ0. Since the map s (cid:55)−→ T ∗δs is weak∗ continuous on L∞(µ), the set µ(E) E 1 (cid:26) (cid:90) (cid:27) V := s ∈ K : 1 T ∗δs(ω)dµ(ω) > λ1 µ(E) E is an open subset containing s0. We note that µ{ω ∈ Ai : T ∗δs(ω) > λ1} > 0 for all s ∈ V . Otherwise, there is s1 ∈ V such that µ{ω ∈ Ai : T ∗δs1(ω) > λ1} = 0. Then T ∗δs1 (ω) (cid:54) λ1 almost everywhere ω ∈ Ai and (cid:90) 1 µ(E) E T ∗δs1(ω)dµ(ω) (cid:54) λ1. This is a contradiction to the fact that s1 is an element of V , which implies that vi(s) > λ0 for all s ∈ V and V ⊂ {s : vi(s) > λ0}. This gives the lower semi-continuity of vi. The upper semi-continuity of ui follows from the fact that −ui is lower semi-continuous. The claim is proved. We deduce then that the set F is co-meager (c.f. see [31, § 32 II. p. 400]). Since the set {s : s ∈ 4 . Without 4 } is nonempty and open, there exists s0 ∈ F ∩ G such that T f (s0) > 1− ε2 K, T f (s) > 1− ε2 loss of generality, we may assume that Because of the denseness of the sequence (T ∗δsk )k, there exists k0 ∈ N such that (cid:107) < T f (sk0 ) =(cid:10)T ∗δsk0 (cid:107)T ∗δs0 − T ∗δsk0 and ε 4 . Fix q ∈ Q such that 1 − 3 4 ε < q < 1 − ε T f (s0) = (cid:104)T ∗δs0 , f(cid:105) > 1 − ε2 4 . 4 2 and let , f(cid:11) > 1 − ε2 C =(cid:8)ω ∈ Ω : T ∗δsk0 (ω) > q(cid:9) . <(cid:10)T ∗δsk0 (cid:90) (cid:90) T ∗δsk0 (cid:17)(cid:90) (cid:90) (ω)f (ω) dµ T ∗δsk0 , f(cid:11) = (ω)f (ω) dµ + (cid:16) (cid:90) Ω\C Ω = (cid:54) f (ω) dµ + 1 − ε 2 f (ω) dµ Ω\C T ∗δsk0 (cid:90) C C =1 − ε 2 f (ω) dµ. Ω\C f (ω) dµ < ε 2 and Ω\C (cid:90) C f (ω) dµ > 1 − ε 2 . (ω)f (ω) dµ Then Hence we have that 1 − ε2 4 (cid:90) THE BISHOP-PHELPS-BOLLOB ´AS THEOREM FOR OPERATORS ON (cid:32)L1(µ) 19 Let Bn = {ω : q < T ∗δsk0 (ω) < n} for each n. Then C =(cid:83)∞ (cid:90) n=1 Bn and there exists n0 such that Hence Bn0 = Ai0 for some i0 and µ(Ai0 ) > 0. This implies that ui0(sk0) (cid:62) q and ui0(s0) (cid:62) q − ε Setting A = Ai0, it is also clear that 4 > 1− ε. f (ω) dµ > 1 − ε 2 . Bn0 (cid:13)(cid:13)(cid:13)(cid:13) fA(cid:107)fA(cid:107) − f (cid:13)(cid:13)(cid:13)(cid:13) < ε. Since ui0 is continuous at s0, there exist an open neighborhood U of s0 and a continuous function h : K −→ [0, 1] such that ui0(s) > 1 − ε for all s ∈ U , h(s0) = 1 and h(U c) = 0. We define a weak∗-continuous map M : K −→ L∞(µ) by (cid:0)ω ∈ Ω, s ∈ K(cid:1). M (s)(ω) = T ∗δs(ω) + χA(ω)h(s)(1 − T ∗δs(ω)) We note that M (s0) = 1 for all ω ∈ A. It is also easy to get that (cid:107)M (s)(ω) − T ∗δs(ω)(cid:107) = (cid:107)χA(ω)h(s)(1 − T ∗δs(ω))(cid:107) < ε and (cid:107)M (s)(cid:107) = 1. sup s∈K Let S be the operator represented by the function M . Then S satisfies condition (2), S and (cid:107)S − T(cid:107) < ε. (cid:16) fA(cid:107)fA(cid:107) (cid:17) (s0) = 1 (cid:3) by T f (s) =(cid:82) f (n)πn(s) dµ(n), where πn be the n-th natural projection on {0, 1}N As shown in [26], the Dunford-Pettis-Phillips Theorem implies that every weakly compact operator T in from L1(µ) to an arbitrary Banach space Y has separable range, hence the range of its weakly compact adjoint T ∗ is also separable and so T satisfies condition (2). On the other hand, there are Indeed, let µ be a Bochner representable operators which do not satisfy the condition (2) (see [26]). strictly positive probability measure on N and consider the operator T ∈ L(L1(µ), C({0, 1}N ) defined . Then T is Bochner representable, while {T ∗δs : s ∈ G} is non-separable in L∞(µ) for every uncountable subset G of {0, 1}N . Finally, let us comment that it is also observed in [26] that if K has a countable dense subset of isolated points, then condition (2) is automatically satisfied for all T ∈ L(L1(µ), C(K)). Actually, in this case, C(K) has the so-called property (β) and then the pair (X, C(K)) has the BPBp for all Banach spaces X [3, Theorem 2.2]. It would be of interest to characterize those topological Hausdorff compact spaces K such that (X, C(K)) has the BPBp for every Banach space X. References [1] M. D. Acosta, Denseness of norm attaining mappings, Rev. R. Acad. Cien. Serie A. Mat. 100 (2006), 9 -- 30. [2] M. D. Acosta, F. J. Aguirre and R. Pay´a, There is no bilinear Bishop-Phelps theorem, Israel J. Math. 93 (1996), 221-227. [3] M. D. Acosta, R. M. Aron, D. Garc´ıa and M. Maestre, The Bishop-Phelps-Bollob´as theorem for operators, J. Funct. Anal. 254 (2008), 2780-2799. [4] M. D. Acosta, J. Becerra-Guerrero, Y. S. Choi, M. Ciesielski, S. K. Kim, H. J. Lee, M. L. Lorenco¸ and M. Mart´ın, The Bishop-Phelps-Bollob´as property for operators between spaces of continuous functions, Preprint (2013). [5] M. D. Acosta, J. Becerra-Guerrero, D. Garc´ıa and M. Maestre, The Bishop-Phelps-Bollob´as Theorem for bilinear forms, Trans. Amer. Math. Soc. (to appear). [6] M. D. Acosta, J. Becerra-Guerrero, D. Garc´ıa, S. K. Kim, and M. Maestre, Bishop-Phelps-Bollob´as property for certain spaces of opertors, preprint (2013). [7] M. D. Acosta, D. Garc´ıa and M. Maestre, A multilinear Lindenstrauss theorem, J. Funct. Anal. 235 (2006), 122-136. [8] R. M. Aron, B. Cascales and O. Kozhushkina, The Bishop-Phelps-Bollob´as theorem and Asplund operators, Proc. [9] R. M. Aron, Y. S. Choi, D. Garc´ıa and M. Maestre, The Bishop-Phelps-Bollob´as Theorem for L(L1(µ), L∞[0, 1]), Amer. Math. Soc. 139 (2011), 3553-3560. Adv. Math. 228 (2011), 617 -- 628. [10] R. M. Aron, Y. S. Choi, S. K. Kim, H. J. Lee and M. Mart´ın, The Bishop-Phelps-Bollob´as version of Lindenstrauss properties A and B, Preprint (2013). 20 CHOI, KIM, LEE, AND MART´IN [11] R. Aron, C. Finet and E. Werner, Some remarks on norm attaining N-linear forms, In: Functions Spaces (K. Jarosz, Ed.), Lecture Notes in Pure and Appl. Math. 172, Marcel-Dekker, NewYork, 1995, 19-28. [12] E. Bishop and R. R. Phelps, A proof that every Banach space is subreflexive, Bull. Amer. Math. Soc. 67 (1961), 97-98. [13] B. Bollob´as, An extension to the Theorem of Bishop and Phelps, Bull. London Math. Soc. 2 (1970), 181-182. [14] J. Bourgain, On dentability and the Bishop-Phelps property, Israel J. Math. 28 (1977), 265-271. [15] B. Cascales, A. J. Guirao, and V. Kadets, A Bishop-Phelps-Bollob´as type theorem for uniform algebras, Preprint. [16] Y. S. Choi, Norm attaining bilinear forms on L1[0, 1], J. Math. Anal. Appl. 211 (1997), 295-300. [17] Y. S. Choi and S. G. Kim, Norm or numerical radius attaining multilinear mappings and polynomials, J. London Math. Soc. 54 (1) (1996), 135-147. [18] Y. S. Choi and S. K. Kim, The Bishop-Phelps-Bollob´as theorem for operators from L1(µ) to Banach spaces with the Radon-Nikod´ym property, J. Funct. Anal. 261 (2011), 1446-1456. [19] Y. S. Choi and S. K. Kim, The Bishop-Phelps-Bollob´as property and lush spaces, J. Math. Anal. Appl. 390 (2012), 549-555. [20] A. Defant and K. Floret, Tensor Norms and Operator Ideals, North-Holland Math. Studies 176, Elsevier, Amster- dam, 1993. [21] J. Diestel and J. J. Uhl, Jr., Vector measures, Mathematical Surveys 15, AMS, 1977. [22] J. L. Doob, Measure Theory. Graduate Text in Math. 143, Springer, 1993. [23] N. Dunford and J. Schwartz, Linear operators, Vol. I, Interscience, N. Y., 1958. [24] C. Finet and R. Pay´a, Norm attaining operators from L1 into L∞, Israel J. Math. 108 (1998), 139-143. [25] A. Iwanik, Norm attaining operators on Lebesgue spaces, Pacific J. Math. 83 (1979), 381-386. [26] A. Iwanik, On norm-attaining operators acting from L1(µ) to C(S), Proceedings of the 10th Winter School on Abstract Analysis (Srn´ı, 1982) Rend. Circ. Mat. Palermo. 2 (1982), 147-152. [27] J. Johnson and J. Wolfe, Norm attaining operators, Studia Math. 65 (1979), 7-19. [28] S. K. Kim, The Bishop-Phelps-Bollob´as Theorem for operators from c0 to uniformly convex spaces, Israel. J. Math. (to appear). [29] S. K. Kim and H. J. Lee, Uniform convexity and Bishop-Phelps-Bollob´as property, Canadian J. Math. (to appear). [30] J. Kim and H. J. Lee, Strong peak points and strongly norm attaining points with applications to denseness and polynomial numerical indices, J. Funct. Anal. 257 (2009), 931-947. [31] K. Kuratowski, Topology, Vol 1, Academic press, 1966. [32] H. E. Lacey, The isometric theory of classical Banach spaces, Springer, 1974. [33] J. Lindenstrauss, On operators which attain their norm, Israel J. Math. 1 (1963), 139-148. [34] J. R. Partington, Norm attaining operators, Israel J. Math. 43 (1982), 273-276. [35] R. Pay´a and Y. Saleh, Norm attaining operators from L1(µ) into L∞(ν), Arch. Math. 75 (2000), 380-388. [36] W. Schachermayer, Norm attaining operators and renorming of Banach spaces, Israel J. Math. 44 (1983), 201-212. [37] W. Schachermayer, Norm attaining operators on some classical Banach spaces, Pacific J. Math. 105 (1983), 427-438. [38] H. H. Schaefer, Banach lattices and positive operators, Springer, 1974. (Choi) Department of Mathematics, POSTECH, Pohang (790-784), Republic of Korea E-mail address: [email protected] (Kim) School of Mathematics, Korea Institute for Advanced Study (KIAS), 85 Hoegiro, Dongdaemun-gu, Seoul 130-722, Republic of Korea E-mail address: [email protected] (Lee) Department of Mathematics Education, Dongguk University - Seoul, 100-715 Seoul, Republic of Korea E-mail address: [email protected] (Mart´ın) Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, E-18071 Granada, Spain E-mail address: [email protected]
1511.07308	1	1511	2015-11-23T16:53:29	Schatten class and Berezin transform of quaternionic linear operators	[ "math.FA" ]	In this paper we introduce the Schatten class of operators and the Berezin transform of operators in the quaternionic setting. The first topic is of great importance in operator theory but it is also necessary to study the second one because we need the notion of trace class operators, which is a particular case of the Schatten class. Regarding the Berezin transform, we give the general definition and properties. Then we concentrate on the setting of weighted Bergman spaces of slice hyperholomorphic functions. Our results are based on the S-spectrum for quaternionic operators, which is the notion of spectrum that appears in the quaternionic version of the spectral theorem and in the quaternionic S-functional calculus.	math.FA	math	Schatten class and Berezin transform of quaternionic linear operators Fabrizio Colombo Politecnico di Milano Jonathan Gantner Politecnico di Milano Dipartimento di Matematica Dipartimento di Matematica Via E. Bonardi, 9 20133 Milano, Italy Via E. Bonardi, 9 20133 Milano, Italy [email protected] [email protected] Tim Janssens University of Antwerp Department of Mathematics Middelheimlaan, 1 2020 Antwerp, Belgium [email protected] Abstract In this paper we introduce the Schatten class of operators and the Berezin transform of operators in the quaternionic setting. The first topic is of great importance in operator theory but it is also necessary to study the second one because we need the notion of trace class operators, which is a particular case of the Schatten class. Regarding the Berezin transform, we give the general definition and properties. Then we concentrate on the setting of weighted Bergman spaces of slice hyperholomorphic functions. Our results are based on the S-spectrum for quaternionic operators, which is the notion of spectrum that appears in the quaternionic version of the spectral theorem and in the quaternionic S-functional calculus. AMS Classification: 47A10, 47A60, 30G35. Key words: Schatten class of operators, Spectral theorem and the S-spectrum, Berezin transform of quater- nionic operators, Weighted Bergman spaces of slice hyperholomorphic functions. 1 Introduction The study of quaternionic linear operators was stimulated by the celebrated paper of G. Birkhoff and J. von Neumann [13] on the logic of quantum mechanics, where they proved that there are essentially two possible ways to formulate quantum mechanics: the well known one using complex numbers and also one using quaternions. The development of quaternionic quantum mechanics has been done by several authors in the past. Without claiming completeness, we mention: S. Adler, see [1], L. P. Horwitz and L. C. Biedenharn, see [31], D. Finkelstein, J. M. Jauch, S. Schiminovich and D. Speiser, see [26], and G. Emch see [23]. The notion of spectrum of a quaternionic linear operator used in the past was the right spectrum. Even though 1 this notion of spectrum gives just a partial description of the spectral analysis of a quaternionic operator, basically it contains just the eigenvalues, several results in quaternionic quantum mechanics have been obtained. Only in 2007, one of the authors introduced with some collaborators the notion of S-spectrum of a quaternionic operator, see the book [20]. This spectrum was suggested by the quaternionic version of the Riesz-Dunford functional calculus (see [21,32] for the classical results), which is called S-functional calculus or quaternionic functional calculus and whose full development was done in [12,18,19]. Using the notion of right spectrum of a quaternionic matrix, the spectral theorem was proved in [24] (see [22] for the classical result). The spectral theorem for quaternionic bounded or unbounded normal operators based on the S-spectrum has been proved recently in the papers [3, 5]. The case of compact operators is in [29]. The literature contains several attempts to prove the quaternionic version of the spectral theorem, but the notion of spectrum in the general case was not made clear, see [34, 37]. The fact that the S-eigenvalues and the right eigenvalues are equal explains why some results in quaternionic quantum mechanic were possible to be obtained without the notion of S-spectrum. More recently, the use of the S-spectrum can be found in [35]. The aim of this paper is to introduce the Schatten class of quaternionic operators and the Berezin transform. In the following, a complex Hilbert space will be denoted by HC while we keep the symbol H for a quaternionic Hilbert space. We recall some facts of the classical theory in order to better explain the quaternionic setting. Suppose that B is a compact self-adjoint operator on a complex separable Hilbert space HC. Then it is well known that there exists a sequence of real numbers {λn} tending to zero and there exists an orthonormal set {un} in HC such that Bx = ∞Xn=1 λnhx, uniun, for all x ∈ HC. When A is a compact operator, but not necessarily self-adjoint, we consider the polar decomposition A = UA, where A = √A∗A is a positive operator and U is a partial isometry. From the spectral representation of compact self-adjoint operators and the polar decomposition of bounded linear operators, we get the representation of the compact operator A as Ax = ∞Xn=1 λnhx, univn, for all x ∈ HC, where vn := U un is an orthonormal set. The positive real numbers {λn}n∈N are called singular values of A. Associated with the singular values of A, we have the Schatten class operators Sp that consists of those operators for which the sequence {λn}n∈N belongs to ℓp for p ∈ (0,∞). In the case p ∈ [1,∞) the sets Sp are Banach spaces. Two important spaces are p = 1, the so called trace class, and p = 2, the Hilbert-Smith operators. The Schatten class operators were apparently introduced to study integral equations with non symmetric kernels as pointed out in the book of I. C. Gohberg and M. G. Krein [30], later on it was discovered that the properties of the singular values play an important role in the construction of the general theory of symmetrically normed ideals of compact operators. Let us point out that in the quaternionic setting, it has been understood just recently what the correct notion of spectrum is that replaces the classical notion of spectrum of operators on a complex Banach space. This spectrum is called S-spectrum and its is also defined for n-tuples of noncommuting operators, see [20]. Precisely, let V be a bilateral quaternionic Banach space and denote by B(V ) the Banach space of all quaternionic bounded linear operators on V endowed with the natural norm. Let T : V → V be a quaternionic bounded linear operator (left or right linear). We define the S-spectrum of T as σS(T ) = {s ∈ H : T 2 − 2Re(s)T + s2I is not invertible in B(V )} 2 where s = s0 + s1e1 + s2e2 + s2e3 is a quaternion and e1, e2 and e3 is the standard basis of the quaternions H, Re(s) = s0 is the real part and s2 is the squared Euclidean norm. The S-resolvent set is defined as ρS(T ) = H \ σS(T ). The notion of S-spectrum for quaternionic operators arises naturally in the slice hyperholomorphic func- tional calculus, called S-functional calculus or quaternionic functional calculus, which is the quaternionic analogue of the Riesz-Dunford functional calculus for complex operators on a complex Banach space. Re- cently, it turned out that also the spectral theorem for quaternionic operators (bounded or unbounded) is based on the S-spectrum. This fact restores for the quaternionic setting the well known fact that both the Riesz-Dunford functional calculus and the spectral theorem are based on the same notion of spectrum. This was not clear for a long time because the literature related to quaternionic linear operators used the notion of right spectrum of a quaternionic operator, which, for a right linear operator T , is defined as σR(T ) = {s ∈ H such that ∃ 0 6= v ∈ V : T v = vs}. This definition gives only rise to the eigenvalues of T . The other possibility is to define the left spectrum σL(T ), where we replace T v = vs by T v = sv. In the case of the right eigenvalues we have a nonlinear operator associated with the spectral problem. In the case of the left spectrum, we have a linear equation, but in both cases there does not exist a suitable notion of resolvent operator which preserves some sort of hyperholomorphicity. Only in the case of the S-spectrum, we can associate to it the two S-resolvent operators. For s ∈ ρS(T ), the left S-resolvent operator and the right S-resolvent operator S−1 L (s, T ) := −(T 2 − 2Re(s)T + s2I)−1(T − sI), S−1 R (s, T ) := −(T − sI)(T 2 − 2Re(s)T + s2I)−1. are operator-valued slice hyperholomorphic functions. Moreover, with the notion of S-spectrum, the usual decompositions of the spectrum turn out to be natural, for example in point spectrum, continuous spectrum and residual spectrum. So one of the main aims of this paper is to define the Schatten classes of quaternionic operators using the notion of S-spectrum and to study their main properties. Here we point out the following fact that shows one of the main differences with respect to the complex case. The trace of a compact complex linear operator is defined as tr A =Xn∈Nhen, Aeni, (1) where (en)n∈N is an orthonormal basis of the Hilbert space HC. It can be shown that the trace is independent of the choice of the orthonormal basis and therefore well-defined. If we try to generalize the above definition to the quaternionic setting, we face the problem that this invariance with respect to the choice of the orthonormal basis does not persist. We have to define the trace in a different way. In fact we have to restrict to operators that satisfy an additional condition: Let J ∈ B(H) be an anti-selfadjoint and unitary operator. For p ∈ (0, +∞], we define the (J, p)-Schatten class of operators Sp(J) as Sp(J) := {T ∈ B0(H) : [T, J] = 0 and (λn(T ))n∈N ∈ ℓp}, where (λn(T ))n∈N denotes the sequence of singular values of T . The Berezin transform is a useful tool to study operators acting on spaces of holomorphic functions such as 3 Bergman spaces. In this paper we treat the case of quaternionic operators acting on spaces of slice hyper- holomorphic functions. We recall some facts of the classical theory, see [33] for more details. To state the definition of the Berezin transform of a bounded operator we recall that: A reproducing kernel Hilbert space on an open set Ω of C, is a Hilbert space HC of functions on Ω such that for every w ∈ Ω the linear functional f → f (w) is bounded on HC. When HC is a reproducing kernel Hilbert space, by the Riesz representation theorem, there exists a unique Kw ∈ HC for every w ∈ Ω such that f (w) = hf, Kwi for all f ∈ HC. The function Kw is called a reproducing : ℓ ∈ L} is an orthonormal basis for HC. Since Kw(z) = Kz(w), we write K(z, w) := Kw(z). The norm of Kw is given by kernel. It is a well known fact that Kw(z) =Pℓ∈L eℓ(w)eℓ(z) where {eℓ and when this norm is different from zero the function kKwk2 = hKw, Kwi = Kw(w) = K(w, w) kw = Kw kKwk is called normalized reproducing kernel at w. Examples of reproducing kernel Hilbert spaces are Hardy spaces, Bergman spaces, Fock spaces and Dirichlet spaces to name a few. Keeping in mind the above definition, we define the Berezin transform of an operator as follows: Let HC be a reproducing kernel Hilbert space of analytic functions on an open set Ω of C and let A be a bounded operator on HC. The Berezin transform of A is defined as eA(w) := hKw, AKwi, for all w ∈ Ω. The importance of the Berezin transform is due to the possibility to deduce properties of the bounded operator A from the properties of the analytic function eA(w). For example, an important consequence is that A = 0 if and only if eA(w) = 0 for all w ∈ Ω. When we consider quaternionic operators, we have to replace the notion of holomorphic function with the notion of slice hyperholomorphic function. Using the theory of function spaces of slice hyperholomorphic functions, it is possible to extend classical results to the quaternionic setting. We treat the case of weighted Bergman spaces on the unit ball D in the quaternions, whose reproducing kernel is defined as follows: consider α > −1 and q, w ∈ D. For q = q0 + iqq1 set qiw = q0 + iwq1, where w = w0 + iww1, and define the slice hyperholomorphic α-Bergman kernel as Kα(q, w) := 1 2 (1 − iqi) 1 (1 − qiw w)2+α + 1 2 (1 + iqi) 1 (1 − qiw w)2+α . This kernel is left slice hyperholomorphic in q and anti right slice hyperholomorphic in w. Moreover, whenever q and w belong to the same complex plane, it reduces to the complex Bergman kernel. The plan of the paper is as follows:. Section 1 contains the introduction. Section 2 contains a subsection in which we discussion two of the possible definitions of slice hyperholomorphic functions and their properties, and a subsection that contains the main results on quaternionic operator theory. Section 3 contains the Schatten class of quaternionic operators. Precisely, we consider the singular values of compact operators, the Schatten class, the dual of the Schatten class and different characterizations of Schatten class operators. Section 4 contains the general theory of the Berezin transform and the case of weighted Bergman spaces with several properties. 4 2 Preliminary results Recent works on Schur analysis in the slice hyperholomorphic setting introduced and studied the quaternionic Hardy spaces H2(Ω), where Ω is the quaternionic unit ball D or the half space H+ of quaternions with positive real part, [7, 8, 11]. The Hardy spaces H p(B) for p > 2 are considered in [36]. The Bergman spaces are treated in [15 -- 17] and for the Fock spaces see [10]. Weighted Bergman spaces, Bloch, Besov and Dirichlet spaces on the unit ball B are studied in [14]. In the following we want discuss two possible definitions of slice hyperholomorphic functions since in both cases the above mentioned function spaces can be defined with minor changes in the proofs. The discussion in the next subsection will clarify the variations of slice hyperholomorphicity. 2.1 Slice hyperholomorphic functions The skew-field of quaternions consists of the real vector space H :=(ξ0 + 3Xℓ=1 ξℓeℓ : ξℓ ∈ R) , where the units e1, e2 and e3 satisfy e2 i = −1 and eℓeκ = −eℓeκ for ℓ, κ ∈ {1, 2, 3} with ℓ 6= κ. The real part ℓ=1 ξℓeℓ and its ℓ=1 ξℓeℓ is defined as Re(x) := ξ0, its imaginary part as x :=P3 of a quaternion x = ξ0 +P3 conjugate as x := Re(x) − x. Each element of the set S := {x ∈ H : Re(x) = 0,x = 1} is a square-root of −1 and is therefore called an imaginary unit. For any i ∈ S, the subspace Ci := {x0 + ix1 : x1, x2 ∈ R} is isomorphic to the field of complex numbers. For i, j ∈ S with i ⊥ j, set k = ij = −ji. Then 1, i, j and k form an orthonormal basis of H as a real vector space and 1 and j form an orthonormal basis of H as a left or right vector space over the complex plane Ci, that is Any quaternion x belongs to such a complex plane: if we set H = Ci + Cij and H = Ci + jCi. ix :=(x/x, any i ∈ S, if x 6= 0 if x = 0, then x = x0 + ixx1 with x0 = Re(x) and x1 = x. The set [x] := {x0 + ix1 : i ∈ S}, is a 2-sphere, that reduces to a single point if x is real. We recall a well known result that we will need in the sequel. Lemma 2.1. Let x, y ∈ H. Then y ∈ [x] if and only if there exists q ∈ H with q = 1 such that y = q−1xq. The notion of slice hyperholomorphicity is the generalization of holomorphicity to quaternion-valued functions that underlies the theory of quaternionic linear operators. We recall the main results on slice hyperholomorphic functions. The proofs of the results stated in this subsection can be found in the book [20]. Definition 2.2. A set U ⊂ H is called (i) axially symmetric if [x] ⊂ U for any x ∈ U and (ii) a slice domain if U is open, U ∩ R 6= 0 and U ∩ Ci is a domain for any i ∈ S. 5 Definition 2.3. Let U ⊂ H be an axially symmetric open set. A real differentiable function f : U → H is called left slice hyperholomorphic if it has the form such that the functions α and β, which take values in H, satisfy the compatibility condition f (x) = α(x0, x1) + ixβ(x0, x1), ∀x = x0 + ixx1 ∈ U and the Cauchy-Riemann-system α(x0, x1) = α(x0, x1) β(x0, x1) = −β(x0,−x1) ∂ ∂x0 ∂ ∂x0 α(x0, x1) = β(x0, x1) β(x0, x1) = − α(x0, x1). ∂ ∂x1 ∂ ∂x1 A function f : U → H is called right slice hyperholomorphic if it has the form f (x) = α(x0, x1) + β(x0, x1)ix, ∀x = x0 + ixx1 ∈ U, such that the functions α and β satisfy (3) and (4). (2) (3) (4) (5) The sets of left and right slice hyperholomorphic functions on U are denoted by SHL(U ) and SHR(U ), respectively. Finally, we say that a function f is left or right slice hyperholomorphic on a closed axially symmetric set K, if there exists an open axially symmetric set U with K ⊂ U such that f ∈ SHL(U ) resp. SHR(U ). Corollary 2.4. Let U ⊂ H be axially symmetric. (i) If f, g ∈ SHL(U ) and a ∈ H, then f a + g ∈ SHL(U ). (ii) If f, g ∈ SHR(U ) and a ∈ H, then af + g ∈ SHR(U ). On axially symmetric slice domains, slice hyperholomorphic functions can be characterized as those functions that lie in the kernel of a slicewise Cauchy-Riemann-operator. As a consequence, the restriction of a slice hyperholomorphic function to a complex plane can be split into two holomorphic components. Definition 2.5. Let f be a function defined on a set U ⊂ H and let i ∈ S. We denote the restriction of f to the complex plane Ci by fi, i.e. fi := fU∩Ci. In order to avoid confusion we point out that, in this paper, indices i and j always refer to restrictions of a function to the respective complex planes. Definition 2.6. Let U ⊂ H be open. We define the following differential operators: for any real differentiable function f : U → H and any x = x0 + ixx1 ∈ U , we set ∂if (x) = ∂if (x) = fix(x) − ix fix(x) + ix and f (x)∂← i = ∂x0 ∂x0 1 1 2(cid:18) ∂ 2(cid:18) ∂ 2(cid:18) ∂ 2(cid:18) ∂ 1 1 ∂x0 ∂x1 ∂ ∂ ∂x1 fix(x)(cid:19) fix(x)(cid:19) fi(x)ix(cid:19) fi(x)ix(cid:19) , fi(x) − fi(x) + ∂ ∂x1 ∂ ∂x1 ∂x0 where the arrow ← indicates that the operators ∂← i f (x)∂ ← i = and ∂← i act from the right. If a function depends on several variables and we want to stress that these operators act in a variable x, then we write ∂ix instead of ∂i etc. 6 i = 0. If U is a slice domain, then f ∈ SHR(U ) if and only if f ∂← Corollary 2.7. Let U ⊂ H be open and axially symmetric. (i) If f ∈ SHL(U ), then ∂if = 0. If U is a slice domain, then f ∈ SHL(U ) if and only if ∂if = 0. (ii) If f ∈ SHR(U ), then f ∂← Lemma 2.8 (Splitting Lemma). Let U ⊂ H be axially symmetric and let i, j ∈ S with i ⊥ j. (i) If f ∈ SHL(U ), then there exist holomorphic functions f1, f2 : U ∩ Ci → Ci such that fi = f1 + f2j. (ii) If f ∈ SHR(U ), then there exist holomorphic functions f1, f2 : U ∩ Ci → Ci such that fi = f1 + jf2. Remark 2.9. Originally, in particular in [20], slice hyperholomorphic functions were defined as functions that satisfy ∂if = 0 resp. f ∂← = 0. In principle, this leads to a larger class of functions, but on axially i symmetric slice domains both definitions are equivalent. Indeed, on an axially symmetric slice domain, ∂if = 0 resp. f ∂← = 0 implies that f satisfies the representation formula (cf. Theorem 2.15), which allows i for a representation of f of the form (2) resp. (5). = 0. i The theory of slice hyperholomorphicity was therefore only developed for functions that are defined on axially symmetric slice domains. However, most results do not directly depend on the fact that the functions are defined on a slice domain, but rather on their representation of the form (2) resp. (5). Hence, the definition given in this paper seems to be more appropriate since it allows to extend the theory also to functions defined on sets that are not connected or do not intersect the real line. Definition 2.10. Let U ⊂ H be axially symmetric. A left slice hyperholomorphic f (x) = α(x0, x1) + ixβ(x0, x1) is called intrinsic if α and β are real-valued. We denote the set of intrinsic functions on U by N (U ). Note that an intrinsic function is both left and right slice hyperholomorphic because β(x0, x1) commutes with the imaginary unit ix. The converse is not true: the constant function x 7→ b ∈ H \ R is left and right slice hyperholomorphic, but it is not intrinsic. The importance of the class of intrinsic functions is due to the fact that the multiplication and com- position with intrinsic functions preserve slice hyperholomorphicity. This is not true for arbitrary slice hyperholomorphic functions. Corollary 2.11. Let U ⊂ H be axially symmetric. (i) If f ∈ N (U ) and g ∈ SHL(U ), then f g ∈ SHL(U ). If f ∈ SHR(U ) and g ∈ N (U ), then f g ∈ SHR(U ). (ii) If g ∈ N (U ) and f ∈ SHL(g(U )), then f ◦ g ∈ SHL(U ). If g ∈ N (U ) and f ∈ SHR(g(U )), then f ◦ g ∈ SHR(U ). series of the formP+∞ Important examples of slice hyperholomorphic functions are power series with quaternionic coefficients: n=0 anxn are right slice hyperholomorphic on their domain of convergence. A power series is intrinsic if and only if its coefficients are real. n=0 xnan are left slice hyperholomorphic and series of the formP∞ Conversely, any slice hyperholomorphic function can be expanded into a power series at any real point. Definition 2.12. The slice-derivative of a function f ∈ SHL(U ) is defined as ∂Sf (x) = lim Cix ∋s→x (s − x)−1(f (s) − f (x)), where limCix ∋s→x g(s) is the limit as s tends to x = x0 + ixx1 ∈ U in Cix . The slice-derivative of a function f ∈ SHR(U ) is defined as f ∂← S (f (s) − f (x))(s − x)−1. (x) = lim Cix ∋s→x 7 If a function depends on several variables and we want to stress that we take the slice-derivative in the variable x, then we write ∂Sx and ∂ ← Sx instead of ∂S and ∂← S . Corollary 2.13. The slice derivative of a left (or right) slice hyperholomorphic function is again left (or right) slice hyperholomorphic. Moreover, it coincides with the derivative with respect to the real part, that is ∂Sf (x) = ∂ ∂x0 f (x) resp. f ∂← S (x) = ∂ ∂x0 f (x). Theorem 2.14. If f is left slice hyperholomorphic on the ball Br(α) with radius r centered at α ∈ R, then f (x) = +∞Xn=0 (x − α)n 1 n! ∂n S f (α) for x ∈ Br(α). If f is right slice hyperholomorphic on B(r, α), then f (x) = 1 n! +∞Xn=0 f ∂n S ← (α)(x − α)n for x ∈ Br(α). As a consequence of the slice structure of slice hyperholomorphic functions, their values are uniquely determined by their values on an arbitrary complex plane. Consequently, any function that is holomorphic on a suitable subset of a complex plane possesses an unique slice hyperholomorphic extension. Theorem 2.15 (Representation Formula). Let U ⊂ H be axially symmetric and let i ∈ S. For any x = x0 + ixx1 ∈ U set xi := x0 + ix1. If f ∈ SHL(U ). Then If f ∈ SHR(U ), then f (x) = 1 2 (1 − ixi)f (xi) + 1 2 (1 + ixi)f (xi). f (x) = f (xi)(1 − iix) 1 2 + f (xi)(1 + iix) 1 2 . Corollary 2.16. Let i ∈ S and let f : O → H be real differentiable, where O is a domain in Ci that is symmetric with respect to the real axis. (i) The axially symmetric hull [O] :=Sz∈O[z] of O is an axially symmetric slice domain. ∂ f = 0, then there exists a unique left slice hyperholomorphic extension extL(f ) f + i ∂ ∂x1 ∂x0 (ii) If f satisfies of f to [O]. (iii) If f satisfies ∂ ∂x0 f + ∂ ∂x1 extR(f ) of f to [O]. f i = 0, then there exists a unique right slice hyperholomorphic extension Remark 2.17. If f has a left and a right slice hyperholomorphic extension, they do not necessarily coincide. Consider for instance the function z 7→ bz on Ci with a constant b ∈ Ci \ R. Its left slice hyperholomorphic extension to H is x 7→ xb, but its right slice hyperholomorphic extension is x 7→ bx. 2.2 Quaternionic linear operators In this section, we consider bounded linear operators on a separable quaternionic Hilbert space, even thought some definitions and results hold also for quaternionic Banach spaces. Definition 2.18. Let H be a quaternionic right vector space together with a scalar product h·,·i : H×H → H with the following properties: 8 (i) positivity: hx, xi ≥ 0 and hx, xi = 0 if and only if x = 0 (ii) right-linearity: hx, ya + zi = hx, yia + hx, zi for all x, y, z ∈ H and all a ∈ H. (iii) quaternionic hermiticity: hx, yi = hy, xi for all x, y ∈ H. If H is complete with respect to the norm then (H,h·,·i) is called a quaternionic Hilbert space. kxkH :=phx, xi, Terms such as orthogonality, orthonormal basis etc. are defined as in the complex case. In the following, we shall always assume that H is separable. We consider now bounded quaternionic right linear operators on H. Definition 2.19. A quaternionic right linear operator T : H → H is called bounded if kTk := supkxk=1 kT xk < +∞. We denote the set of all bounded quaternionic right linear operators on H by B(H). The space B(H) together with the operator norm is a real Banach space. Observe that there does not exists any natural multiplication of operators with scalars in H \ R because there is no left-multiplication defined on H. Hence, the operator T s, which is supposed to act as T s(v) = T (sv), has no meaning. Let T ∈ B(H). For s ∈ H, we set Qs(T ) := T 2 − 2Re(s)T + Is2, where I denotes the identity operator. Definition 2.20. We define the S-resolvent set of an operator T ∈ B(H) as and the S-spectrum of T as ρS(T ) := {s ∈ H : Qs(T )−1 ∈ B(H)} σS(T ) := H \ ρS(T ). Remark 2.21. For operators acting on a two-sided quaternionic Banach space, the left and right S-resolvent operators are defined by formally replacing the variable x in the left and right Cauchy kernel by the operator T . This motivates the definition of the S-resolvent set and the S-spectrum as it is done in Definition 2.20, cf. [20]. Theorem 2.22 (See [20]). Let T ∈ B(H). Then σS(T ) is an axially symmetric, compact and nonempty subset of BkT k(0). In [28], the following natural partition of the S-spectrum was introduced: 1. the spherical point spectrum of T : σpS(T ) := {s ∈ H : ker(Qs(T )) 6= {0}}. 2. the spherical residual spectrum of T : σrS(T ) := {s ∈ H : ker(Qs(T )) = {0}, ran(Qs(T )) 6= H) 3. the spherical continuous spectrum of T : σcS(T ) := {s ∈ H : ker(QS(T )) = {0}, ran(Qs(T )) = H, Qs(T )−1 /∈ B(H)}. 9 Theorem 2.23. Let T ∈ B(H). Then s ∈ σpS(T ) if and only if s is a right eigenvalue of T , i.e. there exits x ∈ H such that T x = xs. The adjoint of an operator and related properties are defined analogue to the complex case. Definition 2.24. Let T ∈ B(H). The adjoint T ∗ of T is the unique operator that satisfies hT ∗x, yi = hx, T yi ∀x, y ∈ H. Definition 2.25. An operator T ∈ B(H) is called (i) selfadjoint, if T ∗ = T (ii) anti-selfadjoint, if T ∗ = −T (iii) positive if hx, T xi ≥ 0 for all x ∈ H (iv) normal if T ∗T = T T ∗ (v) unitary if T ∗ = T −1. If T ∈ B(H) is positive, then there exists a unique positive operator S such that S 2 = T . We denote √T := S. Moreover, for any operator T ∈ B(H), the operator T ∗T is positive and we can define T := √T ∗T , which allows to prove the polar decomposition theorem. Theorem 2.26. Let H be a quaternionic Hilbert space and let T ∈ B(H). Then there exist two unique operators W and P in B(H) such that (i) T = W P (ii) P ≥ 0 (iii) ker(P ) ⊂ ker(W ) (iv) ∀u ∈ ker(P )⊥ : kW uk = kuk. The operators W and P have the following properties (a) P = T (b) If T is normal then W defines a unitary operator in B(ran(T )) (c) W is (anti) self-adjoint if T is. The above theorem is mentioned several times in the literature, for a proof see [28]. In order to investigate linear operators on a quaternionic Hilbert space, it is often useful to define a complex structure on the space that allows to write every vector in terms of two components that belong to a certain complex Hilbert space. If this complex structure is chosen appropriately, then the considered operator is the natural extension of a complex linear operator on the component space. Definition 2.27 (See G. Emch [23]). Let J ∈ B(H) be anti-selfadjoint and unitary and let i ∈ S. We define the complex subspaces HJi + := {x ∈ H : Jx = xi} and HJi − := {x ∈ H : Jx = −xi}. 10 Theorem 2.28. HJi + is a nontrivial complex Hilbert space over Ci with respect to the structure induced by H: its sum is the sum of H, its complex scalar multiplication is the right scalar multiplication of H restricted to Ci and its complex scalar product is the restriction of the scalar product of H to HJi + . An analogous statement holds true for HJi − . Theorem 2.29. Every orthonormal basis (en)n∈N of the complex Hilbert space HJi basis of H. Moreover, it is + is also an orthonormal + ×HJi Jx =Xn∈N enihen, xi, ∀x ∈ H. For a proof of the above two results see [28]. Observe that Theorem 2.29 implies that we can write x ∈ H as x = x1 + x2j with x1, x2 ∈ HJi + . Moreover, it justifies considering J as a left multiplication with the imaginary unit i ∈ S. We may set ix := Jx for any x ∈ H and extend this left scalar multiplication to the entire complex field Ci by setting (a0 + ia1)x = xa0 + Jxa1. Since ix = Jx = Jx0 + Jx1j = x0i + x1ij, we then obtain ax = x0a + x1aj for a ∈ Ci. Note that the choice of the imaginary unit i is arbitrary, but that different imaginary units will lead to different left scalar multiplications, which are of course only defined for scalars in the respective complex plane! We discuss now the relation between quaternionic linear operators on H and complex linear operators + . In order to avoid confusion, we distinguish complex linear operators by a C-subscript. on HJi Theorem 2.30. Let J be an anti-selfadjoint, unitary operator on H and take i ∈ S. For every bounded Ci-linear operator TC ∈ B(HJi + ), there exists a unique right H-linear operator fTC ∈ B(H) that satisfies ∀x ∈ HJi + . The following facts also hold: fTCx = TCx, (a) We have kfTCk = kTCk. (b) It is [fTC, J] = 0, where [fTC, J] :=fTCJ − JfTC denotes the commutator of fTC and J. (c) An operator V ∈ B(H) is equal to fTC for some Ci-linear operator TC ∈ B(HJi (d) We have (fTC)∗ =fT ∗ (e) ]SCTC =fSCfTC. (f ) If SC is a right (resp. left) inverse of TC, then fSC is a right (resp. left) inverse of fTC. Furthermore, given another Ci-linear operator SC ∈ B(HJi + ), we have C. modern language. The above theorem generalizes some results of G. Emch in Section 3 of [23] and formulates them in a + ) if and only if [J, V ] = 0. Definition 2.31. For an anti-selfadjoint unitary operator J, we define the set BJ (H) := {T ∈ B(H) : [T, J] = 0}. Remark 2.32. For T ∈ BJ (H) and a = a0 + ia1 ∈ Ci set aT = a0T + a1JT . If we consider J as a left-multiplication with some imaginary unit i ∈ S, then BJ (H) turns into a Banach algebra over Ci that is isometrically isomorphic to B(HJi + ). Isometric isomorphims between these two spaces are given by ResJi :(BJ (H) → B(HJi 7→ T(cid:12)(cid:12)HJ i T + + ) LiftJi :(B(HJi TC 7→fTC + ) → BJ(H) , (6) where fTC is the unique operator obtained from Theorem 2.30. 11 Corollary 2.33. An operator T ∈ BJ (H) is selfadjoint, anti-selfadjoint, positive, normal or unitary if and only if ResJi(T ) is. 3 Schatten classes of quaternionic linear operators In order to introduce Schatten classes of quaternionic linear operators, we need to define the singular values of compact quaternionic linear operators. For analogue results in the complex case see [38]. 3.1 Singular values of compact operators Definition 3.1. A right linear operator is called compact if it maps bounded sequences to sequences that admit convergent subsequences. We denote the set of all compact quaternionic right linear operators on H by B0(H). We recall the spectral theorem for compact quaternionic operators (see [29]); note that this is a special case of the spectral theorem for arbitrary normal operators, which was recently established in [3]. Theorem 3.2. Given a normal operator T ∈ B0(H) with spherical point spectrum σpS(T ) there exists a Hilbert basis N ⊂ H made of eigenvectors of T such that zλzhz, xi ∀x ∈ H, (7) T x = Xz∈N where λz ∈ H is an eigenvalue relative to the eigenvector z and if λz 6= 0 then there are only a finite number of distinct z′ ∈ N such that λz = λz′ . Moreover the values λz are at most countably many and 0 is their only possible accumulation point. Recall that we are considering only separable Hilbert spaces. Hence, N is always countable and we can write N = {en : n ∈ N}. Remark 3.3. Observe that the eigenvalues (λn)n∈N in the spectral decomposition of a compact quaternionic linear operator are in general not unique. Indeed, if T en = enλn and q ∈ H with q = 1, then T (enq) = (T en)q = enλnq = (enq)q−1λnq and hence the eigenpair (en, λn) can be replaced by (enq, q−1λnq). However, q−1λnq ∈ [λn] and consequently at least the eigenspheres [λn] are uniquely determined. In particular, the eigenvalues can be chosen such that λn ∈ C+ Remark 3.4. Using the spectral theorem, one can define a functional calculus for so-called slice functions. We just need the special case of fractional powers of a positive operator compact operator T : consider its := {s = s0 + is1 ∈ Ci : s1 ≥ 0} for a given imaginary unit i ∈ S, cf. Lemma 2.1. i spectral decomposition T =Pn∈N enλnhen,·i. For p > 0, the operator T p is defined as T p =Xn∈N enλnhen,·i. Consider now an arbitrary compact operator T . Then the operator T is normal and, combining Theo- rem 3.2 applied to T with the polar decomposition, Theorem 2.26, we are capable of finding a Hilbert-basis (en)n∈N and an orthonormal set (σn)n∈N in H such that σnλnhen, xi ∀x ∈ H, (8) T x =Xn∈N where the λn ∈ R+ are the eigenvalues of the operator T in decreasing order, the vectors (en)n∈N form an eigenbasis of T and σn = W en with W unitary such that T = WT. 12 Definition 3.5. We call the set {λn}n∈N the set of singular values of T and the representation (8) the singular value decomposition of T . The following Rayleigh's equation gives a characterization of the singular values of T . Since the proof follows the lines of the complex case, we omit it and refer to [22]. Lemma 3.6. Let T be a positive, compact operator on H and let enλnhen, xi T x =Xn∈N be the spectral decomposition of T , where the eigenvalues {λn}n∈N are given in descending order. Then λn+1 = min y1,...,yn max hx,yii=0 i=1,...,n hx, T xi kxk2 . Also in the quaternionic setting we have: Lemma 3.7. Let T be a positive, compact operator on H. Then λn+1 = min y1,...,yn kT xk. max hx,yii=0 i=1,...,n kxk=1 (9) (10) Proof. In fact λn+1 = min y1,...,yn hx, T xi. max hx,yii=0 i=1,...,n kxk=1 using the Cauchy-Schwarz inequality we see that hx, T xi ≤ kxkkT xk and since we assumed in the above lemma that T is positive, we can consider its spectral decomposition (9) in order to see that λn+1 = min y1,...,yn hx, T xi ≤ min y1,...,yn kT xk ≤ max kT xk max hx,yii=0 i=1,...,n kxk=1 max hx,yii=0 i=1,...,n kxk=1 s Xm≥n+1 = max hx,eii=0 i=1,...,n kxk=1 ≤ λn+1. mhx, emi2 ≤ λn+1 max λ2 hx,eii=0 i=1,...,n kxk=1 hx,eii=0 i=1,...,n kxk=1 s Xm≥n+1 hx, emi2 The singular values of an arbitrary compact operator T are the eigenvalues of the positive operator T. Since the eigenvalues of T can be obtained via the formula (10) and since kT xk = kWTxk = kTxk by Theorem 2.26, the singular values of T also satisfy (10). The formula (10) also allows us to deduce the following corollary, just as in the complex case, cf. [38, Theorem 1.34]. Corollary 3.8. Let T be a compact operator on H. Its singular values satisfy λn+1 = inf F ∈Fn kT − Fk, (11) where Fn is the set of all linear operators on H with rank less than or equal to n. 13 The following proposition is an immediate consequence of (10), cf. [22, Corollary XI.9.3] for the complex case. Corollary 3.9. Let T1 and T2 be compact linear operators on H. Then, for every nonnegative n, m ∈ N0, we have that and λn+m+1(T1 + T2) ≤ λn+1(T1) + λm+1(T2) λn+m+1(T1T2) ≤ λn+1(T1)λm+1(T2). 3.2 Definition of the Schatten class For p ∈ (0, +∞) the Schatten p-class Sp(HC) of operators on a complex Hilbert space HC is defined as the set of compact operators whose singular value sequences are p-summable. For p ≥ 1, the Schatten p-norm, which assigns to each operator the ℓp-norm of its singular value sequence, turns Sp(HC) into a Banach space. However, the analogue approach does not make sense in the quaternionic setting. Major problems appear in particular when trying to define the trace of an operator, which plays a crucial role in the classical theory. The trace of a compact complex linear operator is defined as tr A =Xn∈Nhen, Aeni, (12) where (en)n∈N is an orthonormal basis of the Hilbert space HC. It can be shown that the trace is independent of the choice of the orthonormal basis and therefore well-defined. In general, this is not true in the quaternionic setting. Consider for example a compact normal operator T and its spectral decompositions T =Xn∈N enλnhen,·i =Xn∈N enλnhen,·i, where the orthonormal system (en)n∈N is such that the corresponding eigenvalues (λn)n∈N belong to the com- i and the orthonormal system (en)n∈N is such that the corresponding eigenvalues (λn)n∈N plex halfplane C+ belong to the complex halfplane C+ j with i, j ∈ S and i 6= j, cf. Remark 3.3. Moreover, assume that at least one eigenvalue has nonzero imaginary part. Then λn ∈ C+ i \ R, λn ∈ C+ j \ R. Xn∈N Xn∈N but Obviously, it is hT en, eni =Xn∈N hT en, eni =Xn∈N λnhen, eni =Xn∈N λnhen, eni =Xn∈N Xn∈NhT en, eni 6=Xn∈NhT en, eni. In contrast to the classical theory, we have to restrict ourselves to operators that satisfy an additional restriction: compatibility with a chosen complex left-multiplication. Definition 3.10. Let J ∈ B(H) be an anti-selfadjoint and unitary operator. For p ∈ (0, +∞], we define the (J, p)-Schatten class of operators Sp(J) as Sp(J) := {T ∈ B0(H) : [T, J] = 0 and (λn(T ))n∈N ∈ ℓp}, 14 where (λn(T ))n∈N denotes the sequence of singular values of T and ℓp denotes the space of p-summable resp. bounded sequences. For T ∈ Sp(J) , we define kTkp = Xn∈Nλn(T )p! 1 p if p ∈ (0, +∞) (13) and kTkp = sup n∈N λn(T ) = kTk if p = +∞. Obviously, Sp(J) is a subset of BJ (H), see Definition 2.31. The following lemma shows that it is isomet- rically isomorphic to the Schatten p-class on HJi Lemma 3.11. Let J be a unitary and anti-selfadjoint operator and let i ∈ S. For all T ∈ B0(H) such that [T, J] = 0 and for each n ∈ N, we have that + if J is considered as a left-multiplication with i ∈ S. λn(T ) = λ′ n (ResJi(T )) , n(AC) is the n-th singular value of a Ci-linear operator AC ∈ B(HJi + ). where λ′ Proof. Since T is compact, its restriction ResJi(T ) : HJi can find orthonormal sets (en)n∈N and (σn)n∈N in HJi + → HJi + such that + is a compact operator on HJi + . Hence, we T x =Xn∈N σnλ′ n(cid:16)T(cid:12)(cid:12)HJ i +(cid:17)hen, xi ∀x ∈ HJi + . By the uniqueness of the extension, it follows that this expression holds for all x ∈ H and we deduce λn(T ) = λ′ n(ResJi(T )). Corollary 3.12. Let J be an anti-selfadjoint unitary operator on H and consider it as a left-multiplication with i ∈ S. For any p ∈ (0, +∞], the space Sp(J) is isomorphic to Sp(HJi + . An isomorphism between these spaces is given by T 7→ ResJi(T ). Moreover, kTkp = k ResJi(T )kp. In particular, if p ∈ [1, +∞], then k · kp is actually a norm and Sp(J) is a Banach space over Ci that is even isometrically isomorphic to Sp(HJi + ). + ), the Schatten p-class on HJi Finally, as an immediate consequence of Corollary 3.9, we obtain as in the complex case that any Sp(J) is an ideal of BJ (H). Corollary 3.13. Let p ∈ (0, +∞]. Then Sp(J) is a two-sided ideal of BJ (H), i.e. ST and T S belong to Sp(J) whenever T ∈ Sp(J) and S ∈ BJ (H). 3.3 The dual of the Schatten class In the following, we fix a unitary, anti-selfadjoint operator J and consider it as a left-multiplication with i ∈ S. We define the trace of an operator T ∈ S1(J) and establish some elementary results in order to determine the dual space of Sp(J). Definition 3.14. We define the Ji-trace of an operator T ∈ S1(J) as TrJi(T ) := tr (ResJi(T )) where tr (ResJi(T)) denotes the classical trace of a complex linear operator ResJi(T ) as defined in (12). 15 Corollary 3.15. If T ∈ S1(J), then TrJi(T ) =Xn∈N hen, T eni for any orthornormal basis (en)n∈N of HJi + . (14) If T is positive or selfadjoint, then also in the quaternionic setting there are no restrictions on the choice of the orthonormal basis in (14). Lemma 3.16. If T is a positive and compact operator on H with singular values (λn)n∈N then Xn∈N λn =Xn∈N hen, T eni for each orthonormal basis (en)n∈N of H. If in particular T ∈ S1(J), then (14) holds true for any orthonormal basis of H. Proof. Since T is compact and positive, we have T =Pn∈N ηnλnhηn,·i for some orthonormal basis (ηn)n∈N of H by Theorem 3.2. For any orthonormal basis (en)n∈N of H, we therefore have hen, T eni = Xm∈N λmhηm, eni2. Using Fubini's theorem and Parseval's identity kxk2 =Pn∈N hen, xi2, we are left with Xn∈Nhen, T eni =Xn∈NXm∈N λmhηm, eni2 = Xm∈N λmXn∈N hηm, eni2 = Xm∈N λm. Corollary 3.17. Let T ∈ S1(J) be selfadjoint. Then (14) holds true for any orthonormal basis of H. Proof. By Corollary 2.33, the operator ResJi(T ) is a bounded selfadjoint operator on HJi + and can therefore be decomposed into ResJi(T ) = T+,C − T−,C with positive operators T+,C, T−,C ∈ B(HJi + ). If we set T± = LiftJi(T±,C), then T = T+ − T− with positive operators T+, T− ∈ BJ (H). The additivity of the Ji-trace and Lemma 3.16 imply TrJi(T ) = TrJi(T+) − TrJi(T−) =Xn∈Nhen, T+eni −Xn∈Nhen, T−eni =Xn∈Nhen, T eni for any orthonormal basis (en)n∈N of H. Observe that TrJi(T ) of course depends on the imaginary unit i ∈ S. However, as the next result shows, the choice of the imaginary unit i has no essential impact. Lemma 3.18. Let T ∈ S1(J) and let i, j in S. Then TrJj(T ) = φ(TrJi(T )), where φ is the isomorphism φ(z0 + iz1) = z0 + jz1 between the complex fields Ci and Cj. 16 Proof. Let (en)n∈N be an orthonormal basis of HJi Lemma 2.1. Then φ(z) = q−1zq for any z ∈ Ci. Moreover, (enq)n∈N is an orthonormal basis of HJj + and let q ∈ H with q = 1 such that j = q−1iq, cf. + as J(enq) = (Jen)q = eniq = (enq)q−1iq = (enq)j. Since q = 1, we have q−1 = q and thus, by Corollary 3.15, we obtain φ(TrJi(T )) = q−1 TrJi(T )q = +∞Xn=0 henq, T enqi = TrJj(T ). Lemma 3.19. Suppose 1 ≤ p < +∞ and 1 (i) T S and ST belong to the trace class S1(J) p + 1 q = 1. If T ∈ Sp(J) and S ∈ Sq(J), then (ii) TrJi(T S) = TrJi(ST ) (iii) TrJi(T S) ≤ kTkpkSkq. Proof. Applying Corollary 3.12, we can reduce the statement to the case of operators on a complex Hilbert space, where we know that these results are true. Hence, T ∈ Sp(J) and S ∈ Sq(J) ⇐⇒ ResJi(T ) ∈ Sp(HJi + ) and ResJi(S) ∈ Sq(HJi + ) =⇒ ResJi(T S) ∈ S1(HJi + ) ⇐⇒ T S ∈ S1(J). The second statement follows from the definition: TrJi(T S) = tr (ResJi(T S)) = tr (ResJi(T ) ResJi(S)) = tr (ResJi(S) ResJi(T )) = tr (ResJi(ST )) = TrJi(ST ). The final statement relies on the fact that ResJi and LiftJi are p-norm preserving: TrJi(T S) = tr (ResJi(T S)) = tr (ResJi(T ) ResJi(S)) ≤ k ResJi(T )kpk ResJi(S)kq = kTkpkSkq, which finishes the proof of the lemma. Lemma 3.20. Suppose 1 ≤ p < +∞ and 1 p + 1 q = 1. If T ∈ Sp(J), then kTkp = sup{ trJi(ST ) : kSkq = 1, S ∈ Sq(J)} . 17 Proof. Using the fact that this result is true for operators on a complex Hilbert space and that ResJi is a p-norm preserving isomorphism, we find that kTkp = k ResJi(T )kp = sup(cid:8) tr ((SC ResJi(T )) : kSCkq = 1, SC ∈ Sq(HJi + )(cid:9) = sup{ tr ((ResJi(ST )) : k ResJi(S)kq = 1, S ∈ Sq(J)} = sup{ TrJi(ST ) : kSkq = 1, S ∈ Sq(J)} . We can also use the fact that Sp(J) is isometrically isomorphic to the Schatten p-class of operators on the complex Hilbert space HJi Theorem 3.21. If 1 ≤ p < +∞ and 1 + to determine its dual space. p + 1 q = 1, then Sp(J)∗ = Sq(J) with equal norms and under the pairing hT, Si = TrJi(T S). Proof. First, assume that ξ is a continuous linear functional on Sp(J). By Corollary 3.12, the mapping TC 7→ ξ(LiftJi(TC)) is a linear functional on Sp(HJi with kSCkq = kξ ◦ LiftJi k such that + ) under the pairing hTC, SCi = tr(TCSC), there exists an operator SC ∈ Sq(HJi + ) + ) with kξ ◦ LiftJi k = kξk. Since Sp(HJi + )∗ ∼= Sq(HJi We define now Sξ := LiftJi(SC). Then ξ ◦ LiftJi(TC) = tr(TCSC), ∀TC ∈ HJi + . and kSξkq = kSCkq = kξ ◦ LiftJi k = kξk ξ(T ) = ξ ◦ LiftJi ◦ ResJi(T ) = tr(ResJi(T )SC) = tr(ResJi(T Sξ)) = TrJi(T Sξ). Hence, the mapping Φ : ξ 7→ Sξ is an isometric Ci-linear mapping of Sp(J)∗ into Sq(J). Conversely, it follows from the Ci-linearity of the Ji-trace and (iii) of Lemma 3.19 that, for S ∈ Sq(J), the mapping T 7→ ξS(T ) := TrJi(T S) is a bounded Ci-linear functional on Sp(J) with kξSk ≤ kSkq. Consequently, Φ is even invertible and in turn Sp(J)∗ is isometrically isomorphic to Sq(J). 3.4 Characterizations of Schatten class operators In this section we take a closer look at the singular values of Schatten class operators in order to arrive at necessary and sufficient conditions for an operator to belong to the Schatten class. A crucial observation was made in Corollary 3.12: Lemma 3.22. Let T be a positive and compact operator on H and p ∈ (0, +∞). Then T ∈ Sp(J) ⇐⇒ ResJi(T ) ∈ Sp(HJi + ). Moreover, kTkp p = kT pk1. T ∈ Sp ⇐⇒ T p ∈ S1. 18 Proof. Let T =Pn∈N enhen,·iλn be a spectral decomposition of T . Since the eigenvalues λn are positive, they coincide with the singular values of T . Moreover, as in the case of complex operators, we have σS(T ) = {λn : n ∈ N}∪{0}, cf. [25]. The S-spectral theorem implies that σS(T p) = {λp n : n ∈ N}∪{0}. Consequently, the singular value sequence of T p is (λp n)n∈N, and hence kT pk1 =Xn∈N λp n = kTkp p. Theorem 3.23. If T is a compact operator on H such that [T, J] = 0 and p ∈ (0, +∞) then T ∈ Sp(J) ⇐⇒ Tp = (T ∗T ) 2 ∈ S1(J) ⇐⇒ T ∗T ∈ S p 2 (J). p Moreover As a consequence, we have that kTkp p = kTkp p = kTpk1 = kT ∗Tk p 2 p 2 . Proof. We start by proving the first equivalence (the other ones follow analogously from the corresponding results for complex linear operators, cf. [38, Theorem 1.26]): T ∈ Sp(J) ⇐⇒ T ∈ Sp(J). T ∈ Sp(J) ⇐⇒ ResJi(T ) ∈ Sp(HJi + ) ⇐⇒ ResJi(T )p = ResJi(Tp) ∈ S1(HJi + ) ⇐⇒ Tp ∈ Sp(J). For the equality of the norms we prove the first one (the rest is proven in a similar way): kTkp p = k ResJi(T )kp p = k ResJi(T )kp p = k ResJi(T)kp p = kTkp p. Since ResJi and LiftJi are p-norm preserving, we easily obtain further characterizations of (J, p)-Schatten class operators from the respective results in the complex case, cf. [38, Theorems 1.27 -- 1.29]. Theorem 3.24. Suppose that T is a compact operator on a quaternionic Hilbert-space H with [T, J] = 0 and that p ≥ 1. Then T is in Sp(J) if and only if hen, T enip < +∞ Xn∈N hen, T enip# 1 p for all orthonormal sets (en)n∈N in HJi + . If T is also selfadjoint then kTkp = sup "Xn∈N : {en} orthonormal set in HJi . + Theorem 3.25. Suppose that T is a compact operator on H with [T, J] = 0 and that p ∈ [1, +∞). Then T is in Sp(J) if and only if for all orthonormal sets {en} and {σn} in HJi + . If T is also positive then hσn, T enip < +∞ Xn∈N hσn, T enip# 1 p : {en} and {σn} orthonormal sets in HJi + . 19 kTkp = sup "Xn∈N Theorem 3.26. Suppose that T is a compact operator on H with [T, J] = 0 and that 0 < p ≤ 2. Then, for any orthonormal basis {en} of HJi + , we have kTkp p ≤ ∞Xn=1 ∞Xk=1 hek, T enip. Proposition 3.27. Suppose T is a positive, compact operator on H and x is a unit vector in H. Then • hx, T pxi ≥ hx, T xip for all p ∈ [1, +∞). • hx, T pxi ≤ hx, T xip for all 0 < p ≤ 1. Proof. Let T x =Pn∈N enλnhen, xi be a spectral decomposition for the operator T . Then, for all p > 0 and x ∈ H, we have by the spectral theorem that and thus We also know that for every Hilbert basis (ξn)n∈N of H we have that ∀x ∈ H. hξn, xi2, Let us first assume that p ≥ 1 and let q be its conjugate index. Applying Holder's inequality gives enλp nhen, xi λp nhen, xi2. T px =Xn∈N hx, T pxi =Xn∈N kxk2 =Xn∈N hx, T xi = Xn∈N = Xn∈N ≤ Xn∈N λp λnhen, xi = hx, T pxi 1 p λnhen, xi2! 2 phen, xi 2 q! nhen, xi2! 1 p Xn∈Nhen, xi2! 1 q and since p ≥ 1 taking the p-th power preserves the inequality. pair ( 1 If 0 < p ≤ 1 then we can find q ≥ 1 such that: p + 1 p , q) gives us that q = 1. Using the Holder inequality with the conjugate hx, T pxi =Xn∈N =Xn∈N ≤ Xn∈N = hx, T xip λp nhen, xi2 λp nhen, xi2phen, xi 2 q λnhen, xi2!p Xn∈Nhen, xi2! 1 q and this finishes the proof. 20 Corollary 3.28. Suppose that T is a positive, compact operator on H such that [T, J] = 0 and that (en)n∈N is an orthonormal basis of H. If p ∈ [1, +∞), then the condition hen, T enip < +∞ Xn∈N is necessary for T ∈ Sp(J). If 0 < p ≤ 1 then this condition is sufficient. Proof. From Lemma 3.22 and Lemma 3.16, we have T ∈ Sp(J) ⇐⇒ T p ∈ S1(J) ⇐⇒ Xn∈Nhen, T peni < +∞. Applying Proposition 3.27 gives the result. Theorem 3.29. Suppose T is a compact operator on H with [T, J] = 0 and p ≥ 2, then T ∈ Sp(J) ⇐⇒ Xn∈NkT enkp < +∞ for all orthonormal sets {en} in HJi + . Moreover, kTkp = sup "Xn∈N p kT enkp# 1 {en} orthonormal in HJi . + Proof. Since ResJi is p-norm preserving, this follows immediately from the corresponding results for complex linear operators [38, Theorem 1.33]. 4 The Berezin Transform Let H be a quaternionic reproducing kernel Hilbert space of functions on the unit ball, that is a Hilbert space of left slice hyperholomorphic functions in the unit ball D ⊂ H with the property that for each w ∈ D the point evaluation f 7→ f (w) is a bounded right linear functional on H. From the Riesz representation theorem we know that there exists a unique function Kw ∈ H such that The function f (w) = hKw, fi, ∀f ∈ H. K(q, w) := Kw(q), q, w ∈ D is called the reproducing kernel of H. Lemma 4.1. The functions {Kw w ∈ D} span the entire space H. Proof. This follows immediately from the reproducing property. If f ⊥ Kw for all w ∈ D, then and thus f = 0. f (w) = hKw, fi = 0 Since we consider slice hyperholomorphic functions, we can use their specific structure to prove a stronger result. 21 Lemma 4.2. Let i ∈ S and set Di := D ∩ Ci. The set {Kω : ω ∈ Di} spans the space H. Proof. If f ⊥ Kw for any w ∈ Di, then f (w) = hKw, fi = 0 for all w ∈ Di. The representation formula, Theorem 2.15, then implies f ≡ 0, and hence H = span{Kw : w ∈ Di}. The kernel K(q, w) is obviously left slice hyperholomorphic in q. Furthermore, it is right slice hyperholo- morphic in w, i.e. the mapping w 7→ K(q, w) is right slice hyperholomorphic. This follows from Corollary 2.7 because K(q, w)∂ ← iw = = = 1 2hKq, Kwi(∂w0 + iw∂w1 ) 1 (∂w0 − iw∂w1 )hKw, Kqi 2 1 (∂w0 + iw∂w1 )Kq(w) = 0. 2 Pointing out that K(q, w)2 ≤ K(q, q)K(w, w), we can conclude that, if for each q ∈ D there exists f ∈ H such that f (q) 6= 0, then kKqk2 = K(q, q) and We shall assume this to be true in the following. We can then normalize the reproducing kernels to obtain a family of unit vectors kq by K(q, q) > 0 ∀q ∈ D. kq(w) = for w ∈ D. K(w, q) pK(q, q) For the following discussion, we fix an imaginary unit i ∈ S. Furthermore, we assume that there exists a We call these the normalized reproducing kernels of H. unitary and antiselfadjoint operator J such that HJ i Definition 4.3. Let T be a bounded linear operator on H such that [T, J] = 0. The function + = spanCi{Kq, q ∈ Di}. is called the Berezin transform of T . eT (q) = hkq, T kqi, q ∈ Di Proposition 4.4. The Berezin transform has the following properties: (i) If T is self-adjoint, then eT is real-valued. (ii) If T is positive, then eT is non-negative. (iii) We have fT ∗ = eT . (iv) The mapping T 7→ eT is a contractive Ci-linear mapping from BJ (H) into L∞(Di, H). Proof. For q ∈ Di, it is fT ∗(q) = hkq, T ∗kqi = hT kq, kqi = hkq, T kqi = eT (q) and hence (iii) holds. If T is self-adjoint, this implies eT (q) = eT (q) and so eT (q) is real. For positive T , the definition of positivity immediately implies (ii). 22 The Berezin transform is obviously R-linear. Since T ∈ BJ (H), it maps kq ∈ HJ i + to an element in HJ i + , Thus, the Berezin transform is even Ci-linear. and hence its Berezin transform eT takes values in Ci. For iT = JT , we have fiT (q) = hkq, JT kqi = hkq, T kqii = eT (q)i = ieT (q). eT (q) = hkq, T kqi ≤ kkqkkT kqk ≤ kTkkkqk2 = kTk. Finally, we deduce from the Cauchy-Schwarz-inequality that Hence, eT ∈ L∞(D) with keTk∞ ≤ kTk. Observe that the Berezin Transform of T ∈ BJ (H) coincides with the classical Berezin transform of the operator ResJi(T ) ∈ BJ (H). Since the restriction operator ResJi and the classical Berezin transform are injective (cf. [38, Proposition 6.2.]), we immediately obtain the following Lemma. Lemma 4.5. The Berezin-transform is one-to-one. 5 The Case of Weighted Bergman Spaces In this section we consider the special case of weighted Bergman spaces on the unit ball. A first study of these spaces in the slice hyperholomorphic setting has been done in [14]. We recall the main definitions and results. Definition 5.1. Let i ∈ S and let dmi be the Lebesgue measure on the complex plane Ci. For α > −1, we define the measure dAα,i(z) on the unit ball Di := D ∩ Ci in Ci by dAα,i(z) = α + 1 π (1 − z2)α dmi(z). For p > 0, the weighted slice Bergman space Ap hyperholomorphic functions f on D such that α,i(D) is the quaternionic right vector space of all left slice For f ∈ Ap α,i(D), we define ZDi f (z)p dAα,i(z) < +∞. kfkp,α,i :=(cid:18)ZDi f (z)p dAα,i(z)(cid:19) 1 p . Corollary 5.2. Let i, j ∈ S with i ⊥ j, let f ∈ SHL(D) and write fi = f1 + f2j with holomorphic functions f1, f2 : Di → Ci, cf. Lemma 2.8. Then f ∈ Ap α,i(D) if and only if f1 and f2 belong to the complex Bergman C,α(D), i.e. the space of all holomorphic functions g on Di such thatRDi g(z)p dAα,i(z) < +∞. space Ap Lemma 5.3. Let α > −1, p > 0 and i, j ∈ S. A left slice hyperholomorphic function f belongs to Ap if and only if it belongs to Ap α,i(D) α,j(D). Moreover, kfkp p,α,i ≤ 2max{p,1}kfkp p,α,j ≤ 22 max{p,1}kfkp p,α,i. Definition 5.4. For α > −1 and p > 0, we define the weighted slice hyperholomorphic Bergman space Ap α(D) as the space of all left slice hyperholomorphic functions f such that kfkp,α := sup i∈S kfkp,α,i < +∞. Remark 5.5. Observe that Corollary 5.2 implies that, for each i ∈ S, the spaces Ap the same elements and that their norms are equivalent. α(D) and Ap α,i(D) contain 23 5.1 Further properties of the weighted slice Bergman space A2 As in the complex space, the norm on the slice Bergman space A2 α,i(D) α,i(D) is generated by the scalar product hf, gi2,α,i :=ZDi f (z)g(z) dAα,i(z). (15) α(D) is generated by the chosen scalar product. This is however not true for the slice hyperholomorphic Bergman space A2 α(D): orthogonality and other concepts related to the scalar product will always depend on the complex plane chosen to define the scalar product. Nevertheless, by Lemma 5.3, independently of the choice of the complex plane, the norm topology on Ap Lemma 5.6. Endowed with the scalar product defined in (15), the slice Bergman space A2 a reproducing kernel quaternionic Hilbert space. Proof. If f ∈ A2 α,i(D) and w ∈ D, then we can chose j ∈ S with j ⊥ iw and write the restriction of f to the plane Ciw as fiw = f1 + f2j with holomorphic functions f1, f2 : Diw → Ciw . By Corollary 5.2, the functions f1 and f2 belong to the complex Bergman space A2 C,α(Di). Since point evaluations are continuous functionals on A2 C,α(Di), cf. [38, Theorem 4.14], we have α,i(D) turns into f (w) ≤ f1(w) + f2(w) ≤ C(kf1kC,2,α + kf2kC,2,α) ≤ 2Ckfk2,α,iω ≤ eCkfk2,α,i, where the last equality follows from the equivalence of the Bergman slice norms, cf. Lemma 5.3. Hence, point evaluations are continuous linear functionals on A2 α,i(D). Definition 5.7. Let α > −1. We define the slice hyperholomorphic α-Bergman kernel for q, w ∈ D as Kα(q, w) := 1 2 (1 − iqi) 1 (1 − qiw w)2+α + 1 2 (1 + iqi) 1 (1 − qiw w)2+α . Remark 5.8. Observe that Kα(q, w) is an extension of the complex Bergman kernel KC,α(z, w) = 1 (1 − zw)2+α . Whenever q and w belong to the same complex plane, Kα(q, w) = KC,α(q, w). Moreover, by a direct computation it is easy to see that the kernel Kα(x, w) is left slice hyperholomophic in q and right slice hyperholomorphic in w. Lemma 5.9. The function Kα(·,·) is the reproducing kernel of A2 Proof. Let f ∈ A2 hKw, fi2,α,i = f (w). First consider w ∈ Di. Since KwDi is nothing but the complex Bergman kernel KC,α(·,·), we immediately obtain from Corollary 5.2 that Kw belongs to A2 α,i(D) and that we can write fi = f1 + f2j with f1, f2 ∈ A2 C,α(Di). From Remark 5.8 and the reproducing property of KC,α(·,·) in A2 α,i(D) and set Kw(q) := Kα(q, w) for w ∈ D. We show that Kw ∈ A2 C,α(D), we obtain α,i(D) and α,i(D). KC,α(z, w)f2(z) dAα,i(z)j hKw, fi2,α,i =ZDi =ZDi Kα(z, w)f (z) dAα,i(z) KC,α(z, w)f1(z) dAα,i(z) +ZDi = f1(w) + f2(w)j = f (w). 24 If w /∈ Di, then Theorem 2.15 implies Kw = Kwi(1 − iiw) 1 2 + Kwi(1 + iiw) 1 2 , because of the right slice hyperholomorphicity of Kα(q, w) in w. Hence, the function Kw belongs to A2 because it is a right linear combination of the functions Kwi and Kwi, which belong to A2 argumentation. From the representation formula, we finally also deduce α,i(D) α,i(D) by the above hKw, fiα,i = = 1 2 1 2 (1 − iiw)hKwi, fiα,i + (1 − iwi)f (wi) + 1 2 (1 + iiw)hKwi, fiα,i (1 + iwi)f (wi) = f (w). 1 2 5.2 The Berezin transform on A2 α,i(D) In this section we consider the following fixed unitary and anti-selfadjoint operator Jf := ext(ifi), ∀f ∈ A2 α,i(D) and consider it as a left scalar multiplication with i ∈ S. Corollary 5.10. It is (A2 α,i(D))Ji + = spanCi(Kq : q ∈ Di) ∼= A2 C,α(Di). Proof. Write fi for f ∈ A2 linearity of the extension operator, we then have f = ext(f1) + ext(f2)j. From α,i(D) as fi = f1 + f2j with components that are holomorphic on Di. By the right Jf = ext(ifi) = ext(f1i) + ext(f2ij) = ext(f1)i − ext(f2)ji, we deduce that f ∈ (A2 α,i(D))Ji is therefore an isometric isomorphism between (A2 Since ϕ(Kq) = KC,q and spanCi{KC,q, q ∈ Di} is dense in A2 α,i(D))Ji + if and only if f = ext(f1), i.e. if and only if f2 = 0. The mapping ϕ : f 7→ fi C,α(Di). α,i(D))Ji + and A2 C,α(Di) and so (A2 C,α(Di), we obtain that + ∼= A2 A2 α,i(D) = ϕ−1(cid:0)A2 C,α(Di)(cid:1) = ϕ−1(cid:16)spanCi{KC,q, q ∈ Di}(cid:17) = spanCi{ϕ−1(KC,q), q ∈ Di} = spanCi{Kq, q ∈ Di}. Recall that a bounded linear operator T on A2 some bounded linear operator on TC on the complex Bergman space A2 Theorem 5.11. Let T be a bounded positive operator on A2 values. If where fTC denotes the Berezin transform of the complex operator TC. π(1 − z2)2 dmi dµi(z) = 1 25 α,i(D) satisfies [T, J] = 0 if and only if T = LiftJi(TC) for C,α(Di). In this case eT (z) = fTC(z), α,i(D) and let (λn)n∈N be its sequence of singular is the Ci-Mobius invariant area measure on Di, then Proof. Let (en)n∈N be any orthonormal basis of A2 Set S = √T . Fubini's theorem, the reproducing property of Kz and Parseval's identity imply n=1 λn =P+∞ n=0hen, T eni. +∞Xn=1 hen, T eni = λn = (α + 1)ZDi T (z) dµi(z). +∞Xn=1 kSenk2 = α,i(D). By Lemma 3.16, we haveP+∞ +∞Xn=1ZDi Sen(z)2 dAα,i(z) +∞Xn=1 +∞Xn=1 Sen(z)2 dAα,i(z) =ZDi =ZDi =ZDi hen, SKzi2 dAα,i(z) =ZDi kSKzk2 dAα,i(z) =ZDihKz, T Kzi dAα,i(z) =ZDi eT (z)K(z, z) dAα,i(z) = (α + 1)ZDi eT (z) dµi(z). +∞Xn=1 +∞Xn=1 hKz, Seni2 dAα,i(z) We have the following important consequence: Corollary 5.12. Let J be any unitary anti-selfadjoint operator on A2 BJ (A2 Corollary 5.13. If T ∈ S1(J), then eT is in L1(Di, dµi) and α,i(D)) belongs to the trace class S1(J) if and only if eT ∈ L1(Di, dµi). TrJi(T ) = (α + 1)ZDi eT (z) dµi(z). Proof. Set TC = ResJi(T ), write α,i(D). A positive operator T ∈ TC = TC,1 − TC,2 + i(TC,2 − TC,3) with positive operators TC,ℓ ∈ B(A2 C,α(Di)) and set Tℓ := LiftJi(TC,ℓ) for ℓ = 1, . . . , 4. Then the operators Tℓ are positive and T = T1 − T2 + JT3 − JT4. By the Ci-linearity of the Ji-trace, the Ci-linearity of the Berezin transform and Theorem 5.11, we have TrJi(T ) = TrJi(T1) − TrJi(T2) + i TrJi(T3) − i TrJi(T4) =(α + 1)ZDifT1(z) dµi(z) − (α + 1)ZDifT2(z) dµi(z) + i(α + 1)ZDifT3(z) dµi(z) − i(α + 1)ZDifT4(z) dµi(z) =(α + 1)ZDi eT (z) dµi(z). Theorem 5.14. Let T be a positive operator on A2 α,i(D) such that [T, J] = 0. (i) If 1 ≤ p < +∞ and T ∈ Sp(J), then eT ∈ Lp(Di, dµi). 26 (ii) If 0 < p ≤ 1 and eT ∈ Lp(Di, dµi), then T ∈ Sp(J). Proof. For 1 ≤ p < +∞, Proposition 3.27 implies fT p(z) ≥ (eT (z))p for z ∈ Di. Because T ∈ Sp(J), we have by Lemma 3.22 that T p ∈ S1(J), and hence, we deduce from Theorem 5.11 ZDi(cid:16)eT (z)(cid:17)p dµi(z) ≤ZDifT p(z) dµi(z) = TrJi(T p) < +∞. The second statement is proved in a similar way. Corollary 5.15. If 1 ≤ p < +∞ and T ∈ Sp(J), then eT ∈ Lp(Di, dµi). Proof. This follows from Theorem 5.14 and the fact that we can write T , in terms of the Ci-Banach space structure defined on Sp(J), as a Ci-linear combination of positive operators. See the proof of Corollary 5.13. generated by kz, i.e. For z ∈ Di, we denote by Pz the orthogonal projection of A2 ∀f ∈ A2 Pz(f ) = kzhkz, fi, α,i(D). α,i(D) onto the one-dimensional subspace Obviously, Pz is a positive operator with rank one. Observe that kz ∈ HJi Jkz = kzi, from which we deduce + by Corollary 5.10 and hence JPzf = Jkzhkz, fi = kzihkz, fi = kzh−kzi, fi = kzhJ ∗kz, fi = kzhkz, Jfi = PzJf α,i(D). Hence, [Pz, J] = 0 and Pz ∈ S1(J) as TrJi(Pz) = 1. for all f ∈ A2 Corollary 5.16. Let T ∈ BJ (A2 α,i(D)). Then Proof. Let (en)n∈N be an orthonormal basis of (A2 α,i(D))Ji + with e1 = kz. Then ∀z ∈ Di. eT (z) = TrJi(T Pz), +∞Xn=1 hen, T Pzeni = hkz, T kzi = eT (z). TrJi(T Pz) = Corollary 5.17. Let T ∈ BJ (A2 α,i(D)). Then Proof. By Corollary 5.16, we have eT (z) − eT (w) ≤ kTkkPz − Pwk1, ∀z ∈ Dz. Hence, we deduce from Corollary 3.13 and (iii) in Lemma 3.19 that eT (z) − eT (w) = TrJi(T Pz) − TrJi(T Pw) = TrJi(T (Pz − Pw)). eT (z) − eT (w) ≤ TrJi(T (Pz − Pw)) ≤ kTkkPz − Pwk1. 27 Lemma 5.18. Let z, w ∈ Di. Then while kPz − Pwk =(cid:0)1 − hkw, kzi2(cid:1) 1 kPz − Pwk1 = 2(cid:0)1 − hkw, kzi2(cid:1) 1 2 , 2 . Proof. These equalities follow immediately from the corresponding equalities in the complex case, cf. [38, Lemma 6.10], and the following facts: ResJ,i preserves both the operator norm and the Schatten norm (cf. Remark 2.32 and Corollary 3.12) and hkw, kzi = hkw,i, kz,iiC, where h·,·iC denotes the scalar product of the complex Bergman space A2 Indeed, denoting Pz,C := ResJi(Pz) and Pw,C := ResJi(Pw), we have C,i(Di) and kw,i denotes the restriction of kw to Di. kPz − Pwk = kPz,C − Pw,Ck =(cid:0)1 − hkw,i, kz,iiC2(cid:1) 1 2 =(cid:0)1 − hkw, kzi2(cid:1) 1 2 . The case of the Schatten-norm follows analogously. Finally, we obtain Lipschitz estimates for the Berezin transform analogue to those of the complex case. Theorem 5.19. Let T ∈ BJ (A2 α,i(D)). Then eT (z) − eT (w) ≤ 2√2 + αkTkρ(z, w), ρ(z, w) = z − w 1 − zw ∀z, w ∈ Di, is the pseudo-hyperbolic metric between z and w. Furthermore, the Lipschitz constant 2√2 + α is sharp. Proof. Again, we apply the corresponding result for complex linear operators [38, Theorem 6.11]. Denoting TC = ResJi(T ), we have where where eT (z) − eT (w) = fTC(z) −fTC(w) ≤ 2√2 + αkTCkρ(z, w) = 2√2 + αkTkρ(z, w). Theorem 5.20. Let T ∈ BJ (A2 α,i(D)). Then eT (z) − eT (w) ≤ 2√2 + αkTkβ(z, w), β(z, w) = log 1 2 1 + ρ(z, w) 1 − ρ(z, w) ∀z, w ∈ Di, , is the Bergman metric on Di. Furthermore, the Lipschitz constant 2√2 + α is sharp. Proof. Once more, we deduce this from the corresponding result for complex linear operators [38, Theo- rem 6.11]. Denoting TC = ResJi(T ), we have and this concludes the proof. eT (z) − eT (w) = fTC(z) −fTC(w) ≤ 2√2 + αkTCkβ(z, w) = 2√2 + αkTkβ(z, w). 28 We conclude this paper with a remark on further applications of slice hyperholomorphicity and quater- nionic operators. Remark 5.21. Classical Schur analysis is an important branch of operators theory with several applications in science and in technology, see for example the book [2], the notion of S-spectrum and of S-resolvent operators appear in Schur analysis in the slice hyperholomorphic setting in the realization of Schur functions, see the foundational paper [7]. The literature on Schur analysis in the slice hyperholomorphic setting is nowadays very well developed we mention just some of the main results that are contained in the papers [7 -- 9, 11] and in the book [6]. The main reference for slice hyperholomorphic functions are the books [6, 20, 27]. References [1] S. Adler, Quaternionic Quantum Field Theory, Oxford University Press, 1995. [2] D. Alpay, The Schur algorithm, reproducing kernel spaces and system theory, American Mathemat- ical Society, Providence, RI, 2001, Translated from the 1998 French original by Stephen S. Wilson, Panoramas et Synth`eses. [3] D. Alpay, F. Colombo, D. P. Kimsey, The spectral theorem for quaternionic unbounded normal operators based on the S-spectrum, Preprint 2014, avaliable on arXiv:1409.7010. [4] D. Alpay, F. Colombo, D. P. Kimsey and I. Sabadini, An extension of Herglotz's theorem to the quater- nions, J. Math. Anal. Appl. 421 (2015), 754 -- 778. [5] D. Alpay, F. Colombo, D. P. Kimsey and I. Sabadini. The spectral theorem for unitary operators based on the S-spectrum, Preprint 2014, avaliable on arXiv:1403.0175. [6] D. Alpay, F. Colombo, I. Sabadini, Slice Hyperholomorphic Schur Analysis, Quaderni Dipartimento di Matematica del Politecnico di Milno, QDD209, (2015). [7] D. Alpay, F. Colombo, I. Sabadini, Schur functions and their realizations in the slice hyperholomorphic setting, Integral Equations Operator Theory, 72 (2012), 253 -- 289. [8] D. Alpay, F. Colombo and I. Sabadini, Pontryagin De Branges Rovnyak spaces of slice hyperholomorphic functions, J. Anal. Math., 121 (2013), 87-125. [9] D. Alpay, F. Colombo and I. Sabadini, Krein-Langer factorization and related topics in the slice hyper- holomorphic setting, J. Geom. Anal., 24 (2014), 843 -- 872. [10] D. Alpay, F. Colombo, I. Sabadini, G. Salomon, Fock space in the slice hyperholomorphic setting, in Hypercomplex Analysis: New Perspectives and Applications, (2014), 43 -- 59. [11] D. Alpay, F. Colombo, I. Lewkowicz, I. Sabadini, Realizations of slice hyperholomorphic generalized contractive and positive functions, Milan J. Math. 83 (2015), 91-44. [12] D. Alpay, F. Colombo, J. Gantner and I. Sabadini, A new resolvent equation for the S-functional calculus, J. Geom. Anal., 25 (2015), 1939 -- 1968. [13] G. Birkhoff and J. von Neumann, The logic of quantum mechanics, Ann. of Math., 37 (1936), 823-843. [14] C.M.P. Castillo Villalba, F. Colombo, J. Gantner, J.O. Gonz´alez-Cervantes, Bloch, Besov and Dirichlet Spaces of Slice Hyperholomorphic Functions, Complex anal. Oper. theory 9 (2015), 479-517. [15] F. Colombo, J. O. Gonz´alez-Cervantes, M. E. Luna-Elizarraras, I. Sabadini, M. Shapiro, On two approaches to the Bergman theory for slice regular functions, Advances in hypercomplex analysis, 3954, Springer INdAM Ser., 1, Springer, Milan, 2013. 29 [16] F. Colombo, J. O. Gonz´alez-Cervantes, I. Sabadini, On slice biregular functions and isomorphisms of Bergman spaces, Complex Var. Elliptic Equ., 57 (2012), 825 -- 839. [17] F. Colombo, J. O. Gonz´alez-Cervantes, I. Sabadini, The C-property for slice regular functions and applications to the Bergman space, Compl. Var. Ell. Equa., 58 (2013), 1355 -- 1372. [18] F. Colombo, I. Sabadini, On some properties of the quaternionic functional calculus, J. Geom. Anal., 19 (2009), 601-627. [19] F. Colombo, I. Sabadini, On the formulations of the quaternionic functional calculus, J. Geom. Phys., 60 (2010), 1490 -- 1508. [20] F. Colombo, I. Sabadini, and D. C. Struppa, Noncommutative functional calculus. Theory and applica- tions of slice regular functions, volume 289 of Progress in Mathematics. Birkhauser/Springer Basel AG, Basel, 2011. [21] N. Dunford, J. Schwartz. Linear Operators, part I: General Theory , J. Wiley and Sons (1988). [22] N. Dunford and J. Schwartz. Linear Operators, part II: Spectral theory , J. Wiley and Sons (1988). [23] G. Emch, M´ecanique quantique quaternionienne et relativit´e restreinte, I, Helv. Phys. Acta, 36 (1963), 739 -- 769. [24] D. R. Farenick and B. A. F. Pidkowich, The spectral theorem in quaternions, Linear Algebra Appl., 371 (2003), 75 -- 102. [25] M. Fashandi, Compact operators on quaternionic Hilbert spaces Facta Univ. Ser. Math. Inform. 28, no. 3 (2013), 249 -- 256 [26] D. Finkelstein, J. M. Jauch, S. Schiminovich and D. Speiser, Foundations of quaternion quantum me- chanics, J. Mathematical Phys., 3 (1962), 207 -- 220. [27] G. Gentili, C. Stoppato, D. C. Struppa, Regular functions of a quaternionic variable. Springer Mono- graphs in Mathematics. Springer, Heidelberg, 2013. [28] R. Ghiloni, V. Moretti and A. Perotti, Continuous slice functional calculus in quaternionic Hilbert spaces, Rev. Math. Phys., 25 (2013), 1350006, 83 pp. [29] R. Ghiloni, V. Moretti and A. Perotti, Spectral properties of compact normal quaternionic operators, in Hypercomplex Analysis: New Perspectives and Applications Trends in Mathematics, 133 -- 143, (2014). [30] I. C. Gohberg, M. G. Krein, Introduction to the theory of linear nonselfadjoint operators Translated from the Russian by A. Feinstein. Translations of Mathematical Monographs, Vol. 18 American Mathematical Society, Providence, R.I. 1969 xv+378 pp. [31] L. P. Horwitz and L. C. Biedenharn, Quaternion quantum mechanics: Second quantization and gauge fields, Annals of Physics, 157 (1984), 43217488. [32] W. Rudin, Real and complex Analysis, Third Edition, McGraw-Hill Book Co., New York, (1987) [33] K. Stroethoff, The Berezin transform and operators on spaces of analytic functions (English summary) Linear operators (Warsaw, 1994), 361380, Banach Center Publ., 38, Polish Acad. Sci., Warsaw, 1997 [34] C. S. Sharma and T. J. Coulson, Spectral theory for unitary operators on a quaternionic Hilbert space, J. Math. Phys., 28 (1987), 1941 -- 1946. [35] K. Thirulogasanthar, S.T. Ali, A class of vector coherent states defined over matrix domains, J. Math. Phys, 44 (11), 5070 -- 5083. 30 [36] G. Sarfatti, Elements of function theory in the unit ball of quaternions, PhD thesis, Universit´a di Firenze, 2013. [37] K. Viswanath, Normal operators on quaternionic Hilbert spaces, Trans. Amer. Math. Soc., 162 (1971), 337 -- 350. [38] K. Zhu, Operator Theory in Function Spaces, 2nd edition, volume 138 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2007 31
1708.02788	1	1708	2017-08-09T11:26:38	Fractional powers of the parabolic Hermite operator. Regularity properties	[ "math.FA" ]	Let $\mathcal{L}= \partial_t- \Delta_x+\|x\|^2$. Consider its Poisson semigroup $e^{-y\sqrt{\mathcal{L}}}$. For $\alpha >0$ define the Parabolic Hermite-Zygmund spaces $$ \Lambda^\alpha_{\mathcal{L}}=\left\{f: \:f\in L^\infty(\mathbb{R}^{n+1})\:\; {\rm and} \:\; \left\\|\partial_y^k e^{-y\sqrt{\mathcal{L}}} f \right\\|_{L^\infty(\mathbb{R}^{n+1})}\leq C_k y^{-k+\alpha},\;\: {\rm with }\, k=[\alpha]+1, y>0. \right\}, $$ with the obvious norm. It is shown that these spaces have a pointwise description of H\"older type. The fractional powers $\mathcal{L}^{\pm \beta}$ are well defined in these spaces and the following regularity properties are proved: \begin{eqnarray} \alpha, \beta >0, \quad \\|\mathcal{L}^{-\beta} f\\|_{ \Lambda^{\alpha+2\beta}_{\mathcal{L}}}\le C \\|f\\|_{ \Lambda^\alpha_{\mathcal{L}}}. \end{eqnarray} \begin{eqnarray} 0< 2\beta < \alpha, \quad \\|\mathcal{L}^\beta f\\|_{\Lambda_{\mathcal{L}}^{\alpha-2\beta}}\le C \\|f\\|_{\Lambda^\alpha_{\mathcal{L}}}. \end{eqnarray} Parallel results are obtained for the Hermite operator $- \Delta +\|x\|^2.$ The proofs use in a fundamental way the semigroup definition of the operators $\mathcal{L}^{\pm \beta}$ and $(-\Delta+\|x\|^2)^{\pm \beta}$. The non-convolution structure of the operators produce an extra difficulty of the arguments.	math.FA	math	FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES MARTA DE LE ´ON-CONTRERAS AND JOS´E L. TORREA Abstract. Let L = ∂t − ∆x + x2. Consider its Poisson semigroup e−y define the Parabolic Hermite-Zygmund spaces √L. For α > 0 ΛαL = f : f ∈ L ∞ (Rn+1) and ≤ Cky −k+α, with k = [α] + 1, y > 0. (cid:26) (cid:13)(cid:13)(cid:13)∂k y e √L −y f (cid:13)(cid:13)(cid:13)L∞(Rn+1) (cid:27) , with the obvious norm. It is shown that these spaces have a pointwise description of Holder type. The fractional powers L±β are well defined in these spaces and the following regularity properties are proved: α, β > 0, 0 < 2β < α, (cid:107)L−βf(cid:107)Λα+2βL (cid:107)Lβf(cid:107)Λα−2βL ≤ C(cid:107)f(cid:107)ΛαL . ≤ C(cid:107)f(cid:107)ΛαL . Parallel results are obtained for the Hermite operator −∆ + x2. The proofs use in a fun- damental way the semigroup definition of the operators L±β and (−∆ + x2)±β. The non-convolution structure of the operators produce an extra difficulty of the arguments. 1. Introduction Treatises dealing with Lipschitz and Holder spaces have been the object in quite a lot papers and books along the last hundred years. In general they can be considered as the classes between the space of continuos functions and the space of C1 (differentiable with continuous derivatives) functions, this is the case of Cα, 0 < α < 1. Also they can be considered as the spaces which fill the interval between the classes Ck and Ck+1, this is the case of the spaces Ck,α, k ∈ N, 0 < α < 1. The importance of the smoothness of the functions in the classical theory of Fourier series drove, in a natural way, to analyze the validity of different theorems for the case of Lipschiz functions. We refer to the classical book of Zygmund, [21], to see the role played by these classes in classical Fourier Analysis. In Harmonic Analysis the classes became important as spaces in which some operators are well defined and satisfy some boundedness properties, we refer to the book of E. Stein, [13], in order to have a detailed description from a Harmonic Analysis point of view. In differential equations, Lipschitz continuity is the key of the Picard-Lindelof theorem for the existence and uniqueness of the solution to an initial value problem. Results about regularity properties with respect to Holder classes, Cα(Rn) and Ck,α(Rn) , are one of the important matters in the theory of partial differential equations. For elliptic operators they can be used to obtain classical solutions of second order elliptic equations of the form Lu = f (see for instance [5, Chapter 6]). Moreover, in certain measure spaces without notion of derivative, the Lipschitz classes are a good substitute of the space C∞ in order to define distributions, and some 2010 Mathematics Subject Classification. Primary 42C05; Secondary 35K08, 42B35. Key words and phrases. Semigroups. Fractional laplacian. Lipschitz Holder Zygmund spaces. Holder estimates. Research partially supported by grant MTM2015-66157-C2-1-P (MINECO/FEDER). 1 2 DE LE ´ON-CONTRERAS AND TORREA abstract Harmonic Analysis can be performed. This is of special importance in spaces of homogeneous type, see [10]. Finally they are object of study in their own by researchers in Functional Analysis, see [6]. The outbreak produced by the paper of L. Caffarelli and L.Silvestre about the fractional laplacian, [2], has given way to a flowering of papers analyzing the classical properties of the elliptic operators but in the case of these "new" fractional operators. In particular regularity properties for the operator (−∆)σ were proved in [12]. For elliptic operators in divergence form see [3]. In the case of the Harmonic oscillator H = −∆ + x2, the classes CαH(Rn) were defined in [16], see Definition 3.14, Schauder and Holder estimates were proved in this case. As a shorthand it can be said that, for 0 < α < 1, a Cα function satisfies an inequality of the type f (x) − f (x − y) ≤ Cyα. For α > 1, not an integer, the [α]-order derivatives of the function f satisfy the same kind of inequality. Special mention should deserve the case α = 1, with is described as the Zygmund class f (x + y) + f (x − y) − 2f (x) < cy, see [21, Chapter II] . See also the interesting article [7] and the references there in. These pointwise definitions imply that to prove regularity results of an operator among these spaces we need its pointwise expression. In some (in fact many) cases this can be a rather involved formula, see for example the expressions of (−∆)α and H−σf (x) in [16]. In the 60's of last century the language of the semigroups was used in order to characterize Holder spaces, see [18]. This is specially successful in the case of the Poisson semigroup. The classical reference is E. M. Stein, see [13, Chapter 5]. Being a little bit imprecise it can be said that a function f belongs to a class Λα if (cid:107)∂k posteriori these classes are seen to coincide with the Cα classes. It is interesting to notice that this description also covers the Zygmund class. In the present paper the importance of this picture is based on the fact that in order to prove boundedness properties of operators, one could avoid the long, tedious and sometimes cumbersome computations that are needed when the pointwise expressions are handled. This will be our case. √−∆f(cid:107)L∞(Rn) ≤ Ct−k+α, k ≥ α. A t e−t The characterization of Holder spaces via the Poisson semigroup e−t √−∆ raise the question √ Of(cid:107) y e−y √ L∞(Rn, e−x2 πn/2 ) O is the Poisson semigroup associated to the operator O. 2 ∆ − x · ∇, the so-called Gaussian Lipschitz spaces were defined in [4] as the ≤ Cky−k+α, k = [α] + 1, where of analyze some Holder spaces associated to different laplacians and to find the pointwise and semigroup estimate characterizations. For the case of the Ornstein-Ulhenbeck operator in Rn, O = 1 collection of functions such that (cid:107)∂k e−y In the particular interval 0 < α < 1, these Gaussian Lipschitz spaces have been recently characterized pointwise in [9]. If S = −∆ + V is the Schrodinger operator in Rn, n ≥ 3, where V satisfies satisfies a reverse Holder inequality for some q > n/2, the classes ΛαS, 0 < α < 1, were defined in [11]. The authors prove that the classes can be described by a Campanato-BMO type condition, boundedness in these spaces of operators like fractional powers of S are considered. For the Hermite operator H = −∆ + x2 in Rn, pointwise Holder spaces, Ck,αH were defined in [16] and boundedness properties of Hermite fractional laplacian, Hα, 0 < α < 1, were considered. In the case of parabolic operators of the type ∂ u(t, x) + u(t, x)+c(t, x)u(t, x)+f (t, x), where a, b, c are real valued and c ≤ 0, some pointwise bi(t, x) ∂ ∂xi Holder classes were introduced in [8]. Where solvability and a priori estimates were proved. √M is used in [17] for defined the corresponding For M = ∂t + ∆, the Poisson semigroup e−y Holder classes. The coincidence with the pointwise classes of Krylov were proved for the α considered in [8]. This semigroup characterization was used to show new regularity properties for fractional powers (∂t + ∆x)±α. ∂t u(t, x) = aij(t, x) ∂2 ∂xi∂xj FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 3 Now we shall present our results. Along this paper we shall deal with the parabolic Hermite operator L := ∂t + H = ∂t − ∆x + x2, x ∈ Rn, t > 0. (1.1) As the operators ∂t and H commute, the heat semigroup e−yL will be the composition of the heat semigroups e−y∂t and e−yH. As these semigroups are well known, see [14] and [1], we shall have a satisfactory description of the operator e−yL. This description will be use in √L, the fractional parabolic order to define, among other operators, the Poisson semigroup e−y Hermite integrals L−β, β > 0 and the fractional parabolic Hermite laplacian Lβ, β > 0. See Section 2. Once the Poisson semigroup, Py is introduced, see Section 2, we define the following √L and α > 0, we con- (cid:27) ΛαL = whose norm is given by (cid:107)f(cid:107)ΛαL := (cid:107)f(cid:107)∞ + C, where C is the infimum of the positive constants Ck above. associated classes of functions. Definition 1.1. [Parabolic Hermite-Zygmund spaces] Let Py = e−y sider the class ≤ Cky−k+α, with k = [α] + 1, y > 0. (cid:13)(cid:13)(cid:13)L∞(Rn+1) f : f ∈ L∞(Rn+1) and (cid:13)(cid:13)(cid:13)∂k yPyf (cid:26) , The operator H can be factorized as H = 1 We will show, in Theorem 1.2 that these classes have a pointwise description. Moreover, a restriction to functions depending only on x, produces a natural Definition 1.3 and a Theorem 1.4 for the case of Hermite operator in Rn. i=1(AiA−i + A−iAi), Ai = ∂xi + xi, A−i = −∂xi + xi. The first order operators A±i play the role, with respect to operator H, of the derivatives ±∂xi with respect to the classical laplacian ∆. See [14], [16]. Theorem 1.2. Let L := ∂t + H = ∂t − ∆x + x2, x ∈ Rn, t > 0. (1) Suppose that 0 < α < 2. Then f ∈ ΛαL if and only if there exists a constant C > 0 (cid:80)n 2 such that (cid:107)f (· − τ,· − z) + f (· − τ,· + z) − 2f (·,·)(cid:107)L∞(Rn+1) ≤ C(τ1/2 + z)α, (τ, z) ∈ Rn+1 and (1 + x)αf ∈ L∞(Rn+1). In this case, if K denotes the least constant C for which the inequality above is true, then (cid:107)u(cid:107)ΛαL := [u]M α + K. Where [f ]M α = (cid:107)(1 + · )αf (·,·)(cid:107)∞. (1.2) (2) Suppose that α > 2. Then f ∈ ΛαL if and only if (cid:17) AiAjf ∈ Λα−2L , i, j = ±1, . . . ,±n, In this case the following equivalence holds (cid:16)(cid:107)AiAjf(cid:107)Λα−2L (cid:107)f(cid:107)ΛαH ∼ ±n(cid:88) i,j=±1 and ∂tf ∈ Λα−2L . + (cid:107)∂tf(cid:107)Λα−2L . −∆x + x2. Definition 1.3. [Hermite-Zygmund spaces] Let Py = e−y class The above results have the following parallel results in the case of Hermite operator H = √H and α > 0, we consider the (cid:27) (cid:26) ≤ Cky−k+α, with k = [α] + 1, y > 0. , ΛαH = g : g ∈ L∞(Rng) and (cid:13)(cid:13)(cid:13)∂k y Pyg (cid:13)(cid:13)(cid:13)L∞(Rn) 4 DE LE ´ON-CONTRERAS AND TORREA whose norm is given by (cid:107)g(cid:107)ΛαH := (cid:107)g(cid:107)∞ + C, where C is the infimum of the positive constants Ck above. Theorem 1.4. Let g ∈ L∞(Rn). (1) Suppose that 0 < α < 2. Then g ∈ ΛαH if and only if (1 + · )αg ∈ L∞(Rn) and there exists a constant C > 0 such that (cid:107)g(· − z) + g(· + z) − 2g(·)(cid:107)L∞(Rn) ≤ Czα, z ∈ Rn. In this case, if K denotes the least constant C for which the inequality above is true, then (cid:107)g(cid:107)ΛαH := [g]M α + K. Where [g]M α = (cid:107)(1 + · )αg(·)(cid:107)∞. (2) Suppose that α > 1. Then g ∈ ΛαH if and only if g ∈ Λα−1H and xig ∈ Λα−1H ∂ ∂xi In this case the following equivalence holds i = 1, . . . , n. (cid:16)(cid:13)(cid:13)(cid:13)(cid:13) ∂ ∂xi n(cid:88) i=1 (cid:13)(cid:13)(cid:13)(cid:13)Λα−1H g (cid:13)(cid:13)(cid:13)xig (cid:13)(cid:13)(cid:13)Λα−1H (cid:17) . + (cid:107)g(cid:107)ΛαH ∼ (cid:107)g(cid:107)∞ + As we said before we shall obtain regularity results of operators associated to L when acting over the classes defined above. We shall consider positive, negative and imaginary powers of the operators L and H, as well as Riesz transforms. For the appropriated definitions see Section 2. Theorem 1.5. Let 0 < 2β < α and f ∈ ΛαL, (respectively g ∈ ΛαH), then Lβf ∈ Λα−2βL (respectively Hβg ∈ Λα−2βH (cid:107)Lβf(cid:107)Λα−2βL (respectively (cid:107)Hβg(cid:107)Λα−2βH ≤ C(cid:107)g(cid:107)ΛαH). ≤ C(cid:107)f(cid:107)ΛαL, ) and Theorem 1.6. Let 0 < α, β. (i) Given f ∈ ΛαL (respectively g ∈ ΛαH), then L−βf ∈ Λα+2βL (respectively H−βg ∈ Λα+2βH ) and (cid:107)L−βf(cid:107)Λα+2βL (ii) If f ∈ L∞(Rn+1), (cid:107)L−βf(cid:107)ΛβL ≤ C(cid:107)f(cid:107)ΛαL, (respectively (cid:107)H−βg(cid:107)Λα+2βH (respectively g ∈ L∞(Rn)), then ≤ C(cid:107)f(cid:107)∞, (respectively (cid:107)H−βg(cid:107)ΛβH ≤ C(cid:107)g(cid:107)∞). ≤ C(cid:107)g(cid:107)ΛαH). We also get the boundedness of the multiplier operator of the Laplace transform type on the spaces ΛαL and ΛαH. We recall to the reader that the imaginary powers λiγ are examples of multipliers of Laplace transform type. In [11], this result is proved for every Schrodinger operator when 0 < α < 1. Theorem 1.7. Let a be a bounded function on [0,∞) and consider (cid:90) ∞ m(λ) = λ1/2 e−sλ1/2 a(s)ds, λ > 0. Then, for every α > 0, the multiplier operator of the Laplace transform type m(L) (respec- tively m(H)) is bounded from ΛαL (respectively ΛαH) into itself. 0 In [11], this result is proved for every Schrodinger operator when 0 < α < 1. FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 5 Theorem 1.8. Consider the Parabolic Hermite Riesz transforms of order m ≥ 1 defined by Rν = (Aν1±1Aν2±2 . . . Aνn±n)L−m/2 and Rm = ∂m t L−m where νi ≥ 0, i = 1, . . . , n and ν = ν1 + ··· + νn = m. Let α > 0, then Rν and Rm are bounded from ΛαL into itself. A parallel result holds for the operators (Aν1±1Aν2±2 . . . Aνn±n)H−m/2 when acting on the spaces ΛαH. See [14], [16] and [19] and the references there in for more information about the hermitian Riesz transforms AjH−1/2. Apart from the above regularity results, our semigroup language allows us to get some . Suppose that maximum principle. Theorem 1.9. [Maximum principle] Let 0 < β < 1, α > 2β and f ∈ Λα/2,α t,Hx Then Lβf (t0, x0) ≤ 0. (1) f (t0, x0) = 0 for some (t0, x0) ∈ Rn+1, and (2) f (t, x) ≥ 0 for t ≤ t0, x ∈ Rn. Moreover, Lβf (t0, x0) = 0 if and only if f (t, x) = 0 for t ≤ t0 and x ∈ Rn. The organization of the paper is as follows. In Section 2 we present the mains objects like Poisson semigroup and fractional powers of operators. We observe that as the operator L is not positive, the standard definitions have to be adapted to this complex case. In Section 3 we show the coincidence of the spaces ΛαL and ΛαH with some Holder pointwise spaces defined previously in [8] and [17] in the parabolic and Hermite settings. Section 4 is devoted to the proof of Theorems 1.2 and 1.4. Theorems 1.5, 1.6, 1.7, 1.8 and 1.9 are proved in Sections 5 and 6. Finally in Section 7 we collect some inequalities needed along the paper. The non-convolution structure of our operators, produces non trivial difficulties and technical computations that we have to solve in each case. This is common to the parabolic case L and the Hermite case H. We present the computations and the results in such a way that the parabolic case includes as particular case the Hermite case. This will be clarified in the subsections called Elliptic Hermite setting included at the end the corresponding Sections. Along this paper, we will use the variable constant convention, in which C denotes a constant that may not be the same in each appearance. The constant will be written with subindexes if we need to emphasize the dependence on some parameters. For functions g ∈ Lp(Rn), the heat semigroup e−τH has the pointwise expression 2. Preliminary considerations. e−τHg(x) = e− x−z2 4 coth τ e− x+z2 (2π sinh 2τ )n/2 4 tanh τ g(z) dz, Rn (cid:90) (cid:17) (cid:90) e−τ ∂tf (t,·) (x), moreover (cid:90) see [14], [20]. The operator ∂t in (1.1) is taking care of the past, in other words its heat semigroup is given by e−τ ∂tϕ(t) = ϕ(t − τ ). Hence for functions f ∈ C1 Lp(Rn)(R) we have e−τLf (t, x) = e−τH(cid:16) e−τLf (t, x) = e−τH(f (t − τ,·))(x) = (2.3) The Fourier-Hermite transform of a function f ∈ L1(Rn+1) can be defined as Rn tanh τ e− x−z2 4 coth τ e− x+z2 (2π sinh 2τ )n/2 4 f (t − τ, z) dz. (2.4) F(f )(ρ, µ) = f (t, x)e−iρthµ(x)dtdx, ρ ∈ R, µ ∈ Nn 0 . Rn+1 6 Where hµ(x) =(cid:81)n defined by DE LE ´ON-CONTRERAS AND TORREA j=1 hµj (xj), x = (x1, . . . , xn) ∈ Rn. For k ∈ N, hk is the Hermite function hk(t) = (−1)k (2kk!π1/2)1/2 Hk(t) e−t2/2, t ∈ R. Here Hk denotes the Hermite polynomial of degree k (see [20]). These functions are eigen- vectors of the Hermite operator H. In fact Hhµ = (2µ + n) hµ. Consequently for functions f ∈ L1(Rn+1) we have (2.5) Given z ∈ C with (cid:60)z ≥ 0, by analytic continuation it can be seen that F(e−τLf )(ρ, µ) = e−τ (iρ+2µ+n)F(f )(ρ, µ), ρ ∈ R, µ ∈ Nn. √ e−t Hence for f ∈ L1(Rn+1) we have z = √ y 2 π √ e−y iρ+2µ+nF(f )(ρ, µ) = √ y 2 π This last expression can be written as F(e−y √Lf )(ρ, µ) = √ y 2 π e−y2/4τ e−τ z dτ τ 3/2 . e−y2/4τ e−τ iρ+2µ+nF(f )(ρ, µ) dτ τ 3/2 . e−y2/4τF(e−yLf )(ρ, µ) dτ τ 3/2 . 0 (cid:90) ∞ (cid:90) ∞ (cid:90) ∞ 0 0 The Fourier transform defined in (2.4) is an isometry in L2(Rn+1) and in particular we have , in the L2(Rn+1) sense (2.6) Pyf (t, x) = e−y √Lf (t, x) = √ y 2 π e−y2/4τ e−τLf (t, x) dτ τ 3/2 . (cid:90) ∞ 0 For functions f good enough, formulas (2.3) and 2.6 give the following pointwise expression e−y2/4τ e− x−z2 4 coth τ e− x+z2 (2π sinh 2τ )n/2 4 tanh τ f (t − τ, z) dz , x ∈ Rn, t ∈ R. dτ τ 3/2 Rn (cid:90) ∞ (cid:90) (2.7) Pyf (t, x) = √ y 2 π 0 On the other hand ye−y2/4τ e− x−z2 coth τ e− x+z2 4 (2π sinh 2τ )n/2τ 3/2 4 tanh τ χ{τ >0} ≤ C 4τ e− y2 τ e− x−z2 4τ τ n/2 y τ 1/2 χ{τ >0} = Φy(τ, x − z). As Φy belongs to L1(Rn+1), the formula (2.7), defining the Parabolic Poisson Hermite integral , remains valid for any f ∈ Lp(Rn+1), 1 ≤ p ≤ ∞, (x, t) ∈ Rn+1. Moreover this integral satisfies a Parabolic Hermite Laplace equation as the following Proposition shows. Proposition 2.10. Assume f ∈ L∞(Rn+1). Then Pyf (t, x) satisfies the equation (2.8) yPyf (t, x) − LPyf (t, x) = 0, (x, t) ∈ Rn+1. ∂2 Proof. We observe that (cid:16) (cid:12)(cid:12)(cid:12)∂y2 ye−y2/4τ e− x−z2 (cid:12)(cid:12)(cid:12)∆x (cid:16) + coth τ e− x+z2 4 (2π sinh 2τ )n/2τ 3/2 4 tanh τ ye−y2/4τ e− x−z2 χ{τ >0} coth τ e− x+z2 4 (2π sinh 2τ )n/2τ 3/2 4 tanh τ (cid:17)(cid:12)(cid:12)(cid:12) χ{τ >0} (cid:17)(cid:12)(cid:12)(cid:12) FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 7 coth τ e− x+z2 4 (2π sinh 2τ )n/2τ 3/2 4 tanh τ χ{τ >0} (cid:17)(cid:12)(cid:12)(cid:12) (cid:16) (cid:12)(cid:12)(cid:12)∂τ + ≤ C τ ye−y2/4τ e− x−z2 e− y2 τ e− x−z2 4τ τ n/2 4τ χ{τ >0}. Hence, for y > 0 and x − z > 0, the function ye−y2/4τ e smooth in all its variables. In particular we can write − x−z2 − x+z2 4 (2π sinh 2τ )n/2τ 3/2 coth τ e 4 tanh τ χ{τ >0} is Pyf (t, x) = √ y 2 π R Rn e−y2/4(t−τ ) e− x−z2 4 coth(t−τ )e− x+z2 (2π sinh 2(t − τ ))n/2 4 tanh(t−τ ) f (τ, z) dz χ{t−τ >0} dτ (t − τ )3/2 . The above estimates also show that we can interchange the derivatives with the integral for for y > 0 and x − z > 0. Hence the Proposition follows since the kernel of this last integral (cid:3) satisfies the equation (2.8). Remark 2.11. The proof of the previous Lemma also shows that for functions f ∈ L∞(Rn+1) we can write (cid:90) (cid:90) (cid:90) Py(τ, x, z)f (t − τ, x − z)dzdτ (2.9) Pyf (t, x) = = (cid:90) (cid:90) Rn+1 √ y 2 π R Rn e− z2 4 coth τ e− 2x−z2 (2π sinh(2τ ))n/2τ 3/2 tanh τ e− y2 4τ 4 f (t − τ, x − z)dz χ{τ >0} dτ. As we have noticed in (2.5), the infinitesimal generator, L, of the semigroup e−τL is not positive. This forced us to use some complex variable technique in order to give a sense to the powers of the operator L. Given a non necessarily positive operator L, formulas to define L±α, where 0 < α < 1, were considered in [1], [15] and [17]. Given 0 < β , we recall the following two integrals related with the Gamma function: (cid:90) ∞ e−ttβ dt t (cid:90) ∞ (cid:0)e−t − 1(cid:1)[β]+1 dt , 0 cβ = Cβ = (2.10) It is well known that Cβ = Γ(β) for all 0 < β and cβ = Γ(−β) for 0 < β < 1. The following Lemma was proved in [1]. Lemma 2.12. Let 0 < β < 1 and −π/2 ≤ ϕ0 ≤ π/2. Consider the ray in the complex plane rayϕ0 := {z = reiϕ0 : 0 < r < ∞}. Then t1+β . 0 Γ(β) = rayϕ0 e−zzβ dz z , and Γ(−β) = (e−z − 1) dz z1+β . rayϕ0 (cid:90) (cid:90) For 0 < β < 1, the absolutely convergent integrals in (2.10) can be interpreted as integrals of the functions F (t) = e−ttβ−1 and G(t) = (e−t − 1)/t1+β along the "complex" path {z = t : 0 < t < ∞}. The proof of the Lemma is based in the Cauchy Integral Theorem applied to the functions F (z) = e−zzβ−1 and G(z) = (e−z − 1)/z1+β. Both functions are analytic for z (cid:54)= 0. For the integrals defined in (2.10) we could state a parallel Lemma to 2.12, by choosing H(z) = (e−z − 1)[β]+/z1+β. The proof follows the same steps. We leave the details to the reader. We have the following Corollary. 8 DE LE ´ON-CONTRERAS AND TORREA Corollary 2.13. Let β > 0 and λ a complex number with (cid:60)λ ≥ 0. Then (cid:90) ∞ λ−β = 1 Γ(β) 0 eλttβ dt t , and λβ = 1 cβ (eλt − 1)[β]+1 dt t1+β . We use the last Corollary to define define the negative and positive fractional powers of (cid:90) ∞ 0 (cid:17)[2β]+1 the operator L as where c2β =(cid:82) ∞ 0 (cid:90) ∞ (cid:16) Lβf (t, x) = 1 c2β 0 e−τL1/2 − I f (t, x) dτ τ 1+2β , (e−τ − 1)[2β]+1 (cid:90) ∞ dτ τ 1+2β . Also, for β > 0, e−τL1/2 1 L−βf (t, x) = Γ(2β) 0 f (t, x) dτ τ 1−2β . Observe that for good enough functions F(L±βf )(ρ, µ) = (iρ + 2µ + n)±βF(f )(ρ, µ), ρ ∈ R, and µ ∈ Nn. (cid:90) ∞ (cid:90) 2.1. Elliptic Hermite setting. Given g ∈ L∞(Rn), consider the function f (t, x) = g(x), then formula (2.7) becomes e−y2/4τ e− x−z2 4 coth τ e− x+z2 (2π sinh 2τ )n/2 4 tanh τ dτ τ 3/2 Pyf (t, x) = √ y π 0 2 Rn g(z) dz = Pyg(x). • For functions g ∈ L∞(Rn), Pyg(x) satisfies the equation ∂2 (2.11) Where Pyg(x) is the Poisson semigroup associated to the operator H = −∆x + x2. The thoughts developed along this section show that: y Pyg(x)−HPyg(x) = 0, x ∈ Py(τ, x, z)dτ = Py(x, z), for all x, z ∈ Rn, where Py is the Poisson kernel associated to L and Py is the Poisson kernel associated to the harmonic oscillator, H. • Let β > 0, for g good enough, • Identities (2.9) and (2.11) give that Rn+1. (cid:90) R (cid:90) (cid:16) (cid:17)[2β]+1 Hβg(x) = is well defined and (cid:100)Hβg(µ) = (2µ + n)β g(µ), µ ∈ Nn, with g(µ) =(cid:82) g(x) e−τH1/2 − Id dτ τ 1+2β , 1 c2β • Let β > 0, for good enough functions g, (cid:90) ∞ 1 H−βg(x) = dτ τ 1−2β (cid:92)H−βg(µ) = (2µ + n)−β g(µ), µ ∈ Nn. e−τL1/2 Γ(2β) g(x) 0 is well defined and Rn g(x)hµ(x)dx. 3. Coincidence of Parabolic Hermite-Zygmund with Parabolic Hermite-Holder spaces. We shall begin by recalling the following definition, it can be found in [16]. Definition 3.14. [Hermite Holder spaces] Let 0 < α < 1. We consider the space of functions CαH(Rn) = {f : (1 + · )αf (·) ∈ L∞(Rn), and (cid:107)f (· + z) − f (·)(cid:107)L∞(Rn) ≤ Azα} with associated norm (cid:107)f(cid:107)CαH = [f ]M α + [f ]CαH. FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 9 Where [f ]M α = (cid:107)(1 + · )αf (·)(cid:107)∞ and [f ]CαH = sup z>0 For α > 1 and not integer, we say that f ∈ CαH(Rn), if there exist the derivatives of order [α] and the norm (cid:107)f(cid:107)CαH := [f ]M α−[α] + [Ai1 . . . Aimf ]M α−[α] + [Ai1 . . . Ai[α]f ] (cid:88) (cid:88) zα . , Cα−[α]H (cid:107)f (· + z) − f (·)(cid:107)∞ 1≤i1,...,i[α]≤n 1≤i1,...,im≤n 1≤m≤[α] is finite. Some parabolic Holder spaces were considered by N. Krylov, see [8]. Namely (i) Let 0 < α < 1, Cα/2,α was defined as the set of bounded functions such that [f ]Cα/2,α = sup (τ,z)(cid:54)=(0,0) (cid:107)f (· − τ,· − z) − f (·,·)(cid:107)L∞(Rn+1) (τ1/2 + z)α < ∞. (ii) For 1 < α < 2, f ∈ Cα/2,α if ∂xif ∈ Cα/2−1/2,α−1 and f (·, x) ∈ Cα/2(R) uniformly on x. (iii) Let 0 < α < 1, C1+α/2,2+α if ∂2 xif and ∂tf belong to Cα/2,α. These Krylov's definitions together with Definition 3.14 drive us to consider the following definition. Definition 3.15. [Parabolic Hermite Holder spaces] • Let 0 < α < 1. We say that f ∈ Cα/2,α t,H if f ∈ Cα/2,α and (1 + x)αf (t, x) < ∞, In this case, (cid:107)f(cid:107) [f ]M α = sup (t,x)∈Rn+1 = [f ]Mα + [f ] Cα/2,α t,H if A±if ∈ Cα/2−1/2,α−1 • For 1 < α < 2, f ∈ Cα/2,α on x. • For 2 < α < 3 we say that a function f ∈ Cα/2,α function ∂tf belong to Cα/2−1,α−2 t,H . t,H . t,H Cα/2,α t,H and f (·, x) ∈ Cα/2(R) uniformly t,H , if the functions A±iA±jf and the In the next result we will show that the functions in Cα/2,α t,H , 0 < α < 1, can be taken to be continuous, so the inequality f (t − τ, x + z) − f (t, x) ≤ C(τ 1/2 + z)α holds for every x ∈ Rn, t ∈ R. Proposition 3.16. For 0 < α < 1, every f ∈ Cα/2,α measure zero so that it becomes continuous. Proof. Let f ∈ Cα/2,α hypothesis on f , Lemma 7.34 (i) and Lemma 7.33 (3) we have Pyf (t, x) − f (t, x) ≤ t,H (Rn+1). We will follow the ideas in Stein [13, page 142]. By the t,H (Rn+1) can be modified on a set of Py(τ, x, z)(f (t − τ, x − z) − f (t, x))dτ dz Py(τ, x, z)dτ dz − 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Rn+1 (cid:16)(cid:90) (cid:90) ∞ Cα/2,α t,H Rn 0 ≤ [f ] ye− y2+z2 τ cτ (τ 1/2 + z)α n+3 2 dτ dz Rn+1 √L1(t, x)−1 (cid:12)(cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107) yα. Cα/2,α t,H (cid:12)(cid:12)(cid:12)(cid:12) + (cid:17) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)f (t, x) (cid:18)(cid:90) (cid:12)(cid:12)(cid:12)(cid:12)e−y +(cid:107)f(cid:107)∞ 10 DE LE ´ON-CONTRERAS AND TORREA In particular, we conclude that Pyf converges uniformly to f as y goes to zero. As Pyf is (cid:3) continuous, f can be taken to be continuous. Now we shall show that, for 0 < α < 1, the pointwise Definition 3.15 is equivalent to the Definition 1.1 given by using of Poisson semigroup. Theorem 3.17. [0 < α < 1, Parabolic Hermite-Holder = Parabolic Hermite-Zygmund] Let 0 < α < 1. Then Cα/2,α t,H = ΛαL, with equivalence of norms. Proof. For f ∈ Cα/2,α t,H (Rn+1), we write y∂yPyf (t, x) = y∂yPy(τ, x, z)(f (t − τ, x − z) − f (t, x))dτ dz Rn+1 (cid:90) By Lemma 7.34 (i) we have I1 ≤ y∂yPy(τ, x, z)f (t − τ, x − z) − f (τ, x)dz + f (t, x) Rn+1 y∂yPy(τ, x, z)dτ dz = I1 + I2. (cid:90) (cid:90) Rn+1 ≤ C(cid:107)f(cid:107) (cid:90) ∞ (cid:12)(cid:12)(cid:12)f (t, x) 0 (cid:90) (cid:90) ∞ Rn C0,α/2,α H y∂y(ye−y2/4τ ) 0 (cid:90) ∞ √ 1 2 π 0 dτ τ 3/2 (cid:16) ye− y2+z2 τ cτ (τ 1/2 + z)α n+3 2 dτ dz ≤ C(cid:107)f(cid:107) yα. Cα/2,α t,H Regarding I2, as = 0 we can write I2 = y∂y(ye−y2/4τ ) e−τL1(t, x) − 1 (cid:17) dτ τ 3/2 (cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107) yα. Cα/2,α t,H Where in the last inequality we have used Lemma 7.33 (3). Conversely, suppose that f ∈ ΛαL. We can write f (t + τ, x + z) − f (t, x) = (Pyf (t + τ, x + z)−Pyf (t, x)) + (f (t + τ, x + z)−Pyf (t + τ, x + z)) + (Pyf (t, x)− f (t, x)). Let y = τ 1/2 + z. For the second summand we have (cid:13)(cid:13)(cid:13)f (t + τ, x + z) − Pyf (t + τ, x + z) (cid:13)(cid:13)(cid:13)(cid:13)− (cid:90) y 0 (cid:13)(cid:13)(cid:13)∞ = (cid:90) y ∂Py(cid:48)f (t + τ, x + z) ∂y(cid:48) dy(cid:48)(cid:13)(cid:13)(cid:13)(cid:13)∞ ≤ C(cid:107)f(cid:107) C0,α/2,α H 0 y(cid:48)−1+αdy(cid:48) = C(cid:107)f(cid:107)ΛαLyα = C(cid:107)f(cid:107)ΛαL(τ 1/2 + z)α. A similar estimate can be performed for the third summand. On the other hand by the Mean Value Theorem and Lemma 4.21, we have Pyf (t + τ, x + z) − Pyf (t, x) ≤ Pyf (t + τ, x + z) − Pyf (t + τ, x) + Pyf (t + τ, x) − Pyf (t, x) (3.12) ≤ ∇xPyf (t + τ, x + θz)z + ∂tPyf (t + λτ, x)τ. We observe that by the semigroup property, integration by parts and Lemma 7.34 (ii), we have(cid:12)(cid:12)(cid:12)∂xi∂yPyf (t, x) (cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:90) Rn+1 ∂xiPy/2(τ, x, z)∂yPyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 11 Hence as by Lemma 7.34 we have ∂xiPyf (t, x) ≤ C/y, then Finally we shall see that (1 + x)αf ∈ L∞(Rn+1). Given k ∈ N, a direct application of (cid:90) y ≤ Rn+1 + Rn+1 Pyf (t + τ, x + z) − Pyf (t, x) ≤ C(cid:107)f(cid:107)ΛαL(τ 1/2 + z)α. (cid:12)(cid:12)(cid:12) Py/2(τ, x, z)∂xi∂yPyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz (cid:90) (cid:12)(cid:12)(cid:12)(∂xi + ∂zi)Py/2(τ, x, z)∂yPyf (t − τ, x − z)(cid:12)(cid:12)y/2 (cid:12)(cid:12)(cid:12)dτ dz ≤ Cy−2+α. (cid:12)(cid:12)(cid:12)(cid:90) ∞ ∂y(cid:48)∂xiPy(cid:48)f (t, x)dy(cid:48)(cid:12)(cid:12)(cid:12) ≤ Cy−1+α. (cid:12)(cid:12)(cid:12)∂xiPyf (t, x) (cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) can be handled in a parallel way, this time using point (iv) of (cid:12)(cid:12)(cid:12)∂tPyf (t, x) (cid:12)(cid:12)(cid:12) ≤ Cy−2+α. Then going back to (3.12) we have (cid:12)(cid:12)(cid:12)∂tPyf (t, x) (cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107)∞y−(k+γ+s), s > 0. Moreover by the semigroup (cid:12)(cid:12)(cid:12)∂k yPyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz. For k = i Pyf (t, x) =(cid:82) (cid:12)(cid:12)(cid:12)∂k (cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107)ΛαLy−(k+k−α). Then (cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107)ΛαLy−(k−α). Now for x > 1 and 0 < α < 1 (cid:12)(cid:12)(cid:12)xk i Pyf (t, x) (cid:16)Pyf (t, x) − P 1x f (t, x)(cid:17) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + C(cid:107)f(cid:107)ΛαL (cid:90) 1x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + C(cid:107)f(cid:107)ΛαL ≤ C. (cid:90) 1x Pyf (t, x) ≤ xα sup 0<y< 1x f (t, x) + P 1x ∂z1Pz1f (t, x)dz1 i Py/2(τ, x, z)∂k i Pyf (t, x) y xγ i Pyf (t, x) y xk Rn+1 xk −(1−α) 1 z dz1 (cid:3) [α] + 1, the hypothesis and Lemma 7.34 (ii) give The derivative Lemma 7.34, we get Lemma 7.34 (ii) gives property we have ∂k y xk an iterated integration gives we have xαf (t, x) ≤ xα sup 0<y< 1x ≤ xα sup 0<y< 1x ≤ xα sup 0<y< 1x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) y y 3.1. Elliptic Hermite setting. Let g an L∞(Rn) function. Consider, as in Remark 2.1 the function f (t, x) = g(x). It is t,H . Moreover, as Pyf (t, x) = Pyg(x), g ∈ ΛαH clear that if g ∈ CαH if and only if f ∈ Cα/2,α if and only if f ∈ ΛαL. Hence, for 0 < α < 1, Proposition 3.16 and Theorem 3.17 have as consequences the continuity of the functions g ∈ ΛαH and the identity ΛαH = CαH(Rn). 4. Proofs of Theorems 1.2 and 1.4. Coincidence with Holder spaces for α > 1. yPyf (cid:13)(cid:13)(cid:13)∂k (cid:13)(cid:13)(cid:13)L∞(Rn+1) yPyf(cid:13)(cid:13)L∞(Rn+1) (a) (cid:13)(cid:13)∂k (b) (cid:13)(cid:13)∂l yPyf(cid:13)(cid:13)L∞(Rn+1) Remark 4.18. Observe that for bounded functions f , Lemma 7.34 assures that ≤ C(cid:107)f(cid:107)∞y−k. Therefore we can assume in Definition 1.1 that y < 1. Lemma 4.19. Let f ∈ L∞(Rn+1), α > 0, and k, l integers bigger than α. Then, for y > 0, the following conditions are equivalent: ≤ Aky−k+α ≤ Aly−l+α, where Ak and Al are positive constants with Ak ∼ Al. 12 DE LE ´ON-CONTRERAS AND TORREA Proof. Let l = k + 1. By using the semigroup property we have (cid:12)(cid:12)(cid:12)y/2 (cid:12)(cid:12)(cid:12)y/2 ∂yPy(τ, x, z) yPyf (t − τ, x − z) ∂k dτ dz. (cid:90) Rn+1 yPyf (t, x) = ∂l (cid:90) ∞ By Lemma 7.34 (ii) we get (a) =⇒ (b). For the converse, Remark 4.18 allows the integration yPyf (t, x) = (cid:3) ∂k Corollary 4.20. Let α > 0. If f ∈ ΛαL, then for every 0 < β < α, f ∈ ΛβL. z Pzf (t, x)dz, that gives the result. ∂k+1 y For the proof of this Corollary observe that, given kα = [α] + 1, we have (for y < 1) (cid:13)(cid:13)(cid:13)∂kα y Pyf (cid:13)(cid:13)(cid:13) ≤ Akα(cid:107)f(cid:107)ΛαLy−kα+α ≤ Akα(cid:107)f(cid:107)ΛαLy−kα+β. Then, Lemma 4.19 gives the result. Lemma 4.21. Let α > 0, f ∈ ΛαL and k = [α] + 1. (1) For every γ ≥ 0 and m, j ∈ N0 such that γ + m + j ≥ k there exists a constant Cγ,m,j y ∂tPyf(cid:107)∞ ≤ (2) For every m such that m+2 ≥ k, there exists a constant Cm such that (cid:107)∂m xiPyf(cid:107)∞ ≤ Cγ,m,j(cid:107)f(cid:107)ΛαLy−(γ+m+j)+α. such that (cid:107) · γ∂m y ∂j Cy−(m+2)+α. xiPyf (t, x) y ∂j (cid:12)(cid:12)(cid:12)xγ∂m Proof. Observe that the case γ = j = 0 follows from the definition of the space ΛαL, so we will exclude it in the following. Let us analyze the case when m ≥ k. By the semigroup property and integration by parts we have (cid:12)(cid:12)(cid:12) = = xi (cid:90) (cid:12)(cid:12)(cid:12)xγ∂j (cid:90) (cid:12)(cid:12)(cid:12)xγ y Pyf(cid:12)(cid:12)y/2(cid:107)∞ Rn+1 ≤ C(cid:107)∂m Py/2(τ, x, z)∂m (cid:88) Rn+1 (∂xi + ∂zi)jPy/2(τ, x, z)∂m xγ∂p (cid:90) y Pyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz y Pyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) xiPy/2(τ, x, z)dτ dz zi∂q Rn+1 p+q=j ≤ Cγ,m,j(cid:107)f(cid:107)ΛαLy−(γ+m+j)+α. In the last inequality we have use the hypothesis on f and Lemma 7.34 (ii) in each summand. We have chosen s = j + γ in the case p − q + γ ≤ 0. While in the case p − q + γ > 0, we choose s = 2q. Now we prove (2) for m ≥ k. By the semigroup property, the hypothesis on f and Lemma 7.34 (iv) we have y ∂tPyf (t, x) = ∂m (cid:90) ∂τPy/2(τ, x, z)∂m y Pyf(cid:12)(cid:12)y/2(cid:107)∞ (cid:90) Rn+1 Rn+1 ≤ C(cid:107)∂m y Pyf (t − τ, x − z)(cid:12)(cid:12)y/2dτ dz ∂τPy/2(τ, x, z)dτ dz ≤ C(cid:107)f(cid:107)ΛαLy−(m+2)+α. In both cases, for m < k we start from the above estimates for the case m = k and then (cid:3) we perform an k − m iterated integration. Proposition 4.22. Let α > 0. If f ∈ ΛαL, then xαf ∈ L∞(Rn+1). Proof. If α is not an integer we can use the same argument as in the proof of Theorem 3.17. Let α = 1, by using the arguments in the proof of Theorem 3.17, we can obtain (cid:12)(cid:12)(cid:12)x2Pyf (t, x) (cid:12)(cid:12)(cid:12) + (cid:12)(cid:12)(cid:12)x∂vPvf (t, x) (cid:12)(cid:12)(cid:12)v= 1x ≤ (cid:107)f(cid:107)Λ1Ly−1. FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 13 Then, by using that ∂z1Pz1f (t, x) = −(cid:82) 1 z2Pz2f (t, x)dz2 + ∂vPvf (t, x) ∂2 x z1 , we have (cid:12)(cid:12)(cid:12)v= 1x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + (cid:12)(cid:12)(cid:12)xP1/xf (t, x) (cid:18) 1 (cid:19) x − y + C(cid:107)f(cid:107) dz1 (cid:12)(cid:12)(cid:12) Λ1/2,1 t,H xf (t, x) ≤ x sup 0<y< 1x ≤ x sup 0<y< 1x (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)− (cid:90) 1x (cid:16)(cid:90) 1x (cid:90) 1x (cid:90) 1x z2Pz2f (t, x)(cid:12)(cid:12) dz2dz1 + (cid:107)f(cid:107) (cid:12)(cid:12)∂2 z1 y z2Pz2f (t, x)dz2 + ∂vPvf (t, x) ∂2 y z1 (cid:17) x (cid:12)(cid:12)(cid:12)v= 1x (cid:90) 1x Λ1/2,1 t,H sup 0<y< 1x (cid:90) 1x ≤ (cid:107)f(cid:107) Λ1/2,1 t,H x sup 0<y< 1x y z1 z−1 2 dz2dz1 + C(cid:107)f(cid:107) . Λ1/2,1 t,H Since for every 0 < y < 1x we have (cid:90) 1x (cid:90) 1x y z1 x z−1 2 dz2dz1 = x = x (cid:18) (cid:18) 1 (cid:90) 1x (cid:20) y log x log (cid:19) − log z1 (cid:19) (cid:18) 1 (cid:19)(cid:18) 1 x (cid:18) 1 x − y − x − y (cid:19) (cid:18) 1 (cid:19) dz1 x log ≤ C, = xy log(xy) + x (cid:19)(cid:21) 1 x − 1 x − y log y + y we conclude that xf (t, x) ≤ C(cid:107)f(cid:107) . Λ1/2,1 t,H For the cases in which α is an integer bigger that 1, we have to write ∂z1Pz1f in terms of the integral of the derivative of order k, where k = [α] + 1, and proceed analogously. We (cid:3) leave the details to the interested reader. 4.1. Proof of Theorem 1.2. (cid:90) (cid:90) Proof. Proof of epigraph (1) in Theorem 1.2. Let k = [α] + 1. Since Rn+1 yPy(τ, x, z)f (t − τ, x + z)dτ dz = (cid:90) ∞ ∂k  ye− y2 4τ τ 3/2 and ∂k y 0 Rn+1 yPy(τ, x,−z)f (t − τ, x − z)dτ dz, ∂k  dτ = 0 we have (cid:90) (cid:90) (cid:16) yPy(τ, x, z)(f (t − τ, x − z) + f (t − τ, x + z) − 2f (t, x))dτ dz ∂k yPy(τ, x, z) − ∂k (cid:90) ∞ ∂k (cid:17) (cid:0)e−τL1(t, x) − 1(cid:1) dτ yPy(τ, x,−z)  ye− y2 f (t − τ, x − z)dτ dz Rn+1 ∂k y 1 2 + + Rn+1 √ f (t, x) 2 π 4τ τ 3/2 0 yPyf (t, x) = ∂k (4.13) 1 2 = I1 + I2 + I3. 14 DE LE ´ON-CONTRERAS AND TORREA By Lemma 7.34, I1 ≤ C sition 4.22 and the proof of Lemma 7.33 (3) to get e− y2+z2 (τ 1/2 + z)α τ Rn n+k cτ 0 2 dτ τ dz ≤ Cyα−k. For I3 we use Propo-  ye− y2 4τ τ 3/2 (cid:0)e−τL1(t, x) − 1(cid:1) dτ (cid:12)(cid:12)(cid:12)(cid:12) ≤ C[f ]M αyα. yPy(τ, x, z) − ∂k (∂k yk∂k y 0 π √ 2 (cid:90) ∞ (cid:90) ∞ (cid:90) (cid:12)(cid:12)(cid:12)(cid:12) f (t, x)  ye− y2   ye− y2 (cid:90) (cid:90) ∞ 4τ τ 3/2 ∂k y 2 C(cid:107)f(cid:107)∞ ∂k y Regarding I2, we have ykI3 = (cid:90) (cid:90) ∞ Rn+1 2I2 = (cid:90) = Rn ∂k y 4τ τ 3/2 × f (t − τ, x − z)dτ dz. 0 2I2 ≤ C (cid:90) ∞ 0 ≤(cid:124)(cid:123)(cid:122)(cid:125) yPy(τ, x,−z))f (t − τ, x − z)dτ dz e− z2 √ π(2π sinh(2τ ))n/2 e− 2x−z2 4 coth τ (cid:18) 4 tanh τ − e− 2x+z2 4 (cid:19) tanh τ By the Mean Value Theorem applied to the function e− 2x−z2 4 tanh τ we get 4 coth τ e− z2 (sinh(2τ ))n/2 Rn  ye− y2 4τ τ 3/2 (cid:90) Rn (tanh τ )1/2zf (t − τ, x − z)dzdτ e−w2w(tanh τ )1/2 (sinh(2τ ))n/2(coth τ ) dwdτ n+1 z 2 0 0 cτ √ n+1 =w dτ coth τ 2 (tanh τ )1/2 ≤ Ck(cid:107)f(cid:107)∞ ≤ Ck(cid:107)f(cid:107)∞ (sinh(2τ ))n/2(coth τ ) e− y2 τ k/2+1 e− y2 cτ τ k/2 (cid:90) ∞ (cid:90) ∞ We conclude that f ∈ ΛαL. For the converse. If f ∈ ΛαL with α < 1. the result is a consequence of Theorem 3.17. If L = Cα(cid:48)/2,α(cid:48) α ≥ 1, by Theorem 3.17, f ∈ Λα(cid:48) for some α(cid:48) < 1, then (cid:107)y∂yPyf(cid:107)L∞(Rn+1) → 0, t,H as y → 0+. On the other hand, by the proof of Proposition 3.16 we know that (cid:107)Pyf − f(cid:107)L∞(Rn+1) → 0, as y → 0+ Hence we have y(cid:48) ∂2Py(cid:48)f (t, x) dy(cid:48) − y∂yPyf (t, x) + Pyf (t, x). ≤ Cky−k+α. τ α/2 dτ τ (cid:90) y f (t, x) = 0 2 We only do computations for g(t, x) = Pyf (t, x). For the other cases we have to follow the same path. By using Lemma 4.21 we have, for y = τ 1/2 + z, (∂y(cid:48))2 0 g(t − τ, x + z)+g(t − τ, x − z) − 2g(t, x) ≤(cid:12)(cid:12)D2 ≤ [∇xg(t − τ, x + θz) − ∇xg(t − τ, x − λz)]z + 2∂tg(t − ητ, x)τ xg(t − τ, x + νz)(cid:12)(cid:12) (θ + λ)z2 + 2∂tPyf (t − ητ, x)τ ≤ C(cid:107)f(cid:107)ΛαL (τ 1/2 + z)−2+α(z2 + τ ) ≤ C(cid:107)f(cid:107)ΛαL (τ 1/2 + z)α, where 0 < θ, λ < 1, −1 < ν < 1. The fact that (1 + x)αf ∈ L∞(Rn+1) follows from Proposition 4.22. FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 15 (cid:3) For the proof of epigraph (2) in Theorem 1.2, we shall prove the following theorem. Theorem 4.23. Suppose that α > 2. Then f ∈ ΛαL if and only if ∂xif, xif ∈ Λα−1L , i = 1, . . . , n, and ∂tf ∈ Λα−2L . In this case the following equivalence holds ∼ n(cid:88) i=1 (cid:16)(cid:107)∂xif(cid:107)Λα−1L (cid:107)f(cid:107)Λα−2H (cid:17) + (cid:107)xif(cid:107)Λα−1L + (cid:107)∂tf(cid:107)Λα−2L . For the reader's convenience, the proof of this Theorem 4.23 will be divide in several steps. Proposition 4.24. Suppose that f ∈ ΛαL with α > 2. Then, ∂tf ∈ Λα−2L . (cid:12)(cid:12)(cid:12)y=1 Proof. Let 2 < α < 3, by Lemma 4.21 we have (4.14) y ∂z∂tPzf dz + ∂tPyf If y < 1 we have ∂tPyf = (cid:82) 1 (cid:107)∂y∂tPyf(cid:107)∞ ≤ (cid:107)f(cid:107)ΛαLy−3+α (cid:82) y(cid:48) , this implies ∂tPyf is in L∞(Rn+1) uniformly on y. Moreover since ∂tPy(cid:48)f − ∂tPyf ≤ (cid:107)f(cid:107)ΛαL y z−3+αdz → 0 as (y(cid:48), y) → 0, then ∂tPyf converges uniformly when y → 0. As Pyf converges uniformly to f when y → 0, we conclude that ∂tf exists, it is the uniform limit of ∂tPyf = Py∂tf. Hence ∂yPy∂tf = ∂y∂tPyf. The last identity together with inequality (4.14) implies ∂tf ∈ Λα−2L . If α ≥ 3, by Corollary 4.20, the function f ∈ ΛβL for some β < 1. Hence by the thoughts developed before, ∂tf exists and ∂tPyf = Py∂tf . The proof follows the lines of the case 2 < α < 3. (cid:3) Proposition 4.25. Suppose that f ∈ ΛαL with α > 1. Then, Proof. Let 1 < α < 3. By Lemma 4.21 we have integration gives ∈ Λα−1L , i = 1, . . . , n. ∂f ∂xi ≤ C(cid:107)f(cid:107)ΛαLy−3+α. For y < 1, an (cid:13)(cid:13)(cid:13)(cid:13) ∂3Pyf (cid:12)(cid:12)(cid:12) ≤ C(cid:107)f(cid:107)ΛαLy−2+α + C (cid:13)(cid:13)(cid:13)(cid:13)∞ (cid:13)(cid:13)(cid:13) ∂2Pyf ∂y2∂xi ∂y∂xi (cid:12)(cid:12)(cid:12)y=1 (cid:13)(cid:13)(cid:13)∞. (cid:12)(cid:12)(cid:12) ∂2Pyf (t, x) ∂y∂xi (cid:13)(cid:13)(cid:13) ∂f (cid:13)(cid:13)(cid:13)∞ We can proceed as in the proof Proposition 4.24 and we get that ∂xif does exist and C. To prove that ∂f ∂xi that ∈ Λα−1L , we shall see that (cid:107)∂2 ≤ yPy(∂xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−3+α. Observe ∂xi yPy(∂xif )(t, x) = ∂xi∂2 ∂2 yPyf (t, x) − ∂xi∂2 yPy(τ, x, z)f (t − τ, x − z)dτ dz = I + II. (cid:90) Rn+1 By Lemma 4.21 we have that I ≤ C(cid:107)f(cid:107)ΛαLy−3+α = C(cid:107)f(cid:107)ΛαLy−2+(α−1). As f ∈ ΛαL, 1 < α < 3, by Proposition 4.22 we know that xf ∈ L∞(Rn+1). Hence, by Lemma 7.34 (iii) we get that II ≤ C(cid:107)f(cid:107)ΛαLy−3+α = C(cid:107)f(cid:107)ΛαLy−2+(α−1). Suppose now 3 ≤ α < 5. By Corollary 4.20 f ∈ ΛβL for all β < 3. Then, the result just ∈ ΛδL, for all δ < 1. We shall see that yPy(∂xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−4+(α−1). As Py(∂xif ) satisfies (2.8), it is enough to prove that ∈ ΛγL, for all γ < 2 and ∂2f ∂x2 i proved says that ∂f ∂xi (cid:107)∂4 (cid:107)∂2 y (−(cid:80)n j=1 ∂2 xj + x2 + ∂t)Py(∂xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−5+α. = ∂2 y ∂2 (cid:90) − 2 (cid:90) − Rn+1 Rn+1 y ∂xi∂xjPy(τ, x, z)∂xj f (t − τ, x − z)dτ dz ∂2 y ∂xiPy(τ, x, z)∂2 ∂2 xj f (t − τ, x − z)dτ dz. 16 Observe that DE LE ´ON-CONTRERAS AND TORREA (cid:90) ∂2 y ∂2 xjPy(∂xif )(t, x) = ∂2 y ∂2 xj ∂xiPyf (t, x) − (cid:90) xj ∂xiPyf (t, x) − ∂2 xj Rn+1 ∂2 y ∂2 xj ∂xiPy(τ, x, z)f (t − τ, x − z)dτ dz Rn+1 y ∂xiPy(τ, x, z)f (t − τ, x − z)dτ dz ∂2 The first summand is bounded by C(cid:107)f(cid:107)ΛαLy−5+α because of Lemma 4.21. As f and ∂xif are bounded functions, by using Lemma 7.34 (ii) we get the desired boundedness for the second and third summand. Finally Lemma 7.34 (iii) says that the forth summand is bounded by Cy−(1−ν+s), where ν < 1 and s > 0, then by choosing ν and s with s − ν = 4 − α we get the estimate. On the other hand, by using Lemma 7.34 (ii) and (iii) together with the facts that f, ∂xif ∈ L∞(Rn+1) and ∂2f ∂x2 i ∈ Λβ−2L , we get the desired estimate in this case. yPy(∂xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−5+α, we write To prove that (cid:107) · 2∂2 yPy(τ, x, z)f (t − τ, x − z)dτ dz x2∂2 By Lemma 4.21 we know that the first summand is bounded by C(cid:107)f(cid:107)ΛαLy−5+α. For the yPyf (t, x) − x2 ∂xi∂2 Rn+1 yPy(∂xif )(t, x) = x2∂xi∂2 (cid:90) second summand we have (cid:90) x2 Rn+1 ∂xi∂2 yPy(τ, x, z)f (t − τ, x − z)dτ dz ≤ C yPy(τ, x, z)(x − z2 + z2)f (t − τ, x − z)dτ dz, and by Lemma 7.34 (iii) applied to x2f and Lemma 7.34 (ii) we get the desired bound C(cid:107)f(cid:107)ΛαLy−5+α. ∂xi∂2 Rn+1 (cid:90) To get the estimate for (cid:107)∂2 y ∂tPy(∂xif )(t, x) = ∂2 ∂2 y ∂tPy(∂xif )(cid:107)∞, we write y ∂xi∂tPyf (t, x) − (cid:90) Rn+1 ∂xi∂2 yPy(τ, x, z)∂tf (t − τ, x − z)dτ dz. By Proposition 4.24 we know that ∂tf ∈ Λα−2L , 1 ≤ α − 2 < 3. Hence, as ∂2 y ∂xi∂tPyf (t, x) = y ∂xiPy(∂tf )(t, x), by applying Lemma 4.21 (1) we get that the first summand is bounded ∂2 by C(cid:107)f(cid:107)ΛαLy−5+α, and by Lemma 7.34 (iii) applied to ∂tf we get the same bound for the The rest of the cases, 2m + 1 ≤ α < 2m + 3, can be handled analogously by estimating (cid:3) second summand. the norms (cid:107)∂2 xj + x2 + ∂t)mPy(∂xif )(cid:107)∞. We leave the details to the reader. Proposition 4.26. Suppose that f ∈ ΛαL with α > 1. Then, xif ∈ Λα−1L , i = 1, . . . , n. Proof. Consider the case 1 < α < 2. By Proposition 4.22 we know that xif ∈ L∞(Rn+1). In addition, we can write y (−(cid:80) j ∂2 ∂yPy(xif )(t, x) = xi∂yPyf (t, x) − ziPy(τ, x, z)f (t − τ, x − z)dτ dz, (cid:90) Rn+1 FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 17 boundedness of f for the second summand, we get that (cid:107)∂yPy(xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−2+α. and by using Lemma 4.21 for the first summand and Lemma 7.34 together with the Let 2 ≤ α < 3. We have to prove that (cid:107)∂2 yPy(xif )(cid:107)∞ ≤ C(cid:107)f(cid:107)ΛαLy−3+α. As Py(xif ) satisfies (2.8) we have yPy(xif )(cid:13)(cid:13)∞ = (cid:13)(cid:13)∂2 ∂t − n(cid:88) j=1 (cid:105)Py(xif ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ xj − x2) (∂2 n(cid:88) ≤ (cid:107)∂tPy(xif )(cid:107)∞ + (cid:107)∂2 xjPy(xif )(cid:107)∞ + (cid:107) · 2Py(xif )(cid:107)∞. As ∂tf is well defined and bounded, see Proposition 4.24, ∂tPy(xif )(t, x) = Py(τ, x, z)(xi − zi)∂tf (t − τ, x − z)dτ dz = xiPy(∂tf )(t, x) − ziPy(τ, x, z)∂tf (t − τ, x − z)dτ dz. j=1 (cid:90) Rn+1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:104) (cid:90) Rn+1 Therefore, by using Proposition 4.24 and Lemma 4.21 (1) for ∂tf , we get that the first y−1+(α−2) ≤ C(cid:107)f(cid:107)ΛαLy−3+α. For the second summand xjPy(xif )(cid:107)∞, for j = 1, . . . , n. We can write, for every j ∈ summand is bounded by C(cid:107)∂tf(cid:107)Λα−2L we use that ∂tf ∈ L∞(Rn+1) and Lemma 7.34 (ii). To get the bound for (cid:107)∂2 {1, . . . , n}, xjPy(xif )(t, x) = ∂2 ∂2 xj Py(τ, x, z)xif (t − τ, x − z)dτ dz (cid:90) (cid:90) Rn+1 − ∂2 xj (cid:90) Rn+1 (cid:90) (cid:90) − − 2 Rn+1 Rn+1 Py(τ, x, z)zif (t − τ, x − z)dτ dz = xi∂2 xjPyf (t, x) + 2δi,j ∂xj (Py(τ, x, z)f (t − τ, x − z))dτ dz Rn+1 xj (Py(τ, x, z)zi)f (t − τ, x − z)dτ dz ∂2 ∂zi(∂xjPy(τ, x, z)zi)f (t − τ, x − z)dτ dz (cid:90) ∂zj (Py(τ, x, z)zi)∂xj f (t − τ, x − z)dτ dz, where δi,j = 1 if i = j and δi,j = 0 if i (cid:54)= j. Observe that in the last summand we have used integration by parts. As f ∈ ΛαL, by Lemma 4.21 (1) we get that the first summand is bounded by C(cid:107)f(cid:107)ΛαLy−3+α. For the rest of summands we can apply Lemma 7.34 (ii) since f and ∂xif are bounded functions. Rn+1 − In remains the case (cid:107) · 2Py(xif )(cid:107)∞. Observe that, x2Py(xif )(t, x) ≤ C (cid:90) ≤ C (x − z2 + z2)Py(τ, x, z)xi − zif (t − τ, x − z)dτ dz Rn+1 x − z3−αPy(τ, x, z)x − zαf (t − τ, x − z)dτ dz z2Py(τ, x, z)x − zf (t − τ, x − z)dτ dz (cid:90) (cid:90) Rn+1 + C Rn 18 Rn+1 DE LE ´ON-CONTRERAS AND TORREA (cid:90) by estimating the norms (cid:107)(−(cid:80) ≤ C((cid:107)xαf(cid:107)∞ + (cid:107)xf(cid:107)∞) (x3−α + z3−α + z2)Py(τ, x, z)dτ dz ≤ C(cid:107)f(cid:107)ΛαLy−3+α. In the last inequality we have used Lemma 7.34 (ii). It remains For the cases 2m + 1 ≤ α < 2m + 3, with m ≥ 1, we get the result by following the same kind of reasonings, that is, xj + x2 + ∂t)m+1Py(xif )(cid:107)∞. We leave the details for the (cid:3) Proposition 4.27. Let α > 2 and f ∈ L∞(Rn+1) and suppose that ∂xif, xif ∈ Λα−1L , i = 1, . . . , n, and ∂tf ∈ Λα−2L . Then f ∈ ΛαL. Proof. Consider the case 2 ≤ α < 4. We want to see that (cid:107)∂4 Pyf satisfies (2.8), we have that ∂4 is sufficient to prove that xj + x2(cid:17)2 Pyf (t, x). Hence it yPyf(cid:107)∞ ≤ Cy−4+α, and as ∂t −(cid:80)n yPyf (t, x) = interested reader. j=1 ∂2 (cid:16) j ∂2 + 4 Rn+1 (cid:90) (cid:90) (cid:90) + + Rn+1 Rn+1 Py(τ, x, z)xj∂xj f (t − τ, x − z)dτ dz x∂xjPy(τ, x, z)x∂xj f (t − τ, x − z)dτ dz x2∂2 xjPy(τ, x, z)f (t − τ, x − z)dτ dz + x2∂xjPy(∂xj f )(t, x). As the functions f and x∂xj f are bounded, Lemma 7.34 takes care of the first to forth summands. The bound of last summand in (4.15) follows from the fact that ∂xj f ∈ Λα−1L and Lemma 4.21. xjPyf(cid:107)∞ ≤ Cy−4+α, j ∈ {1, . . . , n}, and xjPyf(cid:107)∞ ≤ Cy−4+α, xj (x2Pyf )(cid:107)∞ ≤ Cy−4+α, a) (cid:107)∂4 b) (cid:107)∂2 c) (cid:107)x2∂2 d) (cid:107)x4Pyf(cid:107)∞ ≤ Cy−4+α. e) (cid:107)∂2 t Pyf(cid:107)∞ ≤ Cy−4+α. f) (cid:107)∂tx2Pyf(cid:107)∞ ≤ Cy−4+α. xjPyf(cid:107)∞ ≤ Cy−4+α. g) (cid:107)∂t∂2 (cid:90) Integration by parts gives Rn+1 xj (cid:90) xjPyf (t, x) = ∂3 ∂4 (cid:90) (cid:90) + 2 + Rn+1 Rn+1 + Rn+1 Py(τ, x, z)∂xj f (t − τ, x − z)dτ dz xj (∂xj + ∂zj )Py(τ, x, z)f (t − τ, x − z)dτ dz ∂3 xj (∂xj + ∂zj )Py(τ, x, z)∂xj f (t − τ, x − z)dτ dz ∂2 xj f (t − τ, x − z)dτ dz. ∂xj (∂xj + ∂zj )Py(τ, x, z)∂2 As ∂xj f ∈ Λα−1L , by Lemma 4.21 we get that the first summand is bounded by C(cid:107)∂xj f(cid:107)Λα−1L For the rest of the summands we apply Lemma 7.34 together of the boundedness of the func- tions f, ∂xj f and ∂2 y−4+α. xj f. To prove b), we write xj (x2Pyf )(t, x) = 2Pyf (t, x) + 4 ∂2 (4.15) (cid:90) Rn+1 xj∂xjPy(τ, x, z)f (t − τ, x − z)dτ dz FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 19 To see d), we use that xαf ∈ L∞(Rn+1) and Lemma 7.34 to get (cid:90) x4Pyf (t, x) ≤ C Rn+1 ≤ C((cid:107)xαf(cid:107)∞ + (cid:107)f(cid:107)∞) Rn+1 ≤ C((cid:107)xαf(cid:107)∞ + (cid:107)f(cid:107)∞)y−4+α. (cid:90) Py(τ, x, z)(x − z4−αx − zα + z4)f (t − τ, x − z)dτ dz Py(τ, x, z)(x4−α + z4−α + z4)dτ dz Finally, for the estimates e)-g) observe that (cid:107)∂2 t Pyf(cid:107)∞ = (cid:107)∂tPy(∂tf )(cid:107)∞,(cid:107)∂t∂2 xjPyf(cid:107)∞ = (cid:107)∂2 xjPy(∂tf )(cid:107)∞, and (cid:107)x2∂tPyf(cid:107)∞ = (cid:107)x2Py(∂tf )(cid:107)∞. Hence, by using that ∂tf ∈ Λα−2L we get the result. and Lemma 4.21 (1) For the rest of the values of α we proceed analogously. We leave the details for the (cid:3) interested reader. This is the end of the proof of Propostion 4.27. Propositions 4.24, 4.25, 4.26, 4.27 show the validity of Theorem 4.23. Therefore we have proved Theorem 1.2, epigraph (2). The proof of Theorem 1.4 is now complete. As a consequence of it we get the following characterization of the spaces of Krylov's type introduced in Definition 3.15. Theorem 4.28. Let 0 < α < 3, α not an integer. Then Cα/2,α t,H = ΛαL, with equivalence of norms. Proof. The case 0 < α < 1 was proved in Theorem 3.17. Consider 1 < α < 2. Suppose that f ∈ ΛαL. By Theorem 1.2(1) we know that (1.2) holds, an by taking z = 0 in this inequality we get that f (·, x) ∈ Cα/2(R) uniformly on x. In addition, by Propositions 4.25 and 4.26 we have that (∂xi ± xi)f ∈ Λα−1L = C α 2 ,α t,H . Conversely, t,H . Then, we have that (∂xi ± xi)f (t,·) ∈ Cα−1H suppose that f ∈ C uniformly on t and f (·, x) ∈ Cα/2(R) uniformly on x. Hence, (1 + x)αf ∈ L∞(Rn+1) and f (t − τ, x − z) + f (t − τ, x + z) − 2f (t, x) . Thus, we get that f ∈ C α−1 2 ,α−1 t,H α 2 ,α ≤ f (t − τ, x − z) + f (t − τ, x + z) − 2f (t − τ, x) + 2f (t − τ, x) − f (t, x) ≤ C∇xf (t − τ, x + θz) − ∇xf (t − τ, x − λz)z + Cτ α/2 ≤ Cθ + λα−1zα−1z + Cτ α/2 ≤ C(τ 1/2 + z)α. By Theorem 1.2 (1) we conclude that f ∈ ΛαL. The case 2 < α is a Corollary of Theorem (cid:3) 4.23. 4.2. Elliptic Hermite setting. Again as in the case of subsections 2.1 and 3.1 we handled the functions g(x) and f (t, x) = g(x). The considerations made in that Remarks, together with Theorems 1.2 and 4.23 give the proof of Theorem 1.4. Moreover the following Theorem is also true. Theorem 4.29. If α > 0 is not an integer, we have CαH = ΛαH. 20 DE LE ´ON-CONTRERAS AND TORREA Consider the functions h and ϕ as follows. h(x) =(cid:80)∞ Remark 4.30. There exists a function g ∈ Λ1H(R), but so that sup{x:x∈[0,1],z∈[0,1]} g(x + z)− g(x) ≤ Cz fails for all C. k=1 2−k cos2π2kx and ϕ is a positive differentiable function, with continuous derivative, such that ϕ(x) = 1 when x ∈ [−3, 3], and for any x there exist a constant C with (1 + x)ϕ(x) ≤ C and ϕ(cid:48)(x) ≤ C. It is clear that h(x) ≤ 1, moreover it can be checked, see [21, Theorem 4.9], that (cid:107)h(x + z) + h(x − z) − 2h(x)(cid:107)∞ ≤ Az. (1 + x)g(x) ≤ C. On the other hand by the Mean Value Theorem we have Now we choose the function g(x) = h(x)ϕ(x), then by the properties of h and ϕ we have (cid:12)(cid:12)(cid:12)g(x + z) + g(x − z) − 2g(x) (cid:12)(cid:12)(cid:12)h(x − z) (ϕ(x − z) − ϕ(x + z)) (cid:12)(cid:12)(cid:12) + 2 (cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12) (h(x + z) + h(x − z) − 2h(x)) ϕ(x + z) (cid:12)(cid:12)(cid:12)h(x) (ϕ(x + z) − ϕ(x)) (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ≤ Cz. Now assume that g satisfies g(x + z) − g(x) ≤ Cz. Hence for x, z ∈ [0, 1] we would have h(x+z)−h(x) ≤ Cz. But it is well know that Weierstrass function doesn't satisfy Lipschitz condition, see [21, Theorem 4.9]. + 5. Schauder and Holder estimates Lemma 5.31. Let α, β positive real numbers. (a) Let 0 < 2β < α and f ∈ ΛαL then we have Lβf (t, x) ≤ C < ∞, (t, x) ∈ Rn+1. (b) For every β > 0 and f ∈ L∞(Rn+1) we have L−βf (t, x) ≤ C < ∞, for all (t, x) ∈ Proof. It suffices to consider the case 2β < α < [2β] + 1 = (cid:96). Then (cid:107)(Pνf (t, x) − f (t, x))[2β]+1(cid:107)L∞(Rn+1) = ∂y1 . . . ∂y(cid:96)Py1+...y(cid:96)f (t, x)dy(cid:96) . . . dy1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:90) ν 0 (cid:90) ν 0 . . .(cid:124)(cid:123)(cid:122)(cid:125) (cid:96) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ Rn+1. (cid:90) ν 0 (cid:90) ν 0 . . .(cid:124)(cid:123)(cid:122)(cid:125) (cid:96) ≤ C (y1 + . . . y(cid:96))−(cid:96)+αdy(cid:96) . . . dy1 ≤ Cνα. Then, as 0 < 2β < α, and f, Pνf ∈ L∞(Rn+1) we have (5.16) Lβf (t, x) ≤ cβ να ν1+2β dν + (cid:90) ∞ 1 (cid:90) 1 0 1 ν1+2β ≤ C < ∞. (cid:18)(cid:90) 1 (cid:90) ∞ dν ν1−2β + 0 (cid:19) To prove (b) we use the boundedness of f for ν < 1 and Lemma 7.34 (ii), with sβ > 2β when ν > 1. Thus, L−βf (t, x) = 1 (cid:90) ∞ Γ(2β) 0 Pνf (t, x) dν ν1−2β ≤ Cβ(cid:107)f(cid:107)∞ dν ν1+sβ−2β 1 ≤ Cβ. (cid:3) Proof of Theorem 1.5. Let m = [α − 2β] + 1 and (cid:96) = [2β] + 1. Then, m + (cid:96) = [α − 2β] + 1 + [2β] + 1 > α − 2β + 2β = α, as m + (cid:96) ∈ N we get m + (cid:96) ≥ [α] + 1. Previous Lemma 5.31 and Fubini's Theorem allow us to write (cid:12)(cid:12)(cid:12)∂m y P(Lβf ) (cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)cβ (cid:90) ∞ 0 y Py ∂m (cid:16)(cid:90) ν . . .(cid:124)(cid:123)(cid:122)(cid:125) 0 (cid:96)=[2β]+1 (cid:90) ν 0 wPww=s1+···+s(cid:96)ds1 . . . ds(cid:96) ∂(cid:96) (cid:17) dν ν1+2β (cid:12)(cid:12)(cid:12) FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 21 (cid:16)(cid:90) ν (cid:16)(cid:90) ν 0 0 0 (cid:90) ∞ (cid:90) ∞ (cid:90) y 0 (cid:12)(cid:12)(cid:12)cβ = ≤ Cβ = Cβ . . .(cid:124)(cid:123)(cid:122)(cid:125) . . .(cid:124)(cid:123)(cid:122)(cid:125) (cid:96)=[2β]+1 0 (cid:90) ν (cid:90) ν (cid:90) ∞ 0 (cid:96)=[2β]+1 dν ν1+2β + Cβ (. . . ) w Pww=y+s1+···+s(cid:96)ds1 . . . ds(cid:96) ∂m+(cid:96) (y + s1 + . . . s(cid:96))−(m+(cid:96))+αds1 . . . ds(cid:96) (. . . ) dν ν1+2β = I + II, y (cid:12)(cid:12)(cid:12) (cid:17) dν (cid:17) dν ν1+2β ν1+2β 0 (cid:90) y (cid:90) y 0 0 0 0 (cid:96)=[2β]+1 (cid:90) ν/y (cid:90) ν/y . . .(cid:124)(cid:123)(cid:122)(cid:125) (cid:17)(cid:96) dν (cid:16) ν ν1+2β ≤ Cβy−m+α−(cid:96) (cid:90) ∞ (cid:90) ∞ y y (cid:90) y 0 where in the last inequality we have used that m + (cid:96) ≥ [α] + 1 > α. Now we shall estimate I and II. I ≤ Cβy−m+α ≤ Cβy−m+α (1 + s1 + . . . s(cid:96))−(m+(cid:96))+αds1 . . . ds(cid:96) dν ν1+2β dν ν1+2β−(cid:96) ≤ Cβy−m+α−2β. (cid:16) II ≤ cβ Notice that in the last inequality we have used that 1 + 2β − (cid:96) = 2β − [2β] < 1. On the other hand, (y + ν)−m+α + y−m+α(cid:17) dν ν1+2β = Cy−m+α−2β. While in the case −m+α > ν1+2β ≤ (cid:3) Proof of Theorem 1.6. Let (cid:96) = [α + 2β] + 1 > [α] + 1 > α. Fubini Theorem together with If −m+α ≤ 0 we have II ≤ C 0, as m − α + 2β + 1 = [α − 2β] + 1 − α + 2β + 1 > 1, we get II ≤ C Cy−m+α−2β. ν−m+α dν y−m+α dν (cid:90) ∞ ν1+2β . y y Lemma 5.31 allow to get (cid:13)(cid:13)(cid:13)(cid:13)(cid:90) ∞ 0 (cid:13)(cid:13)(cid:13)(cid:13)∞ (cid:107)∂(cid:96) (cid:90) ∞ yPy(L−βf )(t, x)(cid:107)L∞(Rn+1) = ∂(cid:96) yPyPνf (t, x) dν ν1−2β ≤ C 0 (y + ν)−(cid:96)+α dν ν1−2β ≤ Cy−(cid:96)+α−2β. For (b) we apply Lemma 7.34 (ii), then for (cid:96) = [2β] + 1 we have ∂(cid:96) yPyPνf (t, x) ≤ C (cid:107)f(cid:107)∞ . y(cid:96) (cid:3) Proof of Theorem 1.7. Let f ∈ L∞(Rn+1), by using Lemma 7.34 (i) and (ii), we have min(1, s−2)ds ≤ C. Moreover if f ∈ ΛαL(Rn+1), α > 0 Then we can proceed as before. (cid:12)(cid:12)(cid:12)(cid:90) ∞ e−sL1/2 (cid:90) ∞ (cid:12)(cid:12)(cid:12) ≤ C (cid:16)(cid:90) ∞ (cid:12)(cid:12)(cid:12)(cid:90) ∞ ≤ C f (t, x)a(s)ds 0 0 (cid:12)(cid:12)(cid:12)∂(cid:96) and (cid:96) = [α + 1] + 1 > α + 1, by Fubini's Theorem we have (cid:12)(cid:12)(cid:12)(cid:90) ∞ (cid:17)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) ≤ Cy−(cid:96)+α+1. We have proved that the operator f −→(cid:82) ∞ Psf (t, x)a(s)ds (y + s)−(cid:96)+αds yPy 0 0 0 Then Theorem 1.5 gives the result. 0 e−sL1/2f a(s)ds maps ΛαL(Rn+1) into Λα+1L (Rn+1). (cid:3) (cid:12)(cid:12)(cid:12)w=y+s (cid:12)(cid:12)(cid:12) wPwf (t, x) ∂(cid:96) a(s)ds 22 DE LE ´ON-CONTRERAS AND TORREA Finally the proof of Theorem 1.8 is a direct consequence of Theorem 1.6 and 1.2. 5.1. Elliptic Hermite setting. As we did in the previous Sections, we consider g(x) and f (t, x) = g(x), then it can be easily checked that H±βg(x) = L±βf (t, x) and m(H) = m(L). Hence Remarks 2.1, 3.1 and 3.1 show the Hermite's version of Theorems 1.5, 1.6, 1.7 and 1.8. 6. Maximum and comparison principles. (cid:90) ∞ Proof of Theorem 1.9. Observe that cβ > 0 for [2β]+1 odd and cβ < 0 for [2β]+1 even. On the other hand as the kernel Pν(τ, x, z) is always positive we have Pνf (t, x) ≥ 0, t ≤ t0. If 0 < ν1+2β , then Lβf (t0, x0) ≤ 0. If 1/2 ≤ β < 1, then β < 1/2, Lβf (t0, x0) = ν1+2β , as (P2νf (t0, x0)−2Pνf (t0, x0)) ≤ 0, Lβf (t0, x0) = we obtain that Lβf (t0, x0) ≤ 0. (cid:3) (P2νf (t0, x0)−2Pνf (t0, x0)) (cid:90) ∞ Pνf (t0, x0) 1 cβ 1 cβ dν dν 0 0 The following remark will be used systematically along this manuscript. 7. Computational results Remark 7.32. Let τ > 0. (1) If τ < 1, then sinh τ ∼ τ , cosh τ ∼ C, coth τ ∼ 1 τ and tanh τ ∼ τ . (2) If τ > 1, then sinh τ ∼ eτ , cosh τ ∼ eτ , coth τ ∼ C and tanh τ ∼ C. (3) Given n ∈ N, (cid:96) ∈ N and λ ≥ 0, there exists a positive constant C(cid:96),n,λ such that (sinh τ )n ≤ C(cid:96),n,λ min(τ−n+(cid:96), e−cτ ) ≤ C(cid:96),n,λτ−n+(cid:96)−λ. (coth τ )(cid:96)(sinh τ )n = (tanh τ )(cid:96) (4) Let z ≥ 0 and α ≥ 0 there exist a constant Cα > 0 such that zαe−z ≤ Cαe−z/2. As usual by A ∼ B we mean there exist constants C1, C2 such that C1A ≤ B ≤ C2A. Lemma 7.33. For each x ∈ Rn and τ > 0, we have: 1 x2 e− tanh(2τ ) (cosh(2τ ))n/2 2 . (7.17) (1) e−τL1(t, x) = (2) ∂τ e−τL1(t, x) ≤ C(min{τ, 1} + x2). (3) Given 0 < α < 1, there exists Cα > 0 such that √L1(t, x) − 1 (cid:17) Proof. By using formula (2.3) we have (2π sinh(2τ ))n/2e−τL1(t, x) (cid:12)(cid:12)(cid:12) ≤ Cα(1 + x)αyα. (cid:16) − 1 coth τ (x2 + z2 − 2xz) (cid:12)(cid:12)(cid:12)e−y (cid:90) exp = (cid:17) tanh τ (x2 + z2 + 2xz) dz 4 = exp(− 1 4 Rn (cid:90) × Rn 4 exp (cid:16) − 1 (cid:16) − 1 (cid:16) − 1 exp 4 4 = exp x2(coth τ + tanh τ )) (cid:32)(cid:112)(coth τ + tanh τ )z − coth τ − tanh τ (cid:33)2 (cid:112)(coth τ + tanh τ ) x2(cid:17)(cid:90) (coth τ − tanh τ )2 (coth τ + tanh τ ) (cid:16) 1 (cid:17) exp x 4 Rn − (coth τ − tanh τ )2x2 (coth τ + tanh τ ) (cid:17) dz e− u2 4 du (coth τ + tanh τ )n/2 x2(coth τ + tanh τ ) FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 23 (cid:16) − 1 (cid:17) (sinh(2τ ))n/2 x2 tanh(2τ ) = exp Where we have done the change of variables u = (cid:112)(coth τ + tanh τ )z − coth τ−tanh τ (2 cosh(2τ ))n/2 2nπn/2. √ 2 (coth τ +tanh τ ) x. This concludes the proof of (1). By using the estimates of Remark 7.32, it is easy to show that ∂τ e−τL1(t, x) ≤ C For (3), consider first the case x > 1. By the Mean Value Theorem and parts (1), (2) in e−τL1(t, x) ≤ C(min{τ, 1} + x2). (cid:16) √ 1 (cid:90) ∞ tanh(2τ ) + (1 + tanh2(2τ ))x2(cid:17) (cid:12)(cid:12)(cid:12)(cid:12)(cid:16)(cid:90) 1/x2 (cid:17) ye− y2 (cid:90) 1/x2 (cid:90) ∞ (cid:16)x2y2 (cid:90) ∞ v1/2e−v dv ye− y2 4τ τ 1/2 4τ τ 1/2 v2 + x2τ 1/x2 dτ τ x2y2 + C + π 0 0 c 0 2 ≤ C (cid:12)(cid:12)(cid:12)(cid:12) dτ τ (cid:17) (e−τL1(t, x) − 1) (cid:90) ∞ (cid:90) x2y2 1/x2 4 ye− y2 4τ τ 1/2 dτ τ v1/2e−v dv v (cid:90) x2y2 4 this Lemma we get √L1(t, x) − 1 (cid:12)(cid:12)(cid:12)e−y (cid:12)(cid:12)(cid:12) = C =(cid:124)(cid:123)(cid:122)(cid:125) ≤ Cx2y2 (x2y2)1−α/2 y2 4τ =v v1/2−α/2e−v dv v + Cxαyα 0 x2y2 c v1/2−α/2e−v dv v ≤ CΓ(1/2 − α/2)xαyα. Regarding the case x < 1. Again, by the Mean Value Theorem we get (cid:12)(cid:12)(cid:12)(cid:90) 1 0 dτ τ (cid:12)(cid:12)(cid:12) (cid:90) ∞ (cid:90) 1/x2 (τ + x2)τ ≤ C dτ τ 0 Cx2y2 v1/2e−v dv v2 + C x2y2 4 (e−τL1(t, x) − 1) ye− y2 4τ τ 1/2 ≤ C (cid:90) 1 0 ye− y2 4τ τ 1/2 ≤(cid:124)(cid:123)(cid:122)(cid:125) (cid:90) ∞ x2y2 4 y2 4τ =v ≤ Cxαyα ye− y2 4τ τ 1/2 (cid:90) ∞ (cid:90) ∞ y2 4 y2 4 (cid:90) 1 (cid:19)2 dv (cid:18) y2 + 0 ye− y2 4τ τ 1/2 v v x2τ dτ τ v1/2e−v τ 2 dτ τ v1/2−α/2e−v dv v + Cyα v1/2−α/2e−v dv v ≤ CΓ(1/2 − α/2)yα. On the other hand, ye− y2 4τ τ 1/2 (cid:12)(cid:12)(cid:12)(cid:90) ∞ 1 (e−τL1(t, x) − 1) dτ τ (cid:90) ∞ (cid:12)(cid:12)(cid:12) ≤ C ye− y2 4τ τ 1/2 dτ τ ≤ Cyα 1 ≤ CΓ(1/2 − α/2)yα. (cid:90) ∞ 1 y1−αe− y2 τ 1/2−α/2 4τ dτ τ (cid:3) Lemma 7.34. Let Py(τ, x, z) the Poisson kernel associated with the parabolic harmonic oscillator, L, and given by (2.9). Then, 24 (cid:90) DE LE ´ON-CONTRERAS AND TORREA (cid:12)(cid:12)(cid:12)∂k (cid:12)(cid:12)(cid:12) ≤ Ck e− y2+z2 cτ (cid:12)(cid:12)(cid:12)Py(τ, x, z) (cid:12)(cid:12)(cid:12) ≤ Cy e− y2+z2 cτ (i) There exists a constant C such that for every x, z in Rn and τ > 0, τ−( n+k τ−( n+3 yPy(τ, x, z) 2 +1), for (ii) Let γ, ν ≥ 0, s ≥ 0. For each (cid:96), k, m ∈ N ∪ {0}, there exists a constant Cγ,ν,k,(cid:96),m,s > 0 2 ), and k ≥ 1. such that, for every x ∈ Rn and τ > 0, xγzν∂k for ζ = k + m − (cid:96) − ν + γ and i = 1, . . . , n, j = 1, . . . , n. Cs,α > 0 such that, for every x ∈ Rn and τ > 0, xjPy(τ, x, z)dzdz ≤ (cid:40) y ∂m zi ∂(cid:96) Cγ,ν,(cid:96),k,m,s y−(k+m−(cid:96)−ν+γ+s), Cγ,ν,(cid:96),k,m,s y−s, (iii) Let f such that xαf ∈ L∞(Rn+1), 0 < α ≤ 1 and s ≥ 0. There exists a constant if s ≥ 0, ζ > 0, if s > 0, ζ ≤ 0, Rn+1 (iv) There exists a constant C such that for every x ∈ Rn and τ > 0, yPy(τ, x, z)f (t − τ, x − z)dzdτ ≤ Cs,αy−(1−α+s). (cid:90) ∂τPy(τ, x, z)dzdτ ≤ Cy−2. Rn+1 Proof. Along this proof will use Remark 7.32 and the estimates: (cid:16) (cid:12)(cid:12)(cid:12)∂(cid:96) xi (cid:17)(cid:12)(cid:12)(cid:12) ≤ C(cid:96)e− 2x−z2 tanh τ 8 8τ τ−(k/2+1), e− 2x−z2 tanh τ 4 (tanh τ )(cid:96)/2 and (7.18) (cid:16) ye− y2 ∂k y zi e− z2 coth τ ∂m 4τ τ 3/2 4 (cid:17) ≤ Ck e− y2 ≤ Cme− z2 coth τ 8 (coth τ )m/2. Estimate (i) is consequence of Remark 7.32. In order to prove (ii), as (cid:90) ∂xi∂2 Rn+1 Cτ τ−(k/2+1)e− z2 coth τ C (coth τ )m/2e−x− z 22 tanh τ (tanh τ )(cid:96)/2, zi ∂(cid:96) xjPy(τ, x, z) ≤ ∂k y ∂m again by Remark 7.32, for every λ ≥ 0 we get (sinh τ )n/2 e− y2 C (cid:90) xγzν∂k (cid:90) ∞ (cid:90) Rn+1 ≤ C Rn 0 xjPy(τ, x, z)dzdz zi ∂(cid:96) y ∂m Cτ e− z2 coth τ 2γ + z 2γ)zνe− y2 (x − z (cid:90) ∞ τ (k/2+1)(sinh τ )n/2 (cid:90) C −x− z 2 2 tanh τ e (coth τ )m/2(tanh τ )(cid:96)/2 dτ τ dz ≤ C Rn 0 Cτ e− z2 coth τ e− y2 k+n+m−(cid:96)+γ−ν+λ τ C 2 dz ≤ C dτ τ e− y2 Cτ k+m−(cid:96)+γ−ν+λ 2 dτ τ . (cid:90) ∞ 0 τ The constant C depends on γ, ν, (cid:96), k, m and λ. The result follows by choosing λ = s in the case k + m− (cid:96) + γ − ν > 0, for k + m− (cid:96) + γ − ν ≤ 0 we choose λ = −(k + m− (cid:96) + γ − ν) + s in the case k + m − (cid:96) + γ − ν ≤ 0. (cid:90) Rn+1 For (iii), as xαf ∈ L∞(Rn+1), we have ∂xi∂2 yPy(τ, x, z)f (t − τ, x − z)dzdτ (cid:90) ∞ (cid:90) e− 2x−z2 tanh τ (cid:90) ∞ (cid:90) cτ e− z2 coth τ cτ e− z2 coth τ (tanh τ ) e− y2 e− y2 Rn 1+α 0 4 4 c ≤ C ≤ C Rn 0 τ (sinh(2τ ))n/2 tanh τ2x − z1−α2x − zαf (t − τ, x − z) τ (sinh(2τ ))n/2 dτ τ dz 2 x − zαf (t − τ, x − z) dτ τ dz FRACTIONAL POWERS OF THE PARABOLIC HERMITE OPERATOR. REGULARITY PROPERTIES 25 (cid:90) + C Rn ≤ C[f ]M α (cid:90) ∞ (cid:90) (cid:90) 0 Rn (cid:90) ∞ (cid:90) ∞ 0 + C(cid:107)f(cid:107)∞ (cid:90) ∞ ≤ C 0 τ 0 Rn e− y2 1−α+λ cτ 2 dτ τ . e− y2 cτ e− z2 coth τ 4 1+α 2 zαf (t − τ, x − z) (tanh τ ) τ (sinh(2τ ))n/2 dτ τ dz e− y2 cτ e− z2 coth τ 4 (tanh τ ) 1+α 2 τ (sinh(2τ ))n/2 cτ e− z2 coth τ e− y2 τ (sinh(2τ ))n/2(coth τ )α/2 tanh τ 4 dτ τ dz dτ τ dz The result follows by taking λ = s. We shall prove (iv) in the case of the first derivative, we leave the details for the second derivative to the reader. By using the ideas in the proof of (iii) we have (cid:90) Rn+1 (cid:90) ∞ (cid:90) ∞ 0 (cid:90) (cid:90) ≤ C ≤ C ∂τPy(τ, x, z)dzdτ ye− y2 cτ e− z2 coth τ c e− 2x−z2 tanh τ c Rn τ 3/2(sinh 2τ )n/2 ye− y2 cτ e− z2 coth τ c e− 2x−z2 tanh τ c 0 Rn τ 3/2(sinh 2τ )n/2 ≤ C τ (cid:16) 1 (cid:16) 1 (cid:90) ∞ τ + cosh(2τ ) sinh(2τ ) + + (cid:90) + cosh(2τ ) sinh(2τ ) e− y2 cτ e− z2 cτ τ 1+n/2 0 Rn y2 τ 2 + y2 τ 2 + z2 (sinh τ )2 + z2 (cid:90) ∞ (sinh τ )2 + e− y2 τ cτ 0 dτ ≤ C 1 τ (cid:17) (cid:17) dτ dτ 2x − z2 (cosh τ )2 2x − z2 (cosh τ )2 dτ τ ≤ C y2 . (cid:3) References [1] A. Bernardis, F. J. Mart´ın-Reyes, P. R. Stinga and J. L. Torrea, Maximum principles, extension problem and inversion for nonlocal one-sided equations, J. Differential Equations 260 (2016), 6333 -- 6362. [2] L. A. Caffarelli and L. Silvestre, An extension problem related to the fractional Laplacian, Comm. Partial Differential Equations 32 (2007), 1245 -- 1260. [3] L. A. Caffarelli and P. R. Stinga, Fractional elliptic equations, Caccioppoli estimates and regularity, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire 33 (2016), 767 -- 807. [4] A. E. Gatto and W.O. Urbina, On Gaussian Lipschitz spaces and the boundedness of fractional integrals and fractional derivatives on them, Quaest. Math. 38 (2015), 125. [5] D. Gilbarg and N. S. Trudinger, Elliptic Partial Differential Equations of Second Order, Classics in Mathematics, Springer, Berlin, 2002. [6] N. J. Kalton, Spaces of Lipschitz and Holder functions and their applications, Collect. Math. 55, (2004), 171217. [7] S. G. Krantz, Lipschitz spaces, smoothness of functions, and approximation theory, Exposition. Math. 1(3) (1983), 193 -- 260. [8] N. V. Krylov, Lectures on elliptic and parabolic equations in Holder spaces, Graduate Studies in Mathe- matics, Vol 12, American Mathematical Society, Providence, RI, 1996. [9] L. Liu and P. Sjogren, A characterization of the Gaussian Lipschitz space and sharp estimates for the Ornstein-Uhlenbeck Poisson kernel, Rev. Mat. Iberoam., 32, (2016), 11891210. [10] R.A. Mac´ıas and C. Segovia, Lipschitz functions on spaces of homogeneous type, Adv. Math. 33 (1979), 257-270. [11] T. Ma, P. R. Stinga, J. L. Torrea, and C. Zhang, Regularity properties of Schrodinger operators, J. Math. Anal. Appl., 388, (2012), 817 -- 837. 26 DE LE ´ON-CONTRERAS AND TORREA [12] L. Silvestre, Regularity of the obstacle problem for a fractional power of the Laplace operator, PhD Thesis, The University of Texas at Austin, 2005. [13] E. M. Stein. Singular integrals and differentiability properties of functions. Princeton Mathematical Series, No. 30. Princeton University Press, Princeton, N.J., 1970. [14] K. Stempak and J. L.Torrea, Poisson integrals and Riesz transforms for Hermite function expansions with weights, J. Funct. Anal., 202, (2003), 443 -- 472. [15] P. R. Stinga and J.L.Torrea. Extension problem and Harnack's inequality for some fractional operators, Comm. Partial Differential Equations, 35, (2010), 2092 -- 2122. [16] P. R. Stinga and J.L. Torrea, Regularity theory for the fractional harmonic oscillator, J. Funct. Anal., 260 (2011), 3097-3131. [17] P. R. Stinga and J. L. Torrea, Regularity theory and extension problem for fractional nonlocal parabolic equations and the master equation, To appear in SIAM Journal of Mathematical Analysis. [18] M. Taibleson, On the theory of Lipschitz spaces of distributions on Euclidean n-space, I,II,III J.Math. Mech. 13 (1964), 407 -- 480; 1(1965), 821 -- 840; 15 (1966), 973 -- 981. [19] S. Thangavelu, Riesz transforms and the wave equation for the Hermite operator, Comm. Partial Diff. Eq., 15 (1990), 1199-1215. [20] Sundaram Thangavelu. Lectures on Hermite and Laguerre expansions, volume 42 of Mathematical Notes. Princeton University Press, Princeton, NJ, 1993. With a preface by Robert S. Strichartz. [21] A. Zygmund. Trigonometric series. 2nd ed. Vols. I, II. Cambridge University Press, New York, 1959. Departamento de Matem´aticas, Facultad de Ciencias, Universidad Aut´onoma de Madrid, 28049 Madrid, Spain. E-mail address: [email protected] Departamento de Matem´aticas, Facultad de Ciencias, Universidad Aut´onoma de Madrid, 28049 Madrid, Spain. E-mail address: [email protected]
1812.08979	2	1812	2018-12-25T17:14:10	Composition operators on the spaces of Harmonic Bloch functions	[ "math.FA" ]	In this paper we characterize some basic properties of composition operators on the spaces of harmonic Bloch functions. First we provide some equivalent conditions for boundedness and compactness of composition operators. Then by using these conditions we estimate the essential norm of composition operators. These results extends the similar results that were proven in the literature on the Bloch spaces.	math.FA	math	,, COMPOSITION OPERATORS ON THE SPACES OF HARMONIC BLOCH FUNCTIONS S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN Abstract. In this paper we characterize some basic properties of composition operators on the spaces of harmonic Bloch functions. First we provide some equivalent conditions for boundedness and compactness of composition operators. In the sequel we investi- gate closed range composition operators. These results extends the similar results that were proven for composition operators on the Bloch spaces. 1. Introduction Let D be the open unit disk in the complex plane. For a continuously differentiable complex-valued f (z) = u(z) + iυ(z), z = x + iy, we use the common notation for its formal derivatives: fz = f¯z = 1 2 1 2 (fx − ify), (fx + ify). A twice continuously differentiable complex-valued function f = u + iυ on D is called a harmonic function if and only if the real-valued function u and υ satisfy Laplace's equation ∆u = ∆υ = 0. A direct calculation shows that the Laplacian of f is ∆f = 4fz ¯z. Thus for functions f with continuous second partial derivatives, it is clear that f is harmonic if ana only if ∆f = 0. We consider complex- valued harmonic function f defined in a simply connected domain D ⊂ C. The function f has a canonical decomposition f = h+¯g, where h and g are analytic in D [7]. A planar complex-valued harmonic function f in D is called a harmonic Bloch function if and only if 8 1 0 2 c e D 5 2 ] . A F h t a m [ 2 v 9 7 9 8 0 . 2 1 8 1 : v i X r a βf = sup (z, w) Here βf is the Lipschitz number of f and z,w∈D,z6=w f (z) − f (w) < ∞. (z, w) = arctan h z − w 1 − ¯zw , 2010 Mathematics Subject Classification. 47B33. Key words and phrases. Composition operator, Bloch spaces, Harmonic function. 1 2 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN denotes the hyperbolic distance between z and w in D, where here ρ(z, w) is the pseudo-hyperbolic distance on D. In [3] Colonna proved that βf = sup z∈D (1 − z2)[fz(z) + f¯z(z)]. Moreover, the set of all harmonic Bloch mappings, denoted by the sym- bol HB(1) or HB, forms a complex Banach space with the norm k.k given by kf kHB(1) = f (0) + sup z∈D (1 − z2)[fz(z) + f¯z(z)]. Definition 1.1. For α ∈ (0, ∞), the harmonic α-Bloch space HB(α) consists of complex-valued harmonic function f defined on D such that f HB(α) = sup z∈D (1 − z2)α[fz(z) + f¯z(z)] < ∞, and the harmonic little α-Bloch space HB0(α) consists of all function in HB(α) such that (1 − z2)α[fz(z) + f¯z(z)] = 0. lim z→1 Obviously, when α = 1, we have f HB(α) = βf . Each HB(α) is a Banach space with the norm given by kf kHB(α) = f (0) + sup z∈D (1 − z2)α[fz(z) + f¯z(z)], and HB0(α) is a closed subspace of HB(α). Now we define composition operators. Definition 1.2. Let D be the open unit disk in the complex plane. Let ϕ be an analytic self-map of D, i. e., an analytic function ϕ in D such that ϕ(D) ⊂ D. The composition operator Cϕ induced by such ϕ is the linear map on the spaces of all harmonic functions on the unit disk defined by Cϕf = f oϕ. The composition operators on function spaces were studied by many authors. Some known results about composition operators can be found in [6] and [11]. In this paper we study composition operators on har- monic Bloch-type spaces HB(α). In section 2, by using of Theorem 2.1 in [9], we give a necessary and sufficient condition for the bound- edness of Cϕ on HB(α) for α ∈ (0, ∞), which extends Theorem 3.1 in [9], by Lou. The compactness of Cϕ on analytic Bloch-type spaces were characterized in[10, 9]. In this paper, we deal the compactness of composition operators between the Banach spaces of harmonic function HB(α) and HB0(α). Moreover, we investigate closed range composition operators. Closed range composition operators on the Bloch-type spaces have been studied in [4, 2, 8, 14]). The isometric composition operators on Bloch-type spaces have been studied in a number of papers (such as [5, 3, 12, 13]). For α > 0, and ϕ being an analytic self-map of D, let 3 τϕ,α(z) = (1 − z2)αϕ′(z) (1 − ϕ(z)2)α . We write τϕ if α = 1. We say that a subset G ⊂ D is called sampling set for HB(α) if ∃S > 0 such that for all f ∈ HB(α), (1 − z2)α[fz(z)] + [f¯z(z)] ≥ Sf HB(α). sup z∈G To state the results obtained, we need the following definition. Let ρ(z, w) = ϕz(w) denote the pseudohyperbolic distance (between z and w) on D, where ϕz is a disk automorphism of D that is ϕz(w) = z − w 1 − ¯zw . We say that subset G ⊂ D is an r-net for D for some r ∈ (0, 1) if for each z ∈ D, ∃w ∈ G such that ρ(z, w) < r. For c > 0, let Ωc,α = {z ∈ D : τϕ,α(z) ≥ c}, and let Gc,α = ϕ(Ωc,α). If α = 1, we write Ωc and Gc. Now we recall Montel's theorem for harmonic functions. Theorem 1.3. [1] If {un}∞ n=1 is a sequence of harmonic functions in the region Ω with supn,x∈K un(x) < ∞ for every compact set K ⊂ Ω, then there exists a subsequence, {unj }∞ j=1 converging uniformly on every compact set K ⊂ Ω. Also we recall a very useful theorem that we will use it a lot in this paper. Theorem 1.4. [9] Let 0 < α < ∞. Then there exist f, g ∈ HB(α) such that for all z ∈ D. f ′(z) + g′(z) ≥ 1 (1 − z)α , 2. Main results In this section we study bounded and compact composition operators on HB(α). And then we nvestigate closed range composition operators on HB(α). First we provide some equivalent conditions for bounded- ness of composition operator Cϕ on HB(α). Theorem 2.1. If 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D, then the following statements are equivalent: 4 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN a) Cϕ : HB(α) → HB(α) is bounded. b) (1 − z2)α (1 − ϕ(z)2)α ϕ ′ sup z∈D (z) < ∞. Proof. For the implication a → b, by Theorem 2.1 of [9] we have that for 0 < α < ∞ there exist h, g ∈ B(α) satisfying the inequality ′ h (z) + g ′ (z) ≥ 1 (1 − z)α . If we set f = h + ¯g ∈ HB(α), then f oϕ(z) = hoϕ(z) + goϕ(z) and so by the same method of Theorem 3.1 of [9] we get the proof. For the implication b → a we can do the same as Theorem 3.1 of [9]. (cid:3) In the next theorem we consider the composition operator from HB0(α) into HB(α) and we find some conditions under which Cϕ is bounded. Theorem 2.2. Let 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D. Then the followings are equivalent: a) Cϕ : HB0(α) → HB(α) is bounded. b) (1 − z2)α (1 − ϕ(z)2)α ϕ ′ sup z∈D (z) < ∞. Proof. The proof is similar to the proof of Theorem 3.3 of [9]. Hence we omit the proof. (cid:3) Now we consider the composition operator Cϕ : HB(α) → HB0(α) and we give an equivalent condition to boundedness of Cϕ. Theorem 2.3. If 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D, then the following are equivalent: a) Cϕ : HB(α) → HB0(α) is bounded. b) lim z→1 (1 − z2)α (1 − ϕ(z)2)α ϕ ′ (z) = 0. Proof. By a similar method of the proof of Theorem 3.4 of [9] we get the proof. (cid:3) Finally we provide some conditions for boundedness of the composi- tion operator Cϕ as an operator on HB0(α). Theorem 2.4. If 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D, then the following are equivalent: a) Cϕ : HB0(α) → HB0(α) is bounded. b) ϕ ∈ B0(α) and (1 − z2)α (1 − ϕ(z)2)α ϕ ′ sup z∈D (z) < ∞. Proof. By some simple calculations one can get the proof. (cid:3) 5 A sequence {zn} in D is said to be R-separated if ρ(zn, zm) = zm−zn 1− ¯zmzn R whenever m 6= n. Thus an R-separated sequence consists of points which are uniformly far apart in the pseudohyperbolic metric on D, or equivalently, the hyperbolic balls D(zn, r) = {w : ρ(w, zn) < r} are pairwise disjoint for some r > 0. Evidently, any sequence {zn} in D which satisfies zn → 1 possesses an R-separated subsequence for any R > 0. > Another property of separated sequence is contained in the next propo- sition. Proposition 2.5. [10]. There is an absolute constant R > 0 such that if {zn} is R-separated, then for every bounded sequence {λn} there is an f ∈ B such that (1 − zn2)f ′(zn) = λn for all n. Since every sequence {zn} with zn → 1 contains an R-separated subsequence {znk}, it follows that there is an f ∈ B such that (1 − znk2)f ′(znk) = 1 for all k. Now we begin investigating compactness of the composition operator Cϕ in different cases. First we provide some equivalent conditions for compactness of Cϕ as an operator on HB(α). Theorem 2.6. Let 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D. Then we have the followings equivalent conditions: a) Cϕ : HB(α) → HB(α) is compact. b) and lim 1 − ϕ(z)2(cid:19)α ϕ(z)→1(cid:18) 1 − z2 1 − ϕ(z)2(cid:19)α z∈D(cid:18) 1 − z2 sup ′ ϕ (z) = 0, ′ (z) < ∞. ϕ Proof. By making use of the proof of Theorem 4.2 of [9] and the Propo- sition 1 of [10] we get the proof of Proposition 1 of [10] (cid:3) Here we prove that the compactness of Cϕ : HB0(α) → HB0(α) and Cϕ : HB(α) → HB0(α) are equivalent and we find an equivalent condition for compacness of Cϕ in these cases. Theorem 2.7. Let 0 < α < ∞, ϕ ∈ H(D) and ϕ(D) ⊆ D. Then the following statements are equivalent: a) The operator Cϕ : HB0(α) → HB0(α) is compact. 6 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN b) The operator Cϕ : HB(α) → HB0(α) is compact. c) lim z→1 (1 − z2)α (1 − ϕ(z)2)α ϕ ′ (z) = 0. Proof. First we prove the implication a → c. If Cϕ : HB0(α) → HB0(α) is compact, then the set K = Cϕ(SHB0(α)) ⊂ HB0(α) compact, in which SHB0(α) = {f ∈ HB0(α) : kf kHB0(α) ≤ 1}. By the Theorem 2.6 we get that sup (1 − z2)α[fz(z) + f¯z(z)] = 1 kf kHB(α)≤1 for all z ∈ D. Moreover we have 0 = lim z→1 sup (1 − z2)α[(f oϕ)z(z) + (f oϕ)¯z(z)] kf kHB(α)≤1 = lim z→1 (1 − z2)α (1 − ϕ(z)2)α ϕ ′ (z) sup (1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))]. kf kHB(α)≤1 So we get the desired result. Now we prove the implication c → b. Let {fn}n∈N ⊂ HB(α) and kfnkHB(α) ≤ 1, for all n. First we obtain that {Cϕfn} has a subse- quence that converges in HB0(α). By Montel's Theorem we have a subsequence {fnk} ⊂ {fn}, that converges uniformly on subsets of D to a harmonic function f . Hence we have (1 − z2)α[fz(z) + f¯z(z)] = lim k→∞ ≤ lim k→∞ (1 − z2)α[(fnk)z(z) + (fnk)¯z(z)] kfnkkHB(α) This means that f ∈ HB(α) with kf kHB(α) ≤ 1. Also we have ≤ 1. (1 − z2)α[(f oϕ)z(z) + (f oϕ)¯z(z)] = ≤ ′ (1 − z2)α (1 − ϕ(z))α ϕ (1 − z2)α (1 − ϕ(z)2)α ϕ (z) ′ (z)kf kHB(α). By these observations we conclude that Cϕf ∈ HB0(α). Also we need to show that lim k→∞ kCϕfnk − Cϕf kHB(α) = 0. Since limz→1 r ∈ (0, 1) such that for z with r < z < 1 we have (1−z2)α (1−ϕ(z)2)α ϕ′(z) = 0, then for any ε > 0, there exists (1 − z2)α (1 − ϕ(z)2)α ϕ ′ (z) < ε 4 . And so for all z with r < z < 1 we have (1 − z2)α((fnk − f ) ◦ ϕ)′(z) = (1 − z2)α{[(fnk)zϕ(z) + (fnk)¯zϕ(z)]} 7 − (1 − z2)α{[fzϕ(z) + f¯zϕ(z)]} ≤ ε 4 (kfnkkHB(α) + kf kHB(α)) ≤ ε 2 . For z with z ≤ r, the set {ϕ(z) : z ≤ r} is a compact subset of D. Since (1 − z2)α[fz(z) + f¯z(z)] = lim k→∞ (1 − z2)α[(fnk)z(z) + (fnk)¯z(z)] and (1 − z2)α((fnk − f ) ◦ ϕ)′(z) ≤ (1 − z2)α{[(fnk)zϕ(z) + (fnk)¯zϕ(z)] (1 − z2)α (1 − ϕ(z)2)α ϕ − [fzϕ(z) + f¯zϕ(z)] × sup z∈D ′ (z). Hence we have (1 − z2)α((fnk − f )oϕ)′(z) → 0 uniformly on {z : z ≤ r}. Therefore (1 − z2)α((fnk − f )oϕ)′(z) < ε 2 for k sufficiently large and {z : z ≤ r}. This completes the proof. The implication b → a is clear. (cid:3) Let (X, d) be a metric space and let ε > 0. We say that A ⊂ X is an ε-net for (X, d), if for all x ∈ X there exists a a in A such that d(a, x) < ε. We characterize the compact subsets of HB0(α) in the next lemma. Lemma 2.8. A closed subset of HB0(α) is compact if and only if it is bounded and satisfies lim z→1 sup f ∈k (1 − z2)α[fz(z) + f¯z(z)] = 0. Proof. suppose that K ⊂ HB0(α) is compact and ε > 0. Then we can choose an ε 2-net f1, f2, ..., fn ∈ K. hence there exists δ, 0 < δ < 1, such that for all z with z > δ we have (1 − z2)α[(fi)z(z) + (fi)¯z(z)] < ε 2 for all 1 ≤ i ≤ n. If f ∈ K, then there exists some fi such that kf − fikHB(α) < ε (1−z2)α[fz(z)+f¯z(z)] ≤ kf −fikHB(α)+(1−z2)α[(fi)z(z)+(fi)¯z(z)] < ε. Therefor we get that 2 and so for all z with z > δ we have lim z→1 sup f ∈k (1 − z2)α[fz(z) + f¯z(z)] = 0. Conversely, let K be a closed and bounded subset of HB0(α) such that lim z→1 sup f ∈k (1 − z2)α[fz(z) + f¯z(z)] = 0. Since K is bounded, then it is relatively compact with respect to the topology of the uniform convergence on compact subsets of the unit disk. If (fn) is a sequence in K, then by Montel's Theorem we have 8 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN a subsequence {fnk} ⊂ {fn} which converges uniformly on compact subsets of D to a harmonic function f . Also {f ′ nk} converges uniformly to f ′ on compact subsets of D. For every ε > 0 we can find δ > 0 such that for all z with z > δ we have (1 − z2)α[(fnk)z(z) + (fnk)¯z(z)] < ε 2 for any integer k > 0. Therefor (1 − z2)α[fz(z) + f¯z(z)] < ε z with z > δ. So 2, for all sup z>δ (1 − z2)α[(fnk − f )z(z) + (fnk − f )¯z(z)] ≤ sup z>δ (1 − z2)α[(fnk)z(z) + (fnk)¯z(z)] (1 − z2)α[fz(z) + f¯z(z)] + sup z>δ < ε. Moreover, since (fnk) converges uniformly on compact subsets of D to f and (f ′ nk) converges uniformly to f ′ on {z : z ≤ δ}, we get that (1 − z2)α[(fnk − f )z(z) + fnk − f )¯z(z)] ≤ ε. sup z≤δ Consequently for k large enough, we have limk→∞ kfnk − f kHB(α) ≤ ε. This completes the proof. (cid:3) In the next theorem we prove that the norm convergence in HB(α) implies the uniform convergence. Theorem 2.9. The norm convergence in HB(α) implies the uniform convergence, that is if {fn} ⊂ HB(α) such that kfn − f kHB(α) → 0, then {fn} converges uniformly to f . Proof. For 0 6= z ∈ D, we have 0 dt fn(z) − f (z) = Z 1 = zZ 1 ≤ zZ 1 0 0 d(fn − f ) (zt)dt d(fn − f ) dς(t) (zt)dt + ¯zZ 1 0 d(fn − f ) d¯ς(t) (zt)dt [(fn − f )ς(t)(zt) + (fn − f ) ¯ς(t)(zt)]dt, in which ς(t) = zt. This gives us fn(z) − f (z) ≤ Z 1 0 [(fn − f )ς(t)(zt) + (fn − f ) ¯ς(t)(zt)] (1 − ς(t)2)α (1 − ς(t)2)αdt ≤ (kfn − f kHB(α))Z 1 0 1 (1 − zt)α dt → 0, when n → ∞. So we get the proof. (cid:3) We say that a subset G ⊂ D is called sampling set for HB(α) if ∃S > 0 such that for all f ∈ HB(α), 9 (1 − z2)α[fz(z)] + [f¯z(z)] ≥ Skf kHB(α). sup z∈G In the next theorem we provide some equivalent conditions for closed- ness of range of the composition operator on HB(α). Theorem 2.10. Let ϕ : D → D, α > 0 and Cϕ : HB(α) → HB(α) be a bounded operator. Then the range of Cϕ : HB(α) → HB(α) is closed if and only if there exists c > 0 such that Gc,α is sampling for HB(α). Proof. Since Cϕ : HB(α) → HB(α) is bounded, then ∃K > 0 such that supz∈D τϕ,α(z) ≤ K. Since every non-constant ϕ is an open map, then the composition operator Cϕ is always one to one. By a basic operator theory result, a one-to-one operator has closed range if and only if it is bounded below. hence if Cϕ has closed range, then Cϕ is bounded below, that is ∃ε > 0 such that for all f ∈ HB(α), kCϕf kHB(α) = sup z∈D (1 − z2)α[(f oϕ)z(z) + (f oϕ)¯z(z)] = sup z∈D τϕ,α(z)(1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))] ≥ εkf kHB(α). Now we show that the set Gc,α is sampling for HB(α) with sampling constant S = ε K . Since Ωc,α = {z ∈ D : τϕ,α(z) ≥ c}, so for any z /∈ Ωc,α and c = ε 2, we have τϕ,α(z)(1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))] ≤ sup z /∈Ωc,α ε 2 kf kHB(α). Therefore we have εkf kHB(α) ≤ sup z∈D τϕ,α(z)(1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))] = sup z∈Ωc,α τϕ,α(z)(1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))] ≤ K sup w∈Gc,α (1 − w2)α[h ′ (w) + g ′ (w)]. Hence supw∈Gc,α(1 − w2)α[h′(w) + g ′(w)] ≥ ε K kf kHB(α). this means that Gc,α is a sampling set for HB(α) with sampling constant S = ε K . Conversely, suppose that Gc,α is a sampling set for HB(α), with sam- pling constant S > 0. So for all f ∈ HB(α) and ε = cS we get the 10 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN followings relations: Skf kHB(α) ≤ sup z∈Ωc,α (1 − ϕ(z)2)α[(f )z(ϕ(z)) + (f )¯z(ϕ(z))] (1 − ϕ(z)2)α[h ′ (ϕ(z)) + g ′ (ϕ(z))] (1 − z2)α[(hoϕ)z(z) + (goϕ)¯z(z)] ≤ = sup z∈Ωc,α 1 c 1 c sup z∈D ≤ kf oϕkHB(α). Therefore εkf kHB(α) ≤ kf oϕkHB(α) = kCϕf kHB(α). Hence Cϕ is bounded below and so Cϕ has closed range. (cid:3) Now we give some other necessary and sufficient conditions for closed- ness of range of Cϕ : HB(α) → HB(α). Theorem 2.11. Let ϕ be a self-map of D, α > 0, and Cϕ : HB(α) → HB(α) be a bounded operator. Then we have the followings: a) If the operator Cϕ : HB(α) → HB(α) has closed range, then there exist c, r > 0 with r < 1, such that Gc,α is an r-net for D. b) If there exist c, r > 0 with r < 1, such that Gc,α contains an open annulus centered at the origin and with outer radius 1, then Cϕ has closed range. Proof. a) For a ∈ D, let ϕa(z) be a function such that ϕa(0) = 0 and ϕ′ a(z) = (ψ ′ a(z))α, where ψa is the disc automorphism of D defined by ψa(z) = a−z 1−¯az . Using the equalities 1 − ρ(z, w)2 = 1 − ψw(z)2 = (1 − z2)ψ ′ w(z) we get kϕa + ¯ϕakHB(α) = sup z∈D (1 − z2)α2ϕ ′ a(z) = 2 sup z∈D (1 − ψa(z)2)α = 2. If we put f = ϕa + ¯ϕa, then we have kCϕf kHB(α) = kf oϕkHB(α) = sup z∈D = sup z∈D (1 − z2)α[(f oϕ)z(z) + (f oϕ)¯z(z)] τϕ,α(z)2(1 − ψa(ϕ(z))2)α. Moreover, by assuming that Cϕ is bounded and has closed range, then there exist K, ε > 0 such that supz∈D τϕ,α(z) = K and kf oϕkHB(α) = sup z∈D τϕ,α(z)2(1 − ψa(ϕ(z))2)α ≥ εkϕa + ¯ϕakHB(α). 11 This implies that ε ≤ sup z∈D ≤ sup z∈D τϕ,α(z)(1 − ψa(ϕ(z))2)α τϕ,α(z) = K. Since 1 − ψa(ϕ(z))2 ≤ 1, then there exists za ∈ D such that and τϕ,α(za) ≥ ε 2 (1 − ψa(ϕ(za))2)α ≥ ε 2K . 1 2 and r = q1 − ( ε 2K ) Thus, for c = ε α , we conclude that for all a ∈ D, there exists za ∈ Ωc,α such that ρ(a, ϕ(za)) < r and so Gc,α is an r-net for D. b) Let Gc,α contains the annulus A = {z : r0 < z < 1} and Cϕ : HB(α) → HB(α) be bounded. Suppose that Cϕ doesn't have closed range, then there exists a sequence {fn} with kfnkHB(α) = 1 and kCϕfnkHB(α) → 0. For each ε > 0, let Nε > 0 such that for all n > Nε we have Since kCϕfnkHB(α) < ε < cε. (1−z2)α[(fn)z(z)+(fn)¯z(z)] = sup z∈D sup z∈D then there exists a sequence {an} in D such that for all n, (1−z2)α[h n(z)+g ′ ′ n(z)] = 1, (1 − an2)α[h ′ n(an) + g ′ n(an)] ≥ 1 2 . Moreover, we have (1 − w2)α[(fn)z(w) + (fn)¯z(w)] sup w∈Gc,α τ −1 ϕ,α(z)τϕ,α(z)(1 − ϕ(z)2)α[(fn)z(ϕ(z)) + (fn)¯z(ϕ(z))] (1 − z2)αϕ ′ (z)[(fn)z(ϕ(z)) + (fn)¯z(ϕ(z))] ≤ = sup z∈Ωc,α 1 c cε c sup z∈D < = ε. If we take ε < 1 2, then we get that each an with n > Nε belongs to (Gc,α)c. Thus an ≤ r0 < 1 and an → a with a ≤ r0. On the other hand, by Montel's Theorem, there exists a subsequence {fnk} such that converges uniformly on compact subsets of D to some function f ∈ HB(α). Hence {f ′ nk} converges to f ′ uniformly on compact subsets of D, and since (1 − w2)α[(fn)z(w) + (fn)¯z(w)] → 0, sup w∈Gc,α 12 S. ESMAEILI, Y. ESTAREMI AND A. EBADIAN when n → ∞ and Gc,α contains a compact subset of D, we conclude that f ′ = 0. This contradicts the fact that (1 − a2)α[h ′ (a) + g ′ (a)] ≥ 1 2 . Therefore Cϕ must be bounded below and consequently it has closed range. (cid:3) References [1] S. Axler, P. Bourdon and W. Ramey, Harmonic function theory, Graduate Texts in Mathematics 137, springer, New York, (1992). [2] H. Chen and P. Gauthier, Boundedness from below of composition operator on α-Bloch spaces, Canadian Mathematical Bulletin, 51 (2008), 195-204. [3] F. Colonna, The Bloch constant of bounded harmonic mappings, Indiana Univ. Math. J., 38, (1989), 829-840. [4] M. Contreras and A. Hernandez-Diaz, Weithed composition operator in Wei- thed Banach spaces of analytic functions, J. Australian Math. Soc., 69, (2000), 41-60. [5] J. M. Cohen and F. Colonna, Preimagrs of one-pointsets of Bloch andnormal functions, Mediterr. J. Math., 3, (2006), 513-532. [6] C. Cowen and B. MacCluer, composition operators on spaces of analytic func- tions, CRC Press, Boca Raton, (1995). [7] P. Duren, Harmonic Mapping in the Plane, Cambridge Univ. Press, (2004). [8] P.Ghatage, D, Zheng, and N. Zorboska, Sampling set and Closed-Range Com- position operators on the Bloch space, Proc. Amer. Math. Soc., 133 (2005), 1371-1377. [9] Z. Lou, Composition operator on Bloch type spaces, Analysis 23(2003), No.1, 81-95. [10] K. Madigan, A. Matheson, Compact composition operator on the Bloch spaces, Trans. Amer. Math. Soc. 347 (1995), 2679-2687. [11] J. H. Shapiro, Composition operators and Classical Function Theory, Springer- Verlag, New York, (1993). [12] c. Xiong, Norm of composition operators on the Bloch space, Bull. Austral. Math. Soc., 70, (2004), 293-299. [13] N. Zorboska, Isometric composition operators on the Bloch-type spaces,C. R. Math. Rep. Acad. Sci. Canada Vol. 3 2007, 9196. [14] N. Zorboska, Isometric and Closed-Range Composition operators between Bloch-type spaces, Int. J. Math. Sci., to appear. S. Esmaeili, Y. estaremi and A. Ebadian E-mail address: [email protected] E-mail address: [email protected] E-mail address: [email protected]. Department of mathematics, Payame Noor university , P. O. Box: 19395-3697, Tehran, Iran.,
1604.07393	1	1604	2016-04-25T07:03:19	Analytic functional calculus for two operators	[ "math.FA", "math.OA", "math.SP" ]	Properties of the mappings \begin{align} C&\mapsto\frac1{(2\pi i)^2}\int_{\Gamma_1}\int_{\Gamma_2}f(\lambda,\mu)\,R_{1,\,\lambda}\,C\, R_{2,\,\mu}\,d\mu\,d\lambda, C&\mapsto\frac1{2\pi i}\int_{\Gamma}g(\lambda)R_{1,\,\lambda}\,C\, R_{2,\,\lambda}\,d\lambda \end{align} are discussed; here $R_{1,\,(\cdot)}$ and $R_{2,\,(\cdot)}$ are pseudo-resolvents, i.~e., resolvents of bounded, unbounded, or multivalued linear operators, and $f$ and $g$ are analytic functions. Several applications are considered: a representation of the impulse response of a second order linear differential equation with operator coefficients, a representation of the solution of the Sylvester equation, and an exploration of properties of the differential of the ordinary functional calculus.	math.FA	math	ANALYTIC FUNCTIONAL CALCULUS FOR TWO OPERATORS V.G. KURBATOV, I.V. KURBATOVA, AND M.N. ORESHINA Abstract. Properties of the mappings 1 (2πi)2 ZΓ1ZΓ2 2πiZΓ g(λ)R1, λ C R2, λ dλ f (λ, µ) R1, λ C R2, µ dµ dλ, C 7→ C 7→ 1 are discussed; here R1, (·) and R2, (·) are pseudo-resolvents, i. e., resolvents of bounded, unbounded, or multivalued linear operators, and f and g are analytic functions. Several applications are considered: a representation of the impulse response of a second order linear differential equation with operator coefficients, a representation of the solution of the Sylvester equation, and an exploration of properties of the differential of the ordinary functional calculus. 1. Introduction Let A and B be matrices of the sizes n× n and m× m respectively. The result p(A, B) of the substitution of these matrices into a polynomial p(λ, µ) = PN i,j=0 cijλiµj of two variables is usually understood to be the mapping C 7→PN i,j=0 cijAiCBj acting on n× m- matrices, or to be the block matrix {aijB}. This article is devoted to extensions and applications of this construction. First, the matrices A and B can be replaced by bounded linear operators acting in (infinite-dimensional) Banach spaces X and Y respectively. In this case, the number of interpretations of the object p(A, B) increases. The most natural abstract interpretation is considering p(A, B) as an operator acting in the completion of the algebraic tensor product [28, 64, 112] X⊗Y with respect to some cross-norm. This interpretation embraces many spaces of functions of two variables. For example [28, 64, 112], L1[a, b]⊗πL1[c, d] is isometrically isomorphic to L1[a, b]×[c, d], and C[a, b]⊗εC[c, d] is isometrically isomorphic to C[a, b] × [c, d]. But unfortunately, L∞[a, b]⊗εL∞[c, d] is isomorphic only to a subspace of the space L∞[a, b] × [c, d]. Another example, which can not be treated directly in terms of tensor products, is the interpretation of p(A, B) as the transformation C 7→ PN i,j=0 cijAiCBj of the operators C : Y → X. From the point of view of applications, the last example seems to be the most important. Therefore, in all cases, we call the operator that corresponds to p(A, B) a transformator ; this term is conventional [51] for the mappings of the type C 7→ PN In order to embrace the last example and some others, our treatment is based on the notion of an extended tensor product (Section 5) proposed in [89]. i,j=0 cijAiCBj acting on operators C. Second, one may replace the polynomial p by an analytic function f . In this case, it is convenient to define f (A, B) by means of a contour integral. For our main interpretation of f (A, B) as a transformator acting on operators C : Y → X, the relevant formula looks 1991 Mathematics Subject Classification. 47A60, 47A80, 47B49, 15A22, 39B42, 34A30. Key words and phrases. Extended tensor products, pseudo-resolvent, Sylvester's equation, differential of the functional calculus, impulse response. 1 ϕ1(f ) = 1 2πiZΓ1 1 2πiZΓ2 as follows: f (A, B)C = 1 (2πi)2 ZΓ1ZΓ2 f (λ, µ)(λ1 − A)−1C(µ1 − B)−1 dµ dλ. () We call a correspondence of the type f 7→ f (A, B) that maps functions f to transformators (operators) f (A, B) a functional calculus. From the algebraic point of view, the functional calculus f 7→ f (A, B) possesses prop- erties of the tensor product ϕ1 ⊗ ϕ2 of two ordinary functional calculi f (λ)(λ1 − A)−1 dλ, ϕ2(f ) = f (µ)(µ1 − B)−1 dµ. To emphasise this fact, we use the notation ϕ1 ⊠ ϕ2 and write (cid:2)(ϕ1 ⊠ ϕ2)f(cid:3)C instead of f (A, B)C when the transformator f (A, B) acts in an extended tensor product. An important and nontrivial property of the transformation ϕ1 ⊠ ϕ2 is the spectral mapping theorem (Theorem 39). Third, the basic properties of the functional calculus ϕ1 ⊠ ϕ2 are preserved when one replaces the resolvents Rλ = (λ1 − A)−1 of operators by pseudo-resolvents [67], i. e. operator-valued functions R(·) that satisfy the Hilbert identity Rλ − Rµ = −(λ − µ)RλRµ. This generalization enables one to cover some additional examples. For example, a special case of a pseudo-resolvent is [67] the resolvent of an unbounded operator, and the most general example of a pseudo-resolvent is the resolvent of a linear relation or, in other terminology, a multivalued linear operator [8, 9, 16, 20, 38, 57, 92]. Moreover, in this article, we adhere to the point of view that a pseudo-resolvent is as a fundamental object as an operator (bounded, unbounded or multivalued) that generates it. The reason for this stems from the fact that when speaking of unbounded operators and linear relations we often actually work with their resolvents. For example, an unbounded operator is a generator of a strongly continuous or analytic semigroup if and only if [36, 67, 127] its resolvent satisfies a special estimate of the decay rate at infinity; in [75, Theorem 2.25] and [106, VIII.7], the natural convergence of unbounded operators is defined as the convergence of their resolvents in norm; and in [105, 107], a function f of unbounded operators A and B is defined as an (unbounded) operator f (A, B) that possesses the following property: there exist sequences of bounded operators An and Bn such that the resolvents of An, Bn, and f (An, Bn) converge in norm to the resolvents of A, B, and f (A, B) respectively. Another argument (not used in this article) is that there is no analogue of unbounded and multivalued operators in Banach algebras, but, nevertheless, there are evident analogues of the resolvents of such operators. This approach enables one to extend the notion of f (A, B) to meromorphic functions f (Theorem 42): we define the result of the action of a meromorphic function on A and B to be a new pseudo-resolvent and do not discuss which operator it is generated by. Along the way, we answer (Corollary 43) the question of the independence of the definition of f (A, B) for unbounded A and B posed in [105, 107] (see the previous paragraph) from the choice of approximating sequences An and Bn. Many important applications are connected with the special cases of construction () and their modifications. For example, it often occurs that the function f depends on the difference or the sum of its arguments: the transformator C 7→ AC − CB generated by the function f (λ, µ) = λ − µ is related to the Sylvester equation (Section 10), and the transformator C 7→ eAtCeBt generated by the function f (λ, µ) = et(λ+µ) is connected with the stability theory of differential equations [5, 23]. 2 The version (Section 8) f [1](A, B)C = 1 2πiZΓ f (λ)(λ1 − A)−1C(λ1 − B)−1 dλ (*) of functional calculus () frequently occurs in applications; it involves functions f of one variable. For example, expression (*) with f (λ) = eλt forms the principal part of the representation of a solution of the second order differential equation(cid:0) d 0 (Section 9). Further (Section 11), the differential of the ordinary functional calculus A 7→ f (A) at a point A can be represented in the form (Theorem 67) dt −A(cid:1)C(cid:0) d dt −B(cid:1)y = It turns out that (Theorems 45) f [1](A, B)C coincides with () provided f [1] is understood to be the divided difference df (·, A) = f [1](A, A). f [1](λ, µ) =  , f (λ)−f (µ) λ−µ f′(λ), 0, if λ 6= µ, if λ = µ, if λ = ∞ or µ = ∞. When choosing the level of generality of our exposition, we proceed from the following principles. First, in order that a specialist in the classical operator theory can use the article we are trying to minimize explicit mention of Banach algebras and linear relations (multivalued operators) at least in main statements. At the same time, when using the operator language, we aspire the maximal generality and, in particular, where possible, consider the case of an arbitrary pseudo-resolvent (and thereby, implicitly, the cases of unbounded operators and linear relations). Second, we are trying not to fall outside the framework of the theory of analytic functions of an operator and thus, for example, do not discuss issues related to the generators of semigroups. Third, as far as possible, we avoid the implicit use of operator pencils λ 7→ λF − G with F 6= 1 because this approach leads to very cumbersome formulae. Finally, we restrict ourselves to the consideration of functions of two variables, suggesting that the generalization to three and more variables will not cause significant difficulties. The literature on the subject under discussion is extremely extensive. Therefore, the bibliography can not be made comprehensive; the presented references reflect authors' tastes and interests. Many additional references can be found in the cited articles and books. Sections 2 -- 5 outlines preliminary information. Here we recall and fix notation and the main facts in a convenient form. In Section 2, the terminology connected with Banach algebras and their properties is recalled. In Section 3, the basic properties of algebras of analytic functions of one and two variables are described. In Section 4, we discuss the notion of pseudo-resolvent and recall the construction of the functional calculus of analytic functions of one variable (Theorems 25 and 26) including the spectral mapping theorem (Theorem 27). In Section 5, the definition of the extended tensor product is given, the main examples are described, and the construction of the functional calculus of operator-valued analytic functions of one variable (Theorems 28 and 29) is recalled as well as the relevant spectral mapping theorem (Theorem 31). In Section 6, we present the construction of the functional calculus () of functions of two variables (Theorems 32, 33, and 34) and prove the corresponding spectral mapping theorem (Theorem 39). In Section 7, we extend these results to meromorphic functions. A well-known example of a meromorphic function of an operator is a polynomial of an unbounded operator (a polynomial has a pole at infinity, and the point at infinity belongs 3 to the extended spectrum of an unbounded operator). This example shows that the result of applying a meromorphic function cannot be a bounded transformator; as a convenient tool for its description we use not the resulting object itself, but its resolvent, and we interpret the extended singular set of this resolvent as its spectrum (Theorem 42). In Section 8, we discuss modified variant () of the functional calculus of functions of two variables. The connection between functional calculi () and () is established as well as some properties of functional calculus (). The subsequent sections are devoted to applications. In Section 9, the pencil λ 7→ λ2E +λF +H of the second order is considered; it is induced by the equation E y(t)+F y(t)+Hy(t) = 0; we assume that the pencil admits a factorization, i. e. the representation in the form of a product of two linear pencils. In such a case, the solution of the differential equation is expressed by a transformation of the kind () (Theorem 54). In Section 10, we discuss the properties of the transformator Q : C 7→ Z generated by the Sylvester equation AZ − ZB = C (Theorem 62). Finally, in Section 11, it is shown that the differential of the ordinary functional calculus A 7→ f (A) is also a kind of transformator () (Theorem 67). 2. Banach algebras In this Section, we clarify the terminology connected with Banach algebras [18, 67, 110] and recall some of their properties. In this article, all linear spaces and algebras are assumed to be complex. Let X and Y be Banach spaces. We denote by B(X, Y ) the set of all bounded linear operators A : X → Y . When X = Y , we use the shorthand symbol B(X). The symbol 1 = 1X stands for the identity operator. We adhere to the following notations: X∗ denotes the conjugate space of X; hx, x∗i denotes the value of the functional x∗ ∈ X∗ on x ∈ X, and hx∗∗, x∗i is the value of x∗∗ ∈ X∗∗ on x∗ ∈ X∗; A∗ denotes the conjugate operator of A ∈ B(X, Y ). The preconjugate of an operator A ∈ B(Y ∗, X∗) is an operator A0 ∈ B(X, Y ) such that (A0)∗ = A. The unit [18, 67, 110] of an algebra B is an element 1 ∈ B such that 1A = A1 for all A ∈ B. If an algebra B has a unit, it is called an algebra with a unit or unital. A subset R of an algebra B is called a subalgebra if R is stable under the algebraic operations (addition, scalar multiplication, and multiplication), i. e. 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. Let B be a non-unital algebra. The set eB = C ⊕ B with the componentwise linear algebra with the unit (1, 0), where 0 is the zero of the algebra B. The algebra eB is called the algebra derived from B by adjoining a unit or an algebra with an adjoint unit. The symbol 1 stands for the element (1, 0), and the symbol α1 + A denotes the element (α, A). If B is a normed algebra, we set k(α, A)k = α + kAk. Clearly, this formula defines a norm on eB. It is also clear that eB is complete provided that B is complete. If B is unital, then eB means the algebra B itself. Theorem 1. Let B be a unital Banach algebra and A, B ∈ B. If A is invertible and operations and the multiplication (α, A)(β, B) = (αβ, αB + βA + AB) is obviously an kBk · kA−1k < 1, then the element A − B is also invertible and (A − B)−1 = A−1 + A−1BA−1 + A−1BA−1BA−1 + A−1BA−1BA−1BA−1 + . . . . 4 In this case kA−1k k(A − B)−1k ≤ 1 − kBk · kA−1k k(A − B)−1 − A−1k ≤ kBk · kA−1k2 1 − kBk · kA−1k k(A − B)−1 − A−1 − A−1BA−1k ≤ kBk2 · kA−1k3 1 − kBk · kA−1k Proof. We consider the series , , . (1) A−1 + A−1BA−1 + A−1BA−1BA−1 + A−1BA−1BA−1BA−1 + . . . . We represent this series in the form A−1(cid:0)1+BA−1+BA−1BA−1+BA−1BA−1BA−1+. . .(cid:1). Since kBA−1k ≤ kBk · kA−1k < 1, the series converges absolutely. We denote its sum by C. It is straightforward to verify that C coincides with the inverse of A − B. Estimates (1) follow from the geometric series formula P∞k=0 qk = 1 1−q , q < 1. For example, let us prove the second estimate: kBkk · kA−1kk+1 = kBk · kA−1k2 1 − kBk · kA−1k k(A − B)−1 − A−1k ≤ ∞Xk=1 . (cid:3) 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). The complement ρ(A) = C \ σ(A) is called the resolvent set of A. The function Rλ = (λ1− A)−1 is called the resolvent of the element A. Proposition 2 ([67, Theorem 4.8.1]). The resolvent R(·) of any element A ∈ B satisfies the Hilbert identity Rλ − Rµ = −(λ − µ)RλRµ, (2) Proposition 3 ([18, ch.1, § 2, . 5], [110, Theorem 10.13]). The spectrum of any element A of a nonzero unital Banach algebra B is a compact and nonempty subset of the complex plane C. λ, µ ∈ ρ(A). Let A and B be algebras. A mapping ϕ : A → B is called [18] a morphism of algebras if ϕ(A + B) = ϕ(A) + ϕ(B), ϕ(αA) = αϕ(A), ϕ(AB) = ϕ(A)ϕ(B). If, in addition, A and B are unital and ϕ(1A) = 1B, ϕ is called a morphism of unital algebras. If A and B are Banach algebras [18, 67, 110] and, in addition, the morphism ϕ is continuous, then ϕ is called a morphism of Banach algebras. A unital subalgebra R of a unital algebra B is called [18, ch. 1, § 1.4] full if every B ∈ R which is invertible in B is also invertible in R. Since the inverse is unique, the last definition is equivalent to the following: if for B ∈ R there exists B−1 ∈ B such that BB−1 = B−1B = 1, then B−1 ∈ R. 5 Example 1. Let X be a Banach space. The set B0(X∗) of all operators that have a preconjugate is a full subalgebra of the algebra B(X∗). Proposition 4 ([18, ch. 1, § 2.5]). The closure of a full subalgebra of a Banach algebra is also a full subalgebra. The closure of the least full subalgebra of a Banach algebra that contains a set M is the least full closed subalgebra that contains M. An algebra B is called commutative if AB = BA for all A, B ∈ B. A character of a unital commutative algebra B [18, ch. 1, § 1.5] is a morphism χ : B → C of unital algebras. A character of a commutative non-unital algebra B [18, ch. 1, § 1.5] is a morphism of (non-unital) algebras χ : B → C. If an algebra B is non-unital, we denote by χ0 the character χ0 : B → C that is equal to zero on all elements of B. We call χ0 the zero character. We stress that the character χ0 exists only if the algebra B is non-unital. Proposition 5. All characters of a non-unital algebra B are extendable uniquely to char- the formula χ(α1 + A) = α + χ(A). Conversely, the restriction of any character of the acters of the algebra eB derived from B by adjoining a unit. The extension is defined by algebra eB to B is a character of the algebra B. In particular, the zero character χ0 is the restriction of the character α1 + A 7→ α; we will denote it by the same symbol χ0. Proof. The proof is obvious. (cid:3) We denote by X(B) the set of all nonzero characters of a commutative algebra B (unital B (including the zero character χ0 if the algebra is non-unital). If an algebra B is unital, or non-unital), and we denote by eX(B) the set of all characters of a commutative algebra then eX(B) obviously coincides with X(B). The set X(B) is called [18] the character space of the algebra B. Theorem 6 ([18, ch. 1, § 3.3, Proposition 3]). Let B be a unital commutative Banach algebra. Then for all A ∈ B, Corollary 7 ([18, ch. 1, § 3, Theorem 1]). Every character of a commutative Banach algebra is continuous; namely, its norm is less than or equal to unity. σ(A) = { χ(A) : χ ∈ eX(B) }. Corollary 8. In a unital commutative Banach algebra B, the spectrum continuously de- pends on an element; more precisely, if A, B ∈ B and kA−Bk < ε, then σ(B) is contained in the ε-neighbourhood of σ(A). 3. Algebras of analytic functions This Section is a preparation for a discussion of analytic functional calculi. Here we 2 collect some preliminaries on algebras of analytic functions defined on subsets of C and C . We denote by C the one-point compactification C ∪ {∞} of the complex plane C, and 2 we denote by C the Cartesian product C × C. Proposition 9. Let σ1, σ2 ⊆ C be closed sets, and let an open set W ⊆ C σ1 × σ2. Then there exist open sets U, V ⊆ C such that σ1 × σ2 ⊆ U × V ⊆ W . Proof. For an arbitrary λ ∈ σ1, we consider the set {λ} × σ2. We consider a finite cover of {λ} × σ2 by the sets of the form Ui × Vi, where Ui, Vi ⊆ C are open and Ui × Vi ⊆ W . We put eU = ∩iUi and eV = ∪iVi. It is clear that the set eU × eV ⊆ W also covers the set {λ} × σ2, namely, {λ} ⊆ eU, σ2 ⊆ eV . contains 6 2 Further, we cover every subset of the form {λ} × σ2 of the set σ1 × σ2 by a set of the form eU ×eV ⊆ W . We choose a finite subcover {eUk ×eVk } and put U = ∪keUk, V = ∩keVk. Obviously, σ1 × σ2 ⊆ U × V ⊆ W . Proposition 10. Let U1, U2 ⊆ C be open sets. Then for every compact set N ⊂ U1 × U2 there exist compact sets N1 ⊆ U1 and N2 ⊆ U2 such that N ⊆ N1 × N2. Proof. It is sufficient to take for N1 the image of the set N under the projection (λ, µ) 7→ λ onto the first coordinate, and for N2 the image of the set N under the projection (λ, µ) 7→ µ onto the second coordinate. (cid:3) (cid:3) 2 Let K be a closed subset of C or C and B be a unital Banach algebra. We denote by O(K, B) the set of all analytic1 [54, 115] functions f : U → B, where U is an open neighbourhood of the set K (it is implied that the neighbourhood U may depend on f ). Two functions f1 : U1 → B and f2 : U2 → B are called equivalent if there exists an open neighbourhood U ⊆ U1 ∩ U2 of the set K such that f1 and f2 coincide on U, i. e. f1(λ) = f2(λ) for all λ ∈ U. It can be easily shown that this is really an equivalence relation. Thus, strictly speaking, elements of O(K, B) are classes of equivalent functions. The notation O(K, C) is abbreviated to O(K). Proposition 11. The set O(K, B) is an algebra with respect to pointwise operations with the unit u(λ) = 1, λ ∈ U ⊃ K. Proposition 12. (a) For f ∈ O(K, B), the following conditions are equivalent: the function f is invertible in the algebra O(K, B); the element f (λ) ∈ B is invertible at all points λ ∈ K; the element f (λ) is invertible at all points λ ∈ U, where U ⊃ K is some open set. (b) The spectrum of a function f ∈ O(K, B) in the algebra O(K, B) is given by the formula [λ∈K σ(cid:0)f (λ)(cid:1). We recall the definition of the natural topology on the algebra O(K, B) [18, ch. 1, § 4.1]. For each open set U ⊃ K, we denote by O(U, B) the linear space of all analytic functions f : U → B. We endow O(U, B) with the topology of compact convergence [17, ch. X, § 3.6], [111, ch. III, § 3]. A fundamental system of neighbourhoods of zero in this topology is formed by the sets T (N, δ) = { f : kf (N)k < δ }, where N ⊂ U is compact and δ > 0; clearly, when N enlarges, the neighbourhood T (N, δ) shrinks; therefore, it is enough to consider only those sets N, the interior of which contains K. There are evident canonical mappings gU : O(U, B) → O(K, B). The mappings gU are not always injective. Nevertheless, by misuse of language, we will regard O(U, B) as subspaces of the space O(K, B). We endow O(K, B) with the inductive topology [111, ch. II, § 6] induced by the mappings gU (one may restrict himself to a decreasing sequence of open sets U ⊃ K). A fundamental system of neighbourhoods of zero in O(K, B) consists of all balanced, absorbent, and convex sets W ⊆ O(K, B) such that the inverse image g−1 U (W ) is a 1A function f : U ⊂ C → B is called analytic at infinity if f can be expanded in a power series f (λ) = P∞ → C is called analytic at (∞,∞) if f can be expanded in a power series f (λ, µ) = P∞ λk µm in a neighbourhood of (∞,∞). → C is called analytic at (λ∗,∞), λ∗ ∈ C if f can be expanded in a power series A function f : U ⊂ C f (λ, µ) =P∞ in a neighbourhood of (λ∗,∞). λk in a neighbourhood of infinity. A function f : U ⊂ C ckm(λ−λ∗)k k,m=0 k,m=0 µm ck k=0 7 2 2 ckm neighbourhood of zero in O(U, B). Thus, for all U ⊃ N ⊃ K such that the inte- rior of the compact set N contains K, the inverse image g−1 U (W ) must contain the set T (N, δ) = { f : kf (N)k < δ } ⊂ O(U, B). We recall [111, ch. 2, Theorem 6.1] that a linear mapping ϕ : O(K, B) → E, where E is a Banach space, is continuous if and only if all the compositions ϕ ◦ gU : O(U, B) → E, for any neighbourhood W ⊆ E of zero there exist where U ⊃ K, are continuous, i. e. a compact set N ⊂ U and a number δ > 0 such that the interior of N contains K and ϕ ◦ gU(cid:0)T (N, δ)(cid:1) ⊆ W . We note that since E is a Banach space, it is sufficient to restrict ourselves to the consideration of ε-neighbourhoods of zero for W . Below, by misuse of language, we denote ϕ ◦ gU by the abbreviated symbol ϕ. Proposition 13 ([69, Proposition 1.3]). Let U1 ⊆ C and U2 ⊆ C be open sets. Then the natural image of the algebraic tensor product O(U1) ⊗ O(U2) is everywhere dense in O(U1 × U2). 4. Pseudo-resolvents and functional calculus We call a mapping that converts functions to operators (or transformators) a func- tional calculus. Of course, the most interesting are functional calculi that possess special properties (for example, morphisms of algebras). A pseudo-resolvent is a function that takes values in a Banach algebra and satisfies the Hilbert identity (2), like a resolvent. Every pseudo-resolvent generates a functional calculus (Theorems 25 and 26) which is a morphism of algebras and possesses the property of preserving the spectrum (Theorem 27). Let B be a Banach algebra and U ⊆ C be a subset. A function (family) λ 7→ Rλ defined on U and taking values in B is called [67, ch. 5, § 2] a pseudo-resolvent if it satisfies the Hilbert identity Rλ − Rµ = −(λ − µ)RλRµ, λ, µ ∈ U. (3) A pseudo-resolvent is called [9, p. 103] maximal if it cannot be extended to a larger set with the preservation of the Hilbert identity (3). Below (Theorem 16) we will see that every pseudo-resolvent can be extended to a unique maximal one. The domain ρ(R(·)) of the maximal extension of a pseudo-resolvent R(·) is called a regular set of the original pseudo-resolvent. The complement σ(R(·)) of the regular set ρ(R(·)) is called [6], [9, p. 103] the singular set. Example 2. The examples of pseudo-resolvents are: (a) the resolvent of an element of a unital Banach algebra (Proposition 14); (b) a constant function λ 7→ N, where N ∈ B is an arbitrary element whose square equals zero (Proposition 23); (c) in particular, the identically zero function; (d) the resolvent of a closed linear operator [67, Theorem 5.8.1]; (e) the resolvent of a linear relation [8, 9, 16, 20, 38, 57, 92]; this example is the most general, because every pseudo-resolvent is a resolvent of some linear relation [9, Theorem 5.2.4], [57, Proposition A.2.4]; (f) direct sums of pseudo-resolvents from the previous examples. A simple example of a maximal pseudo-resolvent is given in the following proposition. Proposition 14 ([89, Proposition 17]). The resolvent of an arbitrary element A ∈ B is a maximal pseudo-resolvent, i. e. it cannot be extended to a set larger than ρ(A) with the preservation of the Hilbert identity (3). 8 We note that identity (3) can be equivalently written in the form (here 1 is an adjoint unit if the original algebra has no unit) (4) Below in this Section, we will adjoin a unit to B when it has no unit. Rλ(cid:0)1 + (λ − µ)Rµ(cid:1) = Rµ. Proposition 15 ([67, Corollary 1 of Theorem 5.8.4]). Let Rλ, Rµ ∈ B be two commuting elements that satisfy identity (3). Then the element 1 + (λ − µ)Rµ ∈ eB is necessarily invertible. Theorem 16 ([67, Theorem 5.8.6]). Every pseudo-resolvent whose domain contains at least one point µ ∈ C can be extended to a maximal pseudo-resolvent; this extension is unique. The domain of the maximal extension is the set of all λ ∈ C such that the element 1 + (λ − µ)Rµ is invertible in eB. This extension can be defined by the formula (5) Rµ. Rλ = Rµ(cid:0)1 + (λ − µ)Rµ(cid:1)−1 =(cid:0)1 + (λ − µ)Rµ(cid:1)−1 We will denote the original pseudo-resolvent and its continuation to a maximal pseudo- resolvent by the same symbol R(·). Moreover, we will generally assume that all pseudo- resolvents under consideration are already extended to maximal pseudo-resolvents. Corollary 17 ([67, Theorem 5.8.2], [20, ch. 6, § 1]). The domain of a maximal pseudo- resolvent is an open set and the maximal pseudo-resolvent is an analytic function (with values in B).2 More precisely, in a neighbourhood of any point µ ∈ ρ(R(·)), the maximal pseudo-resolvent admits the power series expansion Rλ = ∞Xn=0 (µ − λ)nRn+1 µ . Proposition 18 ([67, Theorem 5.8.3]). A pseudo-resolvent R(·) in a Banach algebra B is a resolvent of some element A ∈ B if and only if B is unital and the element Rµ is invertible for at least one (and, consequently, for all) µ ∈ ρ(R(·)). A pseudo-resolvent R(·) in B(X), where X is a Banach space, is a resolvent of some unbounded operator A : D(A) ⊂ X → X if and only if the operator Rµ : X → Im Rµ is invertible for at least one (and, consequently, for all) µ ∈ ρ(R(·)). In this case A = λ1 − (Rλ)−1 for all λ ∈ ρ(R(·)). In a similar way, a linear relation can also be recovered from the value of its resolvent at one point. Thus, the resolvent contains all information about a linear relation or an operator that generates it. On the other hand, the conditions on unbounded operators and linear relations are often imposed in terms of their resolvents (the nonemptiness of the resolvent set, the estimate of decay rate at infinity etc.). Besides, functions of linear relations and unbounded operators are often expressed directly via their resolvents. For this reason, the resolvent can be considered as a more fundamental object than an operator or relation that generates it. This is the viewpoint we adhere to in this article. We fix a pseudo-resolvent R(·). We denote by BR the smallest closed subalgebra of the algebra B that contains all elements Rλ, λ ∈ ρ(R(·)), of the extension of the pseudo- resolvent R(·) to a maximal pseudo-resolvent. Proposition 19 ([89, Proposition 21]). The algebra BR coincides with the closure of the linear span of the family of all elements Rλ, λ ∈ ρ(R(·)), and is commutative. 2We accept that the domain of an analytic function may be disconnected. 9 If the algebra BR does not contain the unit of the algebra B (this is certainly the case if B is not unital), then we will, as usual, denote by eBR the algebra BR with an adjoint unit from B3 (or the algebra eB with an adjoint unit). If BR contains the unit of the algebra B, the symbol eBR is understood to be BR. Proposition 20 ([89, Theorem 22]). The subalgebra eBR is commutative and full. Proposition 21. If a character χ of the algebra BR equals zero at least at one element Rµ, µ ∈ ρ(R(·)), then it is identically equal to zero on BR, i. e. coincides with χ0. Proof. The proof follows from formula (5) and the description of BR (Proposition 19) as the closure of the linear span of the family Rλ, λ ∈ ρ(R(·)). (cid:3) We say that a sequence of maximal pseudo-resolvents Rn, (·) converges to a maximal pseudo-resolvent R(·) if there exists a point µ ∈ C such that all the pseudo-resolvents Rn, (·) are defined at µ (for n sufficiently large) and the sequence Rn, µ converges to Rµ in norm, cf. [75, Theorem 2.25]. The following lemma shows that this definition does not depend on the choice of the point µ ∈ C. Lemma 22. Let a sequence Rn, (·) of maximal pseudo-resolvents converge to a maximal pseudo-resolvent R(·) at a point µ ∈ C (it is assumed that the pseudo-resolvents Rn, µ are defined at the point µ for all n large enough). Then for any point λ ∈ ρ(R(·)), the sequence Rn, λ is defined for all n large enough and converges to Rλ in norm. Moreover, given a compact set Γ ⊂ ρ(cid:0)R(·)(cid:1), the elements Rn, λ are defined for n large enough at all λ ∈ Γ and converges to Rλ uniformly with respect to λ ∈ Γ. Proof. Let Rn, µ converge to Rµ. By Theorem 16, the element 1 + (λ − µ)Rµ is invertible for all λ ∈ Γ. Since the function λ 7→ 1 + (λ − µ)Rµ is continuous, from Theorem 1 it follows that min λ∈Γ(cid:13)(cid:13)(cid:0)1 + (λ − µ)Rµ(cid:1)−1(cid:13)(cid:13) > 0. Since Rn, µ converges to Rµ, the sequence λ 7→ 1 + (λ − µ)Rn, µ converges to λ 7→ 1 + (λ − µ)Rµ uniformly with respect to λ ∈ Γ. Therefore, again by Theorem 1, the elements 1 + (λ − µ)Rn, µ are invertible for all λ ∈ Γ provided n is large enough; in this case, by estimate (1), the inverses also converge uniformly. It remains to apply formula (5). (cid:3) We note that the limit of a sequence of resolvents of bounded operators can be the resolvent of a non-bounded operator, see [105, Lemma 7]. Proposition 23 ([67, Theorem 5.9.2]). Let a pseudo-resolvent R(·) admit an analytic continuation in a neighbourhood of the point ∞.4 Then there exist elements P, A, N ∈ BR such that N 2 = 0, P 2 = P, AP = P A = A, NP = P N = 0 and the expansion of the pseudo-resolvent into the Laurent series with centre ∞ has the form Rλ = −N + P λ + A λ2 + A2 λ3 + A3 λ4 + . . . . (6) one-dimensional subspace is a closed subspace. 3Adjoining 1 ∈ B to BR we obtain a closed subalgebra because the sum of a closed subspace and a 4We recall [54, p. 107] that the possibility of an analytic continuation of f in a neighbourhood of the point ∞ is equivalent to the existence of a bounded analytic continuation of f in a deleted neighbourhood of ∞. 10 We call the extended regular set ¯ρ(R(·)) ⊆ C of a pseudo-resolvent R(·) (in the algebra B) either the regular set ρ(R(·)) or the union ρ(R(·)) ∪ {∞}; more precisely, we add the point ∞ to ¯ρ(R(·)) if the algebra B is unital, the regular set ρ(R(·)) contains a (deleted) neighbourhood of ∞, and limλ→∞ λRλ = 1. We call the extended singular set of the pseudo-resolvent the complement ¯σ(R(·)) = C \ ¯ρ(R(·)) of the extended regular set. Proposition 24. The following properties of a maximal pseudo-resolvent are equivalent: (a) ∞ ∈ ¯ρ(R(·)); (b) the maximal pseudo-resolvent is the resolvent of some element A ∈ BR (see Propo- (c) the algebra B is unital and the subalgebra BR contains the unit of the algebra B. sition 18); Proof. The equivalence of (a) and (b) is proved in [89, Proposition 23]. Let assumption (b) be fulfilled, i. e. Rλ = (λ1 − A)−1, where A ∈ BR. Then, by virtue of Theorem 1, the power series expansion (λ1 − A)−1 = 1 λ + A λ2 + A2 λ3 + . . . holds in a neighbourhood of infinity, which shows that 1 ∈ BR, i. e., assumption (c) holds. Let assumption (c) be fulfilled, i. e. the subalgebra BR contain the unit of the algebra B. There are no identically zero characters on a unital commutative algebra, because χ(1) = 1. Therefore, by Proposition 21, χ(Rµ) 6= 0 for all χ ∈ eX(BR) and µ ∈ ρ(R(·)). Hence, by Theorem 6, all values Rµ of the pseudo-resolvent are invertible. Then, by virtue of Proposition 18, the pseudo-resolvent is the resolvent of some element A ∈ BR, i. e. assumption (b) holds. (cid:3) Below in this Section, we assume that X is a Banach space and we are given a maximal pseudo-resolvent R(·) in B(X). Let σ and Σ be two disjoint closed subsets of C. A contour Γ is called [67, ch. V, § 5.2] an oriented envelope of the set σ with respect to the set Σ if Γ is an oriented boundary of an open set U that contains σ and is disjoint from Σ. Thus, Γ surrounds the set σ in the counterclockwise direction and surrounds the set Σ in the clockwise direction. Theorem 25. Assume that ∞ /∈ ¯σ(R(·)). We define the mapping ϕ : O(cid:0)σ(R(·))(cid:1) → BR by the formula ϕ(f ) = f (λ)Rλ dλ, (7) 1 2πiZΓ where Γ (see left fig. 1) is an oriented envelope of the singular set σ(R(·)) with respect to the point ∞ and the complement of the domain of the function f . We assert that ϕ is a continuous morphism of unital algebras. The morphism ϕ maps the function u(λ) = 1 to the identity operator 1X. The function v1(λ) = λ is mapped by ϕ to the operator A ∈ B(X) that generates the maximal pseudo- resolvent R(·) in accordance with Proposition 24; and the function rλ0(λ) = 1 λ0−λ, where λ0 ∈ ρ(R(·)), is mapped by ϕ to Rλ0. Proof. The proof is analogous to that of the theorem on analytic functional calculus for bounded operators [18, ch. 1, § 4, Theorem 3], [67, Theorem 5.2.5], [110, Theorem 10.27]. (cid:3) When it is desirable to stress that the functional calculus ϕ considered in Theorem 25 is generated by the resolvent of an operator A ∈ B(X), we will use the notation RA, λ instead of Rλ, the notation ϕA instead of ϕ, and the notation f (A) instead of ϕ(f ). 11 ⑦⑦⑦⑦ ✬✩ ✫✪ f ✻Γ ✪ ✫ Figure 1. The contour Γ is an oriented envelope of the set σ with respect to ∞ and the set f (left); the contour Γ is an oriented envelope of the set ¯σ and the point ∞ with respect to the set f (right). ¯σ Theorem 26. Assume that ∞ ∈ ¯σ(R(·)). We define the mapping ϕ : O(cid:0)¯σ(R(·))(cid:1) → eBR by the formula ϕ(f ) = 1 2πiZΓ f (λ)Rλ dλ + f (∞)1, (8) where Γ is an oriented envelope of the extended singular set ¯σ(R(·)) with respect to the complement of the domain of f (see the right fig. 1). We assert that ϕ is a continuous morphism of unital algebras. The morphism ϕ maps the function u(λ) = 1 to the identity operator 1. The function rλ0(λ) = 1 λ0−λ, where λ0 ∈ ρ(R(·)), is mapped by ϕ to Rλ0. Proof. The proof is analogous to that of the theorem on analytic functional calculus for unbounded operators [67, Theorem 5.11.2]. (cid:3) When it is desirable to stress that the functional calculus ϕ considered in Theorem 26 is generated by a pseudo-resolvent R(·), we will use the notation ϕR(·) instead of ϕ and the notation f (R(·)) instead of ϕ(f ). A unified notation for formulae (7) and (8) is suggested in [67]: ✬ ✩ ✻Γ f ⑦⑦⑦ σ The following theorem is a version of the spectral mapping theorem for the case of 1 2πiZΓ ϕ(f ) = f (λ)Rλ dλ + δf (∞)1, where δ = 0 if Γ does not enclose ∞, and δ = 1 if Γ encloses ∞. pseudo-resolvents. Theorem 27. For any function f ∈ O(cid:0)¯σ(R(·))(cid:1) we have the equality σeB(cid:0)ϕ(f )(cid:1) = σeBR(cid:0)ϕ(f )(cid:1) = { f (λ) : λ ∈ ¯σ(R(·)) }. Proof. The proof is analogous to that of the spectral mapping theorem for unbounded operators [67, 5.12.1] and for linear relations [9, 5.2.17]. (cid:3) An analytic functional calculus for bounded operators was created in [34, 35, 123]. It was carried over to unbounded operators in [33, 67] and to linear relations in [8, 9, 20, 38, 57]. 5. Extended tensor products The notion of an extended tensor product is a generalization of the notion of a com- pletion of an algebraic tensor product with respect to a uniform cross-norm [28, 31, 55, 64, 112]. It enables one to extend some constructions which are natural for the usual tensor products to supplementary applications. We recall an example which is the most 12 important for our applications. It is known (see, e. g., [53]) that in the case of finite- dimensional Banach spaces X and Y , the space B(Y, X) can be identified with the tensor product X ⊗ Y ∗. If X and Y are infinite-dimensional, then X ⊗ Y ∗ corresponds only to the subspace of B(Y, X) consisting of operators that have a finite-dimensional image. Therefore the completion of X ⊗ Y ∗ with respect to any reasonable norm cannot coincide with the whole B(Y, X). Nevertheless, B(Y, X) can be represented (see example 3(e) below) as an extended tensor product X ⊠ Y ∗ which enables one to treat it almost as a usual tensor product. The exposition in this Section is based on [89]. We denote by X ⊗ Y the usual tensor product of linear spaces X and Y . In order to Let X and Y be Banach spaces. A norm α(·) = k · kα on X ⊗ Y is called a cross-norm distinguish X ⊗ Y from its extensions, we call X ⊗ Y an algebraic tensor product. if kx ⊗ ykα = kxk · kyk for all x ∈ X and y ∈ Y . We denote by X⊗α Y the completion of the tensor product X ⊗ Y by the cross-norm α. Every element v∗ =Pm l=1 x∗l ⊗ y∗l ∈ X∗ ⊗ Y ∗ induces the linear functional nXk=1 v∗ : hxk, x∗l i · hyk, y∗l i xk ⊗ yk 7→ nXk=1 mXl=1 (9) on the space X ⊗ Y . We define the norm α∗ on X∗ ⊗ Y ∗, conjugate to the cross-norm α, by the formula kv∗kα∗ = sup{ hv, v∗i : v ∈ X ⊗ Y, kvkα ≤ 1 }. A cross-norm α is called ∗-uniform if α∗ is finite and is a cross-norm. The space B(X) ⊗ B(Y ) has the natural structure of an algebra. Every element T = Pm l=1 Al ⊗ Bl ∈ B(X) ⊗ B(Y ) induces the linear operator nXk=1 in X ⊗ Y . A cross-norm α on the space X ⊗ Y induces the norm α of the operator T ∈ B(X) ⊗ B(Y ) by the formula (Alxk) ⊗ (Blyk) xk ⊗ yk 7→ mXl=1 nXk=1 T : kTk α = sup{ kT vk : v ∈ X ⊗ Y, kvkα ≤ 1 }. A cross-norm α is called [112] uniform if α is finite and is a cross-norm. Every uniform cross-norm is ∗-uniform, see [118]. Let X and Y be Banach spaces. We call an extended tensor product [89] of X and Y a collection consisting of three objects: a Banach space X ⊠ Y (which we briefly refer to as the extended tensor product) and two (not necessarily closed) full unital subalgebras B0(X) and B0(Y ) of the algebras B(X) and B(Y ) respectively that satisfy assumptions (A), (B), and (C) listed below. (A) We are given a linear mapping j from the algebraic tensor product X ⊗ Y to X ⊠ Y . In the sequel, we denote j(x ⊗ y) by the symbol x ⊠ y. It is assumed that (10) kx ⊠ ykX⊠Y = kxkX · kykY for all x ∈ X and y ∈ Y . (B) We are given a linear mapping J from the algebraic tensor product X∗ ⊗ Y ∗ to (X ⊠Y )∗. In the sequel, we denote J(x∗⊗ y∗) by the symbol x∗ ⊠ y∗. It is assumed that (11) hx ⊠ y, x∗ ⊠ y∗i = hx, x∗ihy, y∗i 13 for all x∗ ∈ X∗, y∗ ∈ Y ∗, x ∈ X, and y ∈ Y , and kx∗ ⊠ y∗k(X⊠Y )∗ = kx∗kX ∗ · ky∗kY ∗ (12) (13) (14) (15) for all x∗ ∈ X∗ and y∗ ∈ Y ∗. (C) We are given a morphism J of unital algebras from the algebraic tensor product B0(X) ⊗ B0(Y ) to B(X ⊠ Y ). In the sequel, we denote J(A ⊗ B) by the symbol A ⊠ B. It is assumed that (cid:0)A ⊠ B(cid:1)(x ⊠ y) = (Ax) ⊠ (By) for all A ∈ B0(X), B ∈ B0(Y ), x ∈ X, and y ∈ Y , and (cid:0)A ⊠ B(cid:1)∗(x∗ ⊠ y∗) = (A∗x∗) ⊠ (B∗y∗) for all A ∈ B0(X), B ∈ B0(Y ), x∗ ∈ X∗, and y∗ ∈ Y ∗, and kA ⊠ BkB(X⊠Y ) = kAkB(X) · kBkB(Y ) for all A ∈ B0(X) and B ∈ B0(Y ). Example 3. We recall [89] some examples of extended tensor products. (a) Let α be a cross-norm on an algebraic tensor product X ⊗ Y . We take for X ⊠ Y the completion X⊗α Y of the space X ⊗ Y with respect to the cross-norm α, and we In such a case, take for B0(X) and B0(Y ) the algebras B(X) and B(Y ) respectively. assumption (12) means that the cross-norm α is ∗-uniform, and assumption (15) means that the cross-norm α is uniform. (b) Let X and Y be Banach spaces. We denote by K(X, Y ) the Banach space of all bilinear forms K : X × Y → C that are bounded with respect to the norm kKk = sup{ K(x, y) : kxk ≤ 1, kyk ≤ 1 }. In order to represent K(X, Y ) as an extended tensor product X∗ ⊠ Y ∗, we take for B0(X∗) and B0(Y ∗) the subalgebras of algebras B(X∗) and B(Y ∗) consisting of all operators that have a preconjugate. We define the mappings j, J, and J by the rules (extended by linearity) [x∗ ⊠ y∗](x, y) = hx, x∗ihy, y∗i, hx∗∗ ⊠ y∗∗, Ki = K(x∗∗, y∗∗), (cid:2)(A ⊠ B)K(cid:3)(x, y) = K(A0x, B0y), where K is the canonical extension [7] of K to X∗∗ × Y ∗∗. (c) Let X and Y be Banach spaces, and X⊗α Y be a completion of the space X⊗Y with respect to a uniform cross-norm α. The conjugate space (X⊗α Y )∗ can be regarded as an extended tensor product X∗⊠Y ∗ if one takes for B0(X∗) and B0(Y ∗) the subalgebras of the algebras B(X∗) and B(Y ∗) consisting of all operators that have a preconjugate. We notice that this example is a generalization of the previous one, since K(X, Y ) ∼= (X⊗π Y )∗, where π is the largest cross-norm [28, 64, 112]. We define j : X∗ ⊗ Y ∗ → X∗ ⊠ Y ∗ = (X⊗α Y )∗ as the canonical embedding (9). Next, we define J : X∗∗ ⊗ Y ∗∗ → (X∗ ⊠ Y ∗)∗ = (X⊗α Y )∗∗. To this end, we assign to each functional w∗ ∈ X∗ ⊠ Y ∗ = (X⊗α Y )∗ the bilinear form Kw∗(x, y) = hx ⊗ y, w∗i on X × Y . For Pn k=1 x∗∗k ⊗ y∗∗k ∈ X∗∗ ⊗ Y ∗∗, we set x∗∗k ⊗ y∗∗k (cid:17), w∗E = DJ(cid:16) nXk=1 nXk=1 Kw∗(x∗∗k , y∗∗k ), where Kw∗ is the canonical extension of the bilinear form Kw∗ to X∗∗ × Y ∗∗. 14 k=1 A0 k ⊗ B0 k=1 Ak ⊗ Bk(cid:1) ∈ B(cid:0)(X⊗α Y )∗(cid:1) as the conjugate of the We define the operator J(cid:0)Pn operator Pn (d) Let X = L∞[a, b] and Y = L∞[c, d]. By example (c), the space L∞[a, b] × [c, d] can be regarded as the extended tensor product L∞[a, b] ⊠ L∞[c, d] (we recall that the space L∞[a, b] is conjugate of the space L1[a, b]). We notice that one should take for B0(X) and B0(Y ) the subalgebras of the algebras B(X) and B(Y ) consisting of all operators that have a preconjugate. k : X⊗α Y → X⊗α Y . (e) Let X and Y be Banach spaces. We represent the space B(Y, X) as an extended tensor product X ⊠ Y ∗. To this end, we take for B0(X) the whole algebra B(X) and we take for B0(Y ∗) the subalgebra of the algebra B(Y ∗) consisting of all operators that have a preconjugate. We define the mappings j, J, and J by the rules (extended by linearity) (x ⊠ y∗)y = xhy, y∗i, hU, x∗ ⊠ y∗∗i = hy∗∗, U∗x∗i, (A ⊠ B)U = AUB0. Note that in this example the subalgebra B0(Y ∗) can be thought of as B(Y ), but the action of B(Y ) on U ∈ B(Y, X) should be understood as contravariant, i. e., (A1 ⊠ B1)(cid:0)(A2 ⊠ B2)U(cid:1) = A1A2UB2B1. Below in this Section, we assume that we are given an extended tensor product X ⊠ Y of Banach spaces X and Y , and a pseudo-resolvent R(·) in the algebra B0(Y ). Theorem 28 ([89, Theorem 26]). Assume that ∞ /∈ ¯σ(R(·)). We define the mapping Φ : O(cid:0)σ(R(·)), B0(X)(cid:1) → B(X ⊠ Y ) by the formula Φ(F ) = F (λ) ⊠ Rλ dλ, (16) 1 2πiZΓ where Γ is an oriented envelope of the singular set σ(R(·)) of the pseudo-resolvent with respect to the point ∞ and the complement of the domain of F . We assert that Φ is a continuous morphism of unital algebras. For all A ∈ B0(X) and h ∈ O(cid:0)σ(R(·))(cid:1)the morphism Φ maps the function F (λ) = Ah(λ) to the operator A ⊠ ϕ(h), where ϕ is defined as in Theorem 25. We stress that the function F in (16) takes its values in B0(X), but not in C. Theorem 29 ([89, Theorem 27]). Assume that ∞ ∈ ¯σ(R(·)). We define the mapping Φ : O(cid:0)¯σ(R(·)), B0(X)(cid:1) → B(X ⊠ Y ) by the formula Φ(F ) = F (λ) ⊠ Rλ dλ + F (∞) ⊠ 1, (17) 1 2πiZΓ where Γ is an oriented envelope of the extended singular set ¯σ(R(·)) of the pseudo-resolvent with respect to the complement of the domain of F . We assert that Φ is a continuous morphism of unital algebras. Ah(λ) to the operator A ⊠ ϕ(h), where ϕ is defined as in Theorem 26. For all A ∈ B0(X) and h ∈ O(cid:0)σ(R(·))(cid:1)the morphism Φ maps the function F (λ) = Theorem 30 ([89, Theorem 41]). Let F ∈ O(cid:0)¯σ(R(·)), B0(X)(cid:1). We define the operator Φ(F ) by formula (16) if ∞ /∈ ¯σ(R(·)); and we define the operator Φ(F ) by formula (17) if ∞ ∈ ¯σ(R(·)). We assert that the operator Φ(F ) : X ⊠ Y → X ⊠ Y is not invertible if and only if for some λ ∈ ¯σ(R(·)) the operator F (λ) ∈ B0(X) is not invertible. 15 Theorem 31 ([89, Theorem 42]). Let F ∈ O(cid:0)¯σ(R(·)), B0(X)(cid:1). We define the operator Φ(F ) by formula (16) if ∞ /∈ ¯σ(R(·)); and we define the operator Φ(F ) by formula (17) if ∞ ∈ ¯σ(R(·)). We assert that the spectrum of the operator Φ(F ) : X ⊠ Y → X ⊠ Y is given by the formula σ(cid:2)Φ(F )(cid:3) = [λ∈¯σ(R(·)) σ(cid:0)F (λ)(cid:1). 6. Functional calculus ϕ1 ⊠ ϕ2 In this Section, we discuss the product ϕ1 ⊠ ϕ2 of functional calculi ϕ1 and ϕ2 that were defined in Section 4; it acts in the extended tensor product X ⊠ Y . Keeping in mind the space B(Y, X) (see Example 3(e)) as the main example of an extended tensor product, we call transformators operators acting in X ⊠ Y . In this Section, we assume that we are given an extended tensor product X ⊠ Y of Banach spaces X and Y , and we are given pseudo-resolvents R1, (·) and R2, (·) in the algebras B0(X) and B0(Y ) respectively. Theorem 32. Assume that ∞ /∈ ¯σ(R1, (·)) and ∞ /∈ ¯σ(R2, (·)). We define the mapping ϕ1 ⊠ ϕ2 : O(cid:0)σ(R1, (·)) × σ(R2, (·))(cid:1) → B(X ⊠ Y ) by the formula (cid:0)ϕ1 ⊠ ϕ2(cid:1)f = 1 (2πi)2 ZΓ1ZΓ2 f (λ, µ)R1, λ ⊠ R2, µ dµ dλ, where Γ1 and Γ2 are oriented envelopes of the singular sets σ(R1, (·)) and σ(R2, (·)) with respect to the point ∞ and the complements C \ U1 and C \ U2; here U1 × U2 is an open neighbourhood of the set σ(R1, (·))× σ(R2, (·)) that lies in the domain of the function f (see Proposition 9). We assert that ϕ1 ⊠ ϕ2 is a continuous morphism of unital algebras. For all g ∈ O(cid:0)σ(R1, (·))(cid:1) and h ∈ O(cid:0)σ(R2, (·))(cid:1) the morphism ϕ1 ⊠ ϕ2 maps the func- tion f (λ, µ) = g(λ)h(µ) to the transformator ϕ1(g) ⊠ ϕ2(h), where ϕ1 and ϕ2 are scalar functional calculi (Theorem 25) generated by the pseudo-resolvents R1, (·) and R2, (·). Proof. The proof is analogous to that of Theorem 34, see below. (cid:3) 1 1 f (λ, µ)R1, λ ⊠ R2, µ dµ dλ + (2πi)2 ZΓ1ZΓ2 Theorem 33. Assume that ∞ /∈ ¯σ(R1, (·)), but ∞ ∈ ¯σ(R2, (·)). We define the mapping ϕ1 ⊠ ϕ2 : O(cid:0)σ(R1, (·)) × ¯σ(R2, (·))(cid:1) → B(X ⊠ Y ) by the formula 2πiZΓ1 (cid:0)ϕ1 ⊠ ϕ2(cid:1)f = where Γ1 is an oriented envelope of the singular set σ(R1, (·)) with respect to the point ∞ and the complement C \ U1, and Γ2 is an oriented envelope of the extended singular set ¯σ(R2, (·)) with respect to the complement C \ U2; here U1 × U2 is an open neighbourhood of the set σ(R1, (·)) × ¯σ(R2, (·)) that lies in the domain of the function f (see Proposition 9). We assert that ϕ1 ⊠ ϕ2 is a continuous morphism of unital algebras. For all g ∈ O(cid:0)σ(R1, (·))(cid:1) and h ∈ O(cid:0)¯σ(R2, (·))(cid:1) the morphism ϕ1 ⊠ ϕ2 maps the function f (λ, µ) = g(λ)h(µ) to the transformator ϕ1(g) ⊠ ϕ2(h), where ϕ1 and ϕ2 are scalar func- tional calculi (Theorems 25 and 26) generated by the pseudo-resolvents R1, (·) and R2, (·). Proof. The proof is analogous to that of Theorem 34, see below. f (λ,∞)R1, λ ⊠ 1Y dλ, (cid:3) 16 Theorem 34. Assume that ∞ ∈ ¯σ(R1, (·)) and ∞ ∈ ¯σ(R2, (·)). We define the mapping ϕ1 ⊠ ϕ2 : O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1) → B(X ⊠ Y ) by the formula 2πiZΓ1 (cid:0)ϕ1 ⊠ ϕ2(cid:1)f = f (λ,∞)R1, λ ⊠ 1 dλ f (λ, µ)R1, λ ⊠ R2, µ dµ dλ + 1 1 (2πi)2 ZΓ1ZΓ2 2πiZΓ2 1 + f (∞, µ)1 ⊠ R2, µ dµ + f (∞,∞)1X⊠Y , where Γ1 and Γ2 are oriented envelopes of the singular sets ¯σ(R1, (·)) and ¯σ(R2, (·)) with respect to the complements C \ U1 and C \ U2; here U1 × U2 is an open neighbourhood of the set ¯σ(R1, (·)) × ¯σ(R2, (·)) that lies in the domain of the function f (see Proposition 9). We assert that ϕ1 ⊠ ϕ2 is a continuous morphism of unital algebras. For all g ∈ O(cid:0)¯σ(R1, (·))(cid:1) and h ∈ O(cid:0)¯σ(R2, (·))(cid:1) the morphism ϕ1 ⊠ ϕ2 maps the func- tion f (λ, µ) = g(λ)h(µ) to the transformator ϕ1(g) ⊠ ϕ2(h), where ϕ1 and ϕ2 are scalar functional calculi (Theorem 26) generated by the pseudo-resolvents R1, (·) and R2, (·). Proof. For each µ ∈ U2 we consider the operator G(µ) = ϕ2(cid:0)f (·, µ)(cid:1) = 1 2πiZΓ1 f (λ, µ)R1, λ dλ + f (∞, µ)1X. (18) By Theorem 26, for any fixed µ ∈ U2 the correspondence f 7→ G(µ) preserves the three operations: addition, scalar multiplication, and multiplication. We change the interpretation: formula (18) defines a mapping f 7→ G from O(cid:0)σ(R1, (·)) × σ(R2, (·))(cid:1) to O(cid:0)σ(R2, (·)), B0(X)(cid:1). Since the three operations in O(cid:0)σ(R2, (·)), B0(X)(cid:1) are understood in the pointwise sense, it follows that the correspondence f 7→ G is a morphism of algebras. In accordance with Theorem 29 we put Φ1(G) = f (λ, µ)R1, λ dλ + f (∞, µ)1X(cid:17) ⊠ R2, µ dµ 1 1 1 = = G(µ) ⊠ R2, µ dµ + G(∞) ⊠ 1Y 2πiZΓ1 f (λ,∞)R1, λ dλ + f (∞,∞)1X(cid:17) ⊠ 1Y 2πiZΓ1 f (λ, µ)R1, λ dλ(cid:17) ⊠ R2, µ dµ 2πiZΓ2 2πiZΓ2(cid:16) 1 2πiZΓ1 +(cid:16) 1 2πiZΓ2(cid:16) 1 2πiZΓ2(cid:16)f (∞, µ)1X(cid:17) ⊠ R2, µ dµ 2πiZΓ1 +(cid:16) 1 (2πi)2 ZΓ1ZΓ2 2πiZΓ1 f (λ,∞)R1, λ dλ + f (∞,∞)1X(cid:17) ⊠ 1Y 2πiZΓ2 f (λ,∞)R1, λ ⊠ 1Y dλ + f (∞,∞)1X ⊠ 1Y . f (λ, µ)R1, λ ⊠ R2, µ dµ dλ + + = + 1 1 1 1 (19) f (∞, µ)1X ⊠ R2, µ dµ By Theorem 29, the correspondence G 7→ Φ1(G) also preserves the three operations. Clearly, the mapping ϕ1 ⊠ ϕ2 from the formulation of the theorem is the composition of the correspondences f 7→ G and G 7→ Φ1(G), and, by what has been proved, is a morphism of algebras. The continuity is evident. 17 The second statement is verified by direct calculations. (cid:3) When it is desirable to stress that in Theorems 32, 33, and 34, the functional calculus ⊠ ϕ1 ⊠ϕ2 is generated by pseudo-resolvents R1, (·) and R2, (·), we will use the notation ϕR1, (·) ϕR2, (·) instead of ϕ1 ⊠ ϕ2, and we will use the notation f (R1, (·), R2, (·)) instead of (ϕ1 ⊠ ϕ2)(f ). If the pseudo-resolvents R1, (·) and R2, (·) are generated by the operators A and B (see Proposition 24), we will use the notations ϕA ⊠ ϕB and f (A, B). In order to present the definitions of ϕ1 ⊠ ϕ2 from Theorems 32, 33, and 34 in a unified form, it is convenient to use the notation (ϕ1 ⊠ ϕ2)(f ) = f (λ, µ)R1, λ ⊠ R2, µ dµ dλ 1 1 (2πi)2 ZΓ1ZΓ2 2πiZΓ1 2πiZΓ2 1 + δ2 + δ1 f (λ,∞)R1, λ ⊠ 1 dλ (20) f (∞, µ)1 ⊠ R2, µ dµ + δ1δ2f (∞,∞)1 ⊠ 1, where δi = 1 if Γi encloses ∞, and δi = 0 in the opposite case, i = 1, 2. functions. We enumerate the results of the action of ϕ1 ⊠ ϕ2 on some frequently encountered 1 ⊠ R2, µ0. ⊠ 1Y ; the function r2,µ0(λ, µ) = 1 Corollary 35. Under the assumptions of Theorems 32, 33, and 34, the morphism ϕ1 ⊠ϕ2 maps the function u(λ, µ) = 1 to the unit 1 ⊠ 1 of the algebra B(X ⊠ Y ); the function r1,λ0(λ, µ) = 1 λ0−λ, where λ0 ∈ ρ(R1, (·)), is mapped by the morphism ϕ1 ⊠ ϕ2 to the µ0−µ, where µ0 ∈ ρ(R2, (·)), is mapped transformator R1, λ0 by the morphism ϕ1 ⊠ ϕ2 to the transformator 1X ⊠ R2, µ0; the function rλ0,µ0(λ, µ) = (λ0−λ)(µ0−µ) , where λ0 ∈ ρ(R1, (·)) and µ0 ∈ ρ(R2, (·)), is mapped by the morphism ϕ1 ⊠ ϕ2 to the transformator R1, λ0 Under the assumptions of Theorems 32 and 33, the morphism ϕ1 ⊠ϕ2 maps the function c1(λ, µ) = λ to the transformator A ⊠ 1Y , where A is the operator that generates the maximal pseudo-resolvent R1, (·) in accordance with Proposition 24. Under the assumptions of Theorem 32, the morphism ϕ1 ⊠ ϕ2 maps the function c2(λ, µ) = µ to the transformator 1X ⊠ B, where B is the operator that generates the max- imal pseudo-resolvent R1, (·) in accordance with Proposition 24; the function rν0(λ, µ) = ν0−λ∓µ is mapped by the morphism ϕ1 ⊠ϕ2 to the transformator (ν01 ⊠1−A ⊠1∓1 ⊠B)−1 provided ν0 /∈ σ(A) ± σ(B). Proof. We restrict ourselves to proving the last statement. Clearly, the function (λ, µ) 7→ ν0−λ∓µ is mapped by the morphism ϕ1 ⊠ϕ2 to the transformator ν01 ⊠1−A ⊠1∓1 ⊠B. Since ϕ1 ⊠ ϕ2 is a morphism of algebras, the reciprocal function is mapped to the inverse transformator. (cid:3) 1 Theorem 36. Let g ∈ O(cid:0)¯σ(R1, (·))× ¯σ(R2, (·))(cid:1) and f ∈ O(cid:0)g(cid:0)¯σ(R1, (·))× ¯σ(R2, (·))(cid:1)(cid:1). Then the transformator (ϕ1 ⊠ ϕ2)(f ◦ g) is the function f of the transformator (ϕ1 ⊠ ϕ2)(g): (ϕ1 ⊠ ϕ2)(f ◦ g) = 1 2πiZΓ3 f (ν)(cid:0)ν1 ⊠ 1 − (ϕ1 ⊠ ϕ2)(g)(cid:1)−1 dν, where Γ3 is an oriented envelope of the spectrum σ(cid:0)(ϕ1 ⊠ ϕ2)(g)(cid:1). 18 1 f(cid:0)g(λ, µ)(cid:1)R1, λ ⊠ R2, µ dµ dλ (2πi)2 ZΓ1ZΓ2 f(cid:0)g(λ,∞)(cid:1)R1, λ ⊠ 1 dλ f(cid:0)g(∞, µ)(cid:1)1 ⊠ R2, µ dµ + δ1δ2f(cid:0)g(∞,∞)(cid:1)1X⊠Y dν(cid:21)R1, λ ⊠ R2, µ dµ dλ 2πiZΓ3 f (ν) Proof. We have (δ1, δ2 = 0, 1) 1 + δ2 + δ1 (ϕ1 ⊠ ϕ2)(f ◦ g) = 2πiZΓ1 1 2πiZΓ2 (2πi)2 ZΓ1ZΓ2(cid:20) 1 2πiZΓ1(cid:20) 1 2πiZΓ3 2πiZΓ2(cid:20) 1 2πiZΓ3 + δ1 + δ2 = 1 1 1 ν − g(λ, µ) f (ν) ν − g(λ,∞) f (ν) dν(cid:21)R1, λ ⊠ 1 dλ dν(cid:21)1 ⊠ R2, µ dµ + δ1δ2f(cid:0)g(∞,∞)(cid:1)1X⊠Y 1 = ν − g(∞, µ) (here we interchange the order of integration) (2πi)2 ZΓ1ZΓ2 2πiZΓ1 2πiZΓ2 f (ν)(cid:20) f (ν)(cid:20) 1 f (ν)(cid:20) 1 2πiZΓ3 2πiZΓ3 2πiZΓ3 + δ2 + δ1 1 1 1 dµ dλ(cid:21) dν R1, λ ⊠ R2, µ ν − g(λ, µ) dλ(cid:21) dν dµ(cid:21) dν + δ1δ2f(cid:0)g(∞,∞)(cid:1)1X⊠Y R1, λ ⊠ 1 ν − g(λ,∞) 1 ⊠ R2, µ ν − g(∞, µ) (further, by Theorems 32, 33, and 34, it follows that) R1, λ ⊠ 1 ν − g(λ,∞) dλ 1 1 1 1 = 2πiZΓ3 2πiZΓ2 − δ1 2πiZΓ3 2πiZΓ3 2πiZΓ3 f (ν)(cid:20)(cid:0)ν1 ⊠ 1 − (ϕ1 ⊠ ϕ2)(g)(cid:1)−1 − δ2 2πiZΓ1 dµ − δ1δ2g(∞,∞)1X⊠Y(cid:21) dν 1 ⊠ R2, µ ν − g(∞, µ) f (ν)(cid:20) 1 2πiZΓ1 f (ν)(cid:20) 1 2πiZΓ2 f (ν)(cid:0)ν1 ⊠ 1 − (ϕ1 ⊠ ϕ2)(g)(cid:1)−1 dν. (cid:3) R1, λ ⊠ 1 ν − g(λ,∞) 1 ⊠ R2, µ ν − g(∞, µ) + δ1 + δ2 = 1 1 dλ(cid:21) dν dµ(cid:21) dν + δ1δ2f(cid:0)g(∞,∞)(cid:1)1X⊠Y 1 f (λ ± µ)RA, λ ⊠ RB, µ dµ dλ = Corollary 37. Let A ∈ B(X) and B ∈ B(Y ). Let f ∈ O(cid:0)σ(A) ± σ(B)(cid:1). Then 2πiZΓ3 (2πi)2 ZΓ1ZΓ2 where Γ3 is an oriented envelope of σ(A) ± σ(B). Proof. This is a special case of Theorem 36 for g(λ, µ) = λ ± µ. Example 4. Let A ∈ B(X) and B ∈ B(Y ). By Corollary 37 and the formula eλteµt = e(λ+µ)t, one has (cf. [12, 53], [65, Theorem 10.9]) f (ν)(ν1 ⊠ 1 − A ⊠ 1 ∓ 1 ⊠ B)−1 dν, (cid:3) 1 eAt ⊠ eBt = e(A⊠1+1⊠B)t. 19 We proceed to the discussion of spectral mapping theorems. Theorem 38. Let f ∈ O(cid:0)¯σ(R1, (·))×¯σ(R2, (·))(cid:1). Then the transformator(cid:0)ϕ1⊠ϕ2(cid:1)(f ) : X ⊠ Y → X ⊠ Y is not invertible if and only if f (λ, µ) = 0 for at least one couple of points λ ∈ ¯σ(R1, (·)) and µ ∈ ¯σ(R2, (·)). Proof. For each µ ∈ U2, we consider operator (18). By Theorem 27, the following state- ment holds: the operator G(µ) : X → X is not invertible if and only if f (λ, µ) = 0 for at least one λ ∈ ¯σ(R1, (·)). Further, by Theorem 30, operator (19) is not invertible if and only if G(µ) is not invertible for at least one µ ∈ ¯σ(R2, (·)). Combining (in the opposite order) all these results, we arrive at the desired statement. Theorem 39. Let f ∈ O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1). Then the spectrum of the transformator (ϕ1 ⊠ ϕ2)f : X ⊠ Y → X ⊠ Y is given by the formula (cid:3) σ(cid:0)(ϕ1 ⊠ ϕ2)f(cid:1) = { f (λ, µ) : λ ∈ ¯σ(R1, (·)), µ ∈ ¯σ(R2, (·)) }. Proof. We take an arbitrary ν ∈ C. By the definition of the spectrum, the number ν belongs to the set σ(cid:0)(ϕ1 ⊠ ϕ2)f(cid:1) if and only if the transformator ν1X⊠Y −(cid:0)ϕ1 ⊠ ϕ2(cid:1)(f ) is not invertible. We denote by u the function from O(cid:0)¯σ(R1, (·))× ¯σ(R2, (·))(cid:1) that identically equals 1. By Theorems 32, 33, and 34, we have (cid:0)ϕ1 ⊠ ϕ2(cid:1)(u) = 1X⊠Y , whence ν1X⊠Y −(cid:0)ϕ1 ⊠ ϕ2(cid:1)(f ) =(cid:0)ϕ1 ⊠ ϕ2(cid:1)(νu − f ). We apply Theorem 38: the transformator (cid:0)ϕ1 ⊠ ϕ2(cid:1)(νu − f ) is not invertible if and only if νu(λ, µ) − f (λ, µ) = 0 for some λ ∈ ¯σ(R1, (·)) and µ ∈ ¯σ(R2, (·)) or, in other words, ν ∈ { f (λ, µ) : λ ∈ ¯σ(R1, (·)), µ ∈ ¯σ(R2, (·)) }. We denote by BR1, R2 the closure in B(cid:0)B(X, Y )(cid:1) of the set of all transformators (ϕ1 ⊠ ϕ2)f , where f ∈ O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1). Corollary 40. The set BR1, R2 is a full commutative subalgebra of the algebra B(cid:0)B(X, Y )(cid:1) Proof. Clearly, the image under ϕ1 ⊠ ϕ2 of the unital commutative algebra O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1) is a unital commutative subalgebra. Let the transformator (ϕ1 ⊠ ϕ2)f be invertible. By Theorem 39, this means that f (λ, µ) 6= 0 for some λ ∈ ¯σ(R1, (·)) and µ ∈ ¯σ(R2, (·)). Clearly, the inverse of (ϕ1 ⊠ ϕ2)f is the transformator (ϕ1 ⊠ ϕ2) 1 f . of all transformators acting in B(X, Y ). (cid:3) It remains to apply Proposition 4. (cid:3) In [113], an analogue of Theorem 32 was proved in the tensor product of Banach spaces for bounded operators and a polynomial function f . In [69, Theorem 2.4], an analogue of Theorem 32 was proved in the tensor product of Banach spaces for bounded operators and an arbitrary analytic function f ; in [69, Theorem 2.4], an analogue of Theorem 34 was also proved for unbounded operators and analytic functions. See also the initial version [68] of article [69]. An analogue of Theorem 36 for matrices was proved in [85, Theorem 4.4]. There are several versions of Theorem 39 in tensor products of Banach spaces. It was shown in [19] that the spectrum of the tensor product A ⊗ B of two bounded operators acting in a Hilbert space is the set σ(A) × σ(B). For functions f of the form f (λ, µ) = 20 g(λ)h(µ), Theorem 39 was proved in [95]; for polynomial functions f of two variables, a version of Theorem 39 was proved in [60, Theorem 3.3], see also [59]; another equivalent version was proved in [37, Theorem 3.4]. In [69, Theorem 3.2], it was proved an analogue of Theorem 39 for unbounded operators and analytic functions, see also [68]. A modern version of Theorem 39 for matrices can be found in [85, Lemma 4.1]. A functional calculus for the transformator A ⊗ 1 − 1 ⊗ B was first described in [109]. Functions of the transformator A ⊗ 1 ± 1 ⊗ B are also investigated in [11, 12, 48]. 7. Meromorphic functional calculus 2 A meromorphic function of a bounded operator is an unbounded operator or a linear relation (provided a pole of the function is contained in the spectrum). According to our approach, we identify such an object with its resolvent. Let U be an open subset of C and f : U → C. The function f is called [115, ch. IV, § 15.43] meromorphic if: (i) f is analytic on a set U \ M, where M is a nowhere dense closed subset of U, (ii) f cannot be analytically continued to any point of M, (iii) for any point ζ ∈ M there exist a connected neighborhood V of ζ and an analytic function qζ : V → C such that the function pζ = f · qζ is analytic in V ∩ (U \ M) and can be extended analytically into V , and qζ equals zero only on V ∩ M. Clearly, qζ(ζ) = 0. The set M is called the polar set of the function f . It consists of points of two types: if pζ(ζ) 6= 0 (and so limz→ζ f (z) = ∞), then ζ is called a pole; if pζ(ζ) = 0, then ζ is called a point of indeterminacy. In any neighbourhood of a point of indeterminacy, the function f takes any value from C [115]. For example, for the function f (λ, µ) = λµ, the points of indeterminacy are (0,∞) and (∞, 0), for the function f (λ, µ) = λ µ, the points of indeterminacy are (0, 0) and (∞,∞), and for the function f (λ, µ) = λ − µ, the point of indeterminacy is (∞,∞). Assume that we are given an extended tensor product X ⊠ Y of Banach spaces X and Y , and we are given pseudo-resolvents R1, (·) and R2, (·) in the algebras B0(X) and B0(Y ) respectively. We consider a function f that is meromorphic in a neighbourhood U ⊆ C f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) = { f (λ, µ) : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·)) } ¯σ(R1, (·)) × ¯σ(R2, (·)) and has no points of indeterminacy in U. We consider the subset of the set C. The set f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) is compact, being the image under the con- tinuous function f of the compact set ¯σ(R1, (·)) × ¯σ(R2, (·)). Lemma 41. For any ν /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) the set of the set 2 2 (C \ U) ∪ { (λ, µ) ∈ U : f (λ, µ) = ν } and does not intersect ¯σ(R1, (·)) × ¯σ(R2, (·)). Moreover, for any closed set is closed in C 2 W ⊆ C \ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1), the set 2 (C \ U) ∪ { (λ, µ) ∈ U : f (λ, µ) ∈ W } is closed in C 2 and does not intersect ¯σ(R1, (·)) × ¯σ(R2, (·)). 2 \ U is closed, being a complement of an open one. The set { (λ, µ) ∈ Proof. The set C U : f (λ, µ) ∈ W } = f−1(W ) is closed in U, being the inverse image of the closed set W under the continuous function f . This means that limit points of the set f−1(W ) = 21 (21) (22) 2 { (λ, µ) ∈ U : f (λ, µ) ∈ W } either belongs to f−1(W ) or to the complement of C \ U. Thus, set (22) is closed. We show that set (22) is disjoint from ¯σ(R1, (·))× ¯σ(R2, (·)). Actually, if f (λ, µ) = ν ∈ W and (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·)), then ν ∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1), which contradicts the assumption. If (λ, µ) /∈ U, then (λ, µ) /∈ ¯σ(R1, (·)) × ¯σ(R2, (·)) by the definition of U. For all ν ∈ C \ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1), we set ν − f (λ, µ) 1 R1, λ ⊠ R2, µ dµ dλ Sν = (cid:3) 1 R1, λ ⊠ 1 dλ (23) 1 1 (2πi)2ZΓ1ZΓ2 2πiZΓ1 2πiZΓ2 1 + δ2 + δ1 ν − f (λ,∞) 1 ν − f (∞, µ) 1 ⊠ R2, µ dµ + 1 ⊠ 1, δ1δ2 ν − f (∞,∞) morphism ϕ1 ⊠ ϕ2 of the function hν: where Γi is an oriented envelope of the spectrum σ(Ri,(·)); δi = 1 if Γi encloses ∞, and δi = 0 in the opposite case; i = 1, 2. By Lemma 41, the function hν(λ, µ) = ν−f (λ,µ) belongs to O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1). Therefore Sν can be regarded as the image under the We denote by Sν(cid:0)R1, (·), R2, (·)(cid:1) transformator (23) generated by the pseudo-resolvents R1, (·) and R2, (·), and we call Sν the resolvent of the function f of R1, λ and R2, λ. Theorem 42. Let the function f be meromorphic in an open neighbourhood U ⊆ C of the set ¯σ(R1, (·)) × ¯σ(R2, (·)) and have no points of indeterminacy in U. Then the family Sν = (ϕ1 ⊠ ϕ2)hν. 1 2 (24) (25) Equality (24) can be considered as an analogue of the spectral mapping theorem. Proof. We show that, on the set C \ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1), the Hilbert identity holds: Sν1 − Sν2 = −(ν1 − ν2)Sν1Sν2, We note that 1 ν1 − f (λ, µ) − 1 ν2 − f (λ, µ) = − ν1, ν2 /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). (cid:0)ν1 − f (λ, µ)(cid:1)(cid:0)ν2 − f (λ, µ)(cid:1). ν1 − ν2 Applying the morphism ϕ1 ⊠ ϕ2 to this identity we arrive at the Hilbert identity (25). We verify that the pseudo-resolvent S(·) is maximal. The validity of the Hilbert identity implies that σ(S(·)) ⊆ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). To prove the reverse inclusion, we fix an auxiliary point ν ∈ C \ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). By Theorem 16, the pseudo-resolvent S(·) can be extended to points η ∈ C in which the transformator 1+(η−ν)Sν is invertible. By Theorems 32, 33, 34, and 39, and formula (23) we have σ(Sν) =n 1 ν − f (λ, µ) 22 : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·))o. defined by formula (23) is a maximal pseudo-resolvent. In particular, Sν, ν /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1), ¯σ(S(·)) = f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). It follows that σ(1 + (η − ν)Sν) =n1 + ν − f (λ, µ) η − ν =n η − f (λ, µ) ν − f (λ, µ) : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·))o : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·))o. which is equivalent to From this formula it is seen that the transformator 1 + (η − ν)Sν is invertible if and only if 0 /∈n η − f (λ, µ) : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·))o, ν − f (λ, µ) η /∈(cid:8)f (λ, µ) : (λ, µ) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·))(cid:9). Thus, from η /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) it follows that η /∈ σ(S(·)). It remains to analyze the case ν = ∞. We assume that ∞ /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). Then, since the set f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) is closed, a neighbourhood W ⊆ C of infinity is also disjoint from f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). Without loss of generality, we may assume that the neighbourhood W is closed. By definition, Sν is defined for all ν ∈ W \ {∞}; besides, by Lemma 41, we may assume that the contours Γ1 and Γ2 in (23) do not depend on ν ∈ W \ {∞}. We calculate the limit (see Theorems 32, 33, and 34): νSν = lim lim ν→∞ ν 1 (2πi)2ZΓ1ZΓ2 ν − f (λ,∞) ν→∞(cid:18) 1 2πiZΓ1 2πiZΓ2 (ϕ1 ⊠ ϕ2)(cid:16) ν − f (∞, µ) ν 1 ν + δ2 + δ1 R1, λ ⊠ R2, µ dµ dλ ν ν − f (λ, µ) R1, λ ⊠ 1 dλ δ1δ2ν ν − f (∞,∞) 1 ⊠ 1(cid:19) 1 ⊠ R2, µ dµ + ν − f (·,·)(cid:17) = lim ν lim ν→∞ νSν = lim = lim ν→∞ ν ν−f (·,·) converge to u as ν → ∞ uniformly on Γ1×Γ2; here u(λ, µ) = Conversely, let ∞ /∈ ¯σ(S(·)). This means that S(·) is defined in a deleted neighbourhood because the functions 1. Consequently, ∞ /∈ ¯σ(S(·)). W of infinity and (ϕ1 ⊠ ϕ2)u = 1 ⊠ 1, ν→∞ R1, λ ⊠ R2, µ dµ dλ ν − f (λ, µ) R1, λ ⊠ 1 dλ (2πi)2 ZΓ1ZΓ2 ν − f (λ,∞) ν→∞(cid:18) 1 2πiZΓ1 2πiZΓ2 By the definition of S(·), we have that W ∩ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1) = ∅. We show that ∞ /∈ f(cid:0)¯σ(R1, (·)), ¯σ(R2, (·))(cid:1). Assuming the contrary, let f (λ∗, µ∗) = ∞ for a point (λ∗, µ∗) ∈ ¯σ(R1, (·)) × ¯σ(R2, (·)). Then, by Theorem 39, 0 ∈ ¯σ(Sν) for any ν ∈ W . We also have 0 ∈ ¯σ(νSν) for ν ∈ W . By Corollaries 40 and 8, it follows that 1 ⊠ 1(cid:19) = 1 ⊠ 1. ν − f (∞,∞) ν − f (∞, µ) 1 ⊠ R2, µ dµ + δ1δ2ν + δ1 + δ2 (26) 1 1 ν ν 0 ∈ ¯σ(cid:0) lim ν→∞ 23 νSν(cid:1), which contradicts (26). (cid:3) Corollary 43, see below, answers in the affirmative to the question posed in [105, 107] about the independence of the definition of f (A, B) for unbounded operators A and B from the choice of sequences of bounded operators An and Bn, the resolvents of which converge to the resolvents of A and B respectively. Corollary 43. Let the sequences of pseudo-resolvents Rn, 1, (·) and R′n, 1, (·) converge5 to the same pseudo-resolvent R1, (·), and let the sequences of pseudo-resolvents Rn, 2, (·) and R′n, 2, (·) converge to the same pseudo-resolvent R2, (·). Then both the sequence Sν(cid:0)Rn, 1, (·), Rn, 2, (·)(cid:1) and the sequence Sν(cid:0)R′n, 1, (·), R′n, 2, (·)(cid:1) converge to the pseudo-resolvent Sν(cid:0)R1, (·), R2, (·)(cid:1). Proof. We make use of definition (23). By Lemma 22, Rn, 1, (·) and R′n, 1, (·) converge to R1, (·) uniformly on Γ1, and Rn, 2, (·) and R′n, 2, (·) converge to R2, (·) uniformly on Γ2. From formula (23) it is seen that Sν(cid:0)Rn, 1, (·), Rn, 2, (·)(cid:1) and Sν(cid:0)R′n, 1, (·), R′n, 2, (·)(cid:1) converge to Sν(cid:0)R1, (·), R2, (·)(cid:1). The theory of meromorphic functions of one operator had its origin in the polynomial functional calculus for unbounded operators constructed in [124], see an exposition in [36, ch. VII, § 9]. Meromorphic functional calculus of one operator was constructed in [57]. A spectral mapping theorem for a polynomial of a linear relation was proved in [20, Theorem VI.5.4]. (cid:3) Polynomial functions of two unbounded operators were defined in [69, Theorem 3.4]; in particular, a spectral mapping theorem was established, see [69, Theorem 3.13]. Other analogues of the spectral mapping theorem for analytic functions of unbounded operators (including polynomials) were obtained in [107, Theorem 1] and [105, Theorem 4]. 8. Functional calculus ϕ1 ⊡ ϕ2 In this Section, we assume that we are given an extended tensor product X ⊠ Y of Banach spaces X and Y , and we are given pseudo-resolvents R1, (·) and R2, (·) in the algebras B0(X) and B0(Y ) respectively. We define the mapping ϕ1 ⊡ ϕ2 acting on functions f ∈ O(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1) of one variable by the formula (ϕ1 ⊡ ϕ2)f = f (λ)R1, λ ⊠ R2, λ dλ, (27) 1 2πiZΓ where Γ is an oriented envelope of the union ¯σ(R1, (·)) ∪ ¯σ(R2, (·)) of the extended singular sets with respect to the complement of the domain of f . Let U ⊆ C be an open set and f : U → C be an analytic function. We call the divided difference [43, 73] of the function f the function f [1] : U × U → C defined by the formula f [1](λ, µ) =  , f (λ)−f (µ) λ−µ f′(λ), 0, if λ 6= µ, if λ = µ, if λ = ∞ or µ = ∞. (28) 5See the definition on p. 10. 24 Example 5. We give examples of divided differences of some functions: v[1] 1 (λ, µ) = 1, v[1] 2 (λ, µ) = λ + µ, v[1] n (λ, µ) = λn−1 + λn−2µ + · · · + µn−1, v[1] 1/2(λ, µ) = 1 , √λ + √µ 1 , r[1] 1 (λ, µ) = 1 (λ0 − λ)(λ0 − µ) (λ0−λ)n − 1 (λ0−µ)n r[1] n (λ, µ) = − (λ0 − λ) − (λ0 − µ) v[1] n (λ0 − λ, λ0 − µ) (λ0 − λ)n(λ0 − µ)n , = = where v1(λ) = λ, where v2(λ) = λ2, where vn(λ) = λn, where v1/2(λ) = √λ, where r1(λ) = 1 λ0 − λ , where rn(λ) = 1 (λ0 − λ)n . The Taylor series for the divided difference of a function f at a point (λ0, λ0) has the form f [1](λ, µ) = ∞Xn=0 f (n+1)(λ0) (n + 1)! v[1] n+1(λ − λ0, µ − λ0) = ∞Xn=0 f (n+1)(λ0) (n + 1)! nXi=0 (λ − λ0)n−i(µ − λ0)i, where vn(λ) = λn. In particular, for expt(λ) = eλt and exp(1) tn tn t (λ) = λeλt we have exp[1] t (λ, µ) = exp(1) [1] t (λ, µ) = ∞Xn=0 ∞Xn=0 v[1] n+1(λ, µ) = (n + 1)! tn n! v[1] n+1(λ, µ) = ∞Xn=0 tn n! ∞Xn=0 nXi=0 (n + 1)! λn−iµi. nXi=0 λn−iµi, Proposition 44. Let U ⊆ C be an open set and f : U → C be an analytic function. Then the function f [1] is analytic in U × U. Proof. The analyticity at a finite point (λ, µ), λ 6= µ, is evident. The analyticity at the points of the form (λ,∞) and (∞, µ), where λ, µ ∈ C, is also evident. We expand f in the Taylor series about a finite point λ0 6= ∞: f (λ) = ∞Xn=0 It follows that for λ 6= µ close to λ0 one has ∞Xn=1 f [1](λ, µ) = cn(λ − λ0)n. cnv[1] n (λ − λ0, µ − λ0), where v[1] in a neighbourhood of the point (λ0, λ0). Clearly, f [1](λ0, λ0) = f′(λ0). n (λ, µ) = λn−1 + λn−2µ + · · · + µn−1. This series determines an analytic function We expand f in the Laurent series with centre ∞: cn λn . f (λ) = ∞Xn=0 25 This formula shows that for λ 6= µ close to ∞ one has v[1] n (λ, µ) λnµn f [1](λ, µ) = − cn ∞Xn=1 , n (λ, µ) = λn−1 + λn−2µ + · · · + µn−1. This series determines an analytic function where v[1] in a neighbourhood of the point (∞,∞). Clearly, f [1](∞,∞) = 0. Theorem 45. Let f ∈ O(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1). Then6 The spectrum of the transformator (ϕ1 ⊡ ϕ2)f : X ⊠ Y → X ⊠ Y is given by the formula (ϕ1 ⊡ ϕ2)f = (ϕ1 ⊠ ϕ2)f [1]. (cid:3) ⑦ ⑦ ⑦ ⑦ ⑦ ⑦ σ(cid:0)(ϕ1 ⊡ ϕ2)f(cid:1) =(cid:8) f [1](λ, µ) : λ ∈ ¯σ(R1, (·)), µ ∈ ¯σ(R2, (·))(cid:9). ✬✩✬ ✫✪ ✫ ⑦⑦⑦⑦ ✻✻ Γ1 σ1 ¯σ2 f Γ2 ✩ ✪ Figure 2. The contours Γ1 and Γ2 from the proof of Theorem 45. The localization of the complement of the domain of f is marked by f Proof. We take contours Γ1 and Γ2 such that the both are oriented envelopes of ¯σ(R1, (·))∪ ¯σ(R2, (·)) with respect to the complement of the domain of the function f , and Γ2 lies outside of Γ1 (so that λ − µ does not vanish for λ ∈ Γ1 and µ ∈ Γ2), see fig. 2. We make use of the definition: (ϕ1 ⊠ ϕ2)f [1] = f [1](λ, µ)R1, λ ⊠ R2, µ dµ dλ We represent the last integral as the sum of two iterated integrals: By the Cauchy integral formula, for the internal integral in (29) we have 6Strictly speaking, in this formula, f [1] O(cid:2)(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1) ×(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1)(cid:3) into O(cid:0)¯σ(R1, (·)) × ¯σ(R2, (·))(cid:1). 26 R2, µ dµ = R2, λ, λ − µ is understood to be the canonical projection of f [1] ∈ + δ2 + δ1 = 1 1 1 1 (2πi)2 ZΓ1ZΓ2 2πiZΓ1 2πiZΓ2 (2πi)2 ZΓ1ZΓ2 λ − µ 2πiZΓ1 2πi ZΓ2 R1, λ ⊠(cid:16) f (λ) 2πiZΓ2(cid:16)f (µ) 2πi ZΓ1 µ − λ 2πiZΓ2 1 1 1 1 1 + f [1](λ,∞)R1, λ ⊠ 1 dλ f [1](∞, µ)1 ⊠ R2, µ dµ + δ1δ2f [1](∞,∞)1 ⊠ 1 f (λ) − f (µ) R1, λ ⊠ R2, µ dµ dλ. 1 R2, µ dµ(cid:17) dλ λ − µ R1, λ dλ(cid:17) ⊠ R2, µ dµ. (29) (30) and by the Cauchy integral theorem, for the internal integral in (30) we have 1 2πiZΓ1 1 µ − λ R1, λ dλ = 0, since, by the assumption, the contour Γ1 does not surround the singularities of the function λ 7→ 1 µ−λ, µ ∈ Γ2, and the pseudo-resolvent λ 7→ R1, λ. Thus, the original integral takes the form 1 2πiZΓ1 f (λ)R1, λ ⊠ R2, λ dλ =(cid:2)(ϕ1 ⊡ ϕ2)f(cid:3). The second formula follows from 39. Theorem 46. Let A ∈ B(X), B ∈ B(Y ), and f ∈ O(cid:0)σ(A) ∪ σ(B)(cid:1). Then ϕA(f ) ⊠ 1 − 1 ⊠ ϕB(f ) =(cid:2)(ϕA ⊡ ϕB)f(cid:3)(A ⊠ 1 − 1 ⊠ B), where the functional calculi ϕA and ϕB are constructed by A and B respectively. Proof. The proof follows from the identity and Theorems 45 and 32. f (λ) − f (µ) = f [1](λ, µ)(λ − µ) (cid:3) (cid:3) In the following corollary, we describe a representation for the increment of an analytic function. Corollary 47. Let A, B ∈ B(X) and f ∈ O(cid:0)σ(A) ∪ σ(B)(cid:1). Then f (A) − f (B) = f (λ)(λ1 − A)−1(A − B)(λ1 − B)−1 dλ, 1 2πiZΓ where the functional calculi ϕA and ϕB are constructed by A and B respectively, and ϕA ⊠ ϕB acts in the extended tensor product B(X, X), see Example 3(e). For the function f = expt, this formula was found in [126, p. 978]. Proof. We apply the formula from Theorem 46 to the operator C = 1, assuming that X = Y . We have (taking into account that C = 1) (A ⊠ 1 − 1 ⊠ B)C = AC − CB = A − B, (cid:2)(ϕA ⊡ ϕB)f(cid:3)(A − B) = 1 2πiZΓ f (λ)(λ1 − A)−1(A − B)(λ1 − B)−1 dλ, (cid:0)ϕA(f ) ⊠ 1 − 1 ⊠ ϕB(f )(cid:1)C = ϕA(f )C − CϕB(f ) = f (A)C − Cf (B) = f (A) − f (B). (cid:3) One of the primary ideas [30, 40, 65, 88, 91, 99] of approximate calculation of an analytic function f of an operator or a pseudo-resolvent consists in an approximation of f by a polynomial or a rational function. In the case of (ϕ1 ⊡ ϕ2)f , for applying this idea it is necessary to be able to calculate ϕ1 ⊡ ϕ2 at least of monomials and elementary rational functions. Formulae of this kind are presented in Corollary 48 below. Corollary 48. If the pseudo-resolvents R1, (·) and R2, (·) are generated by the operators A and B respectively, then (cid:0)ϕA ⊡ ϕB(cid:1)(vn) = An−1 ⊠ 1 + An−2 ⊠ B + · · · + 1 ⊠ Bn−1, If λ0 ∈ ρ(R1, (·)) ∩ ρ(R2, (·)), then (cid:0)ϕ1 ⊡ ϕ2(cid:1)(r1) = R1, λ0 ⊠ R2, λ0, 27 where r1(λ) = where vn(λ) = λn. 1 λ0 − λ . If, in addition, the extended singular sets of the pseudo-resolvents R1, (·) and R2, (·) are disjoint, then (cid:0)ϕ1⊡ϕ2(cid:1)(rn) = −(cid:0)Rn (λ0 − λ)n . If, in addition, the pseudo-resolvents R1, (·) and R2, (·) are generated by the operators A and B respectively, then ⊠1−1⊠R2, λ0(cid:1)−1, where rn(λ) = 2, λ0(cid:1)(cid:0)R1, λ0 ⊠1−1⊠Rn 1, λ0 1 (cid:0)ϕA ⊡ ϕB(cid:1)(rn) =(cid:0)(λ01 − A)n−1 ⊠ 1 + · · · + 1 ⊠ (λ01 − B)n−1(cid:1)(cid:0)Rn ⊠ Rn A, λ0 B, λ0(cid:1). Proof. It suffices to make use of Example 5 and to apply Theorem 45 and Corollary 35. (cid:3) Corollary 49. Let g, h ∈ O(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1). Then (cid:0)ϕ1 ⊡ ϕ2(cid:1)(gh) =(cid:2)(cid:0)ϕ1 ⊡ ϕ2(cid:1)(g)(cid:3)(cid:0)1 ⊠ ϕ2(h)(cid:1) +(cid:0)ϕ1(g) ⊠ 1(cid:1)(cid:2)(cid:0)ϕ1 ⊡ ϕ2(cid:1)(h)(cid:3), where (gh)(λ) = g(λ)h(λ). Proof. By Theorem 45, this formula is equivalent to the identity (cid:0)ϕ1 ⊠ ϕ2(cid:1)(gh)[1] =(cid:2)(cid:0)ϕ1 ⊠ ϕ2(cid:1)g[1](cid:3)(cid:0)1 ⊠ ϕ2(h)(cid:1) +(cid:0)ϕ1(g) ⊠ 1(cid:1)(cid:2)(cid:0)ϕ1 ⊠ ϕ2(cid:1)h[1](cid:3). We have g(λ)h(λ) − g(µ)h(µ) λ − µ = = = g(λ)h(λ) − g(µ)h(λ) + g(µ)h(λ) − g(µ)h(µ) g(λ)h(λ) − g(µ)h(λ) g(µ)h(λ) − g(µ)h(µ) λ − µ + λ − µ g(λ) − g(µ) λ − µ h(λ) + g(µ) λ − µ h(λ) − h(µ) . λ − µ Taking into account the ability of passages to the limits as λ − µ → 0 and λ − µ → ∞ we arrive at (gh)[1](λ, µ) = g(λ)h[1](λ, µ) + g[1](λ, µ)h(µ). It remains to apply Theorems 32, 33, and 34. (cid:3) The function βg, h : U × U → C defined by the formula βg, h(λ, µ) =  , λ−µ g(λ)h(µ)−h(λ)g(µ) g′(λ)h(µ) − h′(λ)g(µ), 0, if λ 6= µ, if λ = µ, if λ = ∞ or µ = ∞, is similar to the divided difference. It is generated by two analytic functions g, h : U → C. By analogy with [61, 62, 84], we call the function βg, h the Bezoutian. The Bezoutian is a difference-differential analogue of the Wronskian. For example, the Bezoutian of the functions sin and cos is sinc(λ − µ) = sin(λ−µ) λ−µ . We note that the Bezoutian can be expressed in terms of divided differences: βg,h(λ, µ) = g[1](λ, µ)h(µ) − h[1](λ, µ)g(µ). (In particular, this formula and Proposition 44 imply that βg, h is an analytic function.) Conversely, where u(λ) = 1. g[1](λ, µ) = βg,u(λ, µ), 28 Corollary 50. Let g, h ∈ O(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1) and let h(λ) 6= 0 for λ ∈ ¯σ(R1, (·)) ∪ ¯σ(R2, (·)). Then (cid:0)ϕ1 ⊡ ϕ2(cid:1)(cid:16) g h(cid:17) =(cid:2)(cid:0)ϕ1 ⊠ ϕ2(cid:1)(βg,h)(cid:3)(cid:2)ϕ1(h) ⊠ ϕ2(h)(cid:3)−1 , where (cid:16) g h(cid:17)(λ) = g(λ) h(λ) . Proof. The proof is analogous to that of Corollary 49 and follows from the formula h g hi[1] (λ, µ) = βg,h(λ, µ) h(λ)h(µ) . (cid:3) Proposition 51. Let X be a Banach space. We take for an extended tensor product the space B(X, X) (see Example 3(e)), and we take for R1, (·) and R2, (·) the resolvent R(·) of the same operator A ∈ B(X). Let an operator C commute with at least one value Rµ of the pseudo-resolvent R(·). Then (cid:2)(cid:0)ϕA ⊡ ϕA(cid:1)f(cid:3)C = f′(A)C = Cf′(A). Proof. We note that, by virtue of Theorem 16, C commutes with all values Rλ of the pseudo-resolvent R(·). By the definition and commutativity, we have (cid:2)(cid:0)ϕ ⊡ ϕ(cid:1)f(cid:3)(C) = 1 2πiZΓ λC dλ =(cid:16) 1 2πiZΓ f (λ)R2 λ dλ(cid:17)C. f (λ)RλCRλ dλ = 1 f (λ)R2 2πiZΓ λ /∈ σ(cid:0)R(·)(cid:1). 2πiZΓ λ = −R′λ, R2 f (λ)R′λ dλ(cid:17)C =(cid:16) 1 Passing to the limit in the Hilbert identity (3) we obtain the relation Substituting this identity into the previous equality and integrating by parts we obtain (cid:2)(cid:0)ϕ ⊡ ϕ(cid:1)f(cid:3)(C) = −(cid:16) 1 2πiZΓ f′(λ)Rλ dλ(cid:17)C = ϕ(f′)C. (cid:3) We note that the divided differences f [1](A, B) of the operators A and B are also closely related to the calculation of functions of block triangular matrices [24, 25, 52, 65, 101]. 9. The impulse response In subsequent sections, we discuss some applications. In this Section, the previous results are applied to the representation of the impulse response of a second order differential equation. Here we regard the space B(Y, X) (exam- ple 3(e)) as an extended tensor product. Therefore, for example, the action of the trans- formator ϕ1(g) ⊠ ϕ2(h) on the operator C ∈ B(Y, X) results in the operator ϕ1(g)Cϕ2(h). Let X and Y be Banach spaces and E, F, H ∈ B(Y, X). A function λ 7→ λ2E + λF + H, where λ ∈ C, is called [42, 83, 96] a square pencil. The resolvent set of the pencil is the set ρ(E, F, H) of all λ ∈ C such that the operator λ2E + λF + H is invertible. The spectrum is the compliment σ(E, F, H) = C \ ρ(E, F, H) and the resolvent is the function Rλ = (λ2E + λF + H)−1, λ ∈ ρ(E, F, H). (31) The main sources [96, 125] of square pencils are the second order differential equations of the form E y(t) + F y(t) + Hy(t) = 0, 29 (32) where y : R → Y . In this Section, it is always assumed that the operator E is invertible7. We recall the following proposition. Proposition 52 (see, for example, [93, Theorem 16]). Let the operator E be invertible. Then the solution of the initial value problem E y(t) + F y(t) + Hy(t) = 0, y(0) = y0, y(0) = y1 can be represented in the form where 1 y(t) = T (t)Ey0 + T (t)(Ey1 + F y0), 2πiZΓ 2πiZΓ expt(λ)(λ2E + λF + H)−1 dλ, exp(1) t (λ)(λ2E + λF + H)−1 dλ, T (t) = T (t) = 1 Γ is an oriented envelope of the pencil spectrum σ(E, F, H), and expt(λ) = eλt, exp(1) t (λ) = λeλt. It can be shown that the function T is the impulse response, and T is its derivative. A factorization of the pencil is the representation of its resolvent in the form Rλ = R1, λCR2, λ, (33) where R1, (·) and R2, (·) are pseudo-resolvents acting in X and Y respectively, and C ∈ B(Y, X). It is assumed that ρ(R1, (·)) ∩ ρ(R2, (·)) ⊇ ρ(E, F, H). Proposition 53. Let the operator E be invertible. Then we have C = E in formula (33), and the pseudo-resolvents R1, (·) and R2, (·) are the resolvents of some operators A1 and A2. Proof. By Proposition 23, we have R1, λ = −N1 + R2, λ = −N2 + P1 λ P2 λ + + A1 λ2 + A2 λ2 + A2 1 λ3 + A2 2 λ3 + A3 1 λ4 + . . . , A3 2 λ4 + . . . . Hence, R1, λCR2, λ = N1CN2 − P1CN2 + N1CP2 λ + −A1CN2 + P1CP2 − N1CA2 λ2 + . . . . On the other hand, Therefore, Rλ = E λ2 − E−1F E−1 λ3 + . . . . N1CN2 = 0, P1CN2 + N1CP2 = 0, −A1CN2 + P1CP2 − N1CA2 = E. 7We note that even if E is invertible, the multiplication of the equation (32) by E−1 is not always desirable. For example, the operators E, F, H are often assumed [83, 114] to be self-adjoint, but the multiplication by E−1 may cause to the loss of this property. 30 Multiplying the second equation on the left by A1P1 (keeping in mind the identities P 2 = P , AP = P A = A and NP = P N = 0 from Proposition 23), we arrive at A1CN2 = 0. Similarly, multiplying the second equation on the right by A2, we have Substituting these results into the third equality, we obtain N1CA2 = 0. P1CP2 = E. Because of the invertibility of E, it follows that the projectors P1 and P2 coincide with 1, and C = E. Consequently (by the identity NP = P N = 0, see Proposition 23), we have N1 = 0 and N2 = 0. It follows that limλ→∞ λR1, λ = 1 and limλ→∞ λR2, λ = 1. By Proposition 24(c), these means that the pseudo-resolvents R1, (·) and R2, (·) are the resolvents of the operators A1 and A2. (cid:3) Theorem 54. Let the operator E be invertible, and the square pencil admits factoriza- tion (33). Then the impulse response T and its derivative T can be represented in the form T (t) =(cid:0)ϕ1 ⊡ ϕ2(cid:1)(expt)C =(cid:0)ϕ1 ⊠ ϕ2(cid:1)(exp[1] T (t) =(cid:0)ϕ1 ⊡ ϕ2(cid:1)(exp(1) for λ 6= µ. t )C =(cid:0)ϕ1 ⊠ ϕ2(cid:1)(exp(1) [1] t )C, λ−µ t (λ, µ) = λeλt−µeµt )C, where exp(1) [1] t Proof. The proof follows from Proposition 52 and Theorem 45. Corollary 55. The spectra of the transformators C 7→ T (t) and C 7→ T (t) in the algebra B(cid:0)B(Y, X)(cid:1) are equal to (cid:8) exp[1] (λ, µ) : λ ∈ σ(A), µ ∈ σ(B)(cid:9), t (λ, µ) : λ ∈ σ(A), µ ∈ σ(B)(cid:9), (cid:8) exp(1) [1] (cid:3) t respectively. Proof. The proof follows from Theorems 39 and 54. (cid:3) Remark 1. (a) In article [78], for the approximate calculation of expressions of the type t ) it is suggested to use the following representation (written in other (cid:0)ϕ1 ⊠ ϕ2(cid:1)(exp[1] notations and verified directly) exp[1] t (λ, µ) = (eλt + eµt) 2 t(cid:1) tanh(cid:0) λ−µ λ−µ 2 t , λ 6= µ. 2(cid:1)z tanh(cid:0) z 2 By Theorem 32, (cid:0)ϕ1 ⊠ ϕ2(cid:1)(eλt + eµt) is ϕ1(expt) ⊠ 1 + 1 ⊠ ϕ2(expt). By Corollary 37, 2 t(cid:1) the operator (cid:0)ϕ1 ⊠ ϕ2(cid:1)(cid:16) tanh(cid:0) λ−µ (A ⊠ 1 − 1 ⊠ B)t. The function τ is analytic in the circle z < π. computation, it is suggested to use the Taylor polynomials or rational approximations. (cid:17) is the function τ (z) = of the transformator In [78], for its λ−µ 2 t (b) Formulae exp[1] t (λ, µ) = e (λ+µ)t 2 2 t(cid:1) sinh(cid:0) λ−µ λ−µ 2 t , exp[1] t (λ, µ) = eµt e(λ−µ)t − 1 λ − µ , assuming a similar usage, are suggested in book [65, formulae (10.17)]. 31 (c) We present a formula that enables one to apply similar ideas for the calculation of exp(1) [1] t : exp(1)[1] t (λ, µ) = λeλt − µeµt λeλt − µeλt + µeλt − µeµt = = eλt + µ eλt − eµt λ − µ = (34) λ − µ = eλt + µ exp[1] t (λ, µ). λ − µ Corollary 56. Let E = 1. Then T (t + s) = T1(t)T (s) + T (t)T2(s), where T1(t) = ϕ1(expt) = T2(t) = ϕ2(expt) = 1 2πiZΓ1 2πiZΓ2 1 expt(λ)R1, λ dλ, expt(µ)R2, µ dµ. Proof. This is a special case of Corollary 49. (cid:3) Issues related to factorization are widely discussed in the literature [27, 76, 81, 83, 90, 96, 117, 125]. The factorization of an operator pencils of an arbitrary order is discussed in [50, 50, 56, 58, 72, 87, 96, 97, 128, 129]. Estimates of the norms of operators (cid:0)ϕ1 ⊠ ϕ2(cid:1)(f )C are obtained in [44, 45]; special attention is paid to T (t) and T (t). Estimates of the norm of e(A⊗1+1⊗B)t are given in [12]. 10. The transformator Q and the Sylvester equation It often arises the problem of calculating the transformator Q = (ϕ1 ⊠ ϕ2)w, where w(λ, µ) = 1 . λ − µ As a rule, it is equivalent to solving the Sylvester equation. In this Section, we discuss some properties of the transformator Q. Let X ⊠ Y be an extended tensor product of Banach spaces X and Y , and R1, (·) and R2, (·) be pseudo-resolvents in the algebras B0(X) and B0(Y ) respectively. We assume that the extended singular sets ¯σ(R1, (·)) and ¯σ(R2, (·)) are disjoint. We consider the trans- formator Q defined as Q = (ϕ1 ⊠ ϕ2)w, where8 w(λ, µ) = 1 . λ − µ If necessary, we will use the more detailed notation Qϕ1,ϕ2 or QA,B. Proposition 57. We assume that the extended singular sets ¯σ(R1, (·)) and ¯σ(R2, (·)) of the pseudo-resolvents R1, (·) and R2, (·) are disjoint. Then Q = 1 2(cid:0)ϕ1 ⊡ ϕ2(cid:1)(sgn12), (35) 8The function w is meromorphic with the point of indeterminacy (∞,∞). 32 where the function sgn12 is equal to 1 in a neighborhood of the extended singular set ¯σ(R1, (·)) and is equal to −1 in a neighborhood of the extended singular set ¯σ(R2, (·)). The transformator Q can be represented in the form Q = 1 2πiZΓ R1, λ ⊠ R2, λ dλ, (36) where Γ is an oriented envelope of ¯σ(R1, (·)) with respect to ¯σ(R2, (·)). ✬✩ ⑦ ✫✪ ✬✩ ✫✪ ⑦ Γ1 Γ2 σ1 σ2 ✻ ✻ ⑦⑦⑦ ¯σ1 ✬ ✫ ✬✩ ❄⑦ ✫✪ Γ1 Γ2 σ2 ✻ ✩ ✪ Figure 3. Various options of an arrangement of the contours Γ1 and Γ2 and the extended singular sets Proof. It is easy to verify that sgn[1] 12 = 2w. So, formula (35) follows from Theorem 45. We calculate (ϕ1 ⊠ ϕ2)w. To be definite, we assume that ∞ /∈ ¯σ(R2, (·)). We assume that the oriented envelope Γ1 of the set ¯σ(R1,(·)) and the oriented envelope Γ2 of the set In particular, λ − µ is not equal to zero for σ(R2,(·)) are located as shown in Fig. 3. λ ∈ Γ1 and µ ∈ Γ2. We have (note that in representation (20) for the function w, we have δ1 = δ2 = 0) Q = (ϕ1 ⊠ ϕ2)w = 1 2πiZΓ1 = R1,λ ⊠(cid:16) 1 1 1 (2πi)2ZΓ1ZΓ2 2πiZΓ2 R2,µ λ − µ λ − µ dµ(cid:17) dλ = R1,λ ⊠ R2,µ dµ dλ 1 2πiZΓ1 R1,λ ⊠ R2,λ dλ. Obviously, Γ1 is an oriented envelope of ¯σ(R1,(·)) with respect to σ(R2,(·)). Proposition 58 ([13, 63], [86, Lemma 2.2]). Let A ∈ B(X), B ∈ B(Y ), and the embed- dings σ(A) ⊂ { λ ∈ C : Re λ < ρ} and σ(B) ⊂ { λ ∈ C : Re λ > ρ} hold for some ρ ∈ R. Then (cid:3) Q = −Z ∞ 0 eAt ⊠ e−Bt dt. Proof. We begin with the identity w(λ, µ) = −Z ∞ 0 eλte−µt dt. It is valid for λ ∈ U and µ ∈ V provided the neighborhoods U ⊃ σ(A) and V ⊃ σ(B) are sufficiently small. Moreover, we may assume that the integral converges uniformly for λ ∈ U and µ ∈ V . We substitute this integral into the formula (cid:0)ϕ1 ⊠ ϕ2(cid:1)w = 1 (2πi)2 ZΓ1ZΓ2 33 w(λ, µ)RA, λ ⊠ RB, µ dµ dλ from Theorem 32 assuming that Γ1 ⊂ U and Γ2 ⊂ V . By the uniform convergence of the last integral, we may change the order of integration: (cid:0)ϕ1 ⊠ ϕ2(cid:1)(w) = 1 (2πi)2 ZΓ1ZΓ2(cid:16)−Z ∞ (2πi)2 ZΓ1ZΓ2 = −Z ∞ 0 (cid:16) 1 = −Z ∞ 0 0 eAt ⊠ e−Bt dt. (cid:3) eλte−µt dt(cid:17)RA, λ ⊠ RB, µ dµ dλ eλte−µtRA, λ ⊠ RB, µ dµ dλ(cid:17) dt Proposition 59 ([13, Theorem 9.1]). Let the embeddings ¯σ(R1, (·)) ⊂ { λ ∈ C : λ < ρ} and ¯σ(R2, (·)) ⊂ { λ ∈ C : λ > ρ} hold for some ρ > 0. Then Q = − ∞Xn=0 An ⊠ Rn+1 2, 0 , where the operator A ∈ B(X) generates R1, (·) according to Proposition 24. Proof. We consider the identity w(λ, µ) = − ∞Xn=0 λn µn+1 . 1 (2πi)2ZΓ1ZΓ2 It is valid for λ ∈ U and µ ∈ V , where the neighborhoods U ⊃ σ(A) and V ⊃ ¯σ(R2, (·)) are sufficiently small. Moreover, we may assume that the series converges uniformly for λ ∈ U and µ ∈ V . We substitute this series into the formula (cid:0)ϕ1 ⊠ ϕ2(cid:1)(w) = δ w(λ,∞)R1, λ ⊠ 1Y dλ w(λ, µ)R1, λ ⊠ R2, µ dµ dλ + from Theorem 33 assuming that Γ1 ⊂ U and Γ2 ⊂ V . By the uniform convergence of the series (and by w(λ,∞) = 0), we have (2πi)2 ZΓ1ZΓ2 − ∞Xn=0 (cid:0)ϕ1 ⊠ ϕ2(cid:1)(w) = (2πi)2 ZΓ1ZΓ2 ∞Xn=0 = − λn µn+1 RA, λ ⊠ R2, µ dµ dλ λn µn+1 RA, λ ⊠ R2, µ dµ dλ = − 2πiZΓ1 ∞Xn=0 An ⊠ Rn+1 2, 0 . (cid:3) 1 1 and ϕ2(f ) provided Q(C) is known; it is a version of Theorem 46. Theorem 60 below reduces the calculation of(cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C to the calculation of ϕ1(f ) Theorem 60. Let the extended singular sets ¯σ(R1, (·)) and ¯σ(R2, (·)) of the pseudo-resol- vents R1, (·) and R2, (·) be disjoint, and f ∈ O(cid:0)¯σ(R1, (·)) ∪ ¯σ(R2, (·))(cid:1). Then In the special case, where B(Y, X) is taken as extended tensor product (see example 3(e)), (cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C =(cid:2)ϕ1(f ) ⊠ 1 − 1 ⊠ ϕ2(f )(cid:3)Q(C). (cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C = ϕ1(f ) · Q(C) − Q(C) · ϕ2(f ). Proof. By Theorem 45, (cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C =(cid:2)(ϕ1 ⊠ ϕ2)f [1](cid:3)C =(cid:0)ϕ1 ⊠ ϕ2(cid:1)(cid:0)(f ⊗ u)w(cid:1)C −(cid:0)ϕ1 ⊠ ϕ2(cid:1)(cid:0)(u ⊗ f )w(cid:1)C, 34 where (f ⊗u)(λ, µ) = f (λ), (u⊗f )(λ, µ) = f (µ). From Theorems 32, 33, and 34 it follows that (cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C = (ϕ1(f ) ⊠ 1)(cid:2)(cid:0)ϕ1 ⊠ ϕ2(cid:1)(w)C(cid:3) − (1 ⊠ ϕ2(f ))(cid:2)(cid:0)ϕ1 ⊠ ϕ2(cid:1)(w)C(cid:3) = (ϕ1(f ) ⊠ 1)Q(C) − (1 ⊠ ϕ2(f ))Q(C). (cid:3) In Corollary 61 below, we present a version of Theorem 60. It suggests the reverse order of operations, which enables one to apply the transformator Q only once; namely, at first, ϕ1(f ) and ϕ2(f ) are calculated, and then Q(·) is applied. Corollary 61. Let the extended singular sets ¯σ(R1, (·)) and ¯σ(R2, (·)) be disjoint. Then (cid:2)(ϕ1 ⊡ ϕ2)f(cid:3)C = Q(cid:0)ϕ1(f ) · C − C · ϕ2(f )(cid:1). Proof. It is sufficient to note that the transformators ϕ1(f ) ⊠ 1, 1 ⊠ ϕ2(f ), and Q = (cid:3) (cid:0)ϕ1 ⊠ ϕ2(cid:1)(w) commute, and then apply Theorem 60. Let A ∈ B(X) and B ∈ B(Y ). The equation AZ − ZB = C (37) for the unknown Z ∈ B(Y, X) with the free term C ∈ B(Y, X) is called the (continuous) Sylvester equation [13, 70, 119]. The Sylvester equation is connected with a number of applications [5, 10, 23, 42, 71, 102, 120] and is widely discussed in the literature. Theorem 62. Let A ∈ B(X) and B ∈ B(Y ). The Sylvester equation (37) has a unique solution Z ∈ B(Y, X) for any C ∈ B(Y, X) if and only if the spectra of the operators A and B are disjoint. This solution coincides with the operator Q(C). Proof. By Theorem 32 and Corollary 35, the transformator Z 7→ AZ − ZB is equal to (cid:0)ϕ1 ⊠ ϕ2(cid:1)f , where f (λ, µ) = λ− µ. By Theorem 39, its spectrum is equal to σ(A)− σ(B). Therefore the transformator is invertible if and only if 0 /∈ σ(A) − σ(B). By Theorem 32, the inverse transformator is (cid:0)ϕ1 ⊠ ϕ2(cid:1)w, where w(λ, µ) = 1 λ−µ. (cid:3) The equation Z − AZB = C (38) Its theory is is called the (discrete) Sylvester equation [70] or the Stein equation [85]. similar to the theory of equation (37). Theorem 63. Let A ∈ B(X) and B ∈ B(Y ). The Sylvester equation (38) has a unique solution Z ∈ B(Y, X) for any C ∈ B(Y, X) if and only if the product of the spectra of the operators A and B does not contain 1. This solution coincides with the operator (cid:2)(ϕ1 ⊠ ϕ2)(s)(cid:3)(C), where 1 s(λ, µ) = . 1 − λµ Proof. By Theorem 32 and Corollary 35, the transformator Z 7→ Z − AZB is equal to (cid:0)ϕ1⊠ϕ2(cid:1)f , where f (λ, µ) = 1−λµ. By Theorem 39, its spectrum is equal to 1−σ(A)σ(B). Therefore the transformator is invertible if and only if 1 /∈ σ(A)σ(B). By Theorem 32, the inverse transformator is (cid:0)ϕ1 ⊠ ϕ2(cid:1)s. Remark 2. Let us return to equation (37) and discuss the case where A and B are un- bounded operators or linear relations. The natural hypothesis is as follows: if the extended singular sets of the resolvents of A and B are disjoint (and thus A or B is a bounded oper- ator), then equation (37) has a unique solution, which is determined by the transformator Q. The problem is: How one can interpret equation (37)? We discuss some variants. (cid:3) 35 First, we assume that A and B are linear relations with non-empty resolvent sets, and the extended spectra of A and B are disjoint. We assume that C ∈ B(Y, X) and a solution Z ∈ B(Y, X) of equation (37) is of interest. To begin with, we show that without loss of generality one can assume that the inverse operators of A and B are everywhere defined bounded operators. Since the extended spectra of A and B are closed and disjoint, there exists ν /∈ ¯σ(A)∪ ¯σ(B). We rewrite (37) in the form and then in the form (with the invertible coefficients ν1 − A and ν1 − B) −νZ + AZ + νZ − ZB = C, −(ν1 − A)Z + Z(ν1 − B) = C. See [57, Proposition A.1.1, p. 281] or [92, Theorem 36] for a justification of the last equality in the case of linear relations. We consider the case where ∞ ∈ ¯σ(A). Since the relation A is invertible, its range coincides with the whole of X, and the image of the zero is zero (otherwise the left side of equation (37) is not an operator). So, A is an operator (not a relation). We call an operator Z ∈ B(Y, X), whose range is contained in the domain of the operator A (otherwise the domains of the left and the right sides of equation (37) are different), a solution of equation (37) provided it satisfies the equation. Since ∞ /∈ ¯σ(B), B is a bounded linear operator, see Proposition 24. Multiplying (37) by A−1 we obtain Z − A−1ZB = A−1C. (39) By [57, Proposition A.3.1], σ(A−1) = { 1 λ : λ ∈ ¯σ(A) }. By Theorem 63, equation (39) has a unique solution Z for an arbitrary A−1C if 0 /∈ { 1− λµ : λ ∈ σ(A−1), µ ∈ σ(B) } = { 1 − µ λ : λ ∈ ¯σ(A), µ ∈ σ(B) }, which is the case, because the extended spectra of A and B are disjoint. We multiply (39) on the left by A (taking into account that AA−1 = 1). As a result we arrive at the original equation (37). Therefore Z is a solution of equation (37) as well. We discuss the case where ∞ ∈ ¯σ(B). We assume that the domain of the relation B coincides with the whole of Y (otherwise the domains of the left and right sides of equation (37) are different). Since the relation B is invertible, its range coincides with the whole of Y , the kernel is zero, but the image of the zero Im0 B = { x : (0, x) ∈ B } may consist not only of zero. We call an operator Z ∈ B(Y, X), whose kernel contains Im0 B (otherwise the left side of equation (37) is not an operator), a solution of equation (37) provided it satisfies the equation. Multiplying (37) on the right by B−1 we obtain According to [92, Theorem 16] we rewrite this equation in the form AZB−1 − ZBB−1 = CB−1. AZB−1 − Z1Y :Im0 B = CB−1, where 1Y :Im0 B = { (y1, y2) ∈ Y × Y : (0, y1 − y2) ∈ B }. Since the kernel of the operator Z contains Im0 B, the last equation can be rewritten as AZB−1 − Z = CB−1. By Theorem 63, this equation has a unique solution Z for an arbitrary CB−1 if 0 /∈ { 1 − λµ : λ ∈ σ(A), µ ∈ σ(B−1) } = { 1 − λ µ : λ ∈ σ(A), µ ∈ ¯σ(B) }, which is the case, because the extended spectra of A and B are disjoint. Multiplying the last equation on the right by B we obtain AZB−1B − ZB = CB−1B, 36 for all y ∈ Dom B. Let us look for a solution of equation (37) in the form Z = V B−1, where V ∈ B(Y, X) is a new unknown operator. Substituting Z = V B−1 into equation (37) we obtain (40) or AZy − ZBy = Cy AV B−1 − V B−1B = C, AV B−1 − V 1Dom B = C, or (according to [92, Theorem 16] and Ker B = 0) AZ − ZB = C. So, Z is a solution of original equation (37) as well. We consider another case: let B be an invertible unbounded operator with the dense domain Dom B (in particular, ∞ ∈ ¯σ(B)). We call an operator Z ∈ B(Y, X) a solution if where 1Dom B is the identity operator with the domain Dom B. By our definition of a solution, the last equation is equivalent to the equation AV B−1 − V = C. Obviously, it has a unique solution V . Returning to equivalent equation (40) we see that the operator Z = V B−1 is a solution of the original equation. Theorem 62 for matrices was first proved in [122]. An independent proof of its sufficient part for the case of operators was obtained in [21, 82, 109]. For a Hilbert space, a necessary and sufficient condition for the solvability of the Sylvester equation was first obtained in [26], see also [49, c. 54]. An analogue of Theorem 63 for matrices is proved, for example, in [85]. The representation for the solution of the Sylvester equation in the form of contour integral (36) was first published in [109], see also Example 4. Estimates of the solution of the Sylvester equation are given in [44, 46, 47]. The Sylvester equation (37) with unbounded operator coefficients A and B is considered in [2, 3, 39, 79, 94, 104, 116]. A generalization of the transformator Q (Q corresponds to the function w(λ, µ) = 1 is the inverse of the transformator v[1] It is discussed in [15, 41, 46]. n+1(A, B), where v[1] λ−µ ) n (λ, µ) = λn−1 +λn−2µ+· · ·+µn−1. 11. The differential of the functional calculus Let X be a Banach space. The (Fr´echet) differential of a nonlinear transformator f : D(f ) ⊆ B(X) → B(X) at a point A ∈ B(X) is defined to be a linear transformator df (·, A) : B(X) → B(X) depending on the parameter A that possesses the property f (A + ∆A) = f (A) + df (∆A, A) + o(k∆Ak). (41) We assume that a neighborhood of A is contained in the domain D(f ) of the transforma- tor f . We recall standard properties of the differential. Proposition 64 ([4, § 2.2.2], [32, 8.2.1]). Let a transformator g : B(X) → B(X) be differentiable at a point A ∈ B(X) and a transformator f : B(X) → B(X) be differentiable at the point g(A) ∈ B(X). Then the composition f ◦ g is differentiable at the point A, and d(f ◦ g)(·, A) = df(cid:2)dg(·, A), g(A)(cid:3). 37 Corollary 65 ([4, § 2.3.4], [32, 8.2.3]). Let a transformator f : B(X) → B(X) be contin- uously differentiable (i. e. df (·, A) depends on A continuously in norm) in a neighborhood of a point A ∈ B(X) and let the transformator df (·, A) be invertible. Then the inverse transformator of f is defined and differentiable in a neighborhood of the point B = f (A), and the differential of the inverse transformator is equal to the inverse of the original differential: f (λ)(λ1 − A)−1 dλ (42) f (λ)(λ1 − A)−1 dλ + df (∆A, A) + o(k∆Ak). = Rλ(1 − ∆A · Rλ)−1 = (1 − Rλ · ∆A)−1Rλ. df−1(·, B) =(cid:2)df (·, A)(cid:3)−1 Let A ∈ B(X) and f ∈ O(cid:0)σ(A)(cid:1). For the transformator . definition (41) of a differential looks as follows: 1 A 7→ f (A) = 2πiZΓ 2πiZΓ =(cid:0)(λ1 − A) − ∆A(cid:1)−1 1 f (λ)(cid:0)λ1 − (A + ∆A)(cid:1)−1 dλ = 1 2πiZΓ (cid:0)λ1 − (A + ∆A)(cid:1)−1 We note that Based on this formula, we adopt the following definition. Let R(·) be a pseudo-resolvent in the algebra B(X). We call the perturbation of R(·) by an operator ∆A ∈ B(X) the function Tλ = Rλ(1 − ∆A · Rλ)−1 = (1 − Rλ · ∆A)−1Rλ. (43) Remark 3. We note shortly an additional reasoning in favor of definition (43). Let R(·) be the resolvent of a linear relation A, i. e. Rλ = (λ1 − A)−1. We show that (1 − Rλ · ∆A)−1Rλ = (λ1 − A − ∆A)−1. =(cid:2)(λ1− A)(1− Rλ · ∆A)(cid:3)−1 =(cid:2)λ1 − A − (λ1 − A)(λ1 − A)−1 · ∆A(cid:3)−1 Obviously (for details, see [57, Proposition A.1.1] or [92, Proposition 12]), (1 − Rλ · λ (1− Rλ · ∆A)(cid:3)−1 ∆A)−1Rλ =(cid:2)R−1 . Further, since the image of the operator Rλ · ∆A is contained in the image of Rλ, which is equal to the domain of the relation A, by virtue of [92, Theorem 36(a)], we can develop the internal parentheses: (cid:2)λ1 − A − (λ1 − A)Rλ · ∆A)(cid:3)−1 . We note that (λ1−A)(λ1−A)−1 is equal to the relation 1X:Im0 A = { (x1, x2) ∈ X×X : (0, x1−x2) ∈ A}. Obviously, (λ1 − A − 1X:Im0 A · ∆A)−1 = (λ1 − A − ∆A)−1. Proposition 66. For any perturbation ∆A ∈ B(X) the function Tλ = Rλ(1 − ∆A · Rλ)−1 = (1 − Rλ · ∆A)−1Rλ is a pseudo-resolvent. Proof. We verify the Hilbert identity for all λ and µ such that Tλ and Tµ are defined. We have Tλ − Tµ + (λ − µ)TλTµ = (1 − Rλ · ∆A)−1Rλ − Rµ(1 − ∆A · Rµ)−1 + (λ − µ)(1 − Rλ · ∆A)−1RλRµ(1 − ∆A · Rµ)−1 = (1 − Rλ · ∆A)−1(cid:2)Rλ(1 − ∆A · Rµ) − (1 − Rλ · ∆A)Rµ + (λ − µ)RλRµ(cid:3) = (1 − Rλ · ∆A)−1(cid:2)Rλ − Rµ + (λ − µ)RλRµ(cid:3) = 0. (cid:3) 38 We define the differential df (·, R(·)) of the mapping (which is a generalization of (42)) (44) R(·) 7→ f (R(·)) = f (λ)Rλ dλ + δf (∞)1, 1 2πiZΓ by means of the formula 1 2πiZΓ f (λ)Rλ(1 − ∆A · Rλ)−1 dλ + δf (∞)1 1 2πiZΓ f (λ)Rλ dλ + δf (∞)1 + df (∆A, R(·)) + o(k∆Ak). Theorem 67. Let R(·) be a pseudo-resolvent in the algebra B(X), and f ∈ O(cid:0)¯σ(R(·))(cid:1). Then the differential of mapping (44) admits the representation = In other words, df (∆A, R(·)) = f (λ)Rλ ∆A Rλ dλ. 1 2πiZΓ df (·, R(·)) =(cid:0)ϕ ⊡ ϕ(cid:1)(f ), where ϕ is the functional calculus generated by the pseudo-resolvent R(·). Proof. We assume that By Theorem 1, we have k∆Ak · kRλk < 1. f (λ)Rλ ∆A Rλ dλ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Therefore, 1 2πiZΓ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:2)Rλ(1 − ∆A · Rλ)−1 − Rλ(cid:3) − Rλ∆A · Rλ(cid:13)(cid:13) ≤ kRλk3 · k∆Ak2 f (λ)Rλ(1 − ∆A · Rλ)−1 dλ − 1 − kRλk · k∆Ak 2πiZΓ f (λ)Rλ dλ − 2πiZΓ 1 1 . 1 =(cid:13)(cid:13)(cid:13) 2πiZΓ f (λ)(cid:2)Rλ(1 − ∆A · Rλ)−1 − Rλ − Rλ ∆A Rλ(cid:3) dλ(cid:13)(cid:13)(cid:13) 2πZΓ f (λ) kRλk3 · k∆Ak2 1 − kRλk · k∆Ak dλ = o(k∆Ak). (cid:3) ≤ Proposition 68 ([14, Theorem 2.1]). Let A ∈ B(X) and f ∈ O(cid:0)σ(A)(cid:1). Then 1 df (A ∆A − ∆A A, A) = ϕA(f )∆A − ∆AϕA(f ), where the functional calculus ϕA is generated by the operator A. Proof. The proof follows from Theorems 46 and 67. Proposition 69 ([65, Theorem 3.3]). Let g, h ∈ O(cid:0)¯σ(R(·))(cid:1). Then d(gh)(∆A, R(·)) = dg(∆A, R(·)) h(R(·)) + g(R(·)) dh(∆A, R(·)), where (gh)(λ) = g(λ)h(λ). Proof. The proof follows from Corollary 49. Corollary 70 ([110, Theorem 10.36], see also [108]). Let f ∈ O(cid:0)¯σ(R(·))(cid:1). We assume that ∆A commutes with at least one value Rµ of the pseudo-resolvent R(·). Then df (∆A, R(·)) = ϕ(f′) ∆A = ∆A ϕ(f′). 39 (cid:3) (cid:3) Proof. The proof follows from 51. Theorem 71. Let R(·) be a pseudo-resolvent in the algebra B(X) and f ∈ O(cid:0)¯σ(R(·))(cid:1). Then the differential of mapping (44) admits the representation (cid:3) df (∆A, R(·)) = 1 (2πi)2 ZΓ1ZΓ2 f [1](λ, µ)Rλ ∆A Rµ dµ dλ, (45) where the divided difference f [1] is defined by formula (28), the contours Γ1 and Γ2 are oriented envelopes of the extended singular set ¯σ(R(·)) with respect to the complement C \ U, and U is the domain of the function f . The spectrum of the transformator df (·, R(·)) : B(X) → B(X) is given by the formula (46) Proof. Theorem 67 shows that df (∆A, R(·)) is the operator (cid:2)(cid:0)ϕ ⊡ ϕ(cid:1)(f )(cid:3)∆A. By Theo- σ(cid:2)df (·, R(·))(cid:3) =(cid:8) f [1](λ, µ) : λ, µ ∈ ¯σ(R(·))(cid:9). (cid:2)(cid:0)ϕ ⊡ ϕ(cid:1)(f )(cid:3)∆A =(cid:2)(cid:0)ϕ ⊠ ϕ(cid:1)f [1](cid:3)∆A. It remains to apply Theorem 32. rem 45, Formula (46) follows from Theorem 39. (cid:3) Example 6. Let A ∈ B(X). The following corollaries are consequences of Example 5 and Theorem 71. The differential of the transformator v2(A) = A2 is given by the formula and its spectrum at a point A ∈ B(X) is equal to dv2(∆A, A) = A · ∆A + ∆A · A, σ(cid:2)dv2(·, A)(cid:3) =(cid:8) λ + µ : λ, µ ∈ σ(A)(cid:9). dr1(∆A, R(·)) = Rλ0 · ∆A · Rλ0, The differential of the mapping r1(R(·)) = Rλ0 is given by the formula (47) : λ, µ ∈ ¯σ(R(·))o. and its spectrum at a point R(·) is equal to σ(cid:2)dr1(·, R(·))(cid:3) =n The differential of the transformator expt(A) = eAt is given by the formula An−i∆A Ai, d expt(∆A, A) = (n + 1)! and its spectrum at a point A ∈ B(X) is equal to (λ0 − λ)(λ0 − µ) nXi=0 ∞Xn=0 t (λ, µ) : λ, µ ∈ σ(A)(cid:9). σ(cid:2)d expt(·, A)(cid:3) =(cid:8) exp[1] ∞Xn=0 nXi=0 tn n! and its spectrum at a point A ∈ B(X) is equal to (λ, µ) : λ, µ ∈ σ(A)(cid:9). t (·, A)(cid:3) =(cid:8) exp(1) [1] The differential of the transformator exp(1) σ(cid:2)d exp(1) t (∆A, A) = An−i∆A Ai. d exp(1) t (A) = A eAt is given by the formula 1 tn t 40 We note special formulae for the differentials of the transformators expt(A) = eAt and t (A) = AeAt. exp(1) Proposition 72. The differentials of the transformators expt(A) = eAt and expt(A)(1) = AeAt at a point A ∈ B(X) can be calculated by means of the formulae d exp(1) e(t−s)A∆AesA ds =Z t 0 0 d expt(∆A, A) =Z t t (∆A, A) = expt(A)∆A +Z t λ − µZ t eλseµ(t−s)(cid:12)(cid:12)(cid:12) λ − µ s=0 = = s=t 1 1 0 eλt − eµt λ − µ Proof. For λ 6= µ we have expt−s(A)∆A exps(A) ds, expt−s(A)∆A A exps(A) ds. (48) (49) d dsheλseµ(t−s)i ds =Z t 0 eλseµ(t−s) ds. 0 f [1](λ, µ)Rλ ∆A Rµ dµ dλ, By the continuity, the same representation of exp[1] t (λ, µ) holds for all finite λ and µ. Hence, from Theorems 71, 32, and 39 it follows (48). Formula (49) follows from (34). (cid:3) The differentials of inverse functions are defined by the inverse transformators, see Corollary 65. To calculate them and their spectra, one can use the following theorem. Theorem 73. Let R(·) be a resolvent of an operator A ∈ B(X), f ∈ O(cid:0)σ(A)(cid:1), and 0 /∈S(cid:8) f [1](λ, µ) : λ, µ ∈ σ(A)(cid:9). Then the differential of the transformator B 7→ f−1(B) at the point B = f (A) is given by the formula df−1(∆B, B) = 1 (2πi)2ZΓ1ZΓ2 1 f [1](λ, µ) Rλ ∆B Rµ dµ dλ, (50) where the contours Γ1 and Γ2 are oriented envelopes of the spectrum σ(A) with respect to the point ∞ and the complement of the domain of the function f . : λ, µ ∈ σ(A)o. The spectrum of the transformator df−1(·, B) : B(X) → B(X) is given by the formula (51) f [1](λ, µ) 1 Proof. By Theorem 71, σ(cid:2)df−1(·, B)(cid:3) =[n (2πi)2 ZΓ1ZΓ2 1 df (∆A, A) = σ(cid:2)df (·, A)(cid:3) =(cid:8) f [1](λ, µ) : λ, µ ∈ σ(A)(cid:9). Since 0 /∈ S(cid:8) f [1](λ, µ) : λ, µ ∈ σ(A)(cid:9), the transformator df (·, A) is invertible. By Corollary 65, the inverse transformator is the differential of the mapping B 7→ f−1(B), which is the inverse mapping of A 7→ f (A). By Theorem 32, the inverse transformator is given by formula (50). By Theorem 39, its spectrum is given by formula (51). (cid:3) Example 7. Let the spectrum of an operator B ∈ B(X) be contained in C\ (−∞, 0]. The following corollaries are consequences of Example 6 and Theorem 73. From (47) it is clear that the differential dv1/2(∆B, B) of the transformator v1/2(B) = √B satisfies the Sylvester equation Therefore, √B · dv1/2(∆B, B) + dv1/2(∆B, B)√B = ∆B. dv1/2(∆B, B) = Q√B, −√B(∆B), 41 where the transformator Q√B, −√B is constructed by means of the functional calculus generated by the operators √B and −√B. The spectrum of the differential of the trans- formator v1/2(B) = √B at the point B is equal to σ(cid:2)dv1/2(·, B)(cid:3) =n 1 λ + µ : λ, µ ∈pσ(B)o. The spectrum of the differential of the transformator f : B 7→ ln B, where ln denotes the principal value, at the point B is equal to σ(cid:2)df (·, B)(cid:3) =n 1 exp[1] 1 (λ, µ) : λ, µ ∈ ln(cid:0)σ(B)(cid:1)o. Remark 4. The equation (52) with A, B, C, D ∈ B(X) and unknown Z ∈ B(X) is called [70, 79, 80] the Riccati equation. It arises in control theory [5, 103]. The differential dZ = dZ(∆A, ∆B, ∆C, ∆D; Z) of the solution Z of Riccati equation (52) satisfies [70, p. 135] the continuous Sylvester equation AZ + ZB + ZCZ + D = 0 (A + ZC)dZ + dZ(B + CZ) = −∆D − ∆A Z − Z ∆B − Z ∆C Z. So, dZ(∆A, ∆B, ∆C, ∆D; Z) = QA+ZC, −B−CZ(−∆D − ∆A Z − Z ∆B − Z ∆C Z). For pseudo-resolvents generated by bounded operators, Theorem 67 is proved in [110, Theorem 10.38], see also [14, formula (2.3)], [22] and [121]. Representation (45) for ma- trices is given in [85, Theorem 5.1]. Formula (46) for matrices is proved in [65, Theorem 3.9], its weaker version was previously obtained in [77, Lemma 2.1]. Formula (48) was first obtained in [74, formula (1.8)], see also [10, ch. 10, § 14], [65, formula (10.15)], [77, example 2], [98], [100], [126]. Differentials connected with some specific functions f are investigated in [1, 29, 65, 66, 77, 78, 126, 130]; a special attention is paid to estimates of their norms which helps to find the condition number of the transformator A 7→ f (A). Properties of differentials of higher orders are investigated in [14]. References [1] Al-Mohy A. H., Higham N. J., Relton S. D. Computing the Fr´echet de- rivative of the matrix logarithm and estimating the condition number // SIAM J. Sci. Comput. -- -- 2013. -- -- Vol. 35, no. 4. -- -- P. 394 -- 410. -- -- URL: http://dx.doi.org/10.1137/120885991. [2] Albeverio S., Makarov K. A., Motovilov A. K. Graph subspaces and the spectral shift function // Canad. J. Math. -- -- 2003. -- -- Vol. 55, no. 3. -- -- P. 449 -- 503. -- -- URL: http://dx.doi.org/10.4153/CJM-2003-020-7. [3] Albeverio S., Motovilov A. K. Operator Stieltjes integrals with respect to a spectral measure and solutions of some operator equations // Trans. Moscow Math. Soc. -- -- 2011. -- -- Vol. 72, no. 1. -- -- P. 45 -- 77. [4] Alekseev V. M., Tikhomirov V. M., Fomin S. V. Optimal control. Contempo- rary Soviet Mathematics. -- -- Consultants Bureau, New York, 1987. -- -- xiv+309 p. -- -- ISBN: 0-306-10996-4. -- -- URL: http://dx.doi.org/10.1007/978-1-4615-7551-1. [5] Antoulas A. C. Approximation of large-scale dynamical systems. -- -- Philadelphia, PA : Society for Industrial and Applied Mathematics (SIAM), 2005. -- -- Vol. 6 of Advances in Design and Control. -- -- xxvi+479 p. -- -- ISBN: 0-89871-529-6. -- -- URL: http://dx.doi.org/10.1137/1.9780898718713. 42 [6] Arendt W. Approximation of degenerate semigroups // Taiwanese J. Math. -- -- 2001. -- -- Vol. 5, no. 2. -- -- P. 279 -- 295. [7] Arens R. The adjoint of a bilinear operation // Proc. Amer. Math. Soc. -- -- 1951. -- -- Vol. 2. -- -- P. 839 -- 848. [8] Arens R. Operational calculus of linear relations // Pacific J. Math. -- -- 1961. -- -- Vol. 11. -- -- P. 9 -- 23. [9] Baskakov A. G. Representation theory for Banach algebras, Abelian groups, and semigroups in the spectral analysis of linear operators // Journal of Mathematical Sciences. -- -- 2006. -- -- Vol. 137, no. 4. -- -- P. 4885 -- 5036. [10] Bellman R. Introduction to matrix analysis. -- -- Second edition. -- -- New York- Dusseldorf-London : McGraw-Hill Book Co., 1970. -- -- xxiii+403 p. [11] Benzi M., Simoncini V. Approximation of functions of large matrices with Kronecker structure // arXiv:1503.02615. -- -- 2015. -- -- P. 1 -- 21. [12] Benzi M., Simoncini V. Decay bounds for functions of matrices with banded or Kronecker structure // arXiv:1501.07376. -- -- 2015. -- -- P. 1 -- 20. [13] Bhatia R., Rosenthal P. How and why to solve the operator equation AX − XB = Y // Bull. London Math. Soc. -- -- 1997. -- -- Vol. 29, no. 1. -- -- P. 1 -- 21. -- -- URL: http://dx.doi.org/10.1112/S0024609396001828. [14] Bhatia R., Sinha K. B. Derivations, derivatives and chain rules // Linear Algebra Appl. -- -- 1999. -- -- Vol. 302/303. -- -- P. 231 -- 244. [15] Bhatia R., Uchiyama M. The Y // Expo. Math. -- -- http://dx.doi.org/10.1016/j.exmath.2009.02.001. = 2009. -- -- Vol. 27, no. 3. -- -- P. 251 -- 255. -- -- URL: i=0 An−iXBi operator equation Pn [16] Bichegkuev M. S. Conditions for solubility of difference inclusions // Izv. Math. -- -- 2008. -- -- Vol. 72, no. 4. -- -- P. 647 -- 658. [17] Bourbaki N. ´El´ements de math´ematique. Fasc. X. Premi`ere partie. Livre III: Topolo- gie g´en´erale. Chapitre 10: Espaces fonctionnels. Deuxi`eme ´edition, enti`erement refondue. Actualit´es Scientifiques et Industrielles, No. 1084. -- -- Paris : Hermann, 1961. -- -- 96 p. [18] Bourbaki N. ´El´ements de math´ematique. Fasc. XXXII. Th´eories spectrales. Chapitre I: Alg`ebres norm´ees. Chapitre II: Groupes localement compacts commutatifs. Actu- alit´es Scientifiques et Industrielles, No. 1332. -- -- Paris : Hermann, 1967. -- -- iv+166 p. [19] Brown A., Pearcy C. Spectra of tensor products of operators // Proc. Amer. Math. Soc. -- -- 1966. -- -- Vol. 17. -- -- P. 162 -- 166. [20] Cross R. Multivalued linear operators. -- -- New York : Marcel Dekker, Inc., 1998. -- -- Vol. 213 of Monographs and Textbooks in Pure and Applied Mathematics. -- -- x+335 p. -- -- ISBN: 0-8247-0219-0. [21] Daleckiı Y. L. On the asymptotic solution of a vector differential equation // Dok- lady Akad. Nauk SSSR (New Series). -- -- 1953. -- -- Vol. 92. -- -- P. 881 -- 884 (in Russian). [22] Daleckiı Y. L. Differentiation of non-Hermitian matrix functions depending on a parameter // Amer. Math. Soc. Transl., Series 2. -- -- 1965. -- -- Vol. 47. -- -- P. 73 -- 87. [23] Daleckiı Y. L., Kreın M. G. Stability of solutions of differential equations in Ba- nach space. -- -- Providence, RI : American Mathematical Society, 1974. -- -- Vol. 43 of Translations of Mathematical Monographs. -- -- vi+386 p. [24] Davies P. I., Higham N. J. A Schur-Parlett algorithm for computing matrix func- tions // SIAM J. Matrix Anal. Appl. -- -- 2003. -- -- Vol. 25, no. 2. -- -- P. 464 -- 485 (elec- tronic). -- -- URL: http://dx.doi.org/10.1137/S0895479802410815. 43 [25] Davis C. Explicit functional calculus // Linear Algebra and Appl. -- -- 1973. -- -- Vol. 6. -- -- P. 193 -- 199. [26] Davis C., Rosenthal P. Solving linear operator equations // Canad. J. Math. -- -- 1974. -- -- Vol. 26. -- -- P. 1384 -- 1389. [27] Davis G. J. Numerical equation // SIAM J. Sci. Statist. Comput. -- -- 1981. -- -- Vol. 2, no. 2. -- -- P. 164 -- 175. -- -- URL: http://dx.doi.org/10.1137/0902014. quadratic matrix solution of a [28] Defant A., Floret K. Tensor norms and operator ideals. -- -- Amsterdam : North- Holland Publishing Co., 1993. -- -- Vol. 176 of North-Holland Mathematics Studies. -- -- xii+566 p. -- -- ISBN: 0-444-89091-2. [29] Dieci L., Morini B., Papini A. Computational techniques for real logarithms of matrices // SIAM J. Matrix Anal. Appl. -- -- 1996. -- -- Vol. 17, no. 3. -- -- P. 570 -- 593. -- -- URL: http://dx.doi.org/10.1137/S0895479894273614. [30] Dieci L., Papini A. Pad´e approximation for the exponential of a block triangular matrix // Linear Algebra Appl. -- -- 2000. -- -- Vol. 308, no. 1-3. -- -- P. 183 -- 202. -- -- URL: http://dx.doi.org/10.1016/S0024-3795(00)00042-2. [31] Diestel J., Fourie J. H., Swart J. The metric theory of tensor products. -- -- x+278 p. -- -- URL: : American Mathematical Society, Providence, RI ISBN: http://dx.doi.org/10.1090/mbk/052. 978-0-8218-4440-3. -- -- Grothendieck's revisited. 2008. -- -- r´esum´e [32] Dieudonn´e J. Foundations of modern analysis. Pure and Applied Mathematics, Vol. X. -- -- New York -- London : Academic Press, 1960. -- -- xiv+361 p. [33] Dunford N. An ergodic theorem for n-parameter groups // Proceedings of the Na- tional Academy of Sciences. -- -- 1939. -- -- Vol. 25, no. 4. -- -- P. 195 -- 196. [34] Dunford N. Spectral theory // Bull. Amer. Math. Soc. -- -- 1943. -- -- Vol. 49. -- -- P. 637 -- 651. [35] Dunford N. Spectral theory. I. Convergence to projections // Trans. Amer. Math. Soc. -- -- 1943. -- -- Vol. 54. -- -- P. 185 -- 217. [36] Dunford N., Schwartz J. T. Linear operators. Part I. General theory. Wiley Classics Library. -- -- New York : John Wiley & Sons, Inc., 1988. -- -- xiv+858 p. [37] Embry M. R., Rosenblum M. Spectra, tensor products, and linear operator equa- tions // Pacific J. Math. -- -- 1974. -- -- Vol. 53. -- -- P. 95 -- 107. [38] Favini A., Yagi A. Degenerate differential equations in Banach spaces. -- -- New York -- Basel -- Hong Kong : CRC Press, 1998. -- -- xii+314 p. -- -- ISBN: 0-8247-1677-9. XT = A. // J. Math. Mech. -- -- 1969/1970. -- -- Vol. 19. -- -- P. 819 -- 828. [39] Freeman J. M. The tensor product of semigroups and the operator equation SX − [40] Frommer A., Simoncini V. Matrix functions // Model order reduction: theory, re- search aspects and applications. -- -- Berlin : Springer, 2008. -- -- Vol. 13 of Math. Ind. -- -- P. 275 -- 303. -- -- URL: http://dx.doi.org/10.1007/978-3-540-78841-6_13. of the [41] Furuta T. Positive equation j=1 An−jXAj−1 = B // Linear Algebra Appl. -- -- 2010. -- -- Vol. 432, no. 4. -- -- semidefinite solutions operator P. 949 -- 955. -- -- URL: http://dx.doi.org/10.1016/j.laa.2009.10.008. Pn [42] Gantmacher F. R. The theory of matrices. Vol. 1. -- -- Providence, RI : AMS Chelsea Publishing, 1998. -- -- x+374 p. -- -- ISBN: 0-8218-1376-5. [43] Gel′fond A. O. Calculus of finite differences. International Monographs on Advanced Mathematics and Physics. -- -- Delhi : Hindustan Publishing Corp., 1971. -- -- vi+451 p. two non-commuting matrices // P. 504 -- 512. -- -- URL: for Electron. J. Linear Algebra. -- -- http://dx.doi.org/10.13001/1081-3810.1453. functions of 2011. -- -- Vol. 22. -- -- [44] Gil' M. Norm estimates 44 [45] Gil' M. Norm estimates for functions of two non-commuting operators // Rocky Mountain J. Math. -- -- 2015. -- -- Vol. 45, no. 3. -- -- P. 927 -- 940. -- -- URL: http://dx.doi.org/10.1216/RMJ-2015-45-3-927. [46] Gil' M. Norm estimates for tions // J. Math. -- -- http://dx.doi.org/10.1155/2015/524829. 2015. -- -- solutions equa- no. Article ID 524829. -- -- P. 7. -- -- URL: polynomial operator of [47] Gil' M. Resolvents of operators on tensor products of Euclidean spaces // Linear and Multilinear Algebra. -- -- 2015. -- -- P. 1 -- 18. [48] Gil' M. I. Regular functions of operators on tensor products of Hilbert spaces // Integral Equations Operator Theory. -- -- 2006. -- -- Vol. 54, no. 3. -- -- P. 317 -- 331. -- -- URL: http://dx.doi.org/10.1007/s00020-004-1359-8. [49] Gohberg I., Goldberg S., Kaashoek M. A. Classes of linear operators. I. -- -- Birkhauser Verlag, 1990. -- -- Vol. 49 of Operator Theory: Ad- ISBN: 3-7643-2531-3. -- -- URL: Basel vances and Applications. -- -- http://dx.doi.org/10.1007/978-3-0348-7509-7. xiv+468 p. -- -- : [50] Gohberg I., Lancaster P., Rodman L. Matrix polynomials. Computer Science and Applied Mathematics. -- -- New York -- London : Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], 1982. -- -- xiv+409 p. -- -- ISBN: 0-12-287160-X. in Hilbert [51] Gohberg I. C., Kreın M. G. Theory and applications of Volterra op- : American Mathemati- of Mathematical Mono- URL: erators cal Society, graphs. -- -- http://gen.lib.rus.ec/book/index.php?md5=b3b7a4a44343a7b57a1f0a0aec37ce34. Providence, RI of Translations 0821815741,9780821815748. -- -- 1970. -- -- x+430 24 ISBN: space. -- -- p. -- -- Vol. [52] Golub G. H., Van Loan C. F. Matrix computations. Johns Hopkins Studies in the Mathematical Sciences. -- -- Third edition. -- -- Baltimore, MD : Johns Hopkins Univer- sity Press, 1996. -- -- xxx+698 p. -- -- ISBN: 0-8018-5413-X; 0-8018-5414-8. [53] Graham A. Kronecker products and matrix calculus: with applications. Ellis Hor- wood Series in Mathematics and its Applications. -- -- New York : Ellis Horwood Ltd., Chichester; Halsted Press [John Wiley & Sons, Inc.], 1981. -- -- 130 p. -- -- ISBN: 0-85312-391-8. [54] Greene R., Krantz S. Function theory of one complex variable. -- -- Third edition. -- -- Providence, RI : Amer. Math. Soc., 2006. -- -- Vol. 40 of Graduate Studies in Mathe- matics. -- -- x+504 p. -- -- URL: http://dx.doi.org/10.1090/gsm/040. [55] Grothendieck A. Produits tensoriels topologiques et espaces nucl´eaires. Mem. Amer. Math. Soc. no. 16. -- -- Providence, RI : American Mathematical Society, 1966. -- -- 383 p. [56] Guo C.-H., Higham N. J., Tisseur F. Detecting and solving hyperbolic quadratic eigenvalue problems // SIAM J. Matrix Anal. Appl. -- -- 2008/09. -- -- Vol. 30, no. 4. -- -- P. 1593 -- 1613. -- -- URL: http://dx.doi.org/10.1137/070704058. [57] Haase M. The functional calculus for sectorial operators. -- -- Basel 2006. -- -- Verlag, cations. -- -- http://dx.doi.org/10.1007/3-7643-7698-8. xiv+392 p. -- -- Vol. 169 of Operator Theory: : Birkhauser Advances and Appli- ISBN: 978-3-7643-7697-0; 3-7643-7697-X. -- -- URL: [58] Harbarth K., Langer H. A factorization theorem for operator pencils // Integral Equations Operator Theory. -- -- 1979. -- -- Vol. 2, no. 3. -- -- P. 344 -- 364. -- -- URL: http://dx.doi.org/10.1007/BF01682674. [59] Harte R. Spectral mapping theorems: a bluffer's guide. Mathematics. -- -- Cham -- Heidelberg -- New York -- Dordrecht -- London : 2014. -- -- in Springer, ISBN: 978-3-319-05647-0; 978-3-319-05648-7. -- -- URL: Springer Briefs xiv+120 p. -- -- 45 http://dx.doi.org/10.1007/978-3-319-05648-7. [60] Harte R. E. Tensor products, multiplication operators and the spectral mapping theorem // Proc. Roy. Irish Acad., Sect. A. -- -- 1973. -- -- Vol. 73. -- -- P. 285 -- 302. [61] Heinig G. The concepts of a bezoutiant and a resolvent for operator bundles // Funct. Anal. Appl. -- -- 1977. -- -- Vol. 11, no. 3. -- -- P. 241 -- 243. [62] Heinig G., Rost K. Introduction to Bezoutians // Numerical methods for structured matrices and applications. -- -- Basel : Birkhauser Verlag, 2010. -- -- Vol. 199 of Oper. Theory Adv. Appl. -- -- P. 25 -- 118. [63] Heinz E. Beitrage zur Storungstheorie der Spektralzerlegung // Math. Ann. -- -- 1951. -- -- Vol. 123. -- -- P. 415 -- 438. [64] Helemskii A. Y. Lectures and exercises on functional analysis. -- -- Providence, RI : American Mathematical Society, 2006. -- -- Vol. 233 of Translations of Mathematical Monographs. -- -- xviii+468 p. -- -- ISBN: 978-0-8218-4098-6; 0-8218-4098-3. [65] Higham N. J. Industrial Philadelphia, ics ISBN: (SIAM), http://dx.doi.org/10.1137/1.9780898717778. Society xx+425 p. -- -- PA : 2008. -- -- Functions of matrices: theory and computation. -- -- and Applied Mathemat- URL: 978-0-89871-646-7. -- -- for [66] Higham N. J., Lin L. An improved Schur-Pad´e algorithm for fractional powers of a matrix and their Fr´echet derivatives // SIAM J. Matrix Anal. Appl. -- -- 2013. -- -- Vol. 34, no. 3. -- -- P. 1341 -- 1360. -- -- URL: http://dx.doi.org/10.1137/130906118. [67] Hille E., Phillips R. S. Functional analysis and semi-groups. -- -- Providence, RI : Amer. Math. Soc., 1957. -- -- Vol. 31 of American Mathematical Society Colloquium Publications. -- -- xiv+808 p. [68] Ichinose T. On the spectra of tensor products of linear operators in Banach spaces. // J. Reine Angew. Math. -- -- 1970. -- -- Vol. 244. -- -- P. 119 -- 153. [69] Ichinose T. Operational calculus for tensor products of linear operators in Banach spaces // Hokkaido Math. J. -- -- 1975. -- -- Vol. 4, no. 2. -- -- P. 306 -- 334. [70] Ikramov K. D. Numerical solution of matrix equations. Orthogonal methods. -- -- Moscow : Nauka, 1984. -- -- 192 p. -- -- (in Russian). [71] Ikramov K. D. The nonsymmetric eigenvalue problem. -- -- Moscow : Nauka, 1991. -- -- 240 p. -- -- (in Russian). [72] Isaev G. A. Linear factorization of polynomial operator pencils // Mat. Zametki. -- -- 1973. -- -- Vol. 13. -- -- P. 551 -- 559. [73] Jordan C. Calculus of finite differences. -- -- Third edition. -- -- New York : Chelsea Publishing Co., 1965. -- -- xxi+655 p. [74] Karplus R., Schwinger J. A note on saturation in microwave spectroscopy // Phys- ical Review. -- -- 1948. -- -- Vol. 73, no. 9. -- -- P. 1020 -- 1026. [75] Kato T. Perturbation theory for linear operators. Classics in Mathematics. -- -- Berlin : Springer -- Verlag, 1995. -- -- xxii+619 p. -- -- ISBN: 3-540-58661-X. -- -- Reprint of the 1980 edition. [76] Keldysh M. V. On the characteristic values and characteristic functions of certain classes of non-self-adjoint equations // Doklady Akad. Nauk SSSR (N.S.). -- -- 1951. -- -- Vol. 77, no. 1. -- -- P. 11 -- 14. [77] Kenney C. S., Laub A. J. Condition estimates for matrix functions // SIAM J. Matrix Anal. Appl. -- -- 1989. -- -- Vol. 10, no. 2. -- -- P. 191 -- 209. -- -- URL: http://dx.doi.org/10.1137/0610014. [78] Kenney C. S., Laub A. J. A Schur-Fr´echet algorithm for computing the logarithm and exponential of a matrix // SIAM J. Matrix Anal. Appl. -- -- 1998. -- -- URL: (electronic). -- -- 640 -- 663 Vol. P. 19, no. 3. -- -- 46 http://dx.doi.org/10.1137/S0895479896300334. [79] Kostrykin V., Makarov K. A., Motovilov A. K. Existence and uniqueness of solutions to the operator Riccati equation. A geometric approach // Advances in differential equations and mathematical physics (Birmingham, AL, 2002). -- -- Providence, RI : Amer. Math. Soc., 2003. -- -- Vol. 327 of Contemp. Math. -- -- P. 181 -- 198. -- -- URL: http://dx.doi.org/10.1090/conm/327/05814. [80] Kostrykin V., Makarov K. A., Motovilov A. K. On the existence of solutions to the operator Riccati equation and the tan Θ theorem //Integral Equations Operator Theory. -- -- 2005. -- -- Vol. 51, no. 1. -- -- P. 121 -- 140. [81] Kostyuchenko A. G., Shkalikov A. A. Self-adjoint quadratic operator pencils and elliptic problems // Funct. Anal. Appl. -- -- 1983. -- -- Vol. 17, no. 2. -- -- P. 109 -- 128. [82] Kreın M. G. Some new studies in the theory of perturbations of self-adjoint oper- ators // First Math. Summer School, Part I. -- -- Kiev : Naukova Dumka, 1964. -- -- P. 103 -- 187. -- -- (Russian). [83] Kreın M. G., principles in // Integral Equations Operator Theory. -- -- 1978. -- -- Vol. 1, no. 3. -- -- P. 364 -- 399. -- -- URL: http://dx.doi.org/10.1007/BF01682844. H. On damped Langer of mathematical oscillations continua. theory linear some of I the [84] Kreın M. G., Naımark M. A. The method of symmetric and Hermitian forms in the theory of the separation of the roots of algebraic equations // Linear and Multilinear Algebra. -- -- 1981. -- -- Vol. 10, no. 4. -- -- P. 265 -- 308. -- -- URL: http://dx.doi.org/10.1080/03081088108817420. [85] Kressner D. Bivariate matrix functions // Operators and Matrices. -- -- 2014. -- -- Vol. 8, no. 2. -- -- P. 449 -- 466. [86] Kressner D., Tobler C. Krylov subspace methods for linear systems with tensor prod- uct structure // SIAM J. Matrix Anal. Appl. -- -- 2010. -- -- Vol. 31, no. 4. -- -- P. 1688 -- 1714. [87] Krupnik I. N. Decomposition tors // Mat. Zametki. -- -- http://dx.doi.org/10.1007/BF01137549. fac- 1991. -- -- Vol. 49, no. 2. -- -- P. 95 -- 101. -- -- URL: a matrix pencil linear into of [88] Kurbatov V. G., Kurbatova I. V. Krylov subspace methods of approximate solving of differential equations from the point of view of functional calculus // Eurasian Math. J. -- -- 2012. -- -- Vol. 3, no. 4. -- -- P. 53 -- 80. [89] Kurbatov V. G., Kurbatova I. V. Extended tensor products and an operator-valued spectral mapping theorem // Izv. Math. -- -- 2015. -- -- Vol. 79, no. 4. -- -- P. 710 -- 739. [90] Kurbatov V. G., Oreshina M. N. On approximate solution of second order lin- ear differential equation // Vestnik Voronezhskogo gosudarstvennogo universiteta. Seriya: Fizika. Matematika [Proceedings of Voronezh State University. Series: Physics. Mathematics]. -- -- 2003. -- -- no. 2. -- -- P. 173 -- 188. -- -- (in Russian). URL: http://www.vestnik.vsu.ru/content/physmath/2003/02/toc_ru.asp. [91] Kurbatov V. G., Oreshina M. N. Interconnect macromodelling and approximation of matrix exponent // Analog Integrated Circuits and Signal Processing. -- -- 2004. -- -- Vol. 40, no. 1. -- -- P. 5 -- 19. [92] Kurbatova I. V. Some properties of relations // Vestnik fakulteta PMM [Proceedings of the faculty of Applied Mathematics]. -- -- Voronezh : Voronezh State University, 2009. -- -- Vol. 7. -- -- P. 68 -- 89. -- -- (in Russian). URL: https://www.researchgate.net/publication/269112381. linear [93] Kurbatova I. V. Functional calculus generated by a square pencil // Journal of Mathematical Sciences. -- -- 2012. -- -- Vol. 182, no. 5. -- -- P. 646 -- 655. 47 [94] Lan N. T. On the operator equation AX − XB = C with unbounded operators A, B, and C // Abstr. Appl. Anal. -- -- 2001. -- -- Vol. 6, no. 6. -- -- P. 317 -- 328. -- -- URL: http://dx.doi.org/10.1155/S1085337501000665. [95] Lumer G., Rosenblum M. Linear operator equations // Proc. Amer. Math. Soc. -- -- 1959. -- -- Vol. 10. -- -- P. 32 -- 41. [96] Markus A. S. Introduction to the spectral theory of polynomial operator pencils. -- -- Providence, RI : American Mathematical Society, 1988. -- -- Vol. 71 of Translations of Mathematical Monographs. -- -- iv+250 p. -- -- ISBN: 0-8218-4523-3. [97] Maroulas J., Psarrakos P. On factorization of matrix polynomials // 2000. -- -- Vol. 304, no. 1-3. -- -- P. 131 -- 139. -- -- URL: Linear Algebra Appl. -- -- http://dx.doi.org/10.1016/S0024-3795(99)00196-2. of [98] Mathias R. Evaluating tial // Numer. Math. -- -- http://dx.doi.org/10.1007/BF01385857. the Fr´echet derivative the matrix exponen- 1992. -- -- Vol. 63, no. 2. -- -- P. 213 -- 226. -- -- URL: [99] Mathias R. Approximation SIAM J. Matrix Anal. Appl. -- -- URL: http://dx.doi.org/10.1137/0614070. of // 1993. -- -- Vol. 14, no. 4. -- -- P. 1061 -- 1063. -- -- matrix-valued functions [100] Najfeld I., Havel T. F. Derivatives of the matrix exponential and their computa- tion // Advances in Applied Mathematics. -- -- 1995. -- -- Vol. 16. -- -- P. 321 -- 375. [101] Parlett B. A recurrence among the elements of functions of triangular matrices // Linear Algebra and Appl. -- -- 1976. -- -- Vol. 14, no. 2. -- -- P. 117 -- 121. [102] Perturbation theory for matrix equations / M. Konstantinov, Da-Wei Gu, V. Mehr- mann, P. Petkov. -- -- Amsterdam : North-Holland Publishing Co., 2003. -- -- Vol. 9 of Studies in Computational Mathematics. -- -- xii+429 p. -- -- ISBN: 0-444-51315-9. [103] Pervozvanskii A. A. A Course in the Theory of Automatic Control. -- -- Moscow : Nauka, 1986. -- -- 615 p. -- -- (in Russian). [104] Ph´ong V. Q. The operator equation AX − XB = C with unbounded operators A and B and related abstract Cauchy problems // Math. Z. -- -- 1991. -- -- Vol. 208, no. 4. -- -- P. 567 -- 588. -- -- URL: http://dx.doi.org/10.1007/BF02571546. [105] Reed M., Simon B. Tensor products of closed operators on Banach spaces // J. Functional Analysis. -- -- 1973. -- -- Vol. 13. -- -- P. 107 -- 124. [106] Reed M., Simon B. Methods of modern mathematical physics. I. Functional anal- ysis. -- -- Second edition. -- -- New York : Academic Press, Inc. [Harcourt Brace Jo- vanovich, Publishers], 1980. -- -- xv+400 p. -- -- ISBN: 0-12-585050-6. [107] Reed M. C., Simon B. A spectral mapping theorem for tensor products of unbounded operators // Bull. Amer. Math. Soc. -- -- 1972. -- -- Vol. 78. -- -- P. 730 -- 733. [108] Rinehart R. F. The derivative of a matrix function // Proc. Amer. Math. Soc. -- -- 1956. -- -- Vol. 7. -- -- P. 2 -- 5. [109] Rosenblum M. On the operator equation BX−XA = Q // Duke Math. J. -- -- 1956. -- -- [110] Rudin W. Functional analysis. International Series in Pure and Applied Mathemat- Vol. 23. -- -- P. 263 -- 269. ics. -- -- Second edition. -- -- New York : McGraw-Hill, Inc., 1973. -- -- xiii+397 p. [111] Schaefer H. H. Topological vector spaces. -- -- The Macmillan Co., New York; Collier- Macmillan Ltd., London, 1966. -- -- P. ix+294. [112] Schatten R. A. A theory of cross-spaces. -- -- N. J. : Princeton Univ. Press, 1950. -- -- vii+153 p. [113] Schechter M. On the spectra of operators on tensor products // J. Functional Anal- ysis. -- -- 1969. -- -- Vol. 4. -- -- P. 95 -- 99. 48 [114] Seshu S., Reed M. B. Linear graphs and electrical networks. Addison-Wesley Series in the Engineering Sciences. -- -- Mass.-London : Addison-Wesley Publishing Co., Inc., 1961. -- -- x+315 p. [115] Shabat B. V. Introduction to complex analysis. Part II. Functions of several vari- ables. -- -- Providence, RI : Amer. Math. Soc., 1992. -- -- Vol. 110 of Translations of Mathematical Monographs. -- -- x+371 p. [116] Shaw S.-Y., Lin S. C. On the equations Ax = q and SX − XT = 1988. -- -- Vol. 77, no. 2. -- -- P. 352 -- 363. -- -- URL: Q // J. Funct. Anal. -- -- http://dx.doi.org/10.1016/0022-1236(88)90092-4. [117] Shkalikov A. A., Pliev V. T. Compact perturbations of strongly damped operator pencils // Math. Notes. -- -- 1989. -- -- Vol. 45, no. 2. -- -- P. 167 -- 174. [118] Simon B. Uniform crossnorms // Pacific J. Math. -- -- 1973. -- -- Vol. 46. -- -- P. 555 -- 560. [119] Simoncini V. On the numerical = C // BIT. Numerical Mathematics. -- -- 1996. -- -- Vol. 36, no. 4. -- -- P. 814 -- 830. -- -- URL: http://dx.doi.org/10.1007/BF01733793. solution of AX − XB [120] Computational methods for linear matrix equations : Rep. / Universit'a di Bologna ; Executor: V. Simoncini : 2013. -- -- March. -- -- 58 p. [121] Stickel E. On the Fr´echet derivative of matrix functions // Linear Algebra Appl. -- -- 1987. -- -- Vol. 91. -- -- P. 83 -- 88. [122] Sylvester J. Sur l'equations en matrices px = xq // C. R. Acad. Sci. Paris. -- -- 1884. -- -- Vol. 99. -- -- P. 67 -- 71, 115 -- 116. [123] Taylor A. E. Analysis in complex Banach spaces // Bull. Amer. Math. Soc. -- -- 1943. -- -- Vol. 49. -- -- P. 652 -- 669. [124] Taylor A. E. Spectral theory of closed distributive operators // Acta Math. -- -- 1951. -- -- Vol. 84. -- -- P. 189 -- 224. [125] Tisseur quadratic SIAM Rev. -- -- no. 2. -- -- http://dx.doi.org/10.1137/S0036144500381988. F., Meerbergen K. The 43, 2001. -- -- Vol. eigenvalue P. 235 -- 286. -- -- problem // URL: [126] Van Loan C. The sensitivity of the matrix exponential // SIAM J. Numer. Anal. -- -- 1977. -- -- Vol. 14, no. 6. -- -- P. 971 -- 981. [127] Yosida K. Functional analysis. -- -- Sixth edition. -- -- Berlin -- New York : Springer-Ver- lag, 1980. -- -- Vol. 123 of Grundlehren der Mathematischen Wissenschaften [Funda- mental Principles of Mathematical Sciences]. -- -- P. xii+501. -- -- ISBN: 3-540-10210-8. [128] Zayachkovskii V. S., Pankov A. A. Stability of factorization of polynomial operator pencils // Funct. Anal. Appl. -- -- 1983. -- -- Vol. 17, no. 2. -- -- P. 140 -- 142. [129] Zayachkovskii V. S., Pankov A. A. Weak stability of factorizations of operator pen- cils // Journal of Mathematical Sciences. -- -- 1990. -- -- Vol. 52, no. 6. -- -- P. 3548 -- 3550. [130] Zhu W., Xue J., Gao W. The sensitivity of the exponential of an essentially non- negative matrix // J. Comput. Math. -- -- 2008. -- -- Vol. 26, no. 2. -- -- P. 250 -- 258. Department of Mathematical Physics, Voronezh State University, 1, Universitetskaya Square, Voronezh 394036, Russia E-mail address: [email protected] Department of Mathematics, Air Force Academy of the Ministry of Defense of the Russian Federation, 54 "A", Starykh Bol'shevikov Str., Voronezh 394064, Russia E-mail address: la [email protected] Department of Applied Mathematics, Lipetsk State Technical University, 30, Moskov- skaya Str., Lipetsk 398600, Russia E-mail address: M [email protected] 49
1803.04560	2	1803	2018-08-30T19:53:19	How far is the Borel map from being surjective in quasianalytic ultradifferentiable classes?	[ "math.FA" ]	The Borel map takes germs at 0 of smooth functions to the sequence of iterated partial derivatives at 0. In the literature, it is well known that the restriction of this mapping to the germs of quasianalytic ultradifferentiable classes which are strictly containing the real analytic functions can never be onto the corresponding sequence space. In this paper, we are interested in studying how large the image of the Borel map is and we investigate the size and the structure of this image by using different approaches (Baire residuality, prevalence and lineability). We give an answer to this question in the very general setting of quasianalytic ultradifferentiable classes defined by weight matrices, which contains as particular cases the classes defined by a single weight sequence or by a weight function.	math.FA	math	HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE IN QUASIANALYTIC ULTRADIFFERENTIABLE CLASSES? CÉLINE ESSER AND GERHARD SCHINDL Abstract. The Borel map j∞ takes germs at 0 of smooth functions to the sequence of iterated partial derivatives at 0. In the literature, it is well known that the restriction of j∞ to the germs of quasianalytic ultradifferentiable classes which are strictly containing the real analytic functions can never be onto the corresponding sequence space. In this paper, we are interested in studying how large the image of j∞ is and we investigate the size and the structure of this image by using different approaches (Baire residuality, prevalence and lineability). We give an answer to this question in the very general setting of quasianalytic ultradifferentiable classes defined by weight matrices, which contains as particular cases the classes defined by a single weight sequence or by a weight function. 8 1 0 2 g u A 0 3 ] . A F h t a m [ 2 v 0 6 5 4 0 . 3 0 8 1 : v i X r a 1. Introduction In 1895, E. Borel proved that given any sequence (an)n∈N of complex numbers, there exists a in- finitely differentiable function such that f (n)(0) = an for every n ∈ N [8]. This work has been investigated and extended ever since by many authors. In particular, the question has been han- dled in the context of so-called ultradifferentiable classes which are subclasses of smooth functions defined by imposing growth conditions on the derivatives of the functions using weight sequences M , functions ω or matrices M, see [10, 11, 17, 26, 7, 6, 4, 20]. Historically, those classes have been first introduced by using weight sequences, motivated among others by the characterization of the regularity of solutions of the heat equation or of other partial differential equations, see e.g. In order to measure the decay of the Fourier transform of smooth functions with compact support, classes of ultradifferentiable functions have then been defined using weight functions, e.g. see [3] and [18]. In [9], it turned out that such a behavior can also equivalently be expressed by having control on the growth of all the derivatives of the function itself in terms of this weight function and in [5] it has been shown that classes defined in terms of weight sequences and weight functions are in general mutually distinct. Finally, in [19] and [24], classes defined by weight matrices have been considered. It turned out that the weight sequence and weight function frameworks are particular cases of this setting, and this general method allows to treat both classical approaches jointly but also leads to more general classes. We say that an ultradifferentiable class is quasianalytic if the restriction of the Borel map f 7→ (∂αf (0))α∈Nr to this class is injective; this notion plays an important role in many different contexts and applications (e.g. such classes do not contain partitions of unity). It came out of many studies [21]. Date: November 9, 2018. 2010 Mathematics Subject Classification. 26E10, 46A13, 46E10, 54E52. Key words and phrases. Spaces of ultradifferentiable functions, Borel map, quasianalyticity, genericity, Baire category, prevalence, lineability. CE is supported by a F.R.S.-FNRS grant; GS is supported by FWF-Project J 3948-N35. 1 2 C. ESSER AND G. SCHINDL that the restriction of the Borel map to the germs of quasianalytic ultradifferentiable classes which are strictly containing the real analytic functions can never be onto the corresponding sequence space. However, an interesting remaining question is "how far away the Borel map is from being surjective?" This is the question we tackle in this paper: We show that the image of the Borel map is "small" in the corresponding sequence space, using different approaches (as done e.g. in [13]). Let us present these different notions here. First, let us recall the following classical definition which gives a notion of residuality from a topological point of view. Definition 1.0.1. If X is a Baire space, then a subset L ⊂ X is called comeager (or residual) if L contains a countable intersection of dense open sets of X. The complement of a residual set is a meager (or first category) set in X. In order to get result about the "size" of sets from a measure-theoretical point of view, the notion of prevalence can be used. It has been introduced in [12, 15] to give an extension of the concept of "almost everywhere" (for the Lebesgue measure) to metric infinite dimensional spaces (in these spaces, no measure is both σ-finite and translation invariant). Definition 1.0.2. Let X denote a complete metric vector space. A Borel subset B ⊂ X is called Haar-null if there exists a compactly supported probability measure µ such that (1.1) ∀x ∈ X, µ(x + B) = 0. A subset S of X is called Haar-null if it is contained in a Haar-null Borel set. A prevalent set is the complement of a Haar-null set. The following results of [12] and [15] enumerate important basic properties of prevalent sets: • If S is Haar-null, then x + S is Haar-null for any x ∈ X. • If the dimension of X is finite, S is Haar-null if and only if S has Lebesgue measure 0. • Prevalent sets are dense. • Any countable intersection of prevalent sets is prevalent. Remark 1.0.3. A useful way to get that a Borel set is Haar-null is to try the Lebesgue measure on the unit ball of a finite dimensional subspace V . In this context, condition (1.1) is equivalent to In this case, we say that V is a probe for the complement of B. ∀x ∈ X, (x + B) ∩ V is of Lebesgue measure zero. Finally, we will also consider the notion of lineability, introduced in [1]. This notion was motivated by the increasing interest toward the search for large algebraic structures of special objects (see [2] for a review). Definition 1.0.4. A set L in a vector space X is said to be lineable in X if L ∪ {0} contains an infinite dimensional vector space. Note that in the above definition, the considered vector space is generated by the finite linear combinations of the elements of an infinite basis. The present paper is organized as follows: In Section 2, we recall basic definitions and results concerning weight sequences M , the M -ultradifferentiable classes (of germs at 0) and the associated sequence spaces which will be needed. In Section 3, we recall some important elements of the proof of [26, Theorem 3] which gives the non-surjectivity of the Borel map in the quasianalytic setting, HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 3 and we explain how to extend it in different directions. This Theorem allows us to obtain that the image of the Borel map is small (i.e. meager, Haar-null) in the Beurling case and its complement is lineable in both Roumieu and Beurling cases for quasianalytic classes strictly containing the real analytic functions. Let us mention that the above results are obtained in a more general context, since we actually prove that the image of the Borel map defined on any quasianalytic germ class associated with a sequence M is small in any weighted sequence space associated with another quasianalytic weight sequence N (assuming some mild standard assumptions on M and N ). Using the results and techniques developed in the single weight sequence case, we study in Section 4 the weight matrix case as well. Here the weight matrix is a one-parameter family of sequences M = {M (λ) : λ ∈ R>0}, M (λ) again having some mild standard assumptions. All the results from the previous section are transferred to this more general setting with the same generality. Finally, in Section 5, we treat the weight function case ω. This is done by using the weight matrix setting and the fact that with each weight function ω (satisfying some mild standard growth con- ditions), we can associate a weight matrix Ω = {W (λ) : λ ∈ R>0} such that the ultradifferentiable classes defined by ω and Ω coincide (as locally convex vector spaces) and similarly for the corre- sponding sequence spaces, see [24, Sect. 6, Sect. 7], [19, Sect. 5] and [20]. Thus the results in this section can be seen as immediate Corollaries of the previous Section 4. Note that the presentation of this work and the standard assumptions on the weight structures are similar to the ones considered in [20]. Moreover, throughout this paper, we write N = {0, 1, . . .}, E(U ) and Cω(U ) shall denote respectively the class of all C-valued smooth functions and the class of all real analytic functions defined on non-empty open U ⊆ Rr. 2. Weight sequences and germs of ultradifferentiable functions 2.1. Denjoy-Carleman ultradifferentiable classes and their germs. Definition 2.1.1. Let M = (Mp)p∈N ∈ RN >0 be an arbitrary sequence of positive real numbers. Let r ∈ N>0 and U ⊆ Rr be non-empty and open. The M -ultradifferentiable Roumieu type class is defined by E{M}(U ) := {f ∈ E(U ) : ∀ K ⊆ U compact ∃ h > 0, kfkM K,h < +∞}, and the M -ultradifferentiable Beurling type class by E(M)(U ) := {f ∈ E(U ) : ∀ K ⊆ U compact ∀ h > 0, kfkM where (using the standard multi-index notation for the partial derivatives) K,h < +∞}, kfkM K,h := sup α∈Nr,x∈K ∂αf (x) hαMα . As usual, we will write m = (mp)p∈N for mp := Mp p! . Remark 2.1.2. At this point, we want to make the reader aware that the sequence M considered in [26] is precisely the sequence m = (mp)p∈N in the notation of this work. For any compact set K with smooth boundary EM,h(K) := {f ∈ E(K) : kfkM K,h < +∞} is a Banach space. The Roumieu type class is endowed with the projective topology w.r.t. all K ⊆ U compact and the inductive topology w.r.t. h ∈ N>0, whereas the Beurling type class is endowed with the projective topology w.r.t. K ⊆ U compact and w.r.t. 1/h, h ∈ N>0. Hence E(M)(U ) is a Fréchet 4 C. ESSER AND G. SCHINDL space and lim−→h>0 EM,h(K) = lim−→n∈N>0 EM,n(K) is a Silva space, i.e. a countable inductive limit of Banach spaces with compact connecting mappings, see [16, Proposition 2.2]. Note that the special case Mp = p! yields E{M}(U ) = Cω(U ), whereas E(M)(U ) consists of the restrictions of all entire functions provided that U is connected. Definition 2.1.3. The spaces of germs at 0 ∈ Rr of the M -ultradifferentiable functions of Roumieu and Beurling types are defined respectively by and E 0,r {M} := lim−→k∈N>0 E 0,r (M) := lim−→k∈N>0 E{M}(cid:18)(cid:18)− E(M)(cid:18)(cid:18)− 1 k , 1 k , 1 k(cid:19)r(cid:19) , k(cid:19)r(cid:19) . 1 Again, if one considers the sequence Mp = p! in the Roumieu case, we obtain the space of germs of real analytic functions at 0 ∈ Rr; it is denoted by O0,r. Let us now introduce the corresponding spaces of complex sequences. Definition 2.1.4. We define the sequence spaces Λr {M} and Λr (M) by setting and where for any h > 0, Λr Λr {M} =nb ∈ CNr (M) =nb ∈ CNr bα bM h := sup α∈Nr : ∃h > 0,bM : ∀h > 0,bM h < +∞o h < +∞o , bαα! hαMα . hαmα = sup α∈Nr We endow these spaces also with their natural topology: Λ{M} is an (LB)-space and Λ(M) a Fréchet space. Remark 2.1.5. Note that there exists a one-to-one correspondence between Λr weighted formal power series, whose elements F =Pα∈Nr Fαxα satisfy Fα ≤ Chαmα for some C, h > 0 and all α ∈ Nr in the Roumieu case, resp. for all h > 0 small, some C = Ch large and all α ∈ Nr in the Beurling case (as it as been considered in [26, Section 1.2] for the Roumieu case). Finally, let us define the Borel map. For reasons of convenience, the following convention will also be used: we write E[M] if either E{M} or E(M) is considered, but not mixing the cases if statements involve more than one E[M] symbol. We use similar notations for the sequence classes Λr [M]. In both cases, the Borel map j∞ is defined by [M] and the (ring) of (2.1) j∞ : E 0,r [M] −→ Λr [M], j∞(f ) =(cid:18) ∂αf (0) α! (cid:19)α∈Nr . Remark 2.1.6. At this step, let us point out that our definitions of sequence spaces differ from those in [20, Section 2.1]; this is due to technical reasons since we will have to work mainly with the sequence m = ( Mp p! )p∈N instead of M = (Mp)p∈N. HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 5 Remark 2.1.7. Moreover, we outline that we could replace in the definition of the germ spaces the point 0 ∈ Rr by any other point a ∈ Rr and define the corresponding Borel map j∞ a similarly as in (2.1). 2.2. Weight sequences. We consider the following definition, according to [20, Section 2.2]. Definition 2.2.1. A sequence of positive real numbers M = (Mp)p∈N ∈ RN sequence if >0 is called a weight (I) 1 = M0 ≤ M1 (normalization), (II) p 7→ Mp is log-convex, or equivalently M 2 (III) lim inf p→∞(mp)1/p > 0. Recall that mp := Mp p! for every p ∈ N. p ≤ Mp−1Mp+1 for all p ∈ N>0, Let us recall that if M is log-convex and normalized, then M and k 7→ (Mk)1/k are both increasing and MjMk ≤ Mj+k holds for all j, k ∈ N, e.g. see [23, Lemmata 2.0.4, 2.0.6]. Given two (weight) sequences, we write M ≤ N if and only if Mp ≤ Np holds for all p ∈ N and define the relations M (cid:22) N :⇔ ∃ C ≥ 1 ∀ p ∈ N : Mp ≤ CpNp ⇐⇒ sup p∈N>0(cid:18) Mp Np(cid:19)1/p p→∞(cid:18) Mp < +∞, Np(cid:19)1/p M ⊳ N :⇔ ∀ h > 0 ∃ Ch ≥ 1 ∀ p ∈ N : Mp ≤ ChhpNp ⇐⇒ lim = 0. It is straightforward to see that in the above relations we can replace the sequences M and N simultaneously by the sequences m and n. Those relations between weight sequences imply inclusions between ultradifferentiable classes, e.g. see [20, Section 2.2] and the references therein. More precisely, let M be a weight sequence and N arbitrary, then M(cid:22)N if and only if E[M] ⊆ E[N ], which is equivalent to Λr [N ]. In particular, choosing M = (p!)p∈N, we get Cω ⊆ E{N } if and only if lim inf p→+∞(np)1/p > 0. Moreover, if N is a weight sequence, then E{N } ⊆ Cω if and only if supp∈N>0 (np)1/p < +∞. Hence Cω ( E{N } if and only if supp∈N>0 (np)1/p = +∞. Similarly M ⊳N if and only if E{M} ( E(N ), which is equivalent to Λr (N ). In particular, Cω ( E(N ) if and only if limp→+∞(np)1/p = +∞. Definition 2.2.2. A weight sequence M is called quasianalytic if [M] ⊆ Λr {M} ( Λr (Q) +∞Xp=1 Mp−1 Mp = +∞. By using Carleman's inequality (a proof is presented in [23, Proposition 4.1.7]), one can show that +∞Xp=1 Mp−1 Mp = +∞ ⇐⇒ +∞Xp=1 1 (Mp)1/p = +∞. Definition 2.2.3. A subclass Q ⊆ E is called quasianalytic if for any open connected set U ⊆ Rr and each point a ∈ U , the Borel map j∞ a is injective on Q(U ). 6 C. ESSER AND G. SCHINDL In the case Q ≡ E[M] the Denjoy-Carleman theorem characterizes this behavior in terms of the defining weight sequence M . More precisely, it states that E[M] is quasianalytic if and only if M satisfies (Q). Let us moreover mention that E[M] is quasianalytic if and only if there do not exist non-trivial functions in E[M] with compact support, e.g. see [22, Thm. 19.10]. A basic assumption in the proof of the Denjoy-Carleman theorem is condition (II) for M (i.e. M is log-convex). But sometimes it might be convenient to skip this condition and to work in a more general stetting: In this case, one considers admissible regularizations of M (with M0 = 1), see [25, Section 4.3] and the references therein. For this reason, we shall denote by • M lc the log-convex minorant of M , i.e. the largest sequence N such that N is log-convex p )1/p)p∈N>0 is the increasing minorant of ((Mp)1/p)p∈N>0 • M I the sequence for which ((M I and N ≤ M , (and put M I 0 := 1). Consequently, M = M lc if and only if (II), and M I = M if and only if k 7→ (Mk)1/k is increasing. One has M lc ≤ M I ≤ M since a log-convex weight sequence is increasing. We can now recall the following result, see [25, Proposition 4.4] which is based on [14, Theorem 1.3.8] for the Roumieu case, and see [16, Theorem 4.2] for the Beurling case. Proposition 2.2.4. Let M ∈ RN >0 with M0 = 1. The following assertions are equivalent: • E[M] is quasianalytic, • M lc satisfies (Q), • P+∞ p )1/p = +∞. 1 (M I p=1 Remark 2.2.5. We mention that in the following sections, we will study the Borel map j∞ defined in quasianalytic ultradifferentiable classes such that Cω ( E[M] holds true. As already pointed out in [20, Remark 1], the general assumptions (I) − (III) on M are not restricting the generality of our >0 with Cω ⊆ E[M] we have lim inf p→+∞(mp)1/p > considerations. More precisely, for any M ∈ RN 0 in the Roumieu and limp→+∞(mp)1/p = +∞ in the Beurling case (see also [19, Prop. 2.12 (4), (5)]). Then, by [19, Theorem 2.15], we can replace M by M lc without changing the associated ultradifferentiable class whereas only Λr [M] follows (and the weight matrix/function setting is reduced to the sequence case situation as will be seen in the next sections). Remark 2.2.6. Let us point out that all results below also hold true if 0 ∈ Rr is replaced by any other point a ∈ Rr (translation). [M lc] ⊆ Λr 3. The weight sequence case M 3.1. Thilliez's proof for non-surjectivity. Let M be a weight sequence, i.e. satisfying our standard assumptions (I) − (III). To ensure that the real analytic functions/germs are strictly contained in the considered class E[M], we have to assume sup k∈N>0 (mk)1/k = +∞ or lim k→+∞ (mk)1/k = +∞ in the Roumieu or in the Beurling case respectively. The aim of this subsection is to recall the main elements of the proof of [26, Theorem 3], which is based on the original ideas of Carleman [11], to be applicable in our present context. We will also explain how it can be extended: Indeed, in [26] only the Roumieu case has been treated and it has been assumed there that m = (mk)k∈N is log-convex (in that case we say that M is strongly HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 7 log-convex), which implies that k 7→ (mk)1/k is increasing. Consequently, also in the Roumieu case, the assumption for the strict inclusion turns into limk→+∞(mk)1/k = +∞. In our approach we do not want to assume strongly log-convexity on M , or more generally on some or all M (λ) ∈ M in the weight matrix case considered in Section 4 below: This is due to the fact that on the one hand, in general we do not know whether some or all of the sequences W (λ) of the matrix Ω associated with a weight function ω will satisfy this requirement, see Section 5 below for further explanations. On the other hand, in any cases, strongly log-convexity seems to be too strong and superfluous in studying the questions under consideration in this paper (but not for some questions studied in [26]). A second generalization is that we also consider a kind of mixed setting of two (in general different) weight sequences M and N . Let us start by recalling the following representation formula, obtained within the first part of the proof of [26, Theorem 3]. As mentioned before, this result has been obtained by assuming the strongly log-convexity on M . However, by following directly the lines of this proof, the result still holds with the weaker (basic) assumptions on M . Hence, we have the following Theorem. Theorem 3.1.1 (Representation formula, [26]). Let M be a quasianalytic weight sequence. There exist numbers (ωM j,k)j,k∈N such that (3.1) lim k→+∞ ωM j,k = 1, ∀j ∈ N, and such that, given any function f ∈ E 0,1 {M}, one has (3.2) for every x > 0 small enough. f (x) = lim k→+∞ k−1Xj=0 ωM j,k f (j)(0) j! xj Keeping the notations of this Theorem, we directly get the following important result. Corollary 3.1.2. Let M be a quasianalytic weight sequence. If b = (bj)j∈N ∈ CN is a sequence for which there exists a sequence of positive real numbers (an)n∈N decreasing to 0 such that (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 n(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (3.3) lim sup k→+∞ ωM j,kbjaj = +∞ for all n ∈ N, then b /∈ j∞(E 0,1 Proof. Assume by contradiction that we can find f ∈ E 0,1 {M} such that j∞(f ) = b. Using the representation formula (3.2) of Theorem 3.1.1 together with the definition of the Borel map, we get {M}). f (x) = lim k→+∞ k−1Xj=0 j,kbjxj ωM for every x > 0 small enough, hence a contradiction. Remark 3.1.3. Let b = (bj)j∈N be a sequence which does not belong to j∞(E 0,1 r > 1, any sequence eb ∈ CNr satisfyingeb(j,0,...,0) = bj for every j ∈ N is not in the image j∞(E 0,r {M}). Then, for {M}): (cid:3) 8 C. ESSER AND G. SCHINDL Indeed, if one assumes now that there is f ∈ E 0,r restriction mapping R : E 0,r would obtain that j∞(R(f )) = b. {M} such that eb = j∞(f ), then by considering the {M}, the restriction R(f ) of f would belong to E 0,1 {M} and one ։ E 0,1 {M} The following Theorem is a direct generalization of the second part of the proof of [26, Theorem 3]: We consider two weight sequences (different or not) and we treat the Beurling case as well. Theorem 3.1.4. Let M and N be two quasianalytic weight sequences such that supk∈N>0 (nk)1/k = +∞ resp. limk→+∞(nk)1/k = +∞, i.e. O0,r ( E 0,r {M}) ∩ Λr {N } resp. O0,r ( E 0,r (N ). Then, one has j∞(E 0,r [N ] ( Λr [N ] [N ] ( Λr (M)) ∩ Λr [N ] also). In particular, the Borel map j∞ : E 0,r (and hence j∞(E 0,r surjective. Remark 3.1.5. Theorem 3.1.4 is stronger than only having non-surjectivity of j∞ : E 0,r [N ] −→ Λr since M can be any other quasianalytic weight sequence satisfying N(cid:22)M (i.e. much larger than N ). [N ] −→ Λr [N ] is not [N ] This theorem will follow directly from the Corollary 3.1.2 and the next lemma which gives the existence of sequences satisfying (3.3) in any class Λr [N ]. The proof of this lemma reduces to the argument given in the proof of [26, Theorem 3] with the only difference that the sequence M is replaced by the square root of the sequence (nj)j∈N to treat both the Beurling case and the mixed setting. Lemma 3.1.6. Let M and N be two quasianalytic weight sequences such that supk∈N>0 (nk)1/k = +∞ resp. limk→+∞(nk)1/k = +∞, i.e. O0,1 ( E 0,1 [N ] and a0 ∈ (0, 1] such that (N ). There exists F ∈ Λ1 {N } resp. O0,1 ( E 0,1 lim sup k→+∞ = +∞, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 ωM j,kFj aj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) for all 0 < a ≤ a0. Since it will be useful in the next section, let us recall that such a sequence F can be obtained by setting (3.4) Fkp := √nkp and Fj := 0 otherwise, where the increasing sequence (kp)p∈N of natural numbers is chosen such that (3.5) and (3.6) sup p∈N(cid:0)nkp(cid:1) 1 kp−1Xj=0 (cid:12)(cid:12)(cid:12)ωM kp = +∞ resp. lim p→+∞(cid:0)nkp(cid:1) 1 kp = +∞ j,kp − 1(cid:12)(cid:12)(cid:12)Fj ≤ 1, ∀p ∈ N>0. Remark 3.1.7. As defined in (3.4), the sequence F = (Fj)j∈N can never define a real analytic germ, otherwise Fkp = (nkp )1/2 ≤ Chkp should be satisfied for some C, h > 0 and all p ∈ N, a contradiction to (3.5) in the Roumieu and the Beurling case. HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 9 j or mixed signs are allowed : we meet here a situation not treated Remark 3.1.8. Let us note that it is possible to define a sequence F satisfying the assumption of Lemma 3.1.6 with only non-zero elements by setting Fj := lp j for kp−1 < j < kp, where the values lp j are subjected to some precise growth control. This provides some additional information on sequences not contained in the image of the Borel map. In particular, negative values lp in [20]. However, the construction of F depends on the given M . Even if lp j ≡ 0 for all p, j, as in the original proof of [26, Theorem 3] (or as in Lemma 3.1.6 above), then (3.6) and hence the choice of (kp)p∈N is still depending on M . It is not clear how to get rid of this problem in general. Consequently it seems not possible to prove by using this technique the existence of F = (Fj )j∈N ∈ Λ1 [N ] such that F does not belong to j∞(E 0,1 {M}) for any quasianalytic weight sequence M as it has been done in [20, Thm. 2, Thm. 3] by using completely different methods (and sequences F = (Fj )j∈N such that Fj > 0 for all j ∈ N and not defining a real analytic germ). Since this more general definition of F would neither change nor simplify the proofs of the main results below and would unnecessarily complicate the notation we will work with lacunary sequences as stated above, i.e. lp j ≡ 0 for all p, j. 3.2. Generic size of the image of the Borel map. Let M and N be two (in general different) quasianalytic weight sequences. The aim of this section is to study the size of j∞(E 0,r [N ] using the different notions of genericity presented in the introduction. The results will be obtained by applying Corollary 3.1.2, Remark 3.1.3 and Lemma 3.1.6. [M]) in Λr First, let us concentrate on the Beurling case. Indeed, we intend to study the size of the image from the point of view of Baire genericity (resp. prevalence), for which the underlying space needs In the next Theorem, we prove that to be a Baire space (resp. a complete metrizable space). (N ) \ j∞(E 0,r Λr (N ), hence the image of the Borel mapping defined on any quasianalytic class of M -ultradifferentiable germs is "small" in the space Λr (M))) is a "big" set in Λr {M}) (and hence also Λr (N ) \ j∞(E 0,r (N ). Theorem 3.2.1. Let M and N be two quasianalytic weight sequences. Let us assume that limk→+∞(nk)1/k = +∞, i.e. O0,r ( E 0,r meager in Λr (N ), i.e. Λr (N ) \ j∞(E 0,r {M}) contains a countable intersection of dense open sets. (N ). Then, the set j∞(E 0,r {M}) ∩ Λr (N ) is Note that as a particular case, the choice M ≡ N yields that Λr intersection of dense open sets. Proof. By Lemma 3.1.6, we can consider F ∈ Λr k−1Xj=0 lim sup k→+∞ = +∞ (3.7) ωM (N ) and a0 ∈ (0, 1] such that (M) \ j∞(E 0,r {M}) contains a countable for any 0 < a ≤ a0, where Fj = F(j,0,...,0) for any j ∈ N. Let us fix a sequence (ap)p∈N in (0, a0] decreasing to 0. If the set G is defined by (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j,kFjaj(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−1Xj=0 G = \p∈N   b ∈ Λr (N ) : lim sup k→+∞ j,kbjaj ωM p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = +∞  , 10 C. ESSER AND G. SCHINDL we know from Corollary 3.1.2 and Remark 3.1.3 that (3.8) G ⊆ Λr (N ) \ j∞(E 0,r {M}) . In order to get the result, it suffices then to prove that G can be written as a countable intersection of dense open sets of Λr (N ). One has Let us fix p ∈ N, P ∈ N and k ∈ N>0 and let us show that G(p, P, k) is open. Let us consider a sequence (b(l))l∈N of elements of Λr (N ) which does not belong to G(p, P, k) and which converges to b in Λr (N ), and let us show that G(p, P, k) ωM j,kbjaj p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) > P  . k−1Xj=0 b ∈ Λr G = \p∈N \P ∈N \K∈N>0 [k≥K (N ) :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) G(p, P, k) =  (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 ωM j,kbjaj ≤ P. where (3.9) (3.10) Let δ > 0 be arbitrary and fix ε > 0 such that δ Pk−1 j,knjaj j=0 ωM By assumption, there is L ∈ N such that (cid:12)(cid:12)b(l) − b(cid:12)(cid:12)N 1 ≤ ε for all l ≥ L. Then, for all l ≥ L, one has ε < p . ωM j,kbjaj k−1Xj=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j,kb(l) ωM j aj p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p + P p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j,k(b(l) ωM p(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−1Xj=0 j − bj)aj k−1Xj=0 j − bj(cid:12)(cid:12)aj j,k(cid:12)(cid:12)b(l) ωM k−1Xj=0 ωM j,knjaj ≤ ε ≤ p + P ≤ δ + P, hence (3.9) since δ > 0 is arbitrary. It follows that for any p ∈ N, P ∈ N and K ∈ N>0 [k≥K G(p, P, k) is open, and it remains to prove that it is dense in Λr and let us fix ε > 0. It follows from (3.7) that for all K ∈ N, there is k ≥ K such that (N ). So, let us consider an arbitrary b ∈ Λr (N ) k−1Xj=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j,kFjaj ωM P ε . ≥ p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 11 b − (b ± εF)N (N ) follows. h = εFN h , (cid:3) Then, either b + εF or b − εF belongs to Sk≥K G(p, P, k): Otherwise, one would have ≤ 2P ωM j,k(bj − εFj)aj ωM j,k(bj + εFj)aj k−1Xj=0 p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) which contradicts (3.10). Moreover, for any h > 0, one has k−1Xj=0 p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ωM j,kFj aj 2ε(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) and the density of Sk≥K G(p, P, k) in Λr (3.11) Remark 3.2.2. In particular, from Baire's theorem, one gets that the set Λr in Λr (N ). (N )\ j∞(E 0,r {M}) is dense Let us now show that the previous result also holds in the context of prevalence. Theorem 3.2.3. Let M and N be two quasianalytic weight sequences. Let us assume that limk→+∞(nk)1/k = +∞, i.e. O0,r ( E 0,r Haar-null in Λr (N ). (N ). Then, the set j∞(E 0,r {M}) ∩ Λr (N ) is Proof. We use similar notations as in the proof of Theorem 3.2.1. From (3.8), it suffices to prove that G is prevalent in Λr (N ). We already know that it is a Borel set, since it is a countable intersection of the open sets Sk≥K G(p, P, k). Let us prove that each of these sets is prevalent, hence the result since a countable intersection of prevalent sets is prevalent. We use for a probe the space generated (N ), the line L := (cid:8)b + αF : α ∈ R(cid:9) contains at most one element in the set by F. For any b ∈ Λr (N )\Sk≥K G(p, P, k). Indeed, assume that there exist two different such sequences in L associated Λr with the numbers α, β ∈ R, α 6= β. Then, for all k ≥ K, one would have This contradicts the property (3.7) of F. The conclusion follows. Theorem 3.2.1 and Theorem 3.2.3 mean that the image j∞(E 0,r space Λr {M})∩ Λr (N ) for any given quasianalytic sequences M and N such that limk→+∞(nk)1/k = +∞. (N ) is generically small in the In the Roumieu case Λr {N }, the notions of genericity previously used are not well defined. One can however wonder if the image is also "small" and in what sense. Following Remark 3.2.2, a first direction is to obtain that the complement of the image is dense. A second possibility is to use the notion of lineability. Theorem 3.2.4. Let M and N be two quasianalytic weight sequences. Let us assume that supk∈N>0 (nk)1/k = +∞, i.e. O0,r ( E 0,r {N }. Then Λr {N } (and so Λr Λr (M)) too). {N } \ j∞(E 0,r {N } \ j∞(E 0,r {M}) is dense in (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 hence ωM j,k(bj + αFj )aj ≤ P and ωM j,k(bj + βFj )aj p(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−1Xj=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 ωM j,kFj aj ≤ 2P α − β . p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ P (cid:3) 12 C. ESSER AND G. SCHINDL Proof. We proceed as in the proof of Theorem 3.2.1. Let b ∈ Λr {N } denote a sequence whose restriction Fj = F(j,0...,0), j ∈ N, is given by Lemma 3.1.6. For any ε > 0, either b + εF or b − εF belongs to Λr {M}), hence also F which is impossible by Corollary 3.1.2. Let h1, h2 > 0 be such that {M}). Indeed, otherwise 2εF would belong to j∞(E 0,r {N } and let F ∈ Λr {N } \ j∞(E 0,1 If h = max{h1, h2}, one has FN h < +∞ and bN FN h1 < +∞ and bN h2 < +∞. h < +∞, and h ≤ εFN h . b − (b ± εF)N The conclusion follows. Let us now concentrate on the notion of lineability. (cid:3) Theorem 3.2.5. Let M and N be two quasianalytic weight sequences. Let us assume that supk∈N>0 (nk)1/k = +∞ resp. limk→∞(nk)1/k = +∞ , i.e. O0,r ( E 0,r O0,r ( E 0,r (N ). [N ] \ j∞(E 0,r Then Λr Proof. Let F ∈ Λr (3.4). For any λ > 0, we define the sequence Fλ by setting [N ] denote a sequence whose restriction Fj = F(j,0...,0), j ∈ N, is defined using {M}) is lineable in Λr [N ] \ j∞(E 0,r [N ] (and so Λr (M)) too). {N } resp. F λ 0 = 0 and F λ j = Fj jλ , ∀j ∈ N>0. Let us show that Fλ /∈ j∞(E 0,r {M}). From Corollary 3.1.2, it suffices to prove that lim p→+∞ for every a small enough. Remark that kp−1Xj=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j,kp F λ ωM j aj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = +∞ lim p→+∞ kp−1Xj=0 j,kp F λ ωM j aj = lim p→+∞  p−1Xq=0  Fkq kλ q akq + kp−1Xj=1 (ωM j,kp − 1) Fj jλ aj  for all a small enough. By (3.6), the second term of the sum is bounded uniformly by 1, while the first one is divergent using (3.5). Let S denote the subspace of Λr [N ] spanned by the Fλ, λ > 0: the elements of S can be written as LXl=1 αlFλl for some L ∈ N>0, λ1, . . . , λL ∈ R and α1, . . . , αL ∈ C. Let us prove that S \{0} ⊂ Λ[N ]\ j∞(E 0,r So, let us consider L ∈ N>0, 0 < λ1 < ··· < λL in R, α1, . . . , αL ∈ C \ {0}, and {M}). G = LXl=1 αlFλl. HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 13 For any a ∈ (0, 1) small enough, we have (3.12) lim sup k→+∞ k−1Xj=0 j,kGjaj ≥ lim sup ωM p→+∞ kp−1Xj=0 j,kp Gjaj = lim sup ωM p→+∞ kp−1Xj=0 j,kp Gjaj ωM since Gj = 0 if j /∈ {kp : p ∈ N}. Note that kp−1Xj=0 j,kp Gjaj = ωM (3.13) (ωM j,kp − 1)Gjaj + p−1Xq=0 Gkq akq and the first term of this sum can be bounded as follows uniformly in p kp−1Xj=0 j,kp − 1)Gjaj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kp−1Xj=0 (ωM (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kp−1Xj=0 ≤ j,kp − 1Fj(cid:0)α1 + ··· + αL(cid:1) ≤(cid:0)α1 + ··· + αL(cid:1), ωM by using (3.6). However, the partial sums of the power series +∞Xq=0 Gkq akq cannot be bounded: indeed, otherwise one would have and noting that lim sup q→+∞ (cid:12)(cid:12)Gkq(cid:12)(cid:12) 1 kq < +∞ , lim j→+∞ Gj jλ1 Fj = lim j→+∞ α1 + α2jλ1−λ2 + ··· + αljλ1−λL = α1, this would in turn imply that which is impossible from the choice of F (see (3.4) and (3.5)). Hence, using (3.13), one has 1 kq < +∞ , lim sup q→+∞ (cid:12)(cid:12)Fkq(cid:12)(cid:12) j,kp Gjaj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kp−1Xj=0 ωM lim sup p→+∞ = +∞, and together with (3.12), Corollary 3.1.2 and Remark 3.1.3, it gives that G does not belong to the image of the Borel map j∞(E 0,r (cid:3) {M}). The conclusion follows. Remark 3.2.6. The vector subspace constructed in Theorem 3.2.5 has a maximal dimension which is the dimension of the set {λ : λ > 0} in Λr {M}) is maximal-lineable in Λr [N ]. We say that Λr [N ] \ j∞(E 0,r [N ]. 14 C. ESSER AND G. SCHINDL 4.1. General definitions. 4. The weight matrix case M Definition 4.1.1. A weight matrix M associated with R>0 is a (one parameter) family of sequences M := {M (λ) ∈ RN >0 : λ ∈ R>0}, such that ∀ λ ∈ R>0, M (λ) is a weight sequence and M (λ) ≤ M (κ) (which is equivalent to m(λ) ≤ m(κ)) for all λ ≤ κ, M (λ) p p! p := for p ∈ N. where we have put m(λ) A matrix is called constant if M (λ)≈M (κ) (i.e. M (λ)(cid:22)M (κ) and M (λ)(cid:22)M (κ)) for all λ, κ ∈ R>0. We introduce classes of ultradifferentiable function of Roumieu type E{M} and of Beurling type E(M) as follows, see [24, Section 7] and [19, Section 4.2]. Definition 4.1.2. Let r ∈ N>0 and U ⊆ Rr be non-empty and open. The M-ultradifferentiable classes of Roumieu and Beurling types are defined respectively by and E{M}(U ) := \K⊆U [λ∈R>0 E(M)(U ) := \λ∈R>0 For a compact set K ⊆ Rr, one has the representations E{M (λ)}(K) E(M (λ))(U ). E{M}(K) := lim−→λ∈R>0 lim−→h>0 EM (λ),h(K) and so for U ⊆ Rr non-empty open Similarly we get for the Beurling case E{M}(U ) = lim←−K⊆U lim−→λ∈R>0 lim−→h>0 EM (λ),h(K). E(M)(U ) = lim←−K⊆U lim←−λ∈R>0 lim←−h>0 EM (λ),h(K). Consequently, since the sequences of M are pointwise ordered, E(M)(U ) is a Fréchet space and lim−→λ∈R>0 EM (n),n(K) is a Silva space, i.e. a countable inductive limit of Banach spaces with compact connecting mappings. For more details concerning the locally convex topology in this setting we refer to [19, Section 4.2]. lim−→h>0 EM (λ),h(K) = lim−→n∈N>0 Definition 4.1.3. The spaces of germs at 0 ∈ Rr of the (M)-ultradifferentiable functions of Roumieu and Beurling types are defined respectively by E 0,r {M} := lim−→k∈N>0 E{M}(cid:18)(cid:18)− 1 k , 1 k(cid:19)r(cid:19) , HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 15 and E 0,r (M) := lim−→k∈N>0 E(M)(cid:18)(cid:18)− 1 k , 1 k(cid:19)r(cid:19) . Finally, as done in the case of weight sequences, we introduce the corresponding spaces of sequences, and we endow them with their classical topology. Definition 4.1.4. We introduce the sequence classes of Roumieu type Λr {M} := [λ∈R>0 Λr {M (λ)} =nb ∈ CNr : ∃ λ ∈ R>0 ∃h > 0,bM (λ) h and of Beurling type Λr (M) := \λ∈R>0 Λr (M (λ)) =nb ∈ CNr : ∀ λ ∈ R>0 ∀h > 0,bM (λ) h < +∞o , < +∞o . Using notations similar as before, the Borel map j∞ is defined in the weight matrix case by j∞ : E 0,r [M] −→ Λr [M], j∞(f ) =(cid:18) ∂αf (0) α! (cid:19)α∈Nr . In [25, Theorem 4.1], the following result has been obtained (under slightly more general assump- tions on M and using regularizations of M (λ)). Theorem 4.1.5. Let M = {M (λ) : λ ∈ R>0} be a weight matrix. quasianalytic. (i) E{M} is non-quasianalytic if and only if there exists λ0 ∈ R>0 such that E[M (λ0 )] is non- (ii) E(M) is non-quasianalytic if and only if each E[M (λ)] is non-quasianalytic. This result yields and motivates the following definition, see also [20, Section 5.1]. Definition 4.1.6. A weight matrix M is called quasianalytic if for all λ ∈ R>0 the sequence M (λ) is quasianalytic, which means (4.1) ∀ λ ∈ R>0, +∞Xj=1 1 j (cid:17)1/j = +∞. (cid:16)M (λ) In this case both classes E{M} and E(M) and all classes E{M (λ)} resp. E(M (λ)) are quasianalytic too, see Proposition 2.2.4. For the Beurling case E(M) it would be enough to require only that there is some M (λ0) which is quasianalytic since then M (λ) for all λ ≤ λ0 is quasianalytic too and since, by definition of the Beurling type classes, the spaces remain unchanged if we remove from M all (possible non-quasianalytic sequences) M (λ) for λ > λ0. Recently, in [20, Thm. 5, Thm. 6], it has been shown that j∞ restricted to the germs E 0,r [M] can never be onto the corresponding sequence space for any quasianalytic weight matrix M such that O0,r ( E 0,r [M]. 16 C. ESSER AND G. SCHINDL 4.2. Generalization of Thilliez's proof for non-surjectivity. We generalize [26, Theorem 3] resp. Theorem 3.1.4 to the general weight matrix setting (for both types). To ensure O0,r ( E 0,r resp. O0,r ( E 0,r ∀ λ ∈ R>0, (M) we assume ∀ λ ∈ R>0, = +∞, = +∞ k→+∞(cid:16)m(λ) k∈N>0(cid:16)m(λ) k (cid:17)1/k k (cid:17)1/k resp. sup lim {M} k (cid:17)1/k case, one could assume that supk∈N>0(cid:16)m(λ) e.g. see [20, Section 5] and which follows from [19, Proposition 4.6]. Note that in the Roumieu = +∞ only for all λ ≥ λ0 for some λ0 ∈ R>0 (large): Indeed, one can skip in this case all small sequences in the matrix without changing the ultradifferentiable class. Let us start with a generalization of Lemma 3.1.6 working with sequence spaces defined via weight matrices. Lemma 4.2.1. Let M be a quasianalytic weight sequence and N = {N (λ) : λ ∈ R>0} be a quasianalytic weight matrix such that supk∈N>0(cid:16)n(λ) k (cid:17)1/k = +∞ {N } resp. O0,1 ( E 0,1 for all λ ∈ R>0, i.e. O0,1 ( E 0,1 (N ). There exists F ∈ Λ1 [N ] and a0 ∈ (0, 1] such that j,kFj aj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) for all 0 < a ≤ a0. Proof. Note first that the Roumieu case follows immediately from Lemma 3.1.6: indeed, it suffices to fix N (λ0) ∈ N and to use the inclusion Λ1 {N }. So, let us concentrate on the Beurling case. Let 0 < a0 ≤ 1 be arbitrary but from now on fixed. Let us show that there is a strictly increasing sequence (kp)p∈N with k0 ≥ 1 such that the sequence F defined by = +∞ resp. limk→+∞(cid:16)n(λ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 {N (λ0)} ⊆ Λ1 k (cid:17)1/k lim sup k→+∞ = +∞, ωM kp Fkp :=rn(cid:0) 1 p+1(cid:1) kp−1Xj=0 (cid:12)(cid:12)(cid:12)ωM j,kp − 1(cid:12)(cid:12)(cid:12)Fjaj 1 kp kp F lim p→+∞ = +∞. (4.2) satisfies (4.3) and (4.4) and Fj := 0 otherwise, 0 ≤ 1, ∀p ∈ N>0, It is easy to show this by induction, since limk→+∞ ωM been constructed, one may choose kp sufficiently large in order to have both (4.3) and j,k = 1 for all j ∈ N by (3.1). So if kp−1 has =vuut(cid:18)n(cid:0) 1 p+1(cid:1) (cid:19) 1 ≥ p. k (cid:17)1/k This is possible by using the assumption limk→+∞(cid:16)n(λ) = +∞ for all λ ∈ R>0 and guarantees 1 kp kp F kp kp (4.4). HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 17 (N ) follows easily from the definition and the assumption of the lemma. More- That F belongs to Λ1 over, for all a ∈ (0, a0], one has kp−1Xj=0 j,kp Fj aj = ωM p−1Xq=0 Fkq akq + kp−1Xj=0 (cid:0)ωM j,kp − 1(cid:1)Fjaj which implies by (4.3) and (4.4). lim sup k→+∞ k−1Xj=0 ωM j,kFjaj = ∞ (cid:3) kp wise, Fkp =qn(1/(p+1)) Remark 4.2.2. As defined in (4.2), the sequence F can never define a real analytic germ: Other- ≤ Chkp should be satisfied for some C, h > 0 and all p ∈ N, a contradiction to (4.4). The Roumieu case follows immediately by Lemma 3.1.6 and (3.5) for N (λ0). Moreover, as commented in Remark 3.1.8, in the proof of Lemma 4.2.1 the sequence F could be defined by setting Fj := lp j subjected to some precise growth control. This provides some additional information on sequences not contained in the image of the Borel map but is not necessary for the forthcoming proofs. j for kp−1 < j < kp for non-zero values lp By proceeding as in the weight sequence case, Corollary 3.1.2 directly gives the following proposition. Proposition 4.2.3. Let M be a quasianalytic weight sequence and N = {N (λ) : λ ∈ R>0} be a quasianalytic weight matrix such that supk∈N>0(cid:16)n(λ) = +∞ {N } resp. O0,r ( E 0,r for all λ ∈ R>0, i.e. O0,r ( E 0,r {M}) ∩ Λr = +∞ resp. limk→+∞(cid:16)n(λ) (N ). Then, one has k (cid:17)1/k k (cid:17)1/k j∞(E 0,r [N ] ( Λr [N ] [N ] also). [N ] ( Λr (M)) ∩ Λr (and hence j∞(E 0,r Note that in order to get the non-surjectivity of the Borel map in the weight matrix case, we need to get an equivalent of Proposition 4.2.3 working only with quasianalytic weight matrices (and not with a weight sequence). This can be obtained thanks to the following result. Proposition 4.2.4. Let M = {M (λ) : λ ∈ R>0} be a quasianalytic weight matrix. Then there exists a quasianalytic weight sequence L satisfying M (λ)⊳L for all λ ∈ R>0, i.e. E{M} ⊆ E(L) holds true. The aim is to construct a quasianalytic (weight) sequence L lying (strictly) above M by applying some diagonal technique. Unfortunately, it seems that such a construction does not preserve the log-convexity; we can overcome this problem by working with regularizations of L and by applying Proposition 2.2.4. The following idea is motivated by the proof of [25, Prop. 4.7 (i)]. Proof. Let (di)i∈N>0 be a strictly increasing sequence in R, with d1 ≥ 1 and tending to infinity as i → +∞. By the assumptions on M there exists a strictly increasing sequence (ji)i∈N>0 (in N) with j1 = 1 and such that Pji+1−1 )1/j ≥ di, see (4.1). According to this sequence, we put 1 (M (i) j=ji j eL0 = 1 and eLj = dj i M (i) j , if ji ≤ j < ji+1, ∀i ∈ N>0. 18 C. ESSER AND G. SCHINDL j j = d−j ≤ M (i) p (cid:17)1/p for all λ ∈ R>0, we obtain limp→+∞(elp)1/p = +∞ (where elp = First, for any given index λ0 ∈ R>0 (large), we have M (λ0) i eLj for all ji ≤ j < ji+1 and i ≥ λ0, hence M (λ)⊳eL for all λ ∈ R>0 because di → +∞ as i → +∞. So E{M (λ0 )} ⊆ E( eL) and E{M} ⊆ E( eL) follows by definition as a special case by [19, Prop. 4.6 (2)]. Unfortunately we do not see directly if eL is log convex but since di → +∞ as i → +∞, and since 0 < lim inf p→+∞(cid:16)m(λ) Consequently E[ eL] = E[ eLlc], i.e. eL can be replaced by its log-convex minorant for both cases, see Remark 2.2.5, and so M (λ)⊳eLlc for all λ ∈ R>0 too. It remains to show that (Q) holds true for eLlc. By definition of eL, the log-convexity of each M i and since (di)i is (strictly) increasing, we have that j 7→ (eLj)1/j is increasing: If j = ji+1, then for all i ≥ 1, one has (eLji+1−1)1/(ji+1−1) = di(M (i) ji+1 )1/ji+1 = (eLji+1)1/ji+1 . So we have shown eL = eLI and finally ji+1−1Xj=ji ji+1−1)1/(ji+1−1) ≤ di(M (i) ji+1 )1/ji+1 ≤ di+1(M (i+1) The remaining cases are clear. 1 di(M (i) j )1/j ≥ ji+1−1Xj=ji 1 = +∞, 1 j )1/j +∞Xj=1 +∞Xi=1 +∞Xi=1 +∞Xi=1 eLp p! ). = = 1 Propositions 4.2.3 and 4.2.4 imply directly the next main result. by the choice of (ji)i∈N>0 above. Hence by Proposition 2.2.4 we get (Q) for eLlc and the conclusion follows by taking L := eLlc. Theorem 4.2.5. Let M and N be two quasianalytic weight matrices such that supk∈N>0(cid:16)n(λ) +∞ resp. limk→+∞(cid:16)n(λ) {N } resp. O0,r ( E 0,r = +∞ for all λ ∈ R>0, i.e. O0,r ( E 0,r k (cid:17)1/k k (cid:17)1/k (N ). Then, (cid:3) = one has (eLI (eLj)1/j j∞(E 0,r {M}) ∩ Λr [N ] ( Λr [N ] (and hence j∞(E 0,r surjective. (M)) ∩ Λr [N ] ( Λr [N ] also). In particular, the Borel map j∞ : E 0,r [N ] −→ Λr [N ] is not 4.3. Generic size of the image of the Borel map. Thanks to Theorem 4.2.5, we can immedi- ately transfer and generalize all central statements from Section 3.2 to the weight matrix case. Theorem 4.3.1. Let N and M be two quasianalytic weight matrices. Assume that limk→+∞(cid:16)n(λ) +∞ for all λ ∈ R>0, i.e. O0,r ( E 0,r in Λr (N ). Then the image of the Borel map j∞(E 0,r {M})∩Λr (N ) is meager k (cid:17)1/k (N ). = Proof. By Proposition 4.2.4, we know that there is a quasianalytic weight sequence L such that E 0,r {M} ⊂ E 0,r (N ). By Lemma 4.2.1, we can consider F ∈ Λr (N ) is meager in Λr (L) ⊂ E 0,r {L}) ∩ Λr {L}. So, it suffices to prove that the larger set j∞(E 0,r (N ) and a0 ∈ (0, 1] such that (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 j,kFj aj(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = +∞, lim sup k→+∞ ωL HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 19 for all 0 < a ≤ a0, where Fj = F(j,0,...,0) for any j ∈ N. As done in the proof of Theorem 3.2.1 and using Corollary 3.1.2, if (ap)p∈N is a fixed sequence of (0, a0] which decreases to 0, it suffices to prove that the set p(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) > P    (N ) :(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k−1Xj=0 G = \p∈N \P ∈N \K∈N0 [k≥K ωL j,kbjaj (N ). Since the inclusion Λr b ∈ Λr is a countable intersection of dense open sets of Λr (N (λ)) is continuous for any λ ∈ R>0, and from the proof of Theorem 3.2.1, it is clear that G is a countable intersection of open sets. We obtain the density of these sets noting that the equality (3.11) holds true for all h > 0 and all N (λ), λ ∈ R>0. (cid:3) Similarly, we get the generalization of Theorem 3.2.3 to the matrix setting. Theorem 4.3.2. Let M and N be two quasianalytic weight matrices. Assume that limk→+∞(cid:16)n(λ) +∞ for all λ ∈ R>0, i.e. O0,r ( E 0,r null in Λr (N ). Then the image of the Borel map j∞(E 0,r (N ) ֒→ Λr {M})∩ Λr (N ) is Haar- (N ). k (cid:17)1/k (cid:3) k (cid:17)1/k {N } \ j∞(E 0,r {M}) is dense in Λr {N }. {N }. Then, the set Λr Proof. As done before, using Proposition 4.2.4, we can reduce the proof to the case where the weight matrix M is constant. We follow then the lines of the proof of Theorem 3.2.3, where the set G is defined as in the proof of Theorem 4.3.1. Moreover, in the Roumieu case, we have the following result. Theorem 4.3.3. Let M and N be two quasianalytic weight matrices. Assume that supk∈N>0(cid:16)n(λ) +∞ for all λ ∈ R>0, i.e. O0,r ( E 0,r Proof. Again, using Proposition 4.2.4, it suffices to consider the case where the weight matrix M is constant. In the proof of Theorem 3.2.4, let h1, h2 > 0 and λ1, λ2 > 0 be such that FN (λ1) < +∞ and bN (λ2) < +∞. Then we put κ := max{λ1, λ2} and again h := max{h1, h2} to get both FN (κ) Finally, we can obtain an equivalent of these results in the context of lineability. Theorem 4.3.4. Let M and N be two quasianalytic weight matrices. Assume that supk∈N>0(cid:16)n(λ) +∞ for all λ ∈ R>0 resp. O0,r ( E 0,r Proof. As previously, we consider the case where the weight matrix M is constant and we follow simply the proof of Theorem 3.2.5. We close this section with the following observation. = +∞ for all λ ∈ R>0, i.e. O0,r ( E 0,r limk→+∞(cid:16)n(λ) [N ] \ j∞(E 0,r < +∞. The conclusion follows. < +∞ and bN (κ) {M}) is lineable in Λr (N ). Then, the set Λr [N ] \ j∞(E 0,r (M)) too). k (cid:17)1/k [N ] (and so Λr h h (cid:3) (cid:3) {N } resp. h1 h2 k (cid:17)1/k Remark 4.3.5. We have used the proofs from the single weight sequence case of Section 3.2 and transferred them to the more general weight matrix case of this Section 4.3. Alternatively, one could start directly with the weight matrix setting (and give the proofs from Section 3.2 in this general approach) and then obtain the single weight sequence case as an immediate consequence for the constant matrix M = {M}. = = = 20 C. ESSER AND G. SCHINDL 5. The weight function case 5.1. General definitions. In this last part, we will study classes of ultradifferentiable functions defined using weight functions in the sense of Braun-Meise-Taylor, see [9]. As we will see, this case can be reduced to the weight matrix case. First, let us start by recalling the basic definitions. Definition 5.1.1. A function ω : [0, +∞) → [0, +∞) is called a weight function if (i) ω is continuous, (ii) ω is increasing, (iii) ω(t) = 0 for all t ∈ [0, 1] (normalization, w.l.o.g.), (iv) limt→+∞ ω(t) = +∞. In this case, we say that ω has (ω0). Classical additional conditions can be imposed on the considered weight functions. More precisely, let us define the following conditions: (ω1) ω(2t) = O(ω(t)) as t → +∞, (ω2) ω(t) = O(t) as t → +∞, (ω3) log(t) = o(ω(t)) as t → +∞ (⇔ limt→+∞ (ω4) ϕω : t 7→ ω(et) is a convex function on R, (ω5) ω(t) = o(t) as t → +∞. ϕω(t) = 0), t For convenience, we define the set W := {ω : [0, +∞) → [0, +∞) : ω has (ω0), (ω1), (ω3), (ω4)}. Note that (ω2) is sometimes also considered as a general assumption on ω (e.g. see [20, Sect. 4.1]) and note also that (ω5) implies (ω2). For ω ∈ W, we define the Legendre-Fenchel-Young-conjugate of ϕω by ϕ∗ ω(x) := sup{xy − ϕω(y) : y ≥ 0}, x ≥ 0. Definition 5.1.2. Let r ∈ N>0, U ⊆ Rr be a non-empty open set and ω ∈ W. The ω-ultradifferentiable Roumieu type class is defined by E{ω}(U ) := {f ∈ E(U ) : ∀ K ⊆ U compact ∃ l > 0, kfkω K,l < +∞}, and the ω-ultradifferentiable Beuling type class by E(ω)(U ) := {f ∈ E(U ) : ∀ K ⊆ U compact ∀ l > 0, kfkω K,l < +∞}, where we have put kfkω K,l := sup α∈Nr ,x∈K ∂αf (x) l ϕ∗ ω(lα)) exp( 1 . As done in the previous contexts, these spaces are endowed with their natural topologies. Analo- {ω} and E 0,r gously as in the sections above, we also consider the spaces of germs at 0, denoted E 0,r (ω), and the associated spaces of complex sequences Λr (ω). Again, we endow these spaces with their natural topology: Λ{ω} is an (LB)-space and Λ(ω) a Fréchet space. In this setting, the Borel map is given by {ω} and Λr j∞ : E 0,r [ω] −→ Λr [ω], j∞(f ) =(cid:18) ∂αf (0) α! (cid:19)α∈Nr . HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 21 As pointed out in [20, Section 4.2], that to ensure Cω ( E{ω} resp. Cω ( E(ω), one has to assume that lim inf t→+∞ ω(t) t = 0 resp. ω(t) = o(t) as t → +∞, i.e. (ω5), which follows from the characterizations given in [19, Lemm. 5.16, Cor. 5.17] and the fact that the weight ω(t) = t (up to equivalence) defines the class Cω. Moreover, in the present setting, the definition of quasianalyticity takes the following form. Definition 5.1.3. A weight function is called quasianalytic if it satisfies (ωQ) Z +∞ 1 ω(t) t2 dt = +∞. 5.2. Generic size of the image of the Borel map. Naturally, one could wonder if the results of the previous sections concerning the Borel map still hold in the context of weight functions, and if the proofs will require some different techniques and methods in this setting. But we will see that they can be obtained without any additional work! Applying the idea presented below has already been helpful in [20] where a closely related topic has been treated. In [24] and [19, Section 5], a matrix Ω := {W (l) = (W (l) ω ∈ W: This matrix is defined by ω(lj)(cid:19) , W (l) := exp(cid:18) 1 ϕ∗ l j ∀j ∈ N, ∀l > 0, j )j∈N : l > 0} has been associated with each and E[ω] = E[Ω] holds as locally convex vector spaces. Moreover, the following results have been obtained: (i) Each W (l) satisfies the basic assumptions (I) and (II). (ii) ω has in addition (ω2) if and only if some/each W (l) has (III), too. So each W (l) ∈ Ω is a weight sequence according to the requirements from Section 2.2, provided ω ∈ W has (ω2). Moreover, by [19, Corollary 5.8] and [25, Corollary 4.8], one has that the following assertions are equivalent: (i) ω ∈ W is quasianalytic, (ii) Ω is quasianalytic in the sense of Definition 4.1.6, (iii) some/each W (l) satisfies (Q). Similarly, from [20, Proposition 2] (and in the same spirit as in [19, Section 5]), one knows that Λr [Ω] as locally convex spaces, too. [ω] = Λr Consequently, under the assumptions described above, we are able to apply the results from Section 4.3 to the matrix N ≡ Ω, using the sequence L from Proposition 4.2.4 lying above the matrix M ≡ Σ which is associated with a given (arbitrary) quasianalytic weight function σ ∈ W. Theorem 5.2.1. • Let ω ∈ W be a quasianalytic weight function satisfying (ω5). Then, for any quasianalytic weight function σ ∈ W, the set j∞(E 0,r (ω) is meager and Haar-null in Λr (ω). • Let ω ∈ W be a quasianalytic weight function satisfying (ω2) and lim inf t→+∞ {σ}) ∩ Λr ω(t) t = 0. {σ}) is dense in {ω} \ j∞(E 0,r Then, for any quasianalytic weight function σ ∈ W, the set Λr {ω} (and so Λr Λr (σ)) too). {ω} \ j∞(E 0,r 22 C. ESSER AND G. SCHINDL • Let ω ∈ W be a quasianalytic weight function satisfying (ω2) and lim inf t→+∞ (σ)) too). ω(t) t = 0 in the Roumieu resp. (ω5) in the Beurling case. Then, for any quasianalytic weight function σ ∈ W, the set Λr [ω] \ j∞(E 0,r [ω] \ j∞(E 0,r {σ}) is lineable in Λr [ω] (and so Λr Acknowledgement. The authors wish to thank the referee for his useful comments and suggestions which have improved the presentation and the structure of this work. C. Esser is supported by a F.R.S. -FNRSgrant; G. Schindl is supported by FWF-Project J3948- N35, as a part of which he is an external researcher at the Universidad de Valladolid (Spain) for the period October 2016 - September 2018. References [1] R.M. Aron, V.I. Gurariy, and J.B. Seoane-Sepúlveda. Lineability and spaceability of sets of functions on R. Proc. Amer. Math. Soc., 133(3):795 -- 803, 2005. [2] L. Bernal-González, D. Pellegrino, and J.B. Seoane-Sepúlveda. Linear subsets of nonlinear sets in topological vector spaces. Bull. Amer. Math. Soc., 51(1):71 -- 130, 2014. [3] A. Beurling. Quasi-analyticity and general distributions. Lecture 4 and 5, AMS Summer Institute, Stanford, 1961. [4] J. Bonet and R. Meise. On the theorem of Borel for quasianalytic classes. Math. Scand., 112(2):302 -- 319, 2013. [5] J. Bonet, R. Meise, and S. N. Melikhov. A comparison of two different ways to define classes of ultradifferentiable functions. Bull. Belg. Math. Soc. Simon Stevin, 14:424 -- 444, 2007. [6] J. Bonet, R. Meise, and B. A. Taylor. Whitney's extension theorem for ultradifferentiable functions of Roumieu type. Proc. Roy. Irish Acad. Sect. A, 89(1):53 -- 66, 1989. [7] J. Bonet, R. Meise, and B. A. Taylor. On the range of the Borel map for classes of nonquasianalytic functions. In Progress in functional analysis (Peñíscola, 1990), volume 170 of North-Holland Math. Stud., pages 97 -- 111. North-Holland, Amsterdam, 1992. [8] E. Borel. Sur quelques points de la théorie des fonctions. Ann. Sci. Ecole Norm. Sup., 12:9 -- 55, 1895. [9] R. W. Braun, R. Meise, and B. A. Taylor. Ultradifferentiable functions and Fourier analysis. Results Math., 17(3-4):206 -- 237, 1990. [10] T. Carleman. Sur le calcul effectif d'une fonction quasi analytique dont on donne les dérivées en un point. C. R. Acad. Sci. Paris, 176:59 -- 68, 1923. [11] T. Carleman. Les fonctions quasi analytiques. Collection Borel, Gauthier-Villars, Paris, 1926. [12] J.P.R. Christensen. Topology and Borel structure. North-Holland, Amsterdam, 1974. [13] C. Esser. Generic results in classes of ultradifferentiable functions. J. Math. Anal. Appl., 413(1):378 -- 391, 2014. [14] L. Hörmander. The analysis of linear partial differential operators I, Distribution theory and Fourier analysis. Springer-Verlag, 2003. [15] B.R. Hunt, T. Sauer, and J.A. Yorke. Prevalence: a translation-invariant "almost every" on infinite-dimensional spaces. Bull. Amer. Math. Soc. (N.S.), 27(2):217 -- 238, 1992. [16] H. Komatsu. Ultradistributions. I. Structure theorems and a characterization. J. Fac. Sci. Univ. Tokyo Sect. IA Math., 20:25 -- 105, 1973. [17] H.-J. Petzsche. On E. Borel's theorem. Math. Ann., 282(2):299 -- 313, 1988. [18] H.-J. Petzsche and D. Vogt. Almost analytic extension of ultradifferentiable functions and the boundary values of holomorphic functions. Math. Ann., 267:17 -- 35, 1984. [19] A. Rainer and G. Schindl. Composition in ultradifferentiable classes. Studia Mathematica, 224(2):97 -- 131, 2014. [20] A. Rainer and G. Schindl. On the Borel mapping in the quasianalytic setting. Math. Scand., 121(2):293 -- 310, 2017. [21] L. Rodino. Linear Partial Differential Operators in Gevrey Spaces. Word Sci. London, 1993. [22] W. Rudin. Real and complex analysis. 3rd edition, McGraw-Hill Book Company, New York, 1987. [23] G. Schindl. Spaces of smooth functions of Denjoy-Carleman-type, 2009. Diploma Thesis, Universität Wien, available online at http://othes.univie.ac.at/7715/1/2009-11-18_0304518.pdf. [24] G. Schindl. Exponential laws for classes of Denjoy-Carleman-differentiable mappings, 2014. PhD Thesis, Uni- versität Wien, available online at http://othes.univie.ac.at/32755/1/2014-01-26_0304518.pdf. [25] G. Schindl. Characterization of ultradifferentiable test functions defined by weight matrices in terms of their Fourier transform. Note di Matematica, 36(2):1 -- 35, 2016. HOW FAR IS THE BOREL MAP FROM BEING SURJECTIVE? 23 [26] V. Thilliez. On quasianalytic local rings. Expo. Math., 26:1 -- 23, 2008. C. Esser: Université de Liège, Département de Mathématique, Quartier Polytech 1, Allée de la Découverte 12, Bâtiment B37, B-4000 Liège, Belgique E-mail address: [email protected] G. Schindl: Departamento de Álgebra, Análisis Matemático, Geometría y Topología, Universidad de Valladolid, Facultad de Ciencias, Paseo de Belén 7, 47011 Valladolid, Spain. E-mail address: [email protected]
1508.04891	3	1508	2015-09-08T13:28:53	A minimax theorem in infinite-dimensional topological vector spaces	[ "math.FA" ]	In this paper, we obtain a minimax theorem by means of which, in turn, we prove the following result: Let $E$ be an infinite-dimensional reflexive real Banach space, $T:E\to E$ a non-zero compact linear operator, $\varphi:E\to {\bf R}$ a lower semicontinuous, convex and coercive functional, $I\subset {\bf R}$ a compact interval, with $0\in I$, $\psi:I\to {\bf R}$ a lower semicontinuous convex function. Then, for each $r>\varphi(0)$, one has $$\sup_{x\in X}\inf_{\lambda\in I}(\varphi(T(x)-\lambda x)+\psi(\lambda))=r+\psi(0)\ ,$$ where $$X=\{x\in E : \varphi(T(x))\leq r\}\ .$$	math.FA	math	A minimax theorem in infinite-dimensional topological vector spaces BIAGIO RICCERI Abstract: In this paper, we obtain a minimax theorem by means of which, in turn, we prove the following result: Let E be an infinite-dimensional reflexive real Banach space, T : E → E a non-zero compact linear operator, ϕ : E → R a lower semicontinuous, convex and coercive functional, I ⊂ R a compact interval, with 0 ∈ I, ψ : I → R a lower semicontinuous convex function. Then, for each r > ϕ(0), one has sup x∈X inf λ∈I (ϕ(T (x) − λx) + ψ(λ)) = r + ψ(0) , where X = {x ∈ E : ϕ(T (x)) ≤ r} . 5 1 0 2 p e S 8 ] . A F h t a m [ 3 v 1 9 8 4 0 . 8 0 5 1 : v i X r a Key words: Minimax theorem; inf-compactness; lower semicontinuity; connectedness; sequential weak topology; compact linear operator. 2010 Mathematics Subject Classifications: 49J35; 49K35; 46A55; 46B10; 47A10; 52A41. Let E be a topological space and X a non-empty subset of E. A function f : X → R is said to be relatively inf-compact (resp. relatively sequentially inf-compact) in E, provided that, for each r ∈ R, the sub-level set f −1(] − ∞, r]) is relatively compact (resp. sequentially relatively compact) in E, that is its closure in E is compact (resp. sequentially compact). A real-valued function f on a convex subset of a vector space is said to be quasi-convex if, for each r ∈ R, the set f −1(] − ∞, r]) is convex. The aim of this very short note is to highlight the following minimax result: THEOREM 1. - Let E be a real Hausdorff topological vector space and let X ⊆ E be an infinite- dimensional convex set whose interior in its closed affine hull is non-empty. Moreover, let I ⊂ R be a compact interval and f : X × I → R a function which is lower semicontinuous in X × I and quasi-convex in I. Finally, assume that there is a set D ⊂ I, dense in I, such that, for each λ ∈ D, the function f (·, λ) is relatively inf-compact (resp. relatively sequentially inf-compact) in E . Then, one has sup x∈X inf λ∈I f (x, λ) = inf λ∈I sup x∈X f (x, λ) . Theorem 1 can be qualified as unconventional in the sense that, in most of the known minimax theorems, lower semicontinuity and inf-compactness are related to the variable with respect to which one takes the inf, while it is quasi-concavity that one generally assumes with respect to the other variable (see, for instance, [4]). It is natural to ask whether the two assumptions made on the convex set X are necessary. We start just presenting two examples related to such a question. The first example concerns the infinite-dimensionality of X. 1 EXAMPLE 1. - Let E be a finite-dimensional normed space and let f : E × [0, 1] → R be the function defined by for all (x, λ) ∈ E × [0, 1]. f (x, λ) = kxk − λ(kxk2 + 1) Of course, f is convex in [0, 1], inf-compact in E and, just because dim(E) < ∞, continuous in E × [0, 1]. kxk2+1 , we have λ ∈ [0, 1] and f (x, λ) = 0. This implies Further, notice that, for each x ∈ E, taking λ = kxk that On the other hand, we clearly have sup x∈E inf λ∈[0,1] f (x, λ) = 0 . inf λ∈[0,1] sup x∈E f (x, λ) = +∞ . So, the conclusion of Theorem 1 can fail if X if finite-dimensional. The second example deals with the non-emptyness of the interior of X in its closed affine hull. EXAMPLE 2. - Let E be an infinite-dimensional reflexive real Banach space, let X be the open unit ball in E and let ϕ ∈ E∗, with kϕkE ∗ = 1. Consider the function f : X × [0, 1] → R defined by for all (x, λ) ∈ X × [0, 1]. 1 1 − ϕ(x) − λ (cid:18) 1 1 − ϕ(x)(cid:19)2 f (x, λ) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) + 1!(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Consider E equipped with the weak topology. Clearly, the affine hull of X is the whole E and, since dim(E) = ∞, the interior of X in the weak topology is empty. Since, by reflexivity, X is relatively weakly compact, the function f is relatively weakly inf-compact in E. Since ϕ ∈ E∗, f is weakly continuous in X × [0, 1], besides being convex in [0, 1]. As in Example 1, it is seen that and sup x∈X inf λ∈[0,1] f (x, λ) = 0 inf λ∈[0,1] sup x∈X f (x, λ) = +∞ . So, the conclusion of Theorem 1 can fail if the interior of X in its closed affine hull is empty. Our proof of Theorem 1 is fully based on the joint use of three previous results of ours. We now recall them. THEOREM A ([2], Proposition 3). - Let E be a real Hausdorff topological vector space, let X ⊆ E be an infinite-dimensional convex set whose interior in its closed affine hull is non-empty and let K ⊆ E be a relatively compact (resp. relatively sequentially compact) set. Then, the set X \ K is connected. REMARK 1. - Notice that, in [2], such a result was proved for the relatively compact case only. The same proof shows the validity of the result also in the relatively sequentially compact case, in view of the fact that any Hausdorff topological vector space possessing a sequentially compact neighbourhood of 0 is finite-dimensional. THEOREM B ([3], Proposition 5.3) - Let X, Y be two topological spaces, with X connected, and let F : X → 2Y be a lower semicontinuous multifunction with non-empty values. Assume the set is dense in X. {x ∈ X : F (x) is connected} 2 Then, the set is connected. {(x, y) ∈ X × Y : y ∈ F (x)} For a generic set S ⊆ X × I, for each (x, λ) ∈ X × I, we set and Sx = {µ ∈ I : (x, µ) ∈ S} Sλ = {u ∈ X : (u, λ) ∈ S} . THEOREM C ([1], Theorem 2.3). - Let X be a topological space, I ⊆ R a compact interval and S, T ⊆ X × I. Assume that S is connected and Sλ 6= ∅ for all λ ∈ I, while T is closed and Tx is non-empty and connected for all x ∈ X. Then, one has S ∩ T 6= ∅. Proof of Theorem 1. Arguing by contradiction, assume that Fix ρ satisfying and put and sup x∈X inf λ∈I f (x, λ) < inf λ∈I sup x∈X f (x, λ) . sup x∈X inf λ∈I f (x, λ) < ρ < inf λ∈I sup x∈X f (x, λ) (1) S = {(x, λ) ∈ X × I : f (x, λ) > ρ} T = {(x, λ) ∈ X × I : f (x, λ) ≤ ρ} . Since f is lower semicontinuous, the set T is closed. Moreover, for each x ∈ X, the set Tx is non-empty by (1) and connected by the quasi-convexity of f (x, ·). By (1) again, Sλ 6= ∅ for all λ ∈ I. Fix λ ∈ D. Since Sλ = X \ {x ∈ X : f (x, λ) ≤ ρ} and {x ∈ X : f (x, λ) ≤ ρ} is relatively compact (resp. relatively sequentially compact) in E, in view of Theorem A, the set Sλ turns out to be connected. On the other hand, since Sx is open for all x ∈ X, the multifunction λ → Sλ is lower semicontinuous in I. At this point, we can apply Theorem B to realize that the set {(λ, x) ∈ I × X : (x, λ) ∈ S} is connected. But such a set is clearly homeomorphic to S, and so S is connected. As a consequence, each assumption of Theorem C is satisfied, and hence we would have S ∩ T 6= ∅ which is clearly false. Such a △ contradiction completes the proof. We conclude with the following application of Theorem 1. We first introduce a notation. Namely, if Y is a topological space and τ is the topology of Y , we denote by τs the topology on Y whose members are the sequentially open subsets of Y . Let us recall that a set A ⊆ Y is said to be sequentially open if, for every sequence {yn} in Y converging to a point of A, there is ν ∈ N such that yn ∈ A for all n ≥ ν. A functional ϕ on a real normed space is said to be coercive if limkxk→+∞ ϕ(x) = +∞. THEOREM 2. - Let E be an infinite-dimensional reflexive real Banach space, T : E → E a non-zero compact linear operator, ϕ : E → R a lower semicontinuous, convex and coercive functional, I ⊂ R a compact interval, with 0 ∈ I, ψ : I → R a lower semicontinuous convex function. Then, for each r > ϕ(0), one has sup x∈X inf λ∈I (ϕ(T (x) − λx) + ψ(λ)) = r + ψ(0) , 3 where X = {x ∈ E : ϕ(T (x)) ≤ r} . PROOF. By a classical result, ϕ turns out to be continuous with respect to the strong topology of E. Since T (E) is a non-zero linear subspace, the set ϕ(T (E)) is unbounded above. Indeed, if not, ϕ would be constant on T (E), contrary to the coercivity of ϕ. As a consequence, since T (E) is connected, we have From this, we clearly infer that ϕ(T (E)) =(cid:18) inf T (E) ϕ, +∞(cid:20) . ϕ(T (x)) = r . sup x∈X (2) Next, consider the function f : X × I → R defined by f (x, λ) = ϕ(T (x) − λx) + ψ(λ) for all (x, λ) ∈ X × I. Now, denote by τ the weak topology of E. Notice that T , being linear and compact, turns out to be sequentially continuous from E with the topology τ to E with the strong topology. It is easy to check that this is equivalent to the continuity of T from E with the topology τs to E with the strong topology. Of course, (E, τs) is a Hausdorff topological vector space. Now, we are going to apply Theorem 1 to the function f considering E with the topology τs. First, notice that the set X is convex and its interior in τs is non-empty. Actually, X contains the non-empty set T −1(ϕ−1(] − ∞, r[)) which is open in τs, by the remarks above. Next, observe that, for each λ ∈ R, the function x → T (x) − λx, being continuous and linear, is continuous from the weak to the weak topology, and so, a fortiori, from the τs to the weak topology. Of course, this implies that the function (x, λ) → T (x) − λx is continuous from the product of τs and the topology of R to the weak topology. But then, since ϕ is weakly lower semicontinuous, the function f is lower semicontinuous in X × I with respect to the considered topology. Of course, f is convex in I. Finally, by a classical result, the spectrum of T is countable, and so the set, say D, of all λ ∈ I such that x → T (x) − λx is a homeomorphism between E (with the strong topology) and itself is dense in I. Fix λ ∈ D. Of course, since ϕ is coercive, for each ρ ∈ R, the set {x ∈ E : ϕ(T (x) − λx) ≤ ρ} is bounded. Hence, due to the reflexivity of E, the sub-level sets of f (·, λ) are weakly compact and so, by the Eberlein-Smulyan theorem, sequentially weakly compact which is equivalent to sequentially τs-compact. Therefore, each assumption of Theorem 1 is satisfied and hence we have sup x∈X inf λ∈I (ϕ(T (x) − λx) + ψ(λ)) = inf λ∈I sup x∈X (ϕ(T (x) − λx) + ψ(λ)) . Now, observe that if λ ∈ I \ {0}, we have ϕ(T (x) − λx) = +∞ . sup x∈X (3) (4) Indeed, since the τs-interior of X is non-empty and E is reflexive and infinite-dimensional, X turns out to be unbounded. But T (X) is bounded (since ϕ is coercive) and so, since λ 6= 0, kT (x) − λxk = +∞ sup x∈X which yields (4) by the coercivity of ϕ again. At this point, the conclusion follows directly from (2), (3) and △ (4). REMARK 2. - Notice that both infinite-dimensionality of E and compactness of T cannot be dropped in Theorem 2. In this connection, it is enough to take T (x) = x, I = [0, 1], ϕ(x) = kxk, and ψ = 0. REMARK 3. - At present, we do not know any example showing that the reflexivity of E cannot be dropped. However, we conjecture that such an example can be constructed in infinite-dimensional Banach spaces with the Schur property. Acknowledgement. The author has been supported by the Gruppo Nazionale per l'Analisi Matemat- ica, la Probabilit`a e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM). 4 References [1] B. RICCERI, Some topological mini-max theorems via an alternative principle for multifunctions, Arch. Math. (Basel), 60 (1993), 367-377. [2] B. RICCERI, Applications of a theorem concerning sets with connected sections, Topol. Methods Non- linear Anal., 5 (1995), 237-248. [3] B. RICCERI, Nonlinear eigenvalue problems, in "Handbook of Nonconvex Analysis and Applications" D. Y. Gao and D. Motreanu eds., 543-595, International Press, 2010. [4] S. SIMONS, Minimax theorems and their proofs, in "Minimax and Applications", D.-Z. Du and P. M. Pardalos eds., 1-23, Kluwer Academic Publishers, 1995. Department of Mathematics University of Catania Viale A. Doria 6 95125 Catania Italy e-mail address: [email protected] 5
1803.10191	2	1803	2018-04-03T15:34:20	Point-like perturbed fractional Laplacians through shrinking potentials of finite range	[ "math.FA", "math-ph", "math-ph" ]	We reconstruct the rank-one, singular (point-like) perturbations of the $d$-dimensional fractional Laplacian in the physically meaningful norm-resolvent limit of fractional Schr\"{o}dinger operators with regular potentials centred around the perturbation point and shrinking to a delta-like shape. We analyse both the possible regimes, the resonance-driven and the resonance-independent limit, depending on the power of the fractional Laplacian and the spatial dimension. To this aim, we also qualify the notion of zero-energy resonance for Schr\"{o}dinger operators formed by a fractional Laplacian and a regular potential.	math.FA	math	POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS THROUGH SHRINKING POTENTIALS OF FINITE RANGE ALESSANDRO MICHELANGELI AND RAFFAELE SCANDONE Abstract. We construct the rank-one, singular (point-like) perturbations of the d-dimensional fractional Laplacian in the physically meaningful norm- resolvent limit of fractional Schrodinger operators with regular potentials cen- tred around the perturbation point and shrinking to a delta-like shape. We analyse both possible regimes, the resonance-driven and the resonance-inde- pendent limit, depending on the power of the fractional Laplacian and the spatial dimension. To this aim, we also qualify the notion of zero-energy reso- nance for Schrodinger operators formed by a fractional Laplacian and a regular potential. 1. Introduction and background In the last decade an amount of studies focused, in particular in application to the context of fractional quantum mechanics, on linear Schrodinger equations governed by the linear operator (−∆)s/2 + singular perturbation at x0 (1.1) for some fixed point x0 ∈ Rd and some s > 0, that is, Schrodinger equations for a singular perturbation of a fractional power of the Laplacian [15, 17, 6, 13, 19, 21, 9, 16, 18]. Motivated by that, in a recent work in collaboration with A. Ottolini [14] we set up the systematic construction and classification of all the self-adjoint realisations in L2(Rd) of the operators of the form (1.1) through a natural 'restriction-extension' procedure: first one restricts the operator (−∆)s/2 (initially defined, e.g., as a Fourier multiplier) to smooth functions vanishing in neighbourhoods of x0, and then one builds all the operator extensions of such restriction that are self-adjoint on L2(Rd). This approach is surely satisfactory from the point of view of the interpretation of the output operator, which by construction is to be regarded as a point-like per- turbation of the fractional Laplacian through an interaction supported only at x0, say, "(−∆)s/2 + δ(x− x0)". However, it obfuscates an amount of physical meaning, since it does not provide information, as the intuition would make one expect in- stead, on how the actual singular perturbation (1.1) is approximatively realised as a genuine pseudo-differential operator (−∆)s/2 + V (x− x0) with a regular potential V centred around x = 0, with sufficiently short range and strong magnitude. For the non-fractional Laplacian −∆ in L2(Rd), the realisation of a singular perturbation at x0 ∈ Rd by means of approximating Schrodinger operators −∆+Vε with regular potentials Vε spiking up and shrinking around x0 at a spatial scale ε−1 in the limit ε ↓ 0 is known since long for dimension d = 1 [4] , d = 2 [1], and d = 3 [2] (we also refer to [3, 5] for a comprehensive overview), that is, all the dimensions in which non-trivial singular perturbations exist. Date: April 4, 2018. Key words and phrases. Fractional Laplacian. Singular perturbations of differential operators. Schrodinger operators with shrinking potentials. Resolvent limits. Zero-energy resonance. 1 2 A. MICHELANGELI AND R. SCANDONE far, and we solve it in the present work. The analogous question for the fractional Laplacian (−∆)s/2 was unanswered so Not only is it topical in view of the above-mentioned recent mainstream in the literature of fractional Schrodinger equations with singular perturbation, but also it rises up the conceptually new issue of how a local potential Vε can be suitably re- scaled so as to produce the desired perturbation of the non-local operator (−∆)s/2. Let us first of all reconsider what emerges from the construction that, as men- tioned, was recently given in [14]. For s > 0 and d ∈ N, the restriction (−∆)s/2C∞ 0 (Rd\{0}) is a positive symmetric operator on the Hilbert space L2(Rd), hence with equal deficiency indices, and for short we shall just speak of the deficiency index. The number J (s, d) := deficiency index of (−∆)s/2C∞ 0 (Rd\{0}) is finite, and is zero or a strictly positive integer depending on s and d, according to the rule (0, d 2 ] := ( ⇒ J (s, d) = (cid:18)d + n − 1 2 + n − 1, d ( d d 2 + n] n = 1, 2, . . . (1.2) s ∈ I (d) n n = 0 (cid:19) . In the non-fractional case s = 2 this yields the familiar values J (2, 1) = 2, J (2, 2) = 1, J (2, 3) = 1, and J (2, d) = 0 for d > 4. As well known, J (s, d) quantifies the infinite multiplicity of self-adjoint exten- sions of (−∆)s/2C∞ 0 (Rd\{0}) in L2(Rd). By means of standard methods of the Kreın-Visik-Birman theory [8] one sees that the domain of each extension is formed by functions that are canonically decomposed into a regular H s-component and a more singular component, the latter belonging to the J (s, d)-dimensional kernel of ((−∆)s/2C∞ 0 (Rd\{0}))∗ + λ1 for some arbitrarily chosen λ > 0, and satisfy an amount of 'boundary' (or 'contact') conditions between the evaluation at x = 0 of the regular part or of some if its derivatives and the coefficients of the leading sin- gularities of the singular part as x → 0. Each set of boundary conditions identifies uniquely an extension. For concreteness of the presentation, in this work we consider the self-adjoint extensions of (−∆)s/2C∞ 0 (Rd\{0}) in the case of deficiency index 1 only, and for simplicity we omit further the explicit discussion of the 'endpoint' values of s, namely the largest possible value, at given d, compatible with J (s, d) = 1. As expressed by (1.2), this amounts to analysing the regime s ∈ ( 1 2 ) in d = 1, s ∈ (1, 2) in d = 2, s ∈ ( 3 2 ) in d = 3, etc., where the considered intervals are the non-endpoint values of s, the endpoint value being s = d 2 + 1. We shall refer to such cases as the 'J = 1 scenario'. For this scenario we then discuss how to realise the corresponding extensions in the limit of Schrodinger operators with fractional Laplacian and shrinking potentials, say, (−∆)s/2 + Vε as ε ↓ 0. In fact, it will be evident from our discussion that the behaviour and the control of the limit ε ↓ 0 in the J = 1 scenario is technically the very same irrespectively of the dimension, and therefore we will pick up a concrete value of d for the explicit computations, modulo the dichotomy that we now describe. 2 , 5 2 , 3 When J (s, d) = 1, and s 6= d 2 + 1, the space where the above-mentioned sin- 0 (Rd\{0}))∗ + λ1), is the one- gular components run over, namely ker(((−∆)s/2C∞ dimensional space spanned by the Green function Gs,λ of the fractional Laplacian, defined by ((−∆)s/2 + λ)Gs,λ = δ(x) ; POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS 3 for higher deficiency index the kernel is spanned by Gs,λ and other non-H s func- tions. Now, depending on d and s, the Green function Gs,0 may be singular or regular at x = 0: when s < d, Gs,0 has a singularity ∼ x−(d−s), it has a logarith- mic singularity when s = d, and it is continuous at x = 0 when s > d. Omitting the transition case, which does not alter the conceptual scheme of the present dis- cussion and could be easily recovered with analogous arguments to those that we shall use when s < d, we thus distinguish two possibilities in the J = 1 scenario, that we call 2 , d • locally singular, or resonance-driven case: s < d, • locally regular, or resonance-independent case: s > d. We shall explain in a moment the meaning of the 'resonance' jargon: it has to do with how the limit of shrinking potentials must be organised in order to reach a self-adjoint extension of (−∆)s/2C∞ 0 (Rd\{0}) in one case or in the other. Also, let us remark that an analogous dichotomy occurs when the deficiency index of (−∆)s/2C∞ 0 (Rd\{0}) is larger than 1: the singular (non-H s) component of the ele- ments in the domain of the considered self-adjoint extension may or may not display a local singularity as x → 0. In view of the above alternative, we make the following presentational choice. Since in all dimensions d but d = 1 the interval s ∈ ( d 2 + 1) corresponding to deficiency index 1 lies strictly below the transition value s = d that separates the locally regular from the locally singular regime, as a representative of any such value of d for concreteness we choose d = 3: the discussion on the limit of shrinking potentials would then be immediately exportable to any other d > 2. Next to that, we also discuss the case d = 1, where instead the interval s ∈ (1, 2) corresponding to deficiency index 1 contains the transition value s = 1. 0 (Rd\{0}) in the locally regular and the locally singular case differ both in the type of non-regularity of the functions in their domain at x = 0, and, as we shall show in this work, in the type of approximating Schrodinger operators (−∆)s/2 + Vε, meaning, in the scaling chosen for Vε and, most importantly, in the spectral requirements. Extensions in the locally regular case can be reached as ε ↓ 0 through suitably rescaled versions Vε of a given potential V with no further prescription on V but those technical assumptions ensuring that the limit itself is well-posed. Instead, extensions in the locally singular case can only be reached if the unscaled operator (−∆)s/2 + V admits a zero-energy resonance, a spectral behaviour at the bottom of its essential spectrum which we shall define in due time and roughly speaking amounts to the existence of a suitably decaying, non square-integrable, L2 loc-solution f to ((−∆)s/2 + V )f = 0. In a sense that we shall make precise, this difference is due to the fact that a zero-energy resonance is needed in the approximating fractional Schrodinger operator in order to reproduce in the limit the locally singular behaviour in the domain of the considered self-adjoint extension. As mentioned above, self-adjoint extensions of (−∆)s/2C∞ In fact, the phenomenon we have just described is the generalisation for (−∆)s/2 of what is well known for −∆ (i.e., s = 2 in the present notation). When d = 1, the deficiency index of (−∆)C∞ ker(cid:0)((−∆)C∞ 0 (R\{0}) equals 2 and e− 0 (R\{0}))∗ + λ1(cid:1) = span(cid:8) 1√λ √λx, (sign x) e− √λx(cid:9) (in particular, G2,λ(x) =p π 2λ e−√λx), therefore the functions in the above space are less regular than H 2(R) but not locally singular at x = 0. The so-called δ-type extensions, namely those in the domain of which the singular component is e−√λx, can indeed be realised as limits of −∆ + Vε with no spectral requirement needed at energy zero for the unscaled −∆ + V [3, Chapt. I.3]. On the contrary, when d = 3 4 A. MICHELANGELI AND R. SCANDONE the deficiency index of (−∆)C∞ 0 (R3\{0}) equals 1 and ker(cid:0)((−∆)C∞ 0 (R3\{0}))∗ + λ1(cid:1) = span{G2,λ} , √λx e− 4πx thus with a local singularity at x = 0. The self-adjoint extensions of (−∆)C∞ 0 (R3\{0}) can be realised as limits of −∆ + Vε provided that −∆ + V is zero-energy resonant [3, Chapt. I.1]. In the former situation we are in the locally regular, resonant- independent case; in the latter we are in the locally singular, resonant-driven case. G2,λ(x) = , The material of this work is organised as follows. • In Section 2 we define the singular perturbations of the three-dimensional fractional Laplacian and we present the approximation scheme in terms of fractional Schrodinger operators with regular, shrinking potentials. • In Section 3 we present the one-dimensional analogue, including the def- inition of the singular perturbations and the two distinct approximation schemes, for the resonance-driven and the resonance-independent cases. • Section 4 contains the proof of the three-dimensional limit. • Section 5 contains the proof of the one-dimensional limit in the resonance- driven case. From the technical point of view, the argument here is com- pletely analogous to that of 4, as the 3D case too is resonance-driven. resonance-independent case. • Section 6 contains instead the proof of the one-dimensional limit in the • In Section 7 we prove a technical result used in the main proofs, that is, the characterisation of the zero-energy resonant behaviour of the unscaled operator (−∆)s/2 + V . Then, we discuss the occurrence of zero-energy resonances. Let us conclude this Introduction with a few comments about our otherwise standard notation. For an operator T on a Hilbert space, D(T ) denotes its operator domain and, when T is self-adjoint, D[T ] denotes its form domain. We shall denote by 1, resp., by , the identity and the null operator on any of the considered Hilbert spaces. We shall indicate the Fourier transform by bφ or F φ with the convention 2RRd e−ipxφ(x)dx. We shall write A . B for A 6 const. B when the bφ(p) = (2π)− d constant does not depend on the other relevant parameters or variables of both sides of the inequality; for x ∈ Rd we shall write hxi := (1 + x2) 1 2 . 2. Approximation scheme in dimension three In this Section we consider the singular perturbations of the three-dimensional fractional Laplacian and their approximation by means of fractional Schrodinger operators with shrinking potentials. Let us start with the densely defined, closed, positive, symmetric operator (2.1) k(s/2) := (−∆)s/2 ↾ C∞0 (R3 \ {0}) , s > 0 , with respect to the Hilbert space L2(R3). In [14] we presented the construction and classification of the self-adjoint extensions of k(s/2), which we recall here below. Clearly, for small enough powers s, k(s/2) is already self-adjoint, thus with no room for point-like singular perturbations, indeed [14, Lemma A.1] (2.2) D(k(s/2)) = H s 0 (R3 \ {0}) = C∞0 (R3 \ {0}) k kHs = H s(R3) , if s ∈ [0, 3 2 ) . POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS 5 When s increases, the domain of k(s/2) is qualified by an increasing number of conditions for H s-functions, namely [14, Lemma A.2] as a consequence, the adjoint of k(s/2) becomes strictly larger than k(s/2) itself, with an increasingly complicated structure of its domain that reflects the fact that for  γ1, γ2, γ3 ∈ N0 , γ1 + γ2 + γ3 6 n − 1 3 bf (p) dp = 0 n ∈ N ; f ∈ H s(R3) such that n 2 ) , 2 , n + 3 1 pγ2 2 pγ3 := (n + 1 if s ∈ I (3) RR3 pγ1 D(k(s/2)) =  the deficiency index of k(s/2) equals(cid:18)n + 2 D(k(s/2)) = nf ∈ H s(R3)(cid:12)(cid:12)(cid:12)ZR3 bf (p) dp = 0o 1 = ( 3 2 ) one has 2 , 5 structure of the family of its self-adjoint extensions. In particular, in the regime s ∈ I (3) if s ∈ ( 3 2 , 5 2 ) (2.3) s ∈ I (3) n (2.4) 3 (cid:19), and this in turn affects the and k(s/2) has deficiency index 1, which leaves room for a one-(real-)parameter family of self-adjoint extensions. In order to qualify them, for chosen λ > 0 and s ∈ R let us denote the Green's function as 1 (2.5) Gs,λ(x) := 1 x, p ∈ R3 . 2(cid:16) ps + λ(cid:17)∨(x) , (2π) 3 By construction, distributionally. (2.6) ((−∆)s/2 + λ) Gs,λ = δ(x) . Observe that Gs,λ has a local singularity x−(3−s), more precisely [14, Sec. 3], (2.7) with (2.8) Λs := Gs,λ(x) = Λs x(3−s) + Js,λ(x) 3 Γ( 3−s 2 ) 2 2s− 3 (2π) λ 2 Γ( s 2 ) 1 ps(ps + λ)(cid:17)∨ 2(cid:16) Js,λ := − (2π) 3 ∈ C∞(R3) . The following construction/classification Theorem was established in [14]. Theorem 2.1. Let s ∈ ( 3 2 , 5 2 ). (i) The self-adjoint extensions in L2(R3) of the operator k(s/2) form the family (k(s/2) ∞ is its Friedrichs extension, namely the self- adjoint fractional Laplacian (−∆)s/2, and all other (proper) extensions are given, for arbitrary λ > 0, by )α∈R∪{∞}, where k(s/2) α (2.9) D(k(s/2) α g = F λ + F λ(0) α − λ 3 s −1 2πs sin( 3π s ) (k(s/2) α + λ) g = ((−∆)s/2 + λ) F λ . ) =  Gs,λ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) F λ ∈ H s(R3) (2.10) (2.11) k(s/2) α for arbitrary λ > 0. (iii) The resolvent of k(s/2) (2.12) (k(s/2) α σess(k(s/2) α σdisc(k(s/2) α where the eigenvalue E(s) (2.13) (2.14) for arbitrary λ > 0. (iv) Each extension is semi-bounded from below, and 6 A. MICHELANGELI AND R. SCANDONE (ii) For each α ∈ R the quadratic form of the extension k(s/2) α is given by [F λ + κλGs,λ] = k∇sF λk2 L2(R3) − λkF λ + κλGs,λk2 L2(R3) s D[k(s/2) α α 3 s −1 3 s −1 2πs sin( 3π ] = H is given by + λkF λk2 s )(cid:1)κλ2 +(cid:0)α − λ 2 (R3) ∔ span{Gs,λ} L2(R3) +(cid:0)α − λ + λ1)−1 = ((−∆)s/2 + λ1)−1 s )(cid:1)−1 ) = σac(k(s/2)) = [0, +∞) , ) = ( ∅ {E(s) α } α is non-degenerate and is given by α = −(cid:0)2πα s sin(− 3π if α > 0 if α < 0 , s )(cid:1) s Gs,λihGs,λ 2πs sin( 3π E(s) 3−s , . s,λ=E(s) α σsc(k(s/2)) = ∅ , the (non-normalised) eigenfunction being G Our goal now is to qualify each of the extensions given by Theorem 2.1 as suitable limits of approximating fractional Schrodinger operators with finite range poten- tials. It is convenient to introduce the class Rs,d, d ∈ N, s ∈ ( d 2 , d), of measurable functions V : Rd → C such that (2.15) ZZRd×Rd dx dy V (x)V (y) x − y2(d−s) =: kV k2 Rs,d < +∞ . 2 , 5 R2,3 is the well-known Rollnick class on R3. Clearly, Rs,d ⊃ C∞0 (Rd). 2 ) we make the following assumption. For each s ∈ ( 3 Assumption (Is). (i) V : R3 → R is a measurable function in L1(R3,hxi2s−3dx) ∩ Rs,3. (ii) η : R → R+ is a continuous function satisfying η(0) = η(1) = 1 and η(ε) = 1 + ηs ε3−s + o(ε3−s) as ε ↓ 0 for some ηs ∈ R that we call the strength of the distortion factor η. For given V and η satisfying Assumption (Is), let us set (2.16) h(s/2) ε := (−∆)s/2 + Vε , Vε(x) := η(ε) εs V ( x ε ) , ε > 0 . For every ε > 0 the operator h(s/2) and σess(h(s/2) ε ε ) = [0, +∞) (Lemma 4.1(iii)). , defined as a form sum, is self-adjoint on L2(R3) The spectral properties of the unscaled operator (−∆)s/2 + V at the bottom of . In the next Theorem the essential spectrum are crucial for the limit ε ↓ 0 in h(s/2) we qualify the zero-energy behaviour of (−∆)s/2 + V . Theorem 2.2. Let s ∈ ( 3 v := V 1 2 and u := V 1 2 , 5 2 sign(V ). (i) The operator u(−∆)− s 2 v is compact on L2(R3). ε 2 ), V ∈ L1(R3,hxi2s−3dx) ∩ Rs,3, real-valued. Let POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS 7 Assume in addition that (2.17) and define (2.18) Then: u(−∆)− s 2 v φ = −φ for some φ ∈ L2(R3) \ {0} ψ := (−∆)− s 2 v φ . loc(R3) and(cid:0)(−∆)s/2 + V(cid:1)ψ = 0 in the sense of distributions, dx V (x)ψ(x), (ii) ψ ∈ L2 (iii) hv, φiL2 = −ZR3 (iv) ψ ∈ L2(R3) ⇔ hv, φiL2 = 0, in which case ψ ∈ D((−∆)s/2 + V ). When a L2-function φ exists that satisfies (2.17) and the corresponding function loc(R3) \ L2(R3) we say that (−∆)s/2 + V is zero- ψ defined by (2.18) belongs to L2 energy resonant and that ψ is a zero-energy resonance for (−∆)s/2 + V . If for the zero-energy resonant operator (−∆)s/2 + V the eigenvalue −1 of u(−∆)− s 2 v is non-degenerate, then we say that the resonance is simple. Of course, if ψ ∈ L2(R3), then ψ is an eigenfunction of (−∆)s/2 + V with eigenvalue zero. We shall prove Theorem 2.2 in Section 7 together with a discussion of the oc- currence of a zero-energy resonance for (−∆)s/2 + V (Proposition 7.1). Let us now formulate our main result for dimension three. It is the control of the approximation, in the norm resolvent sense, of the singular perturbation operator k(s/2) 2 -th fractional Laplacian and α shrinking potentials Vε around the origin. We shall prove it in Section 4. Theorem 2.3. Let s ∈ ( 3 strength ηs satisfying Assumption (Is), for every ε > 0 let h(s/2) the corresponding self-adjoint Schrodinger operator defined in (2.16) with the s fractional Laplacian and the shrinking potential Vε. 2 ). Given a potential V and a distortion factor η with = (−∆)s/2 + Vε be 2 -th by means of Schrodinger operators with the s 2 , 5 ε (i) If (−∆)s/2 + V is not zero-energy resonant, then h(s/2) (ii) If (−∆)s/2 + V admits a simple zero-energy resonance ψ, then for the norm-resolvent sense on L2(R3). ε ε↓0 −−−→ (−∆)s/2 in α := −ηs(cid:12)(cid:12)(cid:12)ZR3 dx V (x)ψ(x)(cid:12)(cid:12)(cid:12) −2 one has h(s/2) ε ε↓0 −−−→ k(s/2) α in the norm-resolvent sense on L2(R3). In view of the discussion we made in the introductory Section, the two possible alternatives in Theorem 2.3 are the manifestation of the locally singular, resonant- driven nature of the limit: the limit is well-posed for a generic class of potentials V , but it is non-trivial only if additionally (−∆)s/2 + V is zero-energy resonant. By a simple scaling argument one sees that (−∆)s/2 + Vε remains zero-energy resonant for any ε > 0 if the scaling is 'purely geometric', namely with trivial dis- tortion factor, η(ε) ≡ 1. In this case, the signature of the resonance is particularly transparent: as stated in Theorem 2.3(ii), the limit ε ↓ 0 with η(ε) ≡ 1 produces the extension parametrised by α = 0 and we see from Theorem 2.1(iv) that the negative eigenvalue of k(s/2) when α < 0 converges to 0 as α ↑ 0, with the corre- α converging pointwise to Gs,0(x) = Λs x(3−s) (see sponding eigenfunction G (2.7)-(2.8) and (2.14) above); the L2 loc\L2-function Gs,0 can be actually regarded as a zero-energy resonance for k(s/2) α=0 (the local square-integrability following from s ∈ ( 3 s,λ=E(s) α 2 , 5 2 )). 8 A. MICHELANGELI AND R. SCANDONE 3. Approximation scheme in dimension one In this Section we consider the singular perturbations of the one-dimensional fractional Laplacian and their approximation by means of fractional Schrodinger operators with shrinking potentials. As for the 3D case, for λ > 0 and s > 0, we set (3.1) Gs,λ(x) := 1 (2π) 1 2(cid:16) 1 ps + λ(cid:17)∨(x) , x, p ∈ R (whence ((−∆)s/2 + λ) Gs,λ = δ(x) distributionally), and k(s/2) := (−∆)s/2 ↾ C∞0 (R \ {0}) (3.2) as an operator closure with respect to the Hilbert space L2(R). k(s/2) has deficiency index n ∈ N when s ∈ I (1) 2 ) (see (1.2) above), and in the case of := (n − 1 2 , n + 1 n deficiency index 1 one has D(k(s/2)) = nf ∈ H s(R)(cid:12)(cid:12)(cid:12)ZR bf (p) dp = 0o , s ∈ ( 1 2 , 3 2 ) (which can be seen by means of a completely analogous argument to that of [14, Appendix A]). The corresponding one-(real-)parameter family of self-adjoint exten- sions is given by the one-dimensional analogous of Theorem 2.1 [14]. Theorem 3.1. Let s ∈ ( 1 2 ) and 2 , 3 (3.3) (3.4) Θ(s, λ) := ((cid:0)λ1− 1 s s sin ( π − 1 π ln λ s )(cid:1)−1 s 6= 1 s = 1 , λ > 0 . (i) The self-adjoint extensions in L2(R) of the operator k(s/2) form the family (k(s/2) ∞ is its Friedrichs extension, namely the self- adjoint fractional Laplacian (−∆)s/2, and all other (proper) extensions are given, for arbitrary λ > 0, by )α∈R∪{∞}, where k(s/2) α (3.5) g = F λ + F λ(0) α − Θ(s, λ) D(k(s/2) α F λ ∈ H s(R) + λ) g = ((−∆)s/2 + λ) F λ . (k(s/2) α ) =  Gs,λ  (ii) For each α ∈ R the quadratic form of the extension k(s/2) α is given by (3.6) (3.7) k(s/2) α D[k(s/2) α ] = H s 2 (R) ∔ span{Gs,λ} [F λ + κλGs,λ] = k∇sF λk2 L2(R) − λkF λ + κλGs,λk2 L2(R) + λkF λk2 L2(R) +(cid:0)α − Θ(s, λ)(cid:1)κλ2 for arbitrary λ > 0. (iii) The resolvent of k(s/2) α is given by (3.8) (k(s/2) α + λ1)−1 = ((−∆)s/2 + λ1)−1 +(cid:0)α − Θ(s, λ)(cid:1)−1 Gs,λihGs,λ for arbitrary λ > 0. (iv) For each α ∈ R the extension k(s/2) α is semi-bounded from below, and (3.9) σess(k(s/2) α ) = σac(k(s/2) α ) = [0, +∞) , σsc(k(s/2) α ) = ∅ , POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS 9 (3.10) σdisc(k(s/2) α where the eigenvalue −E(s) (3.11) ∅ if s 6= 1, (s − 1) α 6 0 if s 6= 1, (s − 1) α > 0 if s = 1 , ) =  {−E(s) α } {−E(1) α } α is non-degenerate and is given by α =((cid:0)αs sin( π s )(cid:1) s s 6= 1 s = 1 , e−πα 1−s E(s) the (non-normalised) eigenfunction being G s,λ=E(s) α . Our goal is to qualify each of the extensions given by Theorem 3.1 as suitable limits of approximating fractional Schrodinger operators with finite range poten- tials. Unlike the 3D setting, here the regime s ∈ ( 1 2 ) is separated by the transition value s = 1, below which we are in the locally singular case for the Green function (3.1), and above which we are in the locally regular case, in the terminology of Section 1. This will result in different assumptions on the approximating potentials and different schemes for the resolvent limit. 2 , 3 We therefore proceed by splitting our discussion into the two above-mentioned cases. 3.1. Locally singular, resonance-driven case. This is the regime s ∈ ( 1 2 , 1). The Green function Gs,λ has a local singularity, (3.12) with (3.13) Λs := Gs,λ(x) = Λs x(1−s) + Js,λ(x) Γ( 1−s 2 ) (2π) 1 2 2s− 1 λ 2 Γ( s 2 ) 1 ps(ps + λ)(cid:17)∨ 2(cid:16) 1 Js,λ := − (2π) ∈ C∞(R) . We make the following assumption (the class Rs,d was introduced in (2.15)). Assumption (I−s ). s ∈ ( 1 (i) V : R → R is a measurable function in L1(R,hxi2s−1dx) ∩ Rs,1; (ii) η : R → R+ is a continuous function satisfying η(0) = η(1) = 1 and 2 , 1) and moreover: η(ε) = 1 + ηs ε1−s + o(ε1−s) as ε ↓ 0 for some ηs ∈ R that we call the strength of the distortion factor η. For given V and η satisfying Assumption (I−s ), let us set (3.14) h(s/2) ε := (−∆)s/2 + Vε , Vε(x) := η(ε) εs V ( x ε ) , ε > 0 . For every ε > 0 the operator h(s/2) and σess(h(s/2) ε ε ) = [0, +∞) (Lemma 5.2(iii)). , defined as a form sum, is self-adjoint on L2(R3) The zero-energy spectral behaviour of (−∆)s/2 + V , which is crucial for the limit , is described as follows, in analogy with Theorem 2.2. ε ε ↓ 0 in h(s/2) Theorem 3.2. Let s ∈ ( 1 v := V 1 2 and u := V 1 2 sign(V ). (i) The operator u(−∆)− s 2 , 1), V ∈ L1(R,hxi2s−1dx) ∩ Rs,1, real-valued. Let 2 v is compact on L2(R). 10 A. MICHELANGELI AND R. SCANDONE Assume in addition that (3.15) and define (3.16) Then: u(−∆)− s 2 v φ = −φ for some φ ∈ L2(R) \ {0} ψ := (−∆)− s 2 v φ . loc(R) and(cid:0)(−∆)s/2 + V(cid:1)ψ = 0 in the sense of distributions, dx V (x)ψ(x), (ii) ψ ∈ L2 (iii) hv, φiL2 = −ZR (iv) ψ ∈ L2(R) ⇔ hv, φiL2 = 0, in which case ψ ∈ D((−∆)s/2 + V ). We defer to Section 7 the proof of Theorem 3.2 and a discussion of the occurrence of a zero-energy resonance for (−∆)s/2 + V (Proposition 7.2). With the same terminology of Section 2, (−∆)s/2 + V is zero-energy resonant and that ψ is a zero-energy resonance for (−∆)s/2 + V when there exists a non-zero L2-function φ satisfying (3.15) and the corresponding function ψ defined by (3.16) belongs to L2 loc(R) \ L2(R). If, for the zero-energy resonant operator (−∆)s/2 + V , the eigenvalue −1 of u(−∆)− s Here below is our first main result in dimension one, relative to the resonance- driven regime. Theorem 3.3. Let s ∈ ( 1 strength ηs satisfying Assumption (I−s ), for every ε > 0 let h(s/2) the corresponding self-adjoint Schrodinger operator defined in (3.14) with the s fractional Laplacian and the shrinking potential Vε. 2 , 1). Given a potential V and a distortion factor η with = (−∆)s/2 + Vε be 2 -th 2 v is non-degenerate, then the resonance is simple. ε (i) If (−∆)s/2 + V is not zero-energy resonant, then h(s/2) (ii) If (−∆)s/2 + V admits a simple zero-energy resonance ψ, then for the norm-resolvent sense on L2(R). ε ε↓0 −−−→ (−∆)s/2 in α := −ηs(cid:12)(cid:12)(cid:12)ZR dx V (x)ψ(x)(cid:12)(cid:12)(cid:12) −2 one has h(s/2) ε ε↓0 −−−→ k(s/2) α in the norm-resolvent sense on L2(R). We shall prove Theorem 3.3 in Section 5. The alternative in Theorem 3.3 is completely analogous to that of Theorem 2.3, due to the the locally singular, resonant-driven nature of both limits: only for zero-energy resonant operators (−∆)s/2 + V is the limit non-trivial. The signature of the resonance is particularly transparent in the absence of distortion factor: when η(ε) ≡ 1 by scaling one sees that (−∆)s/2 + Vε remains zero-energy resonant for any ε > 0, and we may regard the limit operator k(s/2) α=0 too as zero-energy resonant, for the negative eigenvalue of k(s/2) when α 6= 0 vanishes as α → 0 and the corresponding eigenfunctions becomes (proportional to) the L2 loc\L2-function x−(1−s) (see (3.11) above). 3.2. Locally regular, resonance-independent case. α This is the regime s ∈ (1, 3 2 ). In contrast with the resonance-driven regime, no spectral requirement is now needed on the unscaled fractional operator (−∆)s/2 + V and the scaling in Vε is independent of s. Thus, we make the following assumption. s ). Assumption (I+ (i) V : R → R is a measurable function in L1(R). (ii) η : R+ → R+ is a smooth function satisfying η(1) = 1. POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS11 Correspondingly, we set (3.17) h(s/2) ε := (−∆)s/2 + Vε , Vε(x) := η(ε) ε V ( x ε ) , ε > 0 . , defined as a form sum, is self-adjoint on L2(R3) For every ε > 0 the operator h(s/2) and σess(h(s/2) ε ε ) = [0, +∞) (Lemma 6.1(iii)). Here below is our second main result in dimension one, which, as opposite to Theorem 3.3, takes the following form. Theorem 3.4. Let s ∈ (1, 3 defined according to Assumption (I+ norm-resolvent sense on L2(R), where 2 ). For every ε > 0 let h(s/2) ε s ) and (3.17). Then h(s/2) ε = (−∆)s/2 + Vε be ε↓0 −−−→ k(s/2) in the α α := −(cid:16)η(0)ZR dx V (x)(cid:17)−1 . We shall prove Theorem 3.4 in Section 6. 4. Convergence of the 3D limit The goal of this Section is to prove Theorem 2.3. Let us start with qualifying the following useful operator-theoretic properties. Lemma 4.1. Let V : R3 → R belong to L1(R3) ∩ Rs,3 for some s ∈ ( 3 2 , 3). Then: 2 is a Hilbert-Schmidt operator on 2 ((−∆) s 2 + λ1)−1V 1 L2(R3); (i) for every λ > 0, V 1 (ii) V 1 (iii) the operator (−∆) 2 ≪ (−∆) V ) = [0, +∞). 2 ((−∆) Proof. (i) V 1 s s s 4 in the sense of infinitesimally bounded operators; 2 +V defined as a form sum is self-adjoint and σess((−∆) s 2 + 2 acts as an integral operator with kernel 2 + λ1)−1V 1 Ks,λ(x, y) := V (x) 1 2 Gs,λ(x − y)V (y) 1 2 , dx dy V (x)V (y) 2 is continuous from (0, +∞) to the and its Hilbert-Schmidt norm is estimated as (cid:13)(cid:13)(cid:13)V 1 2 ((−∆) s 2 + λ1)−1V 6 2Λ2 6 2Λ2 sZZR3×R3 skV k2 Rs,3 + 2kJs,λk2 1 2 H.S. = ZZR3×R3 2(cid:13)(cid:13)(cid:13) x2(3−s) + 2kJs,λk2 L∞kV k2 L1 < +∞ , dx dy Ks,λ(x, y)2 L∞ZZR3×R3 dx dy V (x)V (y) having used (2.7)-(2.8) in the second step. s lim 2 ((−∆) 2 + λ1)−1V 1 (ii) The map λ 7→ V 1 space of Hilbert-Schmidt operators, and by dominated convergence λ→+∞ZZR3×R3 dx dy V (x)Gs,λ(x − y)2 V (y) = 0 . ε > (cid:13)(cid:13)(cid:13)V 2(cid:0)(−∆) 2 + λε1(cid:1)−1 = (cid:13)(cid:13)(cid:13)(cid:0)(−∆) 2 + λε1(cid:1)− 1 > (cid:13)(cid:13)(cid:13)(cid:0)(−∆) 2 + λε1(cid:1)− 1 2 ϕiL2 + bεkϕk2 2(cid:13)(cid:13)(cid:13) 2 V (cid:0)(−∆) 2 V (cid:0)(−∆) which implies, for some bε > 0, hϕ, V ϕiL2 6 ε hϕ, (−∆) 2(cid:13)(cid:13)(cid:13) 2 + λε1(cid:1)− 1 2(cid:13)(cid:13)(cid:13) 2 + λε1(cid:1)− 1 ∀ϕ ∈ D[(−∆) V H.S. 2 H.S. L2 op 2 2 , 1 1 s s s s s s Therefore, for arbitrary ε > 0 it is possible to find λε > 0 large enough such that s 2 ] = H s 2 (R3) , 12 A. MICHELANGELI AND R. SCANDONE and hence V 1 2 ≪ (−∆) s 4 . (iii) The statement follows at once from (ii). For chosen s ∈ ( 3 2 , 5 recall from (2.16) that Vε(x) = η(ε) εs V ( x 2 ), ε > 0, and V and η satisfying Assumption (Is), let us v(x) := V (x) vε(x) := Vε(x) 1 2 , 1 2 , ε ) and let us define u(x) := V (x) uε(x) := Vε(x) 1 1 2 sign(V (x)) , 2 sign(Vε(x)) . (cid:3) (4.1) Thus, (4.2) vε(x) = √η(ε) εs/2 v( x ε ) , uε(x) = √η(ε) εs/2 u( x ε ) , vεuε = Vε . ε The Hamiltonian h(s/2) = (−∆)s/2 + Vε defined in (2.16) as a form sum is self- adjoint on L2(R3), as guaranteed by Lemma 4.1(iii). An expression for its resolvent that is convenient in the present context is the Konno-Kuroda identity [12]. One has the following. Lemma 4.2. Let V : R3 → R belong to L1(R3) ∩ Rs,3 for some s ∈ ( 3 2 ). Then 2 , 5 (4.3) ε + λ1(cid:1)−1 (cid:0)h(s/2) −(cid:0)(−∆)s/2 + λ1(cid:1)−1 = (cid:0)(−∆)s/2 + λ1(cid:1)−1 vε(cid:16)1 + uε(cid:0)(−∆)s/2 + λ1(cid:1)−1 − for every ε > 0 and every −λ < 0 in the resolvent set of h(s/2) between bounded operators on L2(R3). ε vε(cid:17)−1 uε(cid:0)(−∆)s/2 + λ1(cid:1)−1 , as an identity s 4 . Both conditions are guaranteed by Lemma 4.1. Proof. The statement is precisely the application of the Konno-Kuroda resolvent identity, for which we follow the formulation presented in [3, Theorem B.1(b)], to the operator (−∆)s/2 + vεuε. For the validity of such identity two conditions are needed: the compactness of uε((−∆)s/2 + λ1)−1vε and the infinitesimal bound V 1 (cid:3) 2 ≪ (−∆) Observe that the invertibility of 1+uε((−∆)s/2+λ1)−1vε (with bounded inverse) It is convenient to manipulate the identity (4.3) further so as to isolate terms in the r.h.s. which are easily controllable in the limit ε ↓ 0. To this aim, let us introduce for each ε > 0 the unitary scaling operator Uε : L2(R3) → L2(R3) defined by (4.4) is part of the statement of the Konno-Kuroda formula (4.3). (Uεf )(x) := 1 ε3/2 f ( x ε ) . Its adjoint clearly acts as (U∗ε f )(x) = ε3/2f (εx). Uε induces the scaling transfor- mations U∗ε vεUε = pη(ε) εs/2 v , (4.5) U∗ε(cid:0)(−∆)s/2 + λ1(cid:1)−1 Uε = εs(cid:0)(−∆)s/2 + λεs whose proof is straightforward. U∗ε uεUε = pη(ε) 1(cid:1)−1 , εs/2 u , Let us also introduce, for each ε > 0 and for each µ > 0 such that −µs belongs to the resolvent set of h(s/2) ε , the operators (4.6) A(s) ε B(s) ε C(s) ε (η(ε))− 1 2 vε Uε := ε− 3−s 1(cid:1)−1 2 (cid:0)(−∆)s/2 + µs 1(cid:1)−1 := η(ε) u(cid:0)(−∆)s/2 + (µε)s 2(cid:0)(−∆)s/2 + µs := U∗ε uε (η(ε))− 1 v ε− 3−s 2 . 1(cid:1)−1 POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS13 We shall see in a moment (Lemma 4.4) that A(s) are Hilbert- Schmidt operators on L2(R3). Most importantly for our purposes, the resolvent of h(s/2) ε takes the following convenient form. ε , and C(s) ε , B(s) ε Lemma 4.3. Under the present assumptions, ε (4.7) + µs 1(cid:1)−1 (cid:0)h(s/2) = (cid:0)(−∆)s/2 + µs ε )−1C(s) for every ε > 0 and every µ > 0 such that −µs belongs to the resolvent set of h(s/2) Proof. In formula (4.3) we set λ = µs and we insert 1 = UεU∗ε in the second summand of the r.h.s. right after ((−∆)s/2 + λ1)−1vε. We then commute U∗ε all the way through by means of the scaling transformations (4.5): this way, we reproduce the product A(s) (cid:3) ε ε3−sη(ε)(1 + B(s) ε ε3−sη(ε)(1 + B(s) ε )−1C(s) ε . − A(s) 1(cid:1)−1 . ε ε The limit ε ↓ 0 can be monitored explicitly for A(s) ε , and C(s) Lemma 4.4. For every ε > 0, A(s) on L2(R3) with limit ε , B(s) ε ε , B(s) ε , and C(s) ε . are Hilbert-Schmidt operators (4.8) (4.9) (4.10) lim ε↓0 lim ε↓0 lim ε↓0 A(s) ε = Gs,µsihv ε = B(s) B(s) 0 = u (−∆)− s 2 v C(s) ε = uihGs,µs in the Hilbert-Schmidt operator norm. Proof. By construction, see (4.2), (4.4), and (4.6) above, Gs,µs (x − y) v( y ε f )(x) = ε− 3−s 2 ε− s 2 ε− 3 (A(s) ε )f ( y ε ) dy 2ZR3 = ZR3 Gs,µs (x − εy) v(y)f (y) dy ∀f ∈ L2(R3) , that is, A(s) ε acts as an integral operator with kernel Gs,µs (x−εy)v(y) . The latter is clearly a function in L2(R3 × R3, dx dy) uniformly in ε, and dominated convergence implies dx dy Gs,µs (x − εy)v(y)2 ε↓0 −−−→ kGs,µsk2 L2kV kL1 kA(s) as well as H.S. = ZZR3×R3 ε k2 hg, A(s) ε fiL2 = ZZR3×R3 dx dy g(x) Gs,µs (x − εy) v(y)f (y) ε↓0 −−−→ hg, Gs,µsiL2hv, fiL2 ∀f, g ∈ C∞0 (R3) . As a consequence, as ε ↓ 0, A(s) ε → Gs,µsihv weakly in the operator topology, and the Hilbert-Schmidt norm of A(s) converges to the Hilbert-Schmidt norm of ε its limit. By a well-known feature of compact operators [20, Theorem 2.21], the combination of these two properties implies that A(s) ε → Gs,µsihv in the Hilbert- Schmidt topology. This proves (4.8). is completely analogous: its integral kernel is u(x)Gs,µs (εx− ε , its integral kernel is η(ε)u(x)Gs,(µε)s (x−y)v(y) and the integral is u(x)Gs,0(x − y)v(y): owing to Lemma 4.1(i) both operators are 0 weakly in kernel of B(s) Hilbert Schmidt, and moreover by dominated convergence B(s) y) and (4.10) is proved by the very same type of argument. The discussion for C(s) Concerning B(s) 0 ε ε → B(s) 14 A. MICHELANGELI AND R. SCANDONE the operator topology and kB(s) ε k2 H.S. → kB(s) 0 k2 [20, Theorem 2.21] the limit (4.9) then holds in the Hilbert-Schmidt norm. H.S. as ε ↓ 0. By the same property (cid:3) It is evident from (4.7) that, in order for the limits (4.8)–(4.10) above to qualify as ε ↓ 0, one needs additional information 0 . By the 0 )−1 exists everywhere defined the behaviour of the resolvent of h(s/2) on the possible failure of invertibility in L2(R3) of the operator 1 + B(s) Fredholm alternative, since B(s) is compact, (1 + B(s) 0 ε and bounded, in which case (4.7) implies at once(cid:0)h(s/2) 1(cid:1)−1 admits an eigenvalue −1. as ε ↓ 0, unless B(s) µs 0 ε Let us then assume that the latter circumstance does occurs, namely condition (2.17) of Theorem 2.2. More precisely, we make the following assumption. + µs 1(cid:1)−1 →(cid:0)(−∆)s/2 + Assumption (IIs). Assumption (Is) holds. B(s) 0 has eigenvalue −1, which is 0 φ = −φ and, in non-degenerate. φ ∈ L2(R3) is a non-zero function such that B(s) addition, heφ, φiL2 = −1, where eφ := (signV )φ. Since heφ, φiL2 = −h(signV )φ, (signV )v(−∆)− s normalisation heφ, φiL2 = −1 is always possible. Under Assumption (IIs), (1 + B(s) ε )−1 becomes singular in the limit ε ↓ 0, with a singularity that now competes with the vanishing factor ε3−s of (4.7). To resolve this competing effect, we need first an expansion of B(s) around ε = 0 to a further order, than the limit (4.9). This expansion holds irrespectively of Assumption (IIs). Lemma 4.5. Let s ∈ ( 3 2 vφiL2 = −k(−∆)− s 2 ) and λ > 0. 4 vφk2 L2, the 2 , 5 ε (4.11) (i) For every x ∈ R3\{0} lim λ↓0 Gs,λ(x) − Gs,0(x) s ))−1λ 3 (2πs sin( 3π s −1 (ii) In the norm operator topology one has = 1 . (4.12) lim ε↓0 1 (µε)3−s(cid:0)B(s) 0 (cid:1) = ε − B(s) ηs µ3−s B(s) 0 + 1 2πs sin( 3π s ) uihv . Here µ > 0 is the constant chosen in the definition (4.6) of B(s) is the constant that is part of Assumption (Is). ε and ηs ∈ R Proof. (i) From (2.5) we write Gs,λ(x) − Gs,0(x) λ 3 s −1 = λ = − whence Gs,λ(x) − Gs,0(x) λ↓0 −−−→ − 3 3 1 1 s −1(2π) 2 ZR3 (2π)3ZR3 (2π)3ZR3 dp 1 dp eix·p (2π) 3 −λ 2ps(ps + λ) 1 , ps(ps + 1) dp ei λ1/sx·p λ 3 s −1 2πs sin( 3π s ) by dominated convergence, since p 7→ (ps(ps +1))−1 is integrable when s ∈ ( 3 ps(ps + 1) (ii) The Hilbert-Schmidt operator 2 , 3). 1 1 = 1 (µε)3−s(cid:0)B(s) 0 (cid:1) − ε − B(s) ηs µ3−s B(s) 0 − 1 2πs sin( 3π s ) uihv POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS15 has integral kernel () u(x)(cid:16) η(ε) − 1 (µε)3−s − ηs µ3−s(cid:17) Gs,(µε)s (x − y) v(y) + + u(x)(cid:16) Gs,(µε)s (x − y) − Gs,0(x − y) (µε)3−s 1 − 2πs sin( 3π s )(cid:17) v(y) . The first summand in () vanishes as ε ↓ 0 for a.e. x, y ∈ R3 as a consequence of Assumption (Is)(ii), and so does the second summand in () as a consequence of (4.11), where we take λ = (µε)s. Moreover, each such summand belongs to L2(R3× R3, dx dy) uniformly in ε, thanks to the assumption (Is)(i) on the potentials v and u. Thus, by dominated convergence, the function () vanishes in L2(R3× R3, dx dy) as ε ↓ 0, and this proves the limit (4.12) in the Hilbert-Schmidt norm. (cid:3) We can now monitor the competing effect in ε3−s(1 + B(s) ε )−1 as ε ↓ 0. Lemma 4.6. Under the Assumptions (Is) and (IIs) one has (4.13) lim ε↓0 in the operator norm topology. ε (cid:1)−1 (µε)3−s(cid:0)1 + B(s) = (cid:16) ηs s (cid:17)−1 µ3−s + hv, φiL22 2πs sin 3π φiheφ Proof. We re-write (4.12) in the form of the expansion (i) where, for short, B(s) ε = B(s) 0 + (µε)3−sB(s) + o(ε3−s) whence also 1 ηs µ3−s B(s) 0 + 2πs sin( 3π s ) uihv , B(s) := ε (cid:1)−1 (µε)3−s(cid:0)1 + B(s) = (cid:16)1 + (µε)3−s(cid:0)1 + (µε)3−s + B(s) 0 (cid:1)−1(cid:0)B(s) − 1 + o(1)(cid:1)(cid:17)−1 0 (cid:1)−1 × (µε)3−s(cid:0)1 + (µε)3−s + B(s) = . × The o(εa)-remainders in (i) and (ii) above are clearly meant in the Hilbert-Schmidt norm. The operator (µε)3−s(1 + (µε)3−s + B(s) 0 )−1 that appears twice in (ii) is of the (ii) form z(1 + T + z1)−1 , z ∈ C \ {0} , for a closed operator T with isolated eigenvalue −1; this is a general setting for which a well-known expansion by Kato is available as z → 0 [10, Sec. 3.6.5], which in the present context (in complete analogy with the argument of the proof of [3, Lemma I.1.2.4]) reads (iii) 0 (cid:1)−1 (µε)3−s(cid:0)1 + (µε)3−s + B(s) = −φiheφ + O(ε3−s) as ε ↓ 0 in the operator norm topology. In practice, (1 + (µε)3−s + B(s) 0 )−1 remains bounded also in the limit ε ↓ 0 when restricted to the orthogonal complement of the eigenspace −1 of B(s) 0 , whereas it becomes singular when restricted to such eigenspace; the magnitude of the singularity is precisely (µε)−(3−s), which is can- celled exactly by the pre-factor (µε)3−s in the l.h.s. of (iii). In fact, by assumption of non-degeneracy, the eigenspace −1 is spanned by φ and P := −φiheφ projects onto span{φ} with P φ = φ, as follows from the normalisation heφ, φiL2 = −1. Next, in order to see that the limit ε ↓ 0 in the r.h.s. of (iv) exists and is a 16 A. MICHELANGELI AND R. SCANDONE Combining (ii) and (iii) above yields = (cid:0)1 + P (B(s) − 1) + O(ε3−s)(cid:1)−1 (iv) as ε ↓ 0 in the operator norm topology. bounded operator, we write explicitly ε (cid:1)−1 (µε)3−s(cid:0)1 + B(s) 1 + P (B(s) − 1) = 1 − φiheφ(cid:16) ηs µ3−s u(−∆)− s µ3−sφiheφ − hv, φiL2 where we used the identities heφ, uiL2 = hφ, viL2 and (cid:10)eφ, u(−∆)− s ∀f ∈ L2(R3) . = 1 + (v) ηs 2 v + Setting the constants , 2πs sin 3π 2πs sin 3π 2πs sin 3π 2πs sin( 3π s ) φihv + φiheφ , 2 vf(cid:11)L2 = (cid:10)v(−∆)− s 2 ueφ, f(cid:11)L2 = (cid:10)(signV )u(−∆)− s = −heφ, fiL2 µ3−s + 1(cid:17)(cid:16) ηs a := (cid:16) ηs s (cid:17)−1 µ3−s + hv, φiL22 s (cid:17)−1 s (cid:16) ηs µ3−s + hv, φiL22 b := − hv, φiL2 (cid:0)1 + P (B(s) − 1)(cid:1)(cid:0)1 + aφiheφ + b φihv(cid:1) = 1 (µε)3−s(cid:0)1 + B(s) = (cid:0)1 + P (B(s) − 1))(cid:1)−1 ε (cid:1)−1 s (cid:17) φ (cid:0)1 + P (B(s) − 1)(cid:1)φ = −(cid:16) ηs µ3−s + hv, φiL22 φ = −(cid:16) ηs s (cid:17)−1 (cid:0)1 + P (B(s) − 1)(cid:1)−1 µ3−s + hv, φiL22 2πs sin 3π 2πs sin 3π lim ε↓0 φ . P in the operator norm topology. Last, from (v), using heφ, φiL2 = −1 and heφ, uiL2 = hφ, viL2 , one finds (vi) and hence the expression (v) allows one to compute explicitly (using again heφ, φiL2 = −1) and therefore to deduce that (1 + P (B(s) − 1))−1 exists and is bounded. This fact allows one to deduce from (iv) that (P + O(ε3−s)) 1 2πs sin( 3π s ) uihv − 1(cid:17) 2 vφ, f(cid:11)L2 Plugging the latter identity into (vi) yields finally (4.13) as a limit in the operator (cid:3) norm. We are now in the condition to prove Theorem 2.3. Proof of Theorem 2.3. Owing to (4.7) we need to determine the limit of ε ε3−sη(ε)(1 + B(s) ε )−1C(s) ε −A(s) as ε ↓ 0. As observed already, if u(−∆)− s expression vanishes with ε and 2 v has no eigenvalue −1, then the above ε (cid:0)h(s/2) + µs 1(cid:1)−1 ε↓0 −−−→ (cid:0)(−∆)s/2 + µs 1(cid:1)−1 in the operator norm. If instead u(−∆)− s be (−∆) 2 v does admit a simple eigenvalue −1, 2 + V zero-energy resonant or not, we are under the Assumption (Is) and s POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS17 (IIs) of the present Section and we can therefore apply the limits (4.8), (4.10), and (4.13). This yields −A(s) ε ε3−sη(ε)(1 + B(s) ε )−1C(s) ε ε↓0 −−−→ − Gs,µsihv ◦(cid:16)ηs + () = − ηs + hv, φiL22 µ3−shv, φiL22 2πs sin 3π s µ3−shv, φiL22 2πs sin 3π s (cid:17)−1 Gs,µsihGs,µs φiheφ ◦ uihGs,µs in the operator norm, having used heφ, uiL2 = hφ, viL2 . Now, if (−∆) 2 + V is not zero-energy resonant, then hv, φiL2 = 0, owing to Theorem 2.2(iv), and the conclusion is again s ε (cid:0)h(s/2) + µs 1(cid:1)−1 ε↓0 −−−→ (cid:0)(−∆)s/2 + µs 1(cid:1)−1 s in the operator norm. This proves part (i) of the present Theorem. (−∆) into (4.7) yields If instead 2 + V is zero-energy resonant, then using hv, φiL2 6= 0 and plugging () back ε (cid:0)h(s/2) + µs 1(cid:1)−1 ε↓0 −−−→ (cid:0)(−∆)s/2 + µs + 1(cid:1)−1 hv, φiL22 − −ηs 1 Gs,µsihGs,µs µ3−s 2πs sin 3π s in the operator norm. Upon setting α := −ηshv, φiL2−2 and λ = µs, and compar- ing the resulting expression with (2.12), this means ε (cid:0)h(s/2) + λ1(cid:1)−1 ε↓0 −−−→(cid:0)(−∆)s/2 + λ1(cid:1)−1 = (k(s/2) + λ1)−1 , +(cid:0)α − λ α 3 s −1 2πs sin( 3π s )(cid:1)−1 Gs,λihGs,λ which proves part (ii) of the Theorem. (cid:3) 5. Convergence of the 1D limit: resonant-driven case. The proof of the limit h(s/2) 2 , 1) (Theorem 3.3) is technically analogous to that in three dimensions. Therefore, based on the detailed discussion of the preceding Section, we only present here the steps of the convergence scheme and a sketch of their proofs. in dimension one when s ∈ ( 1 ε↓0 −−−→ k(s/2) α ε Prior to that, let us set up the key resolvent identity and useful scaling properties with a notation that we can use also in Section 6 when we will deal with the resonant-independent limit. We then keep s ∈ ( 1 2 , 1) ∪ (1, 3 re-write (3.14) and (3.17) as 2 ) generic for a moment and, in a unified form, we (5.1) Vε(x) = η(ε) s+γ 2 ε V ( x ε ) . Taking γ = s in (5.1) yields (3.14) and taking γ = 2− s yields (3.17). Thus, setting (5.2) one has (5.3) vε(x) = v(x) := V (x) vε(x) := Vε(x) 1 2 , 1 2 , √η(ε) ε(s+γ)/4 v( x ε ) , 1 2 sign(V (x)) , 2 sign(Vε(x)) , 1 u(x) := V (x) uε(x) := Vε(x) √η(ε) ε(s+γ)/2 u( x uε(x) = ε ) , vεuε = Vε . A. MICHELANGELI AND R. SCANDONE The 1D analogue Uε : L2(R) → L2(R) of the unitary scaling operator (4.4) acts 18 as (5.4) (Uεf )(x) := 1 ε1/2 f ( x ε ) , which induces the scaling transformations (5.5) U∗ε vεUε = pη(ε) U∗ε uεUε = pη(ε) 1(cid:1)−1 Based on arguments that differ depending on whether s ∈ ( 1 Uε = εs(cid:0)(−∆)s/2 + λεs 2 , 1) or s ∈ (1, 3 2 ) and which we shall prove in due time, the Konno-Kuroda-type resolvent identity u , v , s+γ s+γ ε ε . 4 4 (5.6) = (cid:0)(−∆)s/2 + λ1(cid:1)−1 vε(cid:16)1 + uε(cid:0)(−∆)s/2 + λ1(cid:1)−1 − vε(cid:17)−1 uε(cid:0)(−∆)s/2 + λ1(cid:1)−1 holds as an identity between bounded operators on L2(R) for every ε > 0 and every −λ < 0 in the resolvent set of h(s/2) . Inserting UεU∗ε = 1 into (5.6) and applying ε (5.5) then yields U∗ε(cid:0)(−∆)s/2 + λ1(cid:1)−1 + λ1(cid:1)−1 (cid:0)h(s/2) −(cid:0)(−∆)s/2 + λ1(cid:1)−1 ε η(ε)(1 + B(s) ε )−1C(s) ε , (5.7) ε (cid:0)h(s/2) + λ1(cid:1)−1 = (cid:0)(−∆)s/2 + λ1(cid:1)−1 having defined (5.8) A(s) ε B(s) ε C(s) ε := ε− 2−s−γ 2 s−γ := η(ε) ε := U∗ε uε (η(ε))− 1 2−s−γ 2 − A(s) ε ε (cid:0)(−∆)s/2 + λ1(cid:1)−1 2 u(cid:0)(−∆)s/2 + λεs 1(cid:1)−1 2(cid:0)(−∆)s/2 + λ1(cid:1)−1 ε , B(s) v (η(ε))− 1 2 vε Uε ε− 2−s−γ 2 . We shall see in a moment (Lemma 5.3) that A(s) operators on L2(R). ε , and C(s) ε are Hilbert-Schmidt The following scaling property too is going to be useful in both regimes s ∈ ( 1 2 , 1) and s ∈ (1, 3 2 ). Lemma 5.1. For any s, γ, ε > 0 and any x ∈ R\{0} one has (5.9) 2 Gs,λεs (x) = ε 2 Gs,λ(εx) . 2−s−γ s−γ ε Proof. Owing to (3.1), s−γ 2 Gs,λεs (x) = ε = whence the thesis. 1 2π 1 2π ε ε s−γ 2 ZR ZR 2 2−s−γ dp eipx 1 ps + λεs 1 dp eip(εx) ps + λ = ε 2−s−γ 2 Gs,λ(εx) , (cid:3) We can now start the discussion for the proof of Theorem 3.3, thus working in the regime s ∈ ( 1 2 , 1). First, we have the following properties. Lemma 5.2. Let V : R → R belong to L1(R) ∩ Rs,1 for some s ∈ ( 1 2 , 1). Then: 2 ((−∆) s 2 + λ1)−1V 1 2 is a Hilbert-Schmidt operator on L2(R); (i) for every λ > 0, V 1 (ii) V 1 (iii) the operator (−∆) 2 ≪ (−∆) V ) = [0, +∞). s s 4 in the sense of infinitesimally bounded operators; 2 +V defined as a form sum is self-adjoint and σess((−∆) s 2 + POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS19 Proof. The proof is completely analogous to that of Lemma 4.1 for the 3D case, and is based on the fact that the integral kernel of V 1 2 , namely V (x) 1 2 , belongs to L2(R × R, dx dy), as a direct consequence of the assumption V ∈ L1(R) ∩ Rs,1. (cid:3) 2 Gs,λ(x − y)V (y) 1 2 + λ1)−1V 1 2 ((−∆) s Lemma 5.2 justifies the validity of the resolvent identity (5.6), and hence of the rescaled identity (5.7), owing again to the general argument of [3, Theorem B.1(b)]. Next, we monitor separately the following limits. Lemma 5.3. Let V and η satisfy Assumption (I−s ) for some s ∈ ( 1 ε > 0, the operators A(s) are Hilbert-Schmidt operators on L2(R) with limit 2 , 1). For every defined by (5.1)-(5.3) and (5.8) with γ = s ε , and C(s) ε , B(s) ε (5.10) (5.11) (5.12) lim ε↓0 lim ε↓0 lim ε↓0 A(s) ε = Gs,λihv ε = B(s) B(s) 0 = u (−∆)− s 2 v C(s) ε = uihGs,λ in the Hilbert-Schmidt operator norm. Proof. Completely analogous to the proof of Lemma 4.4, the integral kernels being now (with γ = s) A(s) ε (x, y) = Gs,λ(x − εy) v(y) B(s) ε (x, y) = η(ε) u(x) Gs,λεs (x − y) v(y) C(s) ε (x, y) = u(x) Gs,λ(εx − y) . In particular, owing to (5.9), B(s) ε (x, y) = η(ε) ε1−su(x) Gs,λ(εx − εy) v(y) , and using (3.12)-(3.13) one finds B(s) ε (x, y) −−−→ u(x) 21−sΓ( 1−s ε↓0 2 ) 1 2 Γ( s 2 ) (2π) pointwise almost everywhere. 1 x − y1−s v(y) = B(s) 0 (x, y) (cid:3) 0 ε ε (cid:0)h(s/2) ε ε1−sη(ε)(1 + B(s) is compact, (1 + B(s) + λ1(cid:1)−1 + λ1(cid:1)−1 (5.13) we see that, since B(s) Assumption (II−s ). Assumption (I−s ) holds. B(s) →(cid:0)(−∆)s/2 + λ1(cid:1)−1 = (cid:0)(−∆)s/2 + λ1(cid:1)−1 Before plugging the limits found in Lemma (5.3) into (5.7), that now reads ε )−1C(s) − A(s) 0 )−1 exists everywhere defined and as ε ↓ 0, unless B(s) admits an eigenvalue −1. We then consider the following additional assumption. 0 has eigenvalue −1, which is 0 φ = −φ and, in bounded, in which case(cid:0)h(s/2) non-degenerate. φ ∈ L2(R3) is a non-zero function such that B(s) addition, heφ, φiL2 = −1, where eφ := (signV )φ. Since heφ, φiL2 = −h(signV )φ, (signV )v(−∆)− s normalisation heφ, φiL2 = −1 is always possible. ε )−1 becomes singular in the limit ε ↓ 0, with a singularity that now competes with the vanishing factor ε1−s of (5.13). To resolve this competing effect, we need first to expand B(s) ε around ε = 0 to a further order, than the limit (5.11). This expansion is valid irrespectively of Assumption (II−s ). Lemma 5.4. Let s ∈ ( 1 When Assumption (II−s ) holds, (1 + B(s) 2 vφiL2 = −k(−∆)− s 4 vφk2 2 , 1) and λ > 0. , ε 0 L2, the 20 A. MICHELANGELI AND R. SCANDONE (5.14) (i) For every x ∈ R\{0} lim λ↓0 s ))−1λ 1 (ii) In the norm operator topology one has Gs,λ(x) − Gs,0(x) (s sin( π s −1 = 1 . (5.15) 1 1 lim ε↓0 s −1ε1−s (cid:0)B(s) 0 (cid:1) = ε − B(s) s sin( 3π Here ηs ∈ R is the constant that is part of Assumption (I−s ). ηs s −1 0 + B(s) λ λ 1 1 s ) uihv . Proof. Completely analogous to the proof of Lemma 4.5 for the 3D case. ε )−1 as ε ↓ 0. We can now monitor the competing effect in ε1−s(1 + B(s) Lemma 5.5. Under the Assumptions (I−s ) and (II−s ) one has (cid:3) (5.16) lim ε↓0 in the operator norm topology. ε (cid:1)−1 ε1−s(cid:0)1 + B(s) = (cid:16)ηs + hv, φiL22 s −1s sin π s(cid:17)−1 λ 1 φiheφ Proof. Completely analogous to the proof of Lemma 4.6 for the 3D case. (cid:3) With these preliminaries at hand, we can prove Theorem 3.3. Proof of Theorem 3.3. The argument is the very same as the in the proof of The- orem 2.3 for the 3D case. Thus, the limit is the trivial one unless the potential in the approximating operators satisfy Assumptions (I−s ) and (II−s ), in which case, plugging the limits (5.10), (5.12), and (5.16) into (5.13), one has ε 1 1 + ε↓0 + µs λ1− 1 −ηs s s sin 3π s 1(cid:1)−1 (cid:0)h(s/2) Gs,µsihGs,µs . −−−→ (cid:0)(−∆)s/2 + µs 1(cid:1)−1 hv, φiL22 − The comparison of the limit resolvent above with formulas (3.4) and (3.8) shows finally that the limit resolvent is precisely (k(s/2) + λ1)−1 where the extension parameter satisfies α = −ηsRR dx V (x)ψ(x)−2, and this completes the proof. (cid:3) 6. Convergence of the 1D limit: resonant-independent case. This Section contains the proof of Theorem 3.4. Thus, now s ∈ (1, 3 First, we observe that with L1-potentials the following operator-theoretic prop- formulas (5.1)-(5.9) must be specialised with γ = 2 − s. erties hold. Lemma 6.1. Let V : R → R belong to L1(R) and let s ∈ (1, 3 2 ). Then: 2 ) and α 2 ((−∆) s 2 + λ1)−1V 1 2 is a Hilbert-Schmidt operator on L2(R); (i) for every λ > 0, V 1 (ii) V 1 (iii) the operator (−∆) 2 ≪ (−∆) V ) = [0, +∞). s s 4 in the sense of infinitesimally bounded operators; 2 +V defined as a form sum is self-adjoint and σess((−∆) s 2 + Proof. Since s > 1, (3.1) defines a function dGs,λ ∈ L1(R), whence Gs,λ ∈ C∞(R) (continuous and vanishing at infinity). Therefore, the integral kernel of V 1 2 ((−∆) 2 + λ1)−1V 1 2 Gs,λ(x − y)V (y) 1 2 , belongs to L2(R × R, dx dy), and this holds for any λ > 0. Based on this observation, the rest of the reasoning of the (cid:3) proof of Lemma 4.1 can be repeated verbatim. 2 , namely V (x) 1 s POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS21 Following again the general argument of [3, Theorem B.1(b)], Lemma 6.1 justifies the validity of the resolvent identity (5.6), and hence of the rescaled identity (5.7), that now reads − η(ε) A(s) ε (1 + B(s) ε )−1C(s) ε (6.1) ε (cid:0)h(s/2) + λ1(cid:1)−1 = (cid:0)(−∆)s/2 + λ1(cid:1)−1 for every ε > 0 and every −λ < 0 in the resolvent set of h(s/2) Lemma 6.2. Let V and η satisfy Assumption (I+ ε > 0, the operators A(s) γ = 2 − s are Hilbert-Schmidt operators on L2(R) with limit (6.2) ε , and C(s) ε , B(s) ε ε . s ) for some s ∈ (1, 3 2 ). For every defined by (5.1)-(5.3) and (5.8) with lim ε↓0 A(s) ε = Gs,λihv ε = B(s) B(s) 0 = C(s) ε = uihGs,λ (6.3) (6.4) lim ε↓0 lim ε↓0 η(0) s s sin π s λ1− 1 uihv in the Hilbert-Schmidt operator norm. Proof. The integral kernels are now A(s) ε (x, y) = Gs,λ(x − εy) v(y) B(s) ε (x, y) = η(ε) εs−1 u(x) Gs,λεs (x − y) v(y) C(s) ε (x, y) = u(x) Gs,λ(εx − y) . ε and C(s) For A(s) Schmidt as a consequence of Lemma 6.1. Re-writing ε we reason precisely as in the proof of Lemma 4.4. B(s) ε is Hilbert- B(s) ε (x, y) = η(ε) u(x) Gs,λ(εx − εy) v(y) by means of (5.9), and observing that (3.1) implies one deduces Gs,λ(εx − εy) ε↓0 −−−→ Gs,λ(0) = 1 s s sin π s λ1− 1 , B(s) ε (x, y) ε↓0 −−−→ η(0) s s sin π s λ1− 1 u(x)v(y) . Then a dominated convergence argument, analogous to that used in the proof of (cid:3) Lemma 4.4, proves (6.3). It is now convenient to observe the following (see [7, Lemma 5.1] for an analogous argument). Lemma 6.3. Assume that the data s ∈ (1, 3 set of all the h(s/2) exceptional relation ε 's, and V and η matching Assumption (I+ 2 ), λ > 0 with −λ in the resolvent s ), do not satisfy the (6.5) 1 + Then the operator 1 + B(s) on L2(R). 0 η(0) s s sin π λ1− 1 s ZR dxV (x) = 0 . is invertible with bounded inverse, everywhere defined 22 A. MICHELANGELI AND R. SCANDONE 0 Proof. Since B(s) to prove that the validity of (6.5) is equivalent to B(s) fact, B(s) is compact on L2(R), based on the Fredholm alternative we have 0 having eigenvalue −1. In 0 φ = −φ for some non-zero φ ∈ L2(R) is the same as φ = η(0) s s sin π s λ1− 1 hv, φiL2 u , meaning that φ is not orthogonal to v and φ is a multiple of u. When this is the case, u itself must be an eigenfunction of B(s) 0 with eigenvalue −1, and this is (cid:3) tantamount, owing to the identity above, as the validity of (6.5). For given s, η, and V , the exceptional value of −λ satisfying (6.5) is going to described in Theorem 3.1(iv). As correspond to the negative eigenvalue of k(s/2) we are going to monitor the limit h(s/2) in the resolvent sense, not only must we discard the spectral points −λ not belonging to the resolvent set of all the h(s/2) 's, but also the point −λ given by (6.5). Thus, for our purposes the operator ε 1 + B(s) is always invertible with everywhere defined bounded inverse. α ε↓0 −−−→ k(s/2) α ε 0 In particular, (6.3) implies (6.6) in the operator norm. (1 + B(s) ε )−1 ε↓0 −−−→ (1 + B(s) 0 )−1 Based on the preceding preparatory materials, we can now prove Theorem 3.4. Proof of Theorem 3.4. Since (6.5) is excluded and therefore then plugging the limits (6.2), (6.4), and (6.6) into (6.1) yields (1 + B(s) 0 )−1u = (cid:16)1 + u , λ1− 1 s s sin π η(0)RR dx V (x) s (cid:17)−1 η(0)RR dx V (x) η(0) RR dx V (x) λ1− 1 1 + − s s sin π s Gs,λihGs,λ ε (cid:0)h(s/2) + λ1(cid:1)−1 in the operator norm. Upon setting ε↓0 −−−→ (cid:0)(−∆)s/2 + λ1(cid:1)−1 α := −(cid:16)η(0)ZR −−−→(cid:0)(−∆)s/2 + λ1(cid:1)−1 ε↓0 dx V (x)(cid:17)−1 = (k(s/2) α + λ1)−1 , and comparing the resulting expression with (3.4) and (3.8), one finds ε (cid:0)h(s/2) + λ1(cid:1)−1 + α − 1 1 λ1− 1 s s sin π s Gs,λihGs,λ which completes the proof. (cid:3) 7. Zero-energy resonances for Schrodinger operators with fractional Laplacian The purpose of this Section is two-fold. First, we prove Theorems 2.2 and 3.2, concerning the characterisation of the zero-energy resonant behaviour of (−∆)s/2 + V . Then, we discuss the occurrence of zero-energy resonances, both in one and three dimension. POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS23 Proof of Theorem 2.2. The fact that for a real-valued V ∈ L1(R3)∩Rs,3 the opera- tor u(−∆) 2 v is Hilbert-Schmidt follows from Lemma 4.1(i), thus part (i) is proved. s Let us split 1 x3−s(cid:17)v(y)φ(y) where Λs is the constant defined in (2.8). We now see that ψ1 ∈ L2(R3). To this 1 = dy y Λs (a) ≡ one has ψ(x) = ((−∆)− s Λshv, φiL2 x − y3−s v(y)φ(y) x − y3−s − 2 vφ)(x) = ZR3 x3−s + ΛsZR3 dy(cid:16) Λshv, φiL2 x3−s + ψ1(x) , aim, we observe that setting by := y dx(cid:16) x − y3−s − = y2s−3ZR3 . y2s−3ZR3 x3−s(cid:17)2 x −by 3−sx3−s (cid:17)2 dx(cid:16)x −by 3−s − x3−s x −by 3−sx3−s(cid:17)2 dx(cid:16) (cid:12)(cid:12)x −by 3−s − x3−s(cid:12)(cid:12) . hxi2−s in the last step. Since s ∈ ( 3 x3−s(cid:17)2 . y2s−3 . ZR3 ZR3 hxi2−s (b) 1 1 , As a consequence, 1 1 dx(cid:16) x − y3−s − L2(R3) . ZR3 dx(cid:12)(cid:12)(cid:12)ZR3 dy(cid:16) 6 ZZR3×R3 dx dy(cid:16) . ZR3 x − y3−s − x − y3−s − dy V (y)y2s−3 < +∞ , 1 1 kψ1k2 2 1 x3−s(cid:17)v(y)φ(y)(cid:12)(cid:12)(cid:12) x3−s(cid:17)2 V (y) 1 having used the change of variable x 7→ y x in the first step and the uniform bound 2 ), the last integral above is finite, thus we deduce 2 , 5 as follows from a Cauchy-Schwartz inequality in the second step, from the bound (b) in the third step, and from the assumption V ∈ L1(R3,hxi2s−3dx) in the last step. 2 , then identity (a) implies that ψ ∈ Since x−(3−s) ∈ L2 loc(R3). Moreover, from (2.17) and (2.18) one finds loc(R3), because s > 3 L2 V ψ = vu(−∆)− s 2 vφ = −vφ = −(−∆) s 2 ψ , whence ((−∆) s 2 + V )ψ = 0 distributionally. This completes the proof of part (ii). Using (2.18) and the distributional identity proved in part (ii) one finds hv, φiL2 = ZR3 dx v(x) φ(x) = ZR3 dx ((−∆) s 2 ψ)(x) = −ZR3 dx V (x) ψ(x) , which proves part (iii). Last, the identity (a) also implies that ψ ∈ L2(R3) is equivalent to hv, φiL2 = 0. 2 + V )ψ = 0 holds in the L2-sense, (cid:3) When this is the case, the identity ((−∆) implying that ψ ∈ D((−∆) 2 + V ). This completes the proof of part (iv). s s The proof of Theorem 3.2 proceeds along the same line. 24 A. MICHELANGELI AND R. SCANDONE Proof of Theorem 3.2. Part (i) follows from Lemma 5.2(i). Splitting () ψ(x) = ((−∆)− s 2 vφ)(x) = Λshv, φiL2 x1−s + ψ1(x) , where now Λs is the constant defined in (3.13), and using the assumptions V ∈ L1(R,hxi2s−1dx) we can show that ψ1 ∈ L2(R), following the same argument as in the above proof of Theorem 2.2. Therefore, since x−(1−s) ∈ L2 loc(R) because s > 1 loc(R). Moreover, from (3.15) and (3.16) one finds −V ψ = (−∆) 2 + V )ψ = 0 distributionally. Thus, part (ii) is proved. 2 , from () one deduces that ψ ∈ L2 2 ψ, whence ((−∆) s s Next, (3.16) and the distributional identity of part (ii) imply hv, φiL2 = −ZR dx V (x) ψ(x) , which proves part (iii). Last, the identity (*) also implies that ψ ∈ L2(R) is equivalent to hv, φiL2 = 0, 2 + V )ψ = 0 in the L2-sense, and this implies that ψ ∈ (cid:3) s 2 + V ). Thus, part (iv) is proved. s in which case ((−∆) D((−∆) Let us address now the question of the existence of a potential V such that the fractional Schrodinger operator (−∆)s/2 + V is zero-energy resonant. Proposition 7.1. Let θ ∈ S(R3), with θ > 0. Define ψ := θ ∗ Gs,0 = θ ∗ Λs x3−s , V := − θ ψ , where Λs is the constant defined in (2.8). Then V satisfies part (i) of Assump- tion (Is), and (−∆)s/2 + V is zero-energy resonant on L2(R3), with zero-energy resonance ψ. Proof. By construction ψ > 0, being the convolution of two strictly positive func- tions. Moreover, from bψ(p) = bθ(p) ps one sees that ψ is continuous, as bψ ∈ L1(R3), and that ψ /∈ L2(R3), as bψ /∈ L2(R3) either. Still, for every compact K ⊂ R3 k(θ ∗ Gs,0)1KkL2 6 kθkL 5 kGs,0kL2k1KkL6 < +∞ , 6 as follows by means of a Schwartz and a Young inequality. Thus, (i) ψ ∈ L2 loc(R3) \ L2(R3) . Next, we argue that the leading decay in ψ is x−(3−s). To see that, since θ(x) . hxi−m for any m ∈ N, we write y3−s dy ψ(x) . ZR3 hx − yim 1 1 (ii) = Zy−x> 1 1 2 x hy − xim 1 y3−s dy +Zy−x< 1 2x 1 hy − xim 1 y3−s dy POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS25 and we take non-restrictively x > 1 and m > 4. The first integral in the r.h.s. of (ii) is estimated as 1 y3−s dy 1 hy − xim 1 2x Zy−x> 1 = Zy−x> 1 xm−4ZR3 2 x 2x> 3 y> 3 1 . 2 1 y3−s dy +Zy−x> 1 xmZy< 3 y< 3 y3−s . 2 x 2x dy 2x 1 1 hy − xim 1 hy − xim hy − xi4 + dy 1 y3−s dy xm−4 + 1 xm−s and hence vanishes as x → +∞ faster than any power. Concerning the second integral in the r.h.s. of (ii), since y − x < 1 2x implies y > 1 1 dy 2x, one has x3−s . 1 hy − xim . 1 y3−s dy . x3−sZR3 ψ(x) . hxi−(3−s) . Zy−x< 1 1 2x hy − xim Therefore, (ii) implies (iii) Now, splitting ψ = Λsbθ(0) x3−s + ψ1 , ψ1 = (cid:16)bθ(·) −bθ(0) · s (cid:17)∨ and observing that cψ1 ∈ L2(R3) (owing to the bound bθ(p) −bθ(0) . p, valid for small p), whence also ψ1 ∈ L2(R3), we deduce that in order for the decay ψ1(x) . hxi−(3−s) (whose bound is not square-integrable) implied by (iii) above to be compatible with the square-integrability of the continuous function ψ1, ψ1 itself must decay more that hxi−(3−s), which allows us to conclude that the leading decay in ψ is x−(3−s). Since ψ is continuous, positive, and with polynomial decay x−(3−s) at infinity, then 1/ψ is continuous, positive, and with polynomial growth at infinity. Since θ is a Schwartz function, we conclude that V = θ 1 ψ is continuous and decays at infinity faster than any polynomial. This proves that V satisfies part (i) of Assumption (Is). As a consequence, V ∈ L1(R3,hxi2s−3dx) and also kV kRs,3 . kV kL3/s < +∞. By construction, (−∆)s/2ψ = θ, whence distributionally. This also implies (iv) (v) ((−∆)s/2 + V )ψ = 0 (−∆)−s/2V ψ = −ψ . Writing as usual V = uv with v = V 1 2 and u = V 1 2 sign(V ), let us now set φ := −uψ. Then (v) yields (vi) ψ = (−∆)− s 2 vφ , which is precisely the relation between ψ and φ given by (2.18). φ cannot be iden- tically zero, because so would be ψ, owing to (vi), which is not the case. Moreover, φ is square-integrable, because in the product uψ the L2 loc-function ψ is multiplied by a function that decays more than polynomially at infinity. Therefore, (vii) and, multiplying (v) by u, (viii) as an identity in L2(R3). φ ∈ L2(R3) \ {0} , u(−∆)− s 2 vφ = −φ 26 A. MICHELANGELI AND R. SCANDONE Conditions (i), (iv), (vi), (vii), and (viii) above, owing to Theorem 2.2 and to the present definition of resonance, imply that (−∆)s/2 + V is zero-energy resonant, (cid:3) with zero-energy resonance ψ. The counterpart result in 1D, which we state here below, has a completely anal- ogous proof, that we then omit. Proposition 7.2. Let θ ∈ S(R), with θ > 0. Define ψ := θ ∗ Gs,0 = θ ∗ V := − θ ψ , Λs x1−s , where Λs is the constant defined in (3.13) Then V satisfies part (i) of Assumption (I−s ), and (−∆)s/2+V is zero-energy resonant on L2(R), with zero-energy resonance ψ. Remark 7.3. By means of a more refined discussion, in the same spirit of [11], we can identify the threshold coupling parameter λ ∈ R, for a given potential V in a suitable class, for which (−∆)s/2 + λV is zero-energy resonant. We do not develop this interesting approach here. References [1] S. Albeverio, F. Gesztesy, R. Hø egh-Krohn, and H. Holden, Point interactions in two dimensions: basic properties, approximations and applications to solid state physics, J. Reine Angew. Math., 380 (1987), pp. 87–107. [2] S. Albeverio, F. Gesztesy, and R. Høegh-Krohn, The low energy expansion in nonrela- tivistic scattering theory, Ann. Inst. H. Poincar´e Sect. A (N.S.), 37 (1982), pp. 1–28. [3] S. Albeverio, F. Gesztesy, R. Høegh-Krohn, and H. Holden, Solvable Models in Quan- tum Mechanics, Texts and Monographs in Physics, Springer-Verlag, New York, 1988. [4] S. Albeverio, F. Gesztesy, R. Høegh-Krohn, and W. Kirsch, On point interactions in one dimension, J. Operator Theory, 12 (1984), pp. 101–126. [5] S. Albeverio and P. Kurasov, Singular perturbations of differential operators, vol. 271 of London Mathematical Society Lecture Note Series, Cambridge University Press, Cambridge, 2000. Solvable Schrodinger type operators. [6] E. Capelas de Oliveira and J. J. Vaz, Tunneling in fractional quantum mechanics, J. Phys. A, 44 (2011), pp. 185303, 17. [7] G. Dell'Antonio and A. Michelangeli, Schrodinger operators on half-line with shrinking potentials at the origin, Asymptot. Anal., 97 (2016), pp. 113–138. [8] M. Gallone, A. Michelangeli, and A. Ottolini, Kreın-Visik-Birman self-adjoint exten- sion theory revisited, SISSA preprint 25/2017/MATE (2017). [9] S. Jarosz and J. J. Vaz, Fractional Schrodinger equation with Riesz-Feller derivative for delta potentials, J. Math. Phys., 57 (2016), pp. 123506, 16. [10] T. Kato, Perturbation theory for linear operators, Classics in Mathematics, Springer-Verlag, Berlin, 1995. Reprint of the 1980 edition. [11] M. Klaus and B. Simon, Coupling constant thresholds in nonrelativistic quantum mechanics. I. Short-range two-body case, Ann. Physics, 130 (1980), pp. 251–281. [12] R. Konno and S. T. Kuroda, On the finiteness of perturbed eigenvalues, J. Fac. Sci. Univ. Tokyo Sect. I, 13 (1966), pp. 55–63 (1966). [13] E. K. Lenzi, H. V. Ribeiro, M. A. F. dos Santos, R. Rossato, and R. S. Mendes, Time dependent solutions for a fractional Schrodinger equation with delta potentials, J. Math. Phys., 54 (2013), pp. 082107, 8. [14] A. Michelangeli, A. Ottolini, and R. Scandone, Fractional powers and singular pertur- bations of differential operators, arXiv:1801.08885 (2018). [15] S. I. Muslih, Solutions of a particle with fractional δ-potential in a fractional dimensional space, Internat. J. Theoret. Phys., 49 (2010), pp. 2095–2104. [16] M. M. Nayga and J. P. Esguerra, Green's functions and energy eigenvalues for delta- perturbed space-fractional quantum systems, J. Math. Phys., 57 (2016), pp. 022103, 7. [17] E. C. d. Oliveira, F. S. Costa, and J. J. Vaz, The fractional Schrodinger equation for delta potentials, J. Math. Phys., 51 (2010), pp. 123517, 16. [18] A. Sacchetti, Stationary solutions of a fractional Laplacian with singular perturbation, arXiv:1801.01694 (2018). POINT-LIKE PERTURBED FRACTIONAL LAPLACIANS WITH SHRINKING POTENTIALS27 [19] T. Sandev, I. Petreska, and E. K. Lenzi, Time-dependent Schrodinger-like equation with nonlocal term, J. Math. Phys., 55 (2014), pp. 092105, 10. [20] B. Simon, Trace ideals and their applications, vol. 120 of Mathematical Surveys and Mono- graphs, American Mathematical Society, Providence, RI, second ed., 2005. [21] J. D. Tare and J. P. H. Esguerra, Bound states for multiple Dirac-δ wells in space- fractional quantum mechanics, J. Math. Phys., 55 (2014), pp. 012106, 10. (A. Michelangeli) International School for Advanced Studies – SISSA, via Bonomea 265, 34136 Trieste (Italy). E-mail address: [email protected] (R. Scandone) International School for Advanced Studies – SISSA, via Bonomea 265, 34136 Trieste (Italy). E-mail address: [email protected]
1503.08574	1	1503	2015-03-30T07:45:57	Common hypercyclic vectors for high dimensional families of operators	[ "math.FA", "math.DS" ]	Let $(T\_\lambda)\_{\lambda\in\Lambda}$ be a family of operators acting on a $F$-space $X$, where the parameter space $\Lambda$ is a subset of $\mathbb R^d$. We give sufficient conditions on the family to yield the existence of a vector $x\in X$ such that, for any $\lambda\in\Lambda$, the set $\big\{T\_\lambda^n x;\ n\geq 1\big\}$ is dense in $X$. We obtain results valid for any value of $d\geq 1$ whereas the previously known results where restricted to $d=1$. Our methods also shed new light on the one-dimensional case.	math.FA	math	COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS FRÉDÉRIC BAYART ABSTRACT. Let (Tλ)λ∈Λ be a family of operators acting on a F -space X , where the parameter space Λ is a subset of Rd . We give sufficient conditions on the family to yield the existence of x; n ≥ 1ª is dense in X . We obtain results a vector x ∈ X such that, for any λ ∈ Λ, the set©T n valid for any value of d ≥ 1 whereas the previously known results where restricted to d = 1. Our methods also shed new light on the one-dimensional case. λ 1. INTRODUCTION Let X be a separable F -space (namely a separable topological vector space which carries a complete translation-invariant metric), and let T ∈ L (X ). We say that T is hypercyclic provided there exists a vector x ∈ X such that its orbit O(x, T ) = {T n x; n ≥ 0} is dense in X . The vector x is called a hypercyclic vector for T and the set of hypercyclic vectors for T will be denoted by HC (T ). More generally, let (Tn) be a sequence of operators acting on X . We say that x is hypercyclic for (Tn) if {Tn x; n ≥ 0} is dense in X and we denote by HC (Tn) the set of hypercyclic vectors for (Tn). Hypercyclic operators have been intensively studied in the last few decades (see [6] and [11]). One of the most interesting problem in this field is to find, for a given family of hypercyclic operators, a common hypercyclic vector. It turns out that, as soon as T is hypercyclic, HC (T ) is a residual subset of X . Hence, for any countable set Λ, provided each Tλ, λ ∈ Λ, is hypercyclic, Tλ∈Λ HC (Tλ) is a residual subset of X and in particular is nonempty. When Λ is uncountable, the situation is more difficult and has attracted the attention of many mathematicians . In the litterature, we may find two kinds of results regarding common hyper- cyclicity. • algebraic results: these results were first obtained by Leon and Müller in [14] when they showed that, for any operator T ∈ L (X ) and any θ ∈ R, HC (e i θT ) = HC (T ). This re- sult was extended to C0-semigroup in [8] by Conejero, Müller and Peris: if (Tt )t>0 is a strongly continuous group on X , then for any t > 0, HC (Tt ) = HC (T1). • analytic results: the pioneering work in that direction is due to Abakumov and Gor- don ([1]) who showed that Tλ>1 HC (λB) is nonempty, where B is the (unweighted) [9] who showed thatTλ>1 HC (λB) is residual. Costakis and Sambarino gave a rather general criterion for a family (Tλ)λ∈I indexed by an interval I to have a residual set of common hypercyclic vectors. This criterion may be applied to many classical se- quences of operators, like translation operators τa, a ∈ C\{0}, which are defined on the backward shift on ℓ2. This was shortly later improved by Costakis and Sambarino in Date: August 22, 2018. 1991 Mathematics Subject Classification. 47B37. Key words and phrases. hypercyclic operators. 1 2 FRÉDÉRIC BAYART set of entire functions H(C) by τa( f ) = f (·+ a). More precisely, the criterion shows that Tθ∈[0,2π] HC (τei θ) is nonempty. It turns out that both the algebraic results and the analytic results are one-dimensional re- sults. They show that certain families indexed by a subset of R have a common hypercyclic vector. Sometimes, we can combine the two methods to obtain two-dimensional results. For instance, by the analytic method, you can show thatTλ>1 HC (λB) is residual and by the alge- braic method, you can show that for any θ ∈ R and any λ > 1, HC (e i θλB) = HC (λB). This yields the following two-dimensional result:Tλ>1 HC (λB) is residual. A similar argument is used to prove thatTa∈C\{0} HC (τa) is a residual subset of H(C). It was observed by Borichev (see [1]) that there are dimensional obstructions to the existence of a common hypercyclic vector. Indeed, let Λ ⊂ (1,+∞)2 and for λ = (s, t)∈ Λ, define Tλ = sB ⊕t B acting on ℓ2 ⊕ ℓ2. Then each Tλ is hypercyclic but if Tλ∈Λ HC (Tλ) is nonempty, then Λ has Lebesgue measure zero. See also [17] for other limitations relative to the dimension of the parameter space. However, there are at least two seminal papers where two-dimensional results do appear. The residual, with two successive applications of the algebraic results. The second one is due to Tsirivas in [19] (see also [18]). Tsirivas shows that if (λn) is an increasing sequence of positive first one is due to Shkarin in [17] who has proved thatTa,b∈C∗ HC (bτa) is a residual subset of H(C). The proof combines a two-dimensional analytic result, precisely Tb>0,a∈S1 HC (bτa) is real numbers tending to +∞ such that λn+1/λn goes to 1, thenTa∈C\{0} HC (τλn a) is a residual subset of H(C). This is a two-dimensional analytic result, since we cannot apply the algebraic results when λn 6= n. Both results of Shkarin and Tsirivas are truely "tours de force" which seem specific to the trans- lation operators on H(C) or at least to operators very similar to them. In particular, it is not clear if their arguments may be adapted to higher-dimensional families or to operators acting on a Banach space and not on a Fréchet space. In this paper, we provide a new approach which allows us to prove common hypercyclic results for general high-dimensional families. The very simple main idea is the following. The key point in Borichev's example is the fact that if λnB n x is close to y, then µnB n x cannot be close to y provided µ is far away from λ. Now, if you are working with the group of translations (τa), then f (x+na) and f (x+nb) can be simultaneously close to g even if b is far away from a. Indeed, this just mean that f has to be close to g on the balls centered in −na and in −nb, and these conditions are in some way independent. This will allow us, in order to construct a common hypercyclic vector f , to use the same n for different values of the parameter! Here is our main result. Theorem 1.1. Let (Ta)a∈Rd be a strongly continuous group on Rd with the uniform mixing prop- erty. ThenTa∈Rd\{0} HC (Ta) is a residual subset of X We shall define later the uniform mixing property, but it is a rather natural condition which is satisfied by many operator groups. By applying Theorem 1.1, we will get many new examples of common hypercyclicity which are not reachable with the previously known arguments and for high-dimensional families! COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 3 We shall use two main ingredients in our proof. Firstly we translate the problem of finding a common hypercyclic vector to the problem of finding a suitable covering of compact subsets of Rd . Secondly we give a way to produce such coverings. It is based on a method to split sequences of real numbers which are going to infinity but not too quickly, and in fact our state- ment is more general than Theorem 1.1 since it covers sequences (Tλn a) and not only iterates (Tna). Even for operator groups like the translation group, there are obstructions to the existence of a common hypercyclic vector for the sequences of operators (τλn a), a ∈ C\{0}, which is linked to the growth of the sequence (λn). Indeed, Costakis, Tsirivas and Vlachou have shown in [10] λn > 2, thenTa∈C\{0} HC (τna) is empty. This shows that, in some sense, that, if liminfn→+∞ the result of Tsirivas quoted above is optimal, but leaves open the case λn = q n with q ∈ (1,2]. Using our covering argument, we are able to extend this result to the remaining case and to any operator group! λn+1 Theorem 1.2. Let (Ta)a∈Rd be a strongly continuous operator group on X and let (λn) be an λn > 1. ThenTa∈Rd \{0} HC (Tλn a) increasing sequence of positive real numbers such that liminfn is empty. λn+1 When we add supplementary conditions, the method of Costakis and Sambarino is unefficient to solve certain problems. This is the case if we consider frequent hypercyclicity, a notion in- troduced in [5]. Recall that for a set A ⊂ N, its lower density is defined by dens(A) = liminf N→+∞ N #{n ≤ N ; n ∈ A} , where #B stands for the cardinal number of B. Given a sequence of operators (Tn) of X , we say that x ∈ X is a frequently hypercyclic vector for (Tn) if for any U ⊂ X open and nonempty, the set {n; Tn x ∈ U } has positive lower density and we denote by F HC (Tn) the set of frequently hy- percyclic vectors for (Tn). As before, for a single operator T , F HC (T ) will stand for F HC¡(T n)¢. It was shown in [5] that, for any a ∈ C\{0}, τa acting on H(C) is frequently hypercyclic. Moreover, the algebraic method can be carried on frequent hypercyclicity. In particular, if (Tt )t>0 is a strongly continuous semigroup on X , then for any t > 0, F HC (Tt ) = F HC (T1). This implies in particular that we can find a common frequently hypercyclic vector for all operators (τa)a∈R\{0}, a result first obtained in [5]. The methods introduced in this paper allow us to go further and to prove the following natural result. Theorem 1.3. The setTa∈C∗ F HC (τa) is nonempty. As before, a more general version of Theorem 1.3 will be proved in Section 5. In particular, this version can be applied to all the examples introduced in this paper and to high-dimensional families, which is rather surprizing! In the last two sections of this paper, we give related results. First, we study the existence of a common hypercyclic vector for all multiples of operators living in a high-dimensional operator group, leading to a multidimensional generalization of the above result of Shkarin. Second, we emphasize on the algebraic method, showing that it is also helpful to obtain multidimensional results. 4 FRÉDÉRIC BAYART 2. THE UNIFORM MIXING PROPERTY 2.1. The uniform mixing property and a covering argument. In this section, we introduce our main condition for an operator group to admit a common hypercyclic vector. This condition is an enhancement of the mixing property. Definition 2.1. A group (Ta)a∈Rd acting on X has the uniform mixing property if, for any U ,V nonempty open subsets of X , there exists C > 0 such that, for any p ≥ 1, for any a1,..., ap ∈ Rd with kaik ≥ C and kai − a jk ≥ C for any i , j ∈ {1,... , p} with i 6= j , there exists f ∈ U such that Ta j f ∈ V for any j = 1,..., p. This property is weaker than the Runge property introduced by Shkarin in [17] (see also the forthcoming Section 3.2). Moreover, as it is observed in [7] for a similar property, the Runge property cannot be satisfied for an operator group defined on a Banach space. We shall see later that there exist operator groups defined on Banach spaces and satisfying the uniform mixing property. Our first result says that the existence of a common hypercyclic vector for an operator group can be deduced from the construction of a suitable covering of Rd . We shall see later (Theorem 4.3) that the converse is true. Theorem 2.2. Let (Ta)a∈Rd be a strongly continous group on X with the uniform mixing prop- erty. Let K be a compact subset of Rd \{0} and let (λn) be an increasing sequence of positive real numbers. Assume that, for all ε > 0 and all C > 0, for all N ∈ N, we can find M ≥ N and a finite number (xn,k)N≤n≤M, 1≤k≤pn of elements of K satisfying (A) For any n, m, k, j with (n, k)6= (m, j ), then kλn xn,k − λm xm, jk ≥ C . (B) For any x ∈ K , there exist n, m ∈ {N ,..., M} and k ∈ {1,... , pn} such that ThenTa∈K HC (Tλn a) is a residual subset of X . Proof. Let U ,V be nonempty open subsets of X . It is sufficient to show that kλm x − λn xn,kk < ε. U ∩© f ∈ X ; ∀a ∈ K , ∃n ∈ N, Tλn a f ∈ Vª is nonempty (see for instance [6, Proposition 7.4]). Let g ∈ V and let V ′ be a neighbourhood of zero such that g + V ′ + V ′ ⊂ V . Since (Ta) is strongly continuous, the uniform boundedness principle says that the map (a, f ) 7→ Ta f is continuous. In particular, there exist ε1 > 0 and W a neighbourhood of zero such that Ta(W ) ⊂ V ′ for any a with kak < ε1. Moreover, there exists ε2 > 0 such that Ta(g )− g ∈ V ′ provided kak < ε2. We set ε = min(ε1, ε2). We then set V0 = g + W and we apply the uniform mixing property with U and V0. We get a positive real number C and we choose N such that λNkak > C for any a ∈ K . For these values of ε,C , N , we get points (xn,k) satisfying (A) and (B). Applying the uniform mixing property with the sequence (λn xn,k), we know that there exists f ∈ U such that Tλn xn,k f ∈ V0 for any admissible choice of (n, k). Pick now x ∈ K . We may find m, n, k such that kλm x − λn xn,kk < ε. COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 5 Now we write Tλm x f − g = Tλm x f − Tλm x−λn xn,k g + Tλm x−λn xn,k g − g {z = Tλm x−λn xn,k (Tλn,k xn,k f − g } } + Tλm x−λn xn,k g − g } {z ∈V ′ ∈W {z ∈V ′ ) . We shall apply Theorem 2.2 under the following form. Hence, Tλn x f belongs to V , which yields thatTa∈K HC (Tλn a) is a residual subset of X . Corollary 2.3. Let (Ta)a∈Rd be a strongly continous group on X with the uniform mixing prop- erty. Let K be a compact subset of Rd \{0} and let (λn) be an increasing sequence of positive real numbers. Assume that, for all ε > 0 and all C > 0, there exists γ > 0 such that, for all N ∈ N, for all compact subsets L of K with diam(L) < γ, we can find M ≥ N and a finite number (xn,k)N≤n≤M, 1≤k≤pn of elements of L satisfying (cid:3) (A) For any n, m, k, j with (n, k)6= (m, j ), then (B) For any x ∈ L, there exist n, m ∈ {N ,..., M} and k ∈ {1,... , pn} such that kλn xn,k − λm xm, jk ≥ C . kλm x − λn xn,kk < ε. ThenTa∈K HC (Tλn a) is a residual subset of X . Proof. Let ε,C > 0 and N ∈ N. We get some γ > 0 and we write K as a finite union of compact subsets L1,... , Lq with diam(Li ) < γ. We apply iteratively the construction for each Li with Ni defined as follows: • N1 = N ; • if at Step i , we have constructed elements (xn,k) until n ≤ Mi , we take Ni+1 > Mi any positive integer n such that λnkak > λMi kbk+ C for any a, b ∈ K . This ensures that the separation property (A) keeps beings true on the whole set of points (xn,k). Hence, the whole sequence (xn,k), n = Ni ,..., Mi , i = 1,..., q, satisfies the assumptions (A) and (B) of Theorem 2.2 that we may apply. (cid:3) 2.2. An efficient way to split sequences of positive real numbers. We need now to introduce a condition on positive sequences of real numbers (λn) in order to ensure common hypercyclic- ity of the family (Tλn a). For the translation group on H(C), N. Tsirivas has introduced in [18] a sufficient condition: it suffices that, for any M > 0, there exists a subsequence (µn) of (λn) such that µn+1 − µn ≥ M for any n andPn≥1 µn = +∞. This last condition was not very surprizing, because it was the main property on the whole sequence of integers (n) which was used in the Costakis-Sambarino criterion. For our purpose, we will weaken this condition in order to allow sequences with faster growth. 1 Definition 2.4. We say that an increasing sequence (λn) has property (SG) if, for any B > 0, there exist ρ > 1 and a subsequence (µn) of (λn) such that • µn+1 ≥ ρµn; • for any n0 ∈ N,P+∞n=n0+1 1 µn > B µn0 . 6 FRÉDÉRIC BAYART We first show that many classical sequences have property (SG). Proposition 2.5. Let (λn) be an increasing sequence of positive real numbers tending to +∞ such that λn+1/λn → 1. Then (λn) has property (SG). Proof. Let B > 0. There exists ρ0 > 1 such thatP+∞n=1 ρ−n 0 > B. We set ρ = pρ0. Let p ≥ 1 be such λn ≤ ρ provided n ≥ p. We then set ψ(0) = p and we define ψ(n) for n ≥ 1 by induction that λn+1 using the following formula: Then λψ(n+1) ≥ ρλψ(n) and λψ(n+1) ≤ ρλψ(n+1)−1 ≤ ρ2λψ(n) = ρ0λψ(n). Setting µn = λψ(n), we immediately get the conclusion. (cid:3) ψ(n + 1) = inf©m ≥ ψ(n); λm ≥ ρλψ(n)ª. 1 µn > B µn0 contained in the following technical lemma. To be able to produce coverings in arbitrary large dimensions, we will need to be able to iterate arbitrary many times. The precise statement that we need is the propertyP+∞n=n0+1 Lemma 2.6. Let d ≥ 1 and A > 0. There exists B := B(d, A) > 0 such that, if (µn) is an increasing sequence of positive real numbers such that, for any n0 ∈ N,P+∞n=n0+1 , if s ≥ 1 is such that , then we can find s1 ∈ N, subsets Er of Nr−1 for r = 2,... , d + 1, maps sr : Er → N for Ps n=1 r = 2,... , d and a one-to-one map φ : Ed+1 → {0,... , s} such that µn ≥ B µn0 1 1 µn ≥ B µ0 • for any r = 2,..., d + 1, Er =©(k1,... , kr−1); k1 < s1, k2 < s2(k1),..., kr−1 ≤ sr−1(k1,... , kr−2)ª. • for any r = 1,..., d , for any (k1,..., kr−1) ∈ Er , µφ(k1,...,kr−1, j ,0,...,0) ≥ sr (k1,...,kr−1) 1 Xj=1 A µφ(k1,...,kr−1,0,...,0) . • If (k1,..., kd ) > (k′1,... , k′d ) in the lexicographical order, then φ(k1,... , kd ) > φ(k′1,..., k′d ). Proof. We proceed by induction on d. The result is clear for d = 1, setting simply B = A, s1 = s and φ(k) = k. Assume now that the result until step d −1 is known and let us prove it for step d. Let A > 0 and let us set Let (µn) be an increasing sequence of positive integers and s ≥ 1 be such that B := B(d, A) = (A + 2)B(d − 1, A)+ 3. We set φ(0,... ,0) = 0 and we define by induction φ( j + 1,0,... ,0) as the smallest integer N such that 1 µn ≥ B µ0 and s Xn=1 +∞Xn=n0+1 1 µn ≥ B µn0 for any n0 ∈ N. N Xk=φ( j ,0,...,0)+1 1 µk ≥ B(d − 1, A) µφ( j ,0,...,0) . B(d − 1, A) µφ( j ,0,...,0) ≤ φ( j+1,0,...,0) Xk=φ( j ,0,...,0)+1 1 µk ≤ B(d − 1, A)+ 1 µφ( j ,0,...,0) . In particular, (1) COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 7 We stop when j + 1 = s1 where s1 is the smallest integer t such that (A + 1)B(d − 1, A)+ 2 φ(t ,0,...,0) . Xk=φ(0,...,0)+1 1 µk ≥ µφ(0,...,0) In particular, φ(s1,0,...,0) Xk=φ(0,...,0)+1 1 µk ≤ (A + 1)B(d − 1, A)+ 2 µφ(0,...,0) (A + 2)B(d − 1, A)+ 3 µφ(0,...,0) ≤ + . φ(s1,0,...,0) X k=φ(s1−1,0,...,0)+1 1 µk We claim that φ(s1,0,...,0) ≤ s. Indeed, φ(s1,0,...,0) Xk=φ(0,...,0)+1 1 µk ≤ (A + 2)B(d − 1, A)+ 3 µφ(0,...,0) B(d, A) µφ(0,...,0) ≤ ≤ 1 µk . s Xk=1 We now apply the induction hypothesis for j = 0,... , s1 − 1 by using the left part of (1). We get maps φ j , sr, j . We just set φ( j , k1,..., kd ) = φ j (k2,..., kd ), sr ( j , k2,..., kr−1) = sr−1, j (k2,..., kr−1). The only thing which remains to be done is to show that This can be done by observing that 1 µφ( j ,0,...,0) ≥ A µφ(0,...,0) . s1Xj=1 1 µφ( j ,0,...,0) ≥ s1Xj=1 ≥ ≥ ≥ ≥ 1 µk s1−1 φ( j+1,0,...,0) 1 1 1 1 1 µk φ(s1,0,...,0) B(d − 1, A)+ 1 B(d − 1, A)+ 1 Xk=φ( j ,0,...,0)+1 Xj=1 Xk=φ(1,0,...,0)+1 B(d − 1, A)+ 1Ãφ(s1,0,...,0) 1 Xk=1 µk − B(d − 1, A)+ 1µ (A + 1)B(d − 1, A)+ 2− B(d − 1, A)− 1 1 µk! Xk=1 µφ(0,...,0) φ(1,0,...,0) A . µφ(0,...,0) ¶ (cid:3) Lemma 2.6 can be applied for sequences (λn) having property (SG). Corollary 2.7. Let (λn) be a sequence having property (SG). Then for all d ≥ 1 and all A > 0, there exist ρ > 1 and a subsequence (µn) of (λn) such that µn+1 ≥ ρµn for any n ≥ 1 and, for all P > 0, we can find s1 ∈ N, subsets Er of Nr−1 for r = 2,... , d + 1, maps sr : Er → N for r = 2,... , d and a one-to-one map φ : Ed+1 → N such that • for any r = 2,..., d + 1, Er =©(k1,... , kr−1); k1 < s1, k2 < s2(k1),..., kr−1 ≤ sr−1(k1,... , kr−2)ª. 8 FRÉDÉRIC BAYART • for any r = 1,..., d , for any (k1,..., kr−1) ∈ Er , sr (k1,...,kr−1) Xj=1 1 µφ(k1,...,kr−1, j ,0,...,0) ≥ A µφ(k1,...,kr−1,0,...,0) . • φ(0,... ,0) ≥ P . • If (k1,..., kd ) > (k′1,... , k′d ) in the lexicographical order, then φ(k1,... , kd ) > φ(k′1,..., k′d ). 2.3. Common hypercyclic vectors. We are now ready to state and to prove the main result of this section. Theorem 2.8. Let (Ta)a∈Rd be a strongly continuous operator group on X which is uniformly mixing and let (λn) be an increasing sequence of positive real numbers having property (SG). ThenTa∈Rd \{0} HC (Tλn a) is a residual subset of X . Proof. We first show thatTa∈K HC (Tλn a) is a residual subset of X when K is a compact subset of (0,+∞)d . We shall prove that the conditions of Corollary 2.3 are satisfied. Hence, let ε > 0 and C > 0. We then apply Corollary 2.7 to A = C /ε to get some ρ > 1 and some subsequence (µn) of (λn) with µn+1 ≥ ρµn. Since K ⊂ (0,+∞)d , we may find γ > 0 such that, given any a = (a1,..., ad ) ∈ K , ρai − ai − γ > 0 for all i = 1,..., d. Let now L be a compact subset of K with diameter less than γ and let N ∈ N. To simplify the notations, we shall assume that L = Qd i=1[bi , bi + γ] with bi > 0. We apply the properties of the sequence (µn) given by Corollary 2.7 with P ≥ N such that We get maps s1,..., sd and φ. We may now define our covering of L. We set µP inf i=1,...,d (ρbi − bi − γ) ≥ C . n0 = min (k1,...,kd ) φ(k1,..., kd ), m0 = max (k1,...,kd ) φ(k1,... , kd ) and let n ∈ {n0,... , m0}. Then either n is not a φ(k1,..., kd ) and we do nothing. Or n is equal to φ(k1,... , kd ) for a (necessarily) unique (k1,..., kd ). We then define the set {xn,k}1≤k≤pn as α1C µφ(0,...,0) + ε µφ(1,0,...,0) +···+ ε , µφ(k1,0,...,0) ε ε µφ(k1,1,0,...,0) +···+ µφ(k1,k2,...,0) , L ∩½µ b1 + b2 + bd + α2C µφ(k1,0,...,0) + ... µφ(k1,...,kd−1,0) + αdC ε µφ(k1,...,kd−1,1) +···+ ε µφ(k1,...,kd )¶; α1,..., αd ≥ 0¾. We also set ωn = µφ(k1,...,kd ) and we claim that (A) and (B) of Corollary 2.3 are satisfied with ωn instead of λn (but ωn is of course some λm and it would be sufficient to renumber everything). Indeed, let (n, k) 6= (m, j ). We distinguish two cases: • either n 6= m, for instance n < m. In that case, (2) kωm xm, j − ωn xn,kk ≥ ωmb1 − ωn(b1 + γ) ≥ µP (ρb1 − b1 − γ) ≥ C . COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 9 • or n = m = φ(k1,..., kd ). We write xn,k and xn, j as before, with respectively the se- quences (α1,... , αd ) and (β1,..., βd ). Since k 6= j , at least one βi differs from αi . Now, looking at this coordinate, we get kωn xn,k − ωn xn, jk ≥ µφ(k1,...,kd ) µφ(k1,...,ki−1,0,...,0) C ≥ C . Let us now prove (B): let x ∈ L. There exists α1 > 0 such that b1 + α1C µφ(0,...,0) ≤ x1 ≤ b1 + (α1 + 1)C µφ(0,...,0) . Now, by construction of φ, using Corollary 2.7 (recall that A = C /ε), there exists k1 < s1 such that b1 + ε α1C µφ(0,...,0) + α1C µφ(0,...,0) + b1 + µφ(1,0,...,0) +···+ ε µφ(1,0,...,0) +···+ ε µφ(k1,0,...,0) ≤ x1 ≤ . ε µφ(k1+1,0,...,0) This k1 being fixed, there exists α2 ≥ 0 such that b2 + α2C µφ(k1,0,...,0) ≤ x2 ≤ b2 + (α2 + 1)C µφ(k1,0,...,0) . Iterating this construction, we find α1,..., αd ≥ 0 and k1,... , kd such that, for all i = 1,... , d, bi + αi C µφ(k1,...,ki−1,0,...,0) + bi + µφ(k1,...,ki−1,0,...,0) + αi C ε ε µφ(k1,...,ki−1,1,0,...,0) +···+ µφ(k1,...,ki−1,ki ,0,...,0) ≤ xi ≤ µφ(k1,...,ki−1,1,0,...,0) +···+ µφ(k1,...,ki−1,ki+1,0,...,0) . ε ε Let n = φ(k1,... , kd ) and let xn,k corresponding to this value of α1,..., αd . Then, for any i = 1,... , d, kωn x − ωn xn,kk ≤ µφ(k1,...,kd ) × sup i=1,...,d µφ(k1,...,ki+1,0,...,0) ≤ ε. ε Hence, by Corollary 2.3,Ta∈K HC (Tλn a) is a residual subset of X . This works for any compact set K ⊂ (0,+∞)d or, more generally, for any compact set K contained in some open orthant of Rd \{0}. Now, assume that K = {0}d−e × K ′ where K ′ is a compact set of (0,+∞)e and define, for b ∈ Re, Sb = T(0,b). Then (Sb)b∈Re has the uniform mixing property and thus Then, writing Rd \{0} as the countable unionSe=1,...,dSn≥1Sε∈{−1,1}dSS∈P e(d)Qd easily get the conclusion. i=1 Ki ,n,εi (S), we (cid:3) is a residual subset of X . To conclude, let P e(d) be the subsets of {1,... , d} with cardinal number equal to e. For n ≥ 1, ε = ±1 and S ∈ P e(d), define \a∈K HC (Tλn a) = \b∈K ′ HC (Sλnb) Ki ,n,ε(S) =( {0} n , εn¤ £ ε if i ∉ S if i ∈ S. 10 FRÉDÉRIC BAYART Remark 2.9. In property (SG), the condition µn+1 ≥ ρµn is rather unpleasant. It is necessary to separate sufficiently µn xn,k and µm xm,k when n 6= m. If we just looked atTkak=1 HC (Tλn a), we could replace it by the more pleasant condition µn+1 − µn ≥ B: see the section devoted to frequent hypercyclicity and in particular compare (2) above and (5). 3. EXAMPLES 3.1. A sufficient condition. We now give examples of operator groups having the uniform mix- ing property. We first begin by a criterion which can be seen as a strong form of the hypercyclic- ity criterion. Proposition 3.1. Let (Ta)a∈Rd be an operator group on X and let k·k be an F -norm on X . Assume that there exists a dense set D ⊂ X such that, for any f ∈ D and any ε > 0, there exists C > 0 such that, for any N ≥ 1, for any a1,... , aN ∈ Rd with kai − a jk ≥ C if i 6= j and kaik ≥ C , then i=1 Tai f°° < ε. Then (Ta)a∈Rd has the uniform mixing property. °°PN Observe that the hypercyclicity criterion shares the same assumptions restricted to N = 1. For the definition of an F -norm, we refer to [11]. Proof. Let U ,V be nonempty open subsets of X . There exist g , h ∈ D, ε > 0 such that B(g , ε) ⊂ U and B(h, ε) ⊂ V . Applying the assumptions with (g , ε/2) and with (h, ε/2), we get two positive Cg and Ch. We set C = max(Cg ,Ch). Let us consider N ≥ 1 and a1,... , aN ∈ Rd with kai −a jk ≥ C and ka jk ≥ C . We define Setting bi = a j − ai , then kbik ≥ C and kbi − blk = kai − alk ≥ C provided i 6= l . Hence, kTa j f − hk < ε and Ta j f ∈ V . (cid:3) Example 3.2. Let w : Rd → R be a positive bounded and continuous function such that x 7→ w(x+a) is bounded for each a ∈ Rd . For a ∈ Rd and p ≥ 1, let τa be the translation operator w(x) defined on X = Lp (Rd , w(x)d x) by τa f (x) = f (x + a). Assume moreover thatRRd w(x)d x < +∞. ThenTa∈Rd \{0} HC (τa) is a residual subset of X . Proof. Let D ⊂ X be the dense set of compactly supported continuous functions. Let f ∈ D, ε > 0 and let A > 0 be such that the support of f is contained in B(0, A). There exists some C > 0 such that . k f k p ∞ Let N ≥ 1 and a1,... , aN ∈ Rd with kai − a jk ≥ C and kaik ≥ C provided i 6= j . Then • for any a, b ∈ Rd with ka − bk ≥ C , Ta f and Tb f have disjoint support. • Rkxk≥C−A w(x)d x ≤ εp °°°°° Xi=1 ≤Zkxk≥C−A k f k so that (τa)a∈Rd has the uniform mixing property. Tai f°°°°° p ∞w(x)d x ≤ εp N p (cid:3) Clearly, k f − gk < ε/2 < ε. Moreover, let j ∈ {1,... , N }. Then T−ai h. N f = g − Xi=1 Ta j f = Ta j g −Xi6= j Ta j−ai h + h. for any a ∈ Rd with kak ≥ 1. Then (Ta)a∈Rd has the uniform mixing property. Proof. We shall see that the assumptions of Proposition 3.1 are satisfied. The key point is to observe that, if (ai ) is any sequence in Rd such that kai − a jk ≥ 1 for any i 6= j , there exists κd > 0 such that, for any k ≥ 1, Let now f ∈ D, ε > 0 and C ≥ 1. Let a1,..., aN ∈ Rd with kai − a jk ≥ 1 if i 6= j and kaik ≥ C for any i . Then #©i ; 2k ≤ kaik < 2k+1ª ≤ κd 2kd . Tai f°°°°° N Xi=1 °°°°° Xi ; 2k≤kaik<2k+1°°Tai f°° Xi ; 2k≤kaik<2k+1 1 2k p 1 ≤ Xk; 2k≥C ≤ A Xk; 2k≥C ≤ Aκd Xk; 2k≥C COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 11 This improves (even when d = 1) Example 7.20 of [6]. We can also deduce from Proposition 3.1 a useful corollary to get common hypercyclicity. Corollary 3.3. Let (Ta)a∈Rd be an operator group acting on X and let k·k be an F -norm on X . Assume that there exist a dense set D ⊂ X and p > d such that, for any f ∈ D, there exists A > 0 such that kTa f k ≤ A kakp and this is less than ε provided C is large enough. 2k(p−d) (cid:3) 3.2. The Runge property. Our second example deals with groups having the Runge property introduced in [17]. The name "Runge property" is reminiscent for the method of proof of the hypercyclicity of τa on H(C), a 6= 0. Definition 3.4. Let (Ta)a∈Rd be an operator group on X . We say that (Ta)a∈Rd has the Runge property if for any continuous seminorm k·k on X , there exists C > 0 such that, for any N ≥ 1, for any b1,..., bN ∈ Rd with kbi − b jk ≥ C for i 6= j , for any g1,..., g N ∈ X , for any ε > 0, there exists f ∈ X such that kTbi f − gik < ε for each i = 1,..., N . Proposition 3.5. An operator group having the Runge property has the uniform mixing property. Proof. Let U ,V be nonempty open subsets of X , let g , h ∈ X , let k·k be a continuous seminorm on X such that { f ∈ X ; k f − hk < 1} ⊂ U and { f ∈ X ; k f − gk < 1} ⊂ V . For this seminorm k·k and for ε = 1, applying the definition of the Runge property, we find some C > 0. Let N ≥ 1 and a1,... , aN ∈ Rd with kai − a jk ≥ C and kaik ≥ C for any i 6= j . Define b0 = 0 and bi = ai for i = 1,..., N . Then there exists f ∈ X such that kTb0 f − hk < 1 and kTbi f − gk < 1 for i = 1,... , N . In particular, f ∈ U and Tai f ∈ V for any i = 1,..., N , so that (Ta)a∈Rd has the uniform mixing property. (cid:3) In [17], it is shown that the translation group (τa)a∈C acting on H(C) has the Runge property. In several complex variables, the situation is less clear due to the lack of Runge theorem. However, this remains true if we restrict ourselves to translations by vectors in Rd . 12 FRÉDÉRIC BAYART Example 3.6. The operator group (τa)a∈Rd acting on H(Cd ) has the Runge property. Proof. By a result of Khudaiberganov [13], the union of a finite union of disjoint balls in Cd with centers in Rd is polynomially convex. By the Oka-Weil theorem (see for instance [15]), any function holomorphic in a neighbourhood of a compact polynomially convex set K ⊂ Cd can be uniformly approximated by holomorphic polynomials. Hence, let k·k be a continuous seminorm on H(Cd ). There exists A > 0 such that, for any f ∈ H(Cd ), k f k ≤ k f kA := A sup kzk≤A f (z). We set C = 4A and we consider a finite set of points b1,... , bN in Rd and a finite set of holo- morphic functions g1,..., g N such that kbi − b jk ≥ C provided i 6= j . Let Bi be the closed ball Bi = {z ∈ Cd ; kz − bik ≤ A}. Since these balls are pairwise disjoint, K = B1 ∪···∪ BN is polyno- mially convex. Thus, there is a polynomial f such that supz∈Bi f (z)− gi (z + bi ) < ε/A for any i = 1,..., N . We obtain immediately that kTbi f − gik < ε for any i = 1,..., N . (cid:3) It is also very easy to show that the translation group (τa)a∈Rd acting on the Fréchet space Cc(Rd ) of continous functions f : Rd → R with the topology of uniform convergence on com- pact sets satisfies the Runge property. 3.3. Heisenberg translations. We now give an example of a group which is not a translation group. Let d ≥ 2 and let X = H 2(Bd ) be the Hardy space on the (euclidean) unit ball of Cd denoted by Bd . Let φ be an automorphism of Bd . Then the composition operator Cφ defined by Cφ( f ) = f ◦ φ is a bounded operator on H 2(Bd ). When φ has no fixed points in Bd , it has be shown in [12] that Cφ is hypercyclic. A class of automorphisms plays a crucial role in the study of composition operators and of linear fractional maps of the ball, the class of Heisenberg translations (see [2] or [3]). To under- stand these automorphisms, it is better to move on the Siegel upper half-space The Siegel half-space is biholomorphic to Bd via the Cayley map ω defined by Hd =©Z = (z, w) ∈ C× Cd−1; ℑm(z) > w2ª. z + i¶. 1− z¶, ω−1(z, w) =µ z − i ω(z, w) =µi 1+ z 1− z , z + i 2w , i w The Cayley transform extends to a homeomorphism of Bd onto Hd ∪ ∂Hd ∪ {∞}, the one-point compactification of Hd . The image of H 2(Bd ) by the Cayley transform is denoted by H 2(Hd ) : As one easily sees by computing the jacobian of ω, H2(Hd ) is endowed with the norm H 2(Hd ) =©F : Hd → C holomorphic; F ◦ ω−1 ∈ H 2(Bd )ª. kFk2 H 2 = κ2Z∂Hd F (z, w)2 z + i2d d σ∂Hd , where κ is a constant that we will not try to compute. For γ ∈ Cd−1\{0}, the Heisenberg translation with symbol γ is defined by Hγ(z, w) = (z + 2i〈w, γ〉+ iγ2, w+ γ). Hγ is an automorphism of Hd fixing ∞ only and, as already mentioned, C Hγ is hypercyclic on H 2(Hd ). It is also easy to check that (C Hγ )γ∈Cd−1 is a strongly continous group on H 2(Hd ). COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 13 Before to prove this theorem, we need a couple of lemmas. Theorem 3.7. Tγ∈Cd−1\{0} HC (C Hγ) is a residual subset of H 2(Hd ). Lemma 3.8. Let p ≥ 1. The functions (z1−1)p g (z), where g runs over the ball algebra A(Bd ), are dense in H 2(Bd ). Proof. Let h ∈ A(Bd ) and define fk(z) = h(z)Ã1−µ 1+ z1 2 = gk(z)(1− z1)p ¶k!p with gk ∈ A(Bd ). Lebesgue's theorem implies that fk tends to h in H 2(Bd ). We conclude by density of A(Bd ) in H 2(Bd ). (cid:3) Lemma 3.9. Let p ≥ 1 and let Dp =½F ∈ A(Hd ); there exists C > 0 s.t. ∀(z, w) ∈ Hd , F (z, w) ≤ Then Dp is dense in H 2(Hd ). C z + ip¾. Proof. This follows immediately from the previous lemma and the definition of the Cayley map, since ¯¯¯¯ z − i z + i − 1¯¯¯¯ = 2 . z + i (cid:3) Lemma 3.10. There exists a dense set D ⊂ H 2(Hd ) such that, for any F ∈ D, there exists A > 0 such that, for any γ ∈ Cd−1\{0}, γ ≥ 1, (3) A . kC Hγ FkH 2 ≤ γ2d− 3 2 Proof. We shall see that, provided p is large enough, Dp satisfies the conclusions of the lemma. Precisely, let ε ∈ (0,1) be such that (2d − 1)− ε(d − 1) ≥ 2d − 3 2 . We then adjust p so that 2εp ≥ 2d− 3 2 and we pick F ∈ Dp , γ ∈ Cd−1 with γ ≥ 1. To simplify the notations, we shall write during this proof that u . v provided there exists C > 0 such that u ≤ C v where C does not depend on γ (it may depend on F , p or ε). Then kC Hγ Fk2 H 2 . Zw∈Cd−1Zx∈R . Zw∈Cd−1Zx∈R . Zw∈Cd−1Zx∈R F (x + iw2 + 2i〈w, γ〉+ iγ2, w+ γ)2 x + iw2 + i2d 1 d xd w (1+x+w2)2d (x − 2ℑm〈w, γ〉+γ+ w2 + 1)p (1+x+w2)2d (1+γ+ w2)2p d xd w. 1 d xd w We split the integral following the value of γ+ w. Assume first that γ+ w ≤ γε. Then, writing 1 1 1 (1+x+w2)2d × (1+γ+ w2)2p . (1+x+γ2)2d 14 FRÉDÉRIC BAYART and observing that the ball B(−γ, γε) in Cd−1 has volume comparable to γ2ε(d−1), we get . d xd w . γ2ε(d−1) 1 1 Zw+γ≤γεZx∈R (1+x+w2)2d (1+γ+ w2)2p (1+γ2)2d−1 . γ2¡2d− 3 2¢ When γ+ w ≥ γε, we now write (1+x+w2)2d × 1 1 1 (1+γ+ w2)2p . (1+x+w2)2d × 1 (1+γ2ε)2p . We integrate this over Cd−1 × R. Taking into account the inegalities satisfied by ε and p, we get the result of the lemma. (cid:3) Proof of Theorem 3.7. Theorem 3.7 follows immediately from Corollary 3.3 and Lemma 3.10. (cid:3) 4. OBSTRUCTIONS TO COMMON HYPERCYCLICITY 4.1. A converse to the covering property. In this section, we now show that, even if a group (Ta)a∈Rd has the uniform mixing property, we cannot expect thatTa∈Rd \{0} HC (Tλn a) is non- empty provided (λn) grows too fast. We need a condition on the group saying that it is not too quickly mixing. Definition 4.1. We say that an operator group (Ta)a∈Rd is locally separating if, for any A > δ > 0, there exists a nonempty open set U such that TaU ∩U = ∅ provided δ ≤ kak ≤ A. It turns out, that under this condition, the existence of a common hypercyclic vector implies a covering property similar to that of Section 2.1. Lemma 4.2. Let K be a compact subset of Rd \{0}, let (λn)n be an increasing sequence of positive real numbers going to infinity and let (Ta)a∈Rd be a strongly continuous operator group on X which is locally separating. Assume thatTa∈K HC (Tλn a) is nonempty. Then, for all A > δ > 0, for all N ∈ N, we can find M ≥ N and a finite number (yn,k)N≤n≤M,1≤k≤qn of elements of K satisfying (A) for any n, m, k, j , either kλn yn,k − λm ym, jk < δ or kλn yn,k − λm ym, jk > A. (B) K ⊂SM Proof. Let f ∈Ta∈K HC (Tλn a) and let, for n ≥ N , k=1 B³yn,k, δ λn´ . n=NSqn where the open set U is given by the local separation property. Then each Vn is an open subset of K and K is contained inSn≥N Vn. By the compactness of K , there exists M ≥ N such that K ⊂ SM n=N Vn. Let us now consider x ∈ K and let n(x) be the smallest integer n ≥ N such that x ∈ Vn. λn(x)´. By the compactness of K again, we can extract a finite Then K is contained inSx∈K B³x, δ λn´. sequence (yn,k)N≤n≤M such that each yn,k belongs to Vn and K is contained inSn,k B³yn,k, δ Moreover, since Tλn yn,k f ∈ U and Tλm ym, j f ∈ U , it is plain that Tλn yn,k−λm ym, j U ∩U 6= ∅, which implies (A) by the definition of U . (cid:3) When (Ta)a∈Rd has the uniform mixing property, we can close the circle and show that the converse of Theorem 2.2 is true! Vn =©a ∈ K ; Tλn a f ∈ Uª COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 15 Theorem 4.3. Let (Ta)a∈Rd be a strongly continous group on X with the uniform mixing prop- erty. Let K be a compact subset of Rd \{0} and let (λn) be an increasing sequence of positive real numbers. Then the following assertions are equivalent : (1) Ta∈K HC (Tλn a) is a residual subset of X . (2) For all ε > 0 and all C > 0, for all N ∈ N, we can find M ≥ N and a finite number (xn,k)N≤n≤M, 1≤k≤pn of elements of K satisfying (A) For any n, m, k, j with (n, k)6= (m, j ), then (B) For any x ∈ K , there exist n, m ∈ {N ,..., M} and k ∈ {1,... , pn} such that kλn xn,k − λm xm, jk ≥ C . kλm x − λn xn,kk < ε. Proof. That (2) implies (1) has already been proved in Theorem 2.2. Assume now that (1) is satisfied. We first observe that the uniform mixing property implies the local separation prop- erty. Indeed, let A > δ > 0. By the uniform mixing property and Theorem 1.1,Ta∈Rd \{0} HC (Ta) is nonempty. Pick f ∈Ta∈Rd\{0} HC (Ta). Then Ta f 6= f for any a ∈ Rd \{0}; otherwise, the set {Tt f ; t > 0} would be compact, hence nondense. By continuity of the map f 7→ Ta f and by compactness of the corona {a ∈ Rd ; δ ≤ kak ≤ A}, there exists a neighbourhood U of f such that TaU ∩U = ∅ for any a ∈ Rd with δ ≤ kak ≤ A. Hence, we may apply Lemma 4.2 and we are lead to show that the conditions of this lemma imply (2). Let C , ε > 0 and N ∈ N. We apply Lemma 4.2 with A = C and δ = ε/2 to get M and (yn,k). We order N2 using the lexicographical order. We construct by induction sequences (y′j ) and (λ′j ) by setting y′1 = yN,1, λ′1 = λN and, provided y′1,..., y′j , λ′1,... , λ′j have been constructed with y′j = yn,k and λ′j = λn, we set y′j+1 = inf©(m, ℓ) ≥ (n, k); ∀p ≤ j , kλm ym,ℓ − λ′p y′pk ≥ Aª and λ′j+1 is the corresponding λm. If (m, ℓ) does not exist, then we stop the construction. We then rename (y′j ) as (xn,k): for a given n in {N ,..., M}, we set {xn,k } = {y′j ; λ′j = λn}. The con- struction of the sequence (xn,k) immediately implies that (2)(A) is satisfied. Moreover, let x ∈ K . We know that there exists (n, k) with kλn x − λn yn,kk < δ. By construction, there exists (m, ℓ) with m ≤ n such that kλn yn,k −λm xm,ℓk < A. But since xm,ℓ is itself an element of the sequence (yp,u), we have kλn yn,k − λm xm,ℓk < δ. This implies kλn x − λm xm,ℓk < 2δ = ε, so that (2)(B) is satisfied. (cid:3) 4.2. Obstructions. We may now state the main theorem of this section, which is the desired ex- tension of the result of Costakis, Tsirivas and Vlachou. We say that an interval in Rd is nontrivial if it contains at least two points. Theorem 4.4. Let I be a nontrival compact interval in Rd \{0} and let (Ta)a∈Rd be a strongly con- tinuous operator group on X which is locally separating. Let also (λn) be an increasing sequence of positive real numbers such that liminfn λn+1/λn > 1. ThenTa∈I HC (Tλn a) = ∅. Lemma 4.5. Let I , J1,... , Jp be intervals of R. Then I \(J1∪ J2∪···∪ Jp ) is the reunion of s disjoint intervals E1,..., Es with s ≤ p + 1 and E1+···+Es ≥ I−J1−···−Jp. The proof of this theorem will depend heavily on the following easy lemma on intervals of R. 16 FRÉDÉRIC BAYART Proof. We proceed by induction on p, the result being trivial for p = 0. Let E1,... , Es, s ≤ p + 1 be disjoint intervals such that I \(J1 ∪ J2 ∪···∪ Jp ) = E1 ∪···∪ Es. Renumbering the intervals Ei if necessary, we assume that max Ei ≤ min Ei+1. We divide the proof into several cases: • if there exists some j such that Jp+1 ⊂ E j , then set E j \Jp+1 =: E′j ∪ E′′j where E′j and E′′j are intervals. In that case, I \(J1 ∪ J2 ∪···∪ Jp ) = E1 ∪···∪ E j−1 ∪ E′j ∪ E′′j ∪ E j+1 ∪···∪ Es. • if there exist j < k such that min Jp+1 ∈ E j and max Jp+1 ∈ Ek, then E j \Jp+1 := E′j , Ek\Jp+1 := E′k where E′j , E′k are intervals. In that case, I \(J1 ∪ J2 ∪···∪ Jp ) = E1 ∪···∪ E j−1∪ E′j ∪ E′k ∪ Ek+1 ∪···∪ Es. • if min(Jp+1) does not belong to E1∪···∪Es or max(Jp+1) does not belong to E1∪···∪Es, the proof is similar and even simpler. (cid:3) We will use Lemma 4.5 under the form of the following corollary. Corollary 4.6. Let I , J1,... , Jp be intervals of R such that I >Pp contains an interval of length at least 1 j=1Jp. Then I \(J1∪ J2∪···∪ Jp ) We are now ready for the Proof of Theorem 4.4. Let q > 1 be such that, for any j ≥ 1, λ j+1 λ j ≥ q. Let us observe that if (µn) is an increasing sequence of positive real numbers such that {λn; n ≥ 1} is contained in {µn; n ≥ 1}, thenTa∈I HC (Tλn a) ⊂Ta∈I HC (Tµn a). Then, adding some terms to the sequence (λn) if necessary, we may always assume that, for any j ≥ 1, We argue by contradiction and we assume thatTa∈I HC (Tλn a) 6= ∅. To simplify the notations, we shall assume that I ⊂ R. Let m ≥ 1 be such that q m ≥ 2(m + 1) and let δ, A > 0 and N ∈ N be such that p+1³I−Pp j=1Jp´ . λ j+1 λ j ≤ q 2. q ≤ 1 q m−1¶ < 1 8 1 q +···+ δµ1+ A − 4δ q 2(m−1) > 1 1 λN ≤ I.   By Lemma 4.2, there exist M ≥ N and a finite sequence (yn,k)N≤n≤M, 1≤k≤qn such that (A) and (B) are satisfied. For n > M, we set qn = 0. Let n ≥ N be fixed. We set un = 0 if qn = 0. Otherwise, we construct intervals Jn,k and J′n,k as follows. We set k1 = 1 and λn¶ , J′n,1 =µyn,k1 − δ λn , yn,k1 + Let k2 be the first integer k > k1 such that yn,k ∉ J′n,1. Then by (A) λn(yn,k2 − yn,k1) > A. We λn´. We continue this process then set Jn,2 =³yn,k2 − 2δ , yn,k1 + λn´ and J′n,2 =³yn,k2 − δ a finite number of times (until this is impossible). At step s, we require that ks is the smallest Jn,1 =µyn,k1 − , yn,k2 + 2δ , yn,k2 + δ δ λn¶ . 2δ λn 2δ λn λn COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 17 integer k > ks−1 such that yn,k ∉ J′n, j , 1 ≤ j ≤ s − 1. We denote by un the numbers of intervals constructed in this way. For any n ≥ N , we have thus constructed a finite number of intervals (Jn, j )1≤ j≤un of length 4δ/λn such that (4) dist(Jn, j , Jn,k ) > qn[k=1µyn,k − δ λn , yn,k + A − 4δ provided j 6= k λn λn¶ ⊂ Jn,1 ∪···∪ Jn,un . δ 1 We finally set IN = I and, for j ≥ 0, IN+ j+1 = IN+ j \SuN+ j k=1 JN+ j ,k. By (B) and (4), IM+1 is empty. But we will contradict this fact by showing by induction on j ≥ 1 that IN+m j contains an interval . This property is true for m = 0 and assume that it is true at rank j . Let of length at least J ⊂ IN+m j be an interval of length . We claim that, for any n in {N+m j ,..., N+m( j +1)−1}, at most one interval Jn,k can intersect J. Indeed, for k 6= ℓ, λn ≥ dist(Jn,k , Jn,ℓ) > A − 4δ λn = q 2(m−1) > J. λN+m j λN+m j λN+m j 1 1 A − 4δ λN+m j × A − 4δ λN+m j × For n ∈ {N + m j ,..., N + m( j + 1)− 1}, let Kn = ∅ if no interval Jn,k intersect J and let Kn = Jn,k if Jn,k is the unique interval Jn,ℓ intersecting J. Then IN+m( j+1) ⊃ J\(KN+m j ∪···∪ KN+m( j+1)−1). 1 We apply Corollary 4.6: IN+m( j+1) contains an interval of length greater than m + 1¡J−KN+m j −···−Kn+m( j+1)−1¢ ≥ ≥ λN+m j −···− 1 1 4δ λN+m( j+1)−1¶ q m−1¶¶ 1 1 q +···+ 1 4δ λN+m j − m + 1µ (m + 1)λN+m j µ1− 4δµ1+ 2(m + 1)λN+m j 1 ≥ ≥ 1 . λN+m( j+1) This concludes the proof of Theorem 4.4. (cid:3) Remark 4.7. Theorem 4.4 remains true if we replace the condition liminf λn+1/λn > 1 by the following one: there exists p ≥ 1 such that liminf λn+p /λn > 1. Since the uniform mixing property implies the local separation property, we get a kind of du- ality for an operator group with the uniform mixing property. If the sequence (λn) does not increase too quickly, thenTa∈Rd\{0} HC (Tλn a) is nonempty. If the sequence (λn) does increase very quickly, then even for any nontrivial compact interval I in Rd \{0},Ta∈I HC (Tλn a) = ∅. When the compact interval I is "radial" (namely, when it is contained in a line passing through 0), then we can dispense with the local separation property in the statement of Theorem 4.4. Corollary 4.8. Let I be a nontrivial compact interval in Rd \{0} which is contained in a line pass- ing through 0 and let (Ta)a∈Rd be any strongly continuous operator group on X . Let also (λn) be an increasing sequence of positive real numbers such that liminfn λn+1/λn > 1. Then the set Ta∈I HC (Tλn a) is empty. 18 FRÉDÉRIC BAYART Of course, this corollary immediately implies Theorem 1.2. Proof. Let b ∈ Rd \{0} and κ > 1 be such that I = [b, κb]. Then consider the group (S t )t∈R defined by S t = Tκb. ThenTµ∈[1,κ] HC (Sλnµ) =Ta∈I HC (Tλn a). Now, a hypercyclic group defined on R has automatically the local separation property. Indeed, pick f ∈ X such that {S t f ; t ∈ R} is dense in X . Then, for any t ∈ [δ, A] ∪ [−A,−δ], S t f 6= f ; otherwise {S t f ; t ∈ R} would be compact. It is then easy to find by a compactness argument a neighbourhood U of f such that S tU ∩U = ∅ for any t ∈ R with δ ≤ t ≤ A. (cid:3) Question 4.9. Does Theorem 4.4 remains true if we do not assume that (Ta)a∈Rd is locally sepa- rating? Question 4.10. Let K be a compact subset of (0,+∞) and let (λn) = (q n), q > 1. Assume that Ta∈K HC (Tλn a) is nonempty. Can we link q and the Hausdorff dimension of K ? 5. COMMON FREQUENT HYPERCYCLICITY 5.1. A covering of the unit sphere with control. In this section, we study the existence of com- mon frequently hypercyclic vectors for operator groups, proving in particular Theorem 1.3. Our first main argument is a way to divide sequences of integers like in Corollary 2.7, but now with a control on the growth of the function φ (in order to obtain frequent hypercyclicity). We need to introduce a definition. Definition 5.1. We say that an increasing sequence (λn) has property (FHCSG) if • for any n ≥ 1, λn+1 − λn ≥ 1; • for any C > 0, there exists p ∈ N such that, for any N ≥ 1, N+p Xn=N+1 1 λn ≥ C λN . The main difference with property (SG) is that the number of terms of the sum appearing in the last displayed inequality does not depend on N . When a sequence satisfies property (FHCSG), we have the following improved version of Corol- lary 2.7. Lemma 5.2. Let (λn) be a sequence having property (FHCSG). Then for all d ≥ 1 and all A > 0, there exists Q ∈ N such that, for any N ∈ N, we can find s1 ∈ N, subsets Er of Nr−1 for r = 2,..., d+1, maps sr : Er → N for r = 2,... , d and a one-to-one map φ : Ed+1 → N such that • for any r = 2,..., d + 1, Er =©(k1,... , kr−1); k1 < s1, k2 < s2(k1),..., kr−1 ≤ sr−1(k1,... , kr−2)ª. • for any r = 1,..., d , for any (k1,..., kr−1) ∈ Er , λφ(k1,...,kr−1, j ,0,...,0) ≥ sr (k1,...,kr−1) 1 Xj=1 A λφ(k1,...,kr−1,0,...,0) . • For any (k1,..., kd ), (l1,..., ld ) ∈ Ed+1 with (k1,..., kd ) 6= (l1,..., ld ), then φ(k1,..., kd )− φ(l1,... , ld ) ≥ A N ≤ φ(k1,..., kd ) ≤ N +Q − 1. COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 19 Proof. We first observe that there exists ρ > 1 such that, for any n ∈ N, λn+1 ≤ ρλn. Indeed, there exists p ∈ N such that, for any N ∈ N, N+p Xn=N+1 1 λn ≥ 1 λN This yields given by Lemma 2.6. We apply property (FHCSG) with C > 0 such that . Let now d ≥ 1 and A > 0. Let κ ∈ N with κ > A and let B := B(d, A) be λN λN+1 ≥ 1 p . C κµ1+···+ 1 ρκ−1¶ ≥ B(d, A). This gives some p ∈ N. Let N ∈ N and define (µn) by setting µn = λN+nκ. Let s > 0 be such that sκ ≥ p. Then for any r ≥ 0, 1 1 1 1 1 λN+nκ + λN+nκ−1 +···+ λN+nκ−(κ−1)¶ 1 µn ≥ r+s Xn=r+1 1 Xn=r+1µ ρκ−1¶ r+s ρκ−1¶ p Xu=1 ρκ−1¶× λN+r κ 1 1 λN+r κ+u C 1 κµ1+···+ κµ1+···+ κµ1+···+ 1 B(d, A) . µr ≥ ≥ ≥ Hence, we may apply Lemma 2.6 to the sequence (µn) and to s. We get sr , Er and φ. We finally set ψ(k1,..., kd ) = N +κφ(k1,..., kd ). It remains to observe that N ≤ ψ(k1,..., kd ) ≤ N +κs to get that the conclusions of Lemma 5.2 are satisfied, with ψ instead of φ and Q = κs + 1. (cid:3) We deduce from this lemma a covering lemma for the unit sphere of Rd with control of the size of the covering. Lemma 5.3. Let (λn) be an increasing sequence of positive real numbers satisfying property (FHCSG). Then, for any d ≥ 1, for any δ > 0, for any B > 0, there exists q > 0 such that, for any N ∈ N, there exists a finite number of elements (xn,k)n=N,...,N+q−1 of S d−1 such that (A) S d−1 ⊂SN+q−1 (B) If (n, k)6= (m, j ), then kλn xn,k − λm xm, jk ≥ B . n=N Sk B³xn,k, δ λn´; The main difference with the coverings used when we applied Theorem 2.2 is that now we use at most q values of the sequence (λn), whereas this size was not controlled before. Proof. We first observe that S d−1 can be covered by a finite union of sets K1,..., Ku such that, for each j = 1,... , u, there exists a surjective map γ j : [0,1]d−1 → K j which is bilipschitz: ∃c > 0 such that, for any y, z ∈ [0,1]d−1, c−1ky − zk ≤ kγ j (y)− γ j (z)k ≤ cky − zk. Of course, c may be chosen to be independent of j . We apply Lemma 5.2 to d − 1 and to A = δ , B´ to get some Q ≥ 0. Let N ∈ N and let us first show how to cover K1. The sets Er , the max³ c2B maps sr , φ are defined by Lemma 5.2 and let n ∈ {N ,..., N + Q − 1}. Then either n is not equal 20 FRÉDÉRIC BAYART to some φ(k1,..., kd−1) and we do nothing or n = φ(k1,... , kd−1) for a unique (k1,..., kd−1). We then define the set {yn,k} as α1cB λφ(0,...,0) + λφ(1,0,...,0) +···+ λφ(k1,0,...,0) [0,1]d−1 ∩½µ δ/c δ/c , α2cB λφ(k1,0,...,0) + δ/c λφ(k1,1,0,...,0) +···+ δ/c λφ(k1,k2,...,0) , ... αd−1cB λφ(k1,...,kd−2,0) + λφ(k1,...,kd−2,kd−1)¶; α1,... , αd−1 ≥ 0¾. We set, for any n = N ,..., N + Q − 1 and any k, xn,k = γ1(yn,k) and let x ∈ K1, x = γ1(y). Arguing exactly as in the proof of Theorem 2.8 and since A ≥ c2B/δ, we find some (n, k) such that λφ(k1,...,kd−2,1) +···+ δ/c δ/c which implies ky − yn,kk ≤ δ cλn kx − xn,kk ≤ δ λn . Moreover, consider xn,k and xm, j with (n, k) 6= (m, ℓ). Then either n 6= m and by construction of φ, kλn xn,k − λm xm,ℓk ≥ λn − λm ≥ n − m ≥ A ≥ B. (5) Or n = m and in that case, as in the proof of Theorem 2.8, kλn yn,k − λn yn,ℓk ≥ cB which im- mediately yields kλn xn,k − λn xn,ℓk ≥ B. Thus, we have produced a good covering of K1. We produce a similar covering of K2 but starting from N +Q +κ and thus stopping at N +2Q −1+κ where κ ∈ N is such that κ ≥ B. More generally, we do the same for each K j , j = 1,... , u, starting at N + ( j − 1)(Q + κ) and stopping at N + ( j − 1)(Q + κ) + Q − 1. We finally get a net (xn,k)N≤n≤N+(u−1)(Q+κ)+Q−1 of S d−1 satisfying (A) with q = (u − 1)(Q + κ)+ Q − 1. Moreover, (B) is also satisfied since, if xn,k belongs to the covering of K j and xm,ℓ belongs to the covering of Kℓ for ℓ 6= k, then n − m ≥ κ so that kλn xn,k − λm xm,ℓk ≥ B. (cid:3) We now combine the previous covering argument with the production of sets with positive lower density. Lemma 5.4. Let (λn) be an increasing sequence of positive real numbers satisfying property (FHCSG). Let (Bp ) and (δp ) be two sequences of positive real numbers. Then there exist a se- quence (qp) of positive integers, a sequence (Np ) of subsets of N such that (1) for any p ≥ 1, for any N ∈ Np , there exists a finite number (xn,k)N≤n≤N+qp −1 of elements of S d−1 such that • S d−1 ⊂SN+qp−1 • if (n, k) 6= (m, ℓ), N ≤ n, m ≤ N + qp − 1, kλn xn,k − λm xm,ℓk ≥ Bp. λn´ . Sk B³xn,k, δp n=N (2) Each set Np has positive lower density. (3) For any p, r ≥ 1 and any (N , M) ∈ Np × Nr with N 6= M , N − M ≥ (Bp + Br + qp + qr ). (4) For any p ≥ 1, min(Np ) ≥ Bp . COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 21 Proof. For a fixed value of p (and thus of Bp and δp ), let qp > 0 be given by Lemma 5.3. We then apply [6, Lemma 6.19] to the sequence (Np ) defined by Np = Bp + qp. This provides a sequence (Np ) of pairwise disjoint subsets of N such that each Np has positive upper density, min(Np ) ≥ Np and N − M ≥ Np + Nr whenever N 6= M and (N , M)∈ Np × Nr . Finally, condition (1) follows by applying Lemma 5.3 for any N ∈ Np . (cid:3) 5.2. The uniform frequent hypercyclicity criterion. We now give a criterion for an operator group to have a common frequently hypercyclic vector. It is not very surprizing that this crite- rion is a strenghtened version of the frequent hypercyclicity criterion. Definition 5.5. Let (Ta)a∈Rd be an operator group on X and let k·k be an F -norm on X . We say that (Ta) satisfies the uniform frequent hypercyclicity criterion if there exists D ⊂ X dense such that, for any f ∈ D, (1) P Tai f converges for any sequence (ai ) ⊂ Rd with kai − a jk ≥ 1 for any i 6= j . (2) sup°°P Tai f°° tends to zero as C goes to infinity, where the supremum is taken over the sequences (ai ) ⊂ Rd such that kai − a jk ≥ 1 for any i 6= j and kaik ≥ C for any i . We need a very last lemma on sets with positive lower density. Lemma 5.6. Let E ⊂ N with positive lower density and let q ∈ N. Let F ⊂ N be such that, for any n ∈ E , [n, n + q)∩ F is nonempty. Then F has positive lower density. Proof. Write E as an increasing sequence (nk). For all k ≥ 1, there exists at least one element of F in [nk q, n(k+1)q). Hence dens(F ) ≥ dens¡(nk q )¢ ≥ 1 q dens(E) > 0. (cid:3) Theorem 5.7. Let (Ta)a∈Rd be a strongly continuous operator group on X satisfying the uniform frequent hypercyclicity criterion. Let (λn) be an increasing sequence of positive integers satisfying property (FHCSG). ThenTa∈Sd−1 F HC (Tλn a) is nonempty. Proof. Let ( fp) ⊂ D be a dense sequence in X . For any p ≥ 1, let δp > 0 be such that kTa fp− fpk < 2−p provided kak < δp . We then consider a sequence (Bp)p≥1 where Bp ≥ 1 is such that, for any sequence (ai ) ⊂ Rd with kaik ≥ Bp − 1 and kai − a jk ≥ 1 whenever i 6= j , then°°Pi Tai fk°° < 2−p for any k ≤ p. We then apply Lemma 5.3 to these sequences (Bp) and (δp ). For all p ≥ 1, we define the set {ai (p)} as the set of all the λnxn,k for N describing Np and N ≤ n ≤ N+qp−1. Then we may observe that, for any i ≥ 1, kai (p)k ≥ λn ≥ Bp and for i 6= j , kai (p)− a j (p)k ≥ 1. This comes trivially from the lemma if ai (p) = λn xn,k and a j (p) = λm xm,ℓ with N ≤ n, m ≤ N +qp−1 and N ∈ Np (observe that there exists at most one N in Np such that N ≤ n ≤ N + qp − 1). Otherwise, kai (p)− a j (p)k ≥ λn − λm ≥ n − m ≥ inf{N − M; M, N ∈ Np, M 6= N }− qp ≥ Bp ≥ 1. The same proof shows that, if p 6= r , then for any i , j , kai (p)− a j (r )k ≥ Bp + Br . 22 FRÉDÉRIC BAYART By definition of Bp, we know that for each p ≥ 1, the seriesPi≥1 T−ai (p) fp converges and that °°Pi≥1 T−ai (p) fp°° ≤ 2−p . We finally set and we claim that f ∈Ta∈Sd−1 F HC (Tλn a). Indeed, let a ∈ S d−1, g ∈ X and let ε > 0. There exists p ≥ 1 such that k fp − gk < 2−p and 2−p (p + 3) < ε. Moreover, for any N ∈ Np , there exists n ∈ {N ,..., N +qp−1} such that kλn a−λn xn,kk < δp . Let i ≥ 1 be such that ai (p) = λn xn,k. Then f = Xp≥1Xi≥1 T−ai (p) fp Tλn a f − g = Xr<pXj Tλn a−a j (r ) fr +Xj6=i Tλn a−a j (p) fp +¡Tλn a−ai (p) fp − fp¢ +¡ fp − g¢+Xr>pXj Tλn a−a j (r ) fr . Let us set, for any r, j , b j (r ) = λn a − a j (r ). Then kb j (r )− bk(r )k ≥ 1 for any r ≥ 1 and any k 6= j . Moreover, if r 6= p, then kb j (r )k ≥ kai (p)− a j (r )k−kai (p)− λn ak ≥ Bmax(r,p) − 1 and this inequality is also true for r = p and j 6= i . By definition of (Br ) and δp we then get 2−r ≤ (p + 3)2−p < ε. Lemma 5.6 then achieves the proof that f is a frequently hypercyclic vector for Ta. °°Tλn a f − g°° ≤ (p − 1)2−p + 2−p + 2−p + 2−p +Xr>p (cid:3) Wehre (λn) is the whole sequence of integers, we can combine this with an algebraic result of [8] to obtain: We now give examples where the previous theorems may be applied. Theorem 5.8. Let (Ta)a∈Rd be a strongly continuous operator group satisfying the uniform fre- quent hypercyclicity criterion. ThenTa∈Rd \{0} F HC (Ta) is nonempty. Corollary 5.9. Let (Ta)a∈Rd be an operator group acting on X and let k·k be an F -norm on X . Assume that there exist a dense set D ⊂ X and p > d such that, for any f ∈ D, there exists A > 0 such that kTa f k ≤ A kakp for any a ∈ Rd with kak ≥ 1. ThenTa∈Rd \{0} F HC (Ta) is nonempty. Proof. The proof of Corollary 3.3 shows that, under the above assumptions, (Ta)a∈Rd satisfies the uniform frequent hypercyclicity criterion. (cid:3) Corollary 5.10. Let w : Rd → R be a positive bounded and continuous function such that x 7→ w(x+a) is bounded for each a ∈ Rd . For a ∈ Rd , let τa be the translation operator defined by w(x) τa f (x) = f (x + a) defined on X = Lp (Rd , w(x)d x), p ≥ 1. Assume moreover thatRRd w(x)d x < +∞. ThenTa∈Rd \{0} F HC (τa) is nonempty. Proof. Let D ⊂ X be the dense set of compactly supported continuous functions. Let f ∈ D and let A > 0 be such that the support of f is contained in B(0, A). Let (ai ) ⊂ Rd be such that kai − a jk ≥ 1 for any i 6= j . Then it is easy to check that there exists κ > 0 which just depends on the dimension d and of A (and not of the particular choice of the sequence (ai )) such that, COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 23 for any x ∈ Rd , kx − aik ≤ A for at most κ different ai . This implies that, for all x ∈ Rd , the series Pi f (x + ai ) is convergent (the sum is finite) and that¯¯Pi f (x + ai )¯¯ ≤ κk f k∞. Therefore, if we assume moreover that infi kaik ≥ C , then p °°°°°Xi Tai f°°°°° ≤ κpk f k p ∞Zkxk≥C−A w(x)d x and this goes to zero as C goes to +∞. Hence, the sequence (τa)a∈Rd satisfies the uniform frequent hypercyclicity criterion. (cid:3) Corollary 5.11. The composition operators on H 2(Hd ) induced by a Heisenberg translation ad- mit a common frequently hypercyclic vector. Proof. As for Theorem 3.7, we may apply directly Corollary 5.9. (cid:3) 5.3. Uniformly Runge transitive operators groups. We now turn to the case of the translation group on H(Cd ). The translation operators τa, a ∈ C∗, acting on H(C), have very special dy- namical properties due to the strongness of Runge theorem. An abstract framework to mimic what is useful in Runge theorem has been done in linear dynamics for at least two problems: • the problem of common hypercyclicity of the whole operator group, like in [17] or in this paper. The natural generalization here seems the Runge property. • the problem of frequent hypercyclicity; this leads the authors of [7] to introduce the notion of a Runge transitive operator. Since we want a common frequently hypercyclic vector, the right concept seems to be a mixing of these two properties. Definition 5.12. Let (Ta)a∈Rd be an operator group acting on a Fréchet space X . We say that (Ta)a∈Rd is uniformly Runge transitive if there is an increasing sequence (k·kp) of seminorms defining the topology of X and positive integers Ap,s, Cp for p, s ∈ N such that (i) for all p, s ∈ N, f ∈ X and a ∈ Rd with kak ≤ s, kTa f kp ≤ Ap,sk f ks+Cp . (ii) for all p, s ∈ N with s > Cp , for all g , h ∈ X , ε > 0 and for all finite sets (zi ) ⊂ Rd with kzi − z jk ≥ Cp and kzik ≥ s whenever i 6= j , there exists f ∈ X such that k f − gks−Cp < ε and kTzi f − hkp < ε for all i . Theorem 5.13. Let (λn) be an increasing sequence of positive real numbers satisfying property (FHCSG) and let (Ta)a∈Rd be a strongly continuous operator group which is uniformly Runge transitive. ThenTa∈Sd−1 F HC (Tλn a) is nonempty. In particular,Ta∈Rd \{0} F HC (Ta) 6= ∅. Proof. We first fix a dense sequence (hp ) in X and a sequence (εp ) of positive real numbers going to 0. Let (k·kp ) be the sequence of seminorms defining the topology of X coming from the definition of the uniform Runge transitivity. We then consider a decreasing sequence (δp ) of positive real numbers and a sequence (Mp) of integers such that, for any p ≥ 1, • kTahp − hpkp < εp provided kak < δp ; • kTa f kp < εp provided k f kMp < δp and kak < δp . 24 FRÉDÉRIC BAYART Since (Ta)a∈Rd is uniformly Runge transitive, there exist sequences (Ap,s ) and (Cp ) such that (be careful! We have slightly changed the notations of Definition 5.12 to adapt them to our present context. More precisely, k·kp is replaced by k·kMp and s is replaces by λs.) (i) for all p, s ∈ N, f ∈ X and a ∈ Rd with kak ≤ λs, kTa f kMp ≤ Ap,sk f kλs+Cp . (ii) for all p, s ∈ N with λs > Cp , for all g , h ∈ X , all ε > 0, all finite sequences (zi ) ⊂ Rd with kzi − z jk ≥ Cp and kzik ≥ λs whenever i 6= j , there exists f ∈ X with k f − gkλs−Cp < ε and kTzi f − hkMp < ε for all i . We apply Lemma 5.4 with Bp = Cp + p and δp fixed above. We get the sequences (Np ) and (qp) and we writeSp≥1 Np as an increasing sequence (n j ). For any j ≥ 1, there is a unique p j such that n j ∈ Np j . Moreover, λn j ≥ n j ≥ Bp j > Cp j and n j ≥ Bp j ≥ p j . We then define by induction on j a sequence ( f j ) ⊂ X by setting f0 = 0 and f j is such that (6) ηn j k f j − f j−1kλn j −Cp j ≤ 1+ max(Apt ,n; t ≤ j , n ≤ nt + qpt ) kTλn xn,k f j − hp j kMp j < ηn j for all n in {n j ,... , n j + qp j − 1} and all possible k where (η j ) is a sequence of positive real numbers such thatP j≥k η j ≤ δk for any k. It is possible to find such an f j because kλn xn,k − λm xm,ℓk ≥ Cp j if (n, k) 6= (m, ℓ) and kλn xn,kk ≥ λn j for all n = n j ,..., n j + q j −1 and all k. The choice of Bp ensures that (λn j −Cp j ) tends to +∞ as j tends to +∞. Therefore, (6) implies that ( f j ) converges to some f ∈ X . Let us now fix j ≥ 1 and ℓ ≥ j . Then, for all n ∈ {n j ,..., n j + qp j − 1} and all possible k, kTλn xn,k ( fℓ+1 − fℓ)kMp j ≤ Ap j ,n j +qp j k fℓ+1 − fℓkλn j +qp j +Cp j . Now, λnℓ+1 −Cpℓ+1 ≥ λn j + qp j +Cp j since λnℓ+1 − λn j ≥ nℓ+1 − n j ≥ Cpℓ+1 +Cp j + qp j . Hence, °°Tλn xn,k ( fℓ+1 − fℓ)°°Mp j ≤ Ap j ,n j+qp j k fℓ+1 − fℓkλnℓ+1−Cpℓ+1 ≤ ηnℓ+1 by (6). Summing these inequalities, we have then shown that (7) °°Tλn xn,k f − hp j°°Mp j ≤Xl≥ j ηnl ≤ δn j ≤ δp j . Let us now show that f ∈Ta∈Sd−1 F HC (Tλn a). Indeed, let a ∈ S d−1, let p ∈ N and N ∈ Np. There exist n ∈ {N ,..., N + qp − 1} and k such that kλn a − λn xn,kk < δp . Then °°Tλn a f − hp°°p ≤ °°Tλn a−λn xn,k¡Tλn xn,k f − hp¢°°p +°°Tλn a−λn xn,k hp − hp°°p ≤ 2εp where the last inequality comes from (7) and from the definitions of δp and Mp. We now con- (cid:3) clude exactly as in the proof of Theorem 5.7 that f ∈Ta∈Sd−1 F HC (Tλn a). Example 5.14. The group of translations (τa)a∈C acting on H(C) is uniformly Runge transitive. COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 25 Proof. Let δ ∈ (0,1) and set k f kn = sup with a ≤ s, z≤n−δ f (z). We have, for all p, s ∈ N, f ∈ X and a ∈ C kTa f kp = sup z−a≤p−δ f (z) ≤ sup z≤s+p−δ f (z) = k f ks+p and, if (zi ) is a finite sequence with kzi−z jk ≥ p and kzik ≥ s with s > p, the disks D(zi , p−δ) are disjoint and disjoint from the disk D(0, s − p − δ). Hence, Runge's theorem immediately yields the existence of f ∈ X with sup z<s−p−δ f (z)− g (z) < ε and sup z−zi<p−δ f (z)− h(z − zi ) < ε. (cid:3) In the particular case of the translation group on H(C), the proof of Theorem 5.13 could be slightly simplified by using Arakelian's theorem of approximation of holomorphic functions on closed sets. However, our general theorem may be applied in other contexts, like the translation group on C (Rd ) or on H(Cd ), if we restrict ourselves to translations by an element of Rd . 6. MULTIPLES OF A SEMIGROUP In this section, we show that we can get a common hypercyclic vector if we consider the multi- ples of an operator group having the Runge property. As usual, we first need a covering lemma. Lemma 6.1. Let d ≥ 1, δ > 0, B > 0 and I be a compact interval of R. Then there exist an integer r ≥ 0, elements (y j ) j=1,...,r of S d−1, elements (α j ) j=1,...,r of I and integers (n j ) j=1,...,r such that • for all α ∈ I and all y ∈ S d−1, there exists j ∈ {1,... , r } with . α− α j < and ky − y jk ≤ δ n j δ n j • for all j 6= l in {1,... , r }, kn j y j − nl ylk ≥ B. • for all j ∈ {1,... , r }, n j ≥ B . Proof. We combine the Costakis-Sambarino method and the methods of the present paper to obtain the right covering. We begin by applying Lemma 5.3 to the whole sequence of integers; we get some q ∈ N. Without loss of generality, we may assume that B ∈ N. We then define a (m+1)(B+q) . Let s ≥ 1 be the biggest sequence (βm)m≥1 by setting β1 = min(I ) and βm+1 = βm + integer m such that βm ≤ max(I ). For each m = 1,..., s, we then apply Lemma 5.3 with N = Nm := m(B + q). We get elements (xn,k(m)) of S d−1 for Nm ≤ n ≤ Nm + q − 1, k ≤ ω(n). We then rename the xn,k(m) by defining δ ©y j ; j = 1,..., rª :=©xn,k(m); 1 ≤ m ≤ s, Nm ≤ n ≤ Nm + q − 1, k ≤ ω(n)ª. For any j in {1,... , r }, there exists a unique (m, n, k) such that y j = xn,k(m). We then set n j = n and α j = βm. We now verify that the conclusions of Lemma 6.1 are satisfied. Pick (α, y) ∈ I×S d−1. There exists m ∈ {1,... , s} such that α− βm ≤ (m+1)(B+q) . This m being fixed, there exist n ∈ {Nm,... , Nm + q − 1} and k ≤ ω(n) such that ky − xn,k(m)k ≤ δ n . Let j ∈ {1,... , r } be such that y j = xn,k(m) so that n j = n and α j = βm. Since n ≤ (m+1)(B + q), the first part of the conclusions of the lemma is verified. δ 26 FRÉDÉRIC BAYART Suppose now that j 6= l are living in {1,... , r }. If y j = xk,n(m) and yl = xn′,k′(m) for the same m, then the inequality kn j y j − nl ykk ≥ B follows directly from the corresponding inequality of Lemma 5.3. Otherwise, we simply write kn j y j − nl ylk ≥ n j − nl ≥ B. Finally, for any j in {1,... , r }, n j ≥ B from the very definition of (Nm). We need a second lemma related to the continuity of (λ, a) 7→ λTa. It is [17, Lemma 3.5] where it is formulated for d = 2, but the proof is unchanged for greater values of d. We now assume that X is a Fréchet space. (cid:3) Lemma 6.2. Let (Ta)a∈Rd be a strongly operator group on X , let g ∈ X and let k·k be a continuous seminorm on X . Then there exist a continous seminorm · on X and δ > 0 such that k ·k ≤ · and, for any α ∈ R, x ∈ S d−1, n ∈ N and f ∈ X satisfying g − e αnTnx f < 1, we have kg − e βnTny f k < 1 whenever β ∈ R and y ∈ S d−1 are such that α− β < δ We can now give our multidimensional analogue of Shkarin's result. n and ky − xk < δ n . Theorem 6.3. Let (Ta)a∈Rd be a strongly operator group on X with the Runge property. Then Tλ∈C∗, a∈Rd\{0} HC (λTa) is a residual subset of X . Proof. First of all, as we have already done previously, we apply the algebraic results of Leon and Müller and of Conejero, Müller and Peris. They give that HC (λTa) = HC (µTb) provided λ = µ and a = θb for θ > 0. Hence, it is sufficient to show that, for any compact set I ⊂ R, the family of operators {e αTa; α ∈ I , a ∈ S d−1} shares a common hypercyclic vector. For this, we argue as in the proof of Theorem 2.2 by picking two nonempty open subsets U and V of X and by showing that U ∩© f ∈ X ; ∀α ∈ I , ∀a ∈ S d−1, ∃n ∈ N, e nαTna f ∈ Vª 6= ∅. Let g , h ∈ X , k·k be a continuous seminorm on X such that { f ∈ X ; k f − hk < 1} ⊂ U and { f ∈ X ; k f − gk < 1} ⊂ V . Let δ > 0 and let · be a seminorm on X satisfying the conclusions of Lemma 6.2. Let also C > 0 be given by the Runge property for this last seminorm ·. We apply Lemma 6.1 with I , δ > 0 and B = C to get finite sequences (α j ), (y j ) and (n j ). By the Runge property, there exists f ∈ X such that • f − h < 1; • for any j = 1,... , r , e n j α j Tn j y j f − g < 1. Then f ∈ U . Moreover, for any (α, a) ∈ I × S d−1, there exists j such that α− α j < δ/n j and ka − y jk < δ/n j . By the choice of δ and ·, this yields ke n j αTn j a f − gk < 1, which concludes the proof of Theorem 6.3. (cid:3) By considering multiples of a semigroup, we cannot go much further; in particular, we cannot get a common frequently hypercyclic vector for the family©λTa; λ > 0, a ∈ S d−1ª. In fact, this cannot be the case even for an uncountable family of multiples of a single operator, as the following proposition points out. It should be noticed that it improves [5, Theorem 4.5] where T was equal to B the backward shift. However, the proof remains almost identical. COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 27 Proposition 6.4. Let T be a continuous operator acting on the F -space X and let Λ be an un- countable subset of (0,+∞). Then the set of common frequently hypercyclic vectors for the family (λT )λ∈Λ is empty. Proof. Let x∗ be a nonzero linear functional on X and assume by contradiction that x is a com- mon frequently hypercyclic vector for all operators λT , λ ∈ Λ, where Λ is uncountable. Let, for any λ ∈ Λ, Nλ =©n ∈ N; x∗(λnT n x) ∈ (1/2,3/2)ª, δλ = dens(Nλ). For all λ ∈ Λ, δλ > 0. Since Λ is uncountable, this implies that there exist λ 6= µ in Λ such that Nλ ∩ Nµ is infinite (see [5] for details). Now, if n belongs to Nλ ∩ Nµ, then λn(T ∗)n x∗(x) ∈ (1/2,3/2) and µn(T ∗)n x∗(x) ∈ (1/2,3/2). This yields (λ/µ)n ∈ (1/3,3), which is a contradiction since n may be chosen as large as we want. (cid:3) Our argument in this section is really specific to operators groups having the Runge property. In view of [17, Corollary 1.10] and of the other results of the present paper, the following question seems natural. Question 6.5. Let (Ta)a∈Rd be a strongly continuous operator group. Assume that there exist δ ∈ (0,1) and D ⊂ X a dense subset of X such that, for any f ∈ D, kTa f k ≤ C f δkak(d−1). Does there exist a common hypercyclic vector for the family©λTa; δ < λ < 1/δ, a ∈ Rd \{0}ª? 7. A NEW LIGHT ON THE ALGEBRAIC METHOD 7.1. Yet another criterion. We now show how the algebraic method can also lead to common hypercyclic results in high dimension. Theorem 7.1. Let G be a compact topological group, let (Tn,g )(n,g )∈N×G be a strongly continuous operator semigroup on X . Then for any g ∈ G, HC (T1,g ) = HC (T1,1G ). Proof. Let u ∈ HC (T1,1G ), let v ∈ X and let g ∈ G. By [4, Theorem 2.2] (which is itself a conse- quence of results of [16]), there exists a sequence (nk) of integers such that Then (T1,1G )nk u → v and g nk → 1G . (T1,g )nk u − v = Tnk ,g nk u − T0,g nk v + T0,g nk v − v = T0,g nk (Tnk ,1G u − v)+ T0,g nk v − v. Now, since (Tn,g )(n,g )∈N×G is strongly continuous, the map (h, w)∈ G×X 7→ T0,h w is continuous by the uniform boundedness principle. We easily deduce that (T1,g )nk u tends to v. Conversely, let g ∈ G, u ∈ HC (T1,g ) and v ∈ X . Let (nk) be a sequence of integers such that Then we get that T nk 1,1G u tends to v by writing (T1,g )nk u → v and g−nk → 1G . T nk 1,1G u − v = T0,g −nk (Tnk ,g nk u − v)+ T0,g −nk v − v. (cid:3) 28 FRÉDÉRIC BAYART This theorem may be applied with G = T the unit circle and Tn,ξ = ξT n with T any operator on X . It gives back the Leon-Müller theorem. However, as we have promised, it also leads to interesting multidimensional results. 7.2. Hyperbolic automorphisms of the ball. We come back to our discussion on composition operators on the Hardy space of the Siegel upper half-space Hd . Let U ∈ C(d−1)×(d−1) be a uni- tary matrix, let λ > 1 and define φλ,U (z, w) = (λz,U w). The maps φλ,U are automorphisms of Hd (now hyperbolic automorphisms) and it has been shown in [12] that, for any λ > 1 and any U ∈ U(Cd−1) (the set of unitary matrices over Cd−1), the composition operator Cφλ,U is hypercyclic on H 2(Hd ). Theorem 7.2. The setTλ>1, U∈U(Cd−1) HC (Cφλ,U ) is a residual subset of H 2(Hd ). Proof. The proof is divided into two parts, the multidimensional part and the one-dimensional part. For the multidimensional part, we fix λ > 1, we set G = U(Cd−1) and we write Tn,U = Cφλn ,U . This is clearly a strongly continuous semigroup on H 2(Hd ). Hence we may apply Theorem 7.2 and we know that HC (Cφλ,U ) = HC (Cφλ,I ) for any U ∈ U(Cd−1). It is now sufficient to show that, for any λ > 1, HC (Cφλ,I ) = HC (Cφe,I ). This is also due to a semi- group argument! Indeed, define, for a > 0, Ta = Cφea ,I . Then (Ta)a>0 is a strongly continuous semigroup hence by the Conejero-Müller-Peris theorem, HC (Ta) = HC (T1) for any a > 0. (cid:3) The previous result and Theorem 3.7 suggest the following natural question. Question 7.3. Let A the set of all automorphisms of Hd with +∞ as attractive fixed point. Is Tφ∈A HC (Cφ) nonempty? REFERENCES [1] E. Abakumov and J. Gordon, Common hypercyclic vectors for multiples of backward shift, J. Funct. Anal. 200 (2003), 494 -- 504. [2] F. Bayart, Parabolic composition operators on the ball, Advances in Math 233 (2010), 1666 -- 1705. [3] F. Bayart and S. Charpentier, Hyperbolic composition operators on the unit ball, Transactions of the AMS 365 (2013), 911 -- 938. [4] F. Bayart and G. Costakis, Hypercyclic operators and rotated orbits with polynomial phases, J. Lond. Math. Soc. 89 (2014), 663 -- 679. [5] F. Bayart and S. Grivaux, Frequently hypercyclic operators, Trans. Amer. Math. Soc. 358 (2006), no. 11, 5083 -- 5117. [6] F. Bayart and É. Matheron, Dynamics of linear operators, Cambridge Tracts in Math, vol. 179, Cambridge Uni- versity Press, 2009. [7] A. Bonilla and K.-G. Grosse-Erdmann, Frequently hypercyclic operators and vectors, Ergodic Theory Dynamical Systems 27 (2007), 383 -- 404. [8] J. A. Conejero, V. Müller, and A. Peris, Hypercyclic behaviour of operators in a hypercyclic C0-semigroup, J. Funct. Anal. 244 (2007), 342 -- 348. [9] G. Costakis and M. Sambarino, Genericity of wild holomorphic functions and common hypercyclic vectors, Adv. Math. 182 (2004), 278 -- 306. [10] G. Costakis, N. Tsirivas, and V. Vlachou, Non-existence of common hypercyclic entire functions for certain fami- lies of translation operators, preprint (2014). [11] K.-G. Grosse-Erdmann and A. Peris, Linear chaos, Springer, 2011. [12] C. Guangfu, G. Kunyu, and C. Xiaoman, Inner Functions and Cyclic Composition Operators on H 2(Bn ), Jour. Math. Anal. Appl. 250 (2000), 660 -- 669. COMMON HYPERCYCLIC VECTORS FOR HIGH DIMENSIONAL FAMILIES OF OPERATORS 29 [13] G. Khudaiberganov, Polynomial and rational convexity of the union of compact sets of Cn, Soviet. Math. 31 (1987), 111 -- 117. [14] F León-Saavedra and V. Müller, Rotations of hypercyclic and supercyclic operators, Integral Equations Operator Theory 50 (2004), 385 -- 391. [15] R.M. Range, Holomorphic functions and integral representations in several complex variables, Graduate Text in Mathematics, vol. 108, Springer, 1998. [16] S. Shkarin, Universal elements for non-linear operators and their applications, J. of Math. Anal. App. 348 (2008), 193 -- 210. [17] [18] N. Tsirivas, Common hypercyclic functions for translation operators with large gaps, preprint (2014), , Remarks on common hypercyclic vectors, J. Funct. Anal. 258 (2010), 132 -- 160. arXiv:1412.0827. [19] , Existence of common hypercyclic vectors for translation operators, preprint (2014), arXiv:1411.7815. CLERMONT UNIVERSITÉ, UNIVERSITÉ BLAISE PASCAL, LABORATOIRE DE MATHÉMATIQUES, BP 10448, F-63000 CLERMONT- FERRAND - CNRS, UMR 6620, LABORATOIRE DE MATHÉMATIQUES, F-63177 AUBIERE E-mail address: [email protected]
1601.07769	1	1601	2016-01-28T14:17:53	Normal extensions	[ "math.FA" ]	Let $L_0$ be a densely defined minimal linear operator in a Hilbert space $H$. We prove theorem that if there exists at least one correct extension $L_S$ of $L_0$ with the property $D(L_S)=D(L_S^)$, then we can describe all correct extensions $L$ with the property $D(L)=D(L^)$. We also prove that if $L_0$ is formally normal and there exists at least one correct normal extension $L_N$, then we can describe all correct normal extensions $L$ of $L_0$. As an example, the Cauchy-Riemann operator is given.	math.FA	math	NORMAL EXTENSIONS B. N. Biyarov Key words: Formally normal operator, normal operator, correct restriction, correct exten- sion. AMS Mathematics Subject Classification: Primary 47Axx, 47A05; Secondary 47B15. Abstract. Let L0 be a densely defined minimal linear operator in a Hilbert space H. We prove theorem that if there exists at least one correct extension LS of L0 with the property D(LS) = D(L∗ S), then we can describe all correct extensions L with the property D(L) = D(L∗). We also prove that if L0 is formally normal and there exists at least one correct normal extension LN , then we can describe all correct normal extensions L of L0. As an example, the Cauchy-Riemann operator is given. 1 Introduction Let us present some definitions, notation, and terminology. In a Hilbert space H, we consider a linear operator L with domain D(L) and range R(L). By the kernel of the operator L we mean the set Ker L =(cid:8)f ∈ D(L) : Lf = 0(cid:9). Definition 1. An operator L is called a restriction of an operator L1, and L1 is called an extension of an operator L, briefly L ⊂ L1, if: 1) D(L) ⊂ D(L1), 2) Lf = L1f for all f from D(L). Definition 2. A linear closed operator L0 in a Hilbert space H is called minimal if R(L0) 6= H and there exists a bounded inverse operator L−1 0 on R(L0). Definition 3. A linear closed operatorbL in a Hilbert space H is called maximal if R(bL) = H and KerbL 6= {0}. Definition 4. A linear closed operator L in a Hilbert space H is called correct if there exists a bounded inverse operator L−1 defined on all of H. Definition 5. We say that a correct operator L in a Hilbert space H is a correct extension Definition 6. We say that a correct operator L in a Hilbert space H is a boundary correct of minimal operator L0 (correct restriction of maximal operator bL) if L0 ⊂ L (L ⊂bL). extension of a minimal operator L0 with respect to a maximal operator bL if L is simultane- ously a correct restriction of the maximal operatorbL and a correct extension of the minimal operator L0, that is, L0 ⊂ L ⊂bL. At the beginning of the 1950s, Vishik [10] extended the theory of self-adjoint extensions of von Neumann -- Krein symmetric operators to nonsymmetric operators in Hilbert space. At the beginning of the 1980s, M. Otelbaev and his disciples proved abstract theorems that allows us to describe all correct extensions of some minimal operator using any single known correct extension in terms of an inverse operator. Here such extensions need not be restrictions of a maximal operator. Similarly, all possible correct restrictions of some maximal operator that need not be extensions of a minimal operator were described (see [7]). For convenience, we present the conclusions of these theorems. 1 Let bL be a maximal linear operator in a Hilbert space H, let L be any known correct restriction of bL, and let K be an arbitrary linear bounded (in H) operator satisfying the following condition: (1.1) Then the operator L−1 K defined by the formula R(K) ⊂ KerbL. L−1 K f = L−1f + Kf, (1.2) describes the inverse operators to all possible correct restrictions LK of bL, i.e., LK ⊂bL. Let L0 be a minimal operator in a Hilbert space H, let L be any known correct extension of L0, and let K be a linear bounded operator in H satisfying the conditions a) R(L0) ⊂ Ker K, b) Ker (L−1 + K) = {0}, K defined by formula (1.2) describes the inverse operators to all possible then the operator L−1 correct extensions LK of L0. of at least one boundary correct extension L was proved by Vishik in [10]. Let K be a linear bounded (in H) operator satisfying the conditions Let L be any known boundary correct extension of L0, i.e., L0 ⊂ L ⊂ bL. The existence b) R(K) ⊂ KerbL, then the operator L−1 boundary correct extensions LK of L0. K defined by formula (1.2) describes the inverse operators to all possible a) R(L0) ⊂ Ker K, Self-adjoint and unitary operators are particular cases of normal operators. A bounded linear operator N in a Hilbert space H is called normal if it commutes with its adjoint: The theory of bounded normal operators are sufficiently developed. Consider an unbounded linear operator A in a Hilbert space H. N ∗N = N N ∗. Definition 7. A densely defined closed linear operator A in a Hilbert space H is called formally normal if D(A) ⊂ D(A∗), kAf k = kA∗f k for all f ∈ D(A). Definition 8. A formally normal operator A is called normal if D(A) = D(A∗). Normal extensions of formally normal operators have been studied by many authors (see [1], [5], [6], [9]). Questions the existence of a normal extension and the description of the domains of normal extensions of a formally normal operator were considered. The spectral properties of the correct restrictions and extensions were systematically In these works a class of operators K that provides studied by the author (see [2] -- [4]). Volterra, the completeness of root vectors, and the dissipativity of the correct restrictions and extensions were described. The present paper is devoted to the description of correct normal extensions in terms of the operator K. 2 Coincidence criterion of D(L) with D(L∗) We consider a densely defined minimal linear operator L0 in a Hilbert space H. Let M0 be a minimal operator with D(M0) = D(L0) that is connected with L0 by the relation 0 is an 0 is an extension of M0. The following (L0u, v) = (u, M0v) for all u, v from D(L0). Then the maximal operator bL = M ∗ extension of L0, and the maximal operator cM = L∗ statement is true. 2 Assertion 1. If there exists a correct extension LS of the minimal operator L0 with the property D(LS) = D(L∗ S), then the operator LS is the boundary correct extension, i.e., L0 ⊂ LS ⊂bL. Proof. From L0 ⊂ LS it follows that L∗ D(M0) ⊂ D(L∗ S) we have S ⊂ L∗ M0 ⊂ L∗ 0 = cM . From D(LS) = D(L∗ S ⊂ cM . S) and the fact that Then L0 ⊂ LS ⊂bL. The assertion is proved. Let there be one fixed correct extension LS of L0 such that D(LS) = D(L∗ S). Then we can describe the inverses to all boundary correct extensions L in the following form u = L−1f = L−1 S f + Kf for all f ∈ H, (2.1) where K is an arbitrary bounded operator in a Hilbert space H that R(K) ⊂ KerbL and R(L0) ⊂ Ker K. Each such operator K defines one boundary correct extension and there do not exist other boundary correct extensions. operator, we obtain a Hilbert space with the scalar product Let us equip D(bL) with the graph norm uG = (u2 + bLu2)1/2. Since bL is a closed (u, v)G = (u, v) + (bLu,bLv) for all u, v from D(bL). Let us denote this space by G bL. The domain D(LS) of the correct restriction LS is a subspace in G bL. Therefore, there exists a projection operator of G bL on the subspace D(LS). As such a projection operator, we take L−1 S bL. Then the projection ΓLS = I − L−1 S bL of G bL on the subspace KerbL. It is obvious that All boundary correct extensions (2.1) transforms into L−1f = L−1 S f + Kf = L−1 S f for all f from H, where I is the identity operator in H. In virtue of D(L) ⊂ D(bL), we have S f + KbLL−1 Ker ΓLS = D(LS) and R(ΓLS ) = KerbL. S f = (I + KbL)L−1 bLu = f for all f from H, u from D(L) D(L) =(cid:8)u ∈ D(bL) : (I − KbL)u ∈ D(LS)(cid:9). where It is easy to see that the operator K defines the domain of L, as (see [3]) (I − KbL)D(L) = D(LS), (I + KbL)D(LS) = D(L), (I − KbL) = (I + KbL)−1. Therefore, all boundary correct extensions L are differed from fixed boundary correct exten- sion LS only the domain. The bounded (in G bL) operator I − KbL maps D(L) onto D(LS) in a one-to-one fashion. Then the domain of L can be defined as follows: D(L) =(cid:8)u ∈ D(bL) : ΓLS (I − KbL)u = 0(cid:9). 3 There exists one more representation of the domain of L Similarly we can define D(L) =(cid:8)u ∈ D(bL) : ((I − KbL)u, L∗ D(L∗) =(cid:8)u ∈ D(cM) : ΓL∗ S)(cid:9). Sv) = (bLu, v) for all v from D(L∗ S cM S (I − K ∗cM )u = 0(cid:9). S = I − L∗−1 ΓL∗ Now we can formulate the following result: Theorem 2. Let there be a correct extension LS of the minimal operator L0 with D(LS) = D(L∗ S), then any other correct extension L has the property D(L) = D(L∗) if and only if L0 ⊂ L ⊂bL and the operator K from the formula (2.1) satisfies the conditions and and (2.2) R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM), ( ΓLS (I − KbL)u = 0, ΓLS K ∗cM u = KbLu, S bL is the projection defined above. for all u ∈ D(bL) ∩ D(cM), where ΓLS = I − L−1 Proof. Let D(L) = D(L∗). In view of Assertion 1, the operators LS and L turn out to be boundary correct extensions of L0, i.e., L0 ⊂ LS ⊂ bL and L0 ⊂ L ⊂ bL. The inverse to the arbitrary boundary correct extension L has the form (2.1). Then (L∗)−1g = (L∗ S)−1g + K ∗g for all g ∈ H. The condition D(L) = D(L∗) is equivalent to L−1 S f + Kf = (L∗ S)−1g + K ∗g, (2.3) where for each f ∈ H there exists g ∈ H and vice versa, for each g ∈ H there exists f ∈ H that the equality (2.3) is fulfilled. It follows from (2.3) that Then we get R(K ∗) ⊂ D(bL) and R(K) ⊂ D(cM ). R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM). Acting on both sides of equality (2.3) by the operator bL, we obtain f = LS(L∗ g ∈ H. for all Substituting f into (2.3), we obtain the equality S)−1g +bLK ∗g, It follows that This means that L−1 S bLK ∗g + KLS(L∗ S)−1g + KbLK ∗g = K ∗g. (I − L−1 S)−1 + K ∗)g. S bL)K ∗g = KbL((L∗ S bL)K ∗g = KbL(L∗)−1g. (I − L−1 4 If L∗−1 g is replaced by u, then (I − L−1 S bL)K ∗cM u = KbLu, u ∈ D(L∗). to the condition (2.2). Since D(L) = D(L∗) we obtain ΓLS K ∗cM u = KbLu for all u from D(L). This is equivalent We now prove a converse of this theorem. Let L0 ⊂ L ⊂bL and the operator K from the formula (2.1) satisfies the conditions R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM), and (2.2). Hence, it is easy to see that Since Lu = f for all u ∈ D(L), we may replace Lu by f in the second equation of the condition (2.2). Then D(L) ∪ D(L∗) ⊂ D(bL) ∩ D(cM). ΓLS K ∗cM L−1f = Kf for all f ∈ H. Acting on both sides of this equality by the projection ΓL∗ S , we obtain Note that Adding the bounded operator L−1 for all f ∈ H. K ∗cM L−1f = (I − (L∗ S)−1cM )Kf S f + K ∗cM Kf + (L∗ SL−1 S f to both sides, we get K ∗L∗ S)−1cM Kf = Kf. (L∗ S)−1L∗ SL−1 S f + K ∗L∗ SL−1 S f + K ∗cM Kf + (L∗ S)−1cM Kf = Kf + L−1 S f. It follows that If we denote by then we have (L∗ S)−1(L∗ SL−1 S +cM K)f + K ∗(L∗ SL−1 S +cM K)f = L−1f for all f ∈ H, g = L∗ SL−1 S f +cM Kf (L∗)−1g = L−1f for all f ∈ H. for all f ∈ H. It follows that D(L) ⊂ D(L∗). Acting on both sides of the equations (2.2) by the projection ΓL∗ S , we get By the second equation of the given system, we can rewrite this system of equations in the form ( ΓL∗ ( ΓL∗ ΓL∗ S ΓL∗ S (I − KbL)u = 0, S KbLu = K ∗cM u for all u ∈ D(bL) ∩ D(cM). S (I − K ∗cM)u = 0, KbLu = K ∗cM u for all u ∈ D(bL) ∩ D(cM). S KbL(L∗)−1g = K ∗cM (L∗)−1g for all g ∈ H. KbL(L∗)−1g = (I − (LS)−1bL)K ∗g for all g ∈ H. ΓL∗ 5 The first equation of this system means that u belongs to D(L∗). Then we denote L∗u = g. Therefore, u = (L∗)−1g for all g from H. Then the second equation of this system has the form Acting on both sides of this equality by the projection ΓLS , we obtain Note that Adding the bounded operator (L∗ L−1 S LS(L∗ S)−1g + KLS(L∗ It follows that KLS(L∗ S)−1g to both sides, we get S)−1g + KbLK ∗g + L−1 S)−1g + KbLK ∗g + L−1 S bLK ∗g = K ∗g. S bLK ∗g = K ∗g + (L∗ S)−1g. for all g ∈ H. L−1 S (LS(L∗ S)−1 +bLK ∗)g + K(LS(L∗ S)−1 +bLK ∗)g = (L∗)−1g f = (LS(L∗ for all g ∈ H, S)−1 +bLK ∗)g L−1f = (L∗)−1g for all g ∈ H. If we denote by then we have It follows that D(L∗) ⊂ D(L). The theorem is proved. 3 Normality criterion of correct extensions 0 = cM to D(L0). Then bL = M ∗ Let L0 be a formally normal minimal operator in a Hilbert space H. An operator M0 is the restriction of L∗ 0 defines the maximal operator that boundary correct extension. Then the inverses to all boundary correct extensions L of L0 have the form L0 ⊂ bL. Let there be at least one normal correct extension LN of the formally normal minimal operator L0. In view of Assertion 1, we have that L0 ⊂ LN ⊂ bL, i.e., LN is the where K is an arbitrary bounded operator in a Hilbert space H that R(K) ⊂ KerbL and R(L0) ⊂ Ker K. Then the direct operator L acts as u = L−1f = L−1 for all f ∈ H, N f + Kf (3.1) on the domain where the projection ΓLN = I − L−1 that for all f ∈ H, bLu = f N bL is the bounded operator in the space G bL. It is known D(L) =(cid:8)u ∈ D(bL) : ΓLN (I − KbL)u = 0(cid:9), Ker ΓLN = D(LN ) and R(ΓLN ) = KerbL. Theorem 3. Let there be one correct normal extension LN of the formally normal minimal operator L0 in a Hilbert space H. Then any other correct extension L of L0 is normal if and only if L0 ⊂ L ⊂bL and operator K from the formula (3.1) satisfies the conditions: R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM), ( ΓLN (I − KbL)u = 0, ΓLN K ∗cM u = KbLu for all u ∈ D(bL) ∩ D(cM), bLK ∗ = (cM K)∗, N bL is projection on KerbL. 6 and where ΓLN = I − L−1 (3.2) (3.3) Proof. Let L be a normal correct extension of the formally normal operator L0. In view of Theorem 2, the conditions L0 ⊂ L ⊂ bL, R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM ) and (3.2) will be fulfilled. The normality of L−1 follows from the normality of L: L−1(L∗)−1 = (L∗)−1L−1. By virtue of (3.1), we obtain (L−1 N + K)((L∗ N )−1 + K ∗)f = ((L∗ N )−1 + K ∗)(L−1 N + K)f for all f ∈ H. It follows that L−1 N K ∗f + K(L∗ N )−1 + KK ∗f = (L∗ N )−1Kf + K ∗L−1 N f + K ∗Kf. (3.4) K ∗f = LN (L∗ Acting on both sides of the equality (3.4) by the operator bL, we get N )−1Kf +bLK ∗L−1 N f +bLK ∗Kf. N )−1(bLK ∗)∗f + K ∗(bLK ∗)∗f Taking conjugates of both sides of the equality above, we have Kf = K ∗(LN (L∗ N )−1)∗f + (L∗ Acting on both sides by the operator cM , we obtain bLK ∗ = (cM K)∗. cM Kf = (bLK ∗)∗f This is equivalent to for all f ∈ H. for all f ∈ H. Let us prove the converse. Suppose that the conditions of Theorem 3 hold. From the conditions L0 ⊂ L ⊂bL, R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM ) and (3.2), in view of Theorem 2, we have that D(L) = D(L∗). Then for all f ∈ H there exists g ∈ H such that L−1f = (L∗)−1g. It can be rewritten in the form L−1 N f + Kf = (L∗ N )−1g + K ∗g. Acting on both sides by the operator cM , we get N L−1 g = L∗ Substituting g into (3.5), we have Then Let us show that Kf = (L∗ N )−1cM Kf + K ∗L∗ K ∗f = (cM K)∗L−1 N f + (L∗ N L−1 N L−1 N f +cM Kf. N f + K ∗cM Kf N )∗Kf + (cM K)∗Kf (L∗ N L−1 N )∗ = LN (L∗ N )−1. (3.5) for all f ∈ H. for all f ∈ H. (3.6) It is known that if A is a closed operator, B is bounded in H and AB is densely defined in H, then where the overbar denotes the closure operator. Note that (AB)∗ = B ∗A∗, N L−1 L∗ N ⊃ L−1 N L∗ N . 7 Then (L∗ N L−1 N )∗ = (L∗ N )−1LN ⊂ LN (L∗ N )−1. Taking into account the fact that LN (L∗ (L∗ N )−1LN on the dense set D(LN ), then we obtain that N )−1 is the bounded operator that coincides with LN (L∗ N )−1 = (L∗ N )−1LN = (L∗ N L−1 N )∗. Then, taking into account (3.3), the equality (3.6) can be rewritten in the form N f + LN (L∗ K ∗f =bLK ∗L−1 Adding (L∗ N )−1f to both sides of the last equality, we get for all f ∈ H. K ∗f + (L∗ N )−1f = LN L−1 N (L∗ N f + LN (L∗ N )−1Kf +bLK ∗Kf N )−1f +bLK ∗L−1 N )−1Kf +bLK ∗Kf. It follows that Thus The proof is complete. (L∗)−1f = L(L∗)−1L−1f for all f ∈ H. L−1(L∗)−1f = (L∗)−1L−1f for all f ∈ H. operator L−1 S takes part in the explicit form. Sometimes there exists another operator TLS The domain of LS described as the kernel of the projection ΓLS = I − L−1 S bL. Here the defined on D(bL) and has the property Ker ΓLS = Ker TLS . Between these operators have the following relationship If we know TLS v, then ΓLS v is uniquely determined as the solution of the homogeneous TLS ΓLS v = TLS (I − L−1 S bL)v = TLS v − TLS L−1 equation bL(ΓLS v) = 0 with an inhomogeneous condition TLS (ΓLS v) = TLS v. S bLv = TLS v for all v ∈ D(bL). Its unique solvability follows from the correctness of the operator LS. Therefore, it is not necessary to know the explicit form of the operator L−1 S . In the study of differential operators (see [10]) that the operator TLS is realized in the form of the boundary operator. In such cases we say that the domain is described in terms of the boundary operator. For example, in the case of the Dirichlet problem for a differential equation of elliptic type in L2(Ω) that TLS corresponds to the trace operator on the boundary of Ω, i.e., TLS u = u ∂Ω. Therefore it is sufficient to know the form of the boundary operator TLS . Thus we obtain the following Corollary 4. Let there be a correct extension LS of the minimal operator L0 with D(LS) = D(L∗ S), then any other correct extension L has the property D(L) = D(L∗) if and only if L0 ⊂ L ⊂bL, R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM ) and TLS (K ∗cM − KbL)u = 0 D(L) =(cid:8)u ∈ D(bL) : TLS (I − KbL)u = 0(cid:9). for all u ∈ D(L), where TLS is a boundary operator corresponding to the fixed correct extension LS and (3.7) 8 Remark 1. By virtue of the one-to-one mapping of D(LS) onto D(L) : v = (I − KbL)u for all u ∈ D(L), u = (I + KbL)v for all v ∈ D(LS), in practice, sometimes it is more convenient to use the following condition that is equivalent to (3.7): It has the practical convenience because D(LS) is a fixed domain. TLS (K ∗cM − KbL + K ∗cM KbL)v = 0 for all v ∈ D(LS). Similarly, we can rephrase Theorem 3 in the following form (3.8) Corollary 5. Let there be one correct normal extension LN of the formally normal minimal operator L0 in a Hilbert space H. Then any other correct extension L of L0 is normal if and only if L0 ⊂ L ⊂bL, R(K) ∪ R(K ∗) ⊂ D(bL) ∩ D(cM), for all u ∈ D(L), and where TLN is a boundary operator corresponding to the fixed correct extension LN and TLN (K ∗cM − KbL)u = 0 bLK ∗ = (cM K)∗, D(L) =(cid:8)u ∈ D(bL) : TLN (I − KbL)u = 0(cid:9), (3.9) (3.10) (4.1) and K is the operator determining the boundary correct extension L from the formula (3.1). to which corresponds the minimal operator L0 with domain D(L0) =(cid:8)y ∈ W 2 4 The Examples Example 1. We consider the following operator in a Hilbert space L2(0, 1) bLy ≡ y ′′ + y ′ = f, 2 (0, 1) : y(0) = y(1) = y ′(0) = y ′(1) = 0(cid:9). cM y ≡ y ′′ − y ′ = f. 0 and L∗ the operator M0 has the form We define the operator M0 as the restriction of cM on the set D(L0). Then the action of 0 by bL and cM , respectively. Then we have We will denote the maximal operators M ∗ Let the operator LN acts as bL with domain 2 (0, 1). L0 ⊂bL, M0 ⊂ cM and D(bL) = D(cM ) = W 2 D(LN ) =(cid:8)y ∈ D(bL) : y(0) + y(1) = 0, y ′(0) + y ′(1) = 0(cid:9). N ⊂ cM . The inverse operator to LN has the form (1 − et−x)f (t)dt − f (t)dt + et−1f (t)dt. and L0 ⊂ LN ⊂bL, M0 ⊂ L∗ xZ0 y = L−1 N f = e1−x 1 + e 1Z0 We take the operator LN as the fixed correct extensions of L0. Note that D(LN ) = D(L∗ N ) 1 2 1Z0 9 Then ΓLN is defined as And ΓL∗ N = I − (L∗ − 2 ΓLN y = y(0) + y(1) +(cid:18)1 N )−1cM has the following form +(cid:18) ex y(0) + y(1) N y = ΓL∗ 2 2 1 + e e1−x 1 + e(cid:19)[y ′(0) + y ′(1)]. 2(cid:19)[y ′(0) + y ′(1)]. − 1 The correct extension L of L0 with the property D(L) = D(L∗) is a boundary correct extension. Their inverses are described in the following form y = L−1f = L−1 N f + Kf for all f ∈ L2(0, 1), where K is a bounded linear operator in L2(0, 1) with the properties In our case, such operators are exhausted by the following operators R(K) ⊂ KerbL, R(L0) ⊂ Ker K. f (t)(a11 + a12et)dt + e−x f (t)(a21 + a22et)dt, Kf = 1Z0 1Z0 where aij, i, j = 1, 2 are arbitrary complex numbers. Then It is known that the direct operator L acts as bL from (4.1) and the domain has the form In view of Corollary 4, the domain of L can be defined in another way e−tf (t)dt. 1Z0 K ∗f = (a11 + a12ex) f (t)dt + (a21 + a22ex) 1Z0 D(L) =(cid:8)y ∈ D(bL) : ΓLN (I − KbL)y = 0(cid:9). D(L) =(cid:26)y ∈ D(bL) : y(0) + y(1) = (KbLy)(0) + (KbLy)(1), KbLy(cid:17)(0) +(cid:16) d y ′(0) + y ′(1) =(cid:16) d dx dx First, we will find the correct extensions L such that D(L) = D(L∗). Taking into account Remark 1, let the operator K satisfies the condition (3.8). Then we obtain the system of equations: KbLy(cid:17)(1)(cid:27). −4(a11 − a11) − 2(e + 1)(a12 − a12) − 2 (a21 − a21)  4(a11 + a11) + 2(e + 1)h a21 e (e + 1)2 − e + 1 e + a12i · A = 0, (a22 − a22) +h4a12 + 2 a12ha21(e − 1) + a22 i − 2 e ha21 + a22 a22 e + 1 e + 1 2 e e2 − 1 e2 2 2 e 10 a21 + a12 + − − 1 e 1 e [2a21 + a22(1 + e)] − 2a12 − e2 − 1 a12 −4 e e + 1 e a22i · A = 0, i = 0, a22ha21(e − 1) + a22 e2 − 1 2 i = 0, where A = 2(e − 1)a11 + (e2 − 1)a12 + e + 1 e ha21(e − 1) + a22 e2 − 1 2 i. Solutions of the system of equations with respect to aij, i, j = 1, 2, define the operators K that guarantees the equality D(L) = D(L∗). They will correspond to the following cases: y(1) = 0o, I) D(L) =ny ∈ D(bL) : y(0) = 0, II) D(L) =ny ∈ D(bL) : y(0) = III) D(L) =ny ∈ D(bL) : ay(0) + ¯by(1) = 0, where R is the space of real numberso, a − i a + i a ∈ R, a 6= 0, b ∈ C, y(1), y ′(0) = a − i a + i y ′(1), a ∈ R, y(1) = by ′(0) + ay ′(1), b2 = a2, where C is the space of complex numberso. We use the criterion given in Theorem 3 to find all correct normal extensions L of the minimal operator L0. It is easy to verify the formal normality of L0 and the normality of LN . The equality D(L) = D(L∗) is necessary for the normality of L. They correspond to three cases of I) − III) described above. Now, if the operator K satisfies (3.3), then the operator L is a normal. The condition (3.3) is equivalent to the following Therefore, the operator K takes the form a21 = 0, a12 = 0. Kf = a11 1Z0 f (t)dt + a22e−x 1Z0 etf (t)dt. Then operators L which act as bL from (4.1) turn out to be the normal correct extensions and with the domain D(L) =ny ∈ D(bL) : y(0) = a − i a + i y(1), y ′(0) = y ′(1), a − i a + i a ∈ Ro. From three cases of I) − III) are suitable only the case II). Example 2. Let in the Hilbert space L2(Ω), where Ω = {(x, y) : 0 < x < 1, 0 < y < 1}, we consider the minimal operator L0 generated by the Cauchy-Riemann differential operator ∂u ∂x + i ∂u ∂y = f (x, y). (4.2) bLu ≡ Then where TL0 is a boundary operator defined as the trace of function u ∈ W 1 of ∂Ω. 2 (Ω) on the boundary D(L0) =(cid:8)u ∈ W 1 2 (Ω) : TL0u = 0(cid:9), The action of cM will have the form cM u ≡ − ∂u ∂x + i ∂u ∂y = f (x, y). 11 respectively. If we define the boundary operator TLN the following way Domains of the operators bL and cM have the form D(bL) =(cid:8)u ∈ L2(Ω) : bLu ∈ L2(Ω)(cid:9), D(cM ) =(cid:8)u ∈ L2(Ω) : cM u ∈ L2(Ω)(cid:9), u(x, 0) + u(x, 1) (cid:19) for all u ∈ D(bL), D(LN ) =(cid:8)u ∈ D(bL) : TLN u = 0(cid:9), TLN u =(cid:18) u(0, y) + u(1, y) then the operator LN acting as bL with the domain normal, and in addition L0 ⊂ L ⊂bL. is the correct extension of L0. It is easy to verify that L0 is formally normal and LN is We are interested in the normal boundary correct extensions. Let us clarify some prop- erties of the operator K: 2 (Ω); 1) R(K) ⊂ W 1 2) (Kf )(x + iy); 3) (K ∗f )(x − iy). The first property follows from the fact that Assertion 6. The domain of any normal correct extension L of the minimal operator L0 generated by the differential operator (4.2) has the property: D(L) ⊂ W 1 2 (Ω). Proof. It follows from Theorem 2 of Plesner and Rohlin (see [8]). Now we formulate this theorem: "For each pair of adjoint normal operators A and A∗ there exists one and only one pair of self-adjoint operators A1 and A2, satisfying the condition A = A1 + iA2, A∗ = A1 − iA2, where the operators A1 and A2 commute". The second property follows from the condition R(K) ⊂ KerbL. The third property follows from the condition R(L0) ⊂ Ker K. Further from the conditions (3.9) and (3.10) obtain the operators K for which the correct boundary extension L will be normal. It follows from Assertion 6 that L−1 N , K, and L−1 are compact operators in L2(Ω). This means that the normal correct extension L of L0 is the operator of the discrete spectrum. Hence we have that L has a complete orthonormal system of eigenfunctions. For clarity, the check of normality by Theorem 3, we consider the special case. Let K will be an integral operator of the form Kf = 1Z0 1Z0 K(x, y; ξ, η)f (ξ, η)dξdη. It follows from properties 1) and 2) that Kf = 1Z0 1Z0 K(x + iy, ξ + iη)f (ξ, η)dξdη. 12 From the condition (3.3) of Theorem 3, we get that Kf = 1Z0 1Z0 K(x − ξ + i(y − η))f (ξ, η)dξdη. Using the condition (3.2) of Theorem 3 for the operator K, we obtain all normal correct extensions. We will not give this condition on the kernel K(x − ξ + i(y − η)), because of the cumbersome to write. To demonstrate the mechanism of checking the condition (3.2), we consider the special case when K(x − ξ + i(y − η)) = aeiπ(x−ξ+i(y−η)), where a ∈ C is a complex number of the form a = a1 + ia2. Then the condition (3.2) is equivalent to 2a2 + (a2 1 + a2 2)(eπ − e−π) = 0. There are two kinds of solutions of this equation: I. a1 = 0, a2 = 2 e−π − eπ ; II. a2 = −1 ±p1 − [a1(eπ − e−π)]2 eπ − e−π , where a1 ≤ 1 eπ − e−π . Then in the case of II, the correct extension corresponding to the following boundary prob- lem ∂u ∂y ∂u ∂x + i bLu ≡ D(L) =(cid:26)u ∈ W 1 = f (x, y) for all f ∈ L2(Ω), 2 (Ω) : u(0, y) + u(1, y) = 0, 0 ≤ y ≤ 1, u(x, 0) + u(x, 1) = ia(eπ + 1) 1Z0 eiπ(x−ξ)u(ξ, 1)dξ − ia(e−π + 1) 1Z0 eiπ(x−ξ)u(ξ, 0)dξ, 0 ≤ x ≤ 1(cid:27) is normal, where a = a1 + ia2, or in the case of I, the correct extension corresponding to the boundary problem D(L) =(cid:26)u ∈ W 1 2 (Ω) : u(0, y) + u(1, y) = 0, u(x, 0) + u(x, 1) = 2 is normal. 1Z0 eiπ(x−ξ)u(ξ, 1)dξ(cid:27), All normal correct extensions L have a compact inverse operator because of D(L) ⊂ 2 (Ω). Therefore, their eigenfunctions create an orthonormal basis in L2(Ω). In the partic- W 1 ular case when K(x, y; ξ, η) = 2i e−π − eπ · eiπ(x−ξ+i(y−η)), 13 we obtain the orthonormal basis in the following form: and the corresponding eigenvalues uk,n(x, y) =(cid:26) e2nπiy+iπx, λk,n =(cid:26) iπ − 2nπ, e(2k+1)πix+(2n+1)πiy, (2k + 1)πi − (2n + 1)π, n = 0, ±1, ±2, . . . k = ±1, ±2, . . . , n = 0, ±1, ±2, . . . n = 0, ±1, ±2, . . . k = ±1, ±2, . . . , n = 0, ±1, ±2, . . . . Thus, this method allows us to check for normality of an unbounded operator. Prelimi- nary it is necessary to clarify the question of the existence of at least one normal extension. For the existence of a normal extension we need that the minimal operator must be formally normal. Remark 2. If in Example 2 the square area Ω is replaced by the unit circle, then the minimal operator L0 will not be formally normal. Thus in this case, there are no normal extensions of L0 in L2(Ω). Remark 3. When the minimal operator L0 is symmetric and the fixed operator LN is self- adjoint then the conditions of Theorem 3 are equivalent to K = K ∗ and we have all the self-adjoint correct extensions. 14 References [1] G. Biriuk, E. A Coddington, Normal extensions of unbounded formally normal opera- tors, J. of Math. and Mech. 13(4), (1964), 617 -- 634. [2] B. N. Biyarov, On the spectrum of well-defined restrictions and extensions for the Laplase operator, Mat. Zametki, 95, no. 4, (2014), 507 -- 516 (in Russian). English transl. Math. Notes, 95, no. 4, (2014), 463-470. [3] B. N. Biyarov, Spectral properties of correct restrictions and extensions of the Sturm- Liouville operator, Differ. Equations 30(12), (1994), 1863-1868. (Translated from Differ. Uravn. 30(12), (1994), 2027-2032.) [4] B. N. Biyarov, S. A. Dzhumabaev, A criterion for the Volterra property of boundary value problems for Sturm-Liouville equations, Mat. Zametki, 56, no. 1, (1994), 143 -- 146 (in Russian). English transl. in Math. Notes, 56, no. 1, (1994), 751 -- 753. [5] E. A. Coddington, Formally normal operators having no normal extensions, Can., J. Math. 17, (1965), 1030 -- 1040. [6] E. A. Coddington, Normal extensions of formally normal operators, Pacific J. Math. 10, (1960), 1203 -- 1209. [7] B. K. Kokebaev, M. Otelbaev, and A. N. Shynibekov, On questions of extension and restriction of operators, Sov. Math., Dokl. 28, (1983), 259 -- 262. (Translated from Dokl. Akad. Nauk SSSR 271(6), (1983), 1307 -- 1310.) [8] A. I. Plesner and V. A. Rohlin, Spectral theory of linear operators. II, Usp. Mat. Nauk. 1(1), (1946), 71 -- 196 (in Russian); English transl., Amer. Math. Soc. Transl. II. Ser. 62, (1967), 29 -- 175. [9] V. N. Polyakov, A class of formally normal operators, Math. Notes 2(6), (1967), 859 -- 863. (Translated from Mat. Zametki 2(6), (1967), 605 -- 614.) [10] M. I. Vishik, On general boundary problems for elliptic differential equations, Tr. Mosk. Matem. Obs. 1, (1952), 187 -- 246 (in Russian); English transl., Am. Math. Soc., Transl., II, Ser. 24, (1963), 107 -- 172. Bazarkan Nuroldinovich Biyarov Department of Fundamental Mathematics L.N. Gumilyov Eurasian National University 2 Mirzoyan St, 010008 Astana, Kazakhstan E-mail: [email protected] 15
1401.4630	2	1401	2015-02-09T14:02:00	Spectra of Cantor measures	[ "math.FA" ]	Let $\mu_{q, b}$ be the Cantor measure associated with the iterated function system $f_i(x)=x/b+i/q, 0\le i\le q-1$, where $2\le q, b/q\in \Z$. In this paper, we consider spectra and maximal orthogonal sets of the Cantor measure $\mu_{q, b}$ and their rational rescaling. We introduce a quantity to measure level difference between a branch and its subbranch for the labeling tree corresponding to a maximal orthogonal set of the Cantor measure $\mu_{q, b}$, and use certain boundedness property of that quantity as sufficient and necessary conditions for a maximal orthogonal set of the Cantor measure $\mu_{q, b}$ to be its spectrum. We show that the integrally rescaled set $K\Lambda$ is still a spectrum if it is a maximal orthogonal set, and we provide a simple characterization for the integrally rescaled set to be a maximal orthogonal set. As an application of the above characterization, we find all integers $K$ such that $K\Lambda_4$ are spectra of the Cantor measure $\mu_{2, 4}$, where $\Lambda_4:=\{\sum_{n=0}^\infty d_n 4^n: d_n\in \{0, 1\}\}\subset \Z$ is the first known spectrum for the Cantor measure $\mu_{2, 4}$. Finally we discuss rescaling spectra rationally and construct a spectrum $\Lambda$ for the Cantor measure $\mu_{q, b}$ such that $\Lambda/(b-1)$ is a maximal orthogonal set but not a spectrum.	math.FA	math	SPECTRA OF CANTOR MEASURES XIN-RONG DAI Abstract. A spectrum of a probability measure µ is a countable set Λ such that {exp(−2πiλ·), λ ∈ Λ} is an orthogonal basis for L2(µ). In this paper, we consider the problem when a countable set become the spectrum of the Cantor measure. Starting from tree labeling of a maximal orthogonal set, we introduce a new quantity to measure minimal level difference between a branch of the labeling tree and its subbranches. Then we use boundedness and linear increment of that level difference measurement to justify whether a given maximal orthogonal set is a spectrum or not. This together with the tree labeling of a maximal orthogonal set provides fine structures of spectra of Cantor measures. As applications of our justification, we provide a characterization for the integrally expanding set KΛ of a spectrum Λ to be a spectrum again, thereby we find all integers K such that KΛ4 n=0 dn4n : dn ∈ {0, 1}} is the first known spectrum for µ4. Furthermore, we construct a spectrum Λ such that the integrally shrinking set Λ/K is a maximal orthogonal set but not a spectrum for some integer K. are spectra of the 1/4-Cantor measure µ4, where Λ4 := {P∞ 1. Introduction A fundamental problem in harmonic analysis is whether {exp(−2πiλx), λ ∈ Λ} is an orthogonal basis of L2(µ), the space of all square-integrable functions with respect to a probability measure µ. The above probability measure µ is known as a spectral measure and the countable set Λ as its spectrum. Spectral theory for the Lebesgue measures on sets has been studied extensively since it initialed by Fuglede 1974 [11], see [10, 16, 24] and references therein. Recently, He, Lai and Lau [12] proved that a spectral measure is pure type (i.e. either absolutely continuous or singular continuous or counting measure). For singular continu- ous measures, the first spectral measure was found by Jorgenson and Pederson in n=0 dn4n : dn ∈ {0, 1}} is a spectrum of the Bernoulli convolution µ4. Since then, some significant progresses have been made and various new phenomena different from spectral theory for the Lebesgue mea- sure have been discovered [1 -- 9,13 -- 15,23]. For instance, Fourier frames on the unit interval [0, 1) have Beurling dimension one [17], while spectra of a singular mea- sure could have zero Beurling dimension [2]. Here we define the Cantor measure 1998 [14], they proved that Λ4 := {P∞ 2010 Mathematics Subject Classification. primary 42A65, 42B05, secondary 28A78, 28A80. Key words and phrases. Spectral measure, Cantor measure, Bernoulli convolution, spectrum, maximal orthogonal set. The research is partially supported by the NSF of China (No. 11371383) and the Fundamental Research Funds for the Central Universities. 1 µq,b with 2 ≤ q ∈ Z and q < b ∈ R, (1.1) µq,b = 1 q q−1Xi=0 µq,b(cid:0) f −1 i (·)(cid:1), is a self-similar probability measure associated with the iterated function system, fi(x) = x/b + i/q, i = 0, 1, . . . , q − 1. And we call the special case µb := µ2,b the Bernoulli convolutions. In 1998, Jorgen- son and Pederson proved in their seminal paper [14] that Bernoulli convolutions µb with b ∈ 2Z are spectral measures. The converse problem stood for a long time and it was solved in [1] by the author in 2012 after important contributions by Hu and Lau [13]. The complete characterization for Bernoulli convolutions in [1] was recently extended by He, Lau and the author [4] to the Cantor measure µq,b that it is a spectral measure if and only if (1.2) 2 ≤ q, b/q ∈ Z. The Cantor measures µq,b with q and b satisfying (1.2) are few of known singular spectral measures, but the structure of their spectra is little known, even for the Bernoulli convolution µ4. In this paper, we explore fine structure of spectra of these Cantor measures. Our exploration starts from tree structure of a (maximal) orthogonal set Λ, meaning that {exp(−2πiλx), λ ∈ Λ} is a (maximal) orthogonal set of L2(µq,b). In 2009, Dutkay, Han and Sun gave a complete characterization of the maximal or- thogonal sets of the Bernoulli convolution µ4 by introducing a tree labeling tool [6]. Recently, He, Lai and the author developed a tree labeling technique for Cantor measures µq,b [2]. They proved that a countable set is a maximal orthogonal set of the Cantor measure µq,b if and only if it can be labeled as a maximal tree, see Theorem 2.2. Thus maximal orthogonal sets have tree structure and they can be built selecting maximal tree appropriately. While a maximal orthogonal set is not necessarily a spectrum since it may lack of completeness in L2(µq,b). The com- pleteness of maximal orthogonal sets for Cantor measures µq,b is quite challenging, see [2, 4, 6 -- 8, 14, 23] for various sufficient and necessary conditions. In fact, the completeness of exponential sets is a classical problem in Fourier analysis since 1930s', see [18 -- 22, 25] and references therein for historical remarks and recent advances. The main contribution of this paper is to introduce a quantity Dτ,δ to mea- sure minimal level difference between a branch δ of the labeling tree and its sub- branches, see Definition 2.3. We show in Theorem 2.4 that a maximal orthogonal set Λ with maximal tree labeling τ is a spectrum if Dτ,δ is uniform bounded on all tree branches δ, and also in Theorem 2.5 that it is not a spectrum if Dτ,δ increases linearly to the level of the tree branches δ. Unlike spectra of the Lebesgue measure on the unit interval, a spectrum Λ of a singular measure could have the integrally rescaled set KΛ being its spectrum too, see [7, 8, 14] for the Bernoulli convolution µ4. In this paper, we apply our 2 completeness results in Theorems 2.4 and 2.5 to characterize the spectral property of the rescaled set KΛ for a given spectrum Λ of the Cantor measure µq,b via no repetend of K for the labeling tree of Λ. As corollaries, we find all integers K such that KΛ4 are spectra of the Bernoulli convolution µ4, see Corollary 2.7, and we construct a spectrum Λ of the Cantor measure µq,b such that the rescaled set Λ/(b − 1) is its maximal orthogonal set but not its spectrum, see Theorems 2.6, 5.1 and 5.2. This paper is organized as follows. In Section 2, we recall some preliminaries about (maximal) orthogonal sets for Cantor measures, and state our main results. In Sections 3, we consider the problem when a maximal orthogonal set is a spec- trum. In Section 4, we discuss the necessity for a maximal orthogonal set to be a spectrum. In Section 5, we discuss rationally rescaling of a spectrum. 2. Preliminaries and main theorems We start this section from recalling a characterization of orthogonal sets of a probability measure µ via its Fourier transform µ, µ(ξ) :=ZR e−2πiξxdµ(x). expression, is Observe that the zero set of the Fourier transform bµq,b, see (3.6) for its explicit Zq,b = {b ja : a ∈ Z\qZ, 0 ≤ j ∈ Z}. Then a discrete set Λ is an orthogonal set of µq,b if and only if we have the following for orthogonal sets of the Cantor measure µq,b [2, 4]: (2.1) As orthogonal sets (maximal orthogonal sets and spectra) are invariant under trans- lations, in this paper we always normalize them by assuming that Λ − Λ ⊂ Zq,b ∪ {0}. (2.2) 0 ∈ Λ ⊂ Z. To introduce the tree structure of the maximal orthogonal set of the Cantor mea- sure µq,b, we need some notation and concepts. Denote Σq := {0, · · · , q − 1}, and Σn Σn q. Given q := Σq × · · · × Σq , 1 ≤ n ≤ ∞ be the n copies of Σq, and Σ∗ q :=S1≤n<∞ δ = δ1δ2 · · · ∈ Σ∗ and adopt the notation 0∞ = 000 · · · , 0k = 0 · · · 0 q, we define δ′δ is the concatenation of δ′ and δ, q ∪ Σ∞ q . We call an element in Σ∗ q and δ′ ∈ Σ∗ q ∪ Σ∞ {z } n as a tree branch. For each tree branch δ = δ1δ2 · · · , denote q , and when δ ∈ Σ∗ q, δk :=( δ1δ2 · · · δk when δ ∈ Σ∞ (δ0∞)k {z}k for all k ≥ 1. Definition 2.1. For 2 ≤ q, b/q ∈ Z, we say that a mapping τ : Σ∗ 2} is a tree mapping if q → {−1, 0, . . . , b − 3 (i) τ(0n) = 0 for all n ≥ 1, and (ii) τ(δ) ∈ δn + qZ if δ = δ1 · · · δn ∈ Σn q, n ≥ 1, and that a tree mapping τ is maximal if (iii) for any δ ∈ Σ∗ q there exists δ′ ∈ Σ∗ q such that τ((δδ′)n) = 0 for sufficiently large integers n. In [2], He, Lai and the author established the following characterization for a maximal orthogonal set of the Cantor measure µq,b via some maximal tree map- ping. Theorem 2.2. ( [2]) Let 2 ≤ q, b/q ∈ Z. Assume that Λ is a countable set of real numbers containing zero. Then Λ is a maximal orthogonal set of the Cantor measure µq,b if and only if there exists a maximal tree mapping τ such that Λ = Λ(τ), where (2.3) Λ(τ) :=n ∞Xn=1 τ(δn)bn−1 : δ ∈ Σ∗ q such that τ(δm) = 0 for sufficiently large mo. Given a maximal tree mapping τ : Σ∗ q → {−1, 0, . . . , b − 2}, we say that δ ∈ Σn q, n ≥ 1, is a τ-regular branch if τ(δm) = 0 for sufficiently large m. Define Πτ,n : Σ∗ q → R, n ≥ 1, by q ∪ Σ∞ (2.4) Πτ,n(δ) = nXk=1 τ(δk)bk−1. One may verify that the restriction of Πτ,n onto Σn a τ-regular tree branch δ ∈ Σ∗ to n = ∞ by taking limit in (2.4), q is one-to-one for any n ≥ 1. For q, we can extend the definition Πτ,n(δ), n ≥ 1, in (2.4) (2.5) Πτ,∞(δ) := ∞Xk=1 τ(δk)bk−1. Applying the above b-nary expression, we conclude that a maximal orthogonal set of the Cantor measure µq,b is the image of Πτ,∞ for some maximal tree mapping τ, Λ(τ) =(cid:8)Πτ,∞(δ) : δ ∈ Σ∗ q are τ-regular branches(cid:9). This together with Theorem 2.2 suggests that various maximal orthogonal sets of the Cantor measure µq,b could be constructed by selecting maximal tree mapping appropriately. Now we introduce a quantity to measure (minimal) level difference between a tree branch and its subbranches, which plays important role in our study of spectral property of Cantor measures. For δ′ ∈ Σ∗ q for some n ≥ 1, define q and δ ∈ Σn (2.6) Dτ,δ(δ′) = #Aδ(δ′) + Xn j∈Bδ(δ′) (n j − n j−1 − 1), where Aδ(δ′) := {m ≥ 1 : τ(δδ′m) , 0}, Bδ(δ′) := {m ≥ 1 : τ(δδ′m) < qZ}, n0 = 0, and {n j} j≥1 is a strictly increasing sequence of positive integers given by {n j : j ≥ 1} = Aδ(δ′), and #E is the cardinality of a set E. 4 Definition 2.3. Let 2 ≤ q, b/q ∈ Z and τ : Σ∗ tree mapping. Define (2.7) Dτ,δ := inf(cid:8)Dτ,δ(δ′) : δ′ ∈ Σ∗ q(cid:9), δ ∈ Σ∗ q. q → {−1, 0, . . . , b − 2} be a maximal q → {−1, 0, . . . , b − 2}, we say that δ ∈ Given a maximal tree mapping τ : Σ∗ Σn q is a q, n ≥ 1, is a a τ-main branch if τ(δm) = 0 for all m > n. Clearly δ ∈ Σ∗ τ-regular branch if and only if either δ is a τ-main branch or δ0k is for some k ≥ 1; and for any δ ∈ Σ∗ q. For any q, one may verify that the quantity Dτ,δ is the minimal distance to its τ-main δ ∈ Σ∗ subbranches, (2.8) q there exists a τ-main subbranch δδ′, where δ′ ∈ Σ∗ Dτ,δ = inf(cid:8)Dτ,δ(δ′) : δδ′ are τ-main branches(cid:9) < ∞. A challenging problem in spectral theory for the Cantor measure µq,b is when a maximal orthogonal set becomes a spectrum [2, 4, 6 -- 8, 14, 23]. Now we present our main results of this paper. In our first main result, a sufficient condition via q, is provided for a maximal orthogonal set of the Cantor boundedness of Dτ,δ, δ ∈ Σ∗ measure µq,b to its spectrum. Theorem 2.4. Let 2 ≤ q, b/q ∈ Z. If τ : Σ∗ mapping such that (2.9) then Λ(τ) in (2.3) is a spectrum of the Cantor measure µq,b. q → {−1, 0, . . . , b − 2} is a maximal tree Dτ := sup{Dτ,δ : δ ∈ Σ∗ q} < ∞, We believe that the boundedness assumption on Dτ,δ, δ ∈ Σ∗ sufficient condition for a maximal orthogonal set to be a spectrum. shown in the next theorem, the above boundedness condition on Dτ,δ, δ ∈ Σ∗ close to be necessary. Theorem 2.5. Let 2 ≤ q, b/q ∈ Z, τ : Σ∗ q → {−1, 0, . . . , b − 2} be a maximal tree mapping. If there exists a positive number ǫ0 such that for each n ≥ 1 and δ = δ1δ2 · · · δn ∈ Σn (2.10) then Λ(τ) in (2.3) is not a spectrum of the Cantor measure µq,b. q with δn , 0, Dτ,δ ≥ ǫ0n, q, is a very weak In fact, as q, is Finally we apply our completeness results in Theorems 2.4 and 2.5 to the rescaling- invariant problem when the rescaled set KΛ is a spectrum of the Cantor measure µq,b if Λ is. This simple and natural way to construct new spectra from known ones is motivated from the conclusion that if K = 5k for some k ≥ 1, then the rescaled set KΛ4 := {Kλ : λ ∈ Λ4} of the spectrum (2.11) Λ4 :=n ∞Xj=0 d j4 j, d j ∈ {0, 1}o of the Bernoulli convolution µ4 is also a spectrum [7,8,14]. In the next theorem, we show that if the maximal tree mapping τ associated with the spectrum Λ satisfies the boundedness assumption (2.9), then the integrally rescaled set KΛ is a spectrum of the Cantor measure µq,b if and only if it is a maximal orthogonal set. 5 Theorem 2.6. Let 2 ≤ q, b/q ∈ Z, τ : Σ∗ q → {−1, 0, . . . , b − 2} be a maximal tree mapping satisfying (2.9), and Λ(τ) be as in (2.3). Then for any integer K being prime with b, KΛ(τ) is a spectrum of the Cantor measure µq,b if and only if it is a maximal orthogonal set. Applying Theorem 2.6, we find all possible integers K such that KΛ4 are spectra of the Bernoulli convolution µ4, c.f. [7, 8, 14]. Corollary 2.7. Let Λ4 be as in (2.11) and K ≥ 3 be an odd integer. Then KΛ4 is a spectrum of the Bernoulli convolution µ4 if and only if there does not exist a positive integer N such that (2.12) K NXj=1 d j4 j−1 ∈ (4N − 1)Z\{0} for some d j ∈ {0, 1}, 1 ≤ j ≤ N. Given a spectral set Λ of the Cantor measure µq,b, its irrational rescaling set rΛ (i.e., r < Q) is not an orthogonal set (and hence not a spectrum) by (2.2). The next question is when a rational rescaling set rΛ is an orthogonal set, or a maximal orthogonal set, or a spectrum. A necessary condition is that rΛ ⊂ Z by (2.2), but unlike integral rescaling discussed in Theorems 2.6 there are lots of interesting problems unsolved yet. In this paper, we apply Theorems 2.4 and 2.5 to construct a spectrum Λ of the Cantor measure µq,b such that the rescaled set Λ/(b − 1) is its maximal orthogonal set but not its spectrum, see Theorem 5.2. 3. Maximal orthogonal sets and spectra: a sufficient condition In this section, we prove Theorem 2.4. For that purpose, we need several techni- cal lemmas on spectra of the Cantor measure µq,b, a crucial lower bound estimate for its Fourier transformbµq,b, and an identity for multi-channel conjugate quadra- ture filters. For an orthogonal set Λ of L2(µq,b) containing zero, let Then QΛ is a real analytic function on the real line with QΛ(0) = 1, and QΛ(ξ) =Xλ∈Λ heλ, e−ξi2 ≤ ke−λk2 = 1, ξ ∈ R, where the equality holds if Λ is a spectrum. The converse is shown to be true in [2, 14]. This provides a characterization for an orthogonal set of the Cantor measure µq,b to be its spectrum. Lemma 3.1. ( [2, 14]) Let 2 ≤ q, b/q ∈ Z, and let QΛ(ξ) be defined by (3.1). Then an orthogonal set Λ of the Cantor measure µq,b is a spectrum if and only if QΛ(ξ) = 1 for all ξ ∈ R. 6 (3.1) QΛ(ξ) :=Xλ∈Λ dµq,b(ξ + λ)2. For the Cantor measure µq,b, taking Fourier transform at both sides of the equa- tion (1.1) leads to the following refinement equation in the Fourier domain: (3.2) where (3.3) dµq,b(ξ) = Hq,b(ξ/b) ·dµq,b(ξ/b), Hq,b(ξ) := 1 q q−1Xl=0 e−2πilbξ/q is a periodic function with the properties that Hq,b(0) = 1, (3.4) and (3.5) Hq,b(ξ) = 0 if and only if bξ ∈ Z\qZ, H′ q,b( j/b) , 0 for all j ∈ Z. Applying (3.2) repeatedly and then taking limit m → ∞, we obtain an explicit expression for the Fourier transform of the Cantor measure µq,b: (3.6) where (3.7) dµq,b(ξ) = Hm(ξ)dµq,b(ξ/bm) = ∞Yj=1 Hq,b(ξ/b j), m ≥ 1, Hm(ξ) := mYj=1 Hq,b(ξ/b j), m ≥ 1. Let 2 ≤ q, b/q ∈ Z. Define (3.8) r0 := inf ξ≤(b−2)/(b−1) bµq,b(ξ) and r1 := inf 1≤ j≤q−1 inf ξ≤(b−2)/(b−1) ξ−1Hq,b(ξ/b + j/b). Then it follows from (3.4), (3.5) and (3.6) that both r0 and r1 are well-defined and positive, (3.9) Set (3.10) r0 > 0 and r1 > 0. Tb =(cid:16) − 1 b − 1 , b − 2 b − 1(cid:17)/(cid:16) − 1 b(b − 1) , b − 2 b(b − 1)(cid:17). For any m ≥ 1 and d j ∈ {−1, 0, . . . , b − 2}, 1 ≤ j ≤ m, with dm , 0, one may verify that (3.11) (cid:16)ξ + mXj=1 d jb j−1(cid:17)b−m ∈ Tb for all ξ ∈(cid:16) − 1 b − 1 , b − 2 b − 1(cid:17). To prove Theorem 2.4, we need the following two lemmas which are related to the lower bound estimates of dµq,b(ξ + λ) for ξ ∈ Tb and λ ∈ Z. 7 λ = PK Lemma 3.2. Let 2 ≤ q, b/q ∈ Z, µq,b be the Cantor measure in (1.1), and let j=1 dn jbn j −1 for some positive integers n j, 1 ≤ j ≤ K, satisfying 0 =: n0 < n1 < . . . < nK, and for some dn j , 1 ≤ j ≤ K, belonging to the set {−1, 1, 2, . . . , b−2}. Then (3.12) dµq,b(ξ + λ) ≥ rK+1 0 (cid:16) r1 b(b − 1)(cid:17)#B b−P j∈B(n j−n j−1−1), ξ ∈ Tb, where B = {1 ≤ j ≤ K : dn j < qZ} and r0, r1 are given in (3.8). Proof. For 0 ≤ i ≤ K, define ξ0 = ξ and ξi = (ξ +Pi j=1 dn j bn j−1)/bni for 1 ≤ i ≤ K. ξi ∈ Tb for all 0 ≤ i ≤ K Then (3.13) by (3.11). Observe that (3.14) Hq,b(η) ≤ 1 for all η ∈ R and sup bη∈Tb Hq,b(η) < 1. The above observation, together with (3.6), (3.13) and the fact that Hq,b has period q/b, implies niYℓ=ni−1+1(cid:12)(cid:12)(cid:12)Hq,b(cid:0)(ξ + λ)/bℓ(cid:1)(cid:12)(cid:12)(cid:12) = = if dni ∈ qZ; and i−1Xj=1 dn j bn j−1 + dnibni−1(cid:17)b−ℓ(cid:17)(cid:12)(cid:12)(cid:12)(cid:12) niYℓ=ni−1+1(cid:12)(cid:12)(cid:12)(cid:12)Hq,b(cid:16)(cid:16)ξ + ni−ni−1Yℓ′=1 (cid:12)(cid:12)(cid:12)Hq,b(cid:0)ξi−1/bℓ′(cid:1)(cid:12)(cid:12)(cid:12) ≥ dµq,b(ξi−1) ≥ r0 niYℓ=ni−1+1(cid:12)(cid:12)(cid:12)Hq,b(cid:0)(ξ + λ)/bℓ(cid:1)(cid:12)(cid:12)(cid:12) ni−1Yℓ=ni−1+1(cid:12)(cid:12)(cid:12)(cid:12)Hq,b(cid:16)(cid:16)ξ + i−1Xj=1 = (cid:16) ≥ dµq,b(cid:0)ξi−1) ·(cid:12)(cid:12)(cid:12)Hq,b(ξi−1/bni−ni−1 + dni /b)(cid:12)(cid:12)(cid:12) ≥ r0r1ξi−1/bni−ni−1−1 dn jbn j−1(cid:17)b−ℓ(cid:17)(cid:12)(cid:12)(cid:12)(cid:12)(cid:17) ·(cid:12)(cid:12)(cid:12)(cid:12)Hq,b(cid:16)(cid:16)ξ + ≥ r0r1b−ni +ni−1 /(b − 1) i−1Xj=1 if dni < qZ. Combining the above two lower bound estimates with dn jbn j−1 + dnibni−1(cid:17)b−ni(cid:17)(cid:12)(cid:12)(cid:12)(cid:12) (3.15) proves (3.12). dµq,b(ξ + λ) =(cid:16) KYi=1 niYℓ=ni−1+1 Hq,b(cid:0)(ξ + λ)/bℓ(cid:1)(cid:17) ·dµq,b(cid:0)(ξ + λ)/bnK(cid:1) (cid:3) Lemma 3.3. Let 2 ≤ q, b/q ∈ Z, and τ : Σ∗ satisfying (2.9). Then for each δ ∈ ΣM q → R be a maximal tree mapping q , M > 0, there exists δ′ ∈ Σ∗ q such that (3.16) where r = min(cid:16)r0, 1 dµq,b(ξ + Πτ,∞(δδ′)) ≥ r2Dτ+2HM(ξ + Πτ,M(δ)), ξ ∈ Tb, b(b−1)(cid:17) and r0, r1 are defined in (3.8). b , r1 8 Proof. If δ is a τ-main branch, we set δ′ = 0. In this case, dµq,b(ξ + Πτ,∞(δδ′)) = dµq,b(ξ + Πτ,M(δ)) = HM(ξ + Πτ,M(δ)) ·(cid:12)(cid:12)(cid:12)dµq,b(cid:0)(ξ + Πτ,M(δ))/bM(cid:1)(cid:12)(cid:12)(cid:12) ≥ (cid:16) dµq,b(η)(cid:17) · HM(ξ + Πτ,M(δ)) ≥ r0HM(ξ + Πτ,M(δ)), ξ ∈ Tb, η∈(−1/(b−1),(b−2)/(b−1)) inf (3.17) where the second equalities follows from(3.6), while the first inequality holds as b−M(ξ + Πτ,M(δ)) ∈(cid:0) − 1/(b − 1), (b − 2)/(b − 1)(cid:1) for all ξ ∈ Tb. Now consider δ is not a τ-main branch. In this case, define (3.18) δ′ := 0mδ′′, where m ≥ 1 is the smallest integer such that τ(δm+M) , 0, and δ′′ ∈ Σ∗ chosen that the quantities Dτ,δ0m(δ′′) in (2.6) and Dτ,δ0m in (2.7) are the same, (3.19) Let η1 = (ξ + Πτ,M+m(δ0m))/bM+m and η2 = (ξ + Πτ,M(δ))/bM for ξ ∈ Tb. Then Dτ,δ0m(δ′′) = Dτ,δ0m . q is so (3.20) η1 ∈ Tb and η2 ∈(cid:16) − 1 b − 1 , by (3.11) and τ(δ0m) = τ(δm+M) , 0. Write b − 2 b − 1(cid:17) KXj=1 (cid:0)Πτ,∞(δ0mδ′′) − Πτ,M+m(δ0mδ′′)(cid:1)/bM+m = dn jbn j−1 for some integers n j, 1 ≤ j ≤ K, satisfying 1 ≤ n1 < n2 < . . . < nK and some dn j ∈ {−1, 1, 2, . . . , b − 2}, 1 ≤ j ≤ K. Therefore (cid:12)(cid:12)(cid:12)dµq,b(cid:0)ξ + Πτ,∞(δδ′)(cid:1)(cid:12)(cid:12)(cid:12) M+mYl=M+1 = HM(ξ + Πτ,∞(δδ′)) ·(cid:12)(cid:12)(cid:12)(cid:12) Hq,b(cid:0)(ξ + Πτ,∞(δδ′))/bl(cid:1)(cid:12)(cid:12)(cid:12)(cid:12) ·(cid:12)(cid:12)(cid:12)dµq,b(cid:0)(ξ + Πτ,∞(δδ′))/bM+m(cid:1)(cid:12)(cid:12)(cid:12) mYl=1 KXj=1 = HM(ξ + Πτ,M(δ)) ·(cid:12)(cid:12)(cid:12)(cid:12) Hq,b(η2/bl)(cid:12)(cid:12)(cid:12)(cid:12) ·(cid:12)(cid:12)(cid:12)(cid:12)dµq,b(cid:16)η1 + 0r2Dτ,δ0m(cid:12)(cid:12)(cid:12)HM(ξ + Πτ,M(δ))(cid:12)(cid:12)(cid:12), ≥ r0r2Dτ,δ0m (δ′′)dµq,b(η2) · HM(ξ + Πτ,M(δ)) ≥ r2 d jbn j−1(cid:17)(cid:12)(cid:12)(cid:12)(cid:12) where the first inequality follows from (3.6), (3.14) and Lemma 3.2. Combining (3.17) and (3.21) proves (3.16). (3.21) (cid:3) Observe that Hq,b(ξ) in (3.3) satisfies (3.22) q−1Xj=0 Hq,b(ξ + j/b)2 = 1. 9 Hm(ξ + Πτ,m(δ))2 q Xδ∈Σm = Xδ′∈Σk = Xδ′∈Σk q q q−1Xj=0 q−1Xj=0 Hk(ξ + Πτ,k+1(δ′ j))2 ·(cid:12)(cid:12)(cid:12)Hq,b(cid:0)ξ/bk+1 + Πτ,k+1(δ′ j)/bk+1(cid:1)(cid:12)(cid:12)(cid:12) Hk(ξ + Πτ,k(δ′))2 ·(cid:12)(cid:12)(cid:12)Hq,b(cid:0)ξ/bk+1 + Πτ,k(δ′)/bk+1 + j/b(cid:1)(cid:12)(cid:12)(cid:12) 2 2 = 1, To prove Theorem 2.4, we need a similar identity for Hm(ξ), m ≥ 1, with shifts in Πτ,m(Σm q ). Lemma 3.4. Let 2 ≤ q, b/q ∈ Z, τ : Σ∗ 1, be as in (3.7). Then q → R be a tree mapping, and let Hm(ξ), m ≥ (3.23) Xδ∈Σm q Hm(cid:0)ξ + Πτ,m(δ)(cid:1)2 = 1, ξ ∈ R. Proof. For m = 1, Xδ∈Σm q Hm(ξ + Πτ,m(δ))2 = q−1Xj=0 Hq,b(ξ/b + τ( j)/b)2 = q−1Xj=0 Hq,b(ξ/b + j/b)2 = 1, where the last equality follows from (3.22), and the second one holds as Hq,b has period q/b and τ( j) − j ∈ qZ, 0 ≤ j ≤ q − 1, by the tree mapping property for τ. This proves (3.23) for m = 1. Inductively we assume that (3.23) hold for all m ≤ k. Then for m = k + 1, where the first equality holds as Hk+1(ξ) = Hk(ξ)Hq,b(ξ/bk+1), the second one fol- lows from the observations that Hk and Hq,b are periodic functions with period bk−1q and q/b respectively and that Πτ,k+1(δ′ j) = Πτ,k(δ′) + τ(δ′ j)bk ∈ Πτ,k(δ′) + jbk + qbkZ, 0 ≤ j ≤ q − 1, by the tree mapping property for τ, and the last one is true by (3.22) and the induc- tive hypothesis. This completes the inductive proof. (cid:3) We have all ingredients for the proof of Theorem 2.4. Proof of Theorem 2.4. Let Q(ξ) := QΛ(ξ) be the function in (3.1) associated with the maximal orthogonal set Λ := Λ(τ) of L2(µq,b). As Q is an analytic function on the real line, the spectral property for the maximal orthogonal set Λ reduces to proving Q(ξ) ≡ 1 for all ξ ∈ Tb by Lemma 3.1. Suppose, on the contrary, there exists ξ0 ∈ Tb such that (3.24) Q(ξ0) < 1. For n ≥ 1, set (3.25) Λn :=(cid:8)Πτ,∞(δ) : δ ∈ Σn q such that τ is regular on δ(cid:9), 10 and define (3.26) Then Qn(ξ) := Xλ∈Λn dµq,b(ξ + λ)2, ξ ∈ R. lim n→∞ since Λ = Λ(τ) and Σ∗ sequence that converges to Q(ξ), i.e., q = ∪∞ n=1 Λn = Λ and Λn ⊂ Λn+1 for all n ≥ 1, Σn q. This implies that Qn(ξ), n ≥ 1, is an increasing (3.27) lim n→∞ Qn(ξ) = Q(ξ), ξ ∈ R. Thus for sufficiently small ǫ > 0 chosen later, there exists an integer N such that (3.28) Q(ξ0) − ε ≤ QN(ξ0) ≤ Qn(ξ0) ≤ Q(ξ0) < 1 for all n ≥ N. For any δ ∈ Σn q being τ-regular, (3.29) lim m→∞ Hm(ξ + Πτ,m(δ)) = lim m→∞ For any δ ∈ Σn q such that δ is not τ-regular, the set {m ≥ n + 1 : τ(δm) , 0} contains infinite many integers. Denote that set by {m j, j ≥ 1} for some strictly increasing sequence {m j}∞ j=1. Recall that Hm(ξ + Πτ,∞(δ)) = dµq,b(ξ + Πτ,∞(δ)), ξ ∈ R. (3.30) τ(δm j) ∈ qZ ∩ {−1, 1, 2, . . . , b − 2} for all j ≥ 1 by the tree mapping property for τ. Therefore for m j ≤ m < m j+1 with j ≥ 1, Hm(ξ + Πτ,m(δ)) ≤ Hm j(ξ + Πτ,m(δ))(cid:12)(cid:12)(cid:12) = Hm j(ξ + Πτ,m j(δ))(cid:12)(cid:12)(cid:12) ≤ j−1Yk=1(cid:12)(cid:12)(cid:12)Hq,b(cid:0)(ξ + Πτ,m j(δ))/bmk +1(cid:1)(cid:12)(cid:12)(cid:12) j−1Yk=1(cid:12)(cid:12)(cid:12)Hq,b(cid:0)(ξ + Πτ,mk(δ))/bmk +1(cid:1)(cid:12)(cid:12)(cid:12) ≤ (cid:16) sup Hq,b(η)(cid:17) j−1 , ξ ∈ Tb, bη∈Tb = (3.31) where the inequalities follow from (3.11), (3.14) and (3.30), and the equalities hold by the tree mapping property τ and the q/b periodicity of the filter Hq,b. Combining (3.14) and (3.31) proves that (3.32) lim m→∞ Hm(ξ + Πτ,m(δ)) = 0, ξ ∈ Tb if δ ∈ Σn q is not τ-regular. Applying (3.29) and (3.32) with n and ξ replaced by N and ξ0 respectively, we can find a sufficient large integer M ≥ N + 1 such that (3.33) Xδ∈ΣN q HM(ξ0 + Πτ,M(δ))2 ≤ Xλ∈ΛN dµq,b(ξ0 + λ)2 + ε ≤ Q(ξ0) + ε. 11 This together with Lemma 3.4 implies that Xδ∈ΣM q \ΣN q (3.34) where HM(ξ0 + Πτ,M(δ))2 > 1 − Q(ξ0) − ε > 0, ΣM q \ΣN q \ΣN q =(cid:8)δ ∈ ΣM q : δN0∞ , δ0∞(cid:9). q , let λ(δ) = Πτ,∞(δδ′) with δ′ selected as in Lemma q . This implies that q \ΣN Now, for each δ ∈ ΣM 3.3. Observe that λ(δ) − Πτ,M(δ) ∈ bMZ for all δ ∈ ΣM λ(δ1) , λ(δ2) for two distinct δ1, δ2 ∈ ΣM q . Therefore q \ΣN Q(ξ0) =Xλ∈Λ dµq,b(ξ0 + λ)2 ≥ Xλ∈ΛN dµq,b(ξ0 + λ)2 + Xδ∈ΣM ≥ Q(ξ0) − ε + r4Dτ+4 Xδ∈ΣM q \ΣN q q \ΣN q dµq,b(ξ0 + λ(δ))2 HM(ξ + Πτ,M(δ))2 ≥ Q(ξ0) − ε + r4Dτ+4(1 − Q(ξ0) − ε), where the second inequality follows from (3.28) and Lemma 3.3, and the last holds by (3.34). This contradicts to (3.24) by letting ε chosen sufficiently small. (cid:3) 4. Maximal orthogonal sets and spectra: a necessary condition Given a tree mapping τ, define (4.1) Nτ(n) :=( inf0,δ∈Σq Dτ,δ(0∞) infδ∈Σn Dτ,δ(0∞) q\Σn−1 q if n = 1 if n ≥ 2, where Σn following strong version of Theorem 2.5. := {δ′ j : δ′ ∈ Σn−1 q\Σn−1 q q , 1 ≤ j ≤ q − 1}. In this section, we establish the Theorem 4.1. Let 2 ≤ q, b/q ∈ Z, τ : Σ∗ mapping, and let Nτ(n), n ≥ 1, be as in (4.1). Set q → {−1, 0, . . . , b − 2} be a maximal tree r2 := max{Hq,b(ξ) : 1/b ≤ b(b − 1)ξ ≤ b − 2}. IfP∞ < ∞, then Λ(τ) in (2.3) is not a spectrum of L2(µq,b). n=1 r2Nτ(n) For a maximal tree mapping τ satisfying (2.10), 2 ∞Xn=1 r2Nτ(n) 2 ≤ ∞Xn=1 r2ǫ0n 2 < ∞, where the last inequality holds as Hq,b(ξ) < 1 if bξ < qZ. This together with Theorem 4.1 proves Theorem 2.5. Now it remains to prove Theorem 4.1. Proof of Theorem 4.1. Let N0 ≥ 2 be so chosen that Nτ(n) ≥ 1 for all n ≥ N0. The existence follows the series convergence assumption on Nτ(n), n ≥ 1. Take δ ∈ Σn being τ-regular, where n ≥ N0. Write q\Σn−1 q {m ≥ n + 1 : τ(δm) , 0} = {nk : 1 ≤ k ≤ K} 12 for some integers n < n1 < n2 < . . . < nK, where K ≥ Nτ(n). Therefore for ξ ∈ Tb, dµq,b(ξ + Πτ,∞(δ)) = Hn(ξ + Πτ,∞(δ)) ·(cid:12)(cid:12)(cid:12)dµq,b(cid:0)(ξ + Πτ,∞(δ))/bn(cid:1)(cid:12)(cid:12)(cid:12) ≤ Hn(ξ + Πτ,n(δ)) · KYk=1(cid:12)(cid:12)(cid:12)Hq,b(cid:0)(ξ + Πτ,nk(δ))/b−nk −1(cid:1)(cid:12)(cid:12)(cid:12) · Hn(ξ + Πτ,n(δ)) (4.2) ≤ (cid:16) sup η∈Tb ≤ rNτ(n) 2 Hq,b(η/b)(cid:17)K Hn(ξ + Πτ,n(δ)), where the first equality holds by (3.6); the first inequality follows from (3.6), (3.14) and τ(δnk) ∈ qZ, 1 ≤ k ≤ K, by the tree mapping property for τ; the second inequality is true since (ξ + Πτ,nk(δ))/b−nk ∈ Tb by (3.11); and the last inequality follows from the definition of the quality Nτ(n). Let Λn and Qn, n ≥ 1, be as in (3.25) and (3.26) respectively, and set Λ0 = {0} (4.3) 1 − Qn(ξ) = 1 − Qn−1(ξ) − q\Σn−1 ≥ 1 − Qn−1(ξ) − r2Nτ(n) and Q0(ξ) = dµq,b(ξ)2. Then for n ≥ 1 and ξ ∈ Tb, Xδ∈Σn Xδ∈Σn (cid:16)1 − Xλ∈Λn−1 ) ·(cid:0)1 − Qn−1(ξ)(cid:1), ≥ 1 − Qn−1(ξ) − r2Nτ(n) where the first equality holds because = (1 − r2Nτ(n) is τ-regular 2 q\Σn−1 q 2 2 q dµq,b(ξ + Πτ,∞(δ))2 Hn(ξ + Πτ,n(δ))2 dµq,b(ξ + λ)2(cid:17) the first inequality is true by (4.2); and the second inequality follows from Lemma 3.4 and Λn\Λn−1 =(cid:8)Πτ,∞(δ) : δ ∈ Σn dµq,b(ξ + λ)2 ≤ Xδ∈Σn−1 Xλ∈Λn−1 q q\Σn−1 q is τ-regular(cid:9); Hn(ξ + Πτ,n(δ))2, ξ ∈ R, by (3.6) and (3.14). Recall that limn→∞ Qn(ξ) = Q(ξ), ξ ∈ R, by (3.27). Applying (4.3) repeatedly and using the convergence ofP∞ n=1 r2Nτ(n) 2 gives (4.4) On the other hand, 2 1 − Q(ξ) ≥(cid:16) ∞Yn=N0+1(cid:0)1 − r2Nτ(n) dµq,b(ξ + λ)2 < Xδ∈Σ N0 q QN0(ξ) = Xλ∈ΛN0 (cid:1)(cid:17) · (1 − QN0(ξ)), ξ ∈ Tb. HN0(ξ + Πτ,N0(δ))2 = 1, ξ ∈ Tb by (3.6), (3.14) and (3.23). This together with (4.4) proves that Q(ξ) < 1 for all ξ ∈ Tb, and hence Λ = Λ(τ) is a not a spectrum for L2(µq,b) by Lemma 3.1. (cid:3) 13 5. Spectra rescaling In this section, we first prove Theorem 2.6 in Subsection 5.1. We then consider verification of maximal orthogonality of the rescaled set KΛ in Subsection 5.2. In that subsection, we show that the rescaled set KΛ is not a maximal orthogonal set of the Cantor measure µq,b if and only if the labeling tree τ(Σ∗ q) has certain periodic properties (5.1) and (5.3). Theorem 5.1. Let 2 ≤ q, b/q ∈ Z, τ : Σ∗ q → {−1, 0, . . . , b − 2} be a maximal tree mapping, Λ := Λ(τ) be as in (2.3), and let K > 1 be an integer coprime with b. Then KΛ is not a maximal orthogonal set of the Cantor measure µq,b if and only if there exist δ ∈ Σ∞ n=M+1 is a periodic sequence with positive period N, i.e., q and a nonnegative integer M such that {τ(δn)}∞ (5.1) τ(δn) = τ(δn+N), n ≥ M + 1, and that the word W = ω1ω2 · · · ωN defined by (5.2) ω j = τ(δM+ j), 1 ≤ j ≤ N, is a repetend of the recurring b-band decimal expression of i/K for some i ∈ Z\{0}, i.e., (5.3) i K ω jb j−Nn−1 = PN = 0.ωN · · · ω2ω1ωN · · · ω2ω1ωN · · · = j=1 ω jb j−1 bN − 1 . ∞Xn=1 NXj=1 By Theorems 2.6 and 5.1, we see that the rescaled set KΛ is a spectrum if and only if the labeling tree of Λ contains no repetend of K. For the spectrum Λ4 of the Bernoulli convolution µ4 in (2.11), the associated maximal tree mapping τ2,4 on Σ∗ 2 is given by (5.4) τ2,4(δ) = δn for δ = δ1 · · · δn ∈ Σn 2, n ≥ 1. Thus Dτ2,4,δ = 0 for all δ ∈ Σ∗ 2, and the requirement (2.9) is satisfied for the max- imal tree mapping τ2,4. Hence Corollary 2.7 follows immediately from Theorem 2.6 and 5.1. Finally in Subsection 5.3, we construct a spectrum Λ of the Cantor measure µq,b such that Λ/(b − 1), a seemingly denser set than the spectrum Λ, is its maximal orthogonal set but not its spectrum. Theorem 5.2. Consider 2 ≤ q, b/q ∈ Z and b > 4. Define a tree mapping κ : Σ∗ {−1, 0, 1, . . . , b − 2} by q → (5.5) κ(δk+1) = if 1 ≤ δ ≤ q − 1 and k = 0 0 if δ = 0 and k ≥ 0 δ q if 1 ≤ δ ≤ q − 1 and k ∈ {1, 2, · · · , Kδ, 2b} 0 if 1 ≤ δ ≤ q − 1 and Kδ < k , 2b 14 if δ ∈ Σ1 q, where 0 ≤ Kδ ≤ b − 2 is the unique integer such that q(Kδ + 1) + δ ∈ (b − 1)Z; and inductively if k = 0 j q if k ∈ {1, 2, . . . , Kδ, n + 2b − 1} 0 if k > Kδ and k , n + 2b − 1 if δ = δ′ j for some δ′ ∈ Σn−1 2} is the unique integer such that , n ≥ 2 and j ∈ {1, . . . , q − 1}, where Kδ ∈ {0, 1, . . . , b − κ(δk+n) = (cid:16) n−1Xi=1 q (5.6) (5.7) Then κ(δi) + q(Kδ + 1) + j(cid:17) ∈ (b − 1)Z. Λq,b :=(cid:8)Πκ,∞(δ) : δ ∈ Σ∗ q(cid:9) (5.8) is a spectrum of the Cantor measure µq,b, and the rationally rescaled set Λq,b/(b−1) is its maximal orthogonal set but not its spectrum. 5.1. Proof of Theorem 2.6. The necessity is obvious. Now we prove the suffi- ciency. Without loss of generality, we assume K is positive since −Λ is a spectrum (maximal orthogonal set) if and only if Λ is. Let κ be the maximal tree mapping associated with the maximal orthogonal set KΛ of the Cantor measure µq,b. The existence of such a mapping follows from Theorem 2.2 and the assumption on KΛ. Denote the integral part of a real number x by ⌊x⌋. By Theorem 2.4 and the assumption that Dτ < ∞, it suffices to prove that (5.9) q} ≤ (2⌊logb K⌋ + 4)(Dτ + 1), δ ∈ Σ∗ q. inf{Dκ,δ(δ′), δ′ ∈ Σ∗ Take δ ∈ Σn q, n ≥ 1, and let δ1 ∈ Σ∗ q be so chosen that δδ1 is κ-regular. As Πκ,∞(δδ1) ∈ KΛ, there exists ζ ∈ Σn (5.10) q such that KΠτ,n(ζ) − Πκ,n(δ) ∈ bnZ. Let ζ ′ ∈ Σ∗ q be so chosen that ζζ ′ is a τ-main subbranch of ζ and (5.11) Dτ,ζ(ζ ′) = Dτ,ζ , where the existence of such a tree branch ζ ′ follows from (2.8). Therefore the q such that δδ′ is a verification of (5.9) reduces to showing the existence of δ′ ∈ Σ∗ κ-main branch, (5.12) and (5.13) KΠτ,∞(ζζ ′) = Πκ,∞(δδ′), Dκ,δ(δ′) ≤ (2⌊logb K⌋ + 4)(Dτ,ζ + 1). By Theorem 2.2, there exists a κ-main branch δ2 ∈ Σ∗ q such that (5.14) Then (5.15) Πκ,∞(δ2) = KΠτ,∞(ζζ ′). Πκ,n(δ2) − Πκ,n(δ) ∈ KΠτ,n(ζ) − Πκ,n(δ) + bnZ = bnZ 15 by (5.10). This together with one-to-one correspondence of the mapping Πκ,n : Σn q → Z proves δ2 = δδ′ for some δ′ ∈ Σ∗ q. The equation (5.12) follow from (5.14). Now it remains to prove (5.13). With- out loss of generality, we assume that Πκ,∞(δδ′) , Πκ,n(δ), because otherwise Dκ,δ(δ′) = 0 and hence (5.13) follows immediately. Thus we may write (5.16) Πκ,∞(δδ′) = Πκ,n(δ) + LXl=1 dlbn+ml−1 for a strictly increasing sequence {ml}L 2}, 1 ≤ l ≤ L. l=1 of integers and some dl ∈ {−1, 1, . . . , b − Also we may assume that Πτ,∞(ζζ ′) , Πτ,n(ζ), because otherwise KΠτ,∞(ζζ ′) = KΠτ,n(ζ) ∈ K(−bn/(b − 1), (b − 2)bn/(b − 1)) and Πκ,∞(δδ′) < (−bn+mL −1/(b − 1), (b − 2)bn+mL −1/(b − 1)) by (3.11) and (5.16). This together with (5.12) implies that bmL−1 ≤ K and hence Therefore we can write Dκ,δ(δ′) ≤ mL ≤ ⌊logb K⌋ + 1. Πτ,∞(ζζ ′) = Πτ,n(ζ) + NXj=1 c jbn+n j −1, where c j ∈ {−1, 1, . . . , b−2}, 1 ≤ j ≤ N, and {n j}N of integers. j=1 is a strictly increasing sequence To prove (5.13) for the case that Πτ,∞(ζζ ′) , Πτ,n(ζ), we need the following claim: Claim 1: {ml, 1 ≤ l ≤ L} ⊂ ∪N j=0[n j, n j + ⌊logb K⌋ + 1]. Proof. Suppose, on the contrary, that Claim 1 does not hold. Then there exists + ⌊logb K⌋ + 1 < ml < n j0 +1 for some 0 ≤ j0 ≤ N, where 1 ≤ l ≤ L such that n j0 we set n0 = 0 and nN+1 = +∞. Observe that Πκ,n+ml(δδ′) − KΠτ,n+n j0 by (5.12) and the assumption ml < n j0 +1, and (ζζ ′) ∈ bn+ml Z (5.17) Πκ,n+ml(δδ′) − KΠτ,n+n j0 bn+ml−1 − 1 (ζζ ′) ∈ dlbn+ml−1 + ⊂ dlbn+ml−1 + (−bn+ml−1, bn+ml−1) b − 1 [−1, b − 2] − K bn+n j0 − 1 b − 1 [−1, b − 2] (5.18) by the definitions of Πκ,n+ml and Πτ,n+n j0 + logb K + 1 < ml. Combining (5.17) and (5.18) leads to the contradiction that dl ∈ {−1, 1, . . . , b − 2}. This completes the proof of Claim 1. and the assumption n j0 (cid:3) 16 To prove (5.13) for the case that Πτ,∞(ζζ ′) , Πτ,n(ζ), we need another claim: Claim 2: If n j + ⌊logb K⌋ + 1 < n j+1, then there exists l0 such that ml0 = n j+1, ml0−1 ∈ [n j, n j + ⌊logb K⌋ + 1] and dl0 ∈ qZ if and only if c j+1 ∈ qZ. Proof. Let l0 be the smallest integer l with ml ≥ n j+1. By Claim 1, ml0−1 ≤ n j + (δδ′) − KΠτ,n+n j+1(ζζ ′) ∈ bn+n j+1 Z by ⌊logb K⌋ + 1 ≤ n j+1 − 1. Observe that Πκ,n+ml0 (5.12); and Πκ,n+ml0 (δδ′) − KΠτ,n+n j+1(ζζ ′) ∈ dl0bn+ml0 −1 − Kc j+1bn+n j+1−1 + ⊂ dl0bn+ml0 −1 − Kc j+1bn+n j+1−1 + bn+n j+1−1(−1, 1). (−1, b − 2) − bn+ml0 −1 b − 1 (5.19) Kbn+n j b − 1 (−1, b − 2) Thus dl0bml0 −n j+1 − Kc j+1 ∈ bZ. This together, with the assumptions that c j+1 ∈ = n j+1 and dl0 ∈ qZ {−1, 1, . . . , b−2} and that K and b are coprime, implies that ml0 if and only if c j+1 ∈ qZ. From the argument in (5.19), we see that (5.20) Πκ,ml0−1(δδ′) = KΠτ,n j(ζζ ′). Thus ml0−1 ≥ n j, as Πκ,ml0−1(δδ′) ∈ bml0−1(cid:0) − 1/(b − 1), (b − 2)/(b − 1)(cid:1) and KΠτ,n j(ζζ ′) < Kbn j−1(cid:0) − 1/(b − 1), (b − 2)/(b − 1)(cid:1) by (3.11). This completes the proof of Claim 2. (cid:3) Having established the above two claims, let us return to the proof of the in- equality (5.13). Note that if {k ∈ Z : ml0−1 < k < ml0 } 1 ∪N j=0[n j, n j + ⌊logb K⌋ + 1] for some 1 ≤ l0 ≤ L, then by Claim 1, there exists 1 ≤ j0 ≤ N such that ml0−1 ≤ n j0−1 + ⌊logb K⌋ + 1 < n j0 ≤ ml0 . Then ml0 Thus = n j0, ml0−1 ≥ n j0−1 and dl0 ∈ qZ if and only if c j0 ∈ qZ by Claim 2. and thus ∪dl<qZ(ml−1, ml) ⊂(cid:16)∪N Xdl<qZ j=0[n j, n j + ⌊logb K⌋ + 1](cid:17) ∪(cid:16)∪c j <qZ(n j−1, n j)(cid:17) , (ml − ml−1 − 1) ≤ (⌊logb K⌋ + 2)(N + 1) + Xc j <qZ (n j − n j−1 − 1). This together with Claim 1, implies Dκ,δ(δ′) ≤ 2(⌊logb K⌋+2)(N +1)+Xc j <qZ (n j −n j−1 −1) ≤ (2⌊logb K⌋+4)(Dτ,ζ(ζ ′)+1). We get (5.13) and hence complete the proof of Theorem 2.6. 17 5.2. Proof of Theorem 5.1. (⇐=) Let (5.21) λ0 = KΠτ,M(δ) − ibM , where i ∈ Z is given in (5.3). Inductively applying (5.3) proves that (5.22) λ0 = KΠτ,M+N(δ) − ibM+N = · · · = KΠτ,M+nN(δ) − ibM+nN , n ≥ 1. Take λ ∈ Λ. Now we show that exp(−2πiλ0x) is orthogonal to exp(−2πiKλx). q for By the maximality of the tree mapping τ, there exists a τ-main branch ζ ∈ Σm some m ≥ 1 by Theorem 2.2 such that (5.23) λ = Πτ,∞(ζ). Also for sufficiently large n ≥ 1, there exists λn ∈ Λ by the maximality of the tree mapping τ such that λn , λ and (5.24) The reason for λn , λ is that Πτ,M+Nn(δ) , Πτ,M+Nn(ζ) for sufficiently large n by W = ω1 . . . ωN , 0N by (5.3). λn − Πτ,M+Nn(δ) ∈ bM+NnZ. As both λ, λn ∈ Λ, there exists a nonnegative integer l and an integer a ∈ Z\qZ by (2.1) such that (5.25) Now we show that (5.26) λ − λn = abl. l < M + Nn when n is sufficiently large. Suppose, on the contrary, that l ≥ M + Nn. Then (5.27) On the other hand, λ − Πτ,M+Nn(δ) ∈ bM+NnZ. Πτ,M+Nn(δ) ∈ bM+Nn[−1/(b − 1), (b − 2)/(b − 1)] by the tree mapping property for τ. Therefore λ = Πτ,M+Nn(δ) for sufficiently large n, which is a contradiction as Πτ,M+Nn(δ) < bM+N(n−1)(−1/(b − 1), (b − 2)/(b − 1)) by W = ω1 . . . ωN , 0N and the tree mapping property for τ. Combining (5.24), (5.25) and (5.26) and recalling that K and b are co-prime, we obtain that (5.28) Kλ − KΠτ,M+Mn(δ) = abl for some integers 0 ≤ l < M + Nn and a ∈ Z\qZ. Thus the inner product between exp(−2πiλ0x) and exp(−2πiKλx) is equal to zero by (2.1), (5.22) and (5.28). This proves that KΛ is not a maximal orthogonal set as λ ∈ Λ is chosen arbitrarily. (=⇒) By (2.1) and the assumption on the rescaled set KΛ, there exists a max- imal orthogonal set Θ of the Cantor measure µq,b such that (5.29) KΛ Θ ⊂ Z. 18 Take ϑ0 ∈ Θ\(KΛ). Then (5.30) for some κ-main branch ζ0 ∈ Σm associated with the maximal orthogonal set Θ. ϑ0 = Πκ,∞(ζ0) = Πκ,m(ζ0) q , m ≥ 1, where κ is the maximal tree mapping Let τ be the maximal tree mapping in Theorem 2.2 such that Λ = Λ(τ). To establish the necessity, we need the following claim: Claim 3: Let n ≥ 1. For any ζ ∈ Σn q there exists a unique δ ∈ Σn q such that Πκ,n(ζ) − KΠτ,n(δ) ∈ bnZ. Proof. Observe that KΠτ,n(δ1) − KΠτ,n(δ2) < bnZ for all distinct δ1, δ2 ∈ Σn q, (5.31) because b/q ∈ Z, K and b are coprime, and Πτ,n(δ1) − Πτ,n(δ2) = abl for some 0 ≤ l ≤ n − 1 and a < qZ. On the other hand, (5.32) {KΠτ,n(δ) : δ ∈ Σn by (5.29). Combining (5.31) and (5.32) leads to q}+bnZ = KΛ+bnZ ⊂ Θ+bnZ = {Πκ,n(ζ) : ζ ∈ Σn q}+bnZ (5.33) {KΠτ,n(δ) : δ ∈ Σn q} + bnZ = {Πκ,n(ζ) : ζ ∈ Σn q} + bnZ. Then Claim 3 follows from (5.33) and (5.31). (cid:3) To establish the necessity, we need another claim: Claim 4: ϑ0 < KZ. Proof. Suppose, on the contrary, that ϑ0 ∈ KZ. Then for any λ ∈ Λ, there exist a ∈ Z\qZ and 0 ≤ l ∈ Z by (2.1) and (5.29) such that ϑ0 − Kλ = abl. This together with the co-prime assumption between K and b implies that a/K ∈ Z and 0 , ϑ0/K − λ ∈ (a/K)bl. Thus Λ ∪ {ϑ0/K} is an orthogonal set for the measure µq,b by (2.1), which contradicts to the maximality of the set Λ. (cid:3) Now we continue our proof of the necessity. Let N be the smallest positive integer such that (bN − 1)ϑ0/K ∈ Z, where the existence follows from the co-prime property between K and b. By Claim 4, there exists ω j ∈ {−1, 0, . . . , b − 2}, 1 ≤ j ≤ N, such that the word W := ω1ω2 · · · ωN , 0 and (5.34) ϑ0 K for some integer c ∈ Z. Let W ′ = ω′ 2}, 1 ≤ j ≤ N, and = c.ωN · · · ω2ω1ωN · · · ω2ω1 · · · = c + PN j=1 ω jb j−1 bN − 1 N be so chosen that ω′ 1ω′ 2 · · · ω′ j ∈ {−1, 0, . . . , b− 0 (ω′ (5.35) NXj=1 j=1 ω jb j−1 ∈ bN −1 j=1 ω jb j−1 ∈ bN −1 The existence of such a word W ′ follows from the observation that j + ω j)b j−1 = ω jb j−1, ω j ∈ {−1, 0, . . . , b − 2}o =(cid:16)bN − 1 if PN bN − 1 if PN n NXj=1 b − 1 [−1, b − 2](cid:17) ∩ Z. b−1 [−1, 1) b−1 [1, b − 2]. 19 Let n > m/N and set ζnN = ζ00nN−m ∈ ΣnN q . By Claim 3 and the κ-main branch assumption for ζ0, there exists δnN ∈ ΣnN q such that (5.36) KΠτ,nN(δnN) − ϑ0 ∈ bnN Z. Combining (5.34), (5.35) and (5.36) and recalling that K and b are coprime, we obtain (bN − 1)(Πτ,nN(δnN) − c) + NXj=1 ω′ jb j−1 ∈ bnN Z, j=1 ω jb j−1 ∈ bN −1 j=1 ω jb j−1 ∈ bN −1 b−1 [−1, 1) b−1 [1, b − 2]. where Therefore (5.37) c c = if PN c − 1 if PN Πτ,nN(δnN) − c −(cid:16) NXj=1 ω′ By the construction of ω′ j=1 ω′ jb j−1 ∈ bN −1 sufficiently large k, PN b−1 (−1, b − 2) orPN c +(cid:16) NXj=1 ω′ jb j−1(cid:17)(cid:0)1 + bN + · · · + b(n−1)N(cid:1) ∈ bnN Z. j, 1 ≤ j ≤ N,PN j=1 ω′ j=1 ω′ jb j−1 = bN −1 jb j−1 ∈ bN −1 b−1 (−1, b − 2]. If either b−1 (b − 2) and c ≤ 0, then for jb j−1(cid:17)(cid:0)1 + bN + · · · + b(k−1)N(cid:1) = kNXj=1 θ jb j−1 for some θ j ∈ {−1, 0, . . . , b − 2}, 1 ≤ j ≤ kN, as it is contained in [−(bkN − 1)/(b − 1), (bkN − 1)(b − 2)/(b − 1)]. This together with (5.37) implies that Πτ,nN(δnN) = kNXj=1 θ jb j−1 + NXj=1 ω′ jb j−1(cid:0)bkN + · · · + b(n−1)N(cid:1) for n ≥ k. Thus there exists δ ∈ Σ∞ q such that δnN = δnN and τ(δnN+ j) = ω′ j, 1 ≤ j ≤ N for n ≥ k, which proves the desired conclusion. Now consider the case thatPN b−1 (b − 2) and c > 0. In this case, = b − 2 for all 1 ≤ j ≤ N and N = 1 by the selection of the integer N. Further jb j−1 = bN −1 j=1 ω′ ω′ j we obtain from (5.37) that Πτ,n(δn) − c + 1 + nXj=1 b j−1 ∈ bnZ, which implies that there exists δ ∈ Σ∞ sufficiently large n, which proves the desired conclusion. q such that δn = δn and τ(δn) = −1 for 20 5.3. Proof of Theorem 5.2. First we show that Λq,b is a spectrum of the Cantor measure µq,b. Observe that κ is a maximal tree mapping, every δ ∈ Σ∗ q is κ-regular, and Λq,b = Λ(κ). We then obtain from Theorem 2.2 that (5.38) Λq,b is a maximal orthogonal set of the Cantor measure µq,b. From the definition of the maximal tree mapping κ it follows that (5.39) Dκ,δ ≤ Dκ,δ(0∞) ≤ Kδ + 1 ≤ b − 1 for all δ ∈ Σ∗ q, where Kδ is given in (5.7). Therefore the spectral property for Λq,b holds by (5.38), (5.39) and Theorem 2.4. Next we prove that Λq,b/(b − 1) is a maximal orthogonal set for the Cantor measure µq,b. From (2.1) and the spectral property for the set Λq,b We obtain that (5.40) Λq,b − Λq,b ⊂ {b ja : 0 ≤ j ∈ Z, a ∈ Z\qZ} ∪ {0}. On the other hand, and for any δ ∈ Σ∗ q, 0 ∈ Λq,b ⊂ Z (5.41) Πκ,∞(δ) = ∞Xj=1 κ(δ j)b j−1 ∈ ∞Xj=1 κ(δ j) + (b − 1)Z = (b − 1)Z by (5.5) -- (5.7). Combining (5.40) and (5.41) leads to (Λq,b − Λq,b)/(b − 1) ⊂ {b ja : 0 ≤ j ∈ Z, a ∈ Z\qZ} ∪ {0}, and hence Λq,b/(b − 1) is an orthogonal set for the Cantor measure µq,b by (2.1). Now we establish the maximality of the rescaled set Λq,b/(b − 1). Suppose, on the contrary, that there exists λ0 < Λq,b/(b − 1) such that Λq,b := Λq,b/(b − 1) ∪ {λ0} is an orthogonal set for the Cantor measure µq,b. Then (b − 1) Λq,b − (b − 1) Λq,b ⊂ (b − 1)(cid:0){b ja : 0 ≤ j ∈ Z, a ∈ Z\qZ} ∪ {0}(cid:1) ⊂ {b ja : 0 ≤ j ∈ Z, a ∈ Z\qZ} ∪ {0} and (b − 1) Λq,b is an orthogonal set for the Cantor measure µq,b by (2.1). This contradicts the spectral property for Λq,b. Finally we prove that Λq,b/(b − 1) is not a spectrum of the Cantor measure µq,b. Let τq,b : Σ∗ q → {−1, 0, . . . , b − 2} be the maximal tree mapping such that Λq,b/(b − 1) = Λ(τq,b). By Theorem 4.1, the non-spectral property for the set Λq,b/(b − 1) reduces to showing that (5.42) Dτq,b,δ(0∞) ≥ n for all δ ∈ Σn with (5.5) and (5.6) implies the existence of η ∈ Σm q\Σn−1 , n ≥ 2, being τq,b-regular. Recall that Λq,b = Λ(κ). This together q q , m ≥ 1, such that (5.43) (b − 1)Πτq,b,∞(δ) = Πκ,∞(η) = 21 m+b−2Xj=1 d jb j−1 + q · b2m+2b−2, where d j ∈ {0, 1, · · · , q} for all 1 ≤ j ≤ m + b − 2 and dm ∈ {1, . . . , q − 1}. Write (5.44) Πτq,b,∞(δ) = ∞Xj=1 c jb j−1 = MXj=1 c jb j−1 where c j := τq,b(δ j) ∈ {−1, 0, . . . , b − 2} and M ≥ n is so chosen that cM , 0. The existence of such an integer follows from τq,b(δn) ∈ Z\qZ and τq,b(δ j) = 0 for sufficiently large j. Combining (5.43) and (5.44) leads to MXj=1 c jb j−1 = 1 b − 1(cid:16) m+b−2Xj=1 2m+2b−3Xj=m+b−2 d jb j−1 + q · bm+b−2(cid:17) + q b j +(cid:16)0, b − 1(cid:17)bm+b−2, b − 2 ∈ q 2m+2b−3Xj=m+b−2 b j where the last inequality follows as q ≤ b−3. This, together with c j ∈ {−1, 0, 1, . . . , b− 2}, 1 ≤ j ≤ M, implies that M = 2m + 2b − 2 and c j = q, m + b − 2 < j ≤ M. (5.45) On the other hand, for δ ∈ Σn τq,b that cn < qZ. Thus n ≤ m + b − 2 according to (5.45). Therefore q\Σn−1 q it follows from the tree mapping property for Dτq,b,δ(0∞) ≥ M − (m + b − 2) ≥ n. This proves (5.42) and then the conclusion that Λq,b is not a spectrum of the Cantor set µq,b by Theorem 4.1. References [1] X.-R. Dai, When does a Bernoulli convolution admit a spectrum?, Adv. Math., 231(2012), 1681 -- 1693. [2] X.-R. Dai, X.-G. He and C.-K. Lai, Spectral property of Cantor measures with consecutive digits, Adv. Math., 242(2013), 187 -- 208. [3] X.-R. Dai, X.-G. He and C.-K. Lai, Law of pure types and some exotic spectra of fractal spec- tral measures, Geometry and Analysis of Fractals D.-J. Feng and K. S. Lau (eds.), Springer Proceeding in Mathematics & Statistics 88, pp 47 -- 64, Springer-Verlag Berlin Heidelberg, 2014.. [4] X.-R. Dai, X.-G. He and K.-S. Lau, On spectral N-Bernoulli measures, Adv. Math., 259(2014), 511 -- 531. [5] X.-R. Dai and Q. Sun, Spectral measures with arbitrary Hausdorff dimensions, J. Funct. Anal., to appear. [6] D. Dutkay, D. Han and Q. Sun, On spectra of a Cantor measure, Adv. Math., 221(2009), 251 -- 276. [7] D. Dutkay, D. Han and Q. Sun, Divergence of mock and scrambled Fourier series on fractal measures, Trans. Amer. Math. Soc., 366(2014), 2191 -- 2208. [8] D. Dutkay and P. Jorgensen, Fourier duality for fractal measures with affine scales, Math. Comp., 81(2012), 2253 -- 2273. [9] D. Dutkay and C.-K. Lai, Uniformity of measures with Fourier frames, Adv. Math., 252(2014), 684 -- 707. [10] K. J. Falconer, Fractal Geometry, Mathematical Foundations and Applications, Wiley, New York, 1990. 22 [11] B. Fuglede, Commuting self-adjoint partial differential operators and a group theoretic prob- lem, J. Funct. Anal., 16(1974), 101 -- 121. [12] X.-G. He, C.-K. Lai and K.-S. Lau, Exponential spectra in L2(µ), Appl. Comput. Harmon. Anal., 34(2013), 327 -- 338. [13] T.-Y. Hu and K.-S. Lau, Spectral property of the Bernoulli convolutions, Adv. Math., 219(2008), 554 -- 567. [14] P. Jorgensen and S. Pedersen, Dense analytic subspaces in fractal L2 spaces, J. Anal. Math., 75(1998), 185 -- 228. [15] I. Łaba and Y. Wang, On spectral Cantor measures, J. Funct. Anal., 193(2002), 409 -- 420. [16] J. C. Lagarias and Y. Wang, Tiling the line by the translates of one tile, Invent. Math., 124(1996), 341 -- 365. [17] H. Landau, Necessary density conditions for sampling and interpolation of certain entire functions, Acta Math., 117(1967), 37 -- 52. [18] N. Levinson, Gap and Density Theory, Am. Math. Soc. Colloq. Publ., Vol 26., New York 1940. [19] J. Ortega-Cerd`a and K. Seip, Fourier frames, Ann. of Math. (2), 255(2002), 789 -- 806. [20] R. E. A. C. Paley and N. Weiner, Fourier Transform In The Complex Domain, Am. Math. Soc. Colloq. Publ., Vol 19., New York 1934. [21] A. Poltoratski, A problem on completeness of exponentials, Ann. of Math. (2), 178(2013), 983 -- 1016. [22] A. Poltoratski, Spectral gaps for sets and measures, Acta Math., 208(2012), 151 -- 209. [23] R. S. Strichartz, Convergence of mock Fourier series , J. Anal. Math., 99(2006), 333 -- 353. [24] T. Tao, Fuglede's conjecture is false in 5 or higher dimensions, Math. Res. Lett., 11(2004), 251 -- 258. [25] R. M. Young, An Introduction to Nonharmonic Fourier Series, Acdemic, New York, 1980. School of Mathematics and Computational Science, Sun Yat-sen University, Guangzhou, 510275, P. R. China E-mail address: [email protected] 23
1310.1204	1	1310	2013-10-04T09:06:04	Concentration phenomena in high dimensional geometry	[ "math.FA" ]	The purpose of this note is to present several aspects of concentration phenomena in high dimensional geometry. At the heart of the study is a geometric analysis point of view coming from the theory of high dimensional convex bodies. The topic has a broad audience going from algorithmic convex geometry to random matrices. We have tried to emphasize different problems relating these areas of research. Another connected area is the study of probability in Banach spaces where some concentration phenomena are related with good comparisons between the weak and the strong moments of a random vector.	math.FA	math	Concentration phenomena in high dimensional geometry. Olivier Gu´edon Abstract The purpose of this note is to present several aspects of concentration phenomena in high dimensional geometry. At the heart of the study is a geometric analysis point of view coming from the theory of high dimensional convex bodies. The topic has a broad audience going from algorithmic convex geometry to random matrices. We have tried to emphasize different problems relating these areas of research. Another connected area is the study of probability in Banach spaces where some concentration phenomena are related with good comparisons between the weak and the strong moments of a random vector. 1 Convex geometry and log-concave measures A function f : Rn → R+ is said to be log-concave if ∀x, y ∈ Rn , ∀θ ∈ [0, 1], f ((1 − θ)x + θy) ≥ f (x)1−θ f (y)θ Define a measure µ with a log-concave density f ∈ Lloc 1 , the Pr´ekopa-Leindler inequality [68, 59] implies that it satisfies: for every compact sets A, B ⊂ Rn , for every θ ∈ [0, 1], µ((1 − θ)A + θB ) ≥ µ(A)1−θ µ(B )θ . (1) A measure satisfying (1) is said to be log-concave. A complete characterization of log- concave measures is well known. It has been done during the sixties and seventies and it is related to the work of [46, 67, 68, 59, 15], see also the surveys [21, 61]. In [15] it is proved that a measure is log-concave if and only if it is absolutely continuous with respect to the Lebesgue measure on the affine subspace generated by its convex support, with log- concave locally integrable density. Classical examples are the case of product of exponential 2n exp(−x1 ), the Gaussian measure, f (x) = (2π)n/2 exp(−x2 distributions, f (x) = 1 1 2/2), the uniform measure on a convex body, f (x) = 1K (x). Moreover, it is well known [22, 68, 59, 13] that the class of log-concave measures is stable under convolution and linear transformations. It is important to notice that the class of uniform distribution on a convex body is stable under linear transformation but not under convolution. This is one among several reasons why it is preferable to work with log-concave measures. 1 1.1 The hyperplane conjecture or the slicing problem. It is one of the famous conjecture in high dimensional convex geometry. Conjecture 1. (The hyperplane conjecture) There exist a constant C > 0 such that for every n and every convex body K ⊂ Rn of volume 1 and barycenter at the origin, there is a direction θ such that Vol (K ∩ θ⊥ ) ≥ C . Several other formulations are known. For example, it is equivalent to ask if for every convex bodies K1 and K2 with barycenter at the origin such that for every θ ∈ S n−1 Vol (K1 ∩ θ⊥ ) ≤ Vol (K2 ∩ θ⊥ ) then Vol (K1 ) ≤ C Vol (K2 ) ? It is worth noticed that the constant C can not be 1 in high dimension. Indeed, replacing C by 1 in the conclusion lead to the Busemann-Petty problem which is known to be true in dimension n ≤ 4 but false in dimension n ≥ 5 [35]. We refer to [63] for a more detailed presentation of the slicing (cid:90) problem. For a convex body K , define LK by K n L2 K = min T ∈SLn (R) 1 (Vol K )1+ 2 n T x2 2 dx (2) where the minimum is taken over all affine transformations preserving the volume. Ob- serve that LK is invariant under affine transformation. From a result of Hensley [45], the hyperplane conjecture is equivalent to the following question. Does there exist a constant C such that for every dimension n and convex body K ⊂ Rn , LK ≤ C ? The number LK is called the isotropic constant of the convex body K . The minimum in (2) is attained when the ellipsoid E is the inertia matrix associated to K , centered at the barycenter of (cid:90) (cid:90) K . Equivalently, it is attained for B n 2 when K is in isotropic position, that is : 1 Vol K K K xixj dx = δi,j . 1 Vol K xdx = 0, and 1 In isotropic position, we have LK = It is also possible to define the isotropic . 1 constant of a log-concave function f : Rn → R+ . Let X be the random vector in Rn with (Vol K ) n density of probability f , define Lf = LX = (det(CovX ))2/n (f (EX ))1/n (cid:90) (cid:90) (cid:90) It is not difficult to check that it is invariant under linear transformation, LX = LT (X ) for any T ∈ GLn (R) so that we can assume that X is isotropic. This means that f (x)dx = 1, xixj f (x)dx = δi,j xf (x) = 0 and and that LX = f (0)1/n . A theorem of Ball [6] asserts that the hyperplane conjecture is equivalent to the uniform boundedness of the isotropic constant of log-concave random vectors. Theorem 1 (Ball [6]). There exists a constant C such that Lf ≤ C sup LK ≤ sup n,K n,f sup n,K LK where the suprema are taken with respect to every dimension n, every (isotropic) convex bodies and every (isotropic) log-concave functions. 2 The left hand side of the inequality is obvious. There are more important ingredients to prove the right hand side inequality. In particular, it requires to define a convex body (cid:27) (cid:26) from a log-concave measure and to keep a good control of the isotropic constant. Let (cid:90) +∞ f : Rn → R+ be a log-concave function and define a family of set (Kp (f ))p>0 by tp−1f (tx)dt ≥ f (0) x ∈ Rn , p 0 A main step in the proof of Ball’s theorem above is to prove that Kp (f ) is a convex set for every p > 0. Moreover, there are some good relations between Lf and the isotropic constant of Kn+1 (f ). The best known bound today is due to Klartag [50] who proved that for any log-concave function f : Rn → R+ , Lf ≤ cn1/4 . Kp (f ) = (3) 1.2 Computing the volume of a convex body In algorithmic convex geometry [42], an important question is to produce algorithm which may compute the volume of a convex body. It is another illustration of a high dimensional phenomena in convex geometry. To start to understand the question, we need to be more precise. We assume that the convex body K is given by a separation oracle. A separation oracle has the property that if you give a point x ∈ Rn then it answers either that x ∈ K or that x /∈ K and it gives a separating hyperplane between x and K . There are powerful negative results in this direction, see [27, 8]. These results show that the volume of a convex body can not be even approximated by any deterministic polynomial algorithm. Be aware that in fixed dimension n, these problems can easily be solved in polynomial time. However, the degree of those polynomials estimating the running time increases fast with n. What I know about this negative result is that it is based on an important theorem of high dimensional convex geometry due independently to Carl-Pa jor [19], Gluskin [38] (cid:115) and Barany-Furedi [8]. It evaluates the volume of the absolute convex hull of N points (cid:1) log (cid:0)1 + N (cid:19)1/n ≤ C (cid:18) Vol Conv {±u1 , . . . , ±uN } u1 , . . . , uN from the unit sphere in Rn , n Vol B n n 2 In the topic of algorithmic convex geometry, a breakthrough has been done by Dyer-Frieze- Kannan [24]. They proved that the situation changed drastically if we allow randomization and they gave the first polynomial randomized algorithm to calculate the volume of a convex body. It exists today a vast literature on the sub ject and we refer to the paper of Kannan, Lov´asz and Simonovits [49] and to [42, 74, 75] for a better presentation of this topic. In the randomized situation, we are given two numbers ε and η in (0, 1), a convex body K and a separation oracle. Then there exists a randomized algorithm returning a non negative number ζ such that (1 − ε)ζ < Vol K < (1 + ε)ζ . with probability at least 1 − η . What is random in this situation is the number of oracle (cid:33) (cid:32) (cid:18) (cid:19)3 (cid:18) (cid:19) calls and they get (for example) a bound on its expected value. In [49], the algorithm uses log5 n log log . O n5 ε2 1 ε 1 η 3 random call to the separation oracle which was a significant improvement of the previous known results. The strategy from [49] may be described in two steps and each of them led to interesting problems in high dimensional convex geometry. The first step is a rounding procedure in order to insure that the convex body and the Euclidean structure are correctly related. It consists to finding a position such that 2 ⊂ K ⊂ d B n B n 2 where d depends polynomially in the dimension n. A theorem of John [47] asserts that there exists an affine transformation T such that B n 2 is the ellipsoid of 2 ⊂ T (K ) ⊂ nB n √ maximal volume contained in T (K ) and in which case, B n 2 (the distance is reduced to n in the case of symmetric convex bodies). However, there is no known randomized polynomial algorithm that can return the affine transformation T such that B n 2 is close to the ellipsoid of maximal volume contained in K . A classical procedure was to use the ellipsoid algorithm (see [42]) which achieves d = O(n3/2 ). The main idea is to consider the inertia ellipsoid E associated to K . It is possible to approximate it by a (cid:1) √ volume of K , (1 − ε)Vol K , is contained in (cid:0)log 1 randomized polynomial time algorithm. Moreover, the geometric distance between K and E is known to be of the order of n and it is not difficult to prove that a big part of the n E . The question that attracted a ε lot of attention in high dimensional convex geometry was to find a good random way to approximate the inertia ellipsoid. This has been solved recently by Adamczak, Litvak, Pa jor, Tomczak-Jaegermann in [3] and we will give a more detailed description of the problem in the paragraph 1.3. The second step consists of computing the volume of a convex body which is in a nearly isotropic position. They apply a multiphase Monte-Carlo algorithm. They need good bounds on the mixing time of the random walk and this is based on good estimate about an isoperimetric inequality. We will not describe the notions of local conductance here and refer to [48, 49]. We will focus on an isoperimetric problem for convex bodies in isotropic position. The question of describing the ”almost” extremal sets in the isoperimetric inequality is still an open problem, known today as the KLS conjecture [48]. We will give a more detailed description of the problem in the paragraph 1.4. 1.3 Approximation of the inertia matrix Let X be a random vector uniformly distributed on a convex body in Rn . The inertia matrix is given by EX ⊗ X . The simplest procedure to approximate the inertia matrix is to understand how many sample is needed to approximate it. Given ε ∈ (0, 1), the question (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ ε(cid:107)EX ⊗ X (cid:107) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 N(cid:88) is to give an estimate of the smallest number N such that N i=1 where (cid:107) · (cid:107) denotes the operator norm from (cid:96)n 2 to (cid:96)n 2 . Since the procedure is random, we would like to have such result with some fixed positive probability, 1 − η . It is clear that without loss of generality, we can assume that the random vector is isotropic, that is EX = 0 and EX ⊗ X = Id. In terms of random processes, the question becomes: evaluate Xi ⊗ Xi − EX ⊗ X 4 (4) sup y∈Sn−1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ ε. N(cid:88) N such that with the highest possible probability, (cid:104)Xj , y(cid:105)2 − 1 N i=1 √ √ Using the language of random matrices, this is nothing else that evaluating N such that N , belong all the singular values of the random matrix A, with rows X1/ N , . . . , XN / to the interval [1 − ε, 1 + ε]. In [49] Kannan, Lov´asz and Simonovits proved that if N is ε2 η2 then (4) holds true with probability larger than 1 − η . Shortly after, of the order of n2 Bourgain [16] improved this estimate to N ≈ n log3 n . During more than fifteen years, several ε2 η2 groups of people [69, 37, 36, 65, 44] proposed different strategies to improve this result. A breakthrough has been done by Adamczak, Litvak, Pa jor and Tomczak-Jaegermann [3, 4] √ ε2 then (4) holds true with probability at least 1 − e−c who proved that if N ≈ n n . The main achievement in the result is that N is taken of the order of the ambient dimension n and that the probability of the event is not only large but increases extremely fast regarding to the dimension n. It is now of interest to understand what other random vector than log- concave probability distribution satisfy such type of result. It has been recently investigated by Srivastava and Vershynin [71] and generalized to the case of random matrices in [77]. 1.4 Almost extremal sets in the isoperimetric inequality The isoperimetric problem for convex bodies is the following. Let K be a convex body in Rn and µ be the uniform measure on K . Let S be a subset of K and define the boundary measure of S as µ+ (S ) = lim inf ε→0 2 ) − µ(S ) µ(S + εB n ε . This definition is also valid for any measure with log-concave density on Rn . The question is to evaluate the largest possible h such that ∀ S ⊂ K, µ+ (S ) ≥ h µ(S )(1 − µ(S )) (5) Without any assumption on the measure, you can easily imagine a situation where h may be as close to 0 as you wish. In our situation, we made the assumption that the measure is isotropic and log-concave. This avoid a lot of non regular situation. Kannan, Lov´asz and Simonovits (the same group of persons in an earlier paper !) [48] considered this problem and conjectured that up to a universal constant in the inequality (5), the worth set S should be a half space of the same measure than S . Conjecture 2. (The KLS conjecture) There exists c > 0 such that for any dimension n and any isotropic log-concave probability on Rn , ∀ S ⊂ Rn , µ+ (S ) ≥ c µ(S )(1 − µ(S )) 5 SK\SeThis is supported by the fact that in the Gaussian setting, it is known since the work of Sudakov, Tsirelson [73] and independently Borell [14] that the half spaces are the exact (cid:90) a+ε solutions of the isoperimetric problem in the Gauss space: −∞ 2 ) = γn (H + εB n γn (S + εB n min 2 ) = γn (S )=α (cid:90) a where dγn (x) = e−x2 2 /2dx/(2π)n/2 and H is a half space of Gaussian measure −∞ e−t2 /2 dt√ 2π e−t2 /2 dt√ 2π α = . The inequality (5) is called a Cheeger type inequality and h is usually referred as the Cheeger’s constant of the measure µ. Kannan, Lov´asz and Simonovits proved that if µ is a log-concave isotropic probability measure on Rn , then ≈ 1√ h ≥ c EX 2 n where X is a random vector distributed according to µ. The argument is based on a localization technique introduced in [41, 60]. The localization method was further developed in [32, 33]. And Bobkov [10] improved this result to h ≥ (6) c (VarX 2 )1/4 During the last decade, this question has been much investigated. However, it remains an open question. Very few positive results are known. It has been proved only for some classes of convex bodies like the unit balls of (cid:96)n p [70, 55] and a weaker form is proved for random Gaussian polytopes in [30]. It is also known from the work of Buser [18] and Ledoux [56] that in the case of log-concave probability, the Cheeger constant is related to the best constant in the Poincar´e inequality. Let X be the random vector distributed according to µ, let D2 be the largest constant such that for every regular function F : Rn → R, D2 VarF (X ) ≤ E∇F (X )2 (7) 2 then h2 ≈ D2 . More surprinsigly, Milman [62] proved that h2 ≈ D∞ where D∞ is the largest constant such that for every 1-Lipschitz function F : Rn → R, VarF (X ) ≤ 1 D∞ . Both inequalities are easy consequences of Conjecture 2. The difficult part of the proof of these results concern the reverse statement. A recent paper of Gozlan, Roberto and Samson [40] completes also the picture of the different equivalent formulations of the question. We refer to Chapter 3 in [57] and to [41] for a more detailed description of the links between the Poincar´e inequality and concentration of measure and we just emphasize on the fact that Conjecture 2 implies a very strong concentration inequality of the Euclidean norm. 6 Conjecture 3. (The thin shell conjecture) There exists c > 0 such that for any (cid:12)(cid:12)(cid:12)(cid:12) ≥ t (cid:18)(cid:12)(cid:12)(cid:12)(cid:12)X 2 − √ (cid:19) log-concave isotropic probability on Rn , for any t > 0, √ n n It is the purpose of the next section to describe our knowledge about this question and the different aspects of this problem in high dimensional convex geometry. ≤ 2e−c t P √ n − sup t∈R 2 The thin shell concentration and a Berry-Esseen type theorem for convex bodies The classical Berry-Esseen bounds of the central limit theorem asserts that for every in- (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ τ θ2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)P (cid:32) n(cid:88) (cid:33) (cid:90) t dependent random variables x1 , . . . , xn such that Exi = 0, Ex2 i = 1 and Ex3 i = τ , we have e−u2 /2 du√ ∀θ ∈ S n−1 , θixi ≤ t 4 . −∞ 2π √ √ √ i=1 n), we have θ2 n, . . . , 1/ Observe that if θ = (1/ n which gives a very good rate of 4 = 1/ convergence. At the end of the nineties, Ball (in several talks and in [5]) and independently Brehm and Voigt in [17] asked about a possible generalization of the Berry-Esseen Theorem for convex bodies. The question can be stated as follows: for every isotropic convex body (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ αn (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)P (cid:33) (cid:32) n(cid:88) K ⊂ Rn , does there exist a direction θ ∈ S n−1 such that (cid:90) t e−u2 /2 du√ sup t∈R −∞ 2π i=1 with lim+∞ αn = 0, where P is the uniform distribution on K ? The same question can be asked for every isotropic probability P with log-concave density on Rn . A very good presentation of the problem is done in [5]. Indeed the authors propose a satisfactory way to solve the problem and this generated an intense activity during the last (cid:18)(cid:12)(cid:12)(cid:12)(cid:12) X 2√ (cid:12)(cid:12)(cid:12)(cid:12) ≥ εn (cid:19) decade. Anttila, Ball and Perssinaki [5] proved that if there exists a sequence (εn )n≥1 such that lim+∞ εn = 0 and − 1 ≤ εn (9) n for every isotropic probability with log-concave density on Rn , then the Berry-Esseen The- orem for convex bodies holds true. More precisely, they proved that if (9) is true for an isotropic probability P uniformly distributed on a convex body (or with log-concave den- sity) then there exists a set A of directions θ ∈ S n−1 of extremely large probability such that (8) holds true. They also give an affirmative answer to (9) for some classes of uniform p for all p ≥ 1. From a more measures like the uniform measures on the unit ball of (cid:96)n probabilistic point of view, the principle of studying a thin shell estimate to get a Central Limit Theorem appeared already in [76, 23]. More precise formulation of Berry-Esseen type theorem for isotropic random vector satisfying (9) have been established by Bobkov θixi ≤ t − (8) P 7 [9]. The fact that under the isotropicity condition, for most of the directions θ ∈ S n−1 , the random variables (cid:104)X, θ(cid:105) have a common behavior was already observed in [72]. A complete solution for the Central Limit Problem for convex bodies is given by Klartag in [51], while shortly after, an independent proof of (9) was given in [31]. In these papers, the rates for αn and εn are weak, only of the order of some negative power of log n. In fact, the inequality (9) is equivalent to = 0 (10) VarX 2 lim n→∞ n where X is distributed according to the log-concave measure P and this is known to be true in full generality [51, 31]. As it was emphasized by Bobkov and Koldobsky [11], if we look at the particular case of F being · 2 2 in the conjecture of Kannan, Lov´asz and Simonovits, we deduce from inequality (7) the following conjecture. Conjecture 4. (The Variance conjecture) There exists a constant C such that for (cid:19) (cid:0)EX 2 (cid:18) every isotropic log-concave random vector X , we have VarX 2 2 ≤ C EX 2 (cid:1)1/2 . (cid:1)1/4 ≤ (cid:0)EX 4 2 or equivalently 1 + 2 2 It seems to be a very natural conjecture. Denoting X = (x1 , . . . , xn ), it concerns the 1 + . . . + x2 behavior of the random variable x2 n , where here, the classical hypotheses on the entries of X , independence and variance 1, are replaced by the isotropy and log-concavity of the vector. Taking F as the Euclidean norm in (7), this gives another simple conjecture C n Conjecture 5. (The Variance conjecture - bis) There exists a constant C such that for every isotropic log-concave random vector X , VarX 2 ≤ C It is clear that these conjectures are much stronger than inequality (10). Let us come back to Conjecture 3. The stated concentration inequality contains infor- mations for all t > 0 and the discussion is different for t > 10, for t close to 1 and for √ t < 1/10. Let us start to discuss about the large deviation, t > 10. From Borell’s lemma n) ≤ e−ct . [13] (see also [64] Appendix 3), it is easy to deduce that for t > 10, P(X 2 ≥ t But it was proved in [12] that in the case of uniform measure on an unconditional isotropic √ convex body, this can be improved to e−ct n . In this topic, a breakthrough has been done by Paouris [65]. Shortly after, he proved in [66] a small ball inequality which corresponds to the case of t being close to 1. Theorem 2 (Paouris [65, 66]). Let P be an isotropic probability on Rn with log-concave density. Then for every t ≥ 10, And for every ε ∈ (0, 1/10), √ n . n) ≤ e−ct √ √ n) ≤ (cε) n √ P(X 2 ≥ t P(X 2 ≤ ε 8 An important ingredient in his proof is the study of the volume of the sets Kp (f ) defined in (3), the Keith Ball’s bodies introduced in [6] to prove Theorem 1. Once the thin shell estimate (9) has been proved by Klartag [51], see also [31], it was natural to study the rate of convergence. The results have been improved to polynomial estimates in the dimension n by Klartag [52], by Fleury [29] and lastly by E. Milman and myself [43]. As of today, the best known result is the following n(cid:12)(cid:12) ≥ t ∀t ≥ 0, P (cid:0)(cid:12)(cid:12)X 2 − √ n(cid:1) ≤ C exp(−c Theorem 3 (Gu´edon-Milman [43]). √ √ n min(t3 , t)) From this result we deduced that VarX 2 2 ≤ C n5/3 and using (6), this gives a general bound on the Cheeger’s constant of an isotropic log-concave measure h ≥ c n−5/12 . A new relation between the thin shell estimate and the Cheeger’s constant has been recently developed by Eldan [25] where he deduced h ≥ c n−1/3 (log n)−1/2 from Theorem 3. It is time to draw a picture of a high dimensional convex body. This picture was popularized by Vitali Milman. Observe in particular that it is important to draw the convex set as a star shape body with a lot of points very far from the origin and lot of √ points very close to the origin. Think to the example of the cube whose 2n vertices are at n of the origin and whose 2n middle of faces are at Euclidean distance Euclidean distance 1. A high dimensional isotropic convex body. The volume is concentrated in the thin shell of radius n1/2 and width n1/2−1/6 drawn in blue. Since the volume of the convex body is highly concentrated in the blue part of the picture, it maybe explains the difficulty of computing it with some algorithm because you need to generate a random walk in several disconnected parts. The concentration of the mass of a log-concave measure in a Euclidean ball or Euclidean shell is usually understood via the study of the Lp norms of the random variable X 2 . A new proof of the first part of Theorem 2 has been recently given in [2]. It is worth noticing that the general result can be expressed without the isotropic hypothesis. It gives a deep relation of the strong moments of X 2 with its weak moments defined by E(cid:104)z , X (cid:105)p . σ p p (X ) = supz 2≤1 9 (11) Theorem 4 ([65, 2]). There exist c, C > 0 such that for every log-concave random vector X , we have (EX p ∀p ≥ 1, 2 )1/p ≤ C EX 2 + c σp (X ) where σp (X ) = supz 2≤1 (E(cid:104)z , X (cid:105)p )1/p . √ Observe that in isotropic position, EX 2 ≤ (EX 2 2 )1/2 = n. By Borell’s lemma [13] (E(cid:104)z , X (cid:105)p )1/p ≤ C p (cid:0)E(cid:104)z , X (cid:105)2(cid:1)1/2 = C p z 2 (see also [64] Apppendix 3), we know that ∀p ≥ 1, √ √ 2 )1/p ≤ C Hence by Theorem 4, for all p ≥ 1, (EX p ∀t ≥ 1, P (cid:0)X 2 ≥ t n(cid:1) ≤ e−c t n + cp. Take p = t n and Markov √ inequality gives √ This is exactly the first part of Theorem 2. In [43] we also investigated the Lp norms of X 2 . Looking at Conjecture 4 and at the statement of the previous Theorem, we should be able to prove the following weaker conjecture. n . Conjecture 6. (The weak thin shell conjecture) There exists c > 0 such that for every log-concave random vector X , we have (EX p ∀p ≥ 1, 2 )1/p ≤ EX 2 + c σp (X ) (cid:19) (cid:18) This would prove that in isotropic position (EX p 2 )1/p ≤ EX 2 ∀p ≥ 1, 1 + c . p√ n Recently, Eldan and Klartag [26] established a surprising connection between thin shell esti- mates and the slicing problem. They proved that if Conjecture 4 holds true then Conjecture 1 will be also true. It is also known [7] that for an individual log-concave distribution µ, the isotropic constant of µ is bounded by a function of the Cheeger’s constant of µ. Finally, Eldan [25] proved that Conjecture 4 implies Conjecture 2 up to polylogarithmic term. To conclude this paragraph, I would like to advertise that Conjecture 3 is known to be true when the measure is uniformly distributed on the unit ball of a generalized Orlicz space [28] and that Conjecture 4 is valid when the measure is uniformly distributed on an unconditional convex body [53]. 3 Weak and strong moments of a random vector In this section, we will go away from the framework defined by log-concave distributions. We have seen that the thin shell conjectures for log-concave measures are closely related to the comparison of the strong moments of the Euclidean norm with its weak moments. There are more probabilistic questions in this direction. It is natural to consider for which family of random vectors X , we have for any norm (cid:107) · (cid:107) (E(cid:107)X (cid:107)p )1/p ≤ C E(cid:107)X (cid:107) + c sup (E(cid:104)z , X (cid:105)p )1/p (cid:107)z(cid:107)(cid:63)≤1 10 where (cid:107) · (cid:107)(cid:63) is the dual norm of (cid:107) · (cid:107) and C, c > 0 are numerical constants. It is known to be true with C = 1 for Gaussian [20] or Rademacher (see [58] Theorem 4.7) random vectors series. We refer to [55] and [54] where such questions are discussed. In the area of log-concave measures, Lata(cid:32)la [54] asked the following. Conjecture 7. (Weak and strong moments conjecture) There exist C, c > 0 such that for any log-concave random vector X and any norm (cid:107) · (cid:107) we have (E(cid:107)X (cid:107)p )1/p ≤ C E(cid:107)X (cid:107) + c sup (E(cid:104)z , X (cid:105)p )1/p (cid:107)z(cid:107)(cid:63)≤1 where (cid:107) · (cid:107)(cid:63) is the dual norm of (cid:107) · (cid:107). Observe that Theorem 4 tells that this is true for the Euclidean norm. From now on, we will present a few of the results from a joint work with Adamczak, Lata(cid:32)la, Litvak, Pa jor and Tomczak-Jaegermann [1]. We introduce a new class of random vectors. Let p > 0, m = (cid:100)p(cid:101), and λ ≥ 1. We say that a random vector X in a Banach space E satisfies the assumption H (p, λ) if for every linear mapping A : E → Rm such that Y = AX is non-degenerate there exists a gauge (cid:107) · (cid:107) on Rm such that E(cid:107)Y (cid:107) < ∞ and (E(cid:107)Y (cid:107)p )1/p ≤ λ E(cid:107)Y (cid:107). Be aware that in the definition, we chose m = (cid:100)p(cid:101). And observe that any m-dimensional norm may be approximated by a set {ϕi} of 3m linear forms which means that for any (cid:19)1/p ≈ sup (cid:18) norm in Rm , we have (Eϕ(Y )p )1/p . E sup (cid:107)ϕ(cid:107)(cid:63)≤1 i=1,...,3m p (cid:0)Eϕ(Y )2(cid:1)1/2 . We say that a random vector X ∈ E is ψ2 with constant ψ if it satisfies √ ∀p ≥ 2, ∀φ ∈ E ∗ , (Eϕ(Y )p )1/p ≤ ψ From the previous remark, we can easily deduce that ψ2 random vectors satisfy the hypoth- esis H (p, C ψ2 ) for every p with a universal constant C . This implies that any Gaussian or Rademacher random series of vectors satisfy the hypothesis H (p, C ) for every p. The first main result from [1] is Theorem 5. Let p > 0 and λ ≥ 1. If a random vector X satisfies H (p, λ) then 2 )1/p ≤ c (λEX 2 + σp (X )) (EX p (E(cid:107)Y (cid:107)p )1/p ≈ ϕi (Y )p where c is a universal constant. Proof. A very rough idea of the proof is the following. Let X be the random vector in E satisfying H (p, λ), m = (cid:100)p(cid:101), λ ≥ 1. There are three important steps in the proof. By Gaussian Concentration [20], we know that for a standard Gaussian vector G √ (EGEX (cid:104)G, X (cid:105)p )1/p ≤ EG (EX (cid:104)G, X (cid:105)p )1/p + c p σp (X ) (12) 11 We define a new norm on E by (cid:107)z(cid:107) = (EX (cid:104)z , X (cid:105)p )1/p . It is nothing else than the dual norm of the classical Zp bodies [65]. By Gordon min-max theorem [39], we know that if A : (E , (cid:107) · (cid:107)) → Rm ∼ N (0, Id) is a standard Gaussian matrix, √ we have EG (EX (cid:104)G, X (cid:105)p )1/p ≤ EA minz 2=1 (EX (cid:104)z , AX (cid:105)p )1/p + c p σp (X ). The key property H (p, λ) allows to prove the following nice geometric Lemma (EX (cid:104)z , AX (cid:105)p )1/p ≤ λ EX AX 2 . minz 2=1 (13) (14) To conclude, it remains to glue the argument. Observe that 2 )1/p ≈ 1√ (EX p (EGEX (cid:104)G, X (cid:105)p )1/p . p By (12) and (13), we get (EGEX (cid:104)G, X (cid:105)p )1/p (cid:46) 1√ (EX p 2 )1/p ≈ 1√ p p (EX (cid:104)z , AX (cid:105)p )1/p + σp (X ). EA minz 2=1 From (14), we conclude that (EGEX (cid:104)G, X (cid:105)p )1/p ≤ 1√ 2 )1/p ≈ 1√ (EX p (EX (cid:104)z , AX (cid:105)p )1/p + σp (X ) EA minz 2=1 p p EA λ EX AX 2 + σp (X ) (cid:46) λ EX 2 + σp (X ). (cid:46) 1√ p This is the announced result. Of course, we need to understand which random vectors satisfy the hypothesis H (p, λ) with p as large as possible and λ being a constant. It is not only the case for Gaussian, Rademacher random series of vectors, or ψ2 random vectors. It is satisfied by log-concave random vectors (as expected) and more generally by s-concave random vectors for nega- tive s. The class of s-concave random vectors is very general. For example, the Cauchy distributions belong to it. It has been widely studied in the seventies and there is a serie of papers of Borell [15, 13] who characterized it. Let s < 1/n. A probability Borel measure µ on Rn is called s-concave if for every compact sets A, B ⊂ Rn , for every θ ∈ [0, 1], µ((1 − θ)A + θB ) ≥ ((1 − θ)µ(A)s + θµ(B )s )1/s whenever µ(A)µ(B ) > 0. For s = 0, this corresponds to log-concave measures. Borell [15, 13] characterized the class of s-concave measures and proved that any s-concave probability is supported on some convex subset of an affine subspace where it has a density. Assuming the measure has full dimensional support, the density is a γ -concave function 12 s − n. We will concentrate on the case of s-concave random vectors with s < 0. γ = 1 with 1 Denote s = −1/r. Restating the characterization of Borell, we get that when the support generates the whole space, the s-concave measure has a density g of the form g = f −β with β = n + r and f is a positive convex function on Rn . A classical example is to define g from a norm (cid:107) · (cid:107) on Rn : g(x) = c(1 + (cid:107)x(cid:107))−n−r , r > 0. The class of s-concave measures satisfy several important properties. It is decreasing in s which gives for example that any log-concave probability measure is (−1/r)-concave for any r > 0. The linear image of a (−1/r)-concave vector is also (−1/r)-concave. And the Euclidean norm of a (−1/r)-concave random vector has moments of order 0 < p < r . Our main second result [1] is Theorem 6. Let r ≥ 2 and X be a (−1/r)-concave random vector. Then for every 0 < p < r/2, X satisfies the assumption H (p, C ), C being a universal constant. We refer to [1] for its proof. With Theorem 5, we get Theorem 7. Let r ≥ 2 and X be a (−1/r)-concave random vector. Then for every 0 < p < r/2, (EX p 2 )1/p ≤ C (EX 2 + σp (X )). It is not difficult to deduce nice concentration properties of the random variable X 2 when X is an isotropic (−1/r)-concave random vector in Rn . Corollary 1. Let r ≥ 2 and X be an isotropic (−1/r)-concave random vector in Rn . Then (cid:18) c max(1, r/ (cid:19)r/2 n(cid:1) ≤ P(cid:0)X 2 > t √ for every t > 0, √ t where c is a universal constant. n) This concentration property is central in the work of Srivastava and Vershynin [71] about approximation of the covariance matrix. This is why we can go back to the problem studied in section 1.3 about the approximation of the inertia matrix. Corollary 2. Let r ≥ log n and X be a (−1/r)-concave isotropic random vector. Let (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ ε. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 X1 , . . . , XN be independent copies of X . Then for every ε ∈ (0, 1) and every N ≥ C (ε)n, N(cid:88) one has N i=1 Of course, we stated only a few of the results from [1] and refer to it for more general and precise formulations. In particular, we also got some interesting small ball properties in the spirit of the results of Paouris [66]. We also asked if some thin shell concentration could be proved for −1/r-concave random vector and this was done very recently in [34]. Xi ⊗ Xi − I E 13 References [1] Adamczak, R., Gu´edon, O., Lata(cid:32)la, R., Litvak, A. E., Oleszkiewicz, K., Pajor, A., and Tomczak-Jaegermann, N. Moment estimates for convex measures. Electron. J. Probab. 17 (2012), 1–19. [2] Adamczak, R., Lata(cid:32)la, R., Litvak, A. E., Oleszkiewicz, K., Pajor, A., and Tomczak- Jaegermann, N. A short proof of paouris’ inequality. Can. Math. Bul. (to appear). [3] Adamczak, R., Litvak, A. E., Pajor, A., and Tomczak-Jaegermann, N. Quantitative esti- mates of the convergence of the empirical covariance matrix in log-concave ensembles. J. Amer. Math. Soc. 23, 2 (2010), 535–561. [4] Adamczak, R., Litvak, A. E., Pajor, A., and Tomczak-Jaegermann, N. Sharp bounds on the rate of convergence of the empirical covariance matrix. C. R. Math. Acad. Sci. Paris 349, 3-4 (2011), 195–200. [5] Anttila, M., Ball, K., and Perissinaki, I. The central limit problem for convex bodies. Trans. Amer. Math. Soc. 355, 12 (2003), 4723–4735 (electronic). [6] Ball, K. Logarithmically concave functions and sections of convex sets in Rn . Studia Math. 88, 1 (1988), 69–84. [7] Ball, K., and Nguyen, V. H. Entropy jumps for random vectors with log-concave density and spectral gap. http://arxiv.org/abs/1206.5098 (preprint). [8] B´ar´any, I., and Furedi, Z. Computing the volume is difficult. Discrete Comput. Geom. 2, 4 (1987), 319–326. [9] Bobkov, S. G. On concentration of distributions of random weighted sums. Ann. Probab. 31, 1 (2003), 195–215. [10] Bobkov, S. G. On isoperimetric constants for log-concave probability distributions. In Geometric aspects of functional analysis, vol. 1910 of Lecture Notes in Math. Springer, Berlin, 2007, pp. 81–88. [11] Bobkov, S. G., and Koldobsky, A. On the central limit property of convex bodies. In Geometric aspects of functional analysis, vol. 1807 of Lecture Notes in Math. Springer, Berlin, 2003, pp. 44–52. [12] Bobkov, S. G., and Nazarov, F. L. On convex bodies and log-concave probability measures with unconditional basis. In Geometric aspects of functional analysis, vol. 1807 of Lecture Notes in Math. Springer, Berlin, 2003, pp. 53–69. [13] Borell, C. Convex measures on locally convex spaces. Ark. Mat. 12 (1974), 239–252. [14] Borell, C. The Brunn-Minkowski inequality in Gauss space. Invent. Math. 30, 2 (1975), 207–216. [15] Borell, C. Convex set functions in d-space. Period. Math. Hungar. 6, 2 (1975), 111–136. [16] Bourgain, J. Random points in isotropic convex sets. In Convex geometric analysis (Berkeley, CA, 1996), vol. 34 of Math. Sci. Res. Inst. Publ. Cambridge Univ. Press, Cambridge, 1999, pp. 53–58. [17] Brehm, U., and Voigt, J. Asymptotics of cross sections for convex bodies. Beitrage Algebra Geom. 41, 2 (2000), 437–454. [18] Buser, P. A note on the isoperimetric constant. Ann. Sci. ´Ecole Norm. Sup. (4) 15, 2 (1982), 213–230. [19] Carl, B., and Pajor, A. Gel(cid:48) fand numbers of operators with values in a Hilbert space. Invent. Math. 94, 3 (1988), 479–504. [20] Cirel(cid:48) son, B. S., Ibragimov, I. A., and Sudakov, V. N. Norms of Gaussian sample functions. In Proceedings of the Third Japan-USSR Symposium on Probability Theory (Tashkent, 1975) (Berlin, 1976), Springer, pp. 20–41. Lecture Notes in Math., Vol. 550. 14 [21] Das Gupta, S. Brunn-Minkowski inequality and its aftermath. J. Multivariate Anal. 10, 3 (1980), 296–318. [22] Davidovic, J. S., Korenbljum, B. I., and Hacet, B. I. A certain property of logarithmically concave functions. Dokl. Akad. Nauk SSSR 185 (1969), 1215–1218. [23] Diaconis, P., and Freedman, D. Asymptotics of graphical pro jection pursuit. Ann. Statist. 12, 3 (1984), 793–815. [24] Dyer, M., Frieze, A., and Kannan, R. A random polynomial-time algorithm for approximating the volume of convex bodies. J. Assoc. Comput. Mach. 38, 1 (1991), 1–17. [25] Eldan, R. Thin shell implies spectral gap up to polylog via a stochastic localization scheme. http://arxiv.org/abs/1203.0893 (preprint). [26] Eldan, R., and Klartag, B. Approximately Gaussian marginals and the hyperplane conjecture. In Concentration, functional inequalities and isoperimetry, vol. 545 of Contemp. Math. Amer. Math. Soc., Providence, RI, 2011, pp. 55–68. [27] Elekes, G. A geometric inequality and the complexity of computing volume. Discrete Comput. Geom. 1, 4 (1986), 289–292. [28] Fleury, B. Between Paouris concentration inequality and variance conjecture. Ann. Inst. Henri Poincar´e Probab. Stat. 46, 2 (2010), 299–312. [29] Fleury, B. Concentration in a thin Euclidean shell for log-concave measures. J. Funct. Anal. 259, 4 (2010), 832–841. [30] Fleury, B. Poincar´e inequality in mean value for Gaussian polytopes. Probab. Theory Related Fields 152, 1-2 (2012), 141–178. [31] Fleury, B., Gu´edon, O., and Paouris, G. A stability result for mean width of Lp -centroid bodies. Adv. Math. 214, 2 (2007), 865–877. [32] Fradelizi, M., and Gu´edon, O. The extreme points of subsets of s-concave probabilities and a geometric localization theorem. Discrete Comput. Geom. 31, 2 (2004), 327–335. [33] Fradelizi, M., and Gu´edon, O. A generalized localization theorem and geometric inequalities for convex bodies. Adv. Math. 204, 2 (2006), 509–529. [34] Fradelizi, M., Gu´edon, O., and Pajor, A. Spherical thin-shell concentration for convex measures. [35] Gardner, R. J., Koldobsky, A., and Schlumprecht, T. An analytic solution to the Busemann- Petty problem on sections of convex bodies. Ann. of Math. (2) 149, 2 (1999), 691–703. [36] Giannopoulos, A., Hartzoulaki, M., and Tsolomitis, A. Random points in isotropic uncon- ditional convex bodies. J. London Math. Soc. (2) 72, 3 (2005), 779–798. [37] Giannopoulos, A. A., and Milman, V. D. Concentration property on probability spaces. Adv. Math. 156, 1 (2000), 77–106. [38] Gluskin, E. D. Extremal properties of orthogonal parallelepipeds and their applications to the geometry of Banach spaces. Mat. Sb. (N.S.) 136(178), 1 (1988), 85–96. [39] Gordon, Y. Some inequalities for Gaussian processes and applications. Israel J. Math. 50, 4 (1985), 265–289. [40] Gozlan, N., Roberto, C., and Samson, P.-M. From dimension free concentration to poincar´e inequality. http://arxiv.org/abs/1305.4331 (preprint). [41] Gromov, M., and Milman, V. D. Generalization of the spherical isoperimetric inequality to uniformly convex Banach spaces. Compositio Math. 62, 3 (1987), 263–282. [42] Grotschel, M., Lov´asz, L., and Schrijver, A. Geometric algorithms and combinatorial opti- mization, second ed., vol. 2 of Algorithms and Combinatorics. Springer-Verlag, Berlin, 1993. 15 [43] Gu´edon, O., and Milman, E. Interpolating thin-shell and sharp large-deviation estimates for isotropic log-concave measures. Geom. Funct. Anal. 21, 5 (2011), 1043–1068. [44] Gu´edon, O., and Rudelson, M. Lp -moments of random vectors via ma jorizing measures. Adv. Math. 208, 2 (2007), 798–823. [45] Hensley, D. Slicing convex bodies—bounds for slice area in terms of the body’s covariance. Proc. Amer. Math. Soc. 79, 4 (1980), 619–625. [46] Henstock, R., and Macbeath, A. M. On the measure of sum-sets. I. The theorems of Brunn, Minkowski, and Lusternik. Proc. London Math. Soc. (3) 3 (1953), 182–194. [47] John, F. Extremum problems with inequalities as subsidiary conditions. In Studies and Essays Presented to R. Courant on his 60th Birthday, January 8, 1948. Interscience Publishers, Inc., New York, N. Y., 1948, pp. 187–204. [48] Kannan, R., Lov´asz, L., and Simonovits, M. Isoperimetric problems for convex bodies and a localization lemma. Discrete Comput. Geom. 13, 3-4 (1995), 541–559. [49] Kannan, R., Lov´asz, L., and Simonovits, M. Random walks and an O∗ (n5 ) volume algorithm for convex bodies. Random Structures Algorithms 11, 1 (1997), 1–50. [50] Klartag, B. On convex perturbations with a bounded isotropic constant. Geom. Funct. Anal. 16, 6 (2006), 1274–1290. [51] Klartag, B. A central limit theorem for convex sets. Invent. Math. 168, 1 (2007), 91–131. [52] Klartag, B. Power-law estimates for the central limit theorem for convex sets. J. Funct. Anal. 245, 1 (2007), 284–310. [53] Klartag, B. A Berry-Esseen type inequality for convex bodies with an unconditional basis. Probab. Theory Related Fields 145, 1-2 (2009), 1–33. [54] Lata(cid:32)la, R. Weak and strong moments of random vectors. In Marcinkiewicz centenary volume, vol. 95 of Banach Center Publ. Polish Acad. Sci. Inst. Math., Warsaw, 2011, pp. 115–121. [55] Lata(cid:32)la, R., and Wojtaszczyk, J. O. On the infimum convolution inequality. Studia Math. 189, 2 (2008), 147–187. [56] Ledoux, M. A simple analytic proof of an inequality by P. Buser. Proc. Amer. Math. Soc. 121, 3 (1994), 951–959. [57] Ledoux, M. The concentration of measure phenomenon, vol. 89 of Mathematical Surveys and Mono- graphs. American Mathematical Society, Providence, RI, 2001. [58] Ledoux, M., and Talagrand, M. Probability in Banach spaces. Classics in Mathematics. Springer- Verlag, Berlin, 2011. Isoperimetry and processes, Reprint of the 1991 edition. [59] Leindler, L. On a certain converse of Holder’s inequality. II. Acta Sci. Math. (Szeged) 33, 3-4 (1972), 217–223. [60] Lov´asz, L., and Simonovits, M. Random walks in a convex body and an improved volume algorithm. Random Structures Algorithms 4, 4 (1993), 359–412. [61] Maurey, B. In´egalit´e de Brunn-Minkowski-Lusternik, et autres in´egalit´es g´eom´etriques et fonction- nelles. Ast´erisque, 299 (2005), Exp. No. 928, vii, 95–113. S´eminaire Bourbaki. Vol. 2003/2004. [62] Milman, E. On the role of convexity in isoperimetry, spectral gap and concentration. Invent. Math. 177, 1 (2009), 1–43. [63] Milman, V. D., and Pajor, A. Isotropic position and inertia ellipsoids and zonoids of the unit ball of a normed n-dimensional space. In Geometric aspects of functional analysis (1987–88), vol. 1376 of Lecture Notes in Math. Springer, Berlin, 1989, pp. 64–104. 16 [64] Milman, V. D., and Schechtman, G. Asymptotic theory of finite-dimensional normed spaces, vol. 1200 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. With an appendix by M. Gromov. [65] Paouris, G. Concentration of mass on convex bodies. Geom. Funct. Anal. 16, 5 (2006), 1021–1049. [66] Paouris, G. Small ball probability estimates for log-concave measures. Trans. Amer. Math. Soc. 364, 1 (2012), 287–308. [67] Pr´ekopa, A. Logarithmic concave measures with application to stochastic programming. Acta Sci. Math. (Szeged) 32 (1971), 301–316. [68] Pr´ekopa, A. On logarithmic concave measures and functions. Acta Sci. Math. (Szeged) 34 (1973), 335–343. [69] Rudelson, M. Random vectors in the isotropic position. J. Funct. Anal. 164, 1 (1999), 60–72. [70] Sodin, S. An isoperimetric inequality on the lp balls. Ann. Inst. Henri Poincar´e Probab. Stat. 44, 2 (2008), 362–373. [71] Srivastava, N., and Vershynin, R. Covariance estimation for distributions with 2 + ε moments. Ann. Prob. (to appear). [72] Sudakov, V. N. Typical distributions of linear functionals in finite-dimensional spaces of high dimension. Dokl. Akad. Nauk SSSR 243, 6 (1978), 1402–1405. [73] Sudakov, V. N., and Cirel(cid:48) son, B. S. Extremal properties of half-spaces for spherically invariant measures. Zap. Naucn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 41 (1974), 14–24, 165. Problems in the theory of probability distributions, II. [74] Vempala, S. Geometric random walks: a survey. In Combinatorial and computational geometry, vol. 52 of Math. Sci. Res. Inst. Publ. Cambridge Univ. Press, Cambridge, 2005, pp. 577–616. [75] Vempala, S. S. Recent progress and open problems in algorithmic convex geometry. In 30th Interna- tional Conference on Foundations of Software Technology and Theoretical Computer Science, vol. 8 of LIPIcs. Leibniz Int. Proc. Inform. Schloss Dagstuhl. Leibniz-Zent. Inform., Wadern, 2010, pp. 42–64. [76] von Weizsacker, H. Sudakov’s typical marginals, random linear functionals and a conditional central limit theorem. Probab. Theory Related Fields 107, 3 (1997), 313–324. [77] Youssef, P. Estimating the covariance of random matrices. (preprint). http://arxiv.org/abs/1301.6607 Olivier Gu´edon, Universit´e Paris-Est Laboratoire d’Analyse et Math´ematiques Appliqu´ees (UMR 8050). UPEMLV, F-77454, Marne-la-Vall´ee, France e-mail: [email protected] 17
1910.08996	1	1910	2019-10-20T14:45:43	Sobolev Anisotropic inequalities with monomial weights	[ "math.FA" ]	We derive some anisotropic Sobolev inequalities in $\mathbb{R}^{n}$ with a monomial weight in the general setting of rearrangement invariant spaces. Our starting point is to obtain an integral oscillation inequality in multiplicative form.	math.FA	math	SOBOLEV ANISOTROPIC INEQUALITIES WITH MONOMIAL WEIGHTS F. FEO, J. MARTIN, AND M. R. POSTERARO Abstract. We derive some anisotropic Sobolev inequalities in Rn with a monomial weight in the general setting of rearrangement invariant spaces. Our starting point is to obtain an integral oscillation inequality in multiplicative form. 1. Introduction The study of functional and geometric inequalities with monomial weights, i.e. weights defined by dµ(x) := xAdx = x1A1 · · · xnAn dx. (1.1) where A = (A1, A2, . . . , An) is a vector in Rn with Ai ≥ 0 for i = 1, . . . , n, have been considered extensively recently (see for example [11], [12], [3], [9], [30] and the references quoted therein). The interest for this kind of problems appears when Cabr´e and Ros-Oton (motivated by an open question raised by Haim Brezis [6],[7]) studied in [13] the problem of the regularity of stable solutions to reaction-diffusion problems of double revolution in R2. A function u has symmetry of double revolution if u(x, y) = u(x, y), with (x, y) ∈ RD = RA1+1 × RA2+1 (Ai are positive integers), i.e. the function u can be seen as a suitable function in R2, and it is here where the Jacobian x1A1x2A2 appears (see [13] for the details). In [12], the authors established a sharp isoperimetric inequality in (Rn, µ) (see also [9]) which allows them to obtain the following weighted Sobolev inequality. Theorem 1.1. ([12, Theorem 1.3]) Let µ be defined in (1.1), let (1.2) and 1 ≤ p < D. Then for any f ∈ C1 D = n + A1 + · · · + An c (Rn) we have1 (1.3) for some positive constant C, where f p∗ (cid:18)ZRn dµ(cid:19)1/p∗ ≤ C(cid:18)ZRn ∇f pdµ(cid:19)1/p (1.4) p∗ = Dp D − p . As in the unweighted case a scaling argument shows that the exponent p∗ is optimal, in the sense that (1.4) can not hold with any other exponent. Moreover the exponent p∗ is exactly the same as in the classical Sobolev inequality, but in 2000 Mathematics Subject Classification. Primary: 46E30, 26D10. Key words and phrases. anisotropic inequalities, rearrangement invariant space, Sobolev embedding. 1C 1 c (Rn) denotes the space of C 1 functions with compact support in Rn. 1 2 F. FEO, J. MARTIN, AND M. R. POSTERARO this case the "dimension" is given by D (instead of n). If A1 = ... = An = 0, then exponent p∗ and inequality (1.3) are exactly the classical ones. We observe that when p > 1 and Ai < p − 1 for all i = 1, · · · , n the weight in (1.1) belongs to the Muckenhoupt class Ap, i.e. but, in general the monomial weight does not satisfy the Muckenhoupt condition. The main purpose of this paper is to obtain some anisotropic Sobolev inequalities on Rn with monomial weight xA in the general setting of rearrangement invariant spaces (e.g. Lp, Lorentz, Orlicz, Lorentz-Zygmund, etc...). To this end, we will use the "symmetrization by truncation principle", developed by Milman-Mart´ın in [26] (see also [27] and [28]). This method will provide us a family of rearrangement point- wise inequalities between the special difference2 Oµ(f, t) := f ∗∗ µ(t) (called the oscillation of f ) and the product of the rearrangements of the partial derivative of f (see Theorem 3.1 below) that will be the key to obtain anisotropic inequalities. More precisely we will prove that inequality (1.3) with p = 1 is equivalent to the following oscillation inequality: µ (t) − f ∗ (1.5) Z t D (cid:17)∗ 0 (cid:16)Oµ(f, ·) (·)− 1 (s)ds (cid:22)Z t 0 n Yi=1" d dsZ{f >f ∗ µ(s)} fxi dµ!∗ (τ )# Ai +1 D dτ for every t > 0. The rearrangements without subscript µ are rearrangement with respect to Lebesgue measure on (0, ∞), fxi = ∂f and symbol f (cid:22) g means that ∂xi there exists an universal constant c (independent of all parameters involved) such that f ≤ cg. Inequality (1.5) contains the basic information to obtain anisotropic Sobolev in- equalities on rearrangement invariant spaces, since given a rearrangement invariant space X on (Rn, µ), Hardy's inequality (see (2.3) below) implies3 (1.6) Oµ(f, t)t− 1 (cid:13)(cid:13)(cid:13) D(cid:13)(cid:13)(cid:13) ¯X n (cid:22)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Yi=1" d dsZ{f >f ∗ µ(s)} Ai +1 D fxi dµ!∗ (τ )# . (t)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ¯X For example, given p1, · · · , pn ≥ 1, let p be the weighted harmonic mean between p1, · · · , pn, i.e. (1.7) 1 p = 1 D then (1.6) implies (see Theorem 4.3 below) n Xi=1 , Ai + 1 pi n (1.8) kOµ(f, t)t− 1 D k ¯X (p) (cid:22) kfxik Ai+1 D X (pi) , Yi=1 where X (p) = {f : f p ∈ X} endowed with the norm kf kX (p) = kf pk1/p X . In the particular case that X = L1 and p < D, then (1.8) becomes (see Propo- sition 4.5 below) kf kLp∗ µ (cid:22) n Yi=1 kfxik Ai+1 D Lpi µ ∀f ∈ C1 c (Rn). 2f ∗ µ is the decreasing rearrangement of f with respect the measure µ, and f ∗∗ µ (t) = 1 t R t 0 f ∗ µ(s)ds (see Subsection 2.1). 3The spaces ¯X are defined in Section 2.2 below. SOBOLEV ANISOTROPIC INEQUALITIES 3 where ¯p∗ = D ¯p D− ¯p . In particular if p = p1 = · · · = pn, then p = p, ¯p∗ = p∗, and we get (1.9) kf kLp∗ µ (cid:22) which implies (1.3). kfxik Ai+1 D Lp µ ∀f ∈ C1 c (Rn), n Yi=1 In the unweighted case, i.e. A1 = · · · , An = 0, inequality (1.9) is well-known (see e.g. [34], [33] and [23]). Anisotropic inequalities involving Orlicz norm defined using an n-dimensional Young function were also studied in [14]. However, as far we know, our anisotropic inequalities involving rearrangement invariant spaces are new in this context. The paper is organized as follows. In Section 2 we provide a brief review on the rearrangements of functions and the theory of rearrangement invariant spaces. In Section 3 we will prove our main result (Theorem 3.1 below) which establishes the equivalence between (1.3) with p = 1 and (1.5). Finally in Section 4 we use the oscillation inequality (1.5) to derive anisotropic Sobolev inequalities in Rn with monomial weight xA in the general setting of rearrangement invariant spaces, with special attention in the case of Lebesgue spaces, Lorentz spaces, Lorentz-Zygmund spaces, Gamma spaces and the recent class of GΓ spaces. 2. Notations and preliminary results We briefly recall the basic definitions of rearrangements and of rearrangement- invariant (r.i.) spaces referring the reader to [5] and [22]. 2.1. Rearrangement of functions. Let µ an absolutely continuous measure with respect to Lebesgue measure on Rn. For a µ-measurable function u : Rn → R, the distribution function of u is given by µu(s) = µ{x ∈ Rn : u(x) > s} s ≥ 0. The decreasing rearrangement u∗ function from [0, ∞) into [0, ∞] which is equimeasurable with u. Namely, µ of u is the right-continuous non-increasing u∗ µ(t) = inf{s ≥ 0 : µu(s) ≤ t} t ≥ 0. We also define u∗∗ µ : (0, ∞) → (0, ∞) as u∗∗ µ (t) = u∗ µ(s)ds, 1 t Z t 0 Note that u∗∗ µ is also decreasing and u∗ µ ≤ u∗∗ µ (t) ≤ u∗∗ µ , moreover µ (t) + v∗∗ µ (t) (u + v)∗∗ for t > 0. The oscillation of u is defined by Oµ(u, t) := u∗∗ µ (t) − u∗ µ(t). Note that (2.1) is increasing. tOµ(u, t) =Z ∞ u∗ µ(t) µu(s)ds 4 F. FEO, J. MARTIN, AND M. R. POSTERARO When rearrangements are taken with respect to the Lebesgue measure on (0, ∞), we may omit the measure and simply write u∗ and u∗∗, etc... 2.2. Rearrangement invariant spaces. We say that a Banach function space X = X(Rn) on (Rn, µ) is rearrangement-invariant (r.i.) space, if g ∈ X implies that f ∈ X for all µ−measurable functions f such that f ∗ µ, and kf kX = kgkX. µ = g∗ A basic property of rearrangements is the Hardy-Littlewood inequality which tells us that, if u and w are two µ-measurable functions on Rn , then (2.2) ZRn u(x)w(x) dµ ≤Z ∞ 0 u∗ µ(t)w∗ µ(t) dt. An important consequence (2.2) is the Hardy-Littlewood-P´olya principle stating that (2.3) Z r 0 f ∗ µ(s)ds ≤Z r 0 g∗ µ(s)ds ∀r > 0 ⇒ kf kX ≤ kgkX for any r.i. space X. A r.i. space X(Rn) can be represented by a r.i. space on (0, +∞), with Lebesgue measure, ¯X = ¯X(0, ∞), such that kf kX = kf ∗ µk ¯X , for every f ∈ X. A characterization of the norm k · k ¯X is available (see [5, Theorem 4.10 and subsequent remarks]). The space ¯X is called the representation space of X. If X is a r.i space, we have µ ∩ L∞ ⊂ ¯X ⊂ L1 L1 µ + L∞, with continuous embeddings. The associate space X ′ of X is the r.i. space of all measurable functions h for which the r.i. norm given by khkX ′ = sup g6=0 RRn g(x)h(x) dµ(x) kgkX = sup 0 h∗ g6=0 R ∞ µ(s)g∗ kgkX µ(s)ds . In particular the following generalized Holder inequality ZRn g(x)h(x) dµ(x) ≤ kgkX khkX ′ holds. Classically conditions on r.i. spaces are given in terms of the Hardy operators defined by P f (t) = 1 t Z t 0 f (s)ds; Qaf (t) = 1 ta Z ∞ t saf (s) ds s , 0 ≤ a < 1, (if a = 0, we shall write Q instead of Q0). The boundedness of these operators on r.i. spaces can be described in terms of the so called Boyd indices4 defined by ¯αX = inf s>1 ln hX (s) ln s and αX = sup s<1 ln hX (s) ln s , 4Introduced by D.W. Boyd in [8]. SOBOLEV ANISOTROPIC INEQUALITIES 5 where hX (s) denotes the norm of the compression/dilation operator Es on ¯X, defined for s > 0, by Esf (t) = f ∗( t µ with p > 1, then ¯αX = αX = 1 p . It is well known that s ). For example if X = Lp P is bounded on ¯X ⇔ αX < 1, Q is bounded on ¯X ⇔ αX > a. The next two Lemmas will be used in Section 4. Lemma 2.1. (see [22, Page 43]) Let X be a r.i. space and let 0 ≤ θi ≤ 1 such that i=1 θi = 1, then fiθi kX ≤ k n Yi=1 kfikθi X . n Yi=1 Lemma 2.2. Let g, h be two positive measurable functions on (0, ∞) such that g(s) ≤ h∗∗(s), for all s ∈ (0, ∞), Pn (2.4) and Then 0 0 Z t g(s) ≤Z t g∗(s)ds ≤ 4Z t Z t 0 h∗(s)ds, for all t ∈ (0, ∞). h∗(s)ds for all t ∈ (0, ∞). 0 Therefore for any r.i. space X kgkX ≤ 4 khkX . The proof of the previous lemma is implicitly contained in the proof of Theorem 1.2 of [15], a detailed proof can be found in [27]. 2.2.1. Examples. Convexifications of r.i. spaces. A way to construct r.i. spaces is through the so- called p-convexification, which is the generalization of the procedure to construct Lp spaces, 1 < p < ∞, starting from L1. If X is a r.i. space, the p−convexification X (p) of X, (cf. [25]) is the r.i. space defined X (p) = {f : f p ∈ X} endowed with the following norm (2.5) kf kX (p) = kf pk1/p X . The same is true for the functional k·kX hpi defined as (2.6) kf kX hpi =(cid:13)(cid:13)(cid:13)(cid:0)(f p)∗∗(cid:1)1/p(cid:13)(cid:13)(cid:13)X . Spaces X hpi have been introduced in [17] in connection with the study of Sobolev embeddings into rearrangement-invariant spaces defined by a Frostman measure. 6 F. FEO, J. MARTIN, AND M. R. POSTERARO The generalized Lorentz spaces Λp,q(w). Given 1 ≤ p, q < ∞, and w a weight (a positive locally integrable function) on (0, ∞) . The generalized Lorentz spaces Λp,q(w) are defined by measurable functions on (0, ∞) such that (2.7) kf kΛp,q(w) :=(cid:18)Z ∞ 0 (cid:16)t1/pf ∗(t)(cid:17)q w(t) dt t (cid:19)1/q < ∞. If p = q, we write Λp(w) instead Λp,p(w). We denote by W (t) = R t 0 w(s)ds. It is know (see [10]) that Λp,q(w) = Λq(W q/p−1w). A weight w is called a Bp−weight if three is C > 0 such that Z ∞ t w(s) sp ≤ C rp Z t 0 w(s)ds, t > 0. The Bp class satisfies that Br ⊂ Bp if r ≤ p. If w ∈ Bp then (see [32]) kf kΛp(w) ≃(cid:18)Z ∞ 0 f ∗∗(t)pw(t)dt(cid:19)1/p , therefore Lorentz spaces defined by Bp−weights are r.i. spaces. Moreover, since (see [29, Theorem 6.5]) w ∈ Bp ⇒ W q/p−1w ∈ Bq we get that Λp,q(w) is a r.i. space if w ∈ Bp. Given f a µ−measurable function on Rn we define kf kΛp,q µ (w) =(cid:8)f : kf ∗ µkΛp.q(w) < ∞(cid:9) . Typical examples of generalized Lorentz spaces are the Lp-spaces and the Lorentz spaces Lp,q, defined either for p = q = 1 or p = q = ∞ or 1 < p < ∞, 1 ≤ q ≤ ∞ by t1/p−1/qf ∗ < +∞, µ(t)(cid:13)(cid:13)(cid:13)Lq [0,∞) and, more generally, the Lorentz-Zygmund spaces, defined for 1 < p < ∞, 1 ≤ q ≤ ∞ and α ∈ R by kf kLp,q t1/p−1/q(1 + log f )α/qf ∗ < +∞ µ kf kLp,q :=(cid:13)(cid:13)(cid:13) µ (log L)α :=(cid:13)(cid:13)(cid:13) µ(t)(cid:13)(cid:13)(cid:13)Lq [0,∞) Let us notice that, in spite of the notation, the quantities kf kLp,q and kf kLp,q(log L)α need not be norms; however, they can be turned into equivalent norms, when 1 < p < ∞, replacing f ∗ by f ∗∗. The Gamma spaces Γp(w). Let 1 ≤ p < ∞. Let w be an admissible weight, i.e. (2.8) (2.9) Z t 0 w(s)ds < ∞ and Z ∞ t w(s) sp ds < ∞. The Gamma space Γp(w) is the r.i. space defined as the set of measurable functions such that kf kΓp(w) :=(cid:18)Z ∞ 0 f ∗∗(s)pw(s)ds(cid:19)1/p < ∞. Given f a µ−measurable function on Rn we define kf kΓp µ(w) =(cid:8)f : kf ∗ µkΓp(w) < ∞(cid:9) . SOBOLEV ANISOTROPIC INEQUALITIES 7 The GΓp(p, m, w) spaces. Let 1 ≤ p, m < ∞ and let w be a weight satisfying that (2.10) Z ∞ 0 min(s, t)m/pw(t)dt < ∞, s > 0. The GΓ(p, m, w)−spaces are defined by (2.11) f : kf kGΓ(p,m,w) = Z ∞ 0 (cid:18)Z t 0 m/p f ∗(s)pds(cid:19) w(t)dt!1/m These spaces has been introduced in [19] in connection with compact Sobolev type embedding results, since then its turn out to be important and several papers devoted to the study of this spaces have been published (see e.g. [18], [20], [21] and the references quoted therein). GΓ(p, m, w) =  . < ∞  2.3. Some remarks about function spaces defined by oscillations. In this subsection we analyze functional properties of function spaces whose definition in- volves the oscillation of f . The principal difficulty dealing with the functional Oµ(f, t) is its nonlinearity. Therefore function spaces whose definition involves this quantity are not linear spaces. Consider the Hardy type operator defined on positive measurable function on (0, ∞) by ¯Qf (t) =Z ∞ t s1/Df (s) ds s . Theorem 2.3. Let X be a r.i. space on (0, ∞) and let us assume that X does not contain constant functions. i) If αX > 1 D , then kt−1/Df ∗∗(t)kX ≃ kt−1/D[f ∗∗(t) − f ∗(t)]kX . ii) If ¯αX < 1 D , then kf kL∞ (cid:22) kt−1/D[f ∗∗(t) − f ∗(t)]kX + kf kL1 µ+L∞ . Proof. i) Let f ∈ X. Since X does not contain constant functions, f ∗∗(∞) = 0. An elementary computation shows that (−f ∗∗)′ (t) = f ∗∗(t)−f ∗(t) , thus by the fundamental theorem of Calculus we get t Therefore, ds s . t (f ∗∗(t) − f ∗(t)) f ∗∗(t) =Z ∞ t−1/DZ ∞ kt−1/Df ∗∗(t)kX = (cid:13)(cid:13)(cid:13)(cid:13) t−1/DZ ∞ = (cid:13)(cid:13)(cid:13)(cid:13) (cid:22) (cid:13)(cid:13)(cid:13) s−1/D (f ∗∗(s) − f ∗(s))(cid:13)(cid:13)(cid:13)X t t (f ∗∗(s) − f ∗(s)) ds s (cid:13)(cid:13)(cid:13)(cid:13)X The converse inequality is obvious. s1/Ds−1/D (f ∗∗(s) − f ∗(s)) ds s (cid:13)(cid:13)(cid:13)(cid:13)X 1 D ). (since αX > 8 F. FEO, J. MARTIN, AND M. R. POSTERARO ii) Let us assume now that ¯αX < 1 D . We get kf kL∞ + kf kL1 µ+L∞ = f ∗∗(0) − f ∗∗(1) =Z 1 0 (f ∗∗(s) − f ∗(s)) ds s = Z 1 0 s1/D(cid:16)s−1/D (f ∗∗(s) − f ∗(s))(cid:17) ds s ≤ ks−1/D (f ∗∗(s) − f ∗(s)) kXks1/D−1χ[0,1](s)kX ′. It is enough to check that ks1/D−1χ[0,1](s)kX ′ < ∞. Since ¯αX = 1 − αX ′ < 1 can select D , we (2.12) 1 − 1/D < β < αX ′ . By the definition of indices, there is c > 0 such that (2.13) hX ′ (2−k) ≤ c2−kβ ∀k ≥ 0. For any k ≥ 0, write Ik = [2−k−1, 2−k). Since χ[0,1](s) =P∞ ∞ ∞ ks1/D−1χ[0,1](s)kX ′ ≤ ks1/D−1χIk (s)kX ′ ≤ 2k(1−1/D)kχIk (s)kX ′ k=0 χIk (s), we get Xk=0 2k(1−1/D)kD2−k χ[1/2,1)(s)kX ′ 2k(1−1/D)hX ′ (2−k)kχ[1/2,1)(s)kX ′ 2k(1−1/D−β)2kβhX ′(2−k)kχ[1/2,1)(s)kX ′ ∞ ∞ Xk=0 Xk=0 Xk=0 Xk=0 Xk=0 ∞ ∞ = ≤ = ≤ c 2k(1−1/D−β)kχ[1/2,1)(s)kX ′ < ∞ (by (2.13) and (2.12)). (cid:3) 3. Self-improvement of Sobolev inequality Let W 1,1 0 (Rn, µ) be the closure of the space C1 c (Rn) under the norm Inequality (1.3) with p = 1 implies the following anisotropic Sobolev inequality kukW 1,1 0 (∇u + u)dµ. (Rn,µ) =ZRn Xi=1(cid:13)(cid:13)(cid:13)(cid:13) D−1 µ (cid:22) L n D kf k ∂f ∂xi(cid:13)(cid:13)(cid:13)(cid:13)L1 µ (3.1) for f ∈ W 1,1 0 (Rn, µ). Theorem 3.1. Let µ and D be defined as in (1.1) and (1.2) respectively. Let f ∈ W 1,1 (Rn, µ) the following statements hold and are equivalent: 0 i) (Poincar´e inequality) (3.2) kf k D D−1 µ L (cid:22) n Xi=1 kfxikL1 µ . SOBOLEV ANISOTROPIC INEQUALITIES 9 ii) (Poincar´e inequality in multiplicative form) (3.3) kf k D D−1 µ L (cid:22) kfxik Ai+1 D L1 µ . n Yi=1 iii) (Mazya-Talenti's inequality in multiplicative form) The function f ∗ µ is lo- cally absolutely continuous and for all s > 0 we have that (3.4) µ(cid:1)′ s1−1/D(cid:0)−f ∗ (s) (cid:22) n Yi=1 d dsZ{f >f ∗ µ (s)} Ai+1 D . fxi dµ! iv ) (Oscillation inequality in multiplicative form) For all 1 ≤ p < ∞ and for all t > 0 we get (3.5) Z t D (cid:17)∗ 0 (cid:16)Oµ(f, ·)p (·)− p (s)ds (cid:22)Z t 0 n Yi=1" d dsZ{f >f ∗ µ(s)} fxi dµ!∗ (τ )# p( Ai+1 D ) dτ, where rearrangements without subscript µ are taken with respect to Lebesgue measure on (0, ∞). v ) Proof. i) ⇒ ii) kf k D D−1 µ L ,1 (cid:22) kfxik Ai +1 D L1 µ . n Yi=1 We follow a scaling argument as in [33, Lemma 7]. Indeed we apply (3.2) to function w(x) = f (λ1x1, · · · , λnxn) and we obtain (3.3) by choosing λi =Yj6=i kfxikL1 µ . ii) ⇒ iii) D Since L D−1 µ is continuously embedded in L implies the weaker one D D−1 ,∞ µ with constant 1, inequality (3.3) (3.6) sup t>0 t {x ∈ Rn : f (x) > t} D−1 D ≤ nC Let 0 < t1 < t2 < ∞, the truncations of f are defined by kfxik Ai+1 D L1 µ . n Yi=1 f t2 t1 (x) =  0 Observe that if f ∈ W 1,1 f t2 t1 in (3.6) we obtain t2 − t1 f (x) − t1 0 if f (x) > t2, if t1 < f (x) ≤ t2, if f (x) ≤ t1. (Rn, µ) then f t2 t1 ∈ W 1,1 0 (Rn, µ), therefore replacing f by Since sup t>0 sup t>0 t(cid:12)(cid:12)(cid:8)x ∈ Rn :(cid:12)(cid:12)f t2 t(cid:12)(cid:12)(cid:8)x ∈ Rn :(cid:12)(cid:12)f t2 t1 (x)(cid:12)(cid:12) > t(cid:9)(cid:12)(cid:12) t1 (x)(cid:12)(cid:12) > t(cid:9)(cid:12)(cid:12) D−1 D−1 D ≤ nC Ai +1 D . dµ! n Yi=1 ZRn(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∂f t2 t1 ∂xi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) D ≥ (t2 − t1) {x ∈ Rn : f (x) ≥ t2} D−1 D , 10 and we get (3.7) F. FEO, J. MARTIN, AND M. R. POSTERARO (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∂f t2 t1 ∂xi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12) ∂f ∂xi(cid:12)(cid:12)(cid:12)(cid:12) χ{t1<f ≤t2} n Using that fxi ≤ ∇f , (3.7) implies (t2 − t1) {x ∈ Rn : f (x) ≥ t2}1−1/D ≤ nC Yi=1 Z{t1<f ≤t2} (t2 − t1) {x ∈ Rn : f (x) ≥ t2}1−1/D ≤ nCZ{t1<f ≤t2} Ai +1 D . fxi dµ! ∇f dµ and from this inequality the locally absolutely continuity of f ∗ the same argument as in [26, page 137]. µ follows easily using Let s > 0 and h > 0, pick t1 = f ∗ µ(s + h), t2 = f ∗ µ(s), then by (3.7) we get (cid:0)f ∗ µ(s) − f ∗ µ(s + h)(cid:1) s1−1/D ≤ nC Thus, n Yi=1 Z{f ∗ µ (s+h)<f ≤f ∗ µ (s)} Ai+1 D . fxi dµ! µ(s) − f ∗ h (cid:0)f ∗ µ(s + h)(cid:1) s1−1/D ≤ nC n Yi=1 1 hZ{f ∗ µ (s+h)<f ≤f ∗ µ (s)} Ai+1 D . fxi dµ! Letting h → 0 we obtain (3.4). iii) ⇒ iv) Let 1 ≤ p < ∞. For 0 < s < t, we get Oµ(f, t) = ′ ′ 0 1 1 1 = = ≤ (s)ds (s)ds 0 t1/D µ(s) − f ∗ t Z t 0 (cid:0)f ∗ µ(t)(cid:1) ds t Z t 0 (cid:18)Z t (τ ) dτ(cid:19) ds µ(cid:1)′ s (cid:0)−f ∗ t Z t s(cid:0)−f ∗ µ(cid:1) t Z t s1−1/D(cid:0)−f ∗ µ(cid:1) Yi=1 d t Z t dτ Z{f >f ∗ 0  Yi=1" d t Z t dτ Z{f >f ∗  ≤ nCt1/D 0  Yi=1" d t Z t   ≤ nC t1/D t1/D 1 n n n 0 ≤ nC dτ Z{f >f ∗ µ(τ )} Ai +1 D ds (by (3.4)) fxi dµ! fxi dµ# µ(τ )} µ (τ )} ∗ Ai+1 D   fxi dµ#p Ai+1 D   (s)ds (by (2.2)) 1/p ∗ (s)ds  (by Holder). SOBOLEV ANISOTROPIC INEQUALITIES 11 Thus (3.8) (cid:16)t−1/DOµ(f, t)(cid:17)p (cid:22) 1 0  t Z t  n Yi=1" d dτ Z{f >f ∗ µ (τ )} ∗ fxi dµ#p Ai+1 D   (s)ds. Notice that the previous computation shows that (3.9) Oµ(f, t) = 1 t Z t 0 If p = 1, then using (3.9) and (3.8) we get ′ s(cid:0)−f ∗ µ(cid:1) (s) ds. Z t 0 (z)dz(cid:19) ds ds(cid:19) dz ds(cid:19) dz s ′ ′ z 0 0 0 s s−1/D s−1/D Oµ(f, s)s−1/Dds =Z t s (cid:18)Z s z(cid:0)−f ∗ µ(cid:1) =Z t (z)(cid:18)Z t z(cid:0)−f ∗ µ(cid:1) ≤Z t (z)(cid:18)Z ∞ z(cid:0)−f ∗ µ(cid:1) = DZ t z1−1/D(cid:0)−f ∗ µ(cid:1) 0  Yi=1" d ≤ nCDZ t dτ Z{f >f ∗  (z)dz n 0 0 z ′ ′ s−1/D If 1 < p < ∞, then using (3.8), Hardy's Inequalities (see [5, page 124]) and (3.9) we get fxi(x) dµ(x)# µ (·)} ∗ (s) ds. Ai+1 D   Z t 0 (cid:16)Oµ(f, s)s−1/D(cid:17)p ds p (z)dz(cid:19) (z)dz(cid:19)p ′ ds = ′ 0 0 s Z s ds =Z t 0 (cid:18) s−1/D z(cid:0)−f ∗ µ(cid:1) ≤Z t s Z s 0 (cid:18) 1 z1−1/D(cid:0)−f ∗ µ(cid:1) (cid:22)Z t (z)(cid:17)p 0 (cid:16)z1−1/D(cid:0)−f ∗ µ(cid:1) 0  Yi=1" d (cid:22)Z t dτ Z{f >f ∗  µ (·)} n ′ dz ∗ fxi(x) dµ(x)#p Ai +1 D   (s) ds. By Lemma 2.2 and [5, Exercise 10, page 88], we get Z t D (cid:17)∗ 0 (cid:16)Oµ(f, ·)p (·)− p iv) ⇒ v) n 0  Yi=1" d (s) ds (cid:22) Z t dτ Z{f >f ∗   " d ≤ Z t dτ Z{f >f ∗ Yi=1 n 0 µ(·)} µ(·)} ∗ ∗ fxi(x) dµ(x)#p Ai+1 D   fxi(x) dµ(x)#p Ai+1 D   (s) ds (s) ds 12 F. FEO, J. MARTIN, AND M. R. POSTERARO By (2.3) and (2.4), we obtain (3.10) Oµ(f, s)s− 1 (cid:13)(cid:13)(cid:13) D(cid:13)(cid:13)(cid:13)L1 fxi dµ!∗ Moreover we get (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) d dτ Z{f >f ∗ µ(τ )} (3.11) µ (τ )} fxi dµ# fxi dµ!∗ µ(τ )} n n ≤ 4nCD ≤ 4nCD(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  " d dτ Z{f >f ∗ Yi=1 Yi=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) d dτ Z{f >f ∗ (s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1 = Z ∞ = Z ∞ = ZRn 0 fxi dµ. 0 d dτ Z{f >f ∗ dτ Z{f >f ∗ d µ (τ )} µ (τ )} (s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)L1 ∗ Ai+1 D . L1 Ai +1 D   (s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) fxi dµ!∗ fxi dµ! dτ (s) ds Taking into account that ∂ ∂t f ∗∗ µ (t) = −(f ∗∗ µ (t) − f ∗ µ(t))/t and f ∗∗ µ (∞) = 0 (since f ∈ W 1,1 0 (Rn, µ)) by the Fundamental Theorem Calculus we have Therefore kf k D D−1 L ,1 :=Z ∞ 0 µ (s)s− 1 f ∗∗ f ∗∗ µ (t) =Z ∞ t dτ τ . µ(τ )(cid:1) µ (τ ) − f ∗ (cid:0)f ∗∗ (cid:20)Z ∞ D ds =Z ∞ =Z ∞ (cid:0)f ∗∗ D − 1Z ∞ D = 0 0 0 s µ (τ ) − f ∗ (cid:0)f ∗∗ µ (τ ) − f ∗ µ(τ )(cid:1) τ Z τ µ(τ )(cid:1) (cid:20) 1 µ(τ )(cid:1) τ − 1 0 µ (τ ) − f ∗ (cid:0)f ∗∗ dτ D ds τ (cid:21) s− 1 D ds(cid:21) dτ s− 1 D dτ. Conclusion follows by (3.10) and (3.11). v) ⇒ i) It is consequence of the continuous embedding of Lorentz space L D D−1 ,1 µ into Lebesgue space L mean. D D−1 µ and the relation between weighted arithmetic and geometric (cid:3) Remark 3.2. As in the classical case we have that (3.1) implies a better inequality involving the Lorentz norm k · k ,1 (see e.g. [16], [1], [28] and the bibliography therein). D−1 L D Remark 3.3. By (3.8) we get (3.12) t− 1 D Oµ(f, t) (cid:22) 1 t Z t 0 n Yi=1 d dτ Z{f >f ∗ µ(τ )} Ai +1 D fxi dµ! ds := I(t), SOBOLEV ANISOTROPIC INEQUALITIES 13 that is a pointwise oscillation inequality in multiplicative form. Moreover by (3.12) we recover the classical oscillation inequality (see [28] and [26]) As matter of the fact by Holder inequality t− 1 D Oµ(f, t) (cid:22) ∇f ∗∗ (t). (3.13) I(t) ≤ = ≤ n n 0 d Yi=1 1 t Z t Yi=1 1 t Z{f >f ∗ t Z t Yi=1(cid:18) 1 n fxi∗ 0 (cid:22) ∇f ∗∗ (t). dsZ{f >f ∗ µ(s)} Ai +1 D fxi dµ! ds! Ai+1 D fxi dµ! µ(t)} Ai +1 D (by (2.2)) µ (s)ds(cid:19) 4. Anisotropic inequalities in rearrangement invariant spaces In this section starting from the oscillation inequality (3.5) we derive some anisotropic inequalities in Rn in the general setting of rearrangement invariant spaces. We will use throughout this Section the following notation ( fxi)∗ µ(t) := d dsZ{f >f ∗ µ(s)} fxi dµ!∗ (t). In order to prove these kind of results we need the following results. Lemma 4.1. Let 1 < p < ∞ and let v be a weight (a positive locally integrable function on (0, ∞)). Let u(t) = ∂ ∂t 1 +Z 1 t −1 p−1 (s) v p p−1 s ds!1−p . Then, there exists C > 0 such that (cid:18)Z 1 0 f ∗∗(s)pu(s)ds(cid:19) Proof. Since 1/p ≤ C(cid:18)Z 1 0 (Oµ(f, s))p v(s)ds(cid:19) 1/p +(cid:18)Z 1 0 (cid:18)Z t 0 u(s)ds(cid:19) 1/p Z 1 t −1 p−1 v(s) p p−1 s ds!(p−1)/p ≤ 1 +Z 1 t ≤ 1, −1 p−1 p p−1 v s 1/p 0 Z 1 u(s)ds(cid:19) ds!(1−p)/p Z 1 t f ∗(t)ds. −1 p−1 v(s) p p−1 s ds!(p−1)/p the result follows from [4, Lemma 5.4]. (cid:3) Lemma 4.2. (see [31] and [2]) Let f ∈ W 1,1 0 (Rn, µ), then Z t 0 ( fxi)∗ µ(τ )dτ ≤Z t 0 fxi∗ µ (τ )dτ, (t ≥ 0) 14 F. FEO, J. MARTIN, AND M. R. POSTERARO therefore by (2.3) for any r.i space X on (Rn, µ) we have that 4.1. Convexification of r.i. spaces. (cid:13)(cid:13)(cid:13) ( fxi)∗ fxi∗ = kfxikX . µ(cid:13)(cid:13)(cid:13) ¯X ≤(cid:13)(cid:13)(cid:13) µ(cid:13)(cid:13)(cid:13) ¯X 4.1.1. The X (q) convexification. Let X be a r.i. space on Rn and X (q) its q−con- vexification defined in (2.5). In the next theorem we state some anisotropic in- equalities for functions f such that fxi ∈ X (pi) with pi ≥ 1 for i = 1, · · · , n. Theorem 4.3. Let X be a r.i. p1, · · · , pn ≥ 1, then space on (Rn, µ) and f ∈ W 1,1 0 (Rn, µ). If (4.1) kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]k ¯X (p) (cid:22) where p and D are defined in (1.7) and in (1.2), respectively and the involved norms are defined as in (2.5). Moreover i) if αX > p D , then kfxik Ai+1 D X (pi) , n Yi=1 kfxik n Yi=1 Ai +1 D X (pi) ; On the other hand, since (see [5, Proposition 5.13, Chapter 3 ]) α ¯X (p) = αX ¯p and ¯α ¯X (p) = ¯αX ¯p (4.2) and (4.3) follows from Theorem 2.3. (cid:3) Remark 4.4. We stress in the previous proof if we start from (3.13) instead of (3.5) we can not consider the case when p1 = · · · = pn = 1. In the particular case X = L1 µ, the previous Theorem can be detailed as in the following proposition. (4.2) kt−1/Df ∗∗ µ (t)k ¯X (p) (cid:22) ii) if ¯αX < p D , then n (4.3) kf kL∞ (cid:22) Proof. Since ¯X (p) is a r.i space, by Theorem 3.1 part iv) we get Ai+1 kfxik D X (pi ) + kf kL1 µ+L∞ . Yi=1 µ (s) − f ∗ (cid:13)(cid:13)(cid:13)(cid:0)f ∗∗ µ(s)(cid:1) s− 1 D(cid:13)(cid:13)(cid:13) ¯X (p) pi pi n n Ai +1 Ai +1 ¯p (cid:16) ¯p D (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ¯X (p) (cid:22)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) µi Yi=1h( fxi)∗ D (cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ¯X (p) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) µi Yi=1h( fxi)∗ µ(t)(cid:13)(cid:13)(cid:13) Yi=1(cid:13)(cid:13)(cid:13) Yi=1 ( fxi)∗ kfxik ¯X (pi) X (pi) Ai+1 Ai+1 D ≤ ≤ n n D (by (2.4)) (by Lemma 4.2). SOBOLEV ANISOTROPIC INEQUALITIES 15 Proposition 4.5. Let p1, · · · , pn ≥ 1 and f ∈ W 1,1 i) If p < D, then 0 (Rn, µ). (4.4) kf kLp∗,p µ (cid:22) n Yi=1 Ai+1 kfxik L D pi µ , where p is defined as in (1.7) and p∗ = pD D−p . ii) If p = D, then (4.5) Z 1 0 (cid:18) f ∗∗ µ (t) 1 + ln 1 t(cid:19)D dt t !1/D (cid:22) n Yi=1 kfxik Ai+1 D Lpi µ + kf kL1 µ+L∞ . iii) If p > D, then (4.6) kf kL∞ (cid:22) kfxik Ai +1 D Lpi µ + kf kL1 µ+L∞ . n Yi=1 Theorem 4.3. We have only to prove (4.5). = L ¯p µ and αL ¯p = ¯αL ¯p = 1 ¯p , (4.4) and (4.6) follows from If v(t) = 1 in Lemma 4.1, then t µ(cid:1)( ¯p) Proof. Since (cid:0)L1 u(t) = (D − 1)(cid:18) 1  0 f ∗∗ Z 1 1 + ln(cid:0) 1 t(cid:1) µ (t) 1+ln( 1 t )(cid:19)−D t  !D  dt 1 s . Under this choice Lemma 4.1 allows us to get 1/D (cid:22)(cid:18)Z 1 0 (cid:0)f ∗∗ µ (t) − f ∗ t (cid:19) µ(t)(cid:1)D dt 1/D + f ∗∗ µ (1). (cid:3) Remark 4.6. Our result implies Theorem 1.1. Indeed it gives an embedding in Lorentz spaces. If A1 = · · · = An = 0 we obtain the same results of [33] and [23] for what concerns (4.4). Remark 4.7. Let Ω ⊂ Rn be a bounded domain and f ∈ C1 the proof of Proposition 4.5, we get c (Ω). Arguing as in  Z µ(Ω) 0 Since µ (t) f ∗∗ 1 + ln µ(Ω) t !D sup 0<t<µ(Ω) f ∗∗ µ (t) D−1 D (cid:16)1 + ln µ(Ω) t (cid:17) dt 1/D n (cid:22) Yi=1 t   (cid:22) Z µ(Ω) 0 kfxik Ai+1 D pi µ L + 1 µ (Ω)ZΩ f dµ. µ (t) f ∗∗ 1 + ln µ(Ω) t !D 1/D , dt t   we obtain the following anisotropic Trudinger inequality sup 0<t<µ(Ω) f ∗∗ µ (t) t (cid:17) (cid:16)1 + ln µ(Ω) (cid:22) D−1 D n Yi=1 kfxik Ai+1 D Lpi µ + 1 µ (Ω)ZΩ f dµ. 16 F. FEO, J. MARTIN, AND M. R. POSTERARO 4.1.2. The X hqi convexification. Theorem 4.8. Let X be a r.i. space on (Rn, µ) and f ∈ W 1,1 1, then 0 (Rn, µ). If p1, · · · , pn ≥ kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]k ¯X hpi (cid:22) n Yi=1 kfxik Ai +1 D Xhpii , where p and D are defined in (1.7) and in (1.2), respectively and the involved norms are defined as in (2.6). Moreover i) if α ¯X hpi > 1 D , then kt−1/Df ∗∗ µ (t)k ¯X hpi (cid:22) n Yi=1 ii) if ¯α ¯X hpi < 1 D , then kfxik Ai +1 D Xhpii ; kf kL∞ (cid:22) kfxik n Yi=1 Ai+1 D Xhpii + kf kL1 µ+L∞ . Proof. By Theorem 3.1 part iv) with p = ¯p, we get 1 t Z t D (cid:17)∗ 0 (cid:16)Oµ(f, ·)¯p (·)− ¯p (s)ds (cid:22) = ≤ Let X a r.i. space, then n 0 n 1 1 t Z t Yi=1h( fxi)∗ t Z t Yi=1h(cid:16)( fxi)∗ t Z t Yi=1(cid:18) 1 0 h(cid:16)( fxi)∗ n 0 kt−1/D[f ∗∗ µ (t) − f ∗ pi ( Ai+1 D ) ¯p ¯p dτ dτ D ) pi ( Ai+1 D ) µ(τ )i ¯p( Ai +1 µ(τ )(cid:17)pii µ(τ )(cid:17)pii dτ(cid:19) 1/ ¯p(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ¯X D )(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ¯X pi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (s)ds(cid:19) µ(τ )(cid:17)pii dτ(cid:19) µ(τ )(cid:17)pii dτ(cid:19) pi ( Ai+1 ( Ai +1 D ) ¯X 1 1 n µ(t)]k ¯X hpi =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) t Z t (cid:18) 1 D (cid:17)∗ 0 (cid:16)Oµ(f, ·)¯p (·)− ¯p (cid:22)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) t Z t Yi=1(cid:18) 1 0 h(cid:16)( fxi )∗ Yi=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) t Z t (cid:18) 1 0 h(cid:16)( fxi )∗ µ(τ )(cid:13)(cid:13)(cid:13) Yi=1(cid:13)(cid:13)(cid:13) Yi=1 ( Ai+1 D ) ¯Xhpii ( fxi)∗ kfxik ¯Xhpii Ai+1 ≤ ≤ = n n n D (by Lemma 4.2). The statements (4.2) and (4.3) follows from Theorem 2.3. (cid:3) SOBOLEV ANISOTROPIC INEQUALITIES 17 4.2. Generalized Lorentz spaces. In this section we state some anisotropic in- equalities for functions f such that the partial derivatives are in some generalized Lorentz spaces. More precisely as a consequence of Theorem 4.3 we obtain the following theorem. Theorem 4.9. Let w ∈ Bmin(p1,··· ,pn), f ∈ W 1,1 1 and the involved norms are defined as in (2.7). i) If αΛp,q(w) > 1 D , then 0 (Rn, µ), p1, · · · , pn ≥ 1, q1, · · · , qn ≥ kf kΛp∗ ,q µ (w) (cid:22) n Yi=1 Ai +1 kfxik Λ D pi ,qi µ , (w) where ¯p and ¯q are defined as in (1.7) and p∗ = pD ii) If ¯αΛp,q(w) > 1 D , then D−p . kf kL∞ (cid:22) n Yi=1 iii) In the remaining cases we get kfxik Ai +1 Λpi ,qi D µ + kf kL1 µ+L∞ . (w) Ai +1 kfxik Λ D pi ,qi µ + kf kL1 µ+L∞ , (w) 0 (cid:18)Z 1 ∂t 1 +R 1  t where u(t) = ∂ f ∗∗(s)¯qu(s)ds(cid:19) 1/¯q (cid:22) n Yi=1 −1 ¯q−1 (cid:18)s ¯q ¯p − ¯q D −1 w(s)(cid:19) ¯q ¯q−1 s 1−¯q . ds  Proof. Since w ∈ Bmin(p1,··· ,pn) all the spaces Λpi,qi(w) are r.i spaces and from min(p1, · · · , pn) ≤ p it follows that Λp,q(w) is a r.i. space. We note that (4.7) By Theorem 3.1 we obtain 1 = ¯q D n Xi=1 Ai + 1 qi . kt−1/D[f ∗∗ µ (t) − f ∗ Ai +1 D (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Λp,q(w) Yi=1 fxi(t) n 1 p − 1 ¯q 1 pi − 1 qi fxi(t) Ai+1 Λpi ,qi (w) n fxi(τ ) µ(t)]kΛ ¯p,q(w) (cid:22) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Yi=1 = Z ∞ 0 t Yi=1(cid:18)Z ∞ 0 (cid:16)t Yi=1(cid:13)(cid:13)(cid:13) fxi(cid:13)(cid:13)(cid:13) Yi=1 kfxik ≤ ≤ = n n n D µ Ai+1 Λpi ,qi D (by Lemma 4.2). (w) 1 ¯q Ai +1 D !¯q D (cid:17)qi Ai +1 w(t) dt! w(t) dt(cid:19) Ai +1 qi D (by (4.7) and (2.4)) 18 F. FEO, J. MARTIN, AND M. R. POSTERARO Part i) follows form i) of Theorem 2.3, because kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kΛp,q(w) ≃ kt−1/Df ∗∗ µ (t)kΛp,q(w) µ (t)kΛp∗,q(w) ≃ kf ∗∗ ≃ kf kΛp∗ ,q µ (w). Part ii) is consequence of i) of Theorem 2.3. Let us prove now iii). If we denote I := kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kΛp,q(w) we get I = (cid:18)Z ∞ 0 (cid:16)s ≃ Z ∞ 0 (cid:18)s ≥ Z ∞ 0 (cid:18)s ≥ Z ∞ 0 (cid:18)s ≥ Z ∞ 0 (cid:18)s ≥ Z ∞ 0 (cid:18)s ≃ (cid:18)Z ∞ 0 (cid:16)s = (cid:18)Z ∞ 0 1 ¯p − 1 1 ¯p − 1 ¯q 1 ¯p − 1 1 ¯p − 1 1 ¯p − 1 ¯q 1 ¯p − 1 ¯q ¯q ¯q 1 1 1 µ (t) − f ∗ µ (t) − f ∗ µ (t) − f ∗ µ (z) − f ∗ µ(t)](cid:17)∗ w(s) ds(cid:19)1/¯q (s)(cid:17)¯q ¯q (cid:16)t−1/D[f ∗∗ (z) dz(cid:19)¯q s Z s µ(t)](cid:17)∗ 0 (cid:16)t−1/D[f ∗∗ 2sZ 2s (z) dz(cid:19) µ(t)](cid:17)∗ 0 (cid:16)t−1/D[f ∗∗ 2sZ 2s µ(z)](cid:17) dz(cid:19) 0 (cid:16)z−1/D[f ∗∗ 2sZ 2s µ(z)](cid:17) dz(cid:19) s (cid:16)z−1/D[f ∗∗ Z 2s s (cid:16)z−1/D−1(cid:17) dz(cid:19) w(s) ds(cid:19)1/¯q µ(s)](cid:17) ¯q D −1w(s) ds(cid:19)1/¯q µ (z) − f ∗ D [f ∗∗ µ (s) − f ∗ s[f ∗∗ µ (s) − f ∗ µ(s)] ¯p − ¯q 2s 1 ¯q ¯q . ¯q ¯q ¯q w(s) ds!1/¯q w(s) ds!1/¯q w(s) ds!1/¯q w(s) ds!1/¯q w(s) ds!1/¯q 1 ¯p − 1 ¯q − 1 [f ∗∗ µ (s) − f ∗ µ(s)]¯qs Considering as weights (since w ∈ B ¯p) (by (2.1)) v(s) := s ¯q ¯p − ¯q D −1w(s) and u(t) = the result follows from Lemma 4.1. t (cid:16)s 1 +Z 1 ∂ ∂t   ¯q ¯p − ¯q D −1w(s)(cid:17) ¯q−1 ¯q s −1 ¯q−1 1−¯q , ds  (cid:3) The following corollaries follow from the previous theorem considering w = 1 and w(t) = (1 + ln t)α with α ∈ R, respectively, and recalling that ¯αLp,q(log L)α = αLp,q(log L)α = ¯αLp,q = αLp,q = 1 p . Corollary 4.10. Let f ∈ C1 i) If ¯p < D, then c (Rn), p1, · · · , pn ≥ 1 and q1, · · · , qn ≥ 1. kf kLp∗ ,q µ (cid:22) kfxik Ai +1 Lpi ,qi D µ , n Yi=1 where ¯p and ¯q are defined as in (1.7) and p∗ = pD D−p . SOBOLEV ANISOTROPIC INEQUALITIES 19 ii) If ¯p > D, then n kf kL∞(Rn) (cid:22) kfxik Ai+1 Lpi ,qi D µ + kf kL1 µ+L∞ . Yi=1 s !1/¯q ds (cid:22) n Yi=1 n Yi=1 iii) If ¯p = D, then Z 1 0 (cid:18) f ∗∗ µ (s) 1 + ln 1 s(cid:19)¯q Corollary 4.11. Let f ∈ C1 i) If ¯p < D, then kfxik Ai+1 D pi ,qi µ L + kf kL1 µ+L∞ . c (Rn), p1, · · · , pn ≥ 1 , q1, · · · , qn ≥ 1 and α ∈ R. kf kLp∗ ,q µ (log L)α (cid:22) kfxik Ai +1 Lpi ,qi D µ (log L)α . where ¯p and ¯q are defined as in (1.7) and p∗ = pD ii) If ¯p > D, then D−p . kf kL∞(Rn) (cid:22) iii) If ¯p = D, then n Yi=1 Ai+1 kfxik L D pi ,qi µ (log L)α + kf kL1µ+L∞ . Z 1 0 (cid:18) f ∗∗ µ (s) 1 + ln 1 s(cid:19) ¯q(cid:18)1 + ln 1 s(cid:19)α ds s !1/¯q (cid:22) n Yi=1 kfxik Ai+1 Lpi ,qi D µ (log L)α + kf kL1 µ+L∞ . When A1 = · · · = An = 0, i) of Corollary 4.10 is contained in [33] and [23]. For our knowledge the other ones are new. 4.3. The Gamma spaces. Theorem 4.12. Let w be a weight satisfying condition (2.8), f ∈ W 1,1 p1, · · · , pn ≥ 1 and the involved norms are defined as in (2.9). i) If αΓp∗ ,(w) > 1 D , then 0 (Rn, µ), kt−1/Df ∗ µ(t)kΓp∗ (w) (cid:22) kfxik Ai+1 D Γpi µ (w) , n Yi=1 where ¯p is defined as in (1.7) and p∗ = pD ii) If ¯αΓp∗ ,(w) < 1 D , then D−p . kf kL∞ (cid:22) kfxik Ai+1 D pi µ (w) Γ + kf kL1 µ+L∞ . n Yi=1 iii) In the remaining cases we get 1/p∗ (cid:22) −1 p∗−1 u(s)ds(cid:19) (cid:18)s− D w(s)(cid:19) p∗ p∗ p∗−1 s f ∗∗(s)p∗ 0 (cid:18)Z 1 ∂t 1 +R 1  t where u(t) = ∂ Ai +1 kfxik Γ D pi µ (w) + kf kL1 µ+L∞ , n Yi=1 ds  1−p∗ . 20 Proof. F. FEO, J. MARTIN, AND M. R. POSTERARO kt−1/D[f ∗∗ µ (t) − f ∗ 1 p∗ w(t) dt  Ai+1 piD D ds!p∗ w(t) dt! pi (by (4.7) and (2.4)) 0 n n Ai +1 µ(t) ( fxi)∗ ( fxi)∗ µ(t)]kΓp∗ (w) (cid:22) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) D (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Γp∗ ,(w) Yi=1 =  0 1 t Z t Z ∞ Yi=1 Yi=1 Z ∞ t Z t 0 (cid:18) 1 Yi=1(cid:13)(cid:13)(cid:13) µ(cid:13)(cid:13)(cid:13) Yi=1 kfxik Γ ( fxi)∗ ( fxi)∗ D pi µ (w) µ (w) Ai +1 Ai+1 Γpi = ≤ ≤ n n n 0 D Ai +1 µ(s) µ(s)ds(cid:19) (by Lemma 4.2). Part i) follows form i) of Theorem 2.3, since kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kΓp∗ (w) ≃ kt−1/Df ∗∗ µ (t)kΓp∗ (w) ≃ kt−1/Df ∗ µ(t)kΓp∗ (w). Part ii) is consequence of i) of Theorem 2.3. Let us prove now iii). If we denote I := kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kΓp∗ (w) using the same argument that in part iii) of Theorem 4.9 we easily obtain I ≥(cid:18)Z ∞ 0 Considering as weights [f ∗∗ µ (s) − f ∗ µ(s)]¯qs− p∗ D w(s) ds(cid:19)1/p∗ . v(s) := s− p∗ D w(s) and u(t) = ∂ ∂t the result follows from Lemma 4.1. 4.4. The GΓ(m, p, w)-spaces. t (cid:16)s− p∗ D w(s)(cid:17) p∗−1 p∗ s  1 +Z 1  −1 p∗−1 1−p∗ ds  , (cid:3) Theorem 4.13. Let w be a satisfying (2.10), f ∈ W 1,1 the involved norms are defined as in (2.11). i) If αGΓ(p∗,m,w) > 1 D , then 0 (Rn, µ), p1, · · · , pn ≥ 1 and kt−1/Df ∗ µ(t)kGΓ(p∗,m,w) (cid:22) kfxik n Yi=1 where ¯p is defined as in (1.7) and p∗ = pD ii) If ¯αGΓ(p∗,m,w) < 1 D , then D−p . Ai+1 D GΓ(pi,m,w) , kf kL∞ (cid:22) n Yi=1 Ai+1 kfxik D GΓ(pi,m,w) + kf kL1 µ+L∞ . µ+L∞ , 1/m n 0 . n t s (cid:22) Ai+1 D m m−1 −1 m−1 Ai +1 1−m µ(t) kfxik ( fxi)∗ GΓ(pi,m,w) + kf kL1 (cid:18)Z 1 f ∗∗(s)mu(s)ds(cid:19) Yi=1 ∂t 1 +R 1 ds! (cid:16)s− m D w(s)(cid:17) µ(t)]kGΓ(p∗,m,w) (cid:22) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) D (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)GΓ(p∗,m,w) Yi=1 =  0 Z t Z ∞ Yi=1 ≤  Yi=1(cid:18)Z t Z ∞ Yi=1 Z ∞ 0 (cid:18)Z t Yi=1(cid:13)(cid:13)(cid:13) µ(cid:13)(cid:13)(cid:13) Yi=1 ( fxi)∗ ( fxi)∗ ( fxi)∗ ( fxi)∗ GΓ(pi,m,w) GΓ(pi,m,w) µ(s)pi ds(cid:19) µ(s)pi ds(cid:19) µ(s)p∗ Ai+1 kfxik Ai+1 D Ai +1 ≤ = ≤ n n n n 0 0 n 0 0 D (by Lemma 4.2). 1 m D ds!m/p∗ m pi Ai+1 D 1 m w(t) dt  w(t) dt  w(t)! dt! 1 m m pi Ai +1 D SOBOLEV ANISOTROPIC INEQUALITIES 21 iii) In the remaining cases we get where u(t) = ∂ Proof. By Theorem 3.1 we get kt−1/D[f ∗∗ µ (t) − f ∗ Part i) follows form i) of Theorem 2.3, because kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kGΓ(p∗,m,w) ≃ kt−1/Df ∗∗ µ (t)kGΓ(p∗,m,w) Part ii) is consequence of i) of Theorem 2.3. Let us prove now iii). If we denote I := kt−1/D[f ∗∗ µ (t) − f ∗ µ(t)]kGΓ(p∗,m,w) using the method of part iii) of Theorem 4.9 we get I ≥(cid:18)Z ∞ 0 Considering as weights [f ∗∗ µ (s) − f ∗ µ(s)]ms− m D w(s) ds(cid:19)1/m . v(s) := s− m D w(s) and u(t) = the result follows from Lemma 4.1. ∂ ∂t 1 +Z 1 t (cid:0)s− m D w(s)(cid:1) m−1 s m −1 m−1 1−m , ds  (cid:3) Acknowledgments. The first author has been partially supported by FFABR and the second one has been partially supported in part by Grants MTM2016-77635-P, MTM2016-75196-P (MINECO) and by 2017SGR358. 22 F. FEO, J. MARTIN, AND M. R. POSTERARO The second author gratefully acknowledges the generous hospitality of the Uni- versit`a degli Studi di Napoli "Federico II" Dipartimento di Matematica "R. Cac- cioppoli" where much of this work was done. References [1] A. Alberico, A. Cianchi, L. Pick and L. Slav´ıkov´a, Sharp Sobolev type embeddings on the entire Euclidean space. Commun. Pure Appl. Anal. 17 (2018), n. 5, 2011?-2037. [2] A. Alvino and G. Trombetti, Sulle migliori costanti di maggiorazione per una classe di equazioni ellittiche degeneri. Ricerche Mat. 27 (1978), 413 -- 428. [3] D. Bakry, I. Gentil and M. Ledoux. Analysis and Geometry of Markov Diffusion Operators. Grundlehren der Mathematischen Wissenschaften, vol. 348. Springer, Berlin (2013). [4] J. Bastero, M. Milman and F.J. Ruiz. A note on L(∞, q)s spaces and Sobolev embeddings. Indiana Univ. Math. J. 52 (2003), 1215 -- 1230. [5] C. Bennett and R. Sharpley, Interpolation of Operators. Academic Press, Boston, 1988. [6] H. Brezis. Is there failure of the inverse function theorem? Morse theory, Minimax theory and their applications to nonlinear differential equations. In: Proceedings of the Workshop held at the Chinese Academy of Sciences, Beijing, 1999, 2333, New Stud. Adv. Math., 1, Int. Press, Somerville, MA, (2003) [7] H. Brezis and J.L. V´azquez. Blow-up solutions of some nonlinear elliptic problems. Rev. Mat. Univ. Complut. 10 (1997), 443 -- 469 [8] D.W. Boyd. Indices of function spaces and their relationship to interpolation. Canad. J. Math. 21 (1969), 1245 -- 1254. [9] F. Brock , F. Chiacchio and A. Mercaldo. A weighted isoperimetric inequality in an orthant. Potential Anal. 41 (2014), 171 -- 186. [10] M. J. Carro and J. Soria. Weighted Lorentz spaces and the Hardy operator. J. Funct. Anal. 112 (1993), 480 -- 494. [11] X. Cabr´e. Isoperimetric, Sobolev, and eigenvalue inequalities via the Alexandroff-Bakelman- Pucci method: a survey. Chin. Ann. Math., 38B(1) (2017), 201 -- 214. [12] X. Cabr´e and X. Ros-Oton. Sobolev and isoperimetric inequalities with monomial weights. J. Differential Equations, 255 (2013), 4312 -- 4336. [13] X. Cabr´e and X. Ros-Oton. Regularity of stable solutions up to dimension 7 in domains of double revolution. Commun. Partial Differ. Equ. 38 (2013), 135 -- 154. [14] A. Cianchi. A fully anisotropic Sobolev inequality. Pacific J. Math. 196 (2000), n. 2, 283 -- 295. [15] A. Cianchi. Symmetrization and second-order Sobolev inequalities. Ann.Mat. 183 (2004), 45 -- 77. [16] A. Cianchi. Optimal Orlicz-Sobolev embeddings. Rev. Mat. Iberoamericana 20 (2004), 427 -- 474. [17] A. Cianchi, L. Pick and L. Slavikova. Sobolev embeddings. rearrangement-invariant spaces and Frostman measures. Ann. Inst. H. Poincar´e Anal. Non Lin´eaire (to appear). [18] Fiorenza, A.; Rakotoson, J. M. Some estimates in GΓ(p, m, w) spaces. J. Math. Anal. Appl. 340 (2008), 793 -- 805. [19] A. Fiorenza, J.M. Rakotoson, Compactness, interpolation inequalities for small Lebesgue- Sobolev spaces and applications. Calc. Var. Partial Differential Equations 25 (2005) 187 -- 203. [20] A. Fiorenza, M.R. Formica, A. Gogatishvili, T. Kopaliani, and J.M. Rakotoson. Character- ization of interpolation between grand, small or classical Lebesgue spaces. Nonlinear Anal. 177 (2018),422 -- 453. [21] A. Gogatishvili, M. Krepela, L. Pick and F. Soudsk´y. Embeddings of Lorentz-type spaces involving weighted integral means. J. Funct. Anal. 273 (2017), 2939 -- 2980. [22] S. G. Krein, Yu. I. Petunin and E. M. Semenov. Interpolation of linear operators, Transl. Math. Monogr. Amer. Math, Soc. 54, Providence, (1982). [23] V. I. Kolyada and F. J. P´erez. Estimates of difference norms for functions in anisotropic Sobolev spaces. Mathematische Nachrichten 267(1) (2004), 46 -- 64. SOBOLEV ANISOTROPIC INEQUALITIES 23 [24] N. Lam. Sharp Trudinger-Moser inequalities with monomial weights. NoDEA Nonlinear Dif- ferential Equations Appl. 24 (2017), no. 4, Art. 39, 21 pp. [25] J. Lindenstrauss and L. Tzafriri. Classical Banach Spaces II. Function Spaces, Springer- Verlag, Berlin, 1979. [26] J. Martin and M. Milman. Pointwise symmetrization inequalities for Sobolev functions and applications. Adv. Math. 225 (2010),121 -- 199. [27] J. Martin and M. Milman. Self improving Sobolev-Poincar´e inequalities, truncation and sym- metrization. Potential Anal. 29 (2008), 391 -- 408. [28] J. Martin, M. Milman and E. Pustylnik. Sobolev inequalities: Symmetrization and self- improvement via truncation. J. Funct. Anal. 252 (2007), 677 -- 695. [29] C. J. Neugebauer. Weighted norm inequalities for averaging operators of monotone functions. Publ. Mat. 35 (1991), 429 -- 447. [30] V. H. Nguyen. Sharp weighted Sobolev and Gagliardo-Nirenberg inequalities on half-spaces via mass transport and consequences. Proc. Lond. Math. Soc.(3) 111 (2015), n. 1, 127 -- 148. [31] J. Rakotoson. R´earrangement relatif. (French) [Relative rearrangement] Un instrument d'estimations dans les probl`emes aux limites. for limit problems] Math´ematiques & Applications (Berlin) [Mathematics & Applications], 64. Springer, Berlin, 2008. xvi+293 pp. [An estimation tool [32] E. Sawyer. Boundedness of classical operators on classical Lorentz spaces. Studia Math. 96 (1990), 145 -- 158. [33] L. Tartar. Imbedding theorems of Sobolev spaces into Lorentz spaces. BUMI Serie 8 1-B (1998), 479 -- 500. [34] M. Troisi. Teoremi di inclusione per spazi di Sobolev non isotropi. Ricerche Mat. 18 (1969), 3 -- 24. Dipartimento di Ingegneria, Universit`a degli Studi di Napoli "Parthenope ", Centro Direzionale, Isola C4, 80143 Napoli, Italy E-mail address: [email protected] Current address: Universitat Aut`onoma de Barcelona, Department of Mathematics, Bellaterra (Barcelona), Spain E-mail address: [email protected] Universit`a degli Studi di Napoli "Federico II", Dipartimento di Matematica "R. Cac- cioppoli", Complesso Monte S. Angelo, Napoli, Italy E-mail address: [email protected]
1804.02277	1	1804	2018-04-04T16:58:06	Remarks on Generalized Hardy Algebras	[ "math.FA" ]	For a measure space $(\Omega, \Sigma, \mu)$ with a positive finite measure $\mu$, and a positive real number $p$, we define the space $L_p^{+}(\mu)=L_p^{+}$ of all (equivalence classes of) $\Sigma$-measurable complex functions $f$ defined on $\Omega$ such that the function $\left(\log^+\|f\|\right)^p$ is integrable with respect to $\mu $.We define the metric $d_p$ on $L^{+}_p$ which generalizes the metric introduced by Gamelin and Lumer in [G] for the case $p=1$. It is shown that the space $L^{+}_p$ is a topological algebra. On the other hand, one can define on the space $L_p^{+}$ an equivalent $F$-norm $\| \cdot\|_p$ that makes $L_p^{+}$ into an Orlicz space. For the case of the normalized Lebesgue's measure $dt/2\pi$ on $[0,2\pi)$, it follows that the class $N^p(1<p<\infty)$ introduced by I. I. Privalov in [P], may be considered as a generalization of the Smirnov class $N^+$. Furthermore, $N^p(1<p<\infty)$ with the associated modular becomes an Hardy-Orlicz class. Finally, for a strictly positive and measurable on $[0,2\pi)$ function $w$, we define the generalized Orlicz space $L_p^{w}(\mathrm{d}t/2\pi)=L^w_p$ with the modular $\rho^w_p$ given by the function $\psi_w(t,u)=\big(\log(1+uw(t))\big)^p$, with a "weight" $w$. We observe that the space $L^w_p$ is a generalized Orlicz space with respect to the modular $\rho^w_p$. We examine and compare different topologies induced on $L^w_p$ by corresponding "weights" $w$.	math.FA	math	REMARKS ON GENERALIZED HARDY ALGEBRAS ROMEO MESTROVI ´C∗, ZARKO PAVI ´CEVI ´C∗∗ AND NOVO LABUDOVI ´C∗∗∗ Abstract For a measure space (Ω, Σ, µ) with a positive finite measure p (µ) = p of all (equivalence classes of) Σ-measurable complex func- µ, and a positive real number p, we define the space L+ L+ tions f defined on Ω such that the function (cid:0)log+ f (cid:1)p is inte- grable with respect to µ. We define the metric dp on L+ p which generalizes the metric introduced by Gamelin and Lumer in [G] for the case p = 1. It is shown that the space L+ p is a topologi- cal algebra. On the other hand, one can define on the space L+ p an equivalent F -norm · p that makes L+ p into an Orlicz space. For the case of the normalized Lebesgue's measure dt/2π on [0, 2π), it follows that the class N p(1 < p < ∞) introduced by I. I. Privalov in [P], may be considered as a generalization of the Smirnov class N +. Furthermore, N p(1 < p < ∞) with the associated modular becomes an Hardy-Orlicz class. Finally, for a strictly positive and measurable on [0, 2π) function w, we define the generalized Orlicz space Lw p with the p given by the function ψw(t, u) = ( log(1+ uw(t)))p, modular ρw with a "weight" w. We observe that the space Lw p is a general- ized Orlicz space with respect to the modular ρw p . We examine and compare different topologies induced on Lw p by correspond- ing "weights" w. p (dt/2π) = Lw 1 Introduction For a measure space (Ω, Σ, µ) with a nonnegative finite, complete mea- sure µ not vanishing identically, denote by Lp(µ) = Lp (0 < p ≤ ∞) the familiar Lebesgue spaces on Ω. In Section 2, for p > 0, we de- fine the class L+ p of all (equivalence classes of) Σ-measurable p (µ) = L+ 1991 Mathematics Subject Classification Primary 46E30. Secondary 30H05, 46J15. 1 complex functions f defined on Ω such that the function log+ f is in Lp. Every space L+ p is an algebra. For each p > 0, in Section 2 we define the metric dp on L+ p by dp(f, g) = inf t>0 [t + µ ({x ∈ Ω : f (x) − g(x) ≥ t})] +ZΩ(cid:12)(cid:12)(cid:12)(cid:16)log+ f (x)(cid:17)p −(cid:16)log+ g(x)(cid:17)p(cid:12)(cid:12)(cid:12) dµ. The space L+ in [G]. It was proved in [G] that L+ metric d1 is a topological algebra. statement for every space L+ and δp on L+ as the initial metric dp. 1 was introduced in [G, p. 122], with the notation L(µ) 1 with the topology given by the In Section 2 we prove the same p , p > 0. We also define two metrics ρp p , and we show that they induce the same topology on L+ p By analogy with the Hardy algebra H(µ) defined in [G], in Section 3 we define the algebra Hp(µ) with p > 0. It is known (see [G]) that the Smirnov class N + on the unit disk D : z < 1 in the complex plane may be considered as the Hardy algebra H(dθ/2π). The analoguous results are obtained in Section 5 for the algebra N p, p > 1, introduced by I. I. Privalov with the notation Aq in [P]. In Section 4 we note that the function ψ : [0, ∞) 7→ [0, ∞) defined as ψ(t) = ( log(1 + t))p, is an Orlicz function or a ϕ-function. Fur- ther, we observe that the space L+ p (dt/2π) = Lp, p > 0, consisting of all complex-valued functions f , defined and measurable on [0, 2π) for which (kf kp)p := Z 2π 0 (log(1 + f (t)))p dt 2π < +∞. (1.1) is the Orlicz class coinciding with the associated Orlicz space (see [Mu, Definition 1.4, p. 2]), whose generalization we give in Section 6. We prove that the modular convergence k · kp and the norm convergence · p are equivalent. As an application, we show that there does not exist a nontrivial continuous linear functional on the space (Lp, k · kp). For p > 1, following I. I. Privalov (see [P, p. 93]), a function f holomorphic in D, belongs to the class N p, if there holds sup 0≤r<1Z 2π 0 (cid:16)log+ f (reiθ)(cid:17)p dθ 2π < ∞, In Section 5 we note that the algebra N p may be considered as the Hardy-Orlicz space with the Orlicz function ψ defined in Section 4. 2 Identifying a function f ∈ N p with its boundary function f ∗, the space N p is identical with the closure of the space of all functions holomor- phic in D and continuous in ¯D : z ≤ 1 in the space (Lp(dt/2π) ∩ N, · p). On the other hand, (N p, dp) coincides with the space Hp(dθ/2π) defined in Section 3, i.e. with the closure of the disk algebra P ( ¯D) in the space (Lp(dt/2π), dp). Therefore, the space N p may be considered as generalized Hardy algebra. From this fact, it is easy to show that N p is an F -algebra with respect to the F -norm k · kp given by 1.3. In the last section we note that the real function ψw with p > 0, defined on Ω × [0, ∞) by the formula ψw(t, u) = ( log(1 + uw(t)))p, is a Musielak-Orlicz function. For the Lebesgue measure space (Ω, [0, 2π), dt/2π), we denote by Lw p , p > 0, the class of all (equivalence classes of) complex-valued functions f , defined and measurable on [0, 2π) for which p (dt/2π) = Lw (kf kw p )p := Z 2π 0 (log(1 + f (t)w(t)))p dt 2π < +∞. p is the generalized Orlicz class with the modular k · kw p coincides with the associated generalized Orlicz space. k · kw Then Lw p , and Lw p is a modular in the sense of Definition 1.1 in [Mu, p. 1], and by ρw p (f, g) = (kf − gkw p . By [Mu, p. 2, Theorem 1.5, and p. 35, Theorem 7.7], it follows that the functional · w p , is defined an invariant metric on Lw p )min{p,1}, f, g ∈ Lw p defined as f w p = inf(ε > 0 : Z 2π 0 log 1 + f (t)w(t) ε !!p dt 2π ≤ ε) , f ∈ Lw p , is a complete F -norm. We prove that (Lw weight w. p , ρw p ) is an F -space for any p = Lw p ⊂ Lw p , and Lω For two weights w and ω such that log+ (w/ω) ∈ Lp, we show that p if and only if log (w/ω) ∈ Lp. Further, we p by the metric ρω p is stronger than p . If p ≥ 1 and log (w/ω) /∈ Lp, p by the p by the metric Lω prove that the topology defined on Lω that induced on Lω p by the metric ρw then Lω metric ρω p is strictly stronger than that induced on Lω p , and the topology defined on Lω p is a proper subset of Lw 3 p . As an application, we show that if log w ∈ Lp, then (Lp, ρw ρw p ) is an F -algebra, and the metric topologies ρw p and ρp are the same. Finally, if p ≥ 1 and w is a weight such that log+ w ∈ Lp, we give four equivalent necessary and sufficient conditions for the space (Lp, ρw p ) to be an F -algebra. 2 Equivalent metrics on the space L+ p (0 < p < ∞) Let (Ω, Σ, µ) be a measure space, i.e. Ω is a nonempty set, Σ is a σ-algebra of subsets of Ω and µ is a nonnegative finite, complete measure not vanishing identically. Denote by Lp(µ) = Lp (0 < p ≤ ∞) the familiar Lebesgue spaces on Ω. For p > 0, we define the class L+ p of all (equivalence classes of) Σ-measurable complex functions f defined on Ω such that the function log+ f is in Lp, where log+ a = max{log a, 0}, i.e. such that p (µ) = L+ ZΩ(cid:16)log+ f (x)(cid:17)p dµ < +∞. Obviously, L+ q ⊂ L+ xs/se, x ≤ 0, s > 0, we see that Sp>0 Lp ⊂ Tp>0 L+ p for q > p, and from the inequality log+ x ≤ p . Combining the inequalities log+ f + g ≤ log+ f + log+ g + log 2, log+ f g ≤ log+ f + log+ g with (x + y)p ≤ 2max{p−1,0}(xp + yp) and (x + y + z)p ≤ 3max{p−1,0}(xp + yp + zp), respectively, we see that every space L+ p is an algebra with respect to the pointwise addition and multiplication. For each p > 0, we define the metric dp on L+ p by dp(f, g) = inf t>0 [t + µ ({x ∈ Ω : f (x) − g(x) ≥ t})] dµ. (2.1) +ZΩ(cid:12)(cid:12)(cid:12)(cid:16)log+ f (x)(cid:17)p −(cid:16)log+ g(x)(cid:17)p(cid:12)(cid:12)(cid:12) The space L+ 1 was introduced in [G, p. 122], with the notation L(µ) in [G]. In fact, the above metric dp with p = 1 coincides with the Gamelin-Lumer's metric d defined on L+ 1 . It was proved in [G, p. 122, Theorem 2.3] that the space L+ 1 with the topology given by the metric d1 is a topological algebra. The following result is a generalization of the corresponding result for the case p = 1. The proof of this result is 4 completely analogous to those for the case p = 1 given in [G, p. 122]3, and therefore, may be omitted. Theorem 2.1 The space L+ p with the metric dp given by (1.2) is a topological algebra, i.e. a topological vector space with a complete met- ric in which multiplication is continuous. By the inequality ( log(1 + x))p ≤ 2max{p−1,0}((log 2)p + (log+ x)p), (2.2) it follows that f belongs to L+ p if and only if there holds (kf kp)p := ZΩ (log(1 + f (x)))p dµ < ∞. By the inequality log(1 + f + g) ≤ log(1 + f ) + (log(1 + g) and Minkowski's inequality, it follows that for p > 1 the function ρp defined as ρp(f, g) = kf − gkp, f, g ∈ L+ p , (2.3) satisfies the triangle inequality. Combining the above inequality with (x + y)p ≤ xp + yp for 0 < p ≤ 1, we see that the function ρp defined as ρp(f, g) = (kf − gkp)p, f, g ∈ L+ p , 0 < p ≤ 1, satisfies also the triangle inequality. Hence, ρp is an invariant metric on L+ p for all p > 0. Recall that a subset K of Lp forms an uniformly integrable family if for given ε > 0 there exists a δ > 0 so that f (x) dµ < δ for all f ∈ K, ZE whenever E ⊂ Ω with its measure µ(E) < δ. Two metrics (or norms) defined on the same space will be called equivalent if they induce the same topology. For the proof of Theorem 2.3 we will need the following lemma. Lemma 2.2 ([G, p. 122, Theorem 1.3]). Let (fn) be a sequence in L1 and f ∈ L1 such that fn → f in L1. Then a sequence (fn) is a uni- formly integrable family. Conversely, if a sequence (fn) is a uniformly integrable family on Ω, and fn → f in measure, then f belongs to L1 and fn → f in L1. 5 Theorem 2.3 The metric dp defines a topology for L+ alent to the topology defined by the metric ρp. p which is equiv- Proof. Suppose first that ρp(fn, f ) → 0 as n → ∞, where (fn) and f are in L+ p . Then by Chebyshev's inequality, it is easily seen that fn → f in measure on Ω. For simplicity, put (kf kE)max{p,1} = ZE (log(1 + f (x)))p dµ for any measurable set E ⊂ Ω. By the triangle inequality and (2.2), we have kfnkE ≤ kfn − f kE + kf kE ≤ ρp(fn, f ) + 2max{1−1/p,0} × (cid:18)µ(E)(log 2)p +ZE(cid:16)log+ f (x)(cid:17)p (2.4) dµ)1/ max{p,1}(cid:19) . : n ∈ N} form a uniformly integrable family, and by Lemma 2.2, dp(fn, f ) → 0 as n → ∞. This shows that {(cid:16)log+ fn(x)(cid:17)p definition of dp, fn → f in measure, and by Lemma 2.2, {(cid:16)log+ fn(x)(cid:17)p Conversely, assume that dp(fn, f ) → 0 as n → ∞. Then, by the n ∈ N} are uniformly integrable. Replacing in (2.3) f by fn and fn by f − fn, we see that the family { (log(1 + fn(x) − f (x)))p : n ∈ N} is uniformly integrable. Thus, by Lemma 2.2, ρp(fn, f ) → 0 as n → ∞. Hence the metrics dp and ρp are equivalent. : Remark. Using the same argument applied in the proof of Theorem 2.3, it is easy to see that the metrics ρp and dp are equivalent with the metric δp given on L+ p by δp(f, g) = inf t>0 [t + µ ({x ∈ Ω : f (x) − g(x) ≥ t})] +(cid:18)ZΩ(cid:12)(cid:12)(cid:12) log+ f (x) − log+ g(x)(cid:12)(cid:12)(cid:12) p dµ(cid:19)1/ max{p,1} , f, g ∈ L+ p . Remark. In [Y, Remark 5, p. 460], M. Hasumi pointed out that the Yanagihara's metric ρ = ρ1 defines a topology for the space L+ 1 = L(µ), which is equivalent to the metric topology d1 = d used by Gamelin- Lumer in [G, p. 122]. 6 Corollary 2.4 The space L+ p , p > 0, with the topology given by the metric ρp is an F -algebra, i.e. a topological algebra with a complete translation invariant metric ρp. Proof. We see from Theorems 2.1 and 2.3 that it is sufficient to show that the space L+ p is complete with respect to the metric ρp. The completeness of L+ p may be proved by the standard manner as the completeness of the Lebesgue spaces Lp or an arbitrary generalized Orlicz space with a corresponding F -norm (for example, see the proof of Theorem 7.7 in [Mu, p. 35]). In view of this, note that L+ p may be considered as the generalized Orlicz space Lw p with a constant function w(t) ≡ 1 on [0, 2π) (see Section 6). Theorem 2.5 If 0 < p < s < ∞, then L+ induces a topology for L+ on L+ s by the initial metric ds. p , and the metric dp s which is coarser than the topology defined s ⊂ L+ If 0 < p < s < ∞, L+ Proof. sequence in L+ s and f ∈ L+ s ⊂ L+ p is obvious. Let (fn) be a s such that ds(fn, f ) → 0 as n → 0. This in L1, and a uniformly integrable family on Ω. Since 0 < p < s < ∞, it is routine means by definition of ds, that (cid:16)log+ fn(cid:17)s fn → f in measure. Then by Lemma 2.2, a sequence (cid:16)(cid:16)log+ fn(cid:17)s(cid:17) is to verify that (cid:16)(cid:16)log+ fn(cid:17)p(cid:17) is also a uniformly integrable family on Ω. Then by Lemma 2.2, (cid:16)log+ fn(cid:17)p → (cid:16)log+ f (cid:17)p → (cid:16)log+ f (cid:17)s in L1, and hence dp(fn, f ) → 0 as n → 0. 3 A generalization of Hardy algebras Let A be a uniform algebra on a compact Hausdorff space Ω. Let µ be a representing measure for a nonzero complex-valued homomorphism α of A. For 0 < q ≤ ∞, denote by H q(µ) the Hardy space (the closure of A in the Lebesgue space Lq = Lq(µ)). Motivated by the definition of the Hardy algebra H(µ) given in [G, p. 123], here we define the space Hp(µ) = Hp(0 < p < ∞), as the dp-closure of A in the space L+ p defined in the Section 2. Observe that the space H1 coincides with the Hardy algebra H(µ) defined in [G, p. 122]. From Theorem 2.5 we obtain immediately its Hp-analogue. p (µ) = L+ 7 Theorem 3.1 If 0 < p < s < ∞, then Hs ⊂ Hp, and the metric dp induces a topology for Hs which is coarser than the topology defined on Hs by the initial metric ds. Denote by Mα the set of all representing measures for a homomorphism α. For the proof of Theorem 3.2, we will need the following lemma which generalizes Corollary 2.2 in [G, p. 122]. Lemma 3.2 For all 0 < q ≤ ∞ and all p > 0, the topology of Lq is stronger than the topology of the space L+ p . Proof. Clearly, it suffices to consider the case 0 < q ≤ 1. Let (fn) be a sequence in Lq such that fn → f in Lq for some f ∈ Lq. From the inequality log(1 + x) ≤ xs/s, x ≥ 0, 0 < s ≤ 1, it follows that ( log(1 + x))p ≤ pxq/q, and hence ρp(fn, f ) → 0 as n → ∞. Therefore, we conclude by Theorem 2.3 that dp(fn, f ) → 0 as n → ∞, which completes the proof. Theorem 3.3 Suppose that the space Mα is finite-dimensional, and let µ be a core representing measure for a homomorphism α. Then for any fixed 0 < q ≤ ∞, there holds \p>0 Hp\ Lq = H q(µ). (3.1) Furthermore, there holds Hp\ Lq = H q(µ) for 1 ≤ p ≤ q ≤ ∞. (3.2) Proof. By [G, p. 125, Theorem 4.2], for all 0 < q ≤ ∞ we have H1T Lq = H q(µ). This shows that Tp>0 HpT Lq ⊆ H q(µ). Suppose now that f ∈ H q(µ) for some 0 < q ≤ ∞. Then there is a sequence (fn) in A such that fn → f in Lq as n → ∞. By Lemma 3.2, fn → f in L+ p as n → ∞, for each p > 0. Thus, f ∈ Tp>0 Hp, which implies that H q(µ) ⊆ Tp>0 HpT Lq. This yields (3.1). If 1 ≤ p ≤ q ≤ ∞, by Theorem (3.1) we have Hp\ Lq ⊆ H1\ Lq = H q(µ). Conversely, from (3.1) we obtain H q(µ) = Ts>0 HsT Lq ⊆ HpT Lq. This proves (3.2). 8 4 Orlicz spaces L+ p (dt/2π)(0 < p < ∞) The function ψ : [0, ∞) 7→ [0, ∞) defined as ψ(t) = ( log(1 + t))p, is continuous and nondecreasing in [0, ∞), such that ψ(0) = 0, ψ(t) > 0 for t > 0, and limt→+∞ ψ(t) = +∞, is called an Orlicz function or a ϕ-function (see [Mu, p. 4, Examples 1.9]). Further, observe that the space L+ p (dt/2π) = Lp, p > 0, consisting of all complex-valued functions f , defined and measurable on [0, 2π), for which (kf kp)p := Z 2π 0 (log(1 + f (t)))p dt 2π < +∞. is the Orlicz class (see [Mu, p. 5]), whose generalization we give in Section 6. It follows by the dominated convergence theorem that the class Lp coincides with the associated Orlicz space (see [Mu, p. 2, Definition 1.4]), consisting of those functions f ∈ Lp such that Z 2π 0 (log(1 + cf (t)))p dt 2π → 0 as c → 0 + . Furthermore, since by ρp(f, g) = (kf −gkp)min{p,1}, f, g ∈ Lp, is defined an invariant metric on Lp, the function k · kp given by 4.1 is a modular in the sense of Definition 1.1 in [Mu, p. 1]. For any fixed f ∈ Lp, by the monotone convergence theorem, we see that limc→0 ψ(cf ) → 0, and so (Lp, ρp) is a modular space in the sense of Definition 1.4 in [Mu, p. 2]. In other words, the function k · kp is an F -norm. It is well known (see [Mu, Theorem 1.5, p. 2 and Theorem 7.7, p. 35]) that the functional · p defined as f p = inf(ε > 0 : Z 2π 0 log 1 + f (t) ε !!p dt 2π ≤ ε) , f ∈ Lp, is a complete F -norm. Furthermore (see [L, p. 54]), Lp is a completion (closure) of the space of all continuous functions on [0, 2π) in the space (Lp, · p). Theorem 4.1 The F -norms k · kp and · p induce the same topology on the space Lp. In other words, the norm and modular convergences are equivalent. 9 Proof. Since a ϕ-function ψ(t) = ( log(1 + t))p satisfies obviously the so-called ∆2-condition given by φ(2t) ≤ cφ(t) (∆2) with a constant c = 2p, the assertion follows immediately from [L, p. 55, 2.4]. But, we give here a direct proof. Let Kρ(0, r) = {f ∈ Lp : kf kp < r} be a neighborhood of 0 in the space (Lp, k · kp), and let f ∈ Kρ(0, r) be any fixed. Consider a neighborhood Kψ(f, ε/2) = {f ∈ Lp : kf kp < r} of f in the space (Lp, · p), and assume g ∈ Kψ(f, ε/2). Then we have < δ} = σ. g − f δ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) , Thus, from ε/2 > σ we infer that whence we obtain ε 2 < g − f > g − f p = inf {δ > 0 : (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ε/2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ε/2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:12)(cid:12)(cid:12)(cid:12) + log (1 + g − f ε 2 < < ε 2 ε 2 + ε 2 kgkp ≤ kg − f kp + kf kp ε 2(cid:12)(cid:12)(cid:12)(cid:12) ε 2 + r − ε ) + r − ε + r − ε = r. This shows that Kψ(f, ε/2) ⊂ Kρ(0, r). Conversely, let Kψ(0, r) be a neighborhood of 0 in the space (Lp, · p), and let f ∈ Kψ(0, r) be any fixed. Put σ = r − f p, take k an integer so that k > 2/σ, and set ε = σ/2k. Then for all g ∈ Kρ(f, ε) by the triangle inequality, we have < kk(g − f )kp ≤ kkg − f kp ≤ kε = σ 2 . (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) g − f σ/2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Hence, we obtain g − f p ≤ σ/2, and so gp ≤ g − f p + f p ≤ σ 2 + r − σ = r − σ 2 < r. 10 Thus, g ∈ Kψ(0, r), i.e. Kρ(f, ε) ⊂ Kψ(0, r). Hence, the topologies defined on the space Lp by k · kp and · p are the same. As an application of Theorem 4.1, we obtain the following result. Corollary 4.2 There does not exist a nontrivial continuous linear functional on the space (Lp, k · kp). Proof. Since the Orlicz function ψ(t) = ( log(1 + t))p satisfies the con- dition lim inf t→+∞ t−1ψ(t) = 0, it follows by [MO] that there does not exist a nontrivial modular continuous linear functional on the space (Lp, · p). Hence, this is true by Theorem 4.1 for the space (Lp, k · kp). 5 Privalov's spaces N p(1 < p < ∞) As in [G, p. 125], the Hardy algebra H1(dθ/2π) consists of all functions f holomorphic on the unit disk D : z < 1 for which the family nlog+ f (reiθ) : 0 ≤ r < 1o is uniformly integrable on the unit circle T , that is, for a given ε > 0, there exists δ > 0 so that log+ f (reiθ) dθ 2π ZE < ε, 0 ≤ r < 1, for any measurable set E ⊂ [0, 2π) with its Lebesgue measure E < δ. This space is usually called the Smirnov class N +. For p > 1, following I. I. Privalov (see [P, p. 93], where N p is denoted as Aq), a function f holomorphic in D, belongs to the class N p, if there holds sup 0≤r<1Z 2π 0 (cid:16)log+ f (reiθ)(cid:17)p dθ 2π < ∞, For p = 1, the condition (5.1) defines the Nevanlinna class N of holo- morphic functions in D. Recall that for f ∈ N, the radial limit f ∗(eiθ) = lim r→1 f (reiθ) exists for almost every eiθ and log f ∗ ∈ L1 unless f 6≡ 0. It is known (see [Mo]) that N p ⊂ N + ⊂ N. [p>1 11 Observe that the algebra N p may be considered as the Hardy-Orlicz space with the Orlicz function ψ : [0, ∞) 7→ [0, ∞) defined as ψ(t) = ( log(1 + t))p. For more informations on the Hardy-Orlicz spaces see [Mu, Ch. IV, Sec. 20]. Identifying a function f ∈ N with its bound- ary function f ∗, by [L, 3.4, p. 57], the space N p is identical with the closure of the space of all functions holomorphic in D and con- tinuous in ¯D : z ≤ 1 in the space (Lp(dt/2π) ∩ N, · p). On the other hand, (N p, dp) coincides with the space Hp(dθ/2π) defined in Section 3, i.e. with the closure of the disk algebra P ( ¯D) in the space (Lp(dt/2π), dp). Therefore, the space N p may be considered as gener- alized Hardy algebra. From this fact, Theorem 2.3 and Corollary 2.4, it follows immeditately the following result of M. Stoll obtained in [S]. Theorem 5.1 ([S, Theorem 4.2]). For p > 1, the space N p with the topology given by the metric ρp is an F -algebra. Furthermore, the polynomials are dense in (N p, ρp), and hence N p is separable. 6 The generalized Orlicz spaces Lw p Let (Ω, Σ, µ) be a measure space, and let w be a Σ-measurable nonneg- ative function defined on Ω such that w(t) > 0 µ-almost every on Ω. For p > 0, the real function ψw defined on Ω × [0, ∞) by the formula ψw(t, u) = ( log(1 + uw(t)))p is a Musielak-Orlicz function, since it satisfies the Caratheodory condi- tions, i.e. it satisfies the following conditions (see [Mu, p. 33, Definition 7.1]): (i) u 7→ ψw(t, u) is a ϕ-function of the variable u ≥ 0 for µ-almost is a nondecreasing, continuous function of u such that every t, i.e. ψw(t, 0) = 0, ψw(t, u) > 0 for u > 0, limu→∞ ψw(t, u) → ∞; (ii) ψw(t, u) is a Σ-measurable function of t for all u ≥ 0. For the Lebesgue's measure space (Ω, [0, 2π), dt/2π), denote by Lw p , p > 0, the class of all (equivalence classes of) complex-valued functions f , defined and measurable on [0, 2π) for which p (dt/2π) = Lw (log(1 + f (t)w(t)))p dt 2π < +∞. (kf kw p )p := Z 2π 0 12 p is called the generalized Orlicz class with the modular k · kw Lw p (see [Mu, p. 33, Definition 7.2]). It follows by the dominated convergence theorem that the class Lw p coincides with the associated generalized Orlicz space consisting of those functions f ∈ Lw p for which Z 2π 0 (log(1 + cf (t)w(t)))p dt 2π → 0 as c → 0 + . p + kf kw p coincides with the space of all finite elements of Lw Furthermore, from the inequality kcf kw Lw those functions f ∈ Lw w(t) > 0 almost every on [0, 2π) and p , we see that p consisting of p for every c > 0. Since p such that cf ∈ Lw p ≤ kckw kf + gkw p = (f + g)wp ≤ f wp + gwp = kf kw p + kgkw p , p is a modular in the sense of Definition 1.1 of [Mu, p. 1]. k · kw In other words, the function k · kw p (f, g) = (kf − gkw p . By [Mu, p. 2, Theorem 1.5, and p. 35, Theorem 7.7], it follows that the functional · w p , is defined an invariant metric on Lw p is an F -norm. Hence, by ρw p )min{p,1}, f, g ∈ Lw p defined as f w p = inf(ε > 0 : Z 2π 0 log 1 + f (t)w(t) ε !!p dt 2π ≤ ε) , f ∈ Lw p , is a complete F -norm. Theorem 6.1 (Lw and · w p induce the same topology on the space Lw p . p , · w p ) is an F -space. Furthermore, F -norms k · kw p p , c 7→ cf is a continuous mapping from C into Lw Proof. Since the functional · p is a complete F -norm, and hence for a fixed f ∈ Lw p , it suffices (see [DS, p. 51]) to check the continuity of the mapping f 7→ cf from Lw p , for a fixed c ∈ C. Take k an integer with c ≤ k. Then p into Lw cf ≤ kf ≤ kf ; so f 7→ cf is continuous. The second assertion of the theorem can be proved completely analogously as Theorem 4.1. Theorem 6.2 For any weight w, (Lw p , ρw p ) is an F -space. 13 (iii) Let p ≥ 1. If log (w/ω) /∈ Lp, then Lω Furthermore, the topology defined on Lω stronger than that induced on Lω p by the metric ρw p . p is a proper subset of Lw p . p is strictly p by the metric ρω p , ρw p , ρw Proof. To show that (Lw p ) is an F -space, it suffices by Theorem 6.1 to prove that (Lw p , ρw p ) is complete. Let (fn) be a Cauchy sequence in (Lw p ). This means that (wfn) is a Cauchy sequence in (Lp, ρp), and by Corollary 2.4, there is a g ∈ Lp such that ρp(wfn, g) → 0 as n → ∞. Then ρw p (g/w, 0) = p , ρw ρp(g, 0) < ∞, we see that f = g/w ∈ Lw p ) is complete. p (fn, g/w) = ρp(wfn, g) → 0, and since ρw p . Hence, the space (Lw Theorem 6.3 Let w and ω be two weights such that log+ (w/ω) ∈ Lp. Then: p ⊂ Lw p , and Lω (i) Lω p by the metric ρω (ii) The topology defined on Lω p by the metric ρw p . p if and only if log (w/ω) ∈ Lp. p = Lw that induced on Lω p is stronger than p , ρw (iv) Let p ≥ 1. If (Lω p coincides with that induced on Lω p ⊂ Lw then log (w/ω) ∈ Lp, and the topology defined on Lω metric ρω p ) with p ≥ 1 is a complete metric space, p = Lw p by the p by the metric ρw p . p follows immediately from p +kw/ωkp =, p , then since 1/w ∈ p , it follows that log+ (ω/w) ∈ Lp, and so log (w/ω) ∈ Lp. p ⊂ Proof. (i) The inclusion relation Lω the inequality kf kw f ∈ Lω Lw Conversely, if log (w/ω) ∈ Lp, then log+ (ω/w) ∈ Lp, and thus Lw Lω p = kf ωw/ωkp ≤ kf ωkp+kw/ωkpkf kω p , and the fact that w/ω ∈ Lp. If Lω p , which implies Lω p = Lω p = Lw p . p = Lw (ii) Assume that (fn) is a sequence in Lω p , ρp) as n → ∞. This means that ρω p such that fn → f in (Lω p (fn, f ) → 0, or equivalently, ρp(ωfn, ωf ) → 0 as n → ∞. Since w/ω ∈ Lp, by Corol- lary 2.4, we have ρp(wfn, wf ) → 0, as n → ∞. Thus, ρw p (fn, f ) → 0 as n → ∞. p and f ∈ Lω (iii) If log (w/ω) /∈ Lp, the strict inclusion relation Lω p follows from the assertion (i). Since log+ (w/ω) ∈ Lp and log (w/ω) /∈ L1, by [MP, Theorem 3.1], there holds p ⊂ Lw inf P ∈P0Z 2π 0 log 1 + P (eit) w(t) ω(t)!!p dt 2π = 0, or equivalently, that there exists a sequence (Pn) of (analytic) poly- nomials with Pn(0) = 1 such that ρw p (Pnω, 0) → 0 as n → ∞. On the 14 other hand, by the same theorem, there holds inf P ∈P0Z 2π 0 (cid:16)log(cid:16)1 + P (eit)(cid:17)(cid:17)p dt 2π = (log 2)p, and hence the above sequence (Pn) does not converge to 0 in the space (Lp, ρp), or equivalently that a sequence (Pn/ω) does not converge to 0 in the space (Lω p is strictly stronger than the metric topology ρw p . p , ρw p ) is a complete metric space, but log (w/ω) /∈ p ) Hence, the metric topology ρω (iv) Suppose that (Lω p , ρω Lp. We will consider two cases. Case 1. log (w/ω) /∈ L1. By Theorem 6.2, the space (Lω p ) is an F -space, and by the assumption, (Lω p ) is complete, and henceforth it is an F -space. We see from (iii) that the topology defined on Lω p by the metric ρω p by the metric ρw p is strictly stronger than that induced on Lω p . Consider the identity map p , ρw p , ρω j : (Lω p , ρω p ) → (Lω p , ρw p ). Then by (ii), j is continuous. By the open mapping theorem, we conclude that the inverse j−1 of j is also continuous. This shows that the metric topologies ρω p must be the same. A contradiction. log (w/ω) ∈ L1. Since log+ (ω/w) ∈ Lp ⊂ L1, it follows p . Define the outer that log+ (ω/w) /∈ L1, and hence 1/w ∈ Lw function F by p and ρw p \ Lω Case 2. F (z) = exp Z 2π 0 eit + z eit − z log w(t) ω(t)! dt 2π , z ∈ D. Then F ∗(eit) = w(t)/ω(t) at almost every t ∈ [0, 2π). If p = 1, by the canonical factorization theorem for the Smirnov class N + (see [D, Theorem 2.10]), it folows that F belongs to N +. Similarly, if p > 1, by the canonical factorization theorem for the Privalov's spaces N p (see [P, p. 98]), it folows that F belongs to N p. For simplicity, we shall write N 1 instead of N +. By a result of Mochizuki (see [Mo, Theorem 2]), there is a sequence (fn) in N p, p > 1, such that fnF → 1 in N p as n → ∞. For p = 1, put fn = 1/F ∈ N 1 for all n. Hence, in both cases ρp(f ∗ nw/ω, 1) → 0 as n → ∞. This may be written as ρw n/ω, 1/w) → 0 as n → ∞. Since log+ f ∗ n/ω ∈ Lω p for all n. Hence p (f ∗ n ∈ Lp, it follows that f ∗ nF ∗, 1) → 0, or equivalently ρp(f ∗ 15 f ∗ n/ω → 1/w /∈ Lω closed subspace of the space Lw that (Lω second assertion of (iv) follows immediately from (i) and (ii). p is not p . This contradicts the assumption p ) is complete. Hence must be log (w/ω) ∈ Lp, and the p ), and therefore Lω p in the space (Lw p , ρw p , ρw Corollary 6.4 If w is a weight such that log+ w ∈ Lp, then Lp ⊂ Lw p , and Lp = Lw p if and only if log w ∈ Lp. Furthermore, the topology defined on Lp by the metric ρp is stronger than that induced on Lp by the metric ρw p , and if log w ∈ Lp, then these two topologies coincide. Proof. The proof follows directly from the assertions (i) and (ii) of Theorem 6.3, by putting ω(t) ≡ 1 on [0, 2π). Theorem 6.5 If log w ∈ Lp, then (Lp, ρw metric topologies ρw p and ρp are the same. p ) is an F -algebra, and the Proof. Since by Corollary 2.4, (Lp, ρp) is an F -algebra, to show that (Lp, ρw p ) is an F -algebra, it suffices by Corollary 6.4 to prove that p ) is complete. Let (fn) be a Cauchy sequence in (Lp, ρw (Lp, ρw p ). This means that (wfn) is a Cauchy sequence in (Lp, ρp), and by Corollary 2.4, there is a g ∈ Lp such that ρp(wfn, g) → 0 as n → ∞. Then p (fn, g/w) = ρp(wfn, g) → 0, and since ρw ρw p (g/w, 0) = ρp(g, 0) < ∞, we see that f = g/w ∈ Lw p ) is com- plete, and the theorem is proved. p = Lp. Hence, the space (Lp, ρw Remark. It is easy to see that the vector space Lw algebra with respect to pointwise addition and multiplication. p is not necessarily Corollary 6.6 The initial metric topology ρp is a unique ρω topology with log+ w ∈ Lp that makes (Lp, ρw p ) into an F -space. p metric Proof. Suppose that log+ w ∈ Lp, and that (Lp, ρw Consider the identity map p ) is an F -space. j : (Lp, ρp) → (Lp, ρw p ). Then by Corollary 6.4, j is continuous. Since by Theorem 6.2 (Lp, ρp) is an F -space, by the open mapping theorem, we conclude that the in- verse j−1 of j is also continuous. This shows that the metric topologies ρp and ρw p must be the same on Lp. 16 Theorem 6.7 Let p ≥ 1 and let w be a weight such that log+ w ∈ Lp. Then the following statements about w are equivalent. (i) log w ∈ Lp; (ii) Lw p = Lp; (iii) The metrics ρw (iv) (Lp, ρw (v) (Lp, ρw p ) is a complete metric space. p ) is an F -algebra; p and ρp define the same topology on Lp; Proof. (i)⇔(ii). Follows from Corollary 6.4. (iii)⇒(i). This is a consequence of the assertion (iii) of Theorem 6.3, by setting ω(t) ≡ 1 for t ∈ [0, 2π). (i)⇒(iii). This is immediate from Corollary 6.4. (iv)⇒(i). This follows from the assertion (iv) of Theorem 6.3, by setting ω(t) ≡ 1 for t ∈ [0, 2π). (i)⇒(v). This is immediate from Corollary 6.5. (v)⇒(iv). This is obvious. References [DS] N. Dunford and J. T. Schwartz, Linear operators I, Wiley -- Interscience, New York, 1958. [D] P. L. Duren, Theory of H p spaces, Academic Press, New York, 1970. [G] T. W. Gamelin, Uniform algebras, Prentice -- Hall, Englewood Cliffs, New Jersey, 1969. [L] R. Le´sniewicz, On linear functionals in Hardy -- Orlicz spaces, I, Studia Math. 46 (1973), 53 -- 77. [MP] R. Mestrovi´c and Z. Pavi´cevi´c, The logarithmic analogue of Szego's theorem, Acta Sci. Math. (Szeged) 64 (1998), 97 -- 102. [Mo] N. Mochizuki, Algebras of holomorphic functions between H p and N∗, Proc. Amer. Math. Soc. 105 (1989), 898 -- 902. 17 [Mu] J. Musielak, Orlicz spaces and modular spaces, Lecture Notes in Math. 1034, Springer -- Verlag, 1983. [MO] J. Musielak and W. Orlicz, Some remarks on modular spaces, Bull. Acad. Polon. Sci., 7 (1959), 661 -- 668. [S] M. Stoll, Mean growth and Taylor coefficients of some topological algebras of analytic functions, Ann. Polon. Math. 35 (1977), 139 -- 158. [P] I. I. Privalov, Boundary properties of analytic functions, Moscow University Press, Moscow, 1941; 2nd ed., GITTL, Moscow, 1950. (Russian) [Y] N. Yanagihara, Multipliers and linear functionals for the class N +, Trans. Amer. Math. Soc. 180 (1973), 449 -- 461. ∗ Maritime Faculty, University of Montenegro, 85330 Kotor, Mon- tenegro E-mail address: [email protected] ∗∗ Faculty of Science, University of Montenegro, 81000 Podgorica, Montenegro E-mail address: [email protected] ∗∗∗ Faculty of Science, University of Montenegro, 81000 Podgorica, Montenegro 18
1008.4958	1	1008	2010-08-29T21:01:26	Stampacchia's property, self-duality and orthogonality relations	[ "math.FA" ]	We show that if the conclusion of the well known Stampacchia Theorem, on variational inequalities, holds on a Banach space X, then X is isomorphic to a Hilbert space. Motivated by this we obtain a relevant result concerning self-dual Banach spaces and investigate some connections between existing notions of orthogonality and self-duality. Moreover, we revisit the notion of the cosine of a linear operator and show that it can be used to characterize Hilbert space structure. Finally, we present some consequences of our results to quadratic forms and to evolution triples.	math.FA	math	STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS NIKOS YANNAKAKIS Abstract. We show that if the conclusion of the well known Stampacchia Theorem, on variational inequalities, holds on a Banach space X, then X is isomorphic to a Hilbert space. Motivated by this we obtain a relevant result concerning self-dual Banach spaces and investigate some connections between existing notions of orthogonality and self-duality. Moreover, we revisit the notion of the cosine of a linear operator and show that it can be used to characterize Hilbert space structure. Finally, we present some consequences of our results to quadratic forms and to evolution triples. . A F h t a m [ 1 v 8 5 9 4 . 8 0 0 1 : v i X r a 1. Introduction Let H be a real Hilbert space, k · k be its norm, (· , ·) its inner product and let a : H × H → R be a bounded bilinear form. The well known Stampacchia Theorem (also called the Lions-Stampacchia The- orem, see [5], [18] and [19]) states that if the above bilinear form is coercive, i.e. there exists c > 0 such that (1.1) a(x, x) ≥ ckxk2, for all x ∈ H, then for any nonempty, closed, convex subset M of H and h ∈ H, there exists a unique solution x ∈ M , of the variational inequality (1.2) a(x, z − x) ≥ (h, z − x) , for all z ∈ M. Our first aim, in this paper, is to investigate whether Stampacchia's Theorem can be generalized in the broader setting of an arbitrary Banach space X. As we will see, at least in its full generality, this is impossible since its conclusion implies that X has to be isomorphic to a Hilbert space. In the sequel we obtain a relevant result concerning self-dual Banach spaces, i.e. Banach spaces that are isomorphic to their dual spaces. Along the way we see that our approach brings out some connections between existing notions of orthogonality in general normed linear spaces and self-duality. In the last section and motivated by the above, we revisit the cosine of a linear operator (a notion originally introduced by K. Gustafson in [6]) and use it to obtain an additional Hilbert space characterization based on a result of J. R. Partington which can be found in [17]. 2000 Mathematics Subject Classification. Primary 46C15; Secondary 47B99, 46B03. Key words and phrases. Variational inequality, complemented subspace, Hilbert space char- acterization, self-dual Banach space, positive operator, coercive operator, orthogonality relation, cosine of a linear operator, quadratic form, evolution triple. 1 2 NIKOS YANNAKAKIS Finally, we present some consequences of our results to quadratic forms and to evolution triples. 2. Stampacchia's property Let X be a real Banach space, X ∗ be its dual and h· , ·i be their duality product. By M ⊥ we denote the annihilator of a subspace M of X, i.e. M ⊥ = {x∗ ∈ X ∗ : < x∗, x >= 0 , for all x ∈ M } . To obtain the natural analogue of the conclusion of Stampacchia's Theorem in this situation we need the following definition. Definition 2.1. Let X be a real Banach space. We say that X has Stampacchia's property (property (S) for short), if there exists a bounded, bilinear form a : X × X → R such that if M is any nonempty, convex, closed subset of X and x∗ ∈ X ∗, then there exists a unique x ∈ M such that a(x, z − x) ≥ hx∗, z − xi , for all z ∈ M. Recall that a closed subspace M of a Banach space X is complemented in X, if there exists another closed subspace N of X such that X is their direct sum, i.e. M ∩ N = { 0} and X = M + N . Note that the existence of such a closed subspace N is equivalent to the existence of a bounded linear projection from X onto M . Not all closed subspaces of an arbitrary Banach space are complemented. In fact we have the following well-known result by J. Lindenstrauss and L. Tzafriri [14], which we will use in the sequel. Theorem 2.2. A Banach space X is isomorphic to a Hilbert space if and only if all its closed subspaces are complemented. To proceed with our task we need the following simple lemma. Lemma 2.3. Let X be a Banach space and M be a closed subspace of X. If there exists another Banach space Y and a bounded linear operator S : M → Y 1 − 1 and onto Y , which can be extended to the whole of X, then the closed subspace M is complemented in X. Proof. If S : X → Y denotes the extension of S to the whole of X, then it is easy to see that the operator S−1 ◦ S : X → M is the required bounded projection onto M . (cid:3) We can now show that property (S) characterizes Hilbert space structure. Theorem 2.4. A real Banach space X is isomorphic to a Hilbert space if and only if it has property (S). STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS 3 Proof. The neccesity is obvious. We prove that property (S) is also sufficient. To this end let M be any closed subspace of X. We will show that M is complemented. Since M is a closed subspace of X, it is easy to see that property (S) in particular implies that for all x∗ ∈ X ∗, there exists a unique x ∈ M such that (2.1) a(x, z) = hx∗, zi, for all z ∈ M. Define the bounded linear operator T : X → X ∗ , by and let hT x, zi = a(x, z), for all x, z ∈ X π : X ∗ → X ∗/M ⊥ be the natural quotient map. Then the restriction of the operator π ◦ T on the subspace M is 1 − 1 and onto X ∗/M ⊥. To see this first note that if (π ◦ T )x = 0 then T x ∈ M ⊥, i.e. hT x, zi = 0, for all z ∈ M . By the definition of T this implies that a(x, z) = 0, for all z ∈ M . But by hypothesis there is a unique x ∈ M such that a(x, z) = 0, for all z ∈ M , which by the boundedness of a has to be 0. Thus the restriction of the operator π ◦ T on the subspace M is 1 − 1. To show that π ◦ T M is also onto, let h = x∗ + M ⊥, for some x∗ ∈ X ∗. Then by (2.1) we have that there exists a unique x ∈ M such that a(x, z) = hx∗, zi, for all z ∈ M . Hence again by the definition of T we have that hT x, zi = hx∗, zi , for all z ∈ M , i.e. T x − x∗ ∈ M ⊥. Hence (π ◦ T )(x) = h and thus π ◦ T M is onto X ∗/M ⊥. Note now that by its definition the operator S = π ◦ T M can be trivially extended to the whole of X and thus by Lemma 2.3 the closed subspace M is a complemented subspace of X. Since M was arbitrary we get by Theorem 2.2 that X is isomorphic to a Hilbert space. (cid:3) Remark 2.5. A careful look in the above proof shows that if the Banach space X has property (S) and M is any closed subspace of X then X = M ⊕ T −1(M ⊥) , where T is the operator associated to the bilinear form a(·, ·). Remark 2.6. A main hypothesis in Stampacchia's Theorem is the coercivity condi- tion (1.1). As it is well-known (see for example [4], [13]), such a hypothesis cannot hold in an arbitrary Banach space X since if it did, then X would have an equivalent Hilbertian norm induced by the inner product (x, y) = 1 2 [a(x, y) + a(y, x)] and thus would be isomorphic to a Hilbert space. Hence our result implies that there can be no full generalization of Stampacchia's Theorem in an arbitrary Banach space even if one drops the coercivity condition (1.1). 4 NIKOS YANNAKAKIS Remark 2.7. Note that if we are restricted to bounded closed and convex subsets of a Banach space X, then a generalization of Stampacchia's Theorem is possible by just assuming that the bilinear form is strictly positive i.e. a(x, x) > 0 , for all x ∈ X. The proof is a straightforward application of a result due to Brezis [2, Theorem 24], on pseudomonotone operators. It is easy to see that in this case the space X need not be isomorphic to a Hilbert space. Remark 2.8. It seems appropriate to mention here a recent result by E. Ernst and M. Th´era: if as in Remark 2.7 we are restricted to bounded, closed and convex sets and moreover X is a Hilbert space, then the pseudomonotonicity of the operator associated to the bilinear form a(·, ·), is a necessary and sufficient condition for the existence of a solution of the variational inequality (1.2) (see [5, Theorem 3.1]). A similar result for unbounded sets has been obtained by A. Maugeri and F. Raciti in [15]. 3. Self-dual Banach spaces A self-dual Banach space is a Banach space isomorphic to its dual. It is well- known that Hilbert spaces are self-dual although they are far from being the only ones; if Y is any reflexive Banach space then X = Y ⊕ Y ∗ is self-dual. We will now see that our approach in Section 2 can lead us to a result concerning self-dual Banach spaces. The important observation is the fact that the operator T associated to the bilinear form a(·, ·), in the proof of Theorem 2.4, is an isomorphism from X onto X ∗ and hence X is a self-dual space. Our result is the following. Proposition 3.1. Let X be a real, self-dual, Banach space. If the isomorphism T : X → X ∗ onto X ∗ is such that for any closed subspace M of X, the map π ◦ T M is an isomorphism onto X ∗/M ⊥, where π is the natural quotient map from X ∗ onto X ∗/M ⊥, then X is isomorphic to a Hilbert space. Proof. We follow the proof of Theorem 2.4. (cid:3) Remark 3.2. Recalling that the quotient space X ∗/M ⊥ is isomorphic to M ∗ we can rephrase Proposition 3.1 as follows: "Let X be a self-dual space. If the isomorphism between X and X ∗ induces in a natural way (through the natural quotient maps) isomorphisms between all closed subspaces of X and their corresponding duals, then X is isomorphic to a Hilbert space". As one can easily see, a necessary and sufficient condition for π ◦ T M to be an isomorphism (not necessarily onto) from M into X ∗/M ⊥, is the existence of a positive constant c, such that whenever x ∈ X and x∗ ∈ X ∗ are such that hx∗, xi = 0, we have that (3.1) T x + x∗ ≥ cT x . STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS 5 In order to give some geometric intuition to condition (3.1) we recall the following definition. Definition 3.3. Let X be a normed space and x , y ∈ X. We say that x is orthogonal, in the sense of Birkhoff-James, to y if x + λy ≥ x, for all λ ∈ R. For more details about this notion of orthogonality the interested reader is re- ferred to [1] and [11]. It is easy to see that if whenever x ∈ X and x∗ ∈ X ∗ are such that hx∗, xi = 0, we have that T x ⊥ x∗ , in the sense of Birkhoff-James, then T satisfies condition (3.1). As a matter of fact Birkhoff-James orthogonality is not the only orthogonality relation that can be used to guarantee the validity of condition (3.1). To see this we recall that in [17], J. R. Partington has introduced the concept of boundedness for an orthogonality relation in an arbitrary normed space as follows. Definition 3.4. An orthogonality relation ⊥ in a normed linear space is bounded if there exists c > 0 such that if x⊥y then λx + y ≥ cx, whenever λ ≥ c. Several well-known orthogonality relations (for example Birkhoff-James or Dimin- nie orthogonality, see [3] and [17] for more details) are bounded. Definition 3.5. An orthogonality relation ⊥ in a normed linear space is homoge- neous if x⊥y implies that ax⊥by, for all a, b ∈ R. Remark 3.6. In [16] it was shown that if an orthogonality relation ⊥ is homogeneous, then its boundedness is equivalent to the existence of c > 0, such that x⊥y implies x + y ≥ cx. Therefore if whenever x ∈ X and x∗ ∈ X ∗ are such that hx∗, xi = 0, we have that T x ⊥ x∗, for a homogeneous and bounded orthogonality relation ⊥, then T satisfies (3.1). To state our next result we need one more definition. Definition 3.7. An orthogonality relation ⊥, in a normed linear space, is non- degenerate, if x⊥x implies that x = 0. We can now prove the following Hilbert space characterization. Theorem 3.8. A real reflexive Banach space X is isomorphic to a Hilbert space if and only if there exists an isomorphism T : X → X ∗ , onto X ∗, such that (3.2) T x ⊥ x∗, whenever hx∗, xi = 0 , for a non-degenerate, homogeneous and bounded orthogonality relation ⊥ in X ∗. 6 NIKOS YANNAKAKIS Proof. The necessity is obvious. To prove the sufficiency of our claim we will use Proposition 3.1. To this end let M be any closed subspace of X. By (3.2) and the discussion above, the operator T satisfies condition (3.1) and hence π ◦ T M is an isomorphism. It remains to show that π ◦ T M is onto X ∗/M ⊥. Since (π ◦ T )(M ) is closed it is enough to show that it is a dense subspace of X ∗/M ⊥. Assume the contrary i.e. (π ◦ T )(M ) 6= X ∗/M ⊥ . Then by the Hahn-Banach Theorem there exists 0 6= f ∈ (X ∗/M ⊥)∗ such that f (T x) = 0 , for all x ∈ M . Since X is reflexive so is M and hence it is isometrically isomorphic to (X ∗/M ⊥)∗. Therefore there exists x ∈ M , such that hT x, xi = f (T x) = 0 and thus again by (3.2) we get that T x⊥T x. Using the non-degeneracy of ⊥ and the injectivity of T we get that x and consequently f have to be 0, which is a contradiction. Hence π ◦ T M is an isomorphism onto X ∗/M ⊥ and by Proposition 3.1 the self- (cid:3) dual Banach space X is isomorphic to a Hilbert space. 4. The cosine of a linear operator revisited A simple situation where condition (3.1) holds is when there exists c > 0, such that the operator T satisfies (4.1) hT x, xi ≥ cT x2, for all x ∈ X . Recall the following well-known definition. Definition 4.1. Let X be a real Banach space. We say that the linear operator T : D(T ) ⊆ X → X ∗ (i) is positive, if hT x, xi ≥ 0, for all x ∈ D(T ). (ii) is strictly positive, if hT x, xi > 0, for all x ∈ D(T ), with x 6= 0. (iii) is coercive, if there exists c > 0, such that hT x, xi ≥ cx2 , for all x ∈ D(T ). (iv) is symmetric, if hT x, yi = hT y, xi, for all x , y ∈ D(T ). Note that since in all our previous considerations (in Section 3), the operator T was an isomorphism inequality (4.1) would imply that the operator T was actually coercive. In the general case though, operators satisfying (4.1) form a much larger class than that of coercive operators. For example, see [4] and [10] for more details, any positive, everywhere defined and symmetric operator T satisfies (4.1). On the other hand, unlike coercivity (see Remark 2.6) condition (4.1) cannot guarantee on its own - i.e. when T is no longer an isomorphism but just a continuous linear operator - the Hilbertian structure of X. Note that this is still the case even if T has additional nice properties such as symmetry and positivity. It seems therefore quite natural that there may be some room between these two classes. To make things more precise we need the following definition. STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS 7 Definition 4.2. Let X be a real Banach space and let T : D(T ) ⊆ X → X ∗ be a positive linear operator. The cosine of T is defined as follows: (4.2) cos T = inf (cid:26) hT x, xi T x x , for all 0 6= x ∈ D(T ), such that T x 6= 0(cid:27) . Using expression (4.2) one can define the angle φ(T ) of the linear operator T , which has an obvious geometric interpretation: it measures the maximum turning effect of T . The above concepts were introduced, in the context of a complex Hilbert space, by K. Gustafson in [6] and have attracted a lot of interest since then. We refer the interested reader to the book of K. Gustafson and D. Rao [8], for more details. In order for the cosine of an operator to be a reliable tool, distinguishing between operators with different properties, it has to be positive for a large class of linear operators. As one can easily see this is the case for coercive everywhere defined - thus continuous - linear operators. On the other hand things fail dramatically for unbounded linear operators: it was shown by K. Gustafson and B. Zwahlen in [7] and by P. Hess (in a somewhat more general context) in [9], that the cosine of an unbounded linear operator is always 0. To return to our main theme note that if cos T > 0, then T satisfies (4.1). Thus non-coercive operators with positive cosine form the aforementioned intermediate class, between (4.1) and coercivity. It turns out, as we shall see below, that if X is not isomorphic to a Hilbert space then this class is quite small. We need one more definition. Definition 4.3 ([17]). An orthogonality relation ⊥, in a normed linear space X is (i) symmetric, if x⊥y implies y⊥x. (ii) right additive, if x⊥y and x⊥z implies x⊥(y + z). (iii) resolvable, if for any x , y there exists a ∈ R, such that x⊥(ax + y). (iv) continuous, if xn → x, yn → y and xn⊥yn, then x⊥y. It should be noted that an orthogonality relation having all six properties of Definitions 3.5, 3.7 and 4.3 (i.e. except boundedness) exists in any separable Banach space (see Theorem 3 in [17]). If boundedness is added things change drastically as the following result of J. R. Partington [17] illustrates. Theorem 4.4 ([17], Theorem 4). If X is a Banach space and ⊥ is an orthogonal- ity relation in X, that is non-degenerate, symmetric, homogeneous, right additive, resolvable, continuous and bounded, then X is isomorphic to a Hilbert space. In the sequel we identify T ∗ with the restriction on X of the adjoint of the linear operator T : X → X ∗ (which is defined on the whole of X ∗∗). Using Theorem 4.4, we can prove our main result for this section. Theorem 4.5. Let X be a real Banach space, not isomorphic to a Hilbert space and a positive linear operator. If there exists c > 0, such that T ∗x ≤ cT x , for all x ∈ X , (4.3) T : X → X ∗ 8 NIKOS YANNAKAKIS then cos T = 0. Proof. Assume the contrary and let cos T = δ > 0. The linear operator defined by S : X → X ∗ S = 1 2 (T + T ∗) . is strictly positive, everywhere defined and hence continuous. We define the follow- ing orthogonality relation in X: x⊥y , if hSx, yi = 0 . It is easy to see that ⊥ is non-degenerate, symmetric, homogeneous, right additive, resolvable and continuous. To see that ⊥ is also bounded take x⊥y with x 6= 0. Then hx∗, x + yi x + y = sup x∗6=0 hSx, x + yi x∗ = hT x, xi Sx ≥ ≥ ≥ Sx 2hT x, xi T x(1 + c) 2δ 1 + c x , where the second inequality is justified by (4.3). Since ⊥ is homogeneous, by Remark 3.6, the orthogonality relation ⊥ is also bounded. Hence by Theorem 4.4 the Banach space X is isomorphic to a Hilbert space, (cid:3) which is a contradiction. Thus cos T = 0. Remark 4.6. The class of operators satisfying (4.3) is quite large as it includes positive, everywhere defined, symmetric linear operators. Combining this last remark with Theorem 4.5 we can have the following simple Hilbert space characterization. Corollary 4.7. A real Banach space X is isomorphic to a Hilbert space if and only if there exists a positive and symmetric linear operator with cos T > 0. T : X → X ∗ It seems quite interesting that if X is not isomorphic to a Hilbert space then an operator and its adjoint cannot have both positive cosines. Proposition 4.8. Let X be a real Banach space, not isomorphic to a Hilbert space and a positive linear operator with cos A > 0. Then cos A∗ = 0. A : X → X ∗ STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS 9 Proof. Assume cos A = δ > 0 and let x 6= 0. Then hA∗x, yi A∗x = sup y6=0 hAx, xi y ≥ x ≥ δAx . If T = A∗, then T is a positive linear operator that satisfies (4.3). Thus by Theorem 4.5 we get that cos A∗ = 0. (cid:3) 4.1. An application to quadratic forms. Recall that a continuous quadratic form on a normed space X is a function for which there exists a bounded bilinear form q : X → R such that a : X × X → R q(x) = 1 2 a(x, x) . It is well known (see for example [12]) that there exists a one-to-one correspondence between continuous quadratic forms and symmetric linear operators through the formula (4.4) T : X → X ∗ q(x) = 1 2 hT x, xi . Moreover, each continuous quadratic form is everywhere Frechet differentiable and its derivative is equal to 2T , where T is the symmetric operator in (4.4). Using Corollary 4.7, we can have the following result. Proposition 4.9. Let X be a real Banach space, not isomorphic to a Hilbert space and q : X → R a continuous quadratic form. Then for any ε > 0, there exists x ∈ X, such that q(x) < εq′(x)) x . Proof. If q(x) < 0, for some x ∈ X we are done. If this is not the case then the symmetric linear operator T that generates q, is positive and thus by Corollary 4.7 Hence for any ε > 0, there exists x 6= 0, such that cos T = 0 . hT x, xi T x x < ε . Since q(x) = 1 2 hT x, xi and q′ = 2T the result follows. (cid:3) 10 NIKOS YANNAKAKIS 4.2. An application to evolution triples. We end this paper with an application of Theorem 4.5 to evolution triples. Recall that we say that a Banach space is continuously and densely embedded into another Banach space Y , if there exists an injective, bounded linear operator i : X → Y such that i(X) is dense in Y . We have the following Proposition. Proposition 4.10. Let X be a real reflexive Banach space that is continuously and densely embedded into a Hilbert space H and assume that X is not isomorphic to a Hilbert space. Then for any ε > 0, there exists x ∈ X, such that i(x)2 H < εi∗(i(x))X ∗ xX , where i is the embedding operator from X into H. Proof. Since i : X → H is the embedding operator from X into H and the embedding is continuous and dense, then (after identifying H with its dual space H ∗) the embedding i∗ : H → X ∗ is also continuous and dense (we say that X, H and X ∗ form an evolution or a Gelfand triple). Let T : X → X ∗ be defined by T = i∗ ◦ i. Then T is a strictly positive, symmetric operator and by Corollary 4.7 we have that Thus for any ε > 0, there exists x ∈ X, such that cos T = 0 . hT x, xi T xX ∗xX < ε. But hT x, xi = (i(x), i(x))H = i(x)2 H and hence the result follows. (cid:3) A concrete example of the above situation is the following: Example 4.11. Let Ω ⊆ RN , open and bounded and assume p > 2. Then for every ε > 0, there exists f ∈ Lp(Ω), such that f 2 L2(Ω) < εf Lq(Ω)f Lp(Ω) , where 1 p + 1 q = 1. Proof. Let i : Lp(Ω) → L2(Ω) be the identity operator and use the previous propo- sition. (cid:3) Acknowledgment 4.12. The author would like to thank Dr. D. Drivaliaris and Mr. M. Garagai for many fruitful discussions. STAMPACCHIA'S PROPERTY, SELF-DUALITY AND ORTHOGONALITY RELATIONS 11 References [1] D. Amir, Characterizations of inner product spaces, Oper. Theory Adv. Appl., vol. 20, Birkhuser, Basel, 1986. [2] H. Brezis, ´Equations et in´equations non lin´eaires dans les espaces vectoriels en dualit´e, Ann. Inst. Fourier (Grenoble) 18 (1968), 115 -- 175. [3] C. R. Diminnie, A new orthogonality relation for normed linear spaces, Math. Nachr. 114 (1983), 197 -- 203. [4] D. Drivaliaris, N. Yannakakis, Hilbert space structure and positive operators, J. Math Anal. Appl. 305 (2005), no. 2, 560 -- 565. [5] E. Ernst, M. Th´era, A converse to the Lions-Stampacchia theorem, ESAIM Control Optim. Calc. Var. 15 (2009), no. 4, 810 -- 817. [6] K. Gustafson, The angle of an operator and positive operator products, Bull. Amer. Math. Soc., 74 (1968), 488 -- 492. [7] K. Gustafson, B. Zwahlen, On the cosine of unbounded operators, Acta Sci. Math. 30 (1969), 33 -- 34. [8] K. Gustafson, D. Rao, Numerical Range. The field of values of linear operators and matrices. Universitext. Springer-Verlag, New York, 1997. [9] P. Hess, A remark on the cosine of linear operators, Acta Sci. Math. 32 (1971), 267 -- 269. [10] P. Hess, A remark on a class of linear monotone operators, Math. Z. 125 (1972), 104 -- 106. [11] V. Istrat¸escu, Inner product structures. Theory and applications. Mathematics and its Ap- plications, vol. 25, D. Reidel Publishing Co., Dordrecht, 1987. [12] N. Kalton, S. Konyagin, L. Vesel´y, Delta-semidefinite and delta-convex quadratic forms in Banach spaces, Positivity 12 (2008), 221 -- 240. [13] B. L. Lin, On Banach spaces isomorphic to its conjugate, Studies and essays (presented to Yu-why Chen on his 60th birthday, April 1, 1970), Math. Res. Center, Nat. Taiwan Univ., Taipei, (1970), 151 -- 156. [14] J. Lindenstrauss, L. Tzafriri, On the complemented subspaces problem, Israel J. Math. 9 (1971), 263 -- 269. [15] A. Maugeri, F. Raciti, On existence theorems for monotone and nonmonotone variational inequalities, J. Convex Anal. 16 (2009), no. 3-4, 899 -- 911. [16] P. M. Milici´c, On isomorphisms by orthogonality of a normed space and an inner product space, Publ. Inst. Math. (Beograd) (N.S.) 59(73) (1996), 89 -- 94. [17] J. R. Partington, Orthogonality in normed spaces, Bull. Austral. Math. Soc. 33 (1986), no. 3, 449 -- 455. [18] G. Stampacchia, Formes bilin´eaires coercitives sur les ensembles convexes, C. R. Acad. Sci. Paris 258 (1964), 4413 -- 4416. [19] J. L. Lions, G. Stampacchia, Variational inequalities, Comm. Pure Appl. Math. 20 (1967), 493 -- 519. Department of Mathematics, National Technical University of Athens, Iroon Poly- texneiou 9, 15780 Zografou, Greece E-mail address: [email protected]
1207.0666	2	1207	2013-03-26T18:12:20	Continuous slice functional calculus in quaternionic Hilbert spaces	[ "math.FA", "hep-th", "math-ph", "math.CV", "math-ph", "math.OA" ]	The aim of this work is to define a continuous functional calculus in quaternionic Hilbert spaces, starting from basic issues regarding the notion of spherical spectrum of a normal operator. As properties of the spherical spectrum suggest, the class of continuous functions to consider in this setting is the one of slice quaternionic functions. Slice functions generalize the concept of slice regular function, which comprises power series with quaternionic coefficients on one side and that can be seen as an effective generalization to quaternions of holomorphic functions of one complex variable. The notion of slice function allows to introduce suitable classes of real, complex and quaternionic $C^$--algebras and to define, on each of these $C^$--algebras, a functional calculus for quaternionic normal operators. In particular, we establish several versions of the spectral map theorem. Some of the results are proved also for unbounded operators. However, the mentioned continuous functional calculi are defined only for bounded normal operators. Some comments on the physical significance of our work are included.	math.FA	math	CONTINUOUS SLICE FUNCTIONAL CALCULUS IN QUATERNIONIC HILBERT SPACES RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI Abstract. The aim of this work is to define a continuous functional calculus in quaternionic Hilbert spaces, starting from basic issues regarding the notion of spherical spectrum of a normal operator. As properties of the spherical spectrum suggest, the class of continuous functions to consider in this setting is the one of slice quaternionic functions. Slice functions generalize the con- cept of slice regular function, which comprises power series with quaternionic coefficients on one side and that can be seen as an effective generalization to quaternions of holomorphic functions of one complex variable. The notion of slice function allows to introduce suitable classes of real, complex and quater- nionic C ∗ –algebras and to define, on each of these C ∗ –algebras, a functional calculus for quaternionic normal operators. In particular, we establish several versions of the spectral map theorem. Some of the results are proved also for unbounded operators. However, the mentioned continuous functional cal- culi are defined only for bounded normal operators. Some comments on the physical significance of our work are included. Contents 1. Introduction 2. Quaternionic Hilbert spaces 3. Left multiplications, imaginary units and complex subspaces 4. Resolvent and spectrum 5. Real measurable functional calculus for self–adjoint operators and slice nature of normal operators 6. Relevant C ∗ -algebras of slice functions 7. Continuous slice functional calculus for normal operators References 1 7 17 25 35 50 59 69 1. Introduction Functional calculus in quaternionic Hilbert spaces has been focused especially by mathematical physicists, concerning the application of spectral theory to quantum theories (see e.g. [1, 12, 14, 21]). As a careful reading of these works reveals, most part of results have been achieved into a non completely rigorous fashion, leaving some open gaps and deserving further investigation. As a matter of fact, the clas- sic approach for complex Hilbert spaces, starting from the continuous functional calculus and then reaching the measurable functional calculus, has been essentially 2010 Mathematics Subject Classification. 46S10, 47A60, 47C15, 30G35, 32A30, 81R15. Work partially supported by GNSAGA and GNFM of INdAM. 1 2 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI disregarded, while attention has been devoted to the measurable functional cal- culus almost immediately. Furthermore, even more mathematically minded works on this topics, as [34], do not present complete proofs of the claimed statements. Historically, an overall problem was the absence of a definite notion of spectrum of an operator on quaternionic Hilbert spaces. Such a notion has been introduced only few years ago [8] in the more general context of operators on quaternionic Banach modules. Therefore, on the one hand, no systematic investigation on the spectral properties of operators on quaternionic Hilbert spaces has been performed up to now. On the other hand, the many key–results available for non self–adjoint operators have been obtained by means of a Dirac bra–ket like formalism, which formally, but often erroneously, reduces the argumentations to the case of a vague notion of point spectrum. Finally, another interesting issue concerns the existence of a model of quaternions in terms of anti–self adjoint and unitary operators, com- muting with the self–adjoint parts of a given normal operator. Although that model is not completely understood and analyzed, it was extensively exploited in various technical constructions of physicists (see [1] and Theorem 5.14 below). The aim of this work is to provide a foundational investigation of continuous functional calculus in quaternionic Hilbert spaces, particularly starting from basic issues regarding the general notion of spherical spectrum (Definition 4.1) and its general properties (Theorems 4.3 and 4.8 and Propositions 4.5 and 4.7). The general relation between the theory in complex Hilbert space and the one in quaternionic Hilbert spaces will be examined extending some classic known results (see e.g. [12]) to the unbounded operator case (Proposition 3.11). In view of that general aim, up to Section 4.2, we shall not confine ourselves to the bounded operator case, but we shall consider also unbounded operators. However, the proper continuous functional calculus will be discussed for bounded normal operators only, thus postponing the non–bounded case to a work in preparation [18]. In particular, in the first part of the work, the general spectral properties of bounded and unbounded operators on quaternionic Hilbert spaces will be discussed from scratch. The pivotal tool in our investigation is the notion of slice function, whose rele- vance clearly pops up once the notion of spherical spectrum is introduced. As we shall prove, in the continuous case, that notion allows one to introduce suitable classes of real, complex and quaternionic C ∗ –algebras and to define, on each of these C ∗ –algebras, a functional calculus for normal operators. In particular, we establish several versions of the spectral map theorem. As our results show, the in- terplay between continuous slice functions and the space of operators is much more complicated than in the case of continuous functional calculus on complex Hilbert spaces. 1.1. Slice functions, a key result and the main theorems. The concept of slice regularity for functions of one quaternionic variable has been introduced by Gentili and Struppa in [15, 16] and then extended to octonions, Clifford algebras and in general real alternative ∗–algebras in [7, 17, 19, 20]. This function theory comprises polynomials and power series in the quaternionic variable with quater- nionic coefficients on one side. It can then be seen as an effective generalization to quaternions of the theory of holomorphic functions of one complex variable. At the base of the definition of slice regularity, there is the “slice” character of the quaternionic algebra: every element q ∈ H can be decomposed into the form q = α+ β , where α, β ∈ R and  is an imaginary unit in the two–dimensional sphere QUATERNIONIC SLICE FUNCTIONAL CALCULUS 3 S = {q ∈ H q2 = −1}. This decomposition is unique when assuming further that β ≥ 0; otherwise, for non–real q , β and  are determined up to a sign. Equivalently, H is the union of the (commutative) real subalgebras C ≃ C generated by  ∈ S, with the property that Cı ∩ C = R when ı 6= ±, where R denotes the real subalgebra of H generated by 1. The original definition [15, 16] of slice regularity for a quaternionic function f , defined on an open domain Ω of H, requires that, for every  ∈ S, the restriction of f to Ω ∩ C is holomorphic with respect to the complex structure defined by left multiplication by . The approach taken in [19, 20] allows to embed the space of slice regular functions into a larger class, that of continuous slice functions, which corresponds in some sense to the usual complex continuous functions on the complex plane. The first step is to single out a peculiar class of subsets of H, those that are invariant with respect to the action of S. If K ⊂ C is non–empty and invariant under complex conjugation, one defines the circularization ΩK of K (in H) as: ΩK = {α + β ∈ H α, β ∈ R, α + iβ ∈ K,  ∈ S}, and call a subset of H a circular set if it is of the form ΩK for some K. The second step is the introduction of stem functions. Let H ⊗R C be the complexified quaternionic algebra, represented as H ⊗R C = {x + iy x, y ∈ H}, with complex conjugation w = x + iy 7→ w = x − iy . If a function F : K −→ H ⊗R C satisfies the condition F (z ) = F (z ) for every z ∈ K, then F is called a stem function on K. Any stem function F : K −→ H ⊗R C induces a (left) slice function f = I (F ) : ΩK → H: if q = α + β ∈ ΩK ∩ C , with  ∈ S, we set f (q) := F1 (α + iβ ) + F2 (α + iβ ) , where F1 , F2 are the two H–valued components of F . In this approach to the theory, a quaternionic function turns out to be slice regular if and only if it is the slice function induced by a holomorphic stem function. The definition of a continuous slice function of a normal operator is based on a key result, which can interpreted as the operatorial counterpart of the slice character of H and which describes rigorously the just mentioned model of quaternions for bounded normal operators exploited by physicists. Theorem 5.9 in Section 5.4 can be reformulated as follows: Theorem J. Let H be a quaternionic Hilbert space and let B(H) be the set of al l bounded operators of H. Given any normal operator T ∈ B(H), there exist three operators A, B , J ∈ B(H) such that: (i) T = A + J B , (ii) A is self–adjoint and B is positive, (iii) J is anti self–adjoint and unitary, (iv) A, B and J commute mutual ly. Furthermore, the fol lowing additional facts hold: • A and B are uniquely determined by T : A = (T + T ∗) 1 2 and B = T − T ∗ 1 2 . • J is uniquely determined by T on Ker (T − T ∗)⊥ . In the parallelism between this decomposition of T and the slice decomposition of quaternions, real numbers in H correspond to self–adjoint operators, non–negative 4 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI real numbers to positive operators and quaternionic imaginary units to anti self– adjoint and unitary operators. This parallelism suggests a natural way to define the operator f (T ) for a continuous slice function f , at least in the case of H– instrinsic slice functions, i.e. those slice functions such that f ( ¯q) = f (q) for every q ∈ ΩK . Observe that H–intrinsic slice functions leave all slices C invariant. If f is a polynomial slice function, induced by a stem function with polynomials components F1 , F2 ∈ R[X, Y ], then f is H–intrinsic and we define the normal operator f (T ) ∈ B(H) by setting f (T ) := F1 (A, B ) + J F2 (A, B ), and then we extend the definition to continuous H–intrinsic slice functions by den- sity. We obtain in this way a isometric ∗–homomorphism of real Banach C ∗ – algebras (Theorem 7.4). In particular, continuous H–intrinsic slice functions satisfy the spectral map property f (σS (T )) = σS (f (T )), where σS (T ) denotes the spherical spectrum of T . The definition of f (T ) can be extended to other classes of continuous slice func- In this transition, some of the nice properties of the map f 7→ f (T ) are tions. lost, but other interesting phenomena appear. In Theorem 7.8, the slice functions In this case, a ∗– considered are those which leave only one slice C invariant. homomorphism of complex Banach C ∗–algebras is obtained. A suitable form of the spectral map property continues to hold. In Theorem 7.10, we consider circular slice functions, those which satisfy the condition f ( ¯q) = f (q) for every q . Differently from the previous cases, these functions form a non–commutative quaternionic Ba- nach C ∗–algebra. In this case, we still have an isometric ∗–homomorphism, but the spectral map property holds in a weaker form. In Proposition 7.12, we show how to extend the previous definitions of f (T ) to a generic continuous slice function f , but in this case also the ∗–homomorphism property is necessarily lost. Finally, in Section 7.5, we show that the continuous functional calculus defined above, when restricted to slice regular functions, coincides with the functional calculus developed in [8] as a generalization of the classical holomorphic functional calculus. 1.2. Physical significance. As remarked by Birkhoff and von Neumann in their celebrated seminal work on Quantum Logic in 1936 [4], Quantum Mechanics may alternatively be formulated on a Hilbert space where the ground field of complex numbers is replaced for the division algebra of quaternions. Nowadays, the picture is more clear on the one hand and more strict on the other hand, after the efforts started in 1964 by Piron [29] and concluded in 1995 by Sol`er [32], and more recently reformulated by other researchers (see e.g. [2]). Indeed, it has been rigorously established that, assuming that the set of “yes–no” elementary propositions on a given quantum system are described by a lattice that is bounded, orthomodular, atomic, separable, irreducible, verifying the so–called covering property (see [3, 13, 25]) and finally, assuming that certain orthogonal systems exist therein, then the lattice is isomorphic to the lattice of orthogonal pro jectors of a generalised Hilbert space over the fields R, C or over the division algebra of quaternions H. No further possibility is allowed. Actually, the first possibility is only theoretical, since it has been proved that, dealing with concrete quantum systems, the description of the time–reversal operation introduces a complex structure in the field that makes, indeed, the real Hilbert space a complex Hilbert space (see [1]). Therefore, it seems that the only two realistic possibilities allowed by Nature are complex QUATERNIONIC SLICE FUNCTIONAL CALCULUS 5 Hilbert spaces and quaternionic ones. While the former coincides to the mainstream of the contemporary quantum physics viewpoint, the latter has been taken into consideration by several outstanding physicists and mathematical physicists, since the just mentioned paper of Birkhoff and von Neumann. Adler’s book [1] represents a quite complete treatise on that sub ject from the point of view of physics. While all the fundamental results, like Gleason theorem and Wigner theorem, can be re–demonstrated in quaternionic quantum mechanics with minor changes (see [33]), its theoretic formulation differs from the complex formulation in some key points related to the spectral theory, the proper language of quantum mechanics. Perhaps, the most important is the following. Exactly as in the standard approach, observables are represented by (generally unbounded) self–adjoint operators. How- ever, within the complex Hilbert space picture, self–adjoint operators enter the theory also from another route due to the celebrated Stone theorem. Indeed, self– adjoint operators, when multiplied with i, become the generators of the continuous one–parameter groups of unitary operators representing continuous quantum sym- metries. More generally, in complex Hilbert spaces, in view of well–known results due to Nelson [27], anti self–adjoint operators are the building blocks necessary to construct strongly continuous unitary representations of Lie groups of quantum symmetries. This way naturally leads to the quantum version of Noether theorem relating conserved quantities (i times the anti self–adjoint generators representing the Lie algebra of the group) and symmetries (the one–parameter groups obtained by exponentiating the given anti self–adjoint generators) of a given quantum sys- tem. This nice interplay, in principle, should survive the passage from complex to quaternionic context. However, a difficult snag pops up immediately: the relation- ship between self–adjoint operators S and anti self–adjoint operators A is much more complicated in quaternionic Hilbert spaces than in complex Hilbert spaces. Indeed, in the quaternionic case, an identity as A = J S holds, where J is an op- erator replacing the trivial i in complex Hilbert spaces. Nevertheless, A does not uniquely fix the pair J, S so that the interplay of dynamically conserved quantities and symmetries needs a deeper physical investigation in quaternionic quantum me- chanics. This issue affects all the physical construction from scratch as it is already evident from the various mathematically inequivalent attempts to provide a phys- ically sound definition of the momentum operator of a particle. Furthermore, one has to employ quite sophisticated mathematical tools as the quaternionic version of Mackey’s imprimitivity theorem (see [6]). The operator J has to satisfy several constraints, first of all, it has to commute with S and it has to be anti self–adjoint. The polar decomposition theorem, re–formulated in quaternionic Hilbert spaces, provides such an operator, at least for bounded anti self–adjoint operators A. In some cases, it is convenient for technical reasons to look for an operator J that is also unitary and that it is accompanied by two other similar operators I and K , commuting with S , such that they define a representation of the imaginary quater- nions in terms of operators (see Section 2.3 of [1]). This is not assured by the polar decomposition theorem and the existence of such anti self–adjoint and unitary op- erators I , J, K is by no means obvious. This is one of the key issues tackled in this work. Indeed, in Theorem 5.9, we prove that, every normal operator T can always be decomposed as T = A + J B , with A, B self–adjoint and J anti self–adjoint, unitary and commuting with both A and B . Moreover, J can always be written as Lı for some (quaternionic) imaginary unit ı, where H ∋ q 7→ Lq ∈ B(H) is a 6 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI ∗–representation of H in terms of bounded operators, commuting with A and B (see Theorem 5.14). The systematic investigation of the properties of the above–mentioned operators J will enable us to state and prove a quite general proposition (Proposition 3.11 ex- tending previous results by Emch [12]) concerning the extension to the whole quater- nionic Hilbert space and the properties of such extensions of (generally unbounded) operators initially defined on complex subspaces of the quaternionic Hilbert space induced by J . As we have just recalled, when looking at this matter with a mathematically minded intention, the most problematic issue in the known literature on quater- nionic quantum mechanics is the notion of spectrum of an operator. As is known, in the case of self–adjoint operators representing quantum observables, the Borel sets in the spectrum of an observable account for the outcomes of measurements of that observable. The definition of spectrum of an operator in quaternionic Hilbert spaces turns out to be problematic on its own right in view of the fact that quaternions are a non commutative ring and it generates troubles, already studying eigenvalues and eigenvectors: an eigenspace is not a subspace because it is not closed under multipli- cation with scalars. However, at first glance, all that could not appear as serious as it is indeed, because self–adjoint operators should have real spectrum where quater- nionic noncommutativity is ineffective. Nevertheless, as stressed above, self–adjoint operators are not the only class of operators relevant in quaternionic Hilbert spaces for quantum physics, since also unitary and anti self–adjoint play some important role. Therefore, a full fledged notion of spectrum for normal operators should be in- troduced. In spite of some different formulations of the spectral theorem for normal operators [34], a notion of the spectrum and of a generic operator in quaternionic Hilbert spaces as well as a systematic investigation on its properties do not exist. As a matter of fact, mathematical physicists [12, 14, 21] always tried to avoid to face this issue confining their investigations to (bounded) self–adjoint operators and passing to other classes of operators by means of quite ad hoc arguments and very often, dealing with the help of Dirac bra–ket formalism, not completely jus- tified in these contexts. This remark concerns physicists, [1] in particular, who adopt the popular and formal bra–ket procedure, assume the naive starting point where the spectrum is made of eigenvalues even when that approach is evidently untenable and should be handled with the help of some rigged quaternionic Hilbert space machinery similar to Gelfand’s theory in complex Hilbert spaces. While all these approaches are physically sound and there are no doubts that the produced results are physically meaningful, an overall rigorous mathematical formulation of the quaternionic spectral theory still does not exist. The absence of a suitable no- tion of spectrum has generated a lack in the natural development of the functional calculus. As a matter of fact, in the complex Hilbert space theory, the functional calculus theory on the spectrum starts by the definition of a continuous function of a given normal operator and then passes to define the notion of measurable function of the operator. The pro jector–valued measures exploited to formulate the spectral theory, the last step of the story, are subsequently constructed taking advantage of the measurable functional calculus. In quaternionic Hilbert space, instead, a for- mulation of the spectral theorem exists [34] without any systematic investigation of the continuous and measurable functional calculus, nor an explicit definition of the spectrum of an operator. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 7 A pivotal notion in this paper is just that of spherical spectrum of an operator in a quaternionic Hilbert space. We state it (Definition 4.1) by specializing (and re–elaborating in the case of unbounded operators) the definition recently proposed by Colombo, Sabadini and Struppa in [8] for operators in quaternionic two–side Ba- nach modules. We see that, in spite of the different definition (based on a second order polynomial), the properties of the spectrum of an operator (Theorem 4.8) are natural generalizations of those in complex Hilbert spaces. In particular, the point spectrum, regardless an apparently inequivalent definition, turns out to be the set of eigenvalues exactly as in the complex Hilbert space case (Proposition 4.5). Moreover, a nice relationship appears (Proposition 5.11) between the stan- dard notion of spectrum in complex Hilbert spaces and the quaternionic notion of spectrum, for those operators that have been obtained as extensions of operators on complex Hilbert subspaces as pointed out above. The definition and the properties of the spectrum not only allow to construct a natural extension of the continuous functional calculus to the quaternionic Hilbert space case, but they permit us to discover an intriguing interplay of the continuous functional calculus and the theory of slice functions. 2. Quaternionic Hilbert spaces In this part, we summarize some basic notions about the algebra of quaternions, quaternionic Hilbert spaces and operators (even unbounded and defined in proper subspaces) on quaternionic Hilbert spaces. In almost all cases, the proofs of the various statements concerning quaternionic Hilbert spaces and operators thereon are very close to the analogues for the complex Hilbert space theory. Therefore, we shall omit the corresponding proofs barring some comments if necessary. 2.1. Quaternions. The space of quaternions H is the four dimensional real algebra with unity we go to describe. We denote by 0 the null element of H and by 1 the multiplicative identity of H. The space H includes three so–called imaginary units, which we indicate by i, j, k . By definition, they satisfy (we omit the symbol of product of the algebra): (2.1) ii = j j = kk = −1 . ij = −j i = k , ki = −ik = j, j k = −kj = i, The elements 1, i, j, k are assumed to form a real vector basis of H, so that any element q ∈ H takes the form: q = a1 + bi + cj + dk , where a, b, c, and d belong to R and are uniquely determined by q itself. We identify R with the subalgebra generated by 1. In other words, if a ∈ R, we write simply a in place of a1. The product of two such elements is individuated by (2.1), assuming associativity and distributivity with respect to the real vector space sum. In this way, H turns out to be a non–commutative associative real division algebra. Given q = a + bi + cj + dk ∈ H, we recall that: • q := a − bi − cj − dk is the conjugate quaternion of q . • q := √qq = √a2 + b2 + c2 + d2 ∈ R is the norm of q . • Re(q) := 1 2 (q + q) = a ∈ R is the real part of q and Im(q) := 1 2 (q − q) = bi + cj + dk is the imaginary part of q . The element q ∈ H is said to be real if q = Re(q). It is easy to see that q is real if and only if qp = pq for every p ∈ H or, equivalently, if and only if q = q . If q = −q 8 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI or, equivalently, q = Im(q), then q is said to be imaginary. We denote by Im(H) the imaginary space of H and by S the sphere of unit imaginary quaternions: Im(H) := {q ∈ H q = Im(q)} = {bi + cj + dk ∈ H b, c, d ∈ R} and S := {q ∈ Im(H) q = 1} = {bi + cj + dk ∈ H b, c, d ∈ R, b2 + c2 + d2 = 1}. As said in the introduction, the set S is also equal to {q ∈ H q2 = −1}. We define R+ := {a ∈ R a ≥ 0} and we denote by N the set of all non–negative integers. Remark 2.1. (1) Notice that pq = q p, so the conjugation reverses the order of the factors. (2) · is, indeed, a norm on H, when it is viewed as the real vector space R4 . It also satisfies 1 = 1, q = q if q ∈ H and the remarkable identity: pq = pq if p, q ∈ H. (3) For every  ∈ S, denote by C the real subalgebra of H generated by ; that is, C := {α + β ∈ H α, β ∈ R}. Given any q ∈ H, we can write q = α + β for some α, β ∈ R and  ∈ S. Evidently, α and β are uniquely determined: α = Re(q) and β = Im(q). If q ∈ R, then β = 0 and hence  can be chosen arbitrarily in S. Thanks to the Inde- pendence Lemma (see [11, §8.1]), if q ∈ H \ R, then there are only two possibilities: β = ±Im(q) and  = β−1 Im(q). In other words, we have that H = S∈S C and C ∩ Cκ = R for every , κ ∈ S with  6= ±κ. (4) Two quaternions p and q are called conjugated (to each other), if there is s ∈ H \ {0} such that p = sqs−1 . The conjugacy class of q ; that is, the set of all quaternions conjugated with q , is equal to the 2–sphere Sq := Re(q) + Im(q)S of H. More explicitly, if q = a + bi + cj + dk , then we have that Sq = {a + xi + yj + z k ∈ H x, y , z ∈ R, x2 + y 2 + z 2 = b2 + c2 + d2 }. In particular, q and q are always conjugated, because q ∈ Sq . It is worth stressing that the following assertions are equivalent: • p and q are conjugated, • Sp = Sq , • Re(p) = Re(q) and Im(p) = Im(q), • Re(p) = Re(q) and p = q . (5) Equipped with the metric topology induced by the norm · , H results to be complete. Indeed, as a normed real vector space, it is isomorphic to R4 endowed with the standard Euclidean norm. In particular, given a sequence {qn}n∈N in H, if the series Pn∈N qn converges absolutely; that is, Pn∈N qn < +∞, then Pn∈N qn converges to some element q of H and such a series can be arbitrarily re–ordered without affecting its sum q . 2.2. Quaternionic Hilbert spaces. Let recall the definition of quaternionic Hilbert space (see e.g. [5],[26]). Let H be a right H–module; that is, an abelian group with a right scalar multiplication H × H ∋ (u, q) 7→ uq ∈ H QUATERNIONIC SLICE FUNCTIONAL CALCULUS 9 (u + v)q = uq + vq v(pq) = (vp)q and v(p + q) = vp + vq satisfying the distributive properties with respect to the two notions of sum: if u, v ∈ H and p, q ∈ H, and the associative property with respect to the quaternionic product: if v ∈ H and p, q ∈ H. Such a right H–module H is called quaternionic pre–Hilbert space if there exists a Hermitean quaternionic scalar product; that is, a map H × H ∋ (u, v) 7→ huvi ∈ H satisfying the following three properties: • (Right linearity) huvp + wqi = huvip + huwiq if p, q ∈ H and u, v , w ∈ H. • (Quaternionic Hermiticity) huvi = hv ui if u, v ∈ H. • (Positivity) If u ∈ H, then huui ∈ R+ and u = 0 if huui = 0. Suppose that H is equipped with such a Hermitean quaternionic scalar product. Then we can define the quaternionic norm k · k : H −→ R+ of H by setting kuk := phuui if u ∈ H. (2.2) The function k · k is a genuine norm over H, viewed as a real vector space, and the above defined scalar product h··i fulfills the standard Cauchy–Schwarz inequality. Proposition 2.2. The Hermitean quaternionic scalar product h··i satisfies the Cauchy–Schwarz inequality: huvi2 ≤ huui hv vi if u, v ∈ H. (2.3) Moreover, the map H ∋ u 7→ kuk ∈ R+ defined in (2.2) has the fol lowing properties: • kuqk = kuk q if u ∈ H and q ∈ H. • ku + vk ≤ kuk + kvk if u, v ∈ H. • If kuk = 0 for some u ∈ H, then u = 0. • If u, v ∈ H, then the fol lowing polarization identity holds: 4huvi =ku + vk2 − ku − vk2 + (cid:0)kui + vk2 − kui − vk2 (cid:1) i+ + (cid:0)kuj + vk2 − kuj − vk2 (cid:1) j + (cid:0)kuk + vk2 − kuk − vk2 (cid:1) k . We explicitly present the proof of the above statement just to give the flavour of the procedure taking the non–commutativity of the scalars into account. The result is evident if v = 0. Assume v 6= 0. Concerning (2.3), we start from the inequality 0 ≤ hup − vq up − vqi = phuuip + qhv viq − phuviq − qhv uip. Choosing p = hv vi and q = hv ui, we obtain: 0 ≤ hv vi (huuihv vi − huvihv ui) . Since hv vi > 0 and huvihv ui = hv uihv ui = hv ui2 = huvi2 , (2.3) follows immediately. Let us pass to the norm properties. With regards homogeneity, exploiting the fact that huui is real and thus it commutes with all quaternions, we obtain: kuqk2 = huq uqi = qhuuiq = qqhuui = q 2 kuk2. Triangular inequality follows from (2.3). Indeed, bearing in mind that Re(q) ≤ q for every q ∈ H, it holds: (2.4) ku + vk2 =hu + v u + vi = kuk2 + kvk2 + 2 Re(huvi) ≤ ≤kuk2 + kvk2 + 2huvi ≤ ≤kuk2 + kvk2 + 2kukkvk = (kuk + kvk)2 . 10 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (2.5) The polarization identity can be proved by direct inspection. From the positivity and the triangular inequality, it follows that: if u, v ∈ H d(u, v) := ku − vk defines a (translationally invariant) distance. One easily verifies that, in view of the given definitions, the operations of sum of vectors and right multiplication of vectors and quaternions are continuous with respect to the topology induced by d. Similarly, the scalar product is jointly continuous as a map from H × H to H. As soon as we are equipped with a metric topology, it makes sense to state the following definition. Definition 2.3. The quaternionic pre–Hilbert space H is said to be a quaternionic Hilbert space if it is complete with respect to its natural distance d. In what fol lows, given a quaternionic Hilbert space H, we wil l implicitly assume that H 6= {0}. A function f : H → H is right H–linear if f (u + v) = f (u) + f (v) and f (uq) = f (u)q for every u, v ∈ H and q ∈ H. Let H′ be the topological dual space of H consisting of all continuous right H–linear functions on H. Equip H′ with its natural structure of left H–module induced by the left multiplication by H: if f ∈ H′ and q ∈ H, then qf is the element of H′ sending u ∈ H into qf (u) ∈ H. The following result is an immediate consequence of quaternionic Hahn–Banach theorem (see [5, §2.10] and [8, §4.10]). Lemma 2.4. Let H be a quaternionic Hilbert space and let u ∈ H. If f (u) = 0 for every f ∈ H′ , then u = 0. It is possible to re–cast the definition of Hilbert basis even in the quaternionic– Hilbert–space context. Given a subset A of H, we define: A⊥ := {v ∈ H hv ui = 0 ∀u ∈ A}. Moreover, < A > denotes the right H–linear subspace of H consisting of all finite right H–linear combinations of elements of A. Let I be a non–empty set and let I ∋ i 7→ ai ∈ R+ be a function on I . As usual, we can define Pi∈I ai as the following element of R+ ∪ {+∞}: ai := sup (Xi∈J ai (cid:12)(cid:12)(cid:12)(cid:12) J is a non–empty finite subset of I) . Xi∈I It is clear that, if Pi∈I ai < +∞, then the set of all i ∈ I such that ai 6= 0 is at most countable. Given a quaternionic Hilbert space H and a map I ∋ i 7→ ui ∈ H, one can say that the series Pi∈I ui converges absolutely if Pi∈I kuik < +∞. If this happens, then only a finite or countable number of ui is nonzero and the series Pi∈I ui converges to a unique element of H, independently from the ordering of the ui ’s. The next three results can be established following the proofs of their corre- sponding complex versions (see e.g. [25, 30]). Proposition 2.5. Let H be a quaternionic Hilbert space and let N be a subset of H such that, for z , z ′ ∈ N , hz z ′i = 0 if z 6= z ′ and hz z i = 1. Then conditions (a)–(e) listed below are pairwise equivalent. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 11 (a) For every u, v ∈ H, the series Pz∈N huz ihz vi converges absolutely and it holds: huvi = Xz∈N huz ihz vi. (b) For every u ∈ H, it holds: kuk2 = Xz∈N hz ui2 . (c) N ⊥ = {0}. (d) < N > is dense in H. Proposition 2.6. Every quaternionic Hilbert space H admits a subset N , cal led Hilbert basis of H, such that, for z , z ′ ∈ N , hz z ′i = 0 if z 6= z ′ and hz z i = 1, and N satisfies equivalent conditions (a)–(d) stated in the preceding proposition. Two such sets have the same cardinality. Furthermore, if N is a Hilbert basis of H, then every u ∈ H can be uniquely decomposed as fol lows: u = Xz∈N z hz ui, where the series Pz∈N z hz ui converges absolutely in H. Theorem 2.7. If H is a quaternionic Hilbert space and ∅ 6= A ⊂ H, then it holds: ⊥ A⊥ =< A >⊥= < A > hAi = (A⊥ )⊥ and A⊥⊕ < A >= H , = < A >⊥ , where the bar denotes the topological closure and the symbol ⊕ denotes the ortho- gonal direct sum. Taking the previously introduced results into account, the representation Riesz’ theorem extends to the quaternionic case. This result can be proved as in the complex case (see e.g. [31]). Theorem 2.8 (Quaternionic representation Riesz’ theorem). If H is a quaternionic Hilbert space, the map H ∋ v 7→ hv · i ∈ H′ is wel l–posed and defines a conjugate–H–linear isomorphism. 2.3. Operators. First of all, we present the general definition of what we mean by a right H–linear operator. Definition 2.9. Let H be a quaternionic Hilbert space. A right H–linear operator is a map T : D(T ) −→ H such that: T (ua + vb) = (T u)a + (T v)b if u, v ∈ D(T ) and a, b ∈ H, where the domain D(T ) of T is a (not necessarily closed) right H–linear subspace of H. We define the range Ran (T ) of T by setting Ran (T ) := {T u ∈ H u ∈ D(T )}. In what fol lows, by the term “operator”, we mean a “right H–linear operator”. Similarly, by a “subspace”, we mean a “right H–linear subspace”. Let T : D(T ) −→ H and S : D(S ) −→ H be operators. As usual, we write T ⊂ S if D(T ) ⊂ D(S ) and S D(T ) = T . In this case, S is said to be an extension of T . We define the natural domains of the sum T + S and of the composition T S by setting D(T + S ) := D(T ) ∩ D(S ) and D(T S ) := {x ∈ D(S ) S x ∈ D(T )}. Here we use the symbols T S and S x in place of T ◦ S and S (x), respectively. 12 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI The operator T is said to be closed if the graph G (T ) := D(T ) ⊕ Ran (T ) of T is closed in H × H, equipped with the product topology. Finally, T is called closable if it admits closed operator extensions. In this case, the closure T of T is the smallest closed extension. We have the following elementary, though pivotal, result that permits the intro- duction of the notion of bounded operator. The proof is the same as for complex Hilbert spaces (see e.g. [25, 30]). Theorem 2.10. If H is a quaternionic Hilbert space, then an operator T : D(T ) −→ H is continuous if and only if is bounded; that is, there exists K ≥ 0 such that kT uk ≤ K kuk if u ∈ D(T ). Furthermore, a bounded operator is closed. As in the complex case, if T : D(T ) −→ H is any operator, one define kT k by setting (2.6) (2.7) sup u∈D(T )\{0} kT uk kT k := = inf {K ∈ R kT uk ≤ K kuk ∀u ∈ D(T )}. kuk Denote by B(H) the set of all bounded (right H–linear) operators of H: B(H) := {T : H −→ H operator kT k < +∞}. It is immediate to verify that, if T and S are operators in B(H), then the same is true for T + S and T S , and it holds: kT + S k ≤ kT k + kS k and kT S k ≤ kT k kS k. The reader observes that B(H) has a natural structure of real algebra, in which the sum is the usual pointwise sum, the product is the composition and the real scalar multiplication B(H) × R ∋ (T , r) 7→ T r ∈ B(H) is defined by setting (2.8) (T r)(u) := T (u)r. In Section 3.2, we will extend this real algebra structure to a quaternionic two–sided Banach unital C ∗ -algebra structure. The reader observes that definition (2.8) can be repeated for every operator T : D(T ) −→ H. It is worth introducing here a notion, which will be useful later. As usual, we denote by R[X ] the ring of real polynomials in the indeterminate X . For conve- nience, we write the polynomials in R[X ] with coefficients on the right. Given P (X ) = Pd h=0 X hrh in R[X ], we define the operator P (T ) ∈ B(H) as follows: dXh=0 T hrh , (2.9) where T 0 is considered to be equal to the identity operator I : H −→ H of H. The norm of B(H) allows us to define a metric D : B(H) × B(H) −→ R+ on B(H) as follows: P (T ) := (2.10) D(T , T ′) := kT − T ′k if T , T ′ ∈ B(H). Proposition 2.11. Let H be a quaternionic Hilbert space. Equip B(H) with the metric D. The fol lowing assertions hold: (a) B(H) is a complete metric space. In particular, it is a Baire space. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 13 (b) As maps B(H) × B(H) 7→ B(H), the sum and the composition of operators are continuous. The same is true for the real scalar multiplication. (c) The subset of B(H) consisting of elements admitting two–sided inverse in B(H) is open in B(H). (d) The uniform boundedness principle holds: Given any subset F of B(H), if supT ∈F T u < +∞ for every u ∈ H, then supT ∈F kT k < +∞. (e) The open map theorem holds: If T ∈ B(H) is surjective, then T is open. In particular, if T is bijective, then T −1 ∈ B(H). (f ) The closed graph theorem holds: If T : H −→ H is closed, then T ∈ B(H). Proof. The proofs of (a), (b) are elementary. Point (c) is proved as in [31, Theorem 10.12]. Points (d), (e) and (f ) follow from the Baire theorem exactly as in complex Banach space theory: Lemma 2.2.3, Theorem 2.2.4 and Corollary 2.2.5 of [28] prove (e); Theorem 2.2.9 of [28] leads to (d) and Theorem 2.2.7 proves (f ). (cid:3) The notion of adjoint operator is the same as for complex Hilbert spaces. Definition 2.12. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator with dense domain. The adjoint T ∗ : D(T ∗ ) −→ H of T is the unique operator with the following properties (the definition being well posed since D(T ) is dense): D(T ∗ ) := {u ∈ H ∃wu ∈ H with hwu vi = huT vi ∀v ∈ D(T )}. and (2.11) hT ∗uvi = huT vi ∀v ∈ D(T ), ∀u ∈ D(T ∗ ). It is worth noting that an immediate consequence of such a definition is that the operator T ∗ is always closed. Moreover, if T ∈ B(H), then requirement (2.11) alone automatically determines T ∗ as an element of B(H) in view of quaternionic representation Riesz’ theorem (see Theorem 2.8 above). As usual, an operator T : D(T ) −→ H with dense domain is said to be: • symmetric if T ⊂ T ∗ , • anti symmetric if T ⊂ −T ∗ , • self–adjoint if T = T ∗ , • essential ly self–adjoint if it is closable and T is self–adjoint, • anti self–adjoint if T = −T ∗ , • positive, and we write T ≥ 0, if huT ui ∈ R+ for every u ∈ D(T ), • normal if T ∈ B(H) and T T ∗ = T ∗T , • unitary if D(T ) = H and T T ∗ = T ∗T = I. Remark 2.13. (1) By definition, if U is a unitary operator of H, then hU uU vi = huvi for every u, v ∈ H and U is isometric; that is, kU uk = kuk for every u ∈ H. In particular, U belongs to B(H). As for complex Hilbert spaces, an operator U : H −→ H is unitary if and only if it is isometric and surjective. (2) If N , N ′ ⊂ H are Hilbert bases, then there exists a unitary operator U ∈ B(H) such that U (N ) = N ′ . Moreover, if V ∈ B(H) is unitary, then {V z ∈ H z ∈ N } is a Hilbert basis of H. Now we state quaternionic versions of three well–known results for complex Hilbert spaces. The former is an immediate consequence of Theorem 2.7 and of identity (2.11). 14 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (2.12) Proposition 2.14. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator with domain dense in H. We have: Ran (T )⊥ = Ker (T ∗ ) and Ker (T ) ⊂ Ran (T ∗ )⊥ . Furthermore, if D(T ∗ ) is dense in H and T is closed, then we can replace “⊂” with “=”. In particular, this is true when T ∈ B(H). Theorem 2.15. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator with dense domain. The fol lowing facts hold. (a) T ∗ is closed and, if U ∈ B(H ⊕ H) is the unitary operator sending (x, y ) into (−y , x), then the fol lowing orthogonal decomposition holds: H ⊕ H = G (T ) ⊕ U (G (T ∗ )), where the bar denotes the closure in H ⊕ H. (b) T is closable if and only if D(T ∗) is dense in H and T = (T ∗)∗ . (c) If T is closed, injective and has image dense in H, then the same is true for T ∗ and for T −1 : Ran (T ) −→ D(T ). Moreover, under these hypotheses, it holds: (T ∗ )−1 = (T −1 )∗ . (d) If T is closed, then D(T ∗T ) is dense in H and T ∗T is self–adjoint. (e) (Helling–Toeplitz Theorem) If T is self–adjoint or anti self–adjoint and D(T ) = H, then T ∈ B(H). Proof. The proof are the same as for complex Hilbert spaces, since they do not depend on the fact that the Hilbert space is defined over C or H. Points (a) and (b) can be proved exactly as Theorem 5.1.5 in [28]. One can prove (c) exactly as Proposition 5.1.7 in [28] and (d) as point (i) in Theorem 5.1.9 in [28]. Point (e) follows from (a) taking into account the closed graph theorem. (cid:3) Remark 2.16. Exactly as in the case of complex Hilbert space (see e.g. [25, 31]), one can prove that, if S : D(S ) −→ H and T : D(T ) −→ H are operators with D(T ) and D(S ) dense in H, then it hold: (i) S ∗ ⊂ T ∗ if T ⊂ S . (ii) If T ∈ B(H), then T ∗ ∈ B(H), kT ∗k = kT k and kT ∗T k = kT k2. (iii) T ⊂ (T ∗ )∗ if D(T ∗) is dense in H, and T = (T ∗)∗ if T ∈ B(H). (iv) T ∗ + S ∗ ⊂ (T + S )∗ if D(T + S ) is dense in H, and (T + S )∗ = T ∗ + S ∗ if T ∈ B(H). (v) S ∗T ∗ ⊂ (T S )∗ if D(T S ) is dense in H, and (T S )∗ = S ∗T ∗ if T ∈ B(H). (vi) If T is self–adjoint, S is symmetric and T ⊂ S , then S = T . (vii) T is essentially self–adjoint if and only if D(T ∗ ) is dense in H and T ∗ is self–adjoint. In that case, T ∗ = T and T is the only self–adjoint extension of T . (viii) Let T ∈ B(H). If T is bijective and T −1 ∈ B(H), then T ∗ (T −1)∗ = I = (T −1 )∗T ∗ . In this way, T is bijective and T −1 ∈ B(H) if and only if T ∗ is bijective and (T ∗)−1 ∈ B(H). Moreover, in that situation, it holds: (T ∗ )−1 = (T −1)∗ . (ix) (T r)∗ = T ∗r if r ∈ R. The following last proposition deserves an explicit proof as it is different from that in complex Hilbert spaces. Proposition 2.17. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator with dense domain. The fol lowing assertions hold. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 15 (a) If T is closed (for example, if T ∈ B(H)) and there exists a dense subspace L of D(T ) such that huT ui = 0 for every u ∈ L, then T = 0. (b) If huT ui ∈ R for every u ∈ D(T ) (for example, if T ≥ 0), then T is symmetric and hence it is self–adjoint in the case in which D(T ) = H. Proof. (a) Since D(T ) is dense in H, L is dense in H as well. Fix u, v ∈ L. For every q ∈ H, u + vq belongs to L and hence 0 = hu + vq T (u + vq)i = huT viq + ¯qhv T ui. In particular, taking q = 1, we get: huT vi + hv T ui = 0, from which we deduce that huT viq − ¯qhuT vi = 0 for every q ∈ H. Defining a := huT vi = a0 + a1 for some a0 , a1 ∈ R and  ∈ S, we get: 0 = a + a = (a0+a1)+(a0+a1 ) = −2a1+2a0. Therefore we have that a0 = a1 = 0 and hence huT vi = 0. Choosing a sequence L ⊃ {un}n∈N → T v , we infer that kT vk2 = 0 (and hence T v = 0) for every v ∈ L. Finally, if x ∈ D(T ) and L ⊃ {xn }n∈N → x, then we have that T xn = 0 for every n ∈ N, so {(xn , T xn )}n∈N → (x, 0). Since T is closed, it follows that T x = 0. (b) Let us follow the strategy used in the proof of point (a). First, observe that hxT xi = hxT xi = hT xxi for every x ∈ D(T ). Fix u, v ∈ D(T ). For every q ∈ H, we have: 0 =hu + vq T (u + vq)i − hT (u + vq)u + vqi = =(huT vi − hT uvi)q + ¯q(hv T ui − hT v ui). Define b := huT vi − hT uvi. Choosing q = 1, we obtain that bq − ¯qb = 0 for every q ∈ H. Proceeding as above, we infer that b = 0 or, equivalently, that huT vi = hT uvi for every u, v ∈ D(T ). This proves the symmetry of T . (cid:3) 2.4. Square root and polar decomposition of operators. The theorem of existence of the square root of positive bounded operators works exactly as in the case of complex Hilbert spaces. The reason is that the proof exploits the convergence of sequences of real polynomials of operators in the strong operator topology (cf. [12] and [28, Propositions 3.2.11 and 3.2.12, Theorem 3.2.17]). Theorem 2.18. Let H be a quaternionic Hilbert space and let T ∈ B(H). If T ≥ 0, then there exists a unique operator in B(H), indicated by √T , such that √T ≥ 0 and √T √T = T . Furthermore, it turns out that √T commutes with every operator which commutes with T . Even for quaternionic Hilbert spaces, one has the polar decomposition theorem. Before presenting our quaternionic version of such a theorem, we need some prepa- rations. Proposition 2.19. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a normal operator. Then we have that kT uk = kT ∗uk for every u ∈ H and it holds: Ker (T ) = Ker (T ∗ ) and Ran (T ) = Ran (T ∗ ). In particular, H orthogonal ly decomposes as: H = Ker (T ) ⊕ Ran (T ). (2.13) Proof. Given any u ∈ H, we have: kT uk2 = hT uT ui = huT ∗T ui = huT T ∗ui = hT ∗uT ∗ui = kT ∗uk2 . 16 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI This implies immediately that Ker (T ) = Ker (T ∗ ). By combining the latter equality with Proposition 2.14, we obtain immediately the equality Ran (T ) = Ran (T ∗) and the orthogonal decomposition (2.13). (cid:3) Let T ∈ B(H). The reader observes that T ∗T ≥ 0. Indeed, it holds: huT ∗T ui = hT uT ui = kT uk2 ≥ 0 if u ∈ H. As usual, we define T ∈ B(H) by setting T := √T ∗T . By point (b) of Proposition 2.17, the operator T is self–adjoint. Moreover, it has the same kernel of T : kT (u)k2 = hT (u)T (u)i = huT 2(u)i = huT ∗T ui = kT uk2 (2.14) for every u ∈ H. In this way, if T is normal, one also has that (2.15) Ker (T ) = Ker (T ∗) = Ker (T ) and Ran (T ) = Ran (T ∗) = Ran (T ). We are in position to state and prove our quaternionic version of the polar decomposition theorem. To the best of our knowledge, such a theorem has been mentioned and used several times in the mathematical physics literature, without an explicit proof. Theorem 2.20. Let H be a quaternionic Hilbert space and let T ∈ B(H) be an operator. Then there exist, and are unique, two operators W and P in B(H) such that: (i) T = W P , (ii) P ≥ 0, (iii) Ker (P ) ⊂ Ker (W ), (iv) W is isometric on Ker (P )⊥ ; that is, kW uk = kuk for every u ∈ Ker (P )⊥ . Furthermore, W and P have the fol lowing additional properties: (a) P = T . (b) If T is normal, then W defines a unitary operator in B(Ran (T )). (c) If T is normal, then W commutes with T and with al l the operators in B(H) commuting with both T and T ∗ . (d) W is self–adjoint if T is. (e) W is anti self–adjoint if T is. Proof. Firstly, we show that, if there exists a decomposition (i) of T with properties (ii), (iii) and (iv), then it is unique. Thanks to (ii) and Proposition 2.17,(b), P is self–adjoint and hence Ker (P )⊥ = Ran (P ) and T ∗T = P W ∗W P . By using (iv) and the polarization identity (see (2.4)), we know that hW P uW P vi = hP uP vi or, equivalently, 0 = hu(P W ∗W P − P 2)vi = hu(T ∗T − P 2)vi for every u, v ∈ H. This is equivalent to say that T ∗T = P 2 . Theorem 2.18 ensures that P coincides with T and hence (a) holds. In particular, P is unique. Let us show the uniqueness of W . Since H = Ker (P ) ⊕ Ker (P )⊥ = Ker (P ) ⊕ Ran (P ) and W vanishes on Ker (P ), to determine it, it is enough fixing it on Ker (P )⊥ = Ran(P ). Actually, this is done by the requirement T = W P itself, that fixes W on Ran(P ) and thus on the whole Ran(P ), because W is continuous. Let us prove the existence of a polar decomposition. Let P := T . Define W as follows. If v ∈ Ker (P ), then W v := 0. If v ∈ Ran (P ), then W v := T u, where u QUATERNIONIC SLICE FUNCTIONAL CALCULUS 17 is any element of H such that v = P u. Since Ker (P ) = Ker (T ) (see (2.14)), the latter definition is well posed. Using (2.14) again, we infer that W is isometric on Ran (P ): kW vk = kT uk = kT uk = kvk for every v ∈ Ran (P ). By continuity, W can be uniquely defined on the whole Ran (P ) = Ker (P )⊥ . Evidently, W and P satisfy properties (i)–(iv). In the remainder of the proof, we suppose that T is normal. Let us prove (b). The normality of T ensures that Ran (T ) = Ran (P ) (see (2.15)). Since T = W P , it is evident that W (Ran (P )) = Ran (T ). By continuity, we have that W (Ran (T )) = W (Ran (P )) ⊂ Ran (T ). In this way, we have that Ran (T ) ⊂ W (Ran (T )) ⊂ Ran (T ). On the other hand, by (iv), W is isometric on Ker (P )⊥ = Ran (T ) and hence W (Ran (T )) is closed in H. It follows that W (Ran (T )) = Ran (T ). The restriction of W from Ran (T ) into itself turns out to be a surjective isometric operator or, equivalently, a unitary operator. By (iv) and (a), we have that W ∗W = I on Ker (T )⊥ = Ran (T ). In this way, using the equalities T = W T and T ∗T = T T ∗, we infer that T 2 = W T 2W ∗ . Applying W ∗ on the left, it arises: W ∗ T 2 = T 2W ∗ and, taking the adjoint, T 2W = W T 2 . Since T = pT 2 commutes with all of the operators commuting with T 2 , we get that (T W − W T )T = 0 and hence T W = W T on Ran (T ). Since W vanishes on Ker (T ), it follows that T W = W T on the whole H. Suppose now that A ∈ B(H) commute with T and T ∗ . Since A commutes with T , it follows that A preserves the decomposition H = Ker (T ) ⊕ Ran (T ). As we know, the operator T commutes with all operators commuting with T and T ∗ . In particular, it commutes with A. In this way, the equality AW T = W T A (recall that T = W T ) is equivalent to the following one: (AW − W A)T = 0. It follows that A commutes with W on Ran (T ). Since W vanishes on Ker (T ), A trivially commutes with W on Ker (T ). This proves that A commutes with W and hence (c). Let us prove (d). Since T is self–adjoint and commutes with W , T commutes also with W ∗ . In this way, assuming T self–adjoint, we have that 0 = T − T ∗ = (W − W ∗ )T and hence W = W ∗ on Ker (T )⊥ = Ran (T ). Since W preserves the decomposition H = K er(T ) ⊕ Ran(T ) and vanishes on K er(T ), one easily gets that W ∗ vanishes on K er(T ) as well. We infer that W ∗ = W , as desired. It remains to show (d). Exactly as we have done to prove (c), if T is anti self– adjoint, one proves that 0 = T + T ∗ = (W + W ∗ )T . Thus W = −W ∗ on Ker (T )⊥ . Since W = 0 = −W ∗ on Ker (T ), we have that W ∗ = −W . (cid:3) Remark 2.21. What is important to stress here is that, for a normal operator T , the three operators appearing in the decomposition T = W T separately respect decomposition (2.13). In other words, T , W and P separately map Ker (T ) and Ran (T ) into themselves. This circumstance will turn out useful in several proofs along all this work. Remark 2.22. The reader observes that, it being H = Ker (P )⊕ Ker (P )⊥ , preceding condition (iii) can be replaced by the following one: (iii′ ) Ker (P ) = Ker (W ). 3. Left multiplications, imaginary units and complex subspaces In this part, we introduce a notion that is proper of quaternionic modules, since it arises from the non–commutativity of the algebra of quaternions. It is the notion of left scalar multiplication, which makes both H and B(H) quaternionic two–side 18 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI Banach modules. As a byproduct, the notion of (operatorial) imaginary unit is also presented. This notion plays a crucial role in spectral theory over quaternionic Hilbert spaces, giving rise to a nice relationship with the theory in complex Hilbert spaces. This fact will be evident shortly, especially in view of Proposition 3.11. if u ∈ H and q ∈ H. 3.1. Left scalar multiplications. Let us show that it is possible to equip H with a left multiplication with quaternions. It will be a highly non–canonical operation relying upon a choice of a preferred Hilbert basis. So, pick out a Hilbert basis N of H and define the left scalar multiplication of H induced by N as the map H × H ∋ (q , u) 7→ qu ∈ H given by qu := Xz∈N z qhz ui (3.1) Proposition 3.1. The left product defined in (3.1) satisfies the fol lowing properties. (a) q(u + v) = qu + qv and q(up) = (qu)p for every u ∈ H and q , p ∈ H. (b) kquk = q kuk for every u ∈ H and q ∈ H. (c) q(q ′ u) = (qq ′ )u for every u ∈ H and q , q ′ ∈ H. (d) hquvi = huqvi for every u, v ∈ H and q ∈ H. (e) ru = ur for every u ∈ H and r ∈ R. (f ) qz = z q for every z ∈ N and q ∈ H. Consequently, for every q ∈ H, the map Lq : H −→ H, sending u into qu, is an element of B(H). Moreover, the map LN : H −→ B(H), defined by setting LN (q) := Lq , is a norm–preserving real algebra homomorphism, with the additional properties: if r ∈ R and u ∈ H Lr u = ur (3.2) and (Lq )∗ = Lq if q ∈ H. (3.3) Final ly, if N ′ is another Hilbert basis of H and U ∈ B(H) is any unitary operator such that U (N ) = N ′ , then it holds: (f ) LN ′ (q) = U LN (q)U −1 for every q ∈ H. (g) LN = LN ′ if and only if hz z ′ i ∈ R for every z ∈ N and z ′ ∈ N ′ . Proof. All statements straightforwardly follows from Proposition 2.5. However, the proof of (g) deserves some comments. Bearing in mind that z = Pz ′∈N ′ z ′ hz ′ z i, the equality LN = LN ′ holds if and only if Xz ′ ∈N ′ z ′hz ′ z iq = z q = LN (q)(z ) = LN ′ (q)(z ) = Xz ′ ∈N ′ z ′qhz ′ z i for every z ∈ N and q ∈ H. This is equivalent to say that, fixed any z ∈ N and z ′ ∈ N ′ , hz ′ z iq = qhz ′ z i for every q ∈ H, which in turn is equivalent to require that hz ′ z i ∈ R. (cid:3) Remark 3.2. (1) All possible different notions of left scalar multiplication have to coincide when multiplying with real quaternions in view of point (e) of Proposition 3.1, because ur does not depend on the choice of N . (2) Thanks to Proposition 3.1, given a left scalar multiplication H × H ∋ (q , u) 7→ Lq (u) = qu ∈ H, we know that the map H ∋ q 7→ Lq ∈ B(H) is a norm–preserving real algebra homomorphism satisfying (3.2) and (3.3). Conversely, it can be shown QUATERNIONIC SLICE FUNCTIONAL CALCULUS 19 that every norm–preserving real algebra homomorphism from H to B(H) satisfying (3.2) and (3.3) is a left scalar multiplication induced by some Hilbert basis of H (see [18] for the proof ). In what fol lows, by the symbol H ∋ q 7→ Lq , we mean the left scalar multiplication Lq u := Pz∈N z qhz ui of H induced by some fixed Hilbert basis N of H itself. 3.2. Quaternionic two-sided Banach C ∗ -algebra structure on B(H). In order to avoid any possible misunderstanding, we specify the meaning we give to the notion of quaternionic two–sided Banach C ∗ –algebra with unity. Such a notion coincides with the one used in [8] (see also [5]). We call quaternionic two–sided vector space a non–empty subset V equipped with a sum V × V ∋ (u, v) 7→ u + v ∈ V , with a left scalar multiplication H × V ∋ (q , u) 7→ qu ∈ V and with a right scalar multiplication V × H ∋ (u, q) 7→ uq such that V is an abelian group with respect to the sum and it holds: (v1) q(u + v) = qu + qv and (u + v)q = uq + vq , (v2) (q + p)u = qu + pu and u(q + p) = uq + up, (v3) q(pu) = (qp)u and (up)q = u(pq), (v4) (qu)p = q(up), (v5) u = 1u = u1, (v6) ru = ur for every u, v ∈ V , q , p ∈ H and r ∈ R. Such a quaternionic two–sided vector space V is said to be a quaternionic two–sided algebra if, in addition, it is equipped with a product V × V ∋ (u, v) 7→ uv ∈ V such that (a1) u(vw) = (uv)w, (a2) u(v + w) = uv + uw and (u + v)w = uw + vw, (a3) q(uv) = (qu)v and (uv)q = u(vq) for every u, v , w ∈ V and q ∈ H. Moreover, we say that V is with unity, or unital, if there exists an element  of V , the unity of V , such that (a4) u = u = u for every u ∈ V . Suppose that V is a quaternionic two–sided algebra with unity. If uv = vu for every u, v ∈ V , then V is called commutative. A map ∗ : V −→ V is called ∗–involution if it holds: (∗1) (u∗ )∗ = u, (∗2) (u + v)∗ = u∗ + v∗ , (qu)∗ = u∗ q and (uq)∗ = qu∗ , (∗3) (uv)∗ = v∗u∗ for every u, v ∈ V and q ∈ H. Equipping V with a ∗–involution, we obtain a quaternionic two–sided ∗ –algebra with unity. It is easy to verify that ∗ =  and, if u is invertible in V , then (u−1 )∗ = (u∗ )−1 . Finally, the quaternionic two–sided ∗–algebra V with unity is called a quaternionic two–sided Banach C ∗ –algebra with unity if, in addition, V is equipped with a function k · k : V −→ R+ such that (n1) kuk = 0 if and only if u = 0, (n2) ku + vk ≤ kuk + kvk, (n3) kquk = q kuk = kuqk, (n4) kuvk ≤ kuk kvk, (n5) kk = 1, (n∗) ku∗uk = kuk2 20 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI for every u, v ∈ V and q ∈ H, and the metric space obtained equipping V with the distance V × V ∋ (u, v) 7→ ku − vk ∈ R+ is complete. We say that k · k is a quaternionic Banach C ∗ –norm on V . Suppose that V is a quaternionic two–sided Banach C ∗ –algebras with unity . As usual, a subset V ′ of V is a quaternionic two–sided Banach unital C ∗–subalgebra of V if  ∈ V ′ and the restrictions to V ′ of the operations of V define on V ′ a structure of quaternionic two–sided Banach C ∗ –algebras with unity . A map φ : V −→ W between quaternionic two–sided Banach C ∗–algebras with unity is called ∗–homomorphism if it holds: (h1) φ(u + v) = φ(u) + ψ(v), φ(qu) = qφ(u) and φ(uq) = φ(u)q , (h2) φ(u∗ ) = φ(u)∗ , (h3) φ() =  for every u, v ∈ V and q ∈ H. Remark 3.3. Considering R (resp. C for some fixed  ∈ S) instead of H as set of scalars, one defines a real (resp. C ) two–sided Banach unital C ∗ –algebra. Thanks to (v6), the notion of real two–sided Banach unital C ∗–algebra is equivalent to the usual one of real Banach unital C ∗ –algebra. Similarly, complex Banach unital C ∗–algebras correspond to two–sided C –Banach unital C ∗–algebras V having the following property in place of (v6): (v6′ ) cu = uc for every u ∈ V and c ∈ C . The notion of real Banach unital C ∗–subalgebra of a real Banach unital C ∗ – algebra can be defined in the usual way. Since a quaternionic two–sided Banach unital C ∗ –algebra V is also a real (two–sided) Banach unital C ∗–algebra, we can speak about real Banach unital C ∗–subalgebras of V . Evidently, if (v6′ ) holds, then one can speak about C–Banach unital C ∗ –subalgebras of V as well. Consider a quaternionic Hilbert space H and fix a left scalar multiplication H ∋ q 7→ Lq of H. Given a (right H–linear) operator T : D(T ) −→ H and q ∈ H, we define the map qT : D(T ) −→ H by setting (3.4) (qT )u := q(T u). It is immediate to verify that qT is again a (right H–linear) operator. Similarly, if Lq (D(T )) ⊂ D(T ), one can define the (right H–linear) operator T q : D(T ) −→ H as follows: (3.5) (T q)(u) := T (qu). Observe that, if q ∈ R, then the inclusion Lq (D(T )) ⊂ D(T ) is always verified. The given notions of left and right scalar multiplications qT and T q between quaternions and operators are nothing but usual compositions Lq T and T Lq , re- spectively. In general, qT is different from T q . However, if r ∈ R, the operators rT and T r defined here are equal and coincide with the operator T r defined in (2.8). By a direct inspection, one can easily prove that, if D(T ) is dense in H, then (qT )∗ = T ∗q and (T q)∗ = q T ∗ if Lq (D(T )) ⊂ D(T ). Suppose now that T ∈ B(H). By Proposition 3.1(b), it follows immediately that kqT k = q kT k. (3.6) Moreover, it holds also: (3.7) kT qk = q kT k. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 21 Indeed, if q = 0, (3.7) is evident. If q 6= 0, then we have: kT (qu)k kT (v)k kT (v)k 1 kT qk = sup = q kT k. sup = = sup kq−1vk q−1 kvk kuk u6=0 v 6=0 v 6=0 Equalities (3.6) and (3.7) imply that, if T ∈ B(H), then qT and T q belong to B(H). It is now easy to establish the next result. Theorem 3.4. Let H be a quaternionic Hilbert space equipped with a left scalar multiplication. Then the set B(H), equipped with the pointwise sum, with the left and right scalar multiplications defined in (3.4) and (3.5), with the composition as product, with the adjunction T 7→ T ∗ as ∗–involution and with the norm defined in (2.6), is a quaternionic two–sided Banach C ∗–algebra with unity I. Remark 3.5. Independently from the choice of a left scalar multiplication of H, B(H) is always a real Banach C ∗–algebra with unity I. It suffices to consider the right scalar multiplication (2.8), the adjunction T 7→ T ∗ as ∗–involution and the norm defined in (2.6). 3.3. Imaginary units and complex subspaces. Consider a quaternionic Hilbert space H equipped with a left scalar multiplication H ∋ q 7→ Lq . For short, we write Lq u = qu. For every imaginary unit ı ∈ S, the operator J := Lı is anti self–adjoint and unitary; that is, it holds: J ∗ = −J and J ∗J = I. The proof straightforwardly arises from Proposition 3.1. We intend to establish the converse statement: if an operator J ∈ B(H) is anti–self adjoint and unitary, then ı for some left scalar multiplication H ∋ q 7→ L′ J = L′ q of H. To prove the statement, we need a preliminary definition known from the litera- ture [12], which will turn out to be useful several times in this work later. Definition 3.6. Let J ∈ B(H) be an anti self–adjoint and unitary operator and let ı ∈ S. Recall that Cı denotes the real subalgebra of H generated by ı; that is, + and HJ ı Cı := {α + ıβ ∈ H α, β ∈ R}. Define the complex subspaces HJ ı − of H associated with J and ı by setting HJ ı ± := {u ∈ H J u = ±uı}. Remark 3.7. HJ ı ± are closed subsets of H, because u 7→ J u and u 7→ ±uı are continuous. However, they are not (right H–linear) subspaces of H. The names of + and HJ ı HJ ı − are justified by point (e) of Proposition 3.8 below. The aim of this section is to prove the following result. Proposition 3.8. Let H be a quaternionic Hilbert space, let J ∈ B(H) be an anti self–adjoint and unitary operator, and let ı ∈ S. Then the fol lowing facts hold. (a) There exists a left scalar multiplication H ∋ q 7→ Lq of H such that J = Lı . (b) If H ∋ q 7→ Lq and H ∋ q 7→ L′ q are left scalar multiplications of H such that Lı = L′ ı , then Lq = L′ q for every q ∈ Cı . (c) Let H ∋ q 7→ Lq and H ∋ q 7→ L′ q be left scalar multiplications of H induced by Hilbert bases N and N ′ , respectively. Then Lı = L′ ı if and only if hz z ′i ∈ Cı for every z ∈ N and z ′ ∈ N ′ . + 6= {0} and HJ ı (d) HJ ı − 6= {0}. 22 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (e) Identify Cı with C in the natural way. Then HJ ı + is a complex Hilbert space with respect to the structure induced by H: its sum is the sum of H, its complex scalar multiplication is the right scalar multiplication of H restricted to Cı and its Cı–valued Hermitean scalar product coincides with the restriction of the one of H. For short, we say that HJ ı + is a Cı–Hilbert space. An analogous statement holds for HJ ı − . (f ) If N is a Hilbert basis of the Cı–Hilbert space HJ ı + , then N is also a Hilbert basis of H and it holds: J = Xz∈N An analogous statement holds for HJ ı − . z ıhz · i. In order to prove the proposition, we need two lemmata. The first deals with points (d) and (e). Lemma 3.9. Let H be a quaternionic Hilbert space, let J ∈ B(H) be an anti self– adjoint and unitary operator and let ı ∈ S. Then HJ ı + 6= {0}, the restriction of the Hermitean scalar product h· ·i to HJ ı + is Cı–valued and HJ ı + turns out to be a Cı–Hilbert space with respect to the structure induced by H. An analogous statement holds for HJ ı − . Proof. We prove the thesis for HJ ı + , the other case being essentially identical. Firstly, we show that there exists u 6= 0 such that u − J uı 6= 0. Choose a non–null element If u − J uı 6= 0, we are done. Suppose u = J uı. Let  ∈ S such that u of H. ı = −ı. We have that u = J uı = −J uı. Since u 6= 0, we infer that u 6= J uı. Replacing u with u if necessary, we may suppose that u − J uı 6= 0, as desired. Then J (u − J uı) = J u + uı = (u − J uı)ı; that is, J u′ = u′ ı for u′ := u − J uı 6= 0. Therefore HJ ı + 6= {0}. Next we notice that, if v , w ∈ HJ ı + and r, r′ ∈ R, one has: hv wi(r + r′ ı) = hv w(r + r′ ı)i = hv wr + wr′ ıi = = hv wir + hv J wir′ = rhv wi + r′ h−J v wi = rhv wi − r′ hv ıwi = hv(r − r′ ı)wi = (r + r′ ı)hv wi . Therefore, huvi commutes with every element of Cı . This implies that huvi ∈ Cı and that h··i reduces to a standard Hermitean scalar product on HJ ı + , viewed as a complex pre–Hilbert space. Finally, HJ ı + is complete, since H is such and HJ ı + is closed. (cid:3) Lemma 3.10. As a Cı–Hilbert space, H admits the fol lowing direct sum decompo- sition: + ⊕ HJ ı H = HJ ı − . Moreover, if N is a Hilbert basis of the Cı –Hilbert space HJ ı + , then it holds: (a) If  ∈ S is an imaginary unit with ı = −ı (i.e. orthogonal to ı), N  := {z  z ∈ N } is a Hilbert basis of the Cı–Hilbert space HJ ı − . (b) N is a Hilbert basis of H. Proof. The first statement holds because HJ ı + ∩ HJ ı − = {0} and, defining x± := 1 2 (x ∓ J xı) for x ∈ H, one has x = x+ + x− and x± ∈ HJ ı ± as it follows by a direct inspection. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 23 + ∋ y 7→ y  ∈ HJ ı Let us prove (a). As it is easy to verify, the map HJ ı − is a well–defined isometric Cı –anti-linear isomorphism and hence the set N  is or- − . Since u ∈ HJ ı thonormal, because N is such. Choose u ∈ HJ ı + , we have that u = Pz∈N z hz ui. From this equality, it follows that u = Xz∈N z hz ui = Xz∈N z hz ui. This proves that N  is a Hilbert basis of HJ ı − . It remains to show (b). By point (a), one has: x = Xz∈N z hz x+i + Xz∈N z hz x−i = Xz∈N z hz x+ + x− i . So < N > is dense in H and hz z ′ i = δzz ′ for every z , z ′ ∈ N . Thus N is a Hilbert basis of H. (cid:3) Proof of Proposition 3.8. Points (d) and (e) were proved in Lemma 3.9. + . By the definition of H J ı Let us prove (f ). Let N be a Hilbert basis of H J ı + , it holds that J z = z ı for each z ∈ N . Thanks to point (b) of Lemma 3.10, N is also a Hilbert basis of H. In this way, for every u ∈ H, it holds: J u = J Xz∈N z hz ui = Xz∈N J z hz ui = Xz∈N z ıhz ui. Point (a) follows immediately. In fact, if H ∋ q 7→ Lq is the left scalar multiplication induced by N , then J = Lı . Let now H ∋ q 7→ Lq and H ∋ q 7→ L′ q be scalar left multiplications of H such ı . Since L1 = I = L′ that Lı = L′ 1 , for each α, β ∈ R, it follows: Lα+β ı = L1α + Lıβ = L′ 1α + L′ ıβ = L′ α+β ı . This shows (b). Point (c) can be proved similarly to (g) in Proposition 3.1. (cid:3) 3.4. Extension of J to a full left scalar multiplication of H. Finally, we present a technical, but quite useful, proposition we shall employ later several times. This proposition generalizes, and formulates into a modern language, some results presented in Section 3 of [12]. Proposition 3.11. For every Cı–linear operator T : D(T ) −→ HJ ı + with D(T ) ⊂ + , there exists a unique right H–linear operator T : D( T ) −→ H of T with HJ ı D( T ) ⊂ H such that J (D( T )) ⊂ D( T ), D( T ) ∩ HJ ı + = D(T ) and T (u) = T (u) for every u ∈ D(T ). The fol lowing additional facts hold. + ), then T ∈ B(H) and k T k = kT k. (a) If T ∈ B(HJ ı (b) J T = T J . (c) Let V : D(V ) −→ H be a right H–linear operator with D(V ) ⊂ H. Then V is equal to U for some Cı –linear operator U : D(V ) ∩ HJ ı + −→ HJ ı + if and only if J (D(V )) ⊂ D(V ) and J V = V J . + , then D( T ) is dense in H and ( T )∗ = fT ∗ . (d) If D(T ) is dense in HJ ı Furthermore, given a Cı–linear operator S : D(S ) −→ HJ ı + with D(S ) ⊂ HJ ı + , we have: (e) fS T = S T . (f ) If S is a right (resp. left) inverse of T , then S is a right (resp. left) inverse of T . 24 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI An analogous statement holds for HJ ı − . Proof. Choose  ∈ S in such a way that {1, ı, , ı} is a basis of H and ı = −ı. We begin showing that, if T : D( T ) −→ H is a right H–linear extension of T with J (D( T )) ⊂ D( T ) and D( T ) ∩ HJ ı + = D(T ), then T is uniquely determined by T . Denote by Φ : H −→ H the Cı–anti–linear isomorphism of H sending x in x. It ∓ . Since D( T ) is a right H–linear subspace is immediate to verify that Φ(HJ ı ± ) = HJ ı of H, we have that Φ(D( T )) = D( T ). Applying Φ to both members of the equality D( T ) ∩ HJ ı + = D(T ), we obtain D( T ) ∩ HJ ı − = Φ(D(T )). Let x ∈ D( T ). Define a := 1 2 (x − J xı) and b := 1 2 (x + J xı). It is easy to see that a ∈ HJ ı + , b ∈ HJ ı − and x = a + b. Since J (D( T )) ⊂ D( T ), we infer that a ∈ D( T ) ∩ HJ ı + = D(T ) and − = Φ(D( T )). This proves that, as a Cı –linear subspace of H, D( T ) b ∈ D( T ) ∩ HJ ı coincides with D(T ) ⊕ Φ(D(T )). Bearing in mind that b = Φ(b) ∈ HJ ı + , we have that T (b) = − T (b) = −T (b) and hence T (x) = T (a) − T (b). It follows that T is uniquely determined by T , as desired. Concerning the existence of the extension of T with the required properties, we are now forced to define D( T ) := D(T ) ⊕ Φ(D(T )) and T : D( T ) −→ H by setting T (x) := T (a) − T (b), where x ∈ D( T ), a ∈ D(T ), b ∈ Φ(D(T )) and x = a + b. We must verify that J (D( T )) ⊂ D( T ), D( T ) is a right H–linear subspace of H and T is right H–linear. Let x, a and b be as above. Observe that J (D( T )) ⊂ D( T ). Indeed, J x = J a + J b = aı − bı belongs to D( T ), because D(T ) and Φ(D(T )) are Cı–linear subspaces of H. Let q ∈ H. There exist, and are unique, z1 , z2 ∈ Cı such that q = z1 + z2. It holds: xq = (a + b)(z1 + z2 ) = (az1 + bz2 ) + (bz1 + az2 ). Since z1  = z 1 and z2  = z 2 , we get that az1 + bz2 ∈ D(T ) and bz1 + az2 ∈ Φ(D(T )) and then D( T ) is a right H–linear subspace of H. Moreover, we have: T (xq) = T (az1 + bz2) − T ((bz1 + az2)) = = T (a)z1 + T (bz2) − T (bz1 − az2 ) = = T (a)z1 + T (bz2) − T (bz1) + T (a)z2. On the other hand, we obtain the right H–linearity of T as follows: T (x)q = (T (a) − T (b))(z1 + z2 ) = = T (a)z1 − T (b)z1 + T (a)z2  − T (b)z2 = = T (a)z1 − T (b)z1  + T (a)z2 − T (b)z2  = = T (a)z1 − T (bz1 ) + T (a)z2 + T (bz2 ) = = T (a)z1 − T (bz1) + T (a)z2  + T (bz2) = T (xq). Let us prove the remaining points. + ). Then D( T ) = HJ ı (a) Suppose that T ∈ B(HJ ı + ⊕ HJ ı + ⊕ Φ(HJ ı + ) = HJ ı − = H. Let x, a and b be as above and let c ∈ HJ ı + such that b = c. Since z = z  for every z ∈ Cı , we obtain: habi + hbai = haci + hcai = haci − hcai = = haci − hcai = haci − haci = 0 QUATERNIONIC SLICE FUNCTIONAL CALCULUS 25 and hence kxk2 = ka + bk2 = kak2 + kbk2 . By the same argument, we have that k T (x)k2 = kT (a)k2 + kT (b)k2. In particular, it holds: k T (x)k2 ≤ kT k2(kak2 + kbk2) = kT k2kxk2 . This shows that k T k ≤ kT k. On the other hand, we know that HJ ı + 6= {0} and hence T (y ) = T (y ) for some y ∈ HJ ı + \ {0}. It follows that k T k = kT k. (b) Let x = a + b ∈ H, where a ∈ HJ ı + and b ∈ HJ ı − . Observe that J x = aı − bı ∈ D( T ) = D(T ) ⊕ Φ(D( T )) if and only if aı ∈ D(T ) and bı ∈ Φ(D(T )); that is, x ∈ D( T ). These considerations imply that J (D( T )) = D( T ). It follows that D(J T ) = D( T ) = D( T J ). Suppose now x = a + b ∈ D( T ). It holds: T J x = T (aı − bı) = T (aı) + T (bı) = T (a)ı − T (b)ı and J T x = J (T (a) − T (b)) = T (a)ı + T (b)ı = T (a)ı − T (b)ı = T J x. (c) If V = U for some Cı–linear operator U : D(V ) ∩ HJ ı + −→ HJ ı + , then we just know that J (D(V )) ⊂ D(V ) and J V = V J . Let us prove the converse implication. Given x ∈ D(V ) ∩ HJ ı + , it holds: J (V x) = V (J x) = V (xı) = (V x)ı. In other words, we have that V (D(V ) ∩ HJ ı + ) ⊂ HJ ı + . Define the Cı–linear operator U : D(V ) ∩ + −→ HJ ı HJ ı + by setting U u := V u. Since J (D(V )) ⊂ D(V ), D(U ) = D(V ) ∩ HJ ı + and V is a right H–linear extension of U , by the uniqueness of such an extension, we infer that V = U . − and D( T ) = D(T ) ⊕ Φ(D(T )), if D(T ) is dense in HJ ı (d) Since H = HJ ı + ⊕ HJ ı + , then D( T ) is automatically dense in H. Taking the adjoint in both members of the equality T J = J T , using J ∗ = −J and bearing in mind point (v) of Remark 2.16, we infer that J T ∗ = T ∗J , which includes J (D( T ∗)) ⊂ D( T ∗ ). Here T ∗ denotes ( T )∗ . + such that T ∗ = U . It remains Thanks to (c), there exists U : D( T ∗) ∩ HJ ı + −→ HJ ı to prove that U = T ∗ or, equivalently, that T ∗x = T ∗x for every x ∈ D( T ∗) ∩ HJ ı + . This is quite easy to see. Indeed, if x ∈ D( T ∗) ∩ HJ ı + , then T ∗x ∈ HJ ı + , because T ∗ commutes with J . Moreover, by definition of adjoint, it holds: h T ∗xvi = hx T vi for every v ∈ D( T ). In particular, it follows that h T ∗xvi = hxT vi for every v ∈ D(T ), which is equivalent to say that T ∗x = T ∗x, as desired. (e) Let x ∈ D( fS T ). As we have just proved, D( fS T ) = D(S T ) ⊕ Φ(D(S T )) so there exist, and are unique, a ∈ D(S T ) and b ∈ Φ(D(S T )) such that x = a + b. It follows that {T (a), T (b)} ⊂ D(S ) and hence T (x) = T (a) − T (b) ∈ D(S ) ⊕ Φ(D(S )) = D( S ). In other words, x belongs to D( fS T ). This implies that D( fS T ) ⊂ D( S T ). Proceeding similarly, one can prove the converse inclusion. The equality D( fS T ) = D( S T ), just proved, ensures that J (D( S T )) = J (D( fS T )) ⊂ D( fS T ) = D( S T ) and D( S T ) ∩ HJ ı + = D( fS T ) ∩ HJ ı + = D(S T ). In this way, being S T u = S T u for every u ∈ D(S T ), thanks to the uniqueness of the extension, we infer that S T = fS T . (f ) This point follows immediately from (e). (cid:3) 4. Resolvent and spectrum It is not so obvious how to extend the definitions of spectrum and resolvent in quaternionic Hilbert spaces. Let us focus on the simpler situation, where we are 26 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI looking for eigenvalues of a bounded right H–linear operator T on a quaternionic Hilbert space H. Without fixing any left scalar multiplication of H, the equation determining the eigenvalues reads as follows: T u = uq . Here a drawback arises: if q ∈ H \ R is fixed, the map u 7→ uq is not right H–linear, differently from u 7→ T u. Consequently, the eigenspace of q cannot be a right H– linear subspace. Indeed, if λ 6= 0, uλ is an eigenvector of λ−1 qλ instead of q itself. As a tentative way out, one could decide to deal with quaternionic Hilbert spaces equipped with a left scalar multiplication interpreting the eigenvalues equation in terms of such a left scalar multiplication: T u = qu. Now both sides are right H–linear. However, this approach is not suitable for physical applications [1], where self–adjoint operators should have real spectrum. As an elementary example, consider the finite dimensional quaternionic Hilbert space constructed over H = H ⊕ H equipped with the standard scalar product: if u, v , r, s ∈ H. h(r, s)(u, v)i := ru + sv The right H–linear operator represented by the matrix T = (cid:20) 0 −i 0(cid:21) , i where the multiplication is that on the left, is self–adjoint. (See [24] and refer- ences therin for several results in the finite-dimensional case). However, for every quaternion of the form q = a + cj + dk with a, c, d ∈ R and a2 + c2 + d2 = 1, there is an element wq ∈ H \ {0} with T wq = qwq . Although T = T ∗ , it admits non–real eigenvalues. Consequently, in the rest of the paper, we come back to the former approach keeping the constraint, found above, concerning the fact that each eigenvalue q brings a whole conjugation class of the quaternions, the eigensphere : Sq = {λ−1 qλ ∈ H λ ∈ H \ {0}}. As a matter of fact, we adopt the viewpoint introduced in [8], that is invariant under conjugation in the sense just pointed out. 4.1. Spherical resolvent and spectrum and their elementary properties. We are in a position to present one of the most important notion treated in this work, that of spectrum of a quaternionic operator. First of all, given a (right H– linear) operator T : D(T ) −→ H and q ∈ H, we define the associated operator ∆q (T ) : D(T 2 ) −→ H by setting: ∆q (T ) := T 2 − T (q + q) + Iq 2 . The fundamental suggestion by [8] is to systematically replace T − Iq for ∆q (T ) in the definition of resolvent set, resolvent operator and spectrum. The following definitions are a re-adaptation of the corresponding ones stated in [8]. A more accurate comparison appears immediately after the definition. Definition 4.1. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator. The spherical resolvent set of T is the set ρS (T ) ⊂ H of the quaternions q such that the three following conditions hold true: (a) Ker (∆q (T )) = {0}. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 27 (b) Ran (∆q (T )) is dense in H. (c) ∆q (T )−1 : Ran (∆q (T )) −→ D(T 2) is bounded. The spherical spectrum σS (T ) of T is defined by setting σS (T ) := H \ ρS (T ). It decomposes into three disjoint subsets as follows: (i) the spherical point spectrum of T : σpS (T ) := {q ∈ H Ker (∆q (T )) 6= {0}}; (ii) the spherical residual spectrum of T : σrS (T ) := nq ∈ H (cid:12)(cid:12)(cid:12) Ker (∆q (T )) = {0}, Ran (∆q (T )) 6= H o ; (iii) the spherical continuous spectrum of T : σcS (T ) := nq ∈ H (cid:12)(cid:12)(cid:12) Ker (∆q (T )) = {0}, Ran (∆q (T )) = H, ∆q (T )−1 6∈ B(H) o . The spherical spectral radius of T is defined as the following element rS (T ) of R+ ∪ {+∞}: rS (T ) := sup (cid:8)q ∈ R+ (cid:12)(cid:12) q ∈ σS (T )(cid:9). If T u = uq for some q ∈ H and u ∈ H \ {0}, then u is called eigenvector of T with eigenvalue q . Remark 4.2. (1) In [8], a different, but equivalent, definition of spherical resolvent set ρS (T ) is adopted for T ∈ B(H). Therein no decomposition onto the various parts of the spectrum has been introduced. In [8], the following definition (Def. 4.8.1) is stated (straightforwardly adapting it to the quaternionic Hilbert space case, since [8] studies the case of a quaternionic two–sided Banach modules): ρS (T ) = {q ∈ H ∆q (T ) : H −→ H is bijective and ∆q (T )−1 ∈ B(H)}. Actually, that definition is completely equivalent to our Definition 4.1 so that (4.1) is an identity assuming our definition of ρS (T ). Indeed, if ∆q (T ) is bijective and its inverse belongs to B(H), then conditions (a), (b) and (c) are evidently verified. Conversely, suppose that conditions (a), (b) and (c) hold true. Firstly, observe that (c) implies that Ran (∆q (T )) is closed in H. Let Ran (∆q (T )) ⊃ {yn}n∈N → y ∈ H and, for each n ∈ N, let xn := ∆q (T )−1yn ∈ H = D(∆q (T )). The sequence {xn}n∈N of H is a Cauchy sequence. Indeed, it holds: kxn − xmk ≤ k∆q (T )−1k kyn − ymk. It follows that {xn}n∈N → x for some x ∈ H. Since ∆q (T ) is continuous, we infer that ∆q (T )x = y and hence Ran (∆q (T )) is closed in H. Thanks to (a) and (b), ∆q (T ) turns out to be bijective with ∆q (T )−1 ∈ B(H). In the case of an unbounded operator defined on a subspace of H, we stick to Definition 4.1 since, on the one hand, the domains and boundedness of operators are irrelevant in that definition. On the other hand, our definition gives rise to statements that are straightforwardly extensions of the corresponding propositions in the theory in complex Hilbert spaces (see, in particular, (b) and (c) in Theo- rem 4.8 below). This is true in spite of the fact that definitions are substantially different, as they rely upon the second order polynomial ∆q (T ), rather than the first–order polynomial T − Iq . It is worth noticing that, referring to unbounded operators, [8] proposes a for- mally different definition (see Definition 4.15.2 in [8]), whose equivalence with ours, (4.1) 28 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI through a straightforward re–adaptation to the case of a quaternionic Hilbert space, would deserve investigation. It will be done elsewhere. (2) For every q ∈ H, define Eq (T ) := {u ∈ H T u = uq}. It is worth stressing that Eq (T ) is a right H–linear subspace of H if and only if either q ∈ R or q 6∈ R and Eq (T ) = {0}. Indeed, if q ∈ H \ R, u ∈ Eq (T ) \ {0} and p is a quaternion with qp 6= pq , then up 6∈ Eq (T ). (3) If p and q are conjugated quaternions, then ∆p (T ) = ∆q (T ). Indeed, p + p = q + q and p = q (see point (4) of Remark 2.1). We infer that ρS (T ) and σS (T ) are circular. The same is true for σpS , σrS and σcS . Recall that a subset A of H is called circular if it is equal to ΩK for some subset K of C. This is equivalent to say that A satisfies one of the following two equivalent conditions: • If α + ıβ ∈ A for some α, β ∈ R and ı ∈ S, then α + β ∈ A for every  ∈ S. • If q ∈ A and p ∈ H is conjugated to q , then p ∈ A. Equivalently, if q ∈ A, then Sq ⊂ A. (4) The spherical spectrum, or at least its intersection with a complex plane C , has already appeared in the literature, in the most general setting of real ∗–algebras. In [23], it is referred to as Kaplansky’s definition [22] of the spectrum. More generally, the spherical resolvent and the spherical spectrum can be defined for bounded right H–linear operators on quaternionic two–sided Banach modules in a form similar to that introduced above (see [8]). Several properties of resolvents and of spectra of bounded opertors on complex Banach or Hilbert spaces remain valid in that general context. Nevertheless, their proofs are by no means trivial (see [8] again). Here we recall some of these properties in the quaternionic Hilbert setting. Theorem 4.3. Let H be a quaternionic Hilbert space and let T ∈ B(H) be an operator. The fol lowing assertions hold. (a) Let q ∈ H with q > kT k and, for each n ∈ N, let an be the real number de- h=0 qh qn−h . Then the series Pn∈N T nan fined by setting an := q −2n−2 Pn converges absolutely in B(H) (with respect to the operator norm k · k) to ∆q (T )−1 . In particular, it holds: rS (T ) ≤ kT k. (b) σS (T ) is a non–empty compact subset of H. (c) Let P ∈ R[X ] and let P (T ) be the corresponding operator in B(H) defined in (2.9). Then, if T is self–adjoint, the fol lowing spectral map property holds: (4.2) σS (P (T )) = P (σS (T )). (d) For every n ∈ N, we have that σS (T n ) = (σS (T ))n := {qn ∈ H q ∈ σS (T )}. (e) Gelfand’s spectral radius formula holds: n→+∞ kT nk1/n . rS (T ) = lim In particular, if T is normal, then: rS (T ) = kT k. Proof. This result is a consequence of Theorems 4.7.4, 4.7.5, 4.8.11, 4.12.5, 4.12.6 and 4.13.2 of [8]. Actually, these theorems of [8] deals with the more general case of (4.3) QUATERNIONIC SLICE FUNCTIONAL CALCULUS 29 quaternionic two–sided Banach modules. In this way, in order to be able to apply it in our context, it suffices to fix a left scalar multiplication H ∋ q 7→ Lq of H and to equip B(H) with its natural quaternionic two–sided Banach module structure described in Section 3.1. For the convenience of the reader, we give a proof of the theorem. The following proof of point (e) is a detailed version of the corresponding one of [8]. With regards to (d), we simply give the exact reference in [8]. The proofs of the other points are different, and more direct. (a) Let q ∈ H with q > kT k. By direct computations, it is immediate to verify that a0 = q −2 , −(q + q)a0 + q 2a1 = 0 and an−2 − (q + q)an−1 + q 2 an = 0 for every n ≥ 2. Moreover, we have that an ≤ (n + 1)q −n−2 for every n ∈ N and hence lim supn→+∞ kT nank1/n ≤ kT k q −1 < 1. It follows that the series S := Pn∈N T nan converges absolutely in B(H). Furthermore, it holds: ∆q (T )S =S∆q (T ) = Iq2 a0 + T (−(q + q)a0 + q 2 a1 )+ + Xn≥2 T n (an−2 − (q + q)an−1 + q 2an ) = I. This proves that S is the inverse of ∆q (T ) in B(H) and hence rS (T ) ≤ kT k. (b) Firstly, we show that σS (T ) is compact. Bearing in mind the just proved inequality rS (T ) ≤ kT k, it suffices to see that σS (T ) is closed in H or, equivalently, that ρS (T ) is open in H. By point (c) of Proposition 2.11, the subset I of B(H) consisting of all isomorphisms is open. Since the map Θ : H −→ B(H), sending q into ∆q (T ), is continuous, it follows that ρS (T ) = Θ−1 (I ) is open in H. Now fix a left scalar multiplication of H. Let {1, i, j, k} be the standard or- thonormal basis of H. Identify C with Ci in the natural way. Define the map ψ : C ∩ ρS (T ) −→ B(H) by setting, for z = α + iβ with α, β ∈ R, ψ(z ) := ∆z (T )−1 (T − Iz ) = ∆α+iβ (T )−1(T − I(α − iβ )). Choose z ∈ C ∩ ρS (T ) with z > kT k and define the sequence {an}n∈N ⊂ R as in the statement of point (b) with q = z . It is easy to verify that an z − an−1 = z−n−1 if n ≥ 1. In this way, we infer that ψ(z ) = (cid:0)Pn∈N T nan (cid:1) (T − Iz ) = = −Iz−1 − Pn≥1 T n(an z − an−1 ) = − Pn∈N T nz−n−1 , (4.4) where the series Pn∈N T nz−n−1 converges absolutely in B(H). Let r ≥ rS (T ) and let Br be the open ball of C centered at 0 with radius r. Since the map ψ is real analytic with respect to α and β , and holomorphic for z > kT k, it is holomorphic on an open neighborhood U of C \ Br . This means that ∂ψ/∂α and ∂ψ/∂ β exist, are continuous and satisfy: (4.5) ∂ψ ∂α + ∂ψ ∂ β i = 0 in B(H) if z ∈ U. Let F : B(H) −→ H be a continuous right H–linear functional on B(H); that is, an element of B(H)′ , and let f0 , f1 , f2 , f3 : C∩ ρS (T ) −→ R and ℓ, m : C∩ ρS (T ) −→ C be the functions of class C 1 such that F ◦ ψ = f0 + if1 + j f2 + kf3 = ℓ + km 30 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI + + and hence ℓ = f0 + if1 and m = f3 + if2 . Thanks to (4.5), we infer that i(cid:19) = 0 = F (cid:18) ∂ψ ∂ (F ◦ ψ) ∂ (F ◦ ψ) ∂ψ i = + + ∂α ∂ β ∂α ∂ β = (cid:18) ∂ f0 ∂ β (cid:19) , ∂ β (cid:19) + k (cid:18) ∂ f3 ∂ β (cid:19) + j (cid:18) ∂ f2 ∂ β (cid:19) + i (cid:18) ∂ f1 ∂ f2 ∂ f3 ∂ f0 ∂ f1 ∂α − ∂α − ∂α ∂α which is equivalent to say that ℓ and m are holomorphic. Let us complete the proof of (b) by showing that σS (T ) 6= ∅. By (4.4), we have that (ℓ(z )2 + m(z )2 )1/2 = (F ◦ ψ)(z ) ≤ K kF k z −1 if z ≥ 1 + kT k, (4.6) where kF k is the operator norm of F and K := Pn∈N kT kn(1 + kT k)−n ∈ R+ . Suppose that σS (T ) = ∅. Thanks to (4.6), ℓ and m turn out to be bounded entire holomorphic functions with limz →+∞ ℓ(z ) = 0 = limz →+∞ m(z ). Liouville’s theorem ensures that ℓ and m are null, and hence F ◦ ψ is null for every F ∈ B(H )′ . Lemma 2.4 implies that ψ(z ) = 0 for every z ∈ C. It follows that T = Iz for every z ∈ C, which is impossible. (c) If P is constant, then (c) is trivial. Suppose that P has positive degree. Let us prove that P (σS (T )) ⊂ σS (P (T )). Let q ∈ σS (T ). Since T is assumed to be self–adjoint, as we have just said, Theorem 4.8 below, which is completely independent from the proposition we are proving, implies that q ∈ R. It follows that P (q) ∈ R and ∆P (q) (P (T )) = (P (T ) − IP (q))2 . The polynomial P (X ) − P (q) in R[X ] vanishes at X = q and hence there exists θ ∈ R[X ] such that P (X ) − P (q) = (X − q)θ(X ). In particular, we infer that ∆P (q) (P (T )) = (T − Iq)2 (θ(T ))2 = ∆q (T )(θ(T ))2 = (θ(T ))2∆q (T ). (4.7) It follows that P (q) ∈ σS (P (T )); that is, ∆P (q) (P (T )) is not invertible in B(H). Otherwise, thanks to (4.7), ∆q (T ) would be invertible in B(H) as well, contradicting the fact that q ∈ σS (T ). It remains to show that σS (P (T )) ⊂ P (σS (T )). Let q ∈ σS (P (T )). Since P has real coefficients, the operator P (T ) is self–adjoint. Therefore, q ∈ R and ∆q (P (T )) = (P (T ) − Iq)2 . Fix ı ∈ S. By the Fundamental Theorem of Algebra, there exist α0 ∈ R \ {0}, α1 , . . . , αh ∈ R and αh+1 , . . . , αk ∈ Cı \ R for some h, k ∈ N with h ≤ k such that kYℓ=h+1 hYℓ=1 P (X ) − q = α0 (X − αℓ ) ∆αℓ (X ). In particular, it holds: ∆αℓ (T ) ∆q (P (T )) = (P (T ) − Iq)2 = α2 0 kYℓ=h+1 hYℓ=1 (∆αℓ (T ))2 . If ∆αℓ (T ) were invertible in B(H) for every ℓ ∈ {1, . . . , k}, then ∆q (P (T )) would be invertible in B(H) as well, contradicting the fact that q ∈ σS (P (T )). It follows that there exists ℓ ∈ {1, . . . , k} (or better ℓ ∈ {1, . . . , h}) such that αℓ ∈ σS (T ). Since P (αℓ ) − q = 0, q = P (αℓ ) belongs to P (σS (T )), as desired. (d) This point is proved in Theorem 4.12.5 of [8]. (e) Let us follows the proof of Theorem 4.12.6 of [8]. By combining (d) with the last part of (a), we infer that rS (T n ) = rS (T )n ≤ kT nk for every n ∈ N and hence QUATERNIONIC SLICE FUNCTIONAL CALCULUS 31 = − dξ = m(ξ ) ξ − z rS (T ) ≤ lim inf n→+∞ kT nk1/n . In order to prove (4.2), it suffices to show that n→+∞ kT nk1/n ≤ rS (T ). lim sup (4.8) Let us prove this inequality. Consider again the maps ψ : C ∩ ρS (T ) −→ B(H), F ∈ B(H)′ and ℓ, m : C ∩ ρS (T ) −→ C introduced in the proof of point (b). Let r > rS (T ), let Br be the open ball of C centered at 0 with radius r and let ¯Br be its closure in C. Since ℓ and m are holomorphic on an open neighborhood of C \ Br , it holds: 2πi Z∂Br 2πi Z∂Br 1 1 ℓ(ξ ) (F ◦ ψ)(z ) = ℓ(z ) + km(z ) = − dξ − k ξ − z 2π Z∂Br 1 (ℓ(ξ ) + km(ξ ))(ξ − z )−1 i−1 dξ = 2π Z∂Br 1 (F ◦ ψ)(ξ )(ξ − z )−1 i−1 dξ = = − ψ(ξ )(ξ − z )−1 i−1 dξ(cid:19) if z ∈ C \ ¯Br , = F (cid:18)− 2π Z∂Br 1 where the integral R∂Br ψ(ξ )(ξ − z )−1 i−1 dξ can be defined as an operator norm limit of Riemann sums. Lemma 2.4 implies the following Cauchy formula for ψ : 2π Z∂Br 1 if z ∈ C \ ¯Br . ψ(ξ )(ξ − z )−1 i−1 dξ ψ(z ) = − z )−1 = − Pn∈N ξn z−n−1 if ξ ∈ ∂Br and Fix z ∈ C \ ¯Br . Since (ξ − z )−1 = z−1(1 − ξ the series − Pn∈N ξn z−n−1 converges absolutely for ξ ∈ ∂Br , ψ can be expanded into Laurent series on C \ ¯Br as follows: ψ(ξ )ξn i−1 dξ(cid:19) z−n−1 ψ(ξ )ξn z−n−1 i−1 dξ = Xn∈N (cid:18) 1 2π Z∂Br 2π Z∂Br ψ(z ) = Xn∈N 1 Comparing the latter equality with (4.4) and bearing in mind the uniqueness of 2π R∂Br the Laurent series expansion, we infer at once that T n = 1 ψ(ξ )ξn i−1 dξ for every n ∈ N and the series Pn∈N T nz−n−1 converges absolutely for every z ∈ C with z > rS (T ). In particular, limn→+∞ kT nk z −n−1 = 0 and hence the sequence {kT nk z −n}n∈N ⊂ R is bounded. It follows that lim supn→+∞ kT nk1/n ≤ z for every z ∈ C with z > rS (T ). This proves (4.8) and hence (4.2). If T is normal, then it holds: kT 2k2 = k(T 2)∗T 2k = k(T ∗)2T 2k = k(T ∗T )∗(T ∗T )k = kT ∗T k2 = (kT k2)2 and hence kT 2k = kT k2 . Proceeding by induction, one easily shows that kT 2k k1/2k kT k for every k ∈ N. Now (4.3) follows from (4.2). Corollary 4.4. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a self– adjoint operator and let P ∈ R[X ] such that σS (T ) is finite and P vanishes on σS (T ). Then P (T ) is the nul l operator in B(H). = (cid:3) Proof. Since T is self–adjoint, it is immediate to verify that P (T ) is self–adjoint as well. By point (c) of Theorem 4.3, we infer that σS (P (T )) = {0} and hence (4.3) implies that P (T ) = 0, as desired. (cid:3) 32 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI 4.2. Some spectral properties of operators on quaternionic Hilbert spaces. Regardless different definitions with respect to the complex Hilbert space case, the notions of spherical spectrum and resolvent set enjoy some properties which are quite similar to those for complex Hilbert spaces. We go to illustrate them together with some other features that, conversely, are proper to the quaternionic Hilbert space case. First of all, quite remarkably, it turns out that σpS (T ) coincides with the set of eigenvalues of T . Proposition 4.5. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator. Then σpS (T ) coincides with the set of al l eigenvalues of T . Proof. Let u be an eigenvector of T with eigenvalue q ∈ H. By the definition of ∆q (T ), we have that ∆q (T )u = T (T u − uq) − (T u − uq)q = 0 and hence q ∈ σpS (T ), because u 6= 0. Conversely, if q ∈ σpS (T ), then there is v ∈ D(T 2 ) such that 0 = ∆q (T )v = T (T v − vq) − (T v − vq)q . If v ′ := T v − vq = 0, then q is an eigenvalue of T . Otherwise v ′ is an eigenvector of T with eigenvalue q . Since q = sqs−1 for some s 6= 0, T (v ′s) = (v ′ s)q and v ′s 6= 0, we infer that q is an eigenvalue of T as well. (cid:3) Remark 4.6. In our context, the subspace Ker (∆q (T )) has the role of an “eigenspace”. In particular, it holds: Ker (∆q (T )) 6= {0} if and only if Sq is an eigensphere of T . There is an important difference from the standard complex Hilbert space: if T ∈ B(H), then T and T ∗ have always the same spherical spectrum and, as we shall prove later, T and T ∗ are even unitarily equivalent, whenever T is normal. Proposition 4.7. Let H be a quaternionic Hilbert space and let T ∈ B(H) be an operator. Then ρS (T ) = ρS (T ∗ ) and σS (T ) = σS (T ∗ ). Proof. Since ∆q (T )∗ = ∆q (T ∗) for every q ∈ H, point (viii) of Remark 2.16 and point (1) of Remark 4.2 immediately imply that ρS (T ) = ρS (T ∗ ) and hence σS (T ) = σS (T ∗ ). (cid:3) We are now in a position to establish an important result concerning the spec- trum of normal, self–adjoint, anti self–adjoint and unitary operators. That state- ment, at first glance, sounds quite weird, since it declares that the point spectrum of an anti self–adjoint and unitary operator T on H completely fills a sphere. In this way, σpS (T ) has to be uncountable in all cases, even if the quaternionic Hilbert space H is separable. However, differently from the complex Hilbert space case, there is no contradiction now. Indeed, in quaternionic Hilbert spaces, eigenvectors of two different eigenvalues are not mutually orthogonal in general, unless both the eigenvalues are real. Theorem 4.8. Let H be a quaternionic Hilbert space and let T : D(T ) −→ H be an operator with dense domain. The fol lowing assertions hold. (a) If T ∈ B(H) is normal, then we have that (i) σpS (T ) = σpS (T ∗ ), (ii) σrS (T ) = σrS (T ∗ ) = ∅, QUATERNIONIC SLICE FUNCTIONAL CALCULUS 33 (iii) σcS (T ) = σcS (T ∗ ). (b) If T is self–adjoint (not necessarily in B(H)), then σS (T ) ⊂ R and σrS (T ) is empty. (c) If T is anti self–adjoint (not necessarily in B(H)), then σS (T ) ⊂ Im(H) and σrS (T ) is empty. (d) If T ∈ B(H) is unitary, then σS (T ) ⊂ {q ∈ H q = 1}. (e) If T ∈ B(H) is anti self–adjoint and unitary, then σS (T ) = σpS (T ) = S. Proof. (a) Since ∆q (T )∗ = ∆q (T ∗) and T is assumed to be normal, it follows immediately that ∆q (T ) is normal as well. Therefore, thanks to Proposition 2.19, we know that Ker (∆q (T )) = Ker (∆q (T ∗ )). This equality implies at once that σpS (T ) = σpS (T ∗). Let us prove that σrS (T ) = ∅. Suppose on the contrary that σrS (T ) contains a quaternion q . It would follows that ⊥ {0} = Ker (∆q (T )) = Ker (∆q (T ∗ )) = Ran (∆q (T )) 6= {0}, which is a contradiction. Similarly, one can prove that σrS (T ∗) = ∅. Since σS (T ) = σS (T ∗ ) by Proposition 4.7, σpS (T ) = σpS (T ∗), σrS (T ) = σrS (T ∗) and the three components of the spherical spectrum are pairwise disjoint, we infer that σcS (T ) = σcS (T ∗) as well. (b) Consider q = r + ν ∈ H with r ∈ R and ν ∈ Im(H) \ {0}. We intend to show that q ∈ ρS (T ), which is equivalent to σS (T ) ⊂ R. One has: ∆q (T ) = T 2 − T 2r + Ir2 + Iν 2 = (T − Ir)2 + Iν 2 . Since T is self–adjoint, T − Ir is self–adjoint as well. In particular, T − Ir is closed with dense domain. Define the operator Sr : D(T 2) −→ H by setting Sr := (T −Ir)2 . Applying point (d) of Theorem 2.15, we infer that D(Sr ) = D(T 2 ) = D(∆q (T )) is dense in H and Sr is self–adjoint. Consequently, ∆q (T ) = (T − Ir)2 + Iν 2 is self–adjoint as well. Since T − Ir and Sr are self–adjoint, if u ∈ D(T 2), then we have that huSr ui = h(T − Ir)u(T − Ir)ui ≥ 0, k∆q (T )uk2 = hSr u + uν 2 Sr u + uν 2 i = kSr uk2 + 2ν 2 hu Sr ui + ν 4 kuk2 and hence we can write that k∆q (T )uk ≥ ν 2 kuk. (4.9) The latter inequality implies at once that Ker (∆q (T )) = {0} and ∆q (T )−1 : Ran (∆q (T )) −→ D(T 2) is bounded. Bearing in mind Proposition 2.14 and the fact that ∆q (T ) is self–adjoint, we observe: Ran (∆q (T )) = (Ran (∆q (T ))⊥ )⊥ = Ker (∆q (T )∗)⊥ = = Ker (∆q (T ))⊥ = {0}⊥ = H. This proves that q ∈ ρS (T ) and hence that σS (T ) ⊂ R, as desired. The proof of the fact that σrS (T ) = ∅ is similar to the one given above. p ∈ σrS (T ), then we obtain the following contradiction: ⊥ {0} = Ker (∆p (T )) = Ker (∆p (T )∗ ) = Ran (∆p (T )) 6= {0}. (c) Let λ = r + ν ∈ H with r ∈ R \ {0} and ν ∈ Im(H). We must prove that λ ∈ ρS (T ). First, we show that k∆λ (T )uk ≥ r2 kuk if u ∈ D(T 2 ), (4.10) If 34 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI which implies that Ker (∆λ (T )) = {0} and ∆λ (T )−1 : Ran (∆λ (T )) −→ D(∆λ (T )) is bounded. Let u ∈ D(T 2). Since T is anti self–adjoint, we have that hT 2uT ui + hT uT 2ui = 0 = hT uui + huT ui and hT 2uui + huT 2ui = −2kT uk2. In particular, it holds: k∆λ (T )uk2 = kT 2uk2 + (r2 + ν 2 )2 kuk2 + 2(r2 − ν 2 )kT uk2. If r2 − ν 2 ≥ 0, then (4.10) is evident. Suppose that r2 − ν 2 < 0. Since kT uk2 = −huT 2ui ≤ kuk kT 2uk, we have that 2(r2 − ν 2 )kT uk2 ≥ 2(r2 − ν 2 )kuk kT 2uk. It follows that k∆λ (T )uk2 ≥ kT 2uk2 + (r2 + ν 2 )2 kuk2 + 2(r2 − ν 2 )kuk kT 2uk = = (kT 2uk − ν 2kuk)2 + 2r2kuk kT 2uk + (r4 + 2r2 ν 2 )kuk2 ≥ ≥ r4 kuk2 . Inequality (4.10) is proved. Now we show that D(T 2 ) is dense in H and ∆λ (T )∗ = ∆−r+ν (T ). Applying point (d) of Theorem 2.15 to T , we obtain that D(T 2 ) is dense in H and T 2 is self–adjoint. Evidently, the operator T 2 + Ir2 : D(T 2) −→ H is self–adjoint as well. In this way, bearing in mind point (iv) of Remark 2.16, we infer that T 2 + Ir2 = (T − Ir)2 + T 2r = ((T − Ir)2 + T 2r)∗ ⊃ ((T − Ir)2 )∗ + T ∗2r = = ((T − Ir)2 )∗ − T 2r and hence (T + Ir)2 = (T − Ir)2 + T 4r ⊃ ((T − Ir)2 )∗ . By point (v) of Remark 2.16, we have also that ((T − Ir)2 )∗ ⊃ (T ∗ − Ir)2 = (T + Ir)2 . It follows that ((T − Ir)2 )∗ = (T + Ir)2 and hence ∆λ (T )∗ = ((T − Ir)2 + Iν 2 )∗ = (T + Ir)2 + Iν 2 = ∆−r+ν (T ), as desired. Thanks to (4.10), we know that k∆−r+ν (T )uk ≥ r2 kuk for every u ∈ H. In particular, we infer that Ker (∆−r+ν (T )) = {0}. In this way, Proposition 2.14 implies that Ran (∆λ (T )) = (Ran (∆λ (T ))⊥ )⊥ = Ker (∆λ (T )∗ )⊥ = = Ker (∆−r+ν (T ))⊥ = {0}⊥ = H. We have just established that λ ∈ ρS (T ) and hence that σS (T ) ⊂ Im(H). The equality σrS (T ) = ∅ can be proved as in the proof of (a). (d) Since ∆0 (T ) = T 2 , it is obvious that ∆0 (T ) is bijective and its inverse coincides with (T ∗ )2 ∈ B(H). It follows that 0 ∈ ρS (T ). If q > 1 = kT k, then point (a) of Theorem 4.3 ensures that q ∈ ρS (T ). Let 0 < q < 1. We have: ∆q (T ) = T 2 − T (q + q) + Iq 2 = = ((T ∗ )2 − T ∗(q−1 + q−1 ) + Iq−1 2 )T 2 q 2 = = ∆q−1 (T ∗ )T 2 q 2 . The operator T ∗ is unitary and q−1 = q −1 > 1. In this situation, we know that ∆q−1 (T ∗) is bijective and has bounded inverse. It follows that ∆q (T ) is bijective and has bounded inverse as well. This proves that q ∈ ρS (T ). We have just established that, if q ∈ ρS (T ), then q 6= 1, completing the proof of (d). QUATERNIONIC SLICE FUNCTIONAL CALCULUS 35 (e) The fact that σS (J ) ⊂ S is an immediate consequence of (c), (d). Moreover, due to (d) in Proposition 3.8 and Proposition 4.5, every ı ∈ S belongs to σpS (T ). Since σpS (T ) ⊂ σS (T ) the thesis holds. (cid:3) 5. Real measurable functional calculus for self–adjoint operators and slice nature of normal operators Let H be a quaternionic Hilbert space and let T ∈ B(H) be a normal operator. This part focuses on the problem concerning the existence of self–adjoint operators A, B ∈ B(H) and of an anti self–adjoint and unitary operator J ∈ B(H) such that T decomposes as T = A + J B and J commutes with T and T ∗ . We shall prove not only that operators A, B and J exist (Theorem 5.9), but even that there exists a left scalar multiplication H ∋ q 7→ Lq of H such that Lı = J for some ı ∈ S and Lq commutes with A and B for every q ∈ H (Theorem 5.14). As a by–product, we will give a proof of the known fact [34] that normal operators are unitarily similar to their adjoint operators. Furthermore, in Proposition 5.11 and Corollary 5.13, we will also establish the relation between the spherical spectrum of T and the standard spectrum of the restriction of T to the complex Hilbert subspaces HJ ı ± of H defined in Definition 3.6. The way we shall follow to prove the mentioned results relies upon some tools of measurable functional calculus for self–adjoint operators on quaternionic Hilbert spaces, which are interesting on their own right. 5.1. The operator J0 . It happens that, for a fixed normal operator T ∈ B(H), there is an anti self–adjoint operator J0 , isometric on Ran (T − T ∗ ) = Ker (T − T ∗)⊥ and vanishing on Ker (T − T ∗), uniquely determined by T , that commutes with T and T ∗ , and induces an apparently familiar decomposition of T into a complex combination of self–adjoint operators: . (5.2) T = (5.1) T = (T + T ∗) 1 1 + J0 T − T ∗ 2 2 In the special case in which Ker (T − T ∗) = {0}, if we fix an imaginary unit ı of H and a left scalar multiplication H ∋ q 7→ Lq of H with Lı = J0 (see Proposition 3.8), then we can also write: 1 1 (T − T ∗). (T + T ∗) + J0 2ı 2 Notice that, if J0 did not commute with T − T ∗ , then the operator 1 2ı (T − T ∗) could not be self–adjoint. So commutativity plays a crucial role here. Theorem 5.1. Let H be a quaternionic Hilbert space and let T ∈ B(H) be an operator. Then there exists, and is unique, an operator J0 ∈ B(H) such that: (i) J0 is anti self–adjoint, (ii) decomposition (5.1) holds true, (iii) Ker (T − T ∗) = Ker (J0 ), (iv) J0 J ∗ 0 = I on Ker (T − T ∗)⊥ , (v) J0 commutes with T − T ∗ . Moreover, J0 (Ker (T − T ∗)⊥ ) ⊂ Ker (T − T ∗)⊥ and, if T is normal, then J0 commutes also with al l the operators in B(H) commuting with both T and T ∗ . 36 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI Proof. By applying Theorem 2.20 (see also Remark 2.22) to the anti self–adjoint operator T − T ∗ , we obtain an anti self–adjoint operator W ∈ B(H) such that T − T ∗ = W T − T ∗ , Ker (T − T ∗ ) = Ker (T − T ∗ ) ⊂ Ker (W ) and W W ∗ = I on Ker (T − T ∗)⊥ . Moreover, W commutes with T − T ∗ and with all the operators in B(H) commuting with T − T ∗ (and with (T − T ∗)∗ = −(T − T ∗)). In particular, if T is normal, then W commutes also with T and T ∗ . Evidently, J0 := W has the desired properties. Let us show that such an operator J0 is unique. Let J1 ∈ B(H) be an operator satisfying conditions (i)–(v) (with J0 replaced by J1 ). By combining (i), (ii) and (v) with the fact that T − T ∗ is self–adjoint, we infer that 1 1 2 − J1 T − T ∗ T ∗ = (T + T ∗) . 2 It follows that T − T ∗ = J1 T − T ∗ . Define P := T − T ∗ . Bearing in mind (iii), (iv), the positivity of P and the uniqueness of the quaternionic polar decomposition of T − T ∗ (see Theorem 2.20), we infer at once that J0 = J1 . (cid:3) As an immediate consequence, we obtain: Corollary 5.2. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a normal operator, let J0 ∈ B(H) be the operator with the properties listed in the statement of Theorem 5.1 and let ı ∈ S. Suppose that Ker (T − T ∗) = {0}. Then J0 is unitary. Moreover, if H ∋ q 7→ Lq is a left scalar multiplication of H such that Lı = J0 , then 2 (T + T ∗) and 1 both operators 1 2ı (T − T ∗) are self–adjoint and decomposition (5.2) holds. 5.2. Continuous and measurable real functions of a self-adjoint operator. Throughout this section, given a quaternionic Hilbert space H, we consider B(H) as a real Banach algebra with unity I; that is, here B(H) denotes the set of (bounded right H–linear) operators on H, equipped with the pointwise sum, with the real scalar multiplication defined in (2.8), with the composition as product and with the norm defined in (2.6). Let K be a non–empty subset of R. We denote by C (K, R) the commutative real Banach unital algebra of continuous real–valued functions defined on K. As usual, the algebra operations are the natural ones defined pointwisely, the unity 1K is the function constantly equal to 1 and the norm is the supremum one k · k∞ . It is worth recalling the notion of real polynomial function. Definition 5.3. A function p : K −→ R is said to be a real polynomial function if there exists a polynomial P ∈ R[X ] such that p(t) = P (t) for each t ∈ K. Remark 5.4. Since the zero set of a non–null real polynomial is finite and every finite subset is the zero set of a non–null polynomial, we infer that a real polynomial function on K is induced by a unique polynomial in R[X ] if and only if K is infinite. In other words, the homomorphism from R[X ] to C (K, R), sending P into the real polynomial function induced by P itself, is injective if and only if K is infinite. Before stating the next result, we remind the reader that the spherical spec- trum of a self–adjoint bounded operator is a non–empty compact subset of R (see Theorems 4.3(b) and 4.8(b)). QUATERNIONIC SLICE FUNCTIONAL CALCULUS 37 Theorem 5.5. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a self– adjoint operator. Then there exists, and is unique, a continuous homomorphism ΦT : C (σS (T ), R) ∋ f 7→ f (T ) ∈ B(H) of real Banach unital algebras such that: (i) ΦT is unity–preserving; that is, ΦT (1σS (T ) ) = I. (ii) ΦT (id ) = T , where id : σS (T ) ֒→ R denotes the inclusion map. The fol lowing further facts hold true. (a) The operator f (T ) is self–adjoint for every f ∈ C (σS (T ), R). (b) ΦT is isometric; that is, kf (T )k = kf k∞ for every f ∈ C (σS (T ), R). (c) ΦT is positive; that is, f (T ) ≥ 0 if f ∈ C (σS (T ), R) and f (t) ≥ 0 for every t ∈ σS (T ). (d) For every f ∈ C (σS (T ), R), f (T ) commutes with every element of B(H) that commutes with T . Proof. The present proof is organized into two steps. Step I. Suppose that σS (T ) is infinite. Let us prove the uniqueness of ΦT . First, observe that, thanks to (i) and (ii), if p : σS (T ) −→ R is a real polynomial function induced by the (unique) polynomial P ∈ R[X ], then ΦT (p) = P (T ), where P (T ) is the operator in B(H) defined in (2.9). Let Ψ : C (σS (T ), R) −→ B(H) be another unity–preserving continuous homomorphism with Ψ(id ) = T . It follows that the map Ψ − ΦT is continuous and vanishes on every real polynomial function defined on σS (T ). By the Weierstrass approximation theorem, the set of all real polynomial functions on σS (T ) is dense in C (σS (T ), R) and hence Ψ = ΦT . Let us construct ΦT . For every real polynomial function p : σS (T ) −→ R, define ΦT (p) := P (T ), where P is the unique polynomial in R[X ] inducing p. Since T is self–adjoint, the operator P (T ) is also self–adjoint and hence points (c) and (e) of Theorem 4.3 imply that kΦT (p)k = sup{P (r) ∈ R+ r ∈ σS (T )} = kpk∞ . By continuity, ΦT extends uniquely to an isometric homomorphism on the whole C (σS (T ), R), which satisfies (a) and (b). Evidently, by construction, ΦT satisfies also (i) and (ii). Let us prove (c). Given a function f ∈ C (σS (T ), R) with f (t) ≥ 0 for every t ∈ σS (T ), we can apply ΦT to √f ∈ C (σS (T ), R), obtaining the operator √f (T ) ∈ B(H). Since √f (T )√f (T ) = f (T ) and √f (T ) is self–adjoint, it holds: huf (T )ui = hupf (T )pf (T )ui = hpf (T )upf (T )ui ∈ R+ if u ∈ H, as desired. Point (d) is evident if f is a real polynomial function. By continuity and by the Stone–Weierstrass approximation theorem, (d) turns out to be true for every f ∈ C (σS (T ), R). Step II. Suppose now that σS (T ) is finite. The uniqueness of ΦT is easy to see. Indeed, if f : σS (T ) −→ R is a function (which is always continuous in this case), then there exists P ∈ R[X ] such that f (t) = P (t) for each t ∈ σS (T ). In this way, bearing in mind (i) and (ii), we have that ΦT (f ) must be equal to P (T ). Now, Corollary 4.4 ensures that such a definition of ΦT is consistent, because it does not depend on the choice of P , but only on f . Points (a) and (d) are evident. Points (b) and (c) can be proved as in Step I. (cid:3) Next step consists in extending the notion of function f (T ) of a self–adjoint operator T to the case in which f : σS (T ) −→ R is Borel–measurable and bounded. 38 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI In the following, we consider the non–empty subset K of R to be equipped with the relative euclidean topology and we denote by B(K) the Borel σ–algebra of K. Moreover, we indicate by M (K, R) the commutative real Banach unital algebra of bounded Borel–measurable functions from K to R. As before, the algebra operations are the natural ones defined pointwisely, the unity is the function 1K and the norm is the supremum one k · k∞, and not the essential–supremum one, since no measure has been adopted on K. Theorem 5.6. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a self– adjoint operator. Then there exists, and is unique, a continuous homomorphism bΦT : M (σS (T ), R) ∋ f 7→ f (T ) ∈ B(H) of commutative real Banach unital algebras such that: (i) bΦT is unity–preserving and bΦT (id ) = T . (ii) f (T ) is self–adjoint for every f ∈ M (σS (T ), R). (iii) If a sequence {fn}n∈N in M (σS (T ), R) is bounded and converges pointwisely to some f ∈ M (σS (T ), R), then { bΦT (fn )}n∈N → bΦT (f ) in the weak opera- tor topology; that is, {F ( bΦT (fn )(u))}n∈N → F ( bΦT (f )(u)) in H for every u ∈ H and F ∈ H′ . The fol lowing further facts hold. (a) bΦT extends ΦT ; that is, bΦT = ΦT on C (σS (T ), R). (b) bΦT is norm–decreasing; that is, kf (T )k ≤ kf k∞ for every f ∈ M (σS (T ), R). (c) bΦT is positive; that is, f (T ) ≥ 0 if f ∈ M (σS (T ), R) and f (t) ≥ 0 for every t ∈ σS (T ). (d) For every u ∈ H, define the function µu : B(σS (T )) −→ R+ by setting µu (E ) := hu bΦT (χE )ui, where χE is the characteristic function of the Borel subset E of σS (T ). Then each µu is a finite positive σ–additive regular Borel measure on σS (T ) and, for every f ∈ M (σS (T ), R), it holds: kf (T )uk2 = ZσS (T ) f 2 dµu . (e) For every f ∈ M (σS (T ), R), f (T ) commutes with every element of B(H) that commutes with T . (f ) If a sequence {fn}n∈N in M (σS (T ), R) is bounded and converges pointwisely to some f ∈ M (σS (T ), R), then { bΦT (fn )}n∈N → bΦT (f ) in the strong opera- tor topology; that is, { bΦT (fn )(u)}n∈N → bΦT (f )(u) in H for every u ∈ H. Proof. We begin with the existence issue proving en passant the validity of points (a)–(f ). We subdivide this part of the proof into three steps. Step I. Fix u ∈ H. Bearing in mind point (c) of Theorem 5.5, we can define the map (5.3) C (σS (T ), C) ∋ f 7→ huRe(f )(T )ui + ihuIm(f )(T )ui ∈ C, where C (σS (T ), C) denotes the usual complex Banach space of C–valued continuous functions on σS (T ). That map defines a complex linear functional on C (σS (T ), C), which is positive by (c) of Theorem 5.5. Riesz’ theorem for Borel measures on QUATERNIONIC SLICE FUNCTIONAL CALCULUS 39 (5.4) (5.5) f dµu if f ∈ C (σS (T ), R). Hausdorff locally compact spaces (see Theorem 2.14 of [30]) implies that there is a unique positive σ–additive regular Borel measure µu : σS (T ) −→ R+ such that huRe(f )(T )ui + ihuIm(f )(T )ui = ZσS (T ) if f ∈ C (σS (T ), C) f dµu and hence huf (T )ui = ZσS (T ) As a consequence, we obtain: µu (σS (T )) = ZσS (T ) 1σS (T ) dµu = huIui = kuk2 . Step II. Fix u, v ∈ H. Let us prove the following polarization–type identity: 4huf (T )vi =hu + v f (T )(u + v)i − hu − v f (T )(u − v)i+ + (hui + v f (T )(ui + v)i − hui − v f (T )(ui − v)i) i+ + (huj + v f (T )(uj + v)i − huj − v f (T )(uj − v)i) j+ + (huk + v f (T )(uk + v)i − huk − v f (T )(uk − v)i) k . Let α, β , γ , δ ∈ R such that huf (T )vi = α + β i + γ j + δk . By point (a) of Theorem 5.5, we know that f (T ) is self–adjoint and hence it holds: hu + v f (T )(u + v)i − hu − v f (T )(u − v)i = 2huf (T )vi + 2hv f (T )ui = = 2huf (T )vi + 2hf (T )v ui = = 2huf (T )vi + 2huf (T )vi = = 4Re(huf (T )vi) = 4α. (5.6) In particular, we have: hui + v f (T )(ui + v)i − hui − v f (T )(ui − v)i = 4Re(−ihuf (T )vi) = 4β , huj + v f (T )(uj + v)i − huj − v f (T )(uj − v)i = 4Re(−j huf (T )vi) = 4γ , huk + v f (T )(uk + k)i − huk − v f (T )(uk − v)i = 4Re(−khuf (T )vi) = 4δ. Equality (5.6) follows immediately. For every a ∈ {0, 1, 2, 3}, define the signed measure ν (a) u,v : B(σS (T )) −→ R by setting 1 1 ν (a) if E ∈ B(σS (T )). µuea+v (E ) − u,v (E ) := µuea−v (E ) 4 4 For convenience, define e0 := 1, e1 := i, e2 := j and e3 := k . Thanks to (5.4) and (5.6), we infer at once that 3Xa=0 ZσS (T ) If we think of (finite) signed measures as sub–cases of complex measures, then we can exploit the known result (valid in general for complex regular Borel measures in view of the complex measure Riesz’ representation theorem [30, Theorem 6.19]) that two such measures coincide when the corresponding integrals of real continuous if f ∈ C (σS (T ), R). huf (T )vi = f dν (a) u,v ea (5.7) 40 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (5.8) (5.9) compactly supported functions produce the same results. In that way, by (5.7), one easily obtain that u,v + ν (a) ν (a) u,v+v′ = ν (a) u,v′ 3Xa=0 3Xa=0 In addition, since f (T ) is self–adjoint, we have that 3Xa=0 ZσS (T ) if v ′ ∈ H and a ∈ {0, 1, 2, 3}, ν (a) if q ∈ H. u,v ea q f dµ(a) v,u ea = hv f (T )ui = hf (T )uvi = huf (T )vi = 3Xa=0 ZσS (T ) = ZσS (T ) u,v = −µ(a) v,u and µ(a) u,v = µ(0) µ(0) v,u if a ∈ {1, 2, 3}. Moreover, equality (5.5) implies that ν (a) u,v (σS (T )) ≤ kuk kvk if a ∈ {0, 1, 2, 3}. f dµ(0) u,v − f dµ(a) v,u ea ν (a) u,vq ea = and hence (5.10) (5.11) Indeed, it holds: 4ν (a) u,v (σS (T )) = kuea + vk2 − kuea − vk2 = 4Re(huea vi) = = 4Re(−ea huvi) ≤ 4huvi ≤ 4kuk kvk. Viewing µu as a signed measure, one can compare (5.4) with (5.7) using the same uniqueness property established in Riesz’ theorem for complex Borel measures [30, Theorem 6.19], obtaining: H × H ∋ (u, v) 7→ Qf (u, v) := and ν (a) ν (0) u,u = 0 if a ∈ {1, 2, 3}. (5.12) u,u = µu Step III. Let f ∈ M (σS (T ), R). Define the quadratic form 3Xa=0 ZσS (T ) f dν (a) u,v ea ∈ H. In view of (5.8)–(5.11), that quadratic form is right H–linear in the second compo- nent and quaternionic hermitian; that is, Qf (u, vq + v ′ q ′ ) = Qf (u, v)q + Qf (u, v ′ )q ′ and Qf (u, v) = Qf (v , u) if u, v , v ′ ∈ H and q , q ′ ∈ H. Moreover, Qf is bounded: Qf (u, v) ≤ 4kf k∞kuk kvk. Quaternionic Riesz’ theorem (see Theorem 2.8) immediately implies that there exists a unique operator in B(H), we shall indicate by f (T ) again, such that huf (T )vi = Qf (u, v) if u, v ∈ H. That operator is self–adjoint; that is, (ii) is verified. Indeed, it holds: if u, v ∈ H. huf (T )vi = Qf (u, v) = Qf (v , u) = hv f (T )ui = hf (T )uvi The map bΦT : M (σS (T ), R) −→ B(H), sending f into f (T ), satisfies (a); that is, it is an extension of ΦT . This is immediate consequence of (5.7) and of the definition of f (T ). Point (a) and points (i) and (ii) of Theorem 5.5 implies at once (i). (5.13) = f g dν (a) u,v ea if u, v ∈ H. f g dν (a) u,v ea = g dν (a) f (T )u,v ea . QUATERNIONIC SLICE FUNCTIONAL CALCULUS 41 Let us prove that bΦT is a homomorphism of real algebras. By construction, it is evident that it is R–linear. Let x, y ∈ H. Given f , g ∈ M (σS (T ), R), we must show that bΦT (f g ) = bΦT (f ) bΦT (g ). Firstly, suppose that f , g ∈ C (σS (T ), R). We have: 3Xa=0 ZσS (T ) f dν (a) u,g(T )v ea = huΦT (f )ΦT (g )vi = huΦT (f g )vi = 3Xa=0 ZσS (T ) Once again exploiting the uniqueness of the signed measures, we conclude that u,g(T )v = g dν (a) dν (a) u,v for every a ∈ {0, 1, 2, 3}. In this way, if f ∈ M (σS (T ), R) and g ∈ C (σS (T ), R), it holds: 3Xa=0 ZσS (T ) 3Xa=0 ZσS (T ) = hf (T )ug (T )vi = f dν (a) u,g(T )v ea = huf (T )g (T )vi = 3Xa=0 ZσS (T ) It follows that f dν (a) u,v = dν (a) f (T )u,v for every a ∈ {0, 1, 2, 3}. Finally, if f , g ∈ M (σS (T ), R), then we have that 3Xa=0 ZσS (T ) 3Xa=0 ZσS (T ) g dν (a) hu bΦT (f g )vi = f g dν (a) f (T )u,v ea = u,v ea = = h bΦT (f )u bΦT (g )vi = hu bΦT (f ) bΦT (g )vi. Since this is true for every u, v ∈ H, we infer that bΦT (f g ) = bΦT (f ) bΦT (g ), as desired. Point (c) is quite evident. Indeed, if f ∈ M (σS (T ), R) and f ≥ 0, then √f ∈ M (σS (T ), R) and f = √f √f . Therefore, it holds: huf (T )ui = hupf (T )pf (T )ui = hpf (T )upf (T )ui ≥ 0 if u ∈ H. Let us prove (d). First, observe that, if u ∈ H and E ∈ B(σS (T )), then (5.12) and (5.13) imply that u,u ea = ZσS (T ) 3Xa=0 ZσS (T ) hu bΦT (χE )ui = χE dν (a) Moreover, bearing in mind (5.5), we have: kf (T )uk2 = h bΦT (f )u bΦT (f )ui = hu bΦT (f )2ui = hu bΦT (f 2 )ui = = ZσS (T ) f 2 dµu ≤ kf k2 ∞ µu (σS (T )) = kf k2 ∞kuk2 , so that kf (T )k ≤ kf k∞ . This completes the proof of (d) and proves (b). We pass to prove (e). Let f ∈ M (σS (T ), R) and let S ∈ B(H) be an operator which commutes with T . We must show that S bΦT (f ) = bΦT (f ) S . Thanks to point (d) of Theorem 5.5, the latter equality holds if f ∈ C (σS (T ), R). In this way, repeating the argument employed to prove that bΦT is a homomorphism of real χE dµu = µu (E ). 42 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI f dν (a) S ∗ u,v ea = S ∗ u,v = ν (a) algebras, one obtains that ν (a) u,Sv if u, v ∈ H and a ∈ {0, 1, 2, 3}. It follows that 3Xa=0 ZσS (T ) huS bΦT (f )vi = hS ∗u bΦT (f )vi = 3Xa=0 ZσS (T ) f dν (a) u,Sv ea = hu bΦT (f ) S vi, = if u, v ∈ H and hence S bΦT (f ) = bΦT (f ) S . It remains to show (f ) (which implies (iii)). If {fn}n∈N is a bounded sequence in M (σS (T ), R), which converges pointwisely to some f ∈ M (σS (T ), R), then we can apply Lebesgue’s dominated convergence theorem to {fn − f }n∈N . In this way, thanks to (5.3), we obtain: n→+∞ ZσS (T ) (fn − f )2 dµu = 0, n→+∞ k(fn(T ) − f (T ))uk2 = lim lim Now we consider the uniqueness issue. Assume that there is another continuous unity–preserving homomorphism Ψ : M (σS (T ), R) −→ B(H) satisfying (i)–(iii). By Theorem 5.5 and point (i), we have that Ψ(f ) = ΦT (f ) = bΦT (f ) if f ∈ C (σS (T ), R). Let u ∈ H. Define the map ν (Ψ) : B(σS (T )) −→ R+ by setting u ν (Ψ) u (E ) := huΨ(χE )ui if E ∈ B(σS (T )). Thanks to the fact that Ψ is a homomorphism satisfying (ii), we have that such a map ν (Ψ) is well–defined. Indeed, if E ∈ B(σS (T )), it holds: u huΨ(χE )ui = huΨ(χE χE )ui = huΨ(χE )Ψ(χE )ui = hΨ(χE )uΨ(χE )ui ∈ R+ . Observe that ν (Ψ) u (∅) = huΨ(0)ui = 0. Moreover, if {En}n∈N is a sequence of pairwise disjoint sets in B(σS (T )), then point (iii) implies that u (Sn∈N En ) = Du (cid:12)(cid:12)(cid:12) limk→+∞ Ψ (cid:16)Pk n=0 χEn (cid:17) u E = ν (Ψ) = limk→+∞ Pk n=0 hu Ψ (χEn ) u i = = Pn∈N hu Ψ (χEn ) u i = Pn∈N ν (Ψ) u (En ). This proves that ν (Ψ) is a positive σ–additive Borel measure on σS (T ). Notice u that that measure is finite as ν (Ψ) u (σS (T )) = huΨ(1σS (T ) )ui = kuk2 . In view of Theorem 2.18 in [30], such a measure is regular. By standard approximation of measurable functions with simple functions and by point (iii), we infer that huΨ(f )ui = ZσS (T ) f dν (Ψ) if f ∈ M (σS (T ), R). u On the other hand, bearing in mind (5.4), if f ∈ C (σS (T ), R), one has f dµu = hu bΦT (f )ui = huΨ(f )ui = ZσS (T ) ZσS (T ) f dν (Ψ) u . The already exploited uniqueness property implies that µu = ν (Ψ) u . In this way, for every f ∈ M (σS (T ), R), it holds: u − ZσS (T ) hu(Ψ(f ) − bΦT (f ))ui = ZσS (T ) f dν (Ψ) f dµu = 0. (cid:3) 43 QUATERNIONIC SLICE FUNCTIONAL CALCULUS Applying Proposition 2.17, we have that Ψ = bΦA . 5.3. Quaternionic L2 -representation of measurable real functions of a self-adjoint operator. In this section, exploiting the functional calculus previ- ously constructed, we establish the key tool in proving the existence of an anti self–adjoint and unitary operator J commuting with a given normal operator T and with its adjoint T ∗ . Indeed, here we prove the existence of such J in the spe- cial situation in which T is self–adjoint. The existence of J will be established by constructing a suitable L2–realisation of the quaternionic Hilbert space H, where T and every measurable real function f (T ) act as multiplicative operators. Let X be a set and let µ be a positive σ–additive measure on X . Denote by L2 (X, H; µ) the quaternionic Hilbert space of the squared integrable functions f : X −→ H with respect to µ. Here L2(X, H; µ) is considered as a quaternionic right module in which the sum of functions is the standard pointwise sum and the right multiplication by quaternions is defined pointwisely. The scalar product is defined as follows: if f , g ∈ L2 (X, H; µ) and f0 , f1 , f2 , f3 , g0 , g1 , g2 , g3 are the real– valued functions on X such that f = P3 a=0 fa ea and g = P3 a=0 gaea , then we set 3Xa,b=0 ZX (f g ) := ZX fa gb dµ ea eb . f g dµ = Here {e0 , e1 , e2 , e3} is a fixed orthonormal basis of H. In particular, the norm kf kL2 of f in L2 (X, H; µ) is given by kf kL2 = (cid:0)RX f 2 dµ(cid:1)1/2 (see e.g. [5]). Theorem 5.7. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a self– adjoint operator. Then the fol lowing assertions hold. (a) There exists an orthogonal decomposition of H into closed subspaces H = Lℓ∈Λ Hℓ with the fol lowing two properties: (i) f (T )(Hℓ) ⊂ Hℓ for every f ∈ M (σS (T ), R) and ℓ ∈ Λ. (ii) For each ℓ ∈ Λ, there is a finite positive regular Borel measure µℓ on σS (T ) and an isometric isomorphism Uℓ : L2 (σS (T ), H; µℓ ) −→ Hℓ such that ℓ f (T )Hℓ Uℓ (cid:1) (g ) = f g (cid:0)U −1 for every f ∈ M (σS (T ), R) and g ∈ L2 (σS (T ), H; µℓ), where f (T )Hℓ denotes the restriction of f (T ) from Hℓ to Hℓ . (b) There exists J ∈ B(H) such that: J ∗ = −J , J ∗J = I and J f (T ) = f (T )J for every f ∈ M (σS (T ), R). Proof. Let u ∈ H. Denote by Hu the vector subspace of H consisting of all vectors x having the following property: there exists a non–empty finite fami- ly {E1 , . . . , EN } of pairwise disjoint Borel subsets of σS (T ) and q1 , . . . , qN ∈ H such that x = PN j=1 χEj (T )uqj . Indicate by Hu the closure of Hu in H. Let µu : B(σS (T )) −→ R+ be the finite positive regular Borel measure of σS (T ) sending E into huχE (T )ui = kχE (T )uk2 (see point (d) of Theorem 5.6). Define the vector u of L2(σS (T ), H; µu ) analogous to Hu : a function f : σS (T ) −→ H be- subspace L2 u if it can be represented as a finite sum PM longs to L2 j=1 χFj pj , where F1 , . . . , FM are pairwise disjoint Borel subsets of σS (T ) and p1 , . . . , pM are quaternions. We subdivide the remainder of this proof into two steps. 44 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (5.14) χEj (T )uqj if f = χEj qj . V (f ) := qj 2µ(Ej ). Step I. First, suppose that Hu = H for some fixed u ∈ H. Let V : L2 u −→ H be the map defined as follows: NXj=1 NXj=1 It is quite easy to verify that V is well–defined; that is, V (f ) depends only on f , and not on its representation PN j=1 χEj qj . Moreover, it is evident that V is right H–linear. Let us prove that V is isometric. Bearing in mind that χEj χEk = χEj if j = k and χEj χEk = 0 for j 6= k , we infer that χEj qj (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2 = ZσS (T ) NXj,k=1 NXj=1 χEj χEk dµu qj qk = L2 = ZσS (T ) NXj=1 NXj=1 χEj qj 2 dµu = Since χEj (T )χEk (T ) = (χEj χEk )(T ), we have that χEj (T )χEk (T ) = χEj (T ) if j = k and χEj (T )χEk (T ) = 0 for j 6= k . It follows that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) χEj (T )uqj (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2 NXj,k=1 (cid:10)χEj (T )uqj (cid:12)(cid:12) χEk (T )uqk (cid:11) = NXj=1 H NXj,k=1 qj (cid:10)u (cid:12)(cid:12) χEj (T )χEk (T )u(cid:11) qk = NXj=1 NXj=1 qj 2 (cid:10)u (cid:12)(cid:12) χEj (T )u(cid:11) = qj (cid:10)u (cid:12)(cid:12) χEj (T )u(cid:11) qj = χEj qj (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) qj 2µu (Ej ) = (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2 NXj=1 NXj=1 . = L2 u is dense in L2(σS (T ), H; µu ). Indeed, if f = P3 Observe that L2 a=0 fa ea is a func- tion in L2 (σS (T ), H; µu), then we can apply Theorem 1.17 of [30] to each component fa of f , obtaining a sequence of L2 u which converges to f in L2(σS (T ), H; µu ). Bear- ing in mind that V is isometric and its range is dense in H by hypothesis, we have that V uniquely extends to an isometric isomorphism U : L2 (σS (T ), H; µu) −→ H. The construction also implies that (cid:0)U −1χE (T )U (cid:1) (χF ) = χE χF Indeed, we have: χE (T )(U (χF )) = χE (T )χF (T )u = (χE χF )(T )u = U (χE χF ). if E , F ∈ B(σS (T )). = = = (5.15) By combining Theorem 1.17 in [30] and point (f ) of Theorem 5.6, we obtain at once that (U −1f (T )U )(χF ) = f χF if f ∈ M (σS (T ), R). Finally, making use again of Theorem 1.17 of [30], we obtain that (U −1f (T )U )g = f g µu –a.e. in σS (T ) QUATERNIONIC SLICE FUNCTIONAL CALCULUS 45 for every f ∈ M (σS (T ), R) and for every g ∈ L2 (σS (T ), H; µu ). Fix  ∈ S and define the operator Y ∈ B(L2 (σS (T ), H; µu )) by setting Y g :=  g . Evidently, −Y Y is the identity of L2 (σS (T ), H; µu ). Furthermore, we have that Y ∗ = −Y . Indeed, it holds:(f Y g ) = RσS (T ) f g µu = − RσS (T ) f g µu = (−Y f g ) for every f , g ∈ L2 (σS (T ), H; µu). Define J ∈ B(H) as follows: J := U Y U −1 . It follows that J J = −I and, it being U isometric, J ∗ = −J . Let now f ∈ M (σS (T ), R). Since f is real–valued, we have that Y f = f = f  and hence (U −1J f (T )U )(g ) = (Y U −1f (T )U )(g ) = f g = f g = = (U −1 f (T )U Y )(g ) = (U −1f (T )J U )(g ) for every g ∈ L2 (σS (T ), H; µu). It follows that J commutes with f (T ) for every f ∈ M (σS (T ), R). This proves the theorem if Hu = H for some u ∈ H. Step II. Let us consider the case in which Hu 6= H for every u ∈ H. Since H0 = {0}, we have that H 6= {0}. Choose u ∈ H \ {0}. Observe that u = Iu = 1σS (T ) (T )u = χσS (T ) (T )u ∈ Hu . It follows that Hu 6= {0} and hence Hu 6= {0}. Fix f ∈ M (σS (T ), R). Let us show that f (T )(Hu) ⊂ Hu . (5.16) j=1 χEj (T )uqj be an element of Hu and let {sn = POn Let x = PN k=1 χFnk rnk }n∈N be a bounded sequence of real simple functions on σS (T ) converging pointwisely to f . By point (f ) of Theorem 5.6, it follows that NXj=1 NXj=1 NXj=1 (f χEj )(T )uqj = OnXk=0 NXj=1 NXj=1 χEj (T )χFnk (T )uqj rnk = χEj (T )f (T )uqj = lim n→+∞ OnXk=0 NXj=1 = lim χEj ∩Fnk (T )uqj rnk n→+∞ j=1 POn Since each PN k=0 χEj ∩Fnk (T )uqj rnk belongs to Hu , f (T )(x) belongs to Hu . Now inclusion (5.16) follows by the continuity of f (T ). Combining (5.16) with the u ) ⊂ H⊥ fact that f (T ) is self–adjoint, we infer that f (T )(H⊥ u . In this way, applying Step I to the Hu ’s, we see at once that, in order to complete the proof, it suffices to show the existence of a subset {uℓ}ℓ∈Λ of H \ {0} such that H = Lℓ∈Λ Huℓ and Huℓ ⊥ Huℓ′ for every ℓ, ℓ′ ∈ Λ with ℓ 6= ℓ′ . However, this can be done by a standard argument involving Zorn’s lemma. (cid:3) Remark 5.8. Let us focus on the proof of (b). Consider a function h ∈ M (σS (T ), R) with h(t)2 = 1 for every t ∈ σS (T ) and re–define (Y g )(x) := h(x)g (x) for every g ∈ L2 (σS (T ), H; µν ). It easily arises that Y Y = −I and Y ∗ = −Y . Passing to Hu , and finally to the whole H via Zorn’s lemma, one obtains other definitions of J verifying all requirements in (b). Therefore, J turns out to be highly undetermined by T . f (T )χEj (T )uqj = (χEj f )(T )uqj = f (T )(x) = = 5.4. The commuting operator J . We are eventually in a position to prove the existence of an anti self–adjoint and unitary operator J , commuting with any given normal operator T and with its adjoint T ∗ . 46 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI Theorem 5.9. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a normal operator. Then there exists an anti self–adjoint and unitary operator J ∈ B(H) such that (5.17) and T J = J T , T ∗J = J T ∗ (5.18) . T = (T + T ∗ ) 1 1 + J T − T ∗ 2 2 In particular, J is uniquely determined by T on Ker (T − T ∗)⊥ and the operators 2 , T − T ∗ 1 (T + T ∗ ) 1 2 and J commute mutual ly. Proof. Since T is normal, it is easy to verify that both the operators T and T ∗ preserve the decomposition H = Ker (T − T ∗) ⊕ Ker (T − T ∗)⊥ . Furthermore, T and T ∗ coincide with the self–adjoint operator (T + T ∗ ) 1 2 on Ker (T − T ∗). Denote by T ′ ∈ B(Ker (T − T ∗)) the self–adjoint operator of Ker (T − T ∗) obtained restricting (T + T ∗ ) 1 2 . Let J0 be the operator obtained applying Theorem 5.1 to T . Since J0 preserves the above–mentioned decomposition of H, it is anti self–adjoint, unitary on Ker (T − T ∗ )⊥ and satisfies (5.1), in order to complete the proof, it suffices to find an anti self–adjoint unitary operator J ′ 0 in B(Ker (T − T ∗)) commuting with T ′ . Indeed, if one has such an operator J ′ 0 , then J := J ′ 0 ⊕ J0 has the desired properties. On the other hand, J ′ 0 is an operator given by point (b) of Theorem 5.7, applied to T ′ with f equal to the inclusion map σS (T ′ ) ֒→ R. (cid:3) Remark 5.10. If K er(T − T ∗) is not trivial, differently from J0 , the operator J cannot commute with every operator commuting with T − T ∗. Otherwise, J would preserve Ker (T − T ∗ ) and the restriction J ′ of J from Ker (T − T ∗ ) to itself would be an anti self–adjoint and unitary operator in B(Ker (T − T ∗)), commuting with every element of B(K er(T − T ∗)). As one can easily prove, there are no operators with these properties in B(K er(T − T ∗)). Finally, we stress that, as a consequence of Remark 5.8, J ′ 0 is highly undetermined from T and that indeterminacy affect the definition of J itself. To go on, we need a relevant notion we shall extensively use in the remainder of the paper. Recall that, given a subset K of C, we define the circularization ΩK of K (in H ) by setting ΩK := {α + β ∈ H α, β ∈ R, α + iβ ∈ K,  ∈ S}. (5.19) In the following, we denote by σ(B ) and ρ(B ) the standard spectrum and resol- vent set of a bounded operator B of a complex Hilbert space, respectively. Proposition 5.11. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a nor- mal operator, let J ∈ B(H) be an anti self–adjoint and unitary operator satisfying (5.17), let ı ∈ S and let HJ ı ± be the complex subspaces of H associated with J and ı (see Definition 3.6). Then we have that + ) ⊂ HJ ı + and T ∗(HJ ı + ) ⊂ HJ ı (a) T (HJ ı + . denote the Cı–complex operators in B(HJ ı and T ∗ HJ ı Moreover, if T HJ ı + ) obtained + + restricting respectively T and T ∗ to HJ ı + , then it holds: )∗ = T ∗ HJ ı (b) (T HJ ı . + + QUATERNIONIC SLICE FUNCTIONAL CALCULUS 47 ) = σS (T ) ∩ Ci . Here σ(T HJ ı ) ∪ σ(T HJ ı (c) σ(T HJ ı ) is considered as a subset + + + of Cı via the natural identification of C with Cı induced by the real vector isomorphism C ∋ α + iβ 7→ α + ıβ ∈ Cı . (d) σS (T ) = ΩK , where K := σ(T HJ ı ). + An analogous statement holds for HJ ı − . Proof. Recall that HJ ı + = {u ∈ H J u = uı}. Let u ∈ HJ ı + . Since J commute with both T and T ∗ , we have that J (T u) = T J u = (T u)ı and J (T ∗u) = T ∗J u = (T ∗u)ı; that is, T u, T ∗u ∈ HJ ı + . Bearing in mind Lemma 3.9, it follows immediately that and T ∗ HJ ı belongs to B(HJ ı T+ := T HJ ı + ), and (b) holds true. + + Let us pass to prove the following equality ), ) ∩ ρ(T HJ ı ρS (T ) ∩ Cı = ρ(T HJ ı + + which is equivalent to (c). Let q ∈ ρS (T ) ∩ Cı . Since J commutes with ∆q (T ) and ∆q (T ) admits inverse in B(H), it follows that J commutes also with ∆q (T )−1 and hence the restriction ∆q (T )+ of ∆q (T ) from HJ ı + into itself is a well–defined If I+ denotes the identity operator on HJ ı invertible operator in B(HJ ı + , then + ). ∆q (T )+ = (T+ − I+ q)(T+ − I+ q) in B(HJ ı + ). Since ∆q (T )+ is invertible and the operators ∆q (T )+ , T+ − I+ q and T+ − I+ q commute with each other, we infer that both T+ − I+ q and T+ − I+ q are invertible in B(HJ ı + ). In other words, q belongs to ) ∩ ρ(T HJ ı ρ(T HJ ı ) ∩ ρ(T HJ ı ). This proves the inclusion ρS (T ) ∩ Cı ⊂ ρ(T HJ ı ). + + + + ) ∩ ρ(T HJ ı Let us prove the converse inclusion. Fix q ∈ ρ(T HJ ı ). The operators + + + − I+ q)−1 and (T HJ ı + − I+ q)−1 exist in B(HJ ı (T HJ ı + ). Their product is the inverse of ∆q (T )+ . Points (a) and (f ) of Proposition 3.11 immediately imply that ∆q (T ) is invertible in B(H), so that q ∈ ρS (T ). Point (c) is proved. Finally, (d) is a straightforward consequence of (c) and of definition (5.19). (cid:3) Denote by R[X, Y ] the ring of real polynomials in the indeterminates X and Y . As in the case of the ring R[X ] considered in Section 2.3, we write the polynomials In this way, given Q ∈ R[X, Y ], we in R[X, Y ] with coefficients on the right. have that Q(X, Y ) = P(h,k)∈Q X hY k rhk for some non–empty subset Q of N × N, If A, B ∈ B(H), then we define the operator where the rhk ’s are real numbers. Q(A, B ) ∈ B(H) by setting Q(A, B ) = X(h,k)∈Q AhB k rhk . (5.20) Corollary 5.12. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a normal operator, let J ∈ B(H) be an anti self–adjoint and unitary operator satisfying (5.17) and (5.18), let ı ∈ S and let Q ∈ R[X, Y ] be a polynomial such that Q(α, β ) = 0 if α + ıβ ∈ σS (T ). Define A := (T + T ∗) 1 2 and B := T − T ∗ 1 2 . Then Q(A, B ) is the nul l operator in B(H). Proof. Firstly, we observe that the operator F := Q(A, B ) is self–adjoint and com- mutes with J . By applying the spectral radius formula (see (4.3)) and point (c) of Proposition 5.11 to F , we obtain that kF k = sup nq ∈ R+ (cid:12)(cid:12)(cid:12) q ∈ σ(F HJ ı ) o , + 48 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI + ) obtained restricting F to HJ ı denotes the operator in B(HJ ı where F HJ ı + . Points + of A to HJ ı (a) and (b) of Proposition 5.11 imply that the restrictions AHJ ı + and + + define operators in B(HJ ı of B to HJ ı + ). Moreover, we have that AHJ ı B HJ ı = )∗ (cid:1) 1 = (cid:0)T HJ ı )∗ (cid:1) 1 (cid:0)T HJ ı + + + − (T HJ ı 2 and B HJ ı + (T HJ ı 2ı . The latter equality follows + + + + from the definition of HJ ı + (see Definition 3.6) and from the fact that B is equal to −J (T − T ∗) 1 2 . In this way, if f : σ(F HJ ı ) −→ R denotes the polynomial function + sending α + ıβ into Q(α, β ), then the continuous functional calculus theorem for , B HJ ı = f (AHJ ı complex normal operators (see [30, 28, 25]) ensures that F HJ ı ) + + + )) = {0}. It ) = f (σ(T HJ ı and σ(F HJ ı )). By hypothesis, we know that f (σ(T HJ ı + + + ) = {0} and hence kF k = 0; that is, F = 0, as desired. follows that σ(F HJ ı (cid:3) + In the following define C+ := {q ∈ H q = α + ıβ , α, β ∈ R , β ≥ 0} and, similarly, ı C− := {q ∈ H q = α + ıβ , α, β ∈ R , β ≤ 0}. We have: ı Corollary 5.13. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a normal operator, let J ∈ B(H) be an anti self–adjoint and unitary operator satisfying (5.17) and (5.18), and let ı ∈ S. Then it holds: ) = σS (T ) ∩ C+ ı , σ(T H J ı σ(T H J ı (5.21) − + ) = σS (T ) ∩ C− ı and hence ). (5.22) ) = σ(T H J ı σ(TH J ı − + 2 and B := T − T ∗ 1 Proof. Let A := (T + T ∗ ) 1 2 , so that T = A + J B . Denote by T± , A± and B± the operators in B(HJ ı ± ) obtained restricting T , A and B to HJ ı ± , respectively. Observe that T± = A± ± B± ı and B± ≥ 0, because B is so. Therefore, the spectral theorem for self–adjoint operators in complex Hilbert spaces [31, 25] implies that σ(B± ) ⊂ R+ . Let f± : σ(T± ) −→ Cı be the continuous Cı– complex function sending α + ıβ into β . It follows that B± = f± (±T±), where the right–hand side is defined in the standard sense by means of continuous functional calculus for normal operators in complex Hilbert spaces [31, 25]. Moreover, using again the spectral theory in complex Hilbert spaces, we infer that f± (σ(±T± )) = σ(f± (±T±)). In particular, we have: ±f±(σ(T± )) = f± (σ(±T± )) = σ(f± (±T± )) = σ(B± ) and hence ±f±(σ(T± )) ⊂ R+ . The latter inclusion is equivalent to the following ) ⊂ C+ ) ⊂ C− two: σ(T H J ı ı and σ(T H J ı ı . By combining these inclusions with − + Proposition 5.11(c), we immediately obtain equalities (5.21) and (5.22). (cid:3) 5.5. Extension of J to a full left scalar multiplication of H. We conclude this section by proving how it is possible to extend the operator J satisfying Theorem 5.9 to a full left scalar multiplication H ∋ q 7→ Lq of H in such a way that each Lq commutes with (T + T ∗) 1 2 and T − T ∗ 1 2 . Theorem 5.14. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a normal operator and let ı ∈ S. Suppose that T decomposes as fol lows: T = A + J B , where A and B are self–adjoint operators in B(H) and J ∈ B(H) is an anti self– adjoint and unitary operator commuting with A and B (for example, one can QUATERNIONIC SLICE FUNCTIONAL CALCULUS 49 z hz S z ′iq = S (z ′)q = S (z ′ q) = (SLq )(z ′ ). choose J as in the statement of Theorem 5.9 and define A := (T + T ∗) 1 2 and B := T − T ∗ 1 2 ). Then there exists a left scalar multiplication H ∋ q 7→ Lq of H such that Lı = J and, for every q ∈ H, LqA = ALq and LqB = BLq . Proof. Let N be a Hilbert basis of H and let H ∋ q 7→ Lq = Pz∈N z qhz ·i be the left scalar multiplication of H induced by N . Observe that, if S ∈ B(H) is an operator such that hz S z ′i ∈ R for every z , z ′ ∈ N , then Lq S = SLq . Indeed, if z ′ ∈ N , one has: (Lq S )(z ′ ) = Xz∈N z qhz S z ′i = Xz∈N Thanks to this fact, it is sufficient to find a Hilbert basis N of H such that the induced left scalar multiplication H ∋ q 7→ Lq has the following properties: Lı = J and hz Az ′i, hz B z ′ i ∈ R for every z , z ′ ∈ N . Actually, it suffices to construct a Hilbert basis N of the Cı–Hilbert space HJ ı + satisfying hz A+ z ′i, hz B+ z ′ i ∈ R if z , z ′ ∈ N , where A+ and B+ denote the operators in B(HJ ı + ) obtained restricting A and B to HJ ı + , respectively. Indeed, by point (f ) of Proposition 3.8, N is also a Hilbert basis of H and J = Pz∈N z ıhz ·i. Let T+ ∈ B(HJ ı + ) be the normal operator obtained restricting T to HJ ı + . Observe that A+ and B+ are self–adjoint commuting operators and T+ = A+ + ıB+ in B(HJ ı + ). From the Spectral Representation Theorem (see Chap.X, sec.5 of [10]), there exists an orthogonal decomposition Lℓ∈Λ Hℓ of HJ ı + into closed subspaces and, for every ℓ ∈ Λ, a positive σ–additive Borel measure µℓ on σ(T+ ) and an isometric isomorphism Uℓ from L2 ℓ := L2 (σ(T+ ), C; µℓ ) to Hℓ such that Hℓ is invariant under T+ and T ∗ + (and thus under A+ and B+ ), and (5.23) (UℓT (ℓ) + U −1 ℓ ℓ , where T (ℓ) for every fℓ ∈ L2 + is the restriction of T+ from Hℓ into itself. It follows that )(fℓ )(x + ıy ) = (x + ıy )fℓ (x + ıy ) ℓ A(ℓ) (U −1 + Uℓ )(fℓ )(x + ıy ) = xfℓ (x + ıy ) and ℓ B (ℓ) (U −1 + Uℓ)(fℓ )(x + ıy ) = yfℓ (x + ıy ) + and B (ℓ) ℓ , where A(ℓ) for every ℓ ∈ Λ and fℓ ∈ L2 + denotes the restrictions of A+ and B+ from Hℓ into itself. ℓ B (ℓ) ℓ A(ℓ) + Uℓ and B ′ := U −1 Fix ℓ ∈ Λ and define the operators A′ := U −1 + Uℓ ℓ ). Let Pℓ be the set of all orthonormal subsets F of L2 in B(L2 ℓ such that each function in F is real–valued. Equip Pℓ with the partial ordering induced by the inclusion. Since Pℓ is non–empty and inductive, Zorn’s lemma ensures the existence of a maximal element Mℓ of Pℓ . Let us prove that Mℓ is an Hilbert basis of L2 ℓ . ℓ such that g ⊥ Mℓ in L2 Let g ∈ L2 ℓ ; that is, it holds: Re(g )f dµℓ − ı Zσ(T+ ) gf dµℓ = Zσ(T+ ) 0 = Zσ(T+ ) for each f ∈ Mℓ . It follows that Re(g ) ⊥ Mℓ and Im(g ) ⊥ Mℓ . By the maximality of Mℓ , we infer that Re(g ) = 0 = Im(g ) and hence g = 0, as desired. Observe that, Im(g )f dµℓ 50 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI and xf (x + ıy )f ′(x + ıy ) dµℓ ∈ R for every f , f ′ ∈ L2 ℓ , we have: f (A′ f ′ ) dµℓ = Zσ(T+ ) Zσ(T+ ) (5.24) Zσ(T+ ) f (B ′ f ′ ) dµℓ = Zσ(T+ ) yf (x + ıy )f ′(x + ıy ) dµℓ ∈ R. (5.25) Define Nℓ := U (Mℓ) for every ℓ ∈ Λ and N := Sℓ∈Λ Nℓ . The reader observes that, given any ℓ ∈ Λ, (5.24) and (5.25) are equivalent to (5.23) with z , z ′ ∈ Nℓ . On the other hand, by construction, if ℓ, ℓ′ ∈ Λ with ℓ 6= ℓ′ , z ∈ Nℓ and z ′ ∈ Nℓ′ , then hz A′ z ′i = 0 = hz B ′ z ′i. This implies (5.23) and completes the proof. (cid:3) As a consequence, we obtain: Corollary 5.15. Let H be a quaternionic Hilbert space and let T ∈ B(H) be a normal operator. Then there exists a unitary operator U : H −→ H such that U T U ∗ = T ∗ . Proof. Decompose T as in Theorem 5.9: T = (T + T ∗) 1 2 + J T − T ∗ 1 2 . Define 2 and B := T − T ∗ 1 A := (T + T ∗) 1 2 . Choose ı ∈ S. By Theorem 5.14, there exists a left scalar multiplication H ∋ q 7→ Lq of H such that Lı = J and, for every q ∈ H, LqA = ALq and LqB = BLq . Let p ∈ S such that pıp = −ı. Define U := Lp . It holds: U T U ∗ = U (A + J B )U ∗ = LpALp + LpJ LpLpBLp = = A + LpıpB = A + L−ıB = A − J B = T ∗ . The proof is complete. (cid:3) We remind the reader that the preceding result is false in the complex setting: if H is a complex Hilbert space and T is the normal operator on H obtained multiplying the identity operator by i, then it is immediate to verify that there does not exist any unitary operator U on H such that U T U ∗ = T ∗ . 6. Relevant C ∗ -algebras of slice functions Throughout this section, K wil l denote a non–empty subset of C, invariant under complex conjugation. 6.1. Slice functions. We recall basic definitions and results concerning slice func- tions, taking [20] and [19] as references. Consider the complexification HC := H ⊗R C of H. We represent the elements w of HC by setting w = q + ip with q , p ∈ H, where i2 = −1. The product of HC is given by the following equality: (q + ip)(q ′ + ip′ ) = qq ′ − pp′ + i(qp′ + pq ′ ). (6.1) Let us introduce the notion of stem function on K. Definition 6.1. Let F : K −→ HC be a function and let F1 , F2 : K −→ H be the components of F = F1 + iF2 . We say that F is a stem function on K if (F1 , F2 ) forms an even–odd pair with respect to the imaginary part of z ∈ C; that is, F1 ( ¯z ) = F1 (z ) and F2 (z ) = −F2 (z ) for every z ∈ K. If F1 and F2 are continuous, then F is called continuous. QUATERNIONIC SLICE FUNCTIONAL CALCULUS 51 Remark 6.2. Given w = q + ip ∈ HC , denote by w the element q − ip of HC . It is immediate to verify that a function F : K −→ HC is a stem function if and only if it is complex–intrinsic; that is, F (z ) = F (z ) for every z ∈ K such that z ∈ K. As discussed in Remark 2.1(3), the quaternions have the following two properties, which describe their “slice” nature: • H = S∈S C , • C ∩ Cκ = R for every , κ ∈ S with  6= ±κ. In this way, every q ∈ H can be written as follows: q = α + β for some α, β ∈ R and  ∈ S. If q ∈ R, then α = q , β = 0 and  is an arbitrary element of S. If q ∈ H \ R, then q belongs to a unique “slice complex plane” C and hence q has only two expressions: q = α + β = α + (−)(−β ), (6.2) where α = Re(q), β = ±Im(q) and  = ±Im(q)/Im(q). Recall the definition of circularization ΩK of K given in (5.19): ΩK := {α + β ∈ H α, β ∈ R, α + iβ ∈ K,  ∈ S}. We are now in position to define slice functions. Definition 6.3. Each stem function F = F1 + iF2 : K −→ HC on K induces a (left) slice function I (F ) : ΩK −→ H on ΩK as follows: if q = α + β ∈ ΩK for some α, β ∈ R and  ∈ S, then if z = α + iβ ∈ K. I (F )(q) := F1 (z ) + F2 (z ) If F is continuous, then f is called continuous slice functions on ΩK . We denote by S (ΩK , H) the set of all continuous slice functions on ΩK . The notion of slice function I (F ) just given is well–posed. Indeed, if q ∈ R, then F2 (z ) = 0 and hence I (F )(q) = F1 (q), independently from the choice of  in S. Moreover, if q belongs to H \ R and has expressions (6.2), then it holds: I (F )(α + (−)(−β )) = F1 (z ) + (−)F2 (z ) = F1 (z ) + F2 (z ) = I (F )(α + β ), where z = α + iβ . Remark 6.4. (1) Every slice function is induced by a unique stem function. Indeed, if f is induced by some stem function F1 + iF2 ,  is a fixed element of S, z = α + iβ is an arbitrary point of K and q := α + β ∈ ΩK , then F1 (z ) = 1 2 (f (q) + f (q)) and F2 (z ) = − 1 2 (f (q) − f (q)). It follows that f satisfies the following representation formula : f (α + ıβ ) = 1 2 (f (q) + f (q)) − ı 1 2 (f (q) − f (q)) for every ı ∈ S (see Subsection 3.3 of [20] for details). As an immediate consequence, we infer that, if two slice functions on ΩK coincide on ΩK ∩ C for some  ∈ S, then they coincide on the whole ΩK . (2) All continuous slice functions f : ΩK −→ H are continuous in the usual topological sense (see Proposition 7(1) of [20]); that is, S (ΩK , H) ⊂ C (ΩK , H). (3) Given n ∈ N and a ∈ H, the function C ∋ z 7→ z na = Re(z n )a + iIm(z n )a ∈ HC is a stem function inducing the slice function H ∋ q 7→ qn a ∈ H. It follows that all polynomial functions Pd h=0 qh ah on H and all convergent power series Ph∈N qhah on some ball of H are slice functions. 52 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (6.3) The pointwise product of slice functions is not necessarily a slice function. On the contrary, as it is immediate to see, if F = F1 + iF2 and G = G1 + iG2 are stem functions, then their product F G = (F1G1 − F2G2 ) + i(F1G2 + F2G1 ) is again a stem function. This suggests a natural way to define the product of slice functions. Definition 6.5. Given two slice functions f = I (F ) and g = I (G) on ΩK , we define the slice product f · g of f and g as the slice function I (F G) on ΩK . Remark 6.6. The slice product defines an operation on S (ΩK , H). Indeed, for- mula (6.3) ensures that the product of two continuous stem functions is again a continuous stem function. Let us define the H–intrinsic and C –slice functions. Definition 6.7. Let F = F1 + iF2 be a stem function on K and let f : ΩK −→ H be the slice function induced by F . We say that f is a H–intrinsic slice function if f (q) = f (q) for every q ∈ ΩK . This is equivalent to require that F1 and F2 are real–valued (see Lemma 6.8 below). For this reason, we denote by SR (ΩK , H) the subset of S (ΩK , H) consisting of all continuous H–intrinsic slice functions on ΩK . Given  ∈ S, we say that f is a C –slice function if F1 and F2 are C –valued. We denote by SC (ΩK , H) the subset of S (ΩK , H) consisting of all continuous C –slice functions on ΩK . We underline that, in Definition 10 of [20], H–intrinsic slice functions are called real slice functions. Evidently, SR (ΩK , H) ⊂ SC (ΩK , H) for every  ∈ S. H–intrinsic slice functions have nice characterizations. Lemma 6.8. Given a slice function f = I (F1 + iF2 ) : ΩK −→ H, the fol lowing three conditions are equivalent: (i) f is H–intrinsic. (ii) F1 and F2 are real–valued. (iii) f (ΩK ∩ C ) ⊂ C for every  ∈ S. Proof. The equivalence of conditions (ii) and (iii) is proved in Proposition 10 of [20]. In this way, in order to complete the proof, it suffices to show that (i) is equivalent to (ii). Let F = F1 + iF2 be the stem function inducing f , let  ∈ S, let q = α + β ∈ ΩK for some α, β ∈ R and let z := α + iβ ∈ K. If (ii) holds, then we have: f (q) = F1 (z ) + F2 (z ) = F1 (z ) − F2 (z ) = f (q). This proves implication (ii) =⇒ (i). Assume now that f is H–intrinsic. Let F 0 2 (z ) ∈ R, ı ∈ S and p ∈ ΩK such that F2 (z ) = F 0 2 (z ) + ıF 1 2 (z ), F 1 2 (z ) and p = α + ıβ . Since F1 (z ) − ıF2 (z ) = f (p) = f (p) = F1 (z ) − F2 (z )ı, we infer that F1 (z ) − F1 (z ) = ıF2 (z ) − F2 (z )ı = −2F 1 2 (z ) ∈ Im(H) ∩ R = {0}. It follows that F1 (z ) ∈ R, F 1 2 (z ) = 0 and hence F2 (z ) ∈ R as well. (cid:3) QUATERNIONIC SLICE FUNCTIONAL CALCULUS 53 Now we introduce the notion of circular slice function. Definition 6.9. A slice function I (F1 + iF2 ) on ΩK is called circular slice function if F2 vanishes on the whole K. We denote by Sc (ΩK , H) the subset of S (ΩK , H) consisting of all continuous circular slice functions on ΩK . Lemma 6.10. Given a slice function f : ΩK −→ H, the fol lowing two conditions are equivalent: (i) f is a circular slice function. (ii) f (q) = f (q) for every q ∈ ΩK . Proof. The implication (i) =⇒ (ii) is evident. Let F1 + iF2 : K −→ H be the stem function inducing f . Suppose that (ii) holds. Then, for every α, β ∈ R and ı ∈ S with q := α + ıβ ∈ ΩK , we have that F1 (α, β ) − ıF2 (α, β ) = f (q) = f (q) = F1 (α, β ) + ıF2 (α, β ) or, equivalently, ıF2 (α, β ) = 0. It follows immediately that F2 is null and hence f ∈ Sc (ΩK , H). (cid:3) Given a basis of H, the functions in S (ΩK , H), and hence in SC (ΩK , H) and in Sc (ΩK , H), can be expressed in terms of functions in SR (ΩK , H) as follows. Lemma 6.11. Let {1, , κ, δ} be a basis of H. Then the map (cid:0)SR (ΩK , H)(cid:1)4 ∋ (f0 , f1 , f2 , f3) 7→ f0 + f1  + f2κ + f3δ ∈ S (ΩK , H) is bijective. In particular, it fol lows that, given any f ∈ S (ΩK , H), there exist, and are unique, f0 , f1 , f2 , f3 ∈ SR (ΩK , H) such that (6.4) f = f0 + f1  + f2κ + f3 δ. Moreover, it holds: (a) If  ∈ S, then f belongs to SC (ΩK , H) if and only if f2 = f3 = 0. (b) f belongs to Sc (ΩK , H) if and only if f0 , f1 , f2 and f3 are real–valued. 2 )}3 1 + iF ℓ Proof. Let {fℓ = I (F ℓ ℓ=0 be H–intrinsic slice functions in SR (ΩK , H) and let f : ΩK −→ H be the continuous slice function f := f0 + f1  + f2κ + f3 δ . Define the function F : K −→ HC by setting F := (F 0 1 + F 1 1  + F 2 1 κ + F 3 1 δ) + i(F 0 2 + F 1 2  + F 2 2 κ + F 3 2 δ). It is immediate to verify that F is the (unique) stem function inducing the slice function f0 + f1 + f2κ + f3δ = f . Thanks to the uniqueness of F , it follows that the map mentioned in the statement is injective. Let us prove that it is also surjective. Let g be a slice function in S (ΩK , H) induced by the stem function G = G1 + iG2 2}3 1 , Gℓ and let {Gℓ ℓ=0 be the real–valued continuous functions on K such that Gm = G0 m + G1 m  + G2 mκ + G3 if m ∈ {1, 2}. m δ ℓ=0 on K and the H–intrinsic slice functions {gℓ}3 Define the stem functions {Gℓ }3 ℓ=0 in SR (ΩK , H) by setting Gℓ := Gℓ 1 + iGℓ 2 and gℓ := I (Gℓ ). Since I (G1 ) = g1, I (G2 κ) = g2κ, I (G3 δ) = g3δ and G = G0 + G1  + G2κ + G3 δ , we infer that g = g0 + g1  + g2κ + g3δ , which proves the desired surjectivity. Point (a) and the fact that f0 , f1 , f2 and f3 are real–valued when f ∈ Sc (ΩK , H) are easy consequences of the above argument. Finally, suppose that f0 , f1 , f2 and f3 are real–valued. By Lemma 6.8, we know that fℓ(q ) = fℓ (q) for every q ∈ ΩK 54 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI and ℓ ∈ {0, 1, 2, 3}. It follows that f (q) = f (q) for every q ∈ ΩK . Now Lemma 6.10 ensures that f ∈ Sc (ΩK , H). (cid:3) Remark 6.12. The reader observes that, if f , g ∈ SR (ΩK , H), then f · g = g · f ∈ SR (ΩK , H). Indeed, if f is induced by F = F1 + iF2 and g by G = G1 + iG2 with F1 , F2 , G1 , G2 real–valued, then formula (6.3) ensures that also the components of F G are real–valued and F G = GF . It is worth observing that, in this case, f · g coincides with the pointwise product f g (see Remark 7 of [20] for details). By similar considerations, we see that, if f , g ∈ Sc (ΩK , H), then f · g ∈ Sc (ΩK , H) and f · g = f g . However, in general, f · g is different from g · f . Finally, if f , g ∈ SC (ΩK , H), then f · g ∈ SC (ΩK , H) and f · g = g · f . The first assertion follows again from formula (6.3). The second is an immediate consequence of the representation formula for slice functions and of the fact that f · g and g · f coincide on ΩK ∩ C (see Remark 6.4(1)). Moreover, it is immediate to verify that f · g coincides with f g on ΩK ∩ C , but they can be different outside C . 6.2. A quaternionic two-sided Banach unital C ∗ -algebra structure on HC . Let us introduce a structure of quaternionic two–sided Banach unital C ∗ –algebra on HC (see Section 3.2 for the definition). We need that structure later on. Identify H with the subset {q + ip ∈ HC p = 0} of HC and define the quaternionic left and right multiplications (q , w) 7→ qw and (w, q) 7→ wq on HC via formula (6.1). It is immediate to verify that the set HC , equipped with the usual sum, with these left and right scalar multiplications and with product (6.1), is a quaternionic two– sided algebra with unity 1. Define an involution w 7→ w∗ on HC by setting w∗ := q − ip if w = q + ip. (6.5) Observe that, given w = q + ip and y = q ′ + ip′ in HC , it holds: (wy )∗ = (qq ′ − pp′ ) − i (qp′ + pq ′ ) = = (q ′ q − p′ p) − i(p′ q + q ′ p) = y∗w∗ . (6.6) It follows that the involution w 7→ w∗ is a ∗ –involution on HC and hence, equipping HC with such a ∗–involution, we obtain a quaternionic two–sided unital ∗–algebra. Consider the function k · kHC : HC −→ R+ defined by setting kwkHC := (cid:0)q 2 + p2 + 2Im(p q)(cid:1)1/2 (6.7) if w = q + ip. It is worth observing that kw∗kHC = kwkHC , since Im(−pq) = Im(pq), and (q 2 + p2 )1/2 ≤ kwkHC ≤ q + p (6.8) if w = q + ip. Proposition 6.13. The fol lowing assertions hold. (a) Given w = q + ip ∈ HC , we have that kwkHC = sup∈S q + p. Moreover, if Im(p q) = ıβ for some ı ∈ S and β ≥ 0, then kwkHC = q − ıp. (b) The function k · kHC is a quaternionic Banach C ∗–norm on HC . In this way, the set HC , equipped with the usual sum, with the product defined in (6.1), with the quaternionic left and right scalar multiplications induced by such a product, with the ∗–involution defined in (6.5) and with the norm defined in (6.7), is a quaternionic two–sided Banach C ∗ –algebra with unity 1. Proof. (a) Let φ : HC −→ R+ be the function defined by setting φ(w) := kwk2 HC . Fix w = q + ip ∈ HC . We must prove that φ(w) = sup∈S q + p2 . If q = 0, then QUATERNIONIC SLICE FUNCTIONAL CALCULUS 55 this is evident. Suppose that q 6= 0. Let α, β ∈ R and ı ∈ S such that β ≥ 0 and p q = α + ıβ . Observe that φ(w) = q − ıp2 . Indeed, β = Im(p q) and it holds: q − ıp2 = q −2 (q − ıp) ¯q2 = q −2 (cid:0)(q 2 + β )2 + α2 (cid:1) = = q −2 (cid:0)q 4 + q 2 p2 + 2β q 2 (cid:1) = q 2 + p2 + 2β = φ(w). Moreover, for every  ∈ S, we have: q + p2 = q 2 + p2 + 2 Re(p q) = q 2 + p2 + 2 Re((α + ıβ )) = = q 2 + p2 + 2 Re(ı)β ≤ q 2 + p2 + 2β = φ(w). It follows that φ(w) = sup∈S q + p2 = q − ıp2 , as desired. (b) Firstly, observe that formula (6.7) implies at once that kwkHC = 0 if and only if w = 0. The triangular inequality for k · kHC is an immediate consequence of point (a). It is evident that k1kHC = 1. It is also easy to see that kw∗wkHC = kwk2 HC . Indeed, we have that ww∗ = q 2 + p2 − 2iIm(p q) and hence (a) ensures that kww∗ kHC = sup∈S (cid:12)(cid:12)q 2 + p2 + 2Im(p q)(cid:12)(cid:12) = = q 2 + p2 + 2Im(p q) = kwk2 HC , and therefore also kw∗wkHC = kwk2 HC . Choose q ′ ∈ H. It hold: φ(wq ′ ) = qq ′ 2 + pq ′ 2 + 2Im(pq ′ q ′ q) = = q 2 q ′ 2 + p2 q ′ 2 + 2Im(p q) q ′ 2 = φ(w)q ′ 2 and φ(q ′w) = q ′ q 2 + q ′p2 + 2Im(q ′ p q q ′ ) = = q ′ 2 q 2 + q ′ 2 p2 + 2q ′ Im(p q)q ′ = = q ′ 2 q 2 + q ′ 2 p2 + 2q ′ 2 Im(p q) = q ′ 2φ(w). This proves that kwq ′ kHC = q ′ kwkHC = kq ′wkHC . In particular, it follows that k · kHC is a norm on HC ≃ R4 × R4 in the usual sense, and hence it is equivalent to the euclidean one. This ensures that the metric on HC induced by such a norm is complete. Let p′ ∈ H and let y := q ′ + ip′ . In order to complete the proof, it remains to show that kwykHC ≤ kwkHC kykHC or, equivalently, φ(wy ) ≤ φ(w)φ(y ). Bearing in mind that φ(w) = q 2 + p2 + 2Im(p q), φ(y ) = q ′ 2 + p′ 2 + 2Im(p′ q ′ ) and φ(wy ) = φ((qq ′ − pp′ ) + i(qp′ + pq ′ )) = = qq ′ − pp′ 2 + qp′ + pq ′ 2 + 2(cid:12)(cid:12)Im((qp′ + pq ′ )(q ′ q − p′ p))(cid:12)(cid:12), an explicit computation gives that φ(w)φ(y ) − φ(wy ) = 2(a − b) + 4Im(p q)Im(p′ q ′ ) − 4 Re(q Im(p′ q ′ ) p), where a := q 2 Im(p′ q ′ ) + p2 Im(p′ q ′ ) + q ′ 2 Im(p q) + p′ 2 Im(p q), b := (cid:12)(cid:12)q Im(p′ q ′ )q + p Im(p′ q ′ )p + q ′ 2 Im(p q) + p′ 2 Im(p q)(cid:12)(cid:12) The triangular inequality implies that b ≤ a and hence φ(w)φ(y ) − φ(wy ) ≥ 4Im(p q)Im(p′ q ′ ) − 4 Re(q Im(p′ q ′ ) p). 56 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (cid:3) Let ı′ ∈ S and r ∈ R such that Im(p′ q ′ ) = ı′ r. Choose ′ ∈ S in such a way that {1, ı′ , ′ , ı′ ′} is an orthonormal basis of H. Write pq = x0 + i′x1 + j ′x2 + ı′ ′ x3 , where x0 , x1 , x2 , x3 ∈ R. By a direct computation, we obtain: Im(p q)Im(p′ q ′ )2 − (cid:0)Re(q Im(p′ q ′ ) p)(cid:1)2 = r2 (Im(p q)2 − Re(ı′ p q)2 ) = 1 (cid:1) ≥ 0. = r2 (cid:0)(x2 1 + x2 2 + x2 3 ) − x2 It follows that φ(w)φ(y ) − φ(wy ) ≥ 0. 6.3. C ∗ -algebra structures on slice functions. The aim of this section is to define a structure of quaternionic two–sided Banach unital C ∗–algebra on S (ΩK , H). As a consequence, Sc (ΩK , H) will turn out to be a quaternionic two–sided Banach unital C ∗ –subalgebra of S (ΩK , H). Furthermore, restricting the scalars to R and to a fixed C , SR (ΩK , H) will be a commutative real Banach unital C ∗–subalgebra of S (ΩK , H) and SC (ΩK , H) will be a commutative C –Banach unital C ∗ –subalgebra of S (ΩK , H). We refer again the reader to Section 3.2 for the definitions of this kind of algebraic structures. Let f = I (F ) and g = I (G) be slice functions in S (ΩK , H). The usual pointwise sum f + g of f and g defines an element in S (ΩK , H), because F + G is continuous and f + g = I (F ) + I (G) = I (F + G). Equip S (ΩK , H) with the slice product. As observed in Remark 6.12, the slice product of functions in SR (ΩK , H) is again a function in SR (ΩK , H) and coincides with the pointwise product. The same is true for Sc (ΩK , H). The slice product defines an operation on SC (ΩK , H) as well. Furthermore, the slice product on SR (ΩK , H) and on SC (ΩK , H) is commutative. Let us define quaternionic left and right scalar multiplications (q , f ) 7→ q · f and (f , q) 7→ f · q on S (ΩK , H). Given q ∈ H, denote by cq : ΩK −→ H the function constantly equal to q . Such a function belongs to S (ΩK , H). Indeed, it is the slice function induced by the stem function on K constantly equal to q . For convenience, denote c1 also by the symbol 1ΩK . Define: q · f := cq · f and f · q := f · cq , (6.9) where cq · f and f · cq are slice products. It is easy to see that f · q coincides with the pointwise scalar multiplication f q for every q ∈ H. If q ∈ R or f ∈ Sc (ΩK , H), then also q · f is equal to the pointwise scalar multiplication qf . Otherwise, qf is not, in general, a slice function and hence is different from q · f . It is immediate to see that the set S (ΩK , H), equipped with the pointwise sum and with the left and right scalar multiplications (6.9) is a quaternionic two–sided algebra with unity 1ΩK . Let us introduce an involution f 7→ f ∗ on S (ΩK , H). Fix a stem function F = F1 + iF2 on K and define f := I (F ). Denote by F ∗ : K −→ HC the stem function sending z into (F (z ))∗ , where (F (z ))∗ is defined as in (6.5). More explicitly, we have that F ∗ = F1 − iF2 , where Fh : K −→ H is defined by setting Fh (z ) := Fh (z ) with h ∈ {1, 2}. Define: f ∗ := I (F ∗ ). (6.10) Let g be another function in S (ΩK , H). By (6.6), we infer that (f · g )∗ = g ∗ · f ∗ . Given q ∈ H, it is easy to verify that (cq )∗ = cq . It follows that and (f · q)∗ = q · f ∗ . (q · f )∗ = f ∗ · q QUATERNIONIC SLICE FUNCTIONAL CALCULUS 57 Since it is evident that (f + g )∗ = f ∗ + g ∗ , the involution defined in (6.10) is a ∗–involution of S (ΩK , H). if f ∈ SR (ΩK , H) or f ∈ Sc (ΩK , H), then f ∗ ∈ The reader observes that, SR (ΩK , H) or f ∗ ∈ Sc (ΩK , H), respectively. Moreover, in these cases, we have that f ∗ = f , where f : ΩK −→ H denotes the conjugated function sending q into f (q). If f ∈ SC (ΩK , H), then F1 and F2 are C –valued and hence f ∗ ∈ SC (ΩK , H). Consider now the supremum norm k · k∞ : S (ΩK , H) −→ R+ defined by setting q∈ΩK f (q). kf k∞ := sup (6.11) Thanks to point (a) of Proposition 6.13, we infer that kf k∞ = sup z∈K kF (z )kHC . The norm k · k∞ is a quaternionic Banach C ∗ –norm on S (ΩK , H). To see this, we must prove that the set S (ΩK , H), equipped with the distance induced by the norm k · k∞ , is a complete metric space. Let {fn = I (Fn )}n∈N be a Cauchy sequence in S (ΩK , H). Thanks to (6.12), we have that {Fn}n∈N is a Cauchy sequence in C (K, HC ), where HC is equipped with the norm k · kHC and C (K, HC ) with the corresponding supremum norm. Since C (K, HC ) is a (real) Banach space, {Fn}n∈N converges to some continuous stem function F on K. Using (6.12) again, we infer that {fn}n∈N → I (F ) in S (ΩK , H). This shows the completeness of S (ΩK , H). The above discussion proves the following basic result. (6.12) Theorem 6.14. The fol lowing assertions hold: (a) The set S (ΩK , H), equipped with the pointwise sum, with the left and right scalar multiplications defined in (6.9), with the slice product, with the ∗– involution defined in (6.10) and with the supremum norm defined in (6.11), is a quaternionic two–sided Banach C ∗–algebra with unity 1ΩK . (b) The set SR (ΩK , H) is a commutative real Banach unital C ∗–subalgebra of S (ΩK , H). (c) Given  ∈ S, the set SC (ΩK , H) is a commutative C –Banach unital C ∗– subalgebra of S (ΩK , H). (d) The set Sc (ΩK , H) is a quaternionic two–sided Banach unital C ∗–subalgebra of S (ΩK , H). Remark 6.15. Thanks to point (a) of Proposition 6.13, it is immediate to verify that kf k∞ = supq∈ΩK ∩C f (q) for every f ∈ SC (ΩK , H). 6.4. H-intrinsic polynomial density. Recall that R[X, Y ] denotes the ring of real polynomials in the indeterminates X and Y , with coefficients on the right. Let us introduce a particular class of H–intrinsic slice functions (see Definition 6.7). Definition 6.16. Let g : ΩK −→ H be a slice function induced by the stem function G = G1 + iG2 : K −→ HC . We say that g is a polynomial H–intrinsic slice function on ΩK if there exist polynomials Q1 and Q2 in R[X, Y ] such that G1 (z ) = Q1 (α, β ) and G2 (z ) = Q2 (α, β ) for every z = α + iβ ∈ K. We denote by P SR (ΩK , H) the subset of SR (ΩK , H) consisting of all polynomial H–intrinsic slice functions on ΩK . Remark 6.17. (1) In the preceding definition, it is always possible to assume that Q1 is even in Y and Q2 in odd in Y ; that is, Q1 (X, −Y ) = Q1 (X, Y ) and Q2(X, −Y ) = −Q2(X, Y ) in R[X, Y ]. 58 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (2) Bearing in mind points (1) and (3) of Remark 6.4, one can easily prove that, if there exists a polynomial P ∈ R[X ] such that g (q) = P (q) on ΩK , then g ∈ P SR(ΩK , H). However, in general, there exist functions in P SR (ΩK , H), which cannot be expressed in this way. If K has an interior point in C, an example is given by the polynomial H–intrinsic slice functions sending q ∈ ΩK into q ∈ H. (3) It is easy to verify that P SR (ΩK , H) is a commutative real Banach unital C ∗–subalgebra of SR (ΩK , H). We now prove a version of the Weierstrass approximation theorem for functions in SR (ΩK , H). Proposition 6.18. If K is a non–empty compact subset of C invariant under complex conjugation, then P SR(ΩK , H) is dense in SR (ΩK , H). Proof. Let f ∈ SR (ΩK , H) induced by the stem function F = F1 + iF2 and let ε be a positive real number. By the Weierstrass approximation theorem, there exist polynomials P1 , P2 ∈ R[X, Y ] such that α+iβ∈K F1 (α + iβ ) − P1 (α, β ) < sup ε 2 α+iβ∈K F2 (α + iβ ) − P2 (α, β ) < sup Define Q1 , Q2 ∈ R[X, Y ] by setting Q1 (X, Y ) := P1 (X, Y ) + P1 (X, −Y ) 2 ε 2 . and and P2 (X, Y ) − P2 (X, −Y ) Q2 (X, Y ) := . 2 Evidently, Q1 is even in Y and Q2 is odd in Y . Moreover, given any z = α + iβ ∈ K, it holds: (cid:12)(cid:12)(cid:12)(cid:12) = F1 (z ) − Q1 (α, β ) = (cid:12)(cid:12)(cid:12)(cid:12) P1 (α, β ) + P1 (α, −β ) F1 (z ) + F1 (z ) − 2 2 = (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) ≤ F1 (z ) − P1 (α, −β ) F1 (z ) − P1 (α, β ) + 2 2 (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) F1 (z ) − P1 (α, β ) F1 (z ) − P1 (α, −β ) 2 2 ε ε ε < 4 4 2 Similarly, we have that F2 (z ) − Q2 (α, β ) < ε/2. Denote by g the function in P SR (ΩK , H) induced by the stem function G = G1 + iG2 on K defined by setting if z = α + iβ ∈ K. G1 (z ) := Q1 (α, β ) and G2 (z ) := Q2 (α, β ) Thanks to equality (6.12) and to the second inequality of (6.8), we obtain: z∈K kF (z ) − G(z )kHC ≤ kf − gk∞ = sup ≤ sup (F1 (z ) − G1 (z ) + F2 (z ) − G2 (z )) < z∈K The proof is complete. = ε. ε 2 ε 2 + + = . (cid:3) QUATERNIONIC SLICE FUNCTIONAL CALCULUS 59 7. Continuous slice functional calculus for normal operators Throughout this final part, H wil l denote a fixed quaternionic Hilbert space and T wil l be a fixed normal operator in B(H). We remind the reader that B(H) indicates the set of all (bounded right H–linear) operators of H. 7.1. Polynomial H-intrinsic slice functions of a normal operator T . Define the self–adjoint operator A ∈ B(H) and the positive operator B ∈ B(H) by setting 1 1 and B := T − T ∗ A := (T + T ∗ ) . 2 2 Theorem 5.9 ensures the existence of an anti self–adjoint and unitary operator J ∈ B(H) such that T = A + J B , A, B and J commute mutually, J is uniquely determined by T on Ker (T − T ∗)⊥ . (7.1) The reader observes that T ∗ = A − J B and hence B = −J (T − T ∗ ) 1 2 . Denote by K the non–empty compact subset of C, invariant under complex conjugation, such that ΩK = σS (T ). Definition 7.1. Let g ∈ P SR (σS (T ), H) be a polynomial H–intrinsic slice function on σS (T ) and let G : K −→ HC be the stem function inducing g . Choose poly- nomials Q1 , Q2 ∈ R[X, Y ] such that G(α + iβ ) = Q1 (α, β ) + iQ2(α, β ) for every α + iβ ∈ K. We define the operator g (T ) ∈ B(H) by setting (7.2) g (T ) := Q1 (A, B ) + J Q2 (A, B ). The definition of g (T ) just given is consistent as we see in the next result. Lemma 7.2. The definition of g (T ) given in (7.2) depends only on g and T , not on the operator J and on the polynomials Q1 and Q2 we chose. Proof. By Corollary 5.12, the operators Q1(A, B ) and Q2 (A, B ) depend only on g and T . It follows that the operator Q1(A, B ) + J Q2 (A, B ) depends only on g , T and J . Let us prove that it is independent from J . By Remark 6.17(1), we may suppose that Q2 is odd in Y . In this way, we have that Q2 (A, B ) = CB for some C ∈ B(H) and hence Ker (T − T ∗ ) = Ker (B ) ⊂ Ker (Q2 (A.B )). It follows that the operator J Q2 (A, B ) vanishes on Ker (T − T ∗ ) and hence it is uniquely determined by T on Ker (T − T ∗). On the other hand, by (7.1), the operator J Q2 (A, B ) = Q2 (A, B )J is uniquely determined by T on Ker (T − T ∗)⊥ as well. This completes the proof. (cid:3) Proposition 7.3. For every g ∈ P SR (σS (T ), H), g (T ) is a normal operator in B(H), which commutes with J and satisfies the fol lowing equalities: kg (T )k = kgk∞ (7.3) and (7.4) σS (g (T )) = g (σS (T )). 60 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (7.5) Proof. We begin as in the proof of Corollary 5.12. Since A∗ = A, B ∗ = B , J ∗ = −J and A, B and J commute mutually, it is immediate to verify that g (T ) is a normal operator in B(H) and J commutes with Q1 (A, B ), Q2 (A, B ) and hence with g (T ). Fix ı ∈ S and identify C with Cı via the real isomorphism α + iβ 7→ α + ıβ . Apply the spectral radius formula (see (4.3)) and Proposition 5.11(c) to g (T ). We obtain that kg (T )k = sup nq ∈ R+ (cid:12)(cid:12)(cid:12) q ∈ σ(cid:0)g (T )HJ ı + (cid:1) o , + ) obtained restricting g (T ) to HJ ı denotes the operator in B(HJ ı where g (T )HJ ı + . + + ) ⊂ HJ ı + ) ⊂ HJ ı + and B (HJ ı Thanks to Proposition 5.11, we know that A(HJ ı + . the operators in B(HJ ı Denote by AHJ ı and B HJ ı + ) obtained restricting A and B + + 2 , B = −J (T − T ∗) 1 + , respectively. Bearing in mind that A = (T + T ∗) 1 to HJ ı 2 and = (cid:0)T HJ ı )∗ (cid:1) 1 J u = uı for every u ∈ HJ ı + (T HJ ı + , we infer that AHJ ı 2 , B HJ ı = )∗ (cid:1) 1 (cid:0)T HJ ı + + + + + − (T HJ ı 2ı and + = Q1 (cid:0)AHJ ı + (cid:1) + Q2 (cid:0)AHJ ı + (cid:1) ı. , B HJ ı g (T )HJ ı , B HJ ı (7.6) + + + Let G be the stem function on K inducing g and let Q1 , Q2 ∈ R[X, Y ] be poly- nomials such that G(α + iβ ) = Q1 (α, β ) + iQ2 (α, β ) for every α + iβ ∈ K. Consider G as a continuous function from K ⊂ Cı to Cı by setting G(α + ıβ ) = g (α + ıβ ) = Q1(α) + Q2 (α, β ) ı. Combining (7.6) with the continuous functional calculus the- orem for complex normal operators (see [RudinARC,Analysisnow,Moretti), we see = G(T HJ ı at once that g (T )HJ ı ) and + + ) = G(σ(T HJ ı σ(g (T )HJ ı (7.7) )). + + Thanks (7.5), (7.7) and Proposition 5.11(c), we infer that kg (T )k = sup nG(α + ıβ ) ∈ R+ (cid:12)(cid:12)(cid:12) α + ıβ ∈ σ(cid:0)T HJ ı + (cid:1) o = ∈ R+ (cid:12)(cid:12)(cid:12) α + ıβ ∈ σS (T ) ∩ Cıo = = sup n (cid:0)Q1 (α, β )2 + Q2 (α, β )2 (cid:1)1/2 = sup{g (q) ∈ R+ q ∈ σS (T )} = kgk∞ . Using (7.7) and Proposition 5.11(c) again, together with Lemma 6.8(ii) and with the fact that G is the stem function inducing g , we obtain: σS (g (T )) ∩ Cı = σ(g (T )HJ ı ) ∪ σ(g (T )HJ ı )) ∪ G(σ(T HJ ı ) = G(σ(T HJ ı )) = + + + + )) = G (cid:16)σ(T HJ ı )(cid:17) = )) ∪ G(σ(T HJ ı = G(σ(T HJ ı ) ∪ σ(T HJ ı + + + + = g (σS (T ) ∩ Cı ) = g (σS (T )) ∩ Cı . Since σS (g (T )) and g (σS (T )) are both circular subsets of H, it follows that they are equal. (cid:3) 7.2. Continuous H-intrinsic slice functions of a normal operator T . In this section, we assume that the set SR (σS (T ), H) is equipped with the commutative real Banach unital C ∗–algebra structure given in Theorem 6.14(b). Furthermore, we consider B(H) as a real Banach C ∗ –algebra with unity I; that is, B(H) is equipped with the pointwise sum, with the real scalar multiplication defined in (2.8), with the composition as product, with the adjunction T 7→ T ∗ as ∗–involution and with the norm defined in (2.6). QUATERNIONIC SLICE FUNCTIONAL CALCULUS 61 Theorem 7.4. There exists, and is unique, a continuous ∗–homomorphism ΨR,T : SR (σS (T ), H) ∋ f 7→ f (T ) ∈ B(H) of real Banach unital C ∗ –algebras such that: (i) ΨR,T is unity–preserving; that is, ΨR,T (1σS (T ) ) = I. (ii) ΨR,T (id ) = T , where id : σS (T ) ֒→ H denotes the inclusion map. The fol lowing further facts hold true. (a) If f ∈ SR (σS (T ), H) and J ∈ B(H) is an anti self–adjoint and unitary operator satisfying (5.17) and (5.18), then ΨR,T (f ) is normal and commutes with J . (b) ΨR,T is isometric; that is, kf (T )k = kf k∞ for every f ∈ SR (σS (T ), H). (c) For every f ∈ SR (σS (T ), H), the fol lowing continuous H–intrinsic slice spec- tral map property holds: σS (f (T )) = f (σS (T )). Proof. Uniqueness of ΨR,T is simply proved. Suppose that ΨR,T exists. If Ψ : SR (σS (T ), H) −→ B(H) is another continuous ∗ –homomorphism satisfying (i) and (ii), then it coincides with ΨR,T on P SR (σS (T ), H). On the other hand, thanks to Proposition 6.18, the set P SR (σS (T ), H) is dense in SR (σS (T ), H) and hence, by continuity, Ψ and ΨR,T coincide everywhere on SR (σS (T ), H). Let us pass to prove the existence of ΨR,T . First, define the map ψR,T : P SR (σS (T ), H) −→ B(H) by setting ΨR,T (g ) := g (T ) as in Definition 7.1. It is immediate to verify that ψR,T is a unity–preserving ∗–homomorphism of real Banach unital ∗–algebras send- ing id into T . Moreover, thanks to Proposition 7.3, ψR,T satisfies conditions (a), (b) and (c) with “ΨR,T ” and “f ∈ SR (σS (T ), R)” replaced by “ψR,T ” and “g ∈ P SR (σS (T ), R)”, respectively. In particular, ψR,T is isometric and hence, thanks to the density of P SR (σS (T ), H) in SR (σS (T ), H), it admits a unique exten- sion ΨR,T defined on the whole SR (σS (T , H)). Evidently, by continuity, ΨR,T is a ∗–homomorphism of real Banach unital C ∗ –algebras, satisfying (a) and (b). Let us show that ΨR,T verifies (c). Fix f ∈ SR (σS (T ), H). First, we prove that f (σS (T )) ⊂ σS (f (T )). Choose q ∈ σS (T ) and a sequence {gn}n∈N in P SR (σS (T ), H) converging to f . Observe that {gn(q)}n∈N → f (q) in H, {gn (T )}n∈N → f (T ) in B(H) and hence {∆gn (q) (gn (T ))}n∈N → ∆f (q) (f (T )) in B(H). Recall that the set S of operators in B(H), which do not admit a two–sided inverse in B(H), is closed in B(H) (see Proposition 2.11(c)). By (7.4), each operator ∆gn (q) (gn (T )) belongs to S . It follows that ∆f (q) (f (T )) ∈ S or, equivalently, f (q) ∈ σS (f (T )). It remains to prove that σS (f (T )) ⊂ f (σS (T )). Let p 6∈ f (σS (T )). We must show that p 6∈ σS (f (T )); that is, ∆p (f (T )) has a two–sided inverse in B(H). Let ∆p : H −→ H be the polynomial real function sending q into q2 − q(p + p) + p2 . The continuous H–intrinsic slice function f on σS (T ) and ∆p can be composed, and the composition is still a continuous H–intrinsic slice function ∆p f : σS (T ) −→ H. Indeed, if F is the stem function inducing f , then ∆p f is induced by the stem function ∆pF := F 2 − F (p + p) + p2 . The function ∆pf is nowhere zero. Let us prove this assertion. On the contrary, suppose that there exists y ∈ σS (T ) such that ∆pf (y ) = 0. This is equivalent to say that f (y ) ∈ Sp . Since f is a H–intrinsic slice function, it would follow that f (Sy ) = Sp and hence, it being Sy ⊂ σS (T ), we would infer that p ∈ f (σS (T )), which contradicts our hypothesis. Thanks to 62 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI the fact that ∆p f is nowhere zero, we can define the function f ′ : σS (T ) −→ H by setting f ′ (q) := (∆p f (q))−1 . It is immediate to verify that f ′ is a continuous H–intrinsic slice function, the one induced by the stem function sending z into the inverse of ∆pF (z ) in HC . Since ΨR,T is a unity–preserving homomorphism, we infer that (7.8) ∆p (f (T ))f ′ (T ) = ∆pf (T )f ′(T ) = I = f ′ (T )∆p f (T ) = f ′ (T )∆p (f (T )). This means that ∆p (f (T )) has a two–sided inverse in B(H), as desired. (cid:3) We conclude this section with two corollaries we will use later. Corollary 7.5. Let H be a quaternionic Hilbert space, let T ∈ B(H) be a normal operator, let J ∈ B(H) be an anti self–adjoint and unitary operator satisfying (5.17) ) −→ Cı and (5.18), let ı ∈ S and let f ∈ SR (σS (T ), H). Denote by f Cı : σ(T HJ ı + the continuous function obtained restricting f . Then the restriction ΨR,T (f )HJ,ı of + + defines an operator in B(HJ,ı ΨR,T (f ) to HJ,ı + ) and it holds: = f Cı (T HJ ı ΨR,T (f )HJ,ı ), + + ) indicates the operator in B(HJ  where f C (T HJ  + ) defined in the framework of + standard continuous functional calculus in Cı–Hilbert spaces. is a well–defined operator in B(HJ,ı Proof. The fact that ΨR,T (f )HJ,ı + ) follows im- + mediately from Theorem 7.4(a) and Proposition 5.11. By Corollary 5.13, we know ı . Let w ∈ C (σS (T ) ∩ C+ ) = σS (T ) ∩ C+ that σ(T H J ı ı , Cı ). Lemma 6.8 and the + representation formula for slice functions (see Remark 6.4(1)) imply the existence and the unicity of a function W ∈ SR (σS (T ), H) such that W σS (T )∩C+ = w. Us- ı ing again Theorem 7.4(a), we infer that the operator ΨR,T (W ) ∈ B(H) is normal and J commutes both with ΨR,T (W ) and ΨR,T (W )∗ = ΨR,T (W ∗ ). In this way, Proposition 5.11 and Theorem 7.4 ensures that the map C (σS (T ) ∩ C+ + ∈ B(HJ ı ı , Cı ) ∋ w 7→ ΨR,T (W )HJ ı + ) is a Cı–complex ∗–homomorphism, sending 1σS (T )∩C+ into the identity operator on ı + and the inclusion map σS (T ) ∩ C+ HJ ı ֒→ Cı into T HJ ı . By the uniqueness of ı + the continuous functional calculus ∗–homomorphism in Cı–Hilbert spaces, it holds: ) for every w ∈ C (σS (T ) ∩ C+ ΨR,T (W )HJ ı = w(T HJ ı ı , Cı ), as desired. (cid:3) + + Corollary 7.6. Let J ∈ B(H) be an anti self–adjoint and unitary operator sati- sfying (5.17) and (5.18), and let K ∈ B(H) be another anti self–adjoint and unitary operator such that J K = −K J and K commutes both with A = (T + T ∗ ) 1 2 and with B = T − T ∗ 1 2 . Choose f ∈ SR (σS (T ), H) and define f (T ) := ΨR,T (f ) and f ∗(T ) := ΨR,T (f ∗ ). Then it holds: f (T )J = J f (T ), (7.9) f (T )K = K f ∗ (T ). (7.10) Proof. First, suppose that f = I (F1 + iF2 ) ∈ P SR (σS (T ), H). Let Q1 , Q2 ∈ R[X, Y ] such that Fm (α, β ) = Qm (α, β ) for every m ∈ {1, 2} and (α, β ) ∈ K, where K is the subset of C, invariant under complex conjugation, such that ΩK = σS (T ). By Definition 7.1 and Lemma 7.2, we have that f (T ) = Q1 (A, B ) + J Q2 (A, B ). Since J and K commute both with A and with B , and J K = −K J , equalities (7.9) and QUATERNIONIC SLICE FUNCTIONAL CALCULUS 63 (7.10) are evident. Now the corollary follows immediately from Proposition 6.18 and from the continuity of ΨR,T (see Theorem 7.4). (cid:3) Remark 7.7. When T ∈ B(H) is self–adjoint, an easy inspection shows that the functions of the operator T defined in Theorem 5.5 coincide with those defined in Theorems 7.4. 7.3. Continuous C  -slice functions of a normal operator T . We can pass to discuss the analogue for continuous C–slice functions of what done for continuous H–intrinsic slice functions. Fix  ∈ S and equip SC (σS (T ), H) with the structure of commutative (two–sided) C–Banach unital C ∗–algebra given in Theorem 6.14(c). Recall that, given q ∈ C and f ∈ SC (σS (T ), H), the left and right scalar multiplications q · f and f · q in SC (σS (T ), H) are defined as the slice products q · f := cq · f and f · q := f · cq , where cq denotes the slice function on σS (T ) constantly equal to q . We know that q · f and f · q are equal and coincide with the pointwise scalar product f q . In this way, taking into account only the right scalar multiplication (f , q) 7→ f · q , we can consider SC (σS (T ), H) as a standard commutative C –Banach unital C ∗–algebra. Fix an anti self–adjoint and unitary operator J ∈ B(H) satisfying (5.17) and (5.18). We remind the reader that condition (5.17) requires that J commute with T and T ∗ , and condition (5.18) imposes that T decomposes as follows: T = A + J B , 2 and B := T − T ∗ 1 where A := (T + T ∗) 1 2 . Fix a left scalar multiplication H ∋ q 7→ Lq of H such that L = J . Such a left scalar multiplication of H exists by Proposition 3.8(a) and makes B(H) a quaternionic two–sided Banach unital C ∗ –algebra, via Theorem 3.4. Restricting the scalars from H to C , one defines on B(H) a structure of two–sided C –Banach unital C ∗–algebra. It is important to observe that such a structure on B(H) depends only on real scalar multiplication (2.8) and on J , and not on the fixed left scalar multiplication H ∋ q 7→ Lq of H. Indeed, if q = α + β ∈ C with α, β ∈ R, then it holds: qT = T α + J (T β ) and T q = T α + (T β )J. Since J commutes with T , then qT = T q for every q ∈ C and hence B(H) can be considered as a standard C –Banach unital C ∗ –algebra by taking into account only the right scalar multiplication (T , α + β ) 7→ T α + (T β )J . We assume that B(H) is equipped with that structure of C –Banach unital C ∗– algebra. We are now in a position to present our next result. Theorem 7.8. There exists, and is unique, a continuous ∗–homomorphism ΨC ,T : SC (σS (T ), H) ∋ f 7→ f (T ) ∈ B(H) of C –Banach unital C ∗ –algebras such that: (i) ΨC ,T is unity–preserving; that is, ΨR,T (1σS (T ) ) = I. (ii) ΨC ,T (id ) = T , where id : σS (T ) ֒→ H denotes the inclusion map. The fol lowing further facts hold true. (a) ΨC ,T extends ΨR,T in the fol lowing sense. Let f ∈ SC (σS (T ), H) and let f0 and f1 be the unique functions in SR (σS (T ), H) such that f = f0 + f1  64 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (see Lemma 6.11). Then it holds: (7.11) ΨC ,T (f ) = ΨR,T (f0 ) + ΨR,T (f1 )J. (b) For every f ∈ SC (σS (T ), H), ΨC ,T (f ) is normal and commutes with J . (c) For every f ∈ SC (σS (T ), H), the fol lowing continuous C–slice spectral map property holds: σS (f (T )) = Ωf (σS (T )∩C+  ) . (d) ΨC ,T is norm decreasing; that is, kf (T )k ≤ kf k∞ if f ∈ SC (σS (T ), H). More precisely, it holds: (7.12) kf (T )k = kf σS (T )∩C+  k∞ for every f ∈ SC (σS (T ), H). (e) The kernel of ΨC ,T consists of al l functions in SC (σS (T ), H) vanishing on σS (T ) ∩ C+  . More precisely, a function f ∈ Ker (ΨC ,T ) if and only if there exists g ∈ C (σS (T ) ∩ C−  , C ) with g σS (T )∩R = 0 such that 1 (1 + ı) g (α − β ) f (α + ıβ ) = 2 for every α ∈ R, β ∈ R+ and ı ∈ S with α + ıβ ∈ σS (T ). Proof. We begin proving the uniqueness of ΨC ,T . Assume that ΨC ,T exists. Since ∗–homomorphism satisfying (i) and (ii), if a function f = f0 + f1  in it is a C SC (σS (T ), H) is decomposed as in (a), then one has ΨC ,T (f ) = ΨC ,T (f0 ) + ΨC ,T (f1 )J and then, by Theorem 7.4, ΨC ,T (f0 ) = ΨR,T (f0 ) and ΨC ,T (f1 ) = ΨR,T (f1 ). It follows that ΨC ,T is unique. Concerning the existence, we have now a natural way to define ΨC ,T : ΨC ,T (f ) := ΨR,T (f0 ) + ΨR,T (f1 )J if f = f0 + f1  ∈ SC (σS (T ), H) with f0 , f1 ∈ SR (σS (T ), H). Evidently, ΨC ,T sati- sfies (i), (ii) and (a). Denote ΨC ,T (f ) simply by f (T ) for every f ∈ SC (σS (T ), H). ∗–homomorphism. Fix two functions f = f0 + f1  Let us verify that ΨC ,T is a C and g = g0 + g1  in SC (σS (T ), H) decomposed as in (a). For simplicity, denote ΨR,T (f0 ) by f0 (T ) and ΨR,T (f1 ) by f1 (T ). It holds: f · g = f0g0 + f0g1  + f1 ·  · g0 + f1 ·  · g1 ·  = = f0g0 + f0g1  + f1g0  + f1 ·  ·  · g1 = = (f0g0 − f1g1) + (f0g1 + f1g0) and hence, bearing in mind that ΨR,T is a real ∗–homomorphism satisfying point (a) of Theorem 7.4, we have: (7.13) (f · g )(T ) = f0 (T )g0(T ) − f1 (T )g1 (T ) + f0 (T )g1(T )J + f1 (T )g0 (T )J = = f0 (T )g0(T ) + f0 (T )g1 (T )J + f1 (T )J g0(T ) + f1 (T )J g1 (T )J = = (f0 (T ) + f1 (T )J )(g0 (T ) + g1 (T )J ) = f (T )g (T ). Let q = α + β ∈ C with α, β ∈ R. Replacing g with cq in (7.13), we obtain that f · q = (f0α − f1β ) + (f1α + f0β ). Since (f1 (T )α)J = (f1 (T )J )α and −f1 (T )β = QUATERNIONIC SLICE FUNCTIONAL CALCULUS 65 (f1 (T )J β )J , we infer that (f · q)(T ) = (f0 (T )α − f1 (T )β ) + (f1 (T )α + f0 (T )β )J = = (f0 (T ) + f1 (T )J )α + (cid:0)(f0 (T ) + f1 (T )J )β (cid:1)J = = f (T ) q . It remains to show that ΨC ,T preserves the ∗–involutions. Observe that f ∗ = 1 . Since f ∗ 0 −  · f ∗ f ∗ 0 (T ) = f0 (T )∗ , f ∗ 1 (T ) = f1 (T )∗ and J commutes 1 = f ∗ 0 − f ∗ with f ∗ 1 (T ), we have that −f ∗ 1 (T )J = (f1 (T )J )∗ and hence f ∗ (T ) = f ∗ 0 (T ) − f ∗ 1 (T )J = (f0 (T ) + f1(T )J )∗ = f (T )∗ . The operators f0 (T ), f ∗ 0 (T ), f1 (T ), f ∗ 1 (T ) and J commute mutually. This fact implies at once that f (T ) is normal and commutes with J ; that is, (b) is proved. Let us show (c). First, we need to show that = f C (T HJ  f (T )HJ  + + By Corollary 7.5, we have that f0 (T )HJ  = f0 C (THJ  ) and f1 (T )HJ  = f1 C (THJ  + + + + where f0 C and f1 C denote the operators in B(HJ  + ) defined in the (standard) functional calculus in C –Hilbert spaces. It holds: = (cid:0)f0 (T ) + f1 (T )J (cid:1)HJ  = f0 (T )HJ  f (T )HJ  + + + )(cid:17)  = ) + (cid:16)f1 C (T HJ  = f0 C (T HJ  + + = (cid:0)f0 C + (f1 )C (cid:1) (T HJ  ) = f C (T HJ  + + which proves (7.14). We are now in a position to prove (c). Apply Proposition 5.11 (c) to f (T ). We obtain that f1 (T )HJ  + + J HJ  + (7.14) ). ), = ), ). (7.15) ) ∪ σ(f (T )HJ  σS (f (T )) ∩ C = σ(f (T )HJ  + + By combining (7.14), (7.15) and Corollary 5.13 with the continuous functional cal- culus theorem for C normal operators, we infer that )) ∪ σ(f C (T HJ  σS (f (T )) ∩ C = σ(f C (T HJ  + + )) ∪ f C (σ(T HJ  = f C (σ(T HJ  )) = + +  ) ∪ f (σS (T ) ∩ C+ = f (σS (T ) ∩ C+  ). )) = This proves (c). Let us prove (d). Since f (T ) is normal, we can apply the spectral radius formula (see (4.3)) obtaining the equality kf (T )k = sup{q ∈ R+ q ∈ σS (f (T ))}. Piecing together the latter equality with (c) and with Remark 6.15, we infer at once (d). Point (e) is an immediate consequence of (d) and of the representation formula for slice functions (see Remark 6.4(1)). (cid:3) Remark 7.9. Unless the case in which Ker (T − T ∗ ) = {0}, the map ΨC ,T depends on the choice of J . Indeed, J is uniquely determined by T on Ker (T − T ∗ )⊥ (see Theorem 5.9), but it can be chosen in many ways on Ker (T −T ∗) (see Remark 5.10). In this way, if c is the function on σS (T ) constantly equal to , then ΨC ,T (c ) is equal to J and hence it is not uniquely determined by T on Ker (T − T ∗) 6= {0}. 66 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI 7.4. Continuous slice functions of a normal operator T : the circular and general cases. Fix an anti self–adjoint and unitary operator J ∈ B(H) satisfying (5.17) and (5.18). Fix , κ ∈ S with κ = −κ and fix a left scalar multiplication H ∋ q 7→ Lq of H such that L = J and, for every q ∈ H, LqA = ALq and LqB = BLq . The existence of such a left scalar multiplication of H is ensured by Theorem 5.14. Define K := Lκ . By Proposition 3.1, we infer that K is an anti–self adjoint and unitary operator in B(H) such that J K = −K J . Final ly, we equip B(H) with the structure of quaternionic two–sided Banach unital C ∗–algebra induced by the fixed left scalar multiplication H ∋ q 7→ Lq of H, as described in Theorem 3.4. 7.4.1. The circular case. We now consider continuous circular slice functions, which, differently from the cases of continuous H–intrinsic slice and C –slice functions we treated above, concerns non–commutative structures. We assume that the set Sc (σS (T ), H) is equipped with the quaternionic two–sided Banach unital C ∗–algebra structure given in Theorem 6.14(d). Our next result is as follows. Theorem 7.10. There exists, and is unique, a continuous ∗–homomorphism Ψc,T : Sc (σS (T ), H) ∋ f 7→ f (T ) ∈ B(H) of quaternionic two–sided Banach unital C ∗ –algebras such that: (i) Ψc,T is unity–preserving; that is, ΨR,T (1σS (T ) ) = I. (ii) Ψc,T (id ) = T , where id : σS (T ) ֒→ H denotes the inclusion map. The fol lowing further facts hold true. (a) Ψc,T extends ΨR,T and ΨC ,T in the fol lowing sense. Let f ∈ Sc (σS (T ), H) and let f0 , f1 , f2 , f3 be the unique functions in SR (σS (T ), H) such that f = f0 + f1  + f2κ + f3 κ (see Lemma 6.11). Then it holds: Ψc,T (f ) = ΨR,T (f0 ) + ΨR,T (f1 )J + ΨR,T (f2 )K + ΨR,T (f3 )J K = = ΨC ,T (f0 + f1 ) + ΨC ,T (f2 + f3 )K. (b) For every f ∈ Sc (σS (T ), H), Ψc,T (f ) is normal. (c) For every f ∈ Sc (σS (T ), H), the fol lowing continuous circular slice spectral map property holds: σS (f (T )) ⊂ Ωf (σS (T )) . (d) Ψc,T is norm decreasing; that is, kf (T )k ≤ kf k∞ if f ∈ Sc (σS (T ), H). Before presenting the proof of this result, we underline that point (d) of the preceding theorem can be improved as follows: (d′ ) Ψc,T is isometric; that is, kf (T )k = kf k∞ if f ∈ Sc (σS (T ), H). We will prove this stronger property of Ψc,T in the forthcoming paper [18] mak- ing use of a spectral representation theorem for normal operators on quaternionic Hilbert spaces. Proof of Theorem 7.10. Let us prove that Ψc,T is unique. Suppose that Ψc,T exists. Let f = f0 + f1  + f2κ + f3 κ be a function in Sc (σS (T ), H) decomposed as in (a). QUATERNIONIC SLICE FUNCTIONAL CALCULUS 67 Thanks to Theorem 5.9 and to the fact that Ψc,T is a quaternionic ∗–homomorphism satisfying (i) and (ii), we infer immediately the uniqueness of Ψc,T . Let us pass to define Ψc,T assuming the truthfulness of (a): Ψc,T (f ) := f0 (T ) + f1 (T )J + f2 (T )K + f3 (T )J K if f = f0 + f1+ f2κ+ f3κ is a function in Sc (σS (T ), H) decomposed as in (a), where fℓ (T ) := ΨR,T (fℓ ) for every ℓ ∈ {0, 1, 2, 3}. It is evident that Ψc,T satisfies (i) and (ii). Let ℓ ∈ {0, 1, 2, 3}. By Lemma 6.11, we have that fℓ belongs to SR (σS (T ), H) and is real–valued. Thanks to the latter fact, we have that f ∗ ℓ = fℓ = fℓ . Moreover, Corollary 7.6 ensures that fℓ(T )J = J fℓ(T ) and fℓ (T )K = K fℓ (T ). In this way, the functions f and their images Ψc,T (f ) under Ψc,T behave like quaternions. This implies at once that Ψc,T is a ∗ –homomorphism satisfying (b). It remains to prove (c). Indeed, (d) can be easily deduced from (c) as we did at the end of the proof of Theorem 7.8. In order to prove point (c), we follow the strategy used at the end of the proof of Theorem 7.4. Let f = I (F1 ) ∈ Sc (σS (T ), H), let f (T ) := Ψc,T (f ) and let p 6∈ Ωf (σS (T )) . We must show that ∆p (f (T )) has a two–sided inverse in B(H). Let ∆p f : σS (T ) −→ H be the function sending q into f 2(q) − f (q)(p + p) + p2 . Such a function belongs to Sc (σS (T ), H), because it is 1 − F1 (p + p) + p2 . The the slice function induced by the stem function ∆pF1 := F 2 function ∆pf is nowhere zero. Indeed, it there would exist y ∈ σS (T ) such that ∆pf (y ) = 0, then f (y ) ∈ Sp and hence p ∈ Ωf (σS (T )) , which is impossible. In this way, we can define f ′ : σS (T ) −→ H by setting f ′(q) := (∆p f (q))− 1. This function is an element of Sc (σS (T ), H) induced by the stem function (∆pF1 )−1 . Evidently, equalities (7.8) (with f ′ (T ) := Ψc,T (f ′ )) hold also in this situation. The proof is complete. (cid:3) Remark 7.11. The map Ψc,T depends on the choice of J if Ker (T − T ∗) 6= {0} (see Remark 7.9) and always on K . Indeed, K is not uniquely determined by T and Ψc,T (cκ ) = K if cκ is the function on σS (T ) constantly equal to κ. 7.4.2. Continuous slice functions of T . We now come to the general case: we extend the previous definitions of the operator f (T ) to every continuous slice function f ∈ S (σS (T ), H). Given f ∈ S (σS (T ), H), let f0 , f1 , f2 , f3 be the unique functions in SR (σS (T ), H) such that f = f0 + f1  + f2κ + f3 κ (see Lemma 6.11). Define f (T ) := f0 (T ) + f1 (T )J + f2 (T )K + f3 (T )J K, where fℓ(T ) denotes ΨR,T (fℓ ) for every ℓ ∈ {0, 1, 2, 3}. From the definition, it follows immediately that the map f 7→ f (T ) is R–linear. Since ΨR,T is continuous, the map f 7→ f (T ) is also continuous; that is, there exists a positive constant C such that kf (T )k ≤ C kf k∞ for every f ∈ S (σS (T ), H). Furthermore, we have: Proposition 7.12. Given f , g ∈ S (σS (T ), H), the fol lowing facts hold. (a) If f ∈ SC (σS (T ), H) or g ∈ Sc (σS (T ), H), then (f · g )(T ) = f (T )g (T ). (b) If p ∈ C and q ∈ H, then (p · g )(T ) = p g (T ) and (f · q)(T ) = f (T )q . 68 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI (c) Let f = f0 + f1  + f ∗ 3 κ. Then f ∈ S (σS (T ), H) and it holds: 2 κ + f ∗ f (T )∗ = f ∗ (T ) Proof. Let f = (f0 + f1 ) + (f2 + f3 )κ = f 1 + f 2κ, where f 1 := f0 + f1 , f 2 := f2 + f3 ∈ SC (σS (T ), H), and similarly for g = g 1 + g 2κ. From κ = −κ, it follows that k · f = f ′ · k , where f ′ := (f0 − f1 ) + (f2 − f3 )κ. Therefore f · g = (f 1 + f 2κ) · (g 1 + g 2κ) = f 1 · g 1 − f 2 · (g 2 )′ + (cid:0)f 1 · g 2 + f 2 · (g 1 )′ (cid:1) κ and (f · g )(T ) = f 1 (T )g 1(T ) − f 2 (T )(g 2 )′ (T ) + (cid:0)f 1 (T )g 2 (T ) + f 2 (T )(g 1)′ (T )(cid:1) K. On the other hand, from 7.6 we get that f (T )g (T ) = (f 1 (T ) + f 2 (T )K )(g 1(T ) + g 2 (T )K ) = = f 1 (T )g 1(T ) − f 2 (T )(g 2 )∗ (T ) + (cid:0)f 1 (T )g 2 (T ) + f 2 (T )(g 1)∗ (T )(cid:1) K. From Lemma 6.11, we get that if f ∈ SC (σS (T ), H), then f 2 = 0, while if g ∈ Sc (σS (T ), H), then (g 1 )′ = (g 1 )∗ and (g 2 )′ = (g 2 )∗ . In both cases, (f · g )(T ) coincides with f (T )g (T ) and (a) is proved. Part (b) is an immediate consequence of (a). It remains to prove (c). Since f ∗ = (f0 +f1+f ∗ 0 −f ∗ 3 κ)∗ = f ∗ 2 κ+f ∗ 1 −f2κ−f3 κ, (c) is a consequence of the following equality: f (T )∗ = f0 (T )∗ − J f1 (T )∗ − K f2(T )∗ + K J f3(T )∗ = = f ∗ 0 (T ) − f ∗ 1 (T )J − f2 (T )K − f3 (T )J K. This proves the proposition. (cid:3) 7.5. Slice regular functions of a normal operator T . As we recalled in the Introduction, an important subclass of slice functions is the one of slice regular functions, those slice functions that are induced by holomorphic stem functions. They were introduced in [15, 16] as quaternionic power series and later generalized in [7, 20, 19]. A functional calculus for slice regular functions of a (bounded right H–linear) operator on quaternionic two–sided Banach module has been developed in [8] as an effective generalization of the (classical) holomorphic functional calculus. We recall the definition given in [8]. (7.16) f (T )reg := Definition 7.13. [8, Def. 4.10.4] Let V be a quaternionic two–sided Banach mod- ule, let T ∈ B(V ) be an operator and let f : U −→ H be a slice regular function defined on a circular open neighborhood of σS (T ) in H. Fix any  ∈ S and define the element f (T )reg of B(V ) by setting 2π Z∂ (U ∩C ) 1 Here S−1 L (s, x) denotes the Cauchy kernel for (left) slice regular functions. Proposition 7.14. Let T ∈ B(H) be a normal operator and let f : U −→ H be a slice regular function defined on a circular open neighborhood of σS (T ) in H. Then f σS (T ) ∈ S (σS (T ), H) and it holds: f (T )reg = f σS (T ) (T ); (7.17) that is, the two functional calculi coincide if T is normal and f is slice regular. S−1 L (s, T ) ds −1 f (s) . QUATERNIONIC SLICE FUNCTIONAL CALCULUS 69 Proof. Firstly consider the case of a H–intrinsic slice regular f ∈ SR (Ω, H), with Ω open neighbourhood of σS (T ). If f is a polynomial Pm n=0 qnan with real coefficients an , then f (T )reg and f σS (T ) (T ) = ΨR,T (f ) are both equal to Pm n=0 T nan (cf. [8, Theorem 4.8.10]). If f is a rational slice regular function, of the form f = p−1p′ for two H–intrinsic slice regular polynomials p and p′ , with p = I (P ) = I (P1 + iP2 ) 6= 0 on σS (T ) and p−1 = I ((P1 + iP2 )/P 2 ), the algebraic properties stated in Theorem 7.4 and [8, Proposition 4.11.6] give: ΨR,T (p)ΨR,T (f ) = ΨR,T (p′ ) = p′ (T )reg = p(T )reg f (T )reg = ΨR,T (p)f (T )reg , from which it follows that ΨR,T (f ) = f (T )reg , since ΨR,T (p) is invertible. We now show that every f = I (F ) ∈ SR (Ω, H) can be approximated in the supremum norm by rational H–intrinsic slice regular functions (see also [9] for a Runge Theorem for slice regular functions). Let K ⊂ C, invariant with respect to conjugation, such that ΩK = σS (T ). Given ǫ > 0, the classical Runge Theorem assures the existence of a rational function R, with poles outside K, such that F (z ) − R(z ) < ǫ/2 for every z ∈ K. Let R′ (z ) := (R(z ) + R( ¯z ))/2. Then R′ is a rational stem function, whose induced slice function r is regular and H–intrinsic on a neighbourhood of σS (T ). Therefore z∈K F (z ) − R′ (z ) ≤ (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) < ǫ . F (z ) − R(z ) F ( ¯z ) − R( ¯z ) kf − rkσS (T ) = sup 2 2 Using [8, Theorem 4.10.6] and the continuity of f 7→ f (T ), we obtain the equality (7.17) for every f ∈ SR (Ω, H). Now we come to the case of a generic slice regular f ∈ S (Ω, H), decomposed as f = f0 + f1  + f2κ + f3 κ, with fk ∈ SR (Ω, H). From the above and [8, Proposi- tion 4.11.1] we get (7.17): f (T )reg = f0 (T )reg + f1 (T )reg J + f2 (T )regK + f3 (T )reg J K = = f0 σS (T ) (T ) + f1 σS (T ) (T )J + f2 σS (T ) (T )K + f3 σS (T ) (T )J K = = f σS (T ) (T ) . This completes the proof. (cid:3) References [1] Stephen L. Adler, Quaternionic quantum mechanics and quantum fields, International Series of Monographs on Physics, vol. 88, The Clarendon Press Oxford University Press, New York, 1995. MR 1333599 (97d:81001) [2] Diederik Aerts, Quantum axiomatics, Handbook of quantum logic and quantum structures— quantum logic, Elsevier/North-Holland, Amsterdam, 2009, pp. 79–126. MR 2724647 (2011j:81011) [3] Enrico G. Beltrametti and Gianni Cassinelli, The logic of quantum mechanics, Encyclopedia of Mathematics and its Applications, vol. 15, Addison-Wesley Publishing Co., Reading, Mass., 1981, With a foreword by Peter A. Carruthers. MR 635780 (83d:81008) [4] Garrett Birkhoff and John von Neumann, The logic of quantum mechanics, Ann. of Math. (2) 37 (1936), no. 4, 823–843. MR 1503312 [5] F. Brackx, Richard Delanghe, and F. Sommen, Clifford analysis, Research Notes in Math- ematics, vol. 76, Pitman (Advanced Publishing Program), Boston, MA, 1982. MR 697564 (85j:30103) [6] G. Cassinelli and P. Truini, Quantum mechanics of the quaternionic Hilbert spaces based upon the imprimitivity theorem, Rep. Math. Phys. 21 (1985), no. 1, 43–64. MR 802061 (87g:81038) 70 RICCARDO GHILONI, VALTER MORETTI, AND ALESSANDRO PEROTTI [8] [9] [16] [7] Fabrizio Colombo, Irene Sabadini, and Daniele C. Struppa, Slice monogenic functions, Israel J. Math. 171 (2009), 385–403. MR 2520116 (2010e:30039) vol. 289, in Mathematics, , Noncommutative functional calculus, Progress Birkhauser/Springer Basel AG, Basel, 2011, Theory and applications of slice hyperholomor- phic functions. MR 2752913 , The Runge theorem for slice hyperholomorphic functions, Proc. Amer. Math. Soc. 139 (2011), no. 5, 1787–1803. MR 2763766 (2012c:30104) [10] Nelson Dunford and Jacob T. Schwartz, Linear operators. Part II: Spectral theory. Self ad- joint operators in Hilbert space, With the assistance of William G. Bade and Robert G. Bartle, Interscience Publishers John Wiley & Sons New York-London, 1963. MR 0188745 (32 #6181) [11] H.-D. Ebbinghaus, H. Hermes, F. Hirzebruch, M. Koecher, K. Mainzer, J. Neukirch, A. Pres- tel, and R. Remmert, Numbers, Graduate Texts in Mathematics, vol. 123, Springer-Verlag, New York, 1991, With an introduction by K. Lamotke, Translated from the second 1988 German edition by H. L. S. Orde, Translation edited and with a preface by J. H. Ewing, Readings in Mathematics. MR 1415833 (97f:00001) [12] G´erard Emch, M´ecanique quantique quaternionienne et relativit´e restreinte. I, Helv. Phys. Acta 36 (1963), 739–769. MR 0176811 (31 #1083) [13] Kurt Engesser, Dov M. Gabbay, and Daniel Lehmann (eds.), Handbook of quantum logic and quantum structures—quantum logic, Elsevier/North-Holland, Amsterdam, 2009. MR 2724659 (2011e:81007) [14] David Finkelstein, Josef M. Jauch, Samuel Schiminovich, and David Speiser, Foundations of quaternion quantum mechanics, J. Mathematical Phys. 3 (1962), 207–220. MR 0137500 (25 #952) [15] Graziano Gentili and Daniele C. Struppa, A new approach to Cul len-regular functions of a quaternionic variable, C. R. Math. Acad. Sci. Paris 342 (2006), no. 10, 741–744. MR 2227751 (2006m:30095) , A new theory of regular functions of a quaternionic variable, Adv. Math. 216 (2007), no. 1, 279–301. MR 2353257 (2008h:30052) , Regular functions on the space of Cayley numbers, Rocky Mountain J. Math. 40 (2010), no. 1, 225–241. MR 2607115 (2011c:30124) [18] R. Ghiloni, V. Moretti, and A. Perotti, Spectral representations of normal operators in quater- nionic Hilbert spaces, in preparation. [19] R. Ghiloni and A. Perotti, A new approach to slice regularity on real algebras, Hypercomplex analysis and its Applications, Trends Math., Birkhauser, Basel, 2011, pp. 109–124. , Slice regular functions on real alternative algebras, Adv. Math. 226 (2011), no. 2, 1662–1691. MR 2737796 (2012e:30061) [21] L. P. Horwitz and L. C. Biedenharn, Quaternion quantum mechanics: Second quantization and gauge fields, Annals of Physics 157 (1984), 432–488. [22] Irving Kaplansky, Normed algebras, Duke Math. J. 16 (1949), 399–418. MR 0031193 (11,115d) [23] S. H. Kulkarni, Representations of a class of real B∗ -algebras as algebras of quaternion-valued functions, Proc. Amer. Math. Soc. 116 (1992), no. 1, 61–66. MR 1110546 (92k:46089) [24] T. A. Loring, Factorization of Matrices of Quaternions Expositiones Mathematicae 30 (2012), no. 3, 250-267 [25] Valter Moretti, Spectral Theory and Quantum Mechanics, with an introduction to the Alge- braic Formulation, Springer-Verlag, Milano-Berlin, 2012 [26] Chi-Keung Ng, On quaternionic functional analysis. Math. Proc. Cambridge Philos. Soc. 143 (2007), no. 2, 391-406 [27] Edward Nelson, Analytic vectors, Ann. of Math. (2) 70 (1959), 572–615. MR 0107176 (21 #5901) [28] Gert K. Pedersen, Analysis now, Graduate Texts in Mathematics, vol. 118, Springer-Verlag, New York, 1989. MR 971256 (90f:46001) [29] C. Piron, Axiomatique quantique, Helv. Phys. Acta 37 (1964), 439–468. MR 0204048 (34 #3894) [30] Walter Rudin, Real and complex analysis, third ed., McGraw-Hill Book Co., New York, 1987. MR 924157 (88k:00002) [17] [20] QUATERNIONIC SLICE FUNCTIONAL CALCULUS 71 [31] , Functional analysis, second ed., International Series in Pure and Applied Mathe- matics, McGraw-Hill Inc., New York, 1991. MR 1157815 (92k:46001) [32] M. P. Sol`er, Characterization of Hilbert spaces by orthomodular spaces, Comm. Algebra 23 (1995), no. 1, 219–243. MR 1311786 (95k:46035) [33] V. S. Varadara jan, Geometry of quantum theory, second ed., Springer-Verlag, New York, 1985. MR 805158 (87a:81009) [34] K. Viswanath, Normal operations on quaternionic Hilbert spaces, Trans. Amer. Math. Soc. 162 (1971), 337–350. MR 0284843 (44 #2067) Department of Mathematics, University of Trento, I–38123, Povo-Trento, Italy E-mail address : [email protected], [email protected], [email protected]
1108.1463	1	1108	2011-08-06T08:21:28	Construction of pathological maximally monotone operators on non-reflexive Banach spaces	[ "math.FA" ]	In this paper, we construct maximally monotone operators that are not of Gossez's dense-type (D) in many nonreflexive spaces. Many of these operators also fail to possess the Br{\o}nsted-Rockafellar (BR) property. Using these operators, we show that the partial inf-convolution of two BC--functions will not always be a BC--function. This provides a negative answer to a challenging question posed by Stephen Simons. Among other consequences, we deduce that every Banach space which contains an isomorphic copy of the James space $\mathbf{J}$ or its dual $\mathbf{J}^*$, or $c_0$ or its dual $\ell^1$, admits a non type (D) operator.	math.FA	math	Construction of pathological maximally monotone operators on non-reflexive Banach spaces Heinz H. Bauschke∗, Jonathan M. Borwein†, Xianfu Wang‡, and Liangjin Yao§ August 6, 2011 Abstract In this paper, we construct maximally monotone operators that are not of Gossez's dense-type (D) in many nonreflexive spaces. Many of these operators also fail to possess the Brønsted-Rockafellar (BR) property. Using these operators, we show that the partial inf-convolution of two BC -- functions will not always be a BC -- function. This provides a negative answer to a challenging question posed by Stephen Simons. Among other consequences, we deduce that every Banach space which contains an isomorphic copy of the James space J or its dual J∗, or c0 or its dual ℓ1, admits a non type (D) operator. 2010 Mathematics Subject Classification: Primary 47A06, 47H05; Secondary 47B65, 47N10, 90C25 Keywords: Adjoint, BC -- function, Fitzpatrick function, James space, linear relation, max- imally monotone operator, monotone operator, multifunction, operator of type (BR), oper- ator of type (D), operator of type (NI), partial inf-convolution, Schauder basis, set-valued operator, skew operator, space of type (D), uniqueness of extensions, subdifferential operator. ∗Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. †CARMA, University of Newcastle, Newcastle, New South Wales 2308, Australia. E-mail: [email protected]. Distinguished Professor King Abdulaziz University, Jeddah. ‡Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. §Mathematics, Irving K. Barber School, University of British Columbia, Kelowna, B.C. V1V 1V7, Canada. E-mail: [email protected]. 1 1 Preliminaries Throughout this paper, we assume that X is a real Banach space with norm k · k, that X∗ is the continuous dual of X, and that X and X∗ are paired by h·, ·i. As usual, we identify X with its canonical image in the bidual space X∗∗. Furthermore, X × X∗ and (X × X∗)∗ := X∗ × X∗∗ are likewise paired via h(x, x∗), (y∗, y∗∗)i := hx, y∗i + hx∗, y∗∗i, where (x, x∗) ∈ X × X∗ and (y∗, y∗∗) ∈ X∗ × X∗∗. Let A : X ⇒ X∗ be a set-valued operator (also known as a multifunction) from X to X∗, i.e., for every x ∈ X, Ax ⊆ X∗, and let gra A :=(cid:8)(x, x∗) ∈ X × X∗ x∗ ∈ Ax(cid:9) be the graph of A. The domain of A is dom A := (cid:8)x ∈ X Ax 6= ∅(cid:9), and ran A := A(X) for the range of A. Recall that A is monotone if (1) hx − y, x∗ − y∗i ≥ 0, ∀(x, x∗) ∈ gra A ∀(y, y∗) ∈ gra A, and maximally monotone if A is monotone and A has no proper monotone extension (in the sense of graph inclusion). Let A : X ⇒ X∗ be monotone and (x, x∗) ∈ X × X∗. We say (x, x∗) is monotonically related to gra A if hx − y, x∗ − y∗i ≥ 0, ∀(y, y∗) ∈ gra A. We now recall the three fundamental subclasses of maximally monotone operators. Definition 1.1 Let A : X ⇒ X∗ be maximally monotone. Then three key types of monotone operators are defined as follows. (i) A is of dense type or type (D) (1971, [19] and [28]) if for every (x∗∗, x∗) ∈ X∗∗ × X∗ with inf (a,a∗)∈gra A ha − x∗∗, a∗ − x∗i ≥ 0, there exist a bounded net (aα, a∗α)α∈Γ in gra A such that (aα, a∗α)α∈Γ weak×strong con- verges to (x∗∗, x∗). (ii) A is of type negative infimum (NI) (1996, [32]) if sup (a,a∗)∈gra A(cid:0)ha, x∗i + ha∗, x∗∗i − ha, a∗i(cid:1) ≥ hx∗∗, x∗i, ∀(x∗∗, x∗) ∈ X∗∗ × X∗. (iii) A is of "Brønsted-Rockafellar" (BR) type (1999, [37]) if whenever (x, x∗) ∈ X × X∗, α, β > 0 while inf (a,a∗)∈gra A hx − a, x∗ − a∗i > −αβ then there exists (b, b∗) ∈ gra A such that kx − bk < α, kx∗ − b∗k < β. 2 As we shall see below in Fact 2.7, it is now known that the first two classes coincide. This coincidence is central to many of our proofs. Fact 2.11 also shows us that every maximally monotone operator of type (D) is of type (BR)(The converse fails, see Example 4.1(xiii).). Moreover, in reflexive space every maximally monotone operator is of type (D), as is the subdifferential operator of every closed convex function on a Banach space. While monotone operator theory is rather complete in reflexive space -- and for type (D) operators in general space -- the general situation is less clear [11, 9]. Hence our continuing interest in operators which are not of type (D). We shall say a Banach space X is of type (D) [9] if every maximally monotone operator on X is of type (D). At present the only known type (D) spaces are the reflexive spaces; and our work here suggests that there are no non-reflexive type (D) spaces. In [11, Exercise 9.6.3] such spaces were called (NI) spaces and some potential non-reflexive examples were conjectured; all of which are ruled out by our current work. In [11, Theorem 9.79] a variety of the pleasant properties of type (D) spaces was listed. 1.1 More preliminary technicalities Maximal monotone operators have proven to be a potent class of objects in modern Opti- mization and Analysis; see, e.g., [7, 8, 9], the books [6, 11, 13, 27, 33, 35, 31, 42] and the references therein. We adopt standard notation used in these books especially [11, Chapter 2] and [7, 33, 35]: Given a subset C of X, the indicator function of C, written as ιC, is defined at x ∈ X by (2) ιC(x) :=(0, +∞, if x ∈ C; otherwise. The closed unit ball is BX :=(cid:8)x ∈ X kxk ≤ 1(cid:9), and N := {1, 2, 3, . . .}. Let α, β ∈ R. In the sequel it will also be useful to let δα,β be defined by δα,β := 1, if α = β; δα,β := 0, otherwise. For a subset C∗ of X∗, C∗w is the weak∗ closure of C∗. If Z is a real Banach space with ∀s ∈ S}. Given dual Z∗ and a set S ⊆ Z, we denote S⊥ by S⊥ := {z∗ ∈ Z∗ hz∗, si = 0, a subset D of Z∗, we define D⊥ [29] by D⊥ := {z ∈ Z hz, d∗i = 0, ∀d∗ ∈ D}. The adjoint of an operator A, written A∗, is defined by gra A∗ :=(cid:8)(x∗∗, x∗) ∈ X∗∗ × X∗ (x∗, −x∗∗) ∈ (gra A)⊥(cid:9). We say A is a linear relation if gra A is a linear subspace. We say that A is skew if gra A ⊆ gra(−A∗); equivalently, if hx, x∗i = 0, ∀(x, x∗) ∈ gra A. Furthermore, A is symmetric if 3 gra A ⊆ gra A∗; equivalently, if hx, y∗i = hy, x∗i, ∀(x, x∗), (y, y∗) ∈ gra A. We define the symmetric part and the skew part of A via 2A∗ and S := 1 2A + 1 2 A − 1 2 A∗, (3) P := 1 respectively. It is easy to check that P is symmetric and that S is skew. Let A : X ⇒ X∗ be monotone and S be a subspace of X. We say A is S -- saturated [35] if Ax + S⊥ = Ax, ∀x ∈ dom A. We say a maximally monotone operator A : X ⇒ X∗ is unique if all maximally monotone extensions of A (in the sense of graph inclusion) in X∗∗ × X∗ coincide. Let f : X → ]−∞, +∞]. Then dom f := f−1(R) is the domain of f , and f∗ : X∗ → [−∞, +∞] : x∗ 7→ supx∈X(hx, x∗i − f (x)) is the Fenchel conjugate of f . We say f is proper if dom f 6= ∅. Let f be proper. The subdifferential of f is defined by ∂f : X ⇒ X∗ : x 7→ {x∗ ∈ X∗ (∀y ∈ X) hy − x, x∗i + f (x) ≤ f (y)}. For ε ≥ 0, the ε -- subdifferential of f is defined by ∂εf : X ⇒ X∗ : x 7→(cid:8)x∗ ∈ X∗ (∀y ∈ X) hy − x, x∗i + f (x) ≤ f (y) + ε(cid:9). Note that ∂f = ∂0f . We denote by J := JX the duality map, i.e., the subdifferential of the function 1 2 k · k2 mapping X to X∗. Now let F : X × X∗ → ]−∞, +∞]. We say F is a BC -- function (BC stands for "Bigger conjugate") [35] if F is proper and convex with (4) F ∗(x∗, x) ≥ F (x, x∗) ≥ hx, x∗i ∀(x, x∗) ∈ X × X∗. Let Y be another real Banach space. We set PX : X × Y → X : (x, y) 7→ x, and PY : X × Y → Y : (x, y) 7→ y. Let L : X → Y be linear. We say L is a (linear) isomorphism into Y if L is one to one, continuous and L−1 is continuous on ran L. We say L is an isometry if kLxk = kxk, ∀x ∈ X. The spaces X, Y are then isometric (isomorphic) if there exists an isometry (isomorphism) from X onto Y . Let F1, F2 : X × Y → ]−∞, +∞]. Then the partial inf-convolution F1(cid:3)1F2 is the function defined on X × Y by F1(cid:3)1F2 : (x, y) 7→ inf u∈X F1(u, y) + F2(x − u, y). Then F1(cid:3)2F2 is the function defined on X × Y by F1(cid:3)2F2 : (x, y) 7→ inf v∈Y F1(x, y − v) + F2(x, v). In Example 4.1(vi)&(viii) of this paper, we provide a negative answer to the following ques- tion posed by S. Simons [35, Problem 22.12]: 4 Let F1, F2 : X × X∗ → ]−∞, +∞] be proper lower semicontinuous and convex. Assume that F1, F2 are BC -- functions and that λ [PX ∗ dom F1 − PX ∗ dom F2] is a closed subspace of X∗. [λ>0 Is F1(cid:3)1F2 necessarily a BC -- function? We are now ready to set to work. The paper is organized as follows. In Section 2, we collect auxiliary results for future reference and for the reader's convenience. Our main result (Theorem 3.6) is established in Section 3. In Section 4, we provide various applications and extensions including the promised negative answer to Simons' question. Furthermore, we show that every Banach space containing an isomorphic copy of the James space J or of J∗, of ℓ1 or of c0 is not of type (D) (Example 4.1(xi) or Corollary 4.12, Corollary 4.11 and Example 4.13). 2 Auxiliary results Observation: Fact 2.1 (See [26, Proposition 2.6.6(c)]). Let D be a subspace of X∗. Then (D⊥)⊥ = D w* . We now record a famous Banach space result: Fact 2.2 (Banach and Mazur) (See [16, Theorem 5.8, page 240] or [15, Theorem 5.17, page 144]).) Every separable Banach space is isometric to a subspace of C[0, 1]. Now we turn to prerequisite results on Fitzpatrick functions, monotone operators, and linear relations. Fact 2.3 (Fitzpatrick) (See [17, Corollary 3.9 and Proposition 4.2] and [7, 11].) Let A : X ⇒ X∗ be maximally monotone, and set (5) FA : X × X∗ → ]−∞, +∞] : (x, x∗) 7→ sup (a,a∗)∈gra A(cid:0)hx, a∗i + ha, x∗i − ha, a∗i(cid:1), which is the Fitzpatrick function associated with A. Then FA is a BC -- function and FA = h·, ·i on gra A. 5 Fact 2.4 (Simons and Zalinescu) (See [38, Theorem 4.2] or [35, Theorem 16.4(a)].) Let Y be a real Banach space and F1, F2 : X × Y → ]−∞, +∞] be proper, lower semicontinuous, and convex. Assume that for every (x, y) ∈ X × Y , (F1(cid:3)2F2)(x, y) > −∞ and that Sλ>0 λ [PX dom F1 − PX dom F2] is a closed subspace of X. Then for every (x∗, y∗) ∈ X∗ × Y ∗, (F1(cid:3)2F2)∗(x∗, y∗) = min u∗∈X ∗ [F ∗1 (x∗ − u∗, y∗) + F ∗2 (u∗, y∗)] . Fact 2.5 (Simons and Zalinescu) (See [35, Theorem 16.4(b)].) Let Y be a real Banach space and F1, F2 : X ×Y → ]−∞, +∞] be proper, lower semicontinuous and convex. Assume that for every (x, y) ∈ X × Y , (F1(cid:3)1F2)(x, y) > −∞ and thatSλ>0 λ [PY dom F1 − PY dom F2] is a closed subspace of Y . Then for every (x∗, y∗) ∈ X∗ × Y ∗, (F1(cid:3)1F2)∗(x∗, y∗) = min v∗∈Y ∗ [F ∗1 (x∗, v∗) + F ∗2 (x∗, y∗ − v∗)] . Phelps and Simons proved the next Fact 2.6 for unbounded linear operators in [29, Propo- sition 3.2(a)], but their proof can also be adapted for general linear relations. For reader's convenience, we write down their proof. Fact 2.6 (Phelps and Simons) Let A : X ⇒ X∗ be a monotone linear relation. Then (x, x∗) ∈ X × X∗ is monotonically related to gra A if and only if hx, x∗i ≥ 0 and [hy∗, xi + hx∗, yi]2 ≤ 4hx∗, xihy∗, yi, ∀(y, y∗) ∈ gra A. Proof. We have the following equivalences: (x, x∗) ∈ X × X∗ is monotonically related to gra A ⇔ λ2hy, y∗i − λ [hy∗, xi + hx∗, yi] + hx, x∗i = hλy∗ − x∗, λy − xi ≥ 0, ∀λ ∈ R, ∀(y, y∗) ∈ gra A ⇔ hx, x∗i ≥ 0 and [hy∗, xi + hx∗, yi]2 ≤ 4hx∗, xihy∗, yi, ∀(y, y∗) ∈ gra A (by [29, Lemma 2.1]). This completes the proof. (cid:4) Fact 2.7 (Simons / Marques Alves and Svaiter) (See [32, Lemma 15] or [35, Theo- rem 36.3(a)], and [25, Theorem 4.4].) Let A : X ⇒ X∗ be maximally monotone. Then A is of type (D) if and only if it is of type (NI). 6 We next cite some properties regarding the uniqueness of (maximally) monotone extension of a maximally monotone operator to X∗∗ × X∗. Simons showed that every maximally monotone operator of type (NI) is unique in [34]. Recently, Marques Alves and Svaiter contributed the following results: Fact 2.8 (Marques Alves and Svaiter) (See [24, Theorem 1.6].) Let A : X ⇒ X∗ be a maximally monotone linear relation that is not of type (D). Assume that A is unique. Then gra A = dom FA. Fact 2.9 (Marques Alves and Svaiter) (See [25, Corollary 4.6].) Let A : X ⇒ X∗ be a maximally monotone operator such that gra A is not affine. Then A is of type (D) if and only if A is unique. The Gossez operator defined as in Example 4.1(xii) is a maximally monotone and unique operator that is not of type (D) [20]. The definition of operators of type (BR) directly yields the following result. Fact 2.10 Let A : X ⇒ X∗ be maximally monotone and (x, x∗) ∈ X × X∗. Assume that A is of type (BR) and that inf (a,a∗)∈gra Ahx−a, x∗ −a∗i > −∞. Then x ∈ dom A and x∗ ∈ ran A. Additionally, Fact 2.11 (Marques Alves and Svaiter) (See [24, Theorem 1.4(4)] or [23].) Let A : X ⇒ X∗ be a maximally monotone operator. Assume that A is of type (NI). Then A is of type (BR). We shall also need some precise results about linear relations. The first two are elementary. Fact 2.12 (Cross) (See [14, Proposition I.2.8(a)].) Let A : X ⇒ Y be a linear relation. Then (∀(x, x∗) ∈ gra A) Ax = x∗ + A0. Lemma 2.13 Let A : X ⇒ X∗ be a linear relation. Assume that A∗ is monotone. Then ker A∗ ⊆ (ran A∗)⊥. Proof. Let x∗∗ ∈ ker A∗ and then (αx∗∗, 0) ∈ gra A∗, ∀α ∈ R. Then 0 ≤ hαx∗∗ + y∗∗, y∗i = αhx∗∗, y∗i + hy∗∗, y∗i, ∀(y∗∗, y∗) ∈ gra A∗, ∀α ∈ R. Hence hx∗∗, y∗i = 0, (ran A∗)⊥. ∀(y∗∗, y∗) ∈ gra A∗ and thus x∗∗ ∈ (ran A∗)⊥. Thus ker A∗ ⊆ (cid:4) 7 Fact 2.14 (See [4, Theorem 3.1].) Let A : X ⇒ X∗ be a maximally monotone linear relation. Then A is of type (D) if and only if A∗ is monotone. Fact 2.15 (See [41, Theorem 3.1].) Let A : X ⇒ X∗ be a maximally monotone linear relation, and let f : X → ]−∞, +∞] be a proper lower semicontinuous convex function with dom A ∩ int dom ∂f 6= ∅. Then A + ∂f is maximally monotone. Fact 2.16 (Simons) (See [35, Theorem 28.9].) Let Y be a Banach space, and L : Y → X be continuous and linear with ran L closed and ran L∗ = Y ∗. Let A : X ⇒ X∗ be monotone with dom A ⊆ ran L such that gra A 6= ∅. Then A is maximally monotone if, and only if A is ran L -- saturated and L∗AL is maximally monotone. Theorem 2.17 Let Y be a Banach space, and L : Y → X be an isomorphism into X. Let T : Y ⇒ Y ∗ be monotone. Then T is maximally monotone if, and only if (L∗)−1T L−1, mapping X into X∗, is maximally monotone. Proof. Let A = (L∗)−1T L−1. Then dom A ⊆ ran L. Since L is an isomorphism into X, ran L is closed. By [26, Theorem 3.1.22(b)] or [15, Exercise 2.39(i), page 59], ran L∗ = Y ∗. Hence gra(L∗)−1T L−1 6= ∅ if and only if gra T 6= ∅. Clearly, A is monotone. Since {0} × (ran L)⊥ ⊆ gra(L∗)−1 and then by Fact 2.12, A = (L∗)−1T L−1 is ran L -- saturated. By Fact 2.16, A = (L∗)−1T L−1 is maximally monotone if and only if L∗AL = T is maximally monotone. (cid:4) The following consequence will allow us to construct maximally monotone operators that are not of type (D) in a variety of non-reflexive Banach spaces. Corollary 2.18 (Subspaces) Let Y be a Banach space, and L : Y → X be an isomorphism into X. Let T : Y ⇒ Y ∗ be maximally monotone. Assume that T is not of type (D). Then (L∗)−1T L−1 is a maximally monotone operator mapping X into X∗ that is not of type (D). In particular, every Banach subspace of a type (D) space is of type (D). Proof. By Theorem 2.17, (L∗)−1T L−1 is maximally monotone. By Fact 2.7, there exists (y∗∗0 , y∗0) ∈ Y ∗∗ × Y ∗ such that (6) sup (b,b∗)∈gra T(cid:8)hy∗∗0 , b∗i + hy∗0, bi − hb, b∗i(cid:9) < hy∗∗0 , y∗0i. By [26, Theorem 3.1.22(b)] or [15, Exercise 2.39(i), page 59], ran L∗ = Y ∗ and thus there exists x∗0 ∈ X∗ such that L∗x∗0 = y∗0. Let A = (L∗)−1T L−1. Then we have sup (a,a∗)∈gra A(cid:8)hL∗∗y∗∗0 , a∗i + hx∗0, ai − ha, a∗i(cid:9) (Ly,a∗)∈gra A(cid:8)hy∗∗0 , L∗a∗i + hx∗0, Lyi − hLy, a∗i(cid:9) = sup 8 = = sup sup (Ly,a∗)∈gra A(cid:8)hy∗∗0 , L∗a∗i + hL∗x∗0, yi − hy, L∗a∗i(cid:9) (Ly,a∗)∈gra A(cid:8)hy∗∗0 , L∗a∗i + hy∗0, yi − hy, L∗a∗i(cid:9) (y,y∗)∈gra T(cid:8)hy∗∗0 , y∗i + hy∗0, yi − hy, y∗i(cid:9) (by (Ly, a∗) ∈ gra A ⇔ (y, L∗a∗) ∈ gra T ) = sup < hy∗∗0 , y∗0i = hL∗∗y∗∗0 , x∗0i. (by (6)) (7) Thus A is not of type (NI) and hence A = (L∗)−1T L−1 is not of type (D) by Fact 2.7. (cid:4) Note that it follows that X is of type (D) whenever X∗∗ is. 3 Main result We start with several technical tools. To relate Fitzpatrick functions and skew operators we have: Lemma 3.1 Let A : X ⇒ X∗ be a skew linear relation. Then (8) FA = ιgra(−A∗)∩X×X ∗. Proof. Let (x0, x∗0) ∈ X × X∗. We have FA(x0, x∗0) = sup (x,x∗)∈gra A = sup {h(x∗0, x0), (x, x∗)i − hx, x∗i} h(x∗0, x0), (x, x∗)i (x,x∗)∈gra A = ι(gra A)⊥(x∗0, x0) = ιgra(−A∗)(x0, x∗0) = ιgra(−A∗)∩X×X ∗(x0, x∗0). Hence (8) holds. (cid:4) To produce operators not of type (D) but that are of (BR) we exploit: Lemma 3.2 Let A : X ⇒ X∗ be a maximally monotone and linear skew operator. Assume that gra(−A∗) ∩ X × X∗ ⊆ gra A. Then A is of type (BR). 9 Proof. Let α, β > 0 and (x, x∗) ∈ X × X∗ be such that inf (a,a∗)∈gra Ahx − a, x∗ − a∗i > −αβ. Since A is skew, we have (9) inf (a,a∗)∈gra A hx, x∗i − [hx, a∗i + ha, x∗i] = inf (a,a∗)∈gra A hx − a, x∗ − a∗i > −αβ. Thus, hx, a∗i + ha, x∗i = 0, ∀(a, a∗) ∈ gra A and hence (x, x∗) ∈ gra(−A∗). Then by assump- tion, (x, x∗) ∈ gra A. Taking (b, b∗) = (x, x∗), we have kb − xk < α and kb∗ − x∗k < β. Hence A is of type (BR). (cid:4) Corollary 3.3 Let A : X ⇒ X∗ be a maximally monotone and linear skew operator that is not of type (D). Assume that A is unique. Then gra A = gra(−A∗) ∩ X × X∗ and so A is of type (BR). Proof. Apply Fact 2.8, Lemma 3.1 and Lemma 3.2 directly. (cid:4) Proposition 3.4 Let A : X ⇒ X∗ be maximally monotone. Assume that A is of type (NI) and that there exists e ∈ X∗ such that hx∗, xi ≥ he, xi2, ∀(x, x∗) ∈ gra A. Then e ∈ conv ran A. Proof. Suppose e 6∈ conv ran A. Then by the Separation Theorem, there exists x∗∗0 ∈ X∗∗ such that he − x∗, x∗∗0 i ≥ 1 for all x∗ ∈ ran A. Then we have hx∗ − e, x − x∗∗0 i = he − x∗, x∗∗0 i + hx∗ − e, xi, ∀(x, x∗) ∈ gra A ≥ 1 + he, xi2 − he, xi 3 4 t2 − t + 1 = ≥ min t∈R . Thus A is not of type (NI), which contradicts the assumption. (cid:4) The proof of the following result was partially inspired by that [12, Proposition 2.2]. Proposition 3.5 Let A : X ⇒ X∗ be a maximally monotone linear relation. Assume that there exists e ∈ X∗ such that e /∈ ran A and that hx∗, xi ≥ he, xi2, ∀(x, x∗) ∈ gra A. Then A is neither of type (D) nor unique. 10 Proof. By Proposition 3.4, A is not of type (NI) and hence A is not of type (D) by Fact 2.7. Similar to the proof of Proposition 3.4, there exists x∗∗0 ∈ X∗∗ such that he, x∗∗0 i ≥ 1 and x∗∗0 ∈ (ran A)⊥. Let 0 < α < 2. Then we have hx∗ − αe, x − 1 αx∗∗0 i = hαe − x∗, 1 αx∗∗0 i + hx∗ − αe, xi, ∀(x, x∗) ∈ gra A t2 − αt + 1 ≥ 1 + he, xi2 − αhe, xi ≥ min t∈R = 1 − > 0. α2 4 Thus for every 0 < α < 2, ( 1 αx∗∗0 , αe) ∈ X∗∗ × X∗ is monotonically related to gra A. Take 0 < α1 < α2 < 2. Then by Zorn's Lemma, we have a maximally monotone extension, A1 : X∗∗ ⇒ X∗ such that gra A1 ⊇ gra A ∪ {( 1 x∗∗0 , α1e, )}, and we can also obtain a α1 maximally monotone extension, A2 : X∗∗ ⇒ X∗ such that gra A2 ⊇ gra A ∪ {( 1 x∗∗0 , α2e)}. α2 Now we show gra A1 6= gra A2. Suppose to the contrary that gra A1 = gra A2. Then by the monotonicity of A1, we have (10) On the other hand, h 1 α1 x∗∗0 − 1 α2 x∗∗0 , α1e − α2ei ≥ 0. h 1 α1 x∗∗0 − 1 α2 x∗∗0 , α1e − α2ei = (α1 − α2)( 1 α1 < (α1 − α2)( 1 α1 − 1 α2 − 1 α2 )hx∗∗0 , ei ) < 0, which contradicts (10). Hence gra A1 6= gra A2 and thus A is not unique. (cid:4) We are now ready to establish our work-horse Theorem 3.6, which allows us to construct various maximally monotone operators -- both linear and nonlinear -- that are not of type (D). The idea of constructing the operators in the following fashion is based upon [2, Theorem 5.1] and was stimulated by [12]. Theorem 3.6 (Predual constructions) Let A : X∗ → X∗∗ be linear and continuous. Assume that ran A ⊆ X and that there exists e ∈ X∗∗\X such that hAx∗, x∗i = he, x∗i2, ∀x∗ ∈ X∗. Let P and S respectively be the symmetric part and antisymmetric part of A. Let T : X ⇒ X∗ be defined by (11) gra T :=(cid:8)(−Sx∗, x∗) x∗ ∈ X∗, he, x∗i = 0(cid:9) =(cid:8)(−Ax∗, x∗) x∗ ∈ X∗, he, x∗i = 0(cid:9). Let f : X → ]−∞, +∞] be a proper lower semicontinuous and convex function. Set F := f ⊕ f∗ on X × X∗. Then the following hold. 11 (i) A is a maximally monotone operator on X∗ that is neither of type (D) nor unique. (ii) P x∗ = hx∗, eie, ∀x∗ ∈ X∗. (iii) T is maximally monotone and skew on X. (iv) gra T ∗ = {(Sx∗ + re, x∗) x∗ ∈ X∗, r ∈ R}. (v) −T is not maximally monotone. (vi) T is not of type (D). (vii) FT = ιC, where (12) C := {(−Ax∗, x∗) x∗ ∈ X∗}. (viii) T is not unique. (ix) T is not of type (BR). (x) If dom T ∩ int dom ∂f 6= ∅, then T + ∂f is maximally monotone. (xi) F and FT are BC -- functions on X × X∗. (xii) Moreover, [λ>0 λ(cid:0)PX ∗(dom FT ) − PX ∗(dom F )(cid:1) = X∗, while, assuming that there exists (v0, v∗0) ∈ X × X∗ such that (13) f∗(v∗0) + f∗∗(v0 − A∗v∗0) < hv0, v∗0i, then FT (cid:3)1F is not a BC -- function. (xiii) Assume that (cid:2)ran A −Sλ>0 λ dom f(cid:3) is a closed subspace of X and that ∅ 6= dom f∗∗ ◦ A∗X ∗ " {e}⊥. Then T + ∂f is not of type (D). (xiv) Assume that dom f∗∗ = X∗∗. Then T + ∂f is a maximally monotone operator that is not of type (D). 12 Proof. (i): Clearly, A has full domain. Since A is monotone and continuous, A is maximally monotone. By the assumptions that e /∈ X and ran A ⊆ X = X, then by Proposition 3.5, A is neither of type (D) nor unique. See also [1, Theorem 14.2.1 and Theorem 13.2.3] for alternative proof of that A is not of type (D). (ii): Now we show that (14) P x∗ = hx∗, eie, ∀x∗ ∈ X∗. Since h·, eie = ∂( 1 Clearly, A − h·, eie is skew. Then (14) holds. 2 h·, ei2) and by [29, Theorem 5.1], h·, eie is a symmetric operator on X∗. (iii): Let x∗ ∈ X∗ with he, x∗i = 0. Then we have Sx∗ = hx∗, eie + Sx∗ = P x∗ + Sx∗ = Ax∗ ∈ ran A ⊆ X. Thus (11) holds and T is well defined. We have S is skew and hence T is skew. Let (z, z∗) ∈ X × X∗ be monotonically related to gra T . By Fact 2.6, we have 0 = hz, x∗i + h−Sx∗, z∗i = hz + Sz∗, x∗i, ∀x∗ ∈ {e}⊥. Thus by Fact 2.1, we have z + Sz∗ ∈ ({e}⊥)⊥ = span{e} and then (15) z = −Sz∗ + κe, ∃κ ∈ R. By (0, 0) ∈ gra T , (16) κhz∗, ei = h−Sz∗ + κe, z∗i = hz, z∗i ≥ 0. Then by (15) and (ii), (17) Az∗ = P z∗ + Sz∗ = P z∗ + κe − z = [hz∗, ei + κ] e − z. By the assumptions that z ∈ X, Az∗ ∈ X and e /∈ X, [hz∗, ei + κ] = 0 by (17). Then by (16), we have hz∗, ei = κ = 0 and thus (z, z∗) ∈ gra T by (15). Hence T is maximally monotone. (iv): Let (x∗∗0 , x∗0) ∈ X∗∗ × X∗. Then we have (x∗∗0 , x∗0) ∈ gra T ∗ ⇔ hx∗0, Sx∗i + hx∗, x∗∗0 i = 0, ⇔ hx∗, x∗∗0 − Sx∗0i = 0, ⇔ x∗∗0 − Sx∗0 ∈ ({e}⊥)⊥ = span{e} (by Fact 2.1) ⇔ x∗∗0 − Sx∗0 = re, ∀x∗ ∈ {e}⊥ ∀x∗ ∈ {e}⊥ ∃r ∈ R. 13 Thus gra T ∗ = {(Sx∗ + re, x∗) x∗ ∈ X∗, r ∈ R}. (v): Since e /∈ X, we have e 6= 0. Then there exists z∗ ∈ X∗ such that z∗ 6∈ {e}⊥. Then by (ii)&(iv) and the assumption that ran A ⊆ X, we have (Az∗, z∗) = (Sz∗ + he, z∗ie, x∗) ∈ gra T ∗ ∩ X × X∗. Thus we have hAz∗ − x, z∗ − x∗i = hAz∗, z∗i − [hAz∗, x∗i + hx, z∗i] + hx, x∗i = hAz∗, z∗i ≥ 0, ∀(x, x∗) ∈ gra(−T ). Hence (Az∗, z∗) is monotonically related to gra(−T ). Since z∗ /∈ ran(−T ), (Az∗, z∗) /∈ gra(−T ) and then −T is not maximally monotone. (vi): By (iv), T ∗ is not monotone. Then by Fact 2.14, T is not of type (D). (vii): By (iv), we have (z, z∗) ∈ gra(−T ∗) ∩ X × X∗ ⇔ (z, z∗) = (−Sz∗ − re, z∗), ⇔ (z, z∗) = (−Sz∗ − hz∗, eie + [hz∗, ei − r] e, z∗), ⇔ (z, z∗) = (−Az∗ + [hz∗, ei − r] e, z∗), ⇔ (z, z∗) = (−Az∗, z∗), ⇔ (z, z∗) ∈ {(−Ax∗, x∗) x∗ ∈ X∗} = C. z ∈ X, ∃r ∈ R, z∗ ∈ X∗ z ∈ X, ∃r ∈ R, z∗ ∈ X∗ z ∈ X, ∃r ∈ R, z∗ ∈ X∗ (by (ii)) hz∗, ei = r (since z, Az∗ ∈ X and e /∈ X), ∃r ∈ R, z∗ ∈ X∗ Thus by Lemma 3.1, we have FT = ιC. (viii): Since e /∈ X, we have e 6= 0. Then there exists z∗ ∈ X∗ such that z∗ 6∈ {e}⊥. Thus z∗ /∈ ran T . By (vii), z∗ ∈ PX ∗ [dom FT ]. Thus, gra T 6= dom FT . Then by (vi) and Fact 2.8, T is not unique. (ix): Suppose to the contrary that T is of type (BR). Let z∗ be as in the proof of (viii). Then by Lemma 3.1 and (vii), we have (−Az∗, z∗) ∈ gra(−T ∗) ∩ X × X∗ and then inf (a,a∗)∈gra T h−Az∗ − a, z∗ − a∗i = h−Az∗, z∗i > −∞. Then Fact 2.10 shows z∗ ∈ ran T , which contradicts that z∗ /∈ {e}⊥ = ran T . Hence T is not of type (BR). (x): Apply (iii) and Fact 2.15. (xi): Clearly, F is a BC -- function. By (iii) and Fact 2.3, we see that FT is a BC -- function. 14 (xii): By (vii), we have (18) [λ>0 λ(cid:0)PX ∗(dom FT ) − PX ∗(dom F )(cid:1) = X∗. Then for every (x, x∗) ∈ X × X∗ and u ∈ X, by (xi), FT (x − u, x∗) + F (u, x∗) = FT (x − u, x∗) + (f ⊕ f∗)(u, x∗) ≥ hx − u, x∗i + hu, x∗i = hx, x∗i. Hence (19) (FT (cid:3)1F )(x, x∗) ≥ hx, x∗i > −∞. Then by (18), (19) and Fact 2.5, (FT (cid:3)1F )∗(v∗0, v0) = min x∗∗∈X ∗∗ F ∗T (v∗0, x∗∗) + F ∗(v∗0, v0 − x∗∗) ≤ F ∗T (v∗0, A∗v∗0) + F ∗(v∗0, v0 − A∗v∗0) = 0 + F ∗(v∗0, v0 − A∗v∗0) (by (vii)) = (f ⊕ f∗)∗(v∗0, v0 − A∗v∗0) = (f∗ ⊕ f∗∗)(v∗0, v0 − A∗v∗0) = f∗(v∗0) + f∗∗(v0 − A∗v∗0) < hv∗0, v0i (by (13)). Hence FT (cid:3)1F is not a BC -- function. (xiii): By the assumption, there exists x∗0 ∈ dom f∗∗ ◦ A∗X ∗ such that he, x∗0i 6= 0. Let . By [42, Theorem 2.4.4(iii)]), there exists y∗∗∗0 ∈ ∂ε0f∗∗(A∗x∗0). By [42, Theo- ε0 = he,x∗ 0i2 2 rem 2.4.2(ii)]), (20) f∗∗(A∗x∗0) + f∗∗∗(y∗∗∗0 ) ≤ hA∗x∗0, y∗∗∗0 i + ε0. Then by [35, Lemma 45.9] or the proof of [30, Eq.(2.5) in Proposition 1], there exists y∗0 ∈ X∗ such that (21) f∗∗(A∗x∗0) + f∗(y∗0) < hA∗x∗0, y∗0i + 2ε0. Let z∗0 = y∗0 + x∗0. Then by (21), we have f∗∗(A∗x∗0) + f∗(z∗0 − x∗0) < hA∗x∗0, z∗0 − x∗0i + 2ε0 (22) = hA∗x∗0, z∗0i − hA∗x∗0, x∗0i + 2ε0 = hA∗x∗0, z∗0i − hx∗0, Ax∗0i + 2ε0 = hA∗x∗0, z∗0i − 2ε0 + 2ε0 = hA∗x∗0, z∗0i. 15 Then for every (x, x∗) ∈ X × X∗ and u∗ ∈ X, by (xi), FT (x, x∗ − u∗) + F (x, u∗) = FT (x, x∗ − u∗) + (f ⊕ f∗)(x, u∗) ≥ hx, x∗ − u∗i + hx, u∗i = hx, x∗i. Hence (23) (FT (cid:3)2F )(x, x∗) ≥ hx, x∗i > −∞. Then by (23), (vii) and Fact 2.4, (FT (cid:3)2F )∗(z∗0, A∗x∗0) = min y∗∈X ∗ F ∗T (y∗, A∗x∗0) + F ∗(z∗0 − y∗, A∗x∗0) ≤ F ∗T (x∗0, A∗x∗0) + F ∗(z∗0 − x∗0, A∗x∗0) = 0 + F ∗(z∗0 − x∗0, A∗x∗0) (by (vii)) = (f ⊕ f∗)∗(z∗0 − x∗0, A∗x∗0) = f∗(z∗0 − x∗0) + f∗∗(A∗x∗0) < hz∗0, A∗x∗0i (by (22)). (24) Let F0 : X × X∗ → ]−∞, +∞] be defined by (25) (x, x∗) 7→ hx, x∗i + ιgra(T +∂f )(x, x∗). Clearly, FT (cid:3)2F ≤ F0 on X × X∗ and thus (FT (cid:3)2F )∗ ≥ F ∗0 on X∗ × X∗∗. By (24), F ∗0 (z∗0, A∗x∗0) < hz∗0, A∗x∗0i. Hence T + ∂f is not of type (NI) and thus T + ∂f is not of type (D) by Fact 2.7. (xiv): Since dom f∗∗ = X∗∗, dom f = X by [42, Theorem 2.3.3]. By dom f∗∗ = X∗∗ again, (cid:4) dom f∗∗ ◦ A∗αX ∗ = X∗ " {α}⊥. Then apply (x)&(xiii) directly. Remark 3.7 (Grothendieck spaces [11]) In light of part (xiii) of the previous theorem), we record that for a closed convex function dom f = X implies dom f∗∗ = X∗∗ ⇔ X is a Grothendieck space. All reflexive spaces are Grothendieck spaces while all non-reflexive Grothendieck spaces (such as L∞[0, 1]) contain an isomorphic copy of c0. ♦ We are now ready to exploit Theorem 3.6. 4 Examples and applications We begin with the case of c0 and its dual ℓ1. 16 4.1 Applications to c0 Example 4.1 (c0) Let X := c0, with norm k · k∞ so that X∗ = ℓ1 with norm k · k1, and X∗∗ = ℓ∞ with its second dual norm k · k∗. Let α := (αn)n∈N ∈ ℓ∞ with lim sup αn 6= 0, and let Aα : ℓ1 → ℓ∞ be defined by (26) (Aαx∗)n := α2 nx∗n + 2Xi>n αnαix∗i , ∀x∗ = (x∗n)n∈N ∈ ℓ1. Now let Pα and Sα respectively be the symmetric part and antisymmetric part of Aα. Let Tα : c0 ⇒ X∗ be defined by gra Tα :=(cid:8)(−Sαx∗, x∗) x∗ ∈ X∗, hα, x∗i = 0(cid:9) =(cid:8)(−Aαx∗, x∗) x∗ ∈ X∗, hα, x∗i = 0(cid:9) =(cid:8)(cid:0)(−Xi>n αnαix∗i +Xi<n αnαix∗i )n, x∗(cid:1) x∗ ∈ X∗, hα, x∗i = 0(cid:9). (27) Then (i) hAαx∗, x∗i = hα, x∗i2, ∀x∗ = (x∗n)n∈N ∈ ℓ1 and(27) is well defined. (ii) Aα is a maximally monotone operator on ℓ1 that is neither of type (D) nor unique. (iii) Tα is a maximally monotone operator on c0 that is not of type (D). (iv) −Tα is not maximally monotone. (v) Tα is neither unique nor of type (BR). (vi) FTα (cid:3)1(k · k ⊕ ιBX ∗ ) is not a BC -- function. (vii) Tα + ∂k · k is a maximally monotone operator on c0(N) that is not of type (D). (viii) If 1√2 < kαk∗ ≤ 1, then FTα (cid:3)1( 1 2k · k2 ⊕ 1 2k · k2 1) is not a BC -- function. (ix) For λ > 0, Tα + λJ is a maximally monotone operator on c0 that is not of type (D). (x) Let λ > 0 and a linear isometry L mapping c0 to a subspace of C[0, 1] be given. Then both (L∗)−1(Tα +∂k·k)L−1 and (L∗)−1(Tα +λJ)L−1 are maximally monotone operators that are not of type (D). Hence neither c0 nor C[0, 1] is of type (D). (xi) Every Banach space that contains an isomorphic copy of c0 is not of type (D). 17 (xii) Let G : ℓ1 → ℓ∞ be Gossez's operator [20] defined by (cid:0)G(x∗)(cid:1)n :=Xi>n x∗i −Xi<n x∗i , ∀(x∗n)n∈N ∈ ℓ1. Then Te : c0 ⇒ ℓ1 as defined by gra Te := {(−G(x∗), x∗) x∗ ∈ ℓ1, hx∗, ei = 0} is a maximally monotone operator that is not of type (D), where e := (1, 1, . . . , 1, . . .). (xiii) Moreover, G is a unique maximally monotone operator that is not of type (D), but G is of type (BR). Proof. We have α /∈ c0. Since α = (αn)n∈N ∈ ℓ∞ and kAαk ≤ 2kαk2, Aα is linear and continuous and ran Aα ⊆ c0 ⊆ ℓ∞. αnαix∗i ) αnαix∗nx∗i αnαix∗nx∗i x∗n(α2 α2 nx∗n α2 nx∗n nx∗n + 2Xi>n 2 + 2Xn Xi>n 2 +Xn6=i hAαx∗, x∗i =Xn =Xn =Xn = (Xn (i): We have (28) αnx∗n)2 = hα, x∗i2, ∀x∗ = (x∗n)n∈N ∈ ℓ1. Then Theorem 3.6(ii) shows that the symmetric part Pα of Aα is Pαx∗ = hα, x∗iα (for every x∗ ∈ ℓ1). Thus, the skew part Sα of Aα is (Sαx∗)n = (Aαx∗)n − (Pαx∗)n (29) αnαix∗i = α2 nx∗n + 2Xi>n αnαix∗i −Xi<n =Xi>n αnαix∗i −Xi≥1 αnαix∗i . Then by Theorem 3.6, (27) is well defined. (ii): Apply (i) and Theorem 3.6(i) directly. (iii): Combine Theorem 3.6(iii)&(vi). 18 (iv): Apply Theorem 3.6(v) directly. (v): Apply Theorem 3.6(viii)&(ix). (vi) Since α 6= 0, there exists i0 ∈ N such that αi0 6= 0. Let ei0 := (0, . . . , 0, 1, 0, . . .), i.e., the i0th is 1 and the others are 0. Then by (29), we have (30) Then (31) Sαei0 = αi0(α1, . . . , αi0−1, 0, −αi0+1, −αi0+2, . . .). A∗ei0 = Pαei0 − Sαei0 = αi0(0, . . . , 0, αi0, 2αi0+1, 2αi0+2, . . .). Now set v∗0 := ei0 and v0 := 3kαk2 v0 − A∗v∗0 = 3kαk2 ∗ ∗ei0. Thus by (31), ei0 − A∗ei0 = (0, . . . , 0, 3kαk2 ∗ − α2 (32) i0, −2αi0αi0+1, −2αi0αi0+2, . . .). Let f := k · k on X = c0. Then f∗ = ιBX ∗ by [42, Corollary 2.4.16]. We have f∗(v∗0) + f∗∗(v0 − A∗ei0) = ιBX ∗ (ei0) + kv0 − A∗ei0k∗ = k3kαk∗ei0 − A∗ei0k∗ < 3kαk2 ∗ = hv0, v∗0i. (by (32)) Hence by Theorem 3.6(xii), FTα (cid:3)1(k · k ⊕ ιBX ∗ ) is not a BC -- function. (vii): Let f := k · k on X. Since dom f∗∗ = X∗∗. Then apply Theorem 3.6(xiv). (viii): By 1√2 < kαk∗ ≤ 1, take αi02 > 1 2 . Let ei0 be defined as in the proof of (vi). Then take v∗1 := 1 2ei0 and v1 :=(cid:0)1 + 1 2α2 i0(cid:1)ei0. By (31), we have (33) v1 − A∗v∗1 = (0, . . . , 0, 1, −αi0αi0+1, −αi0αi0+2, . . .). Since αi0αj ≤ kαk2 ∗ ≤ 1, ∀j ∈ N, then (34) kv1 − A∗v∗1k∗ ≤ 1. Let f := 1 2k · k2 on X = c0. Then f∗ = 1 2 k · k2 1 and f∗∗ = 1 2 k · k2 ∗. We have f∗(v∗1) + f∗∗(v1 − A∗v∗1) = 1 2kv∗1k2 1 + 1 2kv1 − A∗v∗1k2 ∗ 19 2 ≤ 1 α2 i0 8 + 1 4 + 1 < 2 = hv∗1, v1i. (by (34)) (since α2 i0 > 1/2) Hence by Theorem 3.6(xii), FTα (cid:3)1( 1 2k · k2 ⊕ 1 2k · k2 ∗) is not a BC -- function. (ix): Let λ > 0 and f := λ 2 k · k2 on X = c0. Then f∗∗ = λ . Then apply Theo- rem 3.6(xiv). 2 k · k2 ∗ (x): Since c0 is separable by [26, Example 1.12.6] or [15, Proposition 1.26(ii)], by Fact 2.2, there exists a linear operator L : c0 → C[0, 1] that is an isometry from c0 to a subspace of C[0, 1]. Then combine (vii)&(ix) and Corollary 2.18. (xi) Combine (iii) (or (vii) or (ix)) and Corollary 2.18. (xii): To obtain the result on Te, directly apply (iii) (or see [2, Example 5.2]). (xiii) Now −G is type (D) but G is not [2]. To see that G is unique, note that −G∗ is monotone by Fact 2.14 and so provides the unique maximal extension. Since G is skew and continuous, clearly, −G∗x∗ = Gx∗, ∀x∗ ∈ ℓ1. Then Lemma 3.2 implies that G is of type (BR). The uniqueness of G was also verified in [1, Example 14.2.2]. (cid:4) Remark 4.2 The maximal monotonicity of the operator Te in Example 4.1(xii) was also verified by Voisei and Zalinescu in [39, Example 19] and later a direct proof given by Bueno and Svaiter in [12, Lemma 2.1]. Herein we have given a more concise proof of above results. Bueno and Svaiter also showed that Te is not of type (D) in [12]. They also showed that each Banach space that contains an isometric (isomorphic) copy of c0 is not of type (D) in [12]. Example 4.1(xi) recaptures their result, while Example 4.1(vi)&(viii) provide a negative answer to Simons' [35, Problem 22.12]. ♦ Remark 4.3 (The continuous case) We recall that a Banach space X is a conjugate monotone space if every continuous linear monotone operator on X has a monotone conju- gate. In particular this holds if every continuous linear monotone operator on X is weakly compact. In consequence, a Banach lattice X contains a complemented copy of ℓ1 if and only if it admits a non (D) continuous linear monotone operator, on using Fact 2.14 along with [2, Remark 5.5] and [2, Examples. 5.2 and 5.3]. Thus, in lattices such as c0, c and C[0, 1] only discontinuous linear monotone operators ♦ can fail to be of type (D). This subtlety escaped the current authors for fifteen years. We now turn to a broader class of spaces: 20 4.2 Applications to more general nonreflexive spaces Our results below are facilitated by making use of Schauder basis structure [16]. Definition 4.4 We say (en, e∗n)n∈N in X × X∗ is a Schauder basis of X if for every x ∈ X there exists a unique sequence (αn)n∈N in R such that x = Pn≥1 αnen, where αn = hx, e∗ni and hei, e∗j i = δi,j, ∀i, j ∈ N. Definition 4.5 Let (en, e∗n)n∈N in X × X∗ be a Schauder basis of X. We say the basis is shrinking if span{e∗n n ∈ N} = X∗. In particular, a Banach space with a shrinking basis has a separable dual and so is an Asplund space [16]. ∀x ∈ X; i=1hx, e∗i iei = x, Fact 4.6 (See [16, Lemma 4.7(iii) and Facts 4.11(ii)&(iii)] or [15, Lemma 6.2(iii) and Facts 6.6(ii)&(iii)] .) Let (en, e∗n)n∈N in X × X∗ be a Schauder basis of X. Then (i) limnPn (ii) Pn i=1hx∗, eiie∗i (iii) (e∗n, en)n∈N in X∗ × X∗∗ is a Schauder basis of span{e∗n n ∈ N}. Lemma 4.7 Let (en, e∗n)n∈N in X × X∗ be a Schauder basis of X. Then e∗n lim inf n∈N kenk > 0. i=1hx∗, eiie∗i weak∗ converges to x∗, written as, Pn w⇁ x∗, ∀x∗ ∈ X∗; w⇁ 0 whenever Proof. Let x ∈ X. Since khx, e∗nienk → 0 because of Fact 4.6(i), and since lim inf n∈N kenk > 0, we have hx, e∗ni → 0. Hence e∗n (cid:4) w⇁ 0 as n → ∞. The proof of Example 4.8(i) was inspired by that [3, Proposition 3.5]. Example 4.8 (Schauder basis) Let (en, e∗n)n∈N in X × X∗ be a Schauder basis of X. Assume that for some e ∈ X∗∗ we have n (35) Let A : X ⇒ X∗ be defined by w⇁ e ∈ X∗∗. ei Xi=1 gra A :=(cid:26)(cid:18)Xn (cid:0) −Xi>n hei, y∗i +Xi<n hei, y∗i(cid:1)en, y∗(cid:19) ∈ X × X∗ y∗ ∈ {e}⊥(cid:27). Assume that lim inf kenk > 0. Then the following hold. 21 (i) A is a maximally monotone and linear skew operator. (ii) A is not of type (BR). (iii) A is not of type (D). (iv) A is not unique. (v) Every Banach space containing a copy of X is not of type (D). Proof. (i): First, we show A is skew. Let (y, y∗) ∈ gra A. Then he, y∗i = 0 and ∞ Xn=1(cid:0) −Xi>n hei, y∗i +Xi<n y = (36) Thus, hei, y∗i(cid:1)en. By the assumption that Pn s :=Xi≥1 hei, y∗i = he, y∗i = 0. i=1 ei w⇁ e ∈ X∗∗, we have k k = lim h k = lim k hei, y∗i(cid:1)en, y∗i hei, y∗i(cid:1)en, y∗i hei, y∗i(cid:1)hen, y∗i hei, y∗i(cid:1)hen, y∗i hei, y∗i(cid:1)hen, y∗i hei, y∗i +Xi<n hei, y∗i +Xi<n hei, y∗i +Xi<n hei, y∗i −Xi<n hei, y∗i +Xi≥n hei, y∗i + he2, y∗iXi≥2 hy, y∗i = hXn (cid:0) −Xi>n Xn=1(cid:0) −Xi>n Xn=1(cid:0) −Xi>n Xn=1(cid:0)Xi>n Xn=1(cid:0) Xi≥n+1 k (cid:18)he1, y∗iXi≥1 + he1, y∗iXi≥2 hei, y∗i + he2, y∗iXi≥3 = − lim = − lim = − lim k k k k (37) (by Fact 4.6(i)) (by (36)) hei, y∗i + · · · + hek, y∗iXi≥k hei, y∗i hei, y∗i + · · · + hek, y∗i Xi≥k+1 hei, y∗i(cid:19) = − lim k (cid:18)she1, y∗i + (s − he1, y∗i)he2, y∗i + · · · + (s − hei, y∗i)hek, y∗i k−1 Xi=1 + (s − he1, y∗i)he1, y∗i +(cid:0)s − 2 Xi=1 hei, y∗i(cid:1)he2, y∗i + · · · + (s − 22 k Xi=1 hei, y∗i)hek, y∗i(cid:19) + s hei, y∗i − hei, y∗i2 − he1, y∗ihe2, y∗i − · · · − 2 Xi=1 k−1 Xi=1 k−1 Xi=1 hei, y∗ihek, y∗i(cid:19) = − lim hei, y∗i − he1, y∗ihe2, y∗i − hei, y∗ihe3, y∗i − · · · − hei, y∗ihek, y∗i k Xi=1 k k (cid:18)s Xi=1 k "2s k Xi=1 = − lim hei, y∗i − ( k Xi=1 hei, y∗i)2# k Xi=1 = −(2s2 − s2) = −s2 = 0. (by (36)) Hence A is skew. To show maximality, let (x, x∗) ∈ X × X∗ be monotonically related to gra A. By Fact 2.6, we have (38) By (35), we have (39) hy∗, xi + hx∗, yi = 0, ∀(y, y∗) ∈ gra A. he, e∗ni =Xi≥1 hei, e∗ni = δn,n = 1, ∀n ∈ N. Let y∗ := −e∗1 + e∗n (n ≥ 2) and y := −e1 − 2Pn−1 Hence y∗ ∈ {e}⊥ and (y, y∗) ∈ gra A. Using (38), i=2 ei − en. By (39), we have he, y∗i = 0. − hx, e∗1i + hx, e∗ni − hx∗, e1i − hx∗, eni − 2 hx∗, eii = 0. n−1 Xi=2 Thus, we have (40) hx, e∗ni = hx, e∗1i − hx∗, e1i + hx∗, eni + 2 hx∗, eii. n−1 Xi=1 As Pi≥1hei, z∗i = he, z∗i(∀z∗ ∈ X∗), we have hx∗, eni → 0. Hence, by Lemma 4.7 -- since lim inf kenk > 0 -- and (40), (41) hx∗, eii = hx, e∗1i − hx∗, e1i. −2Xi≥1 Next we show −2Pi≥1hx∗, eii = hx, e∗1i − hx∗, e1i = 0. Let t = Pi≥1hx∗, eii. Then by (40) and (41), hx, e∗nien x =Xn≥1 23 hx∗, eii + hx∗, eni(cid:19)en =Xn≥1(cid:18) − 2Xi≥1 =Xn≥1(cid:18) − 2Xi≥n =Xn≥1(cid:18) −Xi≥n =Xn≥1(cid:18) −Xi≥n hx∗, eii + 2Xi<n hx∗, eii + hx∗, eni(cid:19)en hx∗, eii −Xi≥n hx∗, eii − Xi≥n+1 hx∗, eii + hx∗, eni(cid:19)en hx∗, eii(cid:19)en. (42) Using (0, 0) ∈ gra A, as in the proof of (37), shows 0 ≥ −hx∗, xi = hXn≥1(cid:18)Xi≥n Xn=1(cid:18)Xi≥n hx∗, eii + Xi≥n+1 hx∗, eii + Xi≥n+1 hx∗, eii(cid:19)en, x∗i hx∗, eii(cid:19)en, x∗i k = lim h k = 2t2 − t2 = t2. Hence t = 0. By (42), x =Xn≥1(cid:18) −Xi>n hx∗, eii +Xi<n hx∗, eii(cid:19)en. Hence (x, x∗) ∈ gra A. Thus, A is maximally monotone. (ii): Suppose to the contrary that A is of type (BR). One checks that (e1, e∗1) ∈ gra A∗ and he, e∗1i = limnhPn i=1 ei, e∗1i = 1. Thus, (e1, −e∗1) ∈ gra(−A∗) ∩ X × X∗ and −e∗1 /∈ {e}⊥. Since ran A ⊆ {e}⊥, −e∗1 /∈ ran A. Then inf (a,a∗)∈gra Ahe1 − a, −e∗1 − a∗i = he1, −e∗1i = −1 > −∞. Then by Fact 2.10, −e∗1 ∈ ran A, which contradicts that −e∗1 /∈ ran A. Hence A is not of type (BR). (iii): By Fact 2.11 and (ii), A is not of type (NI) and hence A is not of type (D) by Fact 2.7. Alternative Proof: Clearly, (e, 0) ∈ gra A∗ and thus e ∈ ker A∗. By the proof of (ii), (e1, e∗1) ∈ gra A∗ and he, e∗1i = 1. Hence e /∈ (ran A∗)⊥. Hence A∗ is not monotone by Lemma 2.13. Then Fact 2.14 shows A is not of type (D). (iv): Apply (iii)&(ii) and Corollary 3.3 directly. (v): Combine (i)&(iii) and Corollary 2.18. (cid:4) We shall especially exploit the lovely properties of the James space: 24 Definition 4.9 The James space, J, consists of all the sequences x = (xn)n∈N in c0 with the finite norm kxk := sup n1<···<nk(cid:0)(xn1 − xn2)2 + (xn2 − xn3)2 + · · · + (xnk−1 − xnk )2(cid:1) Fact 4.10 (See [16, page 205] or [15, Claim, page 185].) The space J is constructed to be of codimension-one in J∗∗. Indeed, J∗∗ = J⊕span{e} where e := (1, 1, . . . , 1, . . .) is the constant sequence in c(N) ⊂ ℓ∞. Thus, J is a separable Asplund space, equivalently J∗ is separable [11, 16, 15], and non-reflexive. Inter alia, the basis (en, e∗n)n∈N is a shrinking Schauder basis in J and (e∗n, en)n∈N is a basis for J∗, where en = (0, . . . , 0, 1, 0, . . .), i.e., the nth is 1 and the others are 0. 1 2 . Corollary 4.11 (James space) Let X be the James space, J. Let en be defined as in Fact 4.10, and let A be defined as in Example 4.8. Then A is a maximally monotone and skew operator that is neither of type (BR) nor unique and so A is not of type (D). Hence, every Banach space that contains an isomorphic copy of J is not of type (D). Proof. To apply Example 4.8 we need only verify that (35) holds. To see this is so, we note the Banach-Alaoglu theorem and [16, Proposition 3.103, page 128] or [15, Proposition 3.24, that (cid:0)Pn page 72], we have the vector e = (1, 1, . . . , 1, . . .) is the unique w∗ limit of(cid:0)Pn i=1 ei(cid:1)n∈N lies in BJ∗∗ -- directly from the definition of the norm in J. Now by i=1 ei(cid:1)n∈N. (cid:4) An easier version of the same argument leads to a recovery of part of Example 4.1: Corollary 4.12 (c0) Let X = c0. Let en be defined as in Fact 4.10 and e := (1, 1, . . . , 1, . . .). Let A be defined as in Example 4.8 (thus A = Te in Example 4.1(xii)). Then A is a maximally monotone and skew operator that is neither of type (BR) nor unique and so A is not of type (D). Hence, every Banach space that contains an isomorphic copy of c0 is not of type (D). We finish our set of core examples by dealing with the dual space J∗. defined by Example 4.13 (Shrinking Schauder basis) Let (en, e∗n)n∈N in X × X∗ be a shrinking w⇁ e for some e ∈ X∗∗. Let A : X∗ ⇒ X∗∗ be Schauder basis of X. Assume that Pn gra A =(cid:26)(y∗, y∗∗) ∈ X∗ × X∗∗ w⇁ y∗∗(cid:27). i=1 ei (43) k Xn=1(cid:0)Xi>n hei, y∗i −Xi<n hei, y∗i(cid:1)en Then A is a maximally monotone and linear skew operator, which is of type (BR). In particular, let (en)n∈N and e be defined as in Fact 4.10. Then A + h·, eie is a maximally monotone operator that is neither of type (D) nor unique; and every Banach space containing a copy of J∗ is not of type (D). 25 Proof. Again, we first show A is skew. Let (y∗, y∗∗) ∈ gra A. Then k Xn=1(cid:0)Xi>n hei, y∗i −Xi<n hei, y∗i(cid:1)en w⇁ y∗∗. By the assumption that Pn (44) i=1 ei w⇁ e ∈ X∗∗, we have hei, y∗i = he, y∗i. s :=Xi≥1 Thus, k hy∗∗, y∗i = lim k = lim k = lim k h k k Xn=1(cid:0)Xi>n Xn=1(cid:0)Xi>n Xn=1(cid:0) Xi≥n+1 Xn=1 k hei, y∗i −Xi<n hei, y∗i −Xi<n hei, y∗i +Xi≥n hei, y∗i(cid:1)en, y∗i hei, y∗i(cid:1)hen, y∗i hei, y∗i − s(cid:1)hen, y∗i Xn=1(cid:0) Xi≥n+1 hei, y∗i +Xi≥n k = −s lim k hen, y∗i + lim k = −s2 + (2s2 − s2) = 0 (as in the proof of (37)). (by (44)) hei, y∗i(cid:1)hen, y∗i Hence A is skew. Now we confirm maximality. Let (x∗, x∗∗) ∈ X∗ × X∗∗ be monotonically related to gra A. By Fact 2.6, we have (45) hy∗, x∗∗i + hx∗, y∗∗i = 0, ∀(y∗, y∗∗) ∈ gra A. Fix n ∈ N and set y∗ := e∗n. Then Pk Pk j=n+1 ej. By the assumption that Pk Xj=n+1 Xj=1 ej − n−1 k i=1 ei ej w⇁ 2 n−1 Xj=1 ej + en − e. j=1(cid:0)Pi>jhei, y∗i − Pi<jhei, y∗i(cid:1)ej = Pn−1 w⇁ e, we have j=1 ej − Hence (e∗n, 2Pn−1 j=1 ej + en − e) ∈ gra A. Then by (45), hx∗∗, e∗ni + 2 n−1 Xj=1 hx∗, eji + hx∗, eni − hx∗, ei = 0. 26 Since Pj≥1hx∗, eji = hx∗, ei, we have n−1 (46) hx∗∗, e∗ni = −2 Xj=1 hx∗, eji − hx∗, eni + hx∗, ei =Xj>n By Fact 4.6(ii)&(iii), Pk n=1(cid:0)Pj>nhx∗, eji −Pj<nhx∗, eji(cid:1)en Thus, A is maximally monotone. hx∗, eji −Xj<n hx∗, eji. w⇁ x∗∗. Hence (x∗, x∗∗) ∈ gra A. We next show that A is of type (BR). Let (z∗, z∗∗) ∈ gra(−A∗) ∩ X∗ × X∗∗. Much as in the proof above starting at (45), we have (z∗, z∗∗) ∈ gra A. Thus, gra(−A∗) ∩ X × X∗∗ ⊆ gra A. Then by Lemma 3.2, A is of type (BR). We turn to the particularization. By Fact 4.10, (en, e∗n)n∈N is a shrinking Schauder basis for 2h·, ei2 is maximally monotone. J. By Fact 2.15 since A is maximal, T = A + h·, eie = A + ∂ 1 Since A is skew, we have (47) hx∗, x∗∗i = hx∗, ei2, ∀(x∗, x∗∗) ∈ gra T. Now we claim that (48) Let (y∗, y∗∗) in gra T . Then k e /∈ ran T . hei, y∗i + hej, y∗i(cid:1)ej Xj=1(cid:0)2Xi>j Xj=1(cid:0)hy∗, ei +Xi>j = k k (49) = hy∗, ei ej + Xj=1 Then by (49), hej, y∗i = he, y∗i) k hei, y∗i −Xi<j Xj=1(cid:0)Xi>j hei, y∗i(cid:1)ej hei, y∗i −Xi<j (by Xi≥1 hei, y∗i(cid:1)ej w⇁ y∗∗. (50) lim k hy∗∗, e∗ki = lim k lim h L L hei, y∗i + hej, y∗i(cid:1)ej, e∗ki Xj=1(cid:0)2Xi>j hei, y∗i + hek, y∗i(cid:1) hek, y∗i = he, y∗i). = lim k (cid:0)2Xi>k = 0 (by Xk≥1 27 Then by Fact 4.10, y∗∗ ∈ J and hence ran T ⊆ J. Thus (51) ran T ⊆ J. Since he, e∗ki = 1, ∀k ∈ N, then by Lemma 4.7, e /∈ J. Then by (51), we have (48) holds. Combining (47), (48) and Proposition 3.5, T = A + h·, eie is neither of type (D) nor unique. This suffices to finish the argument. (cid:4) Remark 4.14 (ℓ1) A simpler version of the previous result recovers the original result that ℓ1 admits Gossez type operators. ♦ 5 Conclusion We have provided various tools for the further construction of pathological maximally mono- tone operators and related Fitzpatrick functions. In particular, we have shown -- building on the work of Gossez, Phelps, Simons, Svaiter, Bueno and others, and our own previous work -- that every Banach space which contains an isomorphic copy of either the James space J or its dual J∗, or c0 or its dual ℓ1, admits an operator which is not of type (D). We observe that the type (D) property is preserved by direct sums and subspaces. Since every separable space is isometric to a quotient space of ℓ1 [16, Theorem 5.1, page 237] or [15, Theorem 5.9, page 140], it is not preserved by quotients. Example 5.1 (Summary) We list some of the salient spaces covered by our work: (i) Separable Asplund spaces: both J and c0 afford examples. (ii) Separable spaces whose dual is nonseparable and contain ℓ1: include L1([0, 1]), C([0, 1]) and its superspace L∞([0, 1]). (iii) Separable spaces whose dual is nonseparable but does not contain a copy of ℓ1: these include the James tree space JT [16, page 233] or [15, page 199] as it contains many copies of J (and of ℓ2(N)). One remaining potential type (D) space is Gowers' space [21] which is a non-reflexive Banach space containing neither c0, ℓ1 or any reflexive subspace. ♦ As we saw, the maximally monotone operators in our examples -- with the exception of the Gossez operator -- that are not of type (D) are actually not unique. This raises the question of how in generality to construct maximally monotone linear relations that are not of type (D) but that are unique. 28 5.1 Graphic of classes of maximally monotone operators We capture much of the current state of knowledge in the following diagram in which the notation below is used. " ∗ " refers to skew operators such as T in Theorem 3.6, Tα in Example 4.1, A in Example 4.8, A in Corollary 4.11, and A in Corollary 4.12. " ∗ ∗" refers to the operators such as A&T in Theorem 3.6, Aα&Tα in Example 4.1, A in Example 4.8, A in Corollary 4.11, A in Corollary 4.12, and A + h·, eie in Example 4.13. " ∗ ∗ ∗ " denotes maximally monotone and unique operators with non affine graphs. We let (ANA), (FP) and (FPV) respectively denote the other monotone operator classes "almost negative alignment", "Fitzpatrick-Phelps" and "Fitzpatrick-Phelps-Veronas". Then by [35, 11, 9, 5, 33, 25, 36, 41], we have the following relationships. type (FPV) * Gossez operator type (BR) type (ANA) type (D) type (FP) type (NI) type (ED) uniqueness The following four questions are left open. (i) Is every maximally monotone operator necessarily of type (FPV)? (ii) Is every maximally monotone operator necessarily of type (ANA)? (iii) Is every maximally monotone linear relation necessarily of type (ANA)? (iv) Is every maximally monotone operator of type (BR) necessarily of type (ANA)? The first of these is especially important, being closely related to the sum theorem in general Banach space (see [35, 11, 9, 40]). 29 Acknowledgments. Heinz Bauschke was partially supported by the Natural Sciences and Engineering Research Council of Canada and by the Canada Research Chair Program. Jonathan Borwein was partially supported by the Australian Research Council. Xianfu Wang was partially supported by the Natural Sciences and Engineering Research Council of Canada. 30 References [1] H.H. Bauschke, Projection Algorithms and Monotone Operators, PhD thesis, Simon Fraser University, Department of Mathematics, Burnaby, British Columbia V5A 1S6, Canada, August 1996; Available at http://www.cecm.sfu.ca/preprints/1996pp.html. [2] H.H. Bauschke and J.M. Borwein, "Maximal monotonicity of dense type, local maximal monotonicity, and monotonicity of the conjugate are all the same for continuousi linear operators", Pacific Journal of Mathematics, vol. 189, pp. 1 -- 20, 1999. [3] H.H. Bauschke, X. Wang, and L. Yao, "Examples of discontinuous maximal monotone linear operators and the solution to a recent problem posed by B.F. Svaiter", Journal of Mathematical Analysis and Applications, vol. 370, pp. 224-241, 2010. [4] H.H. Bauschke, J.M. Borwein, X. Wang and L. Yao, "For maximally monotone linear relations, dense type, negative-infimum type, and Fitzpatrick-Phelps type all coincide with monotonicity of the adjoint", submitted; http://arxiv.org/abs/1103.6239v1, March 2011. [5] H.H. Bauschke, J.M. Borwein, X. Wang, and L. Yao, "Every maximally mono- tone operator of Fitzpatrick-Phelps type is actually of dense type", submitted; http://arxiv.org/abs/1104.0750v1, April 2011. [6] H.H. Bauschke and P.L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, Springer-Verlag, 2011. [7] J.M. Borwein, "Maximal monotonicity via convex analysis", Journal of Convex Analy- sis, vol. 13, pp. 561 -- 586, 2006. [8] J.M. Borwein, "Maximality of sums of two maximal monotone operators in general Banach space", Proceedings of the American Mathematical Society, vol. 135, pp. 3917 -- 3924, 2007. [9] J.M. Borwein, "Fifty years of maximal monotonicity", Optimization Letters, vol. 4, pp. 473 -- 490, 2010. [10] J.M. Borwein, "A note on ε-subgradients and maximal monotonicity", Pacific Journal of Mathematics, vol. 103, pp. 307 -- 314, 1982. [11] J.M. Borwein and J.D. Vanderwerff, Convex Functions, Cambridge University Press, 2010. [12] O. Bueno and B.F. Svaiter, "A non-type (D) operator in c0", http://arxiv.org/abs/1103.2349v1, March 2011. 31 [13] R.S. Burachik and A.N. Iusem, Set-Valued Mappings and Enlargements of Monotone Operators, Springer-Verlag, 2008. [14] R. Cross, Multivalued Linear Operators, Marcel Dekker, 1998. [15] M. Fabian, P. Habala, P. H´ajek, V. Montesinos Santaluc´ıa, J. Pelant and V. Zizler, Functional Analysis and Infinite-Dimensional Geometry, CMS/Springer-Verlag, 2001. [16] M. Fabian, P. Habala, P. H´ajek, V. Montesinos and V. Zizler, Banach Space Theory, CMS/ Springer-Verlag, 2010. [17] S. Fitzpatrick, "Representing monotone operators by convex functions", in Work- shop/Miniconference on Functional Analysis and Optimization (Canberra 1988), Pro- ceedings of the Centre for Mathematical Analysis, Australian National University, vol. 20, Canberra, Australia, pp. 59 -- 65, 1988. [18] S. Fitzpatrick and R.R. Phelps, "Bounded approximants to monotone operators on Banach spaces", Annales de l'Institut Henri Poincar´e. Analyse Non Lin´eaire, vol. 9, pp. 573 -- 595, 1992. [19] J.-P. Gossez, "Op´erateurs monotones non lin´eaires dans les espaces de Banach non r´eflexifs", Journal of Mathematical Analysis and Applications, vol. 34, pp. 371 -- 395, 1971. [20] J.-P. Gossez, "On the range of a coercive maximal monotone operator in a nonreflexive Banach space", Proceedings of the American Mathematical Society, vol. 35, pp. 88-92, 1972. [21] W. T. Gowers, "A Banach space not containing c0, l1 or a reflexive subspace", Trans- actions of the American Mathematical Society, vol. .344, pp. 407-420, 1994. [22] M. Marques Alves and B.F. Svaiter, "A new proof for maximal monotonicity of subd- ifferential operators", Journal of Convex Analysis, vol. 15, pp. 345 -- 348, 2008. [23] M. Marques Alves and B.F. Svaiter, "Brøndsted-Rockafellar property and maximality of monotone operators representable by convex functions in non-reflexive Banach spaces", Journal of Convex Analysis, vol. 15, pp. 693-706, 2008. [24] M. Marques Alves and B.F. Svaiter, "Maximal monotone operators with a unique ex- tension to the bidual", Journal of Convex Analysis, vol. 16, pp. 409-421, 2009 [25] M. Marques Alves and B.F. Svaiter, "On Gossez type (D) maximal monotone opera- tors", Journal of Convex Analysis, vol. 17, pp. 1077 -- 1088, 2010. [26] R.E. Megginson, An Introduction to Banach Space Theory, Springer-Verlag, 1998. 32 [27] R.R. Phelps, Convex Functions, Monotone Operators and Differentiability, 2nd Edition, Springer-Verlag, 1993. [28] R.R. Phelps, "Lectures on maximal monotone operators", Extracta Mathematicae, vol. 12, pp. 193 -- 230, 1997; http://arxiv.org/abs/math/9302209v1, February 1993. [29] R.R. Phelps and S. Simons, "Unbounded linear monotone operators on nonreflexive Banach spaces", Journal of Convex Analysis, vol. 5, pp. 303 -- 328, 1998. [30] R.T. Rockafellar, "On the maximal monotonicity of subdifferential mappings", Pacific Journal of Mathematics, vol. 33, pp. 209 -- 216, 1970. [31] R.T. Rockafellar and R.J-B Wets, Variational Analysis, 3rd Printing, Springer-Verlag, 2009. [32] S. Simons, "The range of a monotone operator", Journal of Mathematical Analysis and Applications, vol. 199, pp. 176 -- 201, 1996. [33] S. Simons, Minimax and Monotonicity, Springer-Verlag, 1998. [34] S. Simons, "Five kinds of maximal monotonicity", Set-Valued and Variational Analysis, vol. 9, pp. 391 -- 409, 2001. [35] S. Simons, From Hahn-Banach to Monotonicity, Springer-Verlag, 2008. [36] S. Simons, "Banach SSD Spaces and classes of monotone sets", Journal of Convex Analysis, vol. 18, pp. 227 -- 258, 2011. [37] S. Simons, "Maximal monotone multifunctions of Brøndsted-Rockafellar type", Set- Valued Analysis, vol. 7 pp. 255-294, 1999. [38] S. Simons and C. Zalinescu, "Fenchel duality, Fitzpatrick functions and maximal mono- tonicity", Journal of Nonlinear and Convex Analysis, vol. 6, pp. 1 -- 22, 2005. [39] M.D. Voisei and C. Zalinescu, "Linear monotone subspaces of locally convex spaces", Set-Valued and Variational Analysis, vol. 18, pp. 29 -- 55, 2010. [40] L. Yao, "The sum of a maximal monotone operator of type (FPV) and a maximal monotone operator with full domain is maximally monotone", to appear Nonlinear Analysis. [41] L. Yao, "The sum of a maximally monotone linear relation and the subdifferential of a proper lower semicontinuous convex function is maximally monotone", to appear Set- Valued and Variational Analysis. 33 [42] C. Zalinescu, Convex Analysis in General Vector Spaces, World Scientific Publishing, 2002. 34
1905.04078	1	1905	2019-05-10T11:35:06	Birkhoff--James orthogonality of operators in semi-Hilbertian spaces and its applications	[ "math.FA" ]	In this paper, the concept of Birkhoff--James orthogonality of operators on a Hilbert space is generalized when a semi-inner product is considered. More precisely, for linear operators $T$ and $S$ on a complex Hilbert space $\mathcal{H}$, a new relation $T\perp^B_A S$ is defined if $T$ and $S$ are bounded with respect to the seminorm induced by a positive operator $A$ satisfying ${\\|T + \gamma S\\|}_A\geq {\\|T\\|}_A$ for all $\gamma \in \mathbb{C}$. We extend a theorem due to R. Bhatia and P. \v{S}emrl, by proving that $T\perp^B_A S$ if and only if there exists a sequence of $A$-unit vectors $\{x_n\}$ in $\mathcal{H}$ such that $\displaystyle{\lim_{n\rightarrow +\infty}}{\\|Tx_n\\|}_A = {\\|T\\|}_A$ and $\displaystyle{\lim_{n\rightarrow +\infty}}{\langle Tx_n, Sx_n\rangle}_A = 0$. In addition, we give some $A$-distance formulas. Particularly, we prove \begin{align} \displaystyle{\inf_{\gamma \in \mathbb{C}}}{\\|T + \gamma S\\|}_{A} = \sup\Big\{\|{\langle Tx, y\rangle}_A\|; \, {\\|x\\|}_{A} = {\\|y\\|}_{A} = 1, \, {\langle Sx, y\rangle}_A = 0\Big\}. \end{align} Some other related results are also discussed.	math.FA	math	BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS IN SEMI-HILBERTIAN SPACES AND ITS APPLICATIONS ALI ZAMANI Abstract. In this paper, the concept of Birkhoff -- James orthogonality of op- erators on a Hilbert space is generalized when a semi-inner product is consid- ered. More precisely, for linear operators T and S on a complex Hilbert space H, a new relation T ⊥B A S is defined if T and S are bounded with respect to the seminorm induced by a positive operator A satisfying kT + γSkA ≥ kT kA for all γ ∈ C. We extend a theorem due to R. Bhatia and P. Semrl, by proving that T ⊥B A S if and only if there exists a sequence of A-unit vectors {xn} in H hT xn, SxniA = 0. In addition, we such that kT xnkA = kT kA and lim lim n→+∞ n→+∞ give some A-distance formulas. Particularly, we prove inf γ∈C kT + γSkA = supnhT x, yiA; kxkA = kykA = 1, hSx, yiA = 0o. Some other related results are also discussed. 1. Introduction and preliminaries Let B(H) denote the C ∗-algebra of all bounded linear operators on a complex Hilbert space H with an inner product h·, ·i and the corresponding norm k · k. The symbol I stands for the identity operator on H. If T ∈ B(H), then we denote by R(T ) and N (T ) the range and the kernel of T , respectively, and by R(T ) the norm closure of R(T ). Throughout this article, we assume that A ∈ B(H) is a positive operator and that P is the orthogonal projection onto R(A). Recall that A is called positive if hAx, xi ≥ 0 for all x ∈ H. Such an A induces a positive semidefinite sesquilinear form h·, ·iA : H × H → C defined by hx, yiA = hAx, yi, x, y ∈ H. Denote by k · kA the seminorm induced by h·, ·iA, that is, kxkA =phx, xiA for every x ∈ H. It can be easily seen that k · kA is a norm if and only if A is an injective operator, and that (H, k · kA) is a complete space if and only if R(A) is closed in H. For x, y ∈ H, we say that x and y are A-orthogonal, denoted by x ⊥A y, if hx, yiA = 0. Note that this definition is a natural extension of the usual notion of orthogonality, which represents the I-orthogonality case. Furthermore, we put B A1/2(H) =nT ∈ B(H) : ∃ c > 0 ∀x ∈ H; kT xkA ≤ ckxkAo. Copyright 2018 by the Tusi Mathematical Research Group. Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz. 2010 Mathematics Subject Classification. Primary 46C05; Secondary 47B65, 47L05. Key words and phrases. Positive operator, semi-inner product, A-Birkhoff -- James orthogo- nality, and A-distance formulas. 1 2 A. ZAMANI We say that an operator T ∈ B(H) is A-bounded if T belongs to B can be shown that B neither closed nor dense in B(H) (see [2]). We equip B k · kA defined as follows: A1/2(H). It A1/2(H) is a unital subalgebra of B(H) which, in general, is A1/2(H) with the seminorm kT kA = sup x∈R(A),x6=0 In addition, for T ∈ B kT kA = sup x∈H,kxkA=1 kT xkA kxkA = infnc > 0; kT xkA ≤ ckxkA, x ∈ Ho < ∞. kT xkA = supnhT x, yiA; x, y ∈ H, kxkA = kykA = 1o. A1/2(H), we have Of course, many difficulties arise. For instance, it may happen that kT kA = ∞ for some T ∈ B(H). In addition, not any operator admits an adjoint operator for the semi-inner product h·, ·iA. For more details about this class of operators, we refer the reader to [2]. In recent years, several results covering some classes of operators on a complex Hilbert space(cid:0)H, h·, ·i(cid:1) are extended to(cid:0)H, h·, ·iA(cid:1); see The notion of orthogonality in B(H) can be introduced in many ways ( see, e.g., [13]). When T, S ∈ B(H), we say that T is Birkhoff -- James orthogonal to S, denoted T ⊥B S, if [2, 3] and their references. kT + γSk ≥ kT k for all γ ∈ C. In Hilbert spaces, this orthogonality is equivalent to the usual notion of orthogo- nality. This notion of orthogonality plays a very important role in the geometry of Hilbert space operators. For T, S ∈ B(H), Bhatia and Semrl in [4, Remark 3.1] and Paul in [14, Lemma 2] independently proved that T ⊥B S if and only if there exists a sequence of unit vectors {xn} in H such that lim n→∞ kT xnk = kT k and lim n→∞ hT xn, Sxni = 0. It follows then that if the Hilbert space H is finite-dimensional, T ⊥B S if and only if there is a unit vector x ∈ H such that kT xk = kT k and hT x, Sxi = 0. Recently, some authors extended the well known result of Bhatia -- Semrl (see [6, 18, 19]). Moreover, the papers [18] and [19] show another ways to obtain the Bhatia -- Semrl theorem. Some other authors studied different aspects of orthogo- nality of operators on various Banach spaces and elements of an arbitrary Hilbert C ∗-module; see, for instance,[1, 5, 7, 10, 11, 15, 17, 20]. Now, let us introduce the notion of A-Birkhoff -- James orthogonality of opera- tors in semi-Hilbertian spaces. Definition 1.1. An element T ∈ B onal to another element S ∈ B A1/2(H) is called an A-Birkhoff -- James orthog- A1/2(H), denoted by T ⊥B kT + γSkA ≥ kT kA for all γ ∈ C. A S, if It is a generalization of the notion of Birkhoff -- James of Hilbert space operators. Notice that the A-Birkhoff -- James orthogonality is homogenous, that is, T ⊥B A S ⇔ (αT ) ⊥B A (βS) for all α, β ∈ C. A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 3 The paper is organized as follows: In the next section, we obtain some char- acterizations of A-Birkhoff -- James orthogonality for bounded linear operators in semi-Hilbertian spaces. In particular, for T, S ∈ B A S if and only if there exists a sequence of A-unit vectors {xn} in H such that A1/2(H), we show that T ⊥B lim n→+∞ kT xnkA = kT kA and lim n→+∞ hT xn, SxniA = 0. Furthermore, for the finite-dimensional Hilbert space H, we show that T ⊥B A S if and only if there exists an A-unit vector x ∈ H such that kT xkA = kT kA and hT x, SxiA = 0. The mentioned property extends the Bhatia -- Semrl theorem. In the last section, some formulas for the A-distance of an operator to the class of multiple scalars of another one in semi-Hilbertian spaces are given. In particular, we show that inf γ∈C kT + γSkA = supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo. We then apply it to prove that inf γ∈C kT + γSk2 A = sup kxkA=1 Φ(T,S) A (x), where Φ(T,S) A (x) =(kT xk2 kT xk2 A A − hT x,SxiA2 kSxk2 A if kSxkA 6= 0, if kSxkA = 0. Our results cover and extend the works of Fujii and Nakamoto in [9] and Bhatia and Semrl in [4]. 2. A-Birkhoff -- James orthogonality of operators We first prove a technical lemma that we need in what follows. We use some techniques of [3, Theorem 3.2] to prove this result. In fact, the following lemma extends Magajna's theorem [12]. Lemma 2.1. Let T, S ∈ B A1/2(H). Then the set WA(T, S) =nξ ∈ C; ∃ {xn} ⊂ H, kxnkA = 1, lim n→+∞ kT xnkA = kT kA, and lim n→+∞ hT xn, SxniA = ξo is nonempty, compact, and convex. Proof. Since the seminorm of T ∈ B A1/2(H) is given by kT kA = sup{kT xkA; x ∈ R(A), kxkA = 1}, there exists a sequence of A-unit vectors {xn} in R(A) such that lim n→+∞ kT xnkA = kT kA. Furthermore, using the Cauchy -- Schwarz inequality, we have hT xn, SxniA ≤ kT xnkAkSxnkA ≤ kT kAkSkA. and also for all sufficiently large m. From (2.1) and (2.2), we get ε 2 m, Sxn mkA − kT kA(cid:12)(cid:12) < ε (cid:12)(cid:12)kT xn miA − ξn(cid:12)(cid:12) < (cid:12)(cid:12)hT xn mkA − kT kA(cid:12)(cid:12) < ε (cid:12)(cid:12)kT xn miA − ξ(cid:12)(cid:12) ≤(cid:12)(cid:12)hT xn m, Sxn (2.1) (2.2) 4 A. ZAMANI Hence, {hT xn, SxniA} is a bounded sequence of complex numbers, so there ex- ists a subsequence {hT xnk, SxnkiA} that converges to some ξ0 ∈ C. Thus ξ0 ∈ WA(T, S) and hence WA(T, S) is nonempty. On the other hand, considering the definition of WA(T, S) follows that WA(T, S) ⊂(cid:8)ξ ∈ C; ξ ≤ kT kAkSkA(cid:9). Therefore, to prove that WA(T, S) is compact, it is enough to show that WA(T, S) ξn = ξ. Since ξn ∈ WA(T, S), there is closed. Let ξn ∈ WA(T, S) and let exists a sequence of A-unit vectors {xn mkA = kT kA n→+∞ m} in H such that kT xn lim lim m→+∞ and lim m→+∞ hT xn m, Sxn miA = ξn. Now, let ε > 0. Hence and m, Sxn (cid:12)(cid:12)hT xn miA − ξn(cid:12)(cid:12) + ξn − ξ < ε 2 kT xn mkA = kT kA and = ε lim ε 2 + m→+∞ for all sufficiently large m. Therefore we deduce that lim m→+∞ hT xn m, Sxn miA = ξ. Thus ξ ∈ WA(T, S) and so WA(T, S) is closed. We next show that WA(T, S) is convex. Since H can be decomposed as H = N (A) ⊕ R(A), so every x ∈ H can be written in a unique way into x = y + z with y ∈ N (A) and z ∈ R(A). Furthermore, since A ≥ 0, it follows that N (A) = N (A1/2) which implies that kxkA = kzkA. Thus WA(T, S) =nξ ∈ C; ∃ {(yn, zn)} ⊂ N (A) × R(A), kznkA = 1, lim n→+∞ kT (yn + zn)kA = kT kA, and lim n→+∞ hT yn, SzniA + hT zn, SzniA = ξo. A1/2(H), then T (N (A)) ⊂ N (A) and S(N (A)) ⊂ N (A). Hence, Since T, S ∈ B we get lim kT znkA = kT kA, and lim n→+∞ n→+∞ WA(T, S) =nξ ∈ C; ∃ {zn} ⊂ R(A), kznkA = 1, =nξ ∈ C; ∃ {zn} ⊂ R(A), kznkA = 1, = WA0(eT ,eS), kP T znkA = kP T R(A) k n→+∞ lim A hT zn, SzniA = ξo , and lim n→+∞ hP T zn, P SzniA = ξo This concept is useful in studying linear operators (see [13], and further references therein). The A-minimum modulus of S ∈ B A1/2(H) can be defined by m(S) = infnkSxk : x ∈ H, kxk = 1o. mA(S) = infnkSxkA : x ∈ H, kxkA = 1o. A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 5 where A0 = A R(A), eT = P T R(A) and eS = P S R(A). By [12, Lemma 2.1], we conclude that WA(T, S) is convex. Recall that the minimum modulus of S ∈ B(H) is defined by (cid:3) We are now in a position to establish the main result of this section. To establish the following theorem, we use some ideas of [16, Theorem 2]. Theorem 2.2. Let T, S ∈ B lent: A1/2(H). Then the following conditions are equiva- (i) There exists a sequence of A-unit vectors {xn} in H such that lim n→+∞ kT xnkA = kT kA and lim n→+∞ hT xn, SxniA = 0. (ii) kT + γSk2 (iii) T ⊥B A S. A ≥ kT k2 A + γ2m2 A(S) for all γ ∈ C. Proof. (i)⇒(ii) Suppose that (i) holds. We have kT + γSk2 A ≥ k(T + γS)xnk2 A = kT xnk2 A + γhT xn, SxniA + γhSxn, T xniA + γ2kSxnk2 A for all γ ∈ C and n ∈ N. Thus kT + γSk2 A ≥ kT k2 A + γ2 lim n→∞ sup kSxnk2 A ≥ kT k2 A + γ2m2 A(S) for all γ ∈ C. (ii)⇒(iii) This implication is trivial. (iii)⇒(i) If kSkA = 0, then since T is a seminorm, there exists a sequence of A- kT xnkA = kT kA. So, the Cauchy -- Schwarz lim unit vectors {xn} in H such that inequality implies n→+∞ hT xn, SxniA ≤ kT xnkAkSxnkA ≤ kT kAkSkA = 0. lim n→+∞ Hence, hT xn, SxniA = 0. Now, let kSkA 6= 0. It is enough to show that 0 ∈ WA(T, S), where WA(T, S) is defined in Lemma 2.1. let 0 /∈ WA(T, S). Lemma 2.1 implies that WA(T, S) is a nonempty compact and convex subset of the complex plane C; hence because of the rotation, we may suppose that WA(T, S) is contained in the right half-plane. Therefore there is a line that separates 0 from WA(T, S). In other words, there exists τ > 0 such that ReWA(T, S) > τ . Let Hτ =nx ∈ H; kxkA = 1, and ReWA(T, S) ≤ τ 2o 6 and A. ZAMANI δ = supnkT xkA; x ∈ Hτo. We first claim that δ < kT kA. Suppose δ ≥ kT kA. Hence δ = kT kA. Thus there exists a sequence of vectors {xn} in Hτ such that kT xnkA = kT kA. As xn ∈ Hτ so kxnkA = 1 and ReWA(T, S) ≤ τ 2 . Now the sequence {hT xn, SxniA} is bounded, and hence it has a convergent subsequence, without loss of generality, hT xn, SxniA, we can assume that {hT xn, SxniA} is convergent. If we set ξ = lim n→+∞ then Re(ξ) ≤ τ 2 . Thus δ < kT kA. Let γ0 = max{ −τ }. Then γ0 < 0. We claim that kT + γ0SkA < 2kSk2 A kT kA. Let x be an A-unit vector in H. If x ∈ Hτ , then 2 and this contradicts the fact that ReWA(T, S) > τ , δ−kT kA 2kSkA n→+∞ lim k(T + γ0S)xkA ≤ kT xkA + γ0kSxkA ≤ δ − γ0kSkA kT kA kT kA − δ ≤ δ + 2kSkA kSkA = + δ 2 2 and so k(T + γ0S)xkA ≤ δ 2 + kT kA 2 . If x /∈ Hτ , then we can write T x = (r + it)Sx + y with r, t ∈ R and Sx ⊥A y. Thus 2rkSk2 A ≥ 2rkSxk2 A = 2RehT x, SxiA > τ 2 ≥ −γ0kSk2 A, and hence 2r + γ0 > 0. Now, let us put Since γ2 0 + 2rγ0 < 0, we obtain k(T + γ0S)xk2 A; x /∈ Hτ , kxkA = 1(cid:9). θ := inf(cid:8)kSxk2 A =D(cid:0)(r + γ0) + it(cid:1)Sx + y,(cid:0)(r + γ0) + it(cid:1)Sx + yEA =(cid:0)(r + γ0)2 + t2(cid:1)kSxk2 A; x /∈ Hτ , kxkA = 1(cid:9) 0 + 2rγ0) inf(cid:8)kSxk2 A + (γ2 A + (γ2 A + (γ2 = kT xk2 ≤ kT xk2 ≤ kT k2 0 + 2rγ0)kSxk2 0 + 2rγ0)θ. A + kyk2 A A Hence k(T + γ0S)xk2 A ≤ kT k2 A + (γ2 0 + 2rγ0)θ. Thus in all cases whence k(T + γ0S)xk2 kT + γ0Sk2 2 A ≤ maxn(cid:0) δ A ≤ maxn(cid:0) δ 2 (cid:1)2 A + (γ2 , kT k2 2 + kT kA 2 + kT kA 2 (cid:1)2, kT k2 (cid:1)2 , kT k2 A + (γ2 A + (γ2 0 + 2rγ0)θo, 0 + 2rγ0)θo. Since maxn(cid:0) δ 2 + kT kA kT kA. Therefore we deduce that T 6⊥B the proof is completed. 0 + 2rγ0)θo < kT k2 A, we obtain kT + γ0SkA < A S which contradicts our hypothesis and (cid:3) A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 7 The following corollary gives a direct application of Theorem 2.2 for the case A = I. Corollary 2.3. (see [4, Remark 3.1] and [14, Lemma 2]) Let H be a complex Hilbert space and let T, S ∈ B(H). Then the following statements are equivalent: (i) T ⊥B S. (ii) There exists a sequence of unit vectors {xn} in H such that lim n→+∞ kT xnk = kT k and lim n→+∞ hT xn, Sxni = 0. In what follows, for T ∈ B A1/2(H), we denote MT A the set of all A-unit vectors at which T attains the seminorm k · kA, that is, MT A =nx ∈ H : kxkA = 1, kT xkA = kT kAo. For more information on norm-attaining sets, see [8]. In the next theorem, we consider a finite-dimensional Hilbert space and characterize the A-Birkhoff -- James orthogonality of operators in semi-Hilbertian spaces. Theorem 2.4. Let H be a finite-dimensional Hilbert space and let T, S ∈ B Then the following conditions are equivalent: A1/2(H). (i) There exists x ∈ MT (ii) T ⊥B A S. A such that T x ⊥A Sx. Proof. (i)⇒(ii) Suppose that (i) holds. Then there exists an A-unit vectors x ∈ H such that kT xkA = kT kA and T x ⊥A Sx. Put xn = x for all n ∈ N. So, by the equivalence (i)⇔(iii) in Theorem 2.2, we deduce that T ⊥B A S. (ii)⇒(i) First note that, by using the decomposition H = N (A) ⊕ R(A) and letting A0 = A R(A), it can be seen that the set {x ∈ R(A); kxkA0 = 1} is homeomorphic to the set {x ∈ R(A); kxk = 1}, which is compact since R(A) is finite-dimensional. Thus we get the set {x ∈ R(A); kxkA0 = 1} is compact. Now, suppose that (ii) holds. Put eT = P T R(A) and eS = P S R(A). Therefore, by the equivalence (i)⇔(iii) in Theorem 2.2, there exists a sequence of A0-unit vectors {xn} in R(A) such that and lim n→+∞ lim n→+∞ keT xnkA0 = keT kA0 Since the set {x ∈ R(A); kxkA0 = 1} is compact, hence {xn} has a subsequence heT xn,eSxniA0 = 0. {xnk} that converges to some x ∈ R(A) with kxkA0 = 1. This yields keT xkA0 = heT xnk ,eSxnkiA0 = 0. From this keT xnkkA0 = keT kA0 and heT x,eSxiA0 = lim k→+∞ it follows that x ∈ MT A and T x ⊥A Sx. k→+∞ lim (cid:3) As an immediate consequence of Theorem 2.4, we have the following result. Corollary 2.5. Let H be finite dimensional and let T, S ∈ B following statements are equivalent: A1/2(H). Then the (i) T ⊥B A S. 8 A. ZAMANI (ii) There exists x ∈ MT A such that for every γ ∈ C kT x + γSxk2 A = kT xk2 A + γ2kSxk2 A. 3. Some A-distance formulas In this section we give some formulas for the A-distance of an operator to the class of multiple scalars of another one in semi-Hilbertian spaces. For T, S ∈ kT + γSkA. The following B A1/2(H) we have, by definition, dA(T, CS) := inf γ∈C auxiliary lemma is needed for next results. Lemma 3.1. Let T, S ∈ B A1/2(H). Then there exists ζ0 ∈ C such that dA(T, CS) = kT + ζ0SkA. Proof. If kSkA = 0, then kT + γSkA ≥ kT kA − γkSkA = kT kA, for all γ ∈ C. It is therefore enough to put ζ0 = 0. If kSkA 6= 0, then put D := nγ ∈ C; γ ≤ 2kT kA kSkAo and define f : D → R by the formula f (γ) = kT + γSkA. Clearly, f is continuous and attains its minimum at, say, ζ0 ∈ D (of course, there may be many such points). Then kT + γSkA ≥ kT + ζ0SkA for all γ ∈ D. If γ /∈ D, then γ > 2kT kA kSkA . Since 0 ∈ D, we obtain kT + γSkA ≥ γkSkA − kT kA > 2kT kA − kT kA = kT kA ≥ kT + ζ0SkA. Thus kT + γSkA ≥ kT + ζ0SkA for all γ /∈ D. Therefore, kT + γSkA ≥ kT + ζ0SkA for all γ ∈ C. So, we conclude that inf γ∈C kT + γSkA = kT + ζ0SkA and hence (cid:3) dA(T, CS) = kT + ζ0SkA. The following result is a kind of the Pythagorean relation for bounded operators in semi-Hilbertian spaces. Theorem 3.2. Let T, S ∈ B ζ0 ∈ C, such that A1/2(H) with mA(S) > 0. Then there exists a unique for every γ ∈ C. (cid:13)(cid:13)(T + ζ0S) + γS(cid:13)(cid:13)2 A ≥ kT + ζ0Sk2 A + γ2 m2 A(S) Proof. By Lemma 3.1, there exists ζ0 ∈ C such that or equivalently, inf γ∈C kT + γSkA = kT + ζ0SkA, inf ξ∈C k(T + ζ0S) + ξSkA = kT + ζ0SkA. Thus (T + ζ0S) ⊥B γ ∈ C, we have A S. So, by the equivalence (i)⇔(ii) in Theorem 2.2, for every (cid:13)(cid:13)(T + ζ0S) + γS(cid:13)(cid:13)2 A ≥ kT + ζ0Sk2 A + γ2 m2 A(S). A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 9 Now, suppose that ζ1 is another point satisfying the inequality (cid:13)(cid:13)(T + ζ1S) + γS(cid:13)(cid:13)2 Choose γ = ζ0 − ζ1 to get A ≥ kT + ζ1Sk2 A + γ2 m2 A(S) (γ ∈ C). kT + ζ0Sk2 A =(cid:13)(cid:13)(T + ζ1S) + (ζ0 − ζ1)S(cid:13)(cid:13)2 A + ζ0 − ζ12 m2 A + ζ0 − ζ12 m2 ≥ kT + ζ1Sk2 ≥ kT + ζ0Sk2 A A(S) A(S). Hence 0 ≥ ζ0−ζ12 m2 ζ0 = ζ1. This shows that ζ0 is unique. A(S). Since m2 A(S) > 0, we get ζ0−ζ12 = 0, or equivalently, (cid:3) Here, we establish one of our main results. In fact, in what follows, we provide a version of the Bhatia -- Semrl theorem (see [4, p. 84]) in the setting of operators in semi-Hilbertian spaces. Theorem 3.3. Let T, S ∈ B A1/2(H). Then dA(T, CS) = supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo. Proof. Let x, y ∈ H, kxkA = kykA = 1 and let Sx ⊥A y. The Cauchy -- Schwarz inequality implies hT x, yiA = h(T + γS)x, yiA ≤ k(T + γS)xkAkykA ≤ kT + γSkA for all γ ∈ C. Thus supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo ≤ kT + γSkA supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo ≤ inf supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo ≤ dA(T, CS). γ∈C kT + γSkA. (3.1) for all γ ∈ C and so Hence On the other hand, by Lemma 3.1, there exists ζ0 ∈ C such that dA(T, CS) = kT + ζ0SkA. We assume that ζ0 = 0 (otherwise we just replace T by T + ζ0S). Thus dA(T, CS) = kT kA, or equivalently, T ⊥B A S. Then, by the equivalence (i)⇔(iii) in Theorem 2.2, there exists a sequence of A-unit vectors {xn} in H hT xn, SxniA = 0. Now, let T xn = such that kT xnkA = kT kA and lim n→+∞ n→+∞ lim 10 A. ZAMANI αnSxn + βnyn with Sxn ⊥A yn, kynkA = 1, and αn, βn ∈ C. Then we have A(T, CS) = kT k2 d2 A = lim n→+∞ kT xnk2 A = lim n→+∞DαnSxn + βnyn, αnSxn + βnynEA = lim n→+∞ = lim n→+∞ = lim n→+∞ hαnSxn, αnSxniA + βn2 hT xn − βnyn, αnSxniA + βn2 αnhT xn, SxniA − αnβnhyn, SxniA + βn2 = lim βn2. n→+∞ Consequently, we obtain dA(T, CS) = lim n→+∞ βn = lim n→+∞ hβnyn, yniA whence = lim n→+∞ hT xn − αnSxn, yniA = lim n→+∞ hT xn, yniA ≤ supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo, dA(T, CS) ≤ supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo. dA(T, CS) = supnhT x, yiA; kxkA = kykA = 1, Sx ⊥A yo. From (3.1) and (3.2), we conclude that (3.2) (cid:3) (3.3) For T ∈ B(H), Fujii and Nakamoto in [9] proved that dA(T, CI) can be written in the following form: d(T, CI) =(cid:16) sup kxk=1(cid:0)kT xk2 − hT x, xi2(cid:1)(cid:17)1/2 = sup kxk=1(cid:13)(cid:13)T x − hT x, xix(cid:13)(cid:13), which shows that dA(T, CI) is the supremum over the lengths of all perpendiculars from T x to x, where x passes over the set of unit vectors. In the following theorem, for T, S ∈ B A1/2(H), we show that dA(T, CS) can also be expressed in the form generalizing of (3.3). Theorem 3.4. Let T, S ∈ B A1/2(H). Then d2 A(T, CS) = sup kxkA=1 Φ(T,S) A (x), where Φ(T,S) A (x) =(kT xk2 kT xk2 A A − hT x,SxiA2 kSxk2 A if kSxkA 6= 0 if kSxkA = 0. A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 11 Proof. For every γ ∈ C and every A-unit vector x ∈ H such that kSxkA 6= 0, we have kT x + γSxk2 A − hT x + γSx, SxiA2 kSxk2 A A + γ2kSxk2 = kT xk2 A + 2RehT x, γSxiA hT x, SxiA2 + γ2kSxk4 A + 2kSxk2 ARehT x, γSxiA − kSxk2 A hT x + γSx, SxiA2 = kT xk2 A − hT x, SxiA2 kSxk2 A . Thus Φ(T,S) A (x) = kT x + γSxk2 A − ≤ kT x + γSxk2 kSxk2 A A ≤ kT + γSk2 A. Also, in the case kSxkA = 0 we have Φ(T,S) A (x) = kT xk2 Hence we obtain Φ(T,S) every γ ∈ C. Therefore, A consequently, kxkA=1 A ≤(cid:0)kT x + γSxkA + kγSxkA(cid:1)2 = kT x + γSxk2 (x) ≤ kT + γSk2 sup Φ(T,S) A A for every A-unit vector x ∈ H and A for every γ ∈ C and (x) ≤ kT + γSk2 A ≤ kT + γSk2 A. sup kxkA=1 Φ(T,S) A (x) ≤ inf γ∈C kT + γSk2 A. Thus sup kxkA=1 Φ(T,S) A (x) ≤ d2 A(T, CS). (3.4) Now, take A-unit vectors x, y ∈ H such that Sx ⊥A y. If kSxkA = 0, then hT x, yiA2 ≤ kT xk2 Akyk2 A = Φ(T,S) A (x) ≤ sup Φ(T,S) A (x). kxkA=1 If kSxkA 6= 0, then hT x, yiA2 =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)DT x − ≤DT x − hT x, SxiA kSxk2 A hT x, SxiA 2 Sx, yEA(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Sx, T x − kSxk2 A hT x, SxiA2 kSxk2 A hT x, SxiA kSxk2 A SxEA = kT xk2 A − = Φ(T,S) A (x) ≤ sup kxkA=1 Φ(T,S) A (x). 12 A. ZAMANI So, we conclude that hT x, yiA2 ≤ sup kxkA=1 such that Sx ⊥A y. Therefore, Theorem 3.3 implies that Φ(T,S) A (x) for all A-unit vectors x, y ∈ H d2 A(T, CS) ≤ sup kxkA=1 Φ(T,S) A Now, the result follows from (3.4) and (3.5). (x). (3.5) (cid:3) We close this paper with the following Inf-sup equality in semi-Hilbertian spaces. Theorem 3.5. Let T, S ∈ B A1/2(H). Then sup k(T + γS)xk2 inf γ∈C kxkA=1 A = sup kxkA=1 k(T + γS)xk2 A. inf γ∈C Proof. Let x ∈ H with kxkA = 1. If kSxkA = 0, then k(T + γS)xkA ≥ kT xkA − γkSxkA = kT xkA, for all γ ∈ C. Thus kT xk2 A ≥ inf γ∈C k(T + γS)xk2 A ≥ kT xk2 A, whence inf γ∈C k(T + γS)xk2 A = kT xk2 A. Hence inf γ∈C k(T + γS)xk2 A = Φ(T,S) A (x). If kSxkA 6= 0, then simple computations show that k(T + γS)xk2 A = kSxk2 hT x, SxiA kSxk2 A 2 + kT xk2 A − hT x, SxiA2 kSxk2 A . A(cid:12)(cid:12)(cid:12) + γ(cid:12)(cid:12)(cid:12) A achieves its minimum at − hT x,SxiA Thus k(T + γS)xk2 kSxk2 A A − hT x,SxiA2 A = Φ(T,S) kT xk2 . Hence inf γ∈C x ∈ H. From this, by Theorem 3.4, we conclude that k(T + γS)xk2 kSxk2 A A and the minimum value is (x) for every A-unit vector sup kxkA=1 inf γ∈C k(T + γS)xk2 Φ(T,S) (x) A A = sup kxkA=1 A(T, CS) = d2 = inf γ∈C kT + γSk2 A = inf γ∈C k(T + γS)xk2 A. sup kxkA=1 Acknowledgments. The author would like to thank the referees for their valuable comments which helped to improve the paper. This work was sup- ported by a grant from Shanghai Municipal Science and Technology Commission (18590745200). (cid:3) A-BIRKHOFF -- JAMES ORTHOGONALITY OF OPERATORS 13 References 1. Lj. Arambasi´c and R. Raji´c, The Birkhoff -- James orthogonality in Hilbert C ∗-modules, Lin- ear Algebra Appl. 437 (2012), 1913 -- 1929. 2. M.L. Arias, G. Corach and M.C. Gonzalez, Metric properties of projections in semi- Hilbertian spaces, Integral Equations and Operator Theory, 62(1) (2008), 11 -- 28. 3. H. Baklouti, K. Feki and O. A. M. Sid Ahmed, Joint numerical ranges of operators in semi-Hilbertian spaces, Linear Algebra Appl. 555 (2018), 266 -- 284. 4. R. Bhatia and P. Semrl, Orthogonality of matrices and some distance problems, Linear Algebra Appl. 287(1-3) (1999), 77 -- 85. 5. T. Bottazzi, C. Conde, M.S. Moslehian, P. W´ojcik and A. Zamani, Orthogonality and parallelism of operators on various Banach spaces, to appear in J. Aust. Math, DOI: 10.1017/S1446788718000150. 6. J. Chmieli´nski, T. Stypu la and P. W´ojcik, Approximate orthogonality in normed spaces and its applications, Linear Algebra Appl. 531 (2017), 305 -- 317. 7. J. Chmieli´nski and P. W´ojcik, Approximate symmetry of Birkhoff orthogonality, J. Math. Anal. Appl. 461 (2018), no. 1, 625 -- 640. 8. J. Falc´o, D. Garc´ıa, M. Maestre and P. Rueda, Spaceability in norm-attaining sets, Banach J. Math. Anal. 11 (2017), no. 1, 90 -- 107. 9. M. Fujii and R. Nakamoto, An estimation of the transcendental radius of an operator, Math. Japon. 27 (1982), 637 -- 938. 10. P. Ghosh, D. Sain and K. Paul, On symmetry of Birkhoff -- James orthogonality of linear operators, Adv. Oper. Theory 2, (2017), no. 4, 428 -- 434. 11. D. Kecki´c, Orthogonality in C1 and C∞ spaces and normal derivations, J. Operator Theory 51 (2004), 89 -- 104. 12. B. Magajna, On the distance to finite dimensional subspaces in operator algebras, J. London. Math. Soc. 47 (1993), 516 -- 532. 13. M. S. Moslehian and A. Zamani, Characterizations of operator Birkhoff -- James orthogonal- ity, Canad. Math. Bull. 60 (2017), no. 4, 816 -- 829. 14. K. Paul, Translatable radii of an operator in the direction of another operator, Sci. Math. 2 (1999), 119 -- 122. 15. K. Paul, D. Sain, A. Mal and K. Mandal, Orthogonality of bounded linear operators on complex Banach spaces, Adv. Oper. Theory 3, (2018), no. 3, 699 -- 709. 16. J.G. Stampfli, The norm of derivation, Pacific J. Math. 33 (1970), 737 -- 747. 17. P. W´ojcik, Gateaux derivative of the norm in K(X; Y ), Ann. Funct. Anal. 7 (2016), no. 4, 678 -- 685. 18. P. W´ojcik, The Birkhoff Orthogonality in pre-Hilbert C ∗-modules, Oper. Matrices 10 (2016), no. 3, 713 -- 729. 19. P. W´ojcik, Birkhoff orthogonality in classical M -ideals, J. Aust. Math. Soc. 103 (2017), 279 -- 288. 20. A. Zamani, M.S. Moslehian, M.T. Chien and H. Nakazato, Norm-parallelism and the Davis -- Wielandt radius of Hilbert space operators, to appear in Linear Multilinear Algebra, DOI: 10.1080/03081087.2018.1484422. Department of Mathematics, Farhangian University, Tehran, Iran. E-mail address: [email protected]
1602.05079	1	1602	2016-02-16T16:21:09	Joint spectra and nilpotent Lie algebras of linear transformations	[ "math.FA" ]	Given a complex nilpotent finite dimensional Lie algebra of linear transformations $L$, in a complex finite dimensional vector space $E$, we study the joint spectra $Sp(L,E)$, $\sigma_{\delta,k}(L,E)$ and $\sigma_{\pi,k}(L,E)$. We compute them and we prove that they all coincide with the set of weights of $L$ for $E$. We also give a new interpretation of some basic module operations of the Lie algebra $L$ in terms of the joint spectra.	math.FA	math	JOINT SPECTRA AND NILPOTENT LIE ALGEBRAS OF LINEAR TRANSFORMATIONS ENRICO BOASSO Abstract. Given a complex nilpotent finite dimensional Lie algebra of linear transformations L, in a complex finite dimensional vector space E, we study the joint spectra Sp(L, E), σδ,k(L, E) and σπ,k(L, E). We compute them and we prove that they all coincide with the set of weights of L for E. We also give a new interpretation of some basic module operations of the Lie algebra L in terms of the joint spectra. 1. Introduction The Taylor joint spectrum is one of the most dynamic and powerful subjects of multiparameter spectral theory. This spectrum, which was introduced in 1970 by J. L. Taylor in [9], associates to an n-tuple a = (a1, . . . , an) of mutually commuting bounded linear operators on a Banach space E, i. e., ai ∈ L(E), the algebra of all bounded linear operators on E, and [ai, aj] = 0, ∀i, j, 1 ≤ i, j ≤ n, a compact non empty subset of Cn, which we denote by Sp(a, E). Besides, when n = 1, the Taylor joint spectrum reduces to the classical spectrum of a single operator and, if a = (ai, . . . , ak) is the k-tuple, 1 ≤ k ≤ n − 1, of the first k operators of a, then the Taylor joint spectrum satisfies the so called projection property: Sp(a, E) = πSp(a, E), where π is the projection map of Cn onto Ck. In addition, if E is a complex finite dimensional vector space, M. Ch`o and M. Takaguchi computed in [4] the Taylor joint spectum of an n-tuple a of commuting operators on E and they obtain that Sp(a, E) = σpt(a), the joint point spectrum of a, where σpt(a) = {λ ∈ Cn : ∃ x ∈ E, x 6= 0 : aj(x) = λjx, 1 ≤ j ≤ n}. Besides in [6], A. McIntosh, A. Pryde and W. Ricker, as a consequence of a more general result which concerns infinite dimensional spaces too, also computed the Taylor joint spectrum for an n-tuple of linear transformations on a complex finite dimensional vector space, and they compared it with other joint spectra. In [2] we defined a joint spectrum for complex solvable finite dimensional Lie al- gebras of operators L, acting on a Banach space E, and we denoted it by Sp(L, E). We also proved that Sp(L, E) is a compact non empty subset of L∗, and that the projection property for ideals still holds. Besides, when L is a commutative al- gebra, Sp(L, E) reduces to the Taylor joint spectrum in the following sense. If dim L = n, {ai}(1≤i≤n) is a basis of L and we consider the n-tuple a = (a1, . . . , an), then {(f (a1), . . . , f (an)) : f ∈ Sp(L, E)} = Sp(a, E), i. e., Sp(L, E), in terms of the basis of L∗ dual of {ai}(1≤i≤n), coincides with the Taylor joint spectrum of the n-tuple a. Then, the following question arises naturally. If L is a complex Lie algebra of linear transformations in a complex finite dimensional vector space, what can be said about its joint spectrum? In this article we compute, under the 1 2 ENRICO BOASSO above assumptions, the joint spectrum of a nilpotent Lie algebra, moreover, we extend to the nilpotent case the characterizations of [4] and [6]. In addition, our result has a beautiful feature: Sp(L, E) is the set of all weights of the Lie algebra L for the vector space E. However, we consider a more general situation. In [1] we introduced a family of joint spectra for complex solvable finite dimensional Lie algebras of operators acting on a Banach space, which, when the algebra is a commutative one, reduce to the Slodkowski joint spectra in the same sense that we have explained for the Taylor joint spectrum, see [8]. We also proved the most important spectral properties for these joint spectra: compactness, non emptiness and the projec- tion property. In this article we also compute these joint spectra for complex nilpotent Lie algebras of linear transformations in a complex finite dimensional vector spaces. Moreover, we also prove that all the joint spectra considered in this article coincide with the set of weights of the Lie algebra L for the vector space E. We observe that, though all joint spectra introduced in [1] and [2] were defined for complex solvable Lie algebras, we restrict ourselves to the nilpotent case for, in the solvable non nilpotent case our result fails. The paper is organized as follows. In Section 2 we review several definitions and results of [1] and [2]. We also review several facts related to the theory of complex nilpotent Lie algebras of linear transformations in complex finite dimensional vector spaces. In Section 3 we prove our main theorems and, with them, we give a proof of the compactness and non emptiness of the joint spectra, for the case under consideration, different from the one of [1] and [2]. We also show an example of a solvable non nilpotent Lie algebra of linear transformations where our result fails. In Section 4 we consider some basic module operations of the Lie algebra L in order to compute several joint spectra. 2. Preliminaries We briefly recall several definitions and results related to the spectrum of a Lie algebra, see [2]. Indeed, in [2] we considered complex solvable finite dimensional Lie algebras of linear bounded operators acting on a Banach space, however, for our purpose, in this article we restrict ourselves to the case of complex fi- nite dimensional nilpotent Lie algebras of linear transformations defined on finite dimensional vector spaces. From now on, E denotes a complex finite dimensional vector space, L(E) the algebra of all linear transformations defined on E and L a complex nilpotent finite dimensional Lie subalgebra of L(E)op, i. e., the algebra L(E) with its opposite product. Such an algebra is called a nilpotent Lie algebra of linear transformations in the vector space E. If dim L = n and f is a character of L, i. e., f ∈ L∗ and f (L2) = 0, where L2 = {[x, y] : x, y ∈ L}, then we consider the following chain complex, (E ⊗ ∧L, d(f )), where ∧L denotes the exterior algebra of L, and dp(f ) is as follows, dp(f ) : E ⊗ ∧pL → E ⊗ ∧p−1L, dp(f )ehx1 ∧ · · · ∧ xpi = (−1)k+1e(xk − f (xk))hx1 ∧ . . . ∧ xk ∧ · · · ∧ xpi Xk=1 + X1≤k<l≤p NILPOTENT LIE ALGEBRAS 3 p (−1)k+leh[xk, xl] ∧ x1 . . . ∧ xk ∧ . . . ∧ xl ∧ . . . ∧ xpi, wheremeans deletion. If p ≤ 0 or p ≥ n + 1, we define dp(f ) = 0. If we denote by H∗((E⊗∧L, d(f ))) the homology of the complex (E⊗∧L, d(f )), we may state our first definition. Definition 2.1. With E, L and f as above, the set {f ∈ L∗ : f (L2) = 0, H∗((E ⊗ ∧L, d(f ))) 6= 0} is the joint spectrum of L acting on E and it is denoted by Sp(L, E). As we have said, in [2] we proved that Sp(L, E) is a compact non empty subset of L∗, which reduces to the Taylor joint spectrum when L is a commutative algebra, in the sense explained in the Introduction. Besides, if I is an ideal of L and π denotes the projection map from L∗ to I ∗, then, Sp(I, E) = π(Sp(L, E)), i. e., the projection property for ideals still holds. With regard to this property, we ought to mention the paper of C. Ott [7], who pointed out a gap in the proof of this result and gave another proof of it. In any case, the projection property remains true. We now give the definition of the Slodkowski joint spectra for Lie algebras. As in [1], we give an homological version. However, as we deal with linear transforma- tions defined on finite dimensional vector spaces, we slightly modify our definition in order to adapt it to our case, for a complete exposition of the subject see [8] and [1]. If L and E are as above, let us consider the set Σp(L, E) = {f ∈ L∗ : f (L2) = 0, Hp((E ⊗ ∧L, d(f ))) 6= 0}, where 0 ≤ p ≤ n, and dim L = n. We now state our definition of the S lodkowski joint spectra for Lie algebras. Definition 2.2. With L and E as above, σδ,k and σπ,k are the following joint spectra, σδ,k(L, E) = [0≤p≤k Σp(L, E), σπ,k(L, E) = [k≤p≤n Σp(L, E), where 0 ≤ k ≤ n, and n = dim L. We observe that σδ,n(L, E) = σπ,0(L, E) = Sp(L, E). Besides, in [1] we showed that σδ,k(L, E) and σπ,k(L, E), 0 ≤ k ≤ n, are compact non empty subsets of L∗, which reduce to the S lodkowski joint spectra when L is a commutative algebra. In addition, they also satisfy the projection property for ideals, i. e., if I is an ideal of L and π denotes the projection map from L∗ to I ∗, then, σδ,k(I, E) = π(σδ,k(L, E)), σπ,k(I, E) = π(σπ,k(L, E)), 4 ENRICO BOASSO where k and n are as above. We shall have occasion to use the theory of weight spaces, however, as we deal with complex nilpotent Lie algebras of linear transformations in complex finite dimensional vector spaces, we restrict our revision to the most important results of the theory, essentially Theorems 7 and 12, Chapter II, of the book by N. Jacobson, 'Lie Algebras'. For a complete exposition of the subject see [5, Chapter II]. Let L and E be as above. A weight of L for E is a mapping, α : x → α(x), of L into C such that there exists a non zero vector v in E with the property, (x − α(x))mv,x(v) = 0, for all x in L and where mv,x belongs to N. The set of vectors, zero included,which satisfy this condition is a subspace of E, Eα, called the weight space of E corresponding to the weight α, Eα = {v ∈ E : ∀x ∈ L there exists mv,x ∈ N such that:(x − α(x))mv,x(v) = 0}. As a consequence of our assumptions we have the following properties, see [5, Chapter II, Theorems 7, 12]: (i) the weights are linear functions on L, which vanish on L2, i. e., they are characters of L, (ii) E has only a finite number of distinct weights; the weight spaces are submod- ules, and E is the direct sum of them, (iii) for each weight α, the restriction of any x ∈ L to Eα has only one character- istic root, α(x), with certain multiplicity, (iv) there is a basis of E such that for each weight α the matrices of elements of L in the weight space Eα can be taken simultaneously in the form, xα = (cid:18)α(x) 0 ∗ α(x)(cid:19) . Finally, if L is a complex nilpotent finite dimensional Lie algebra, by [3, Chapter IV, Section 1], there is a Jordan-Hlder sequence of ideals (Li)(1≤i≤n) such that, (i) {0} = L0 ⊆ Li ⊆ Ln = L, (ii) dim Li = i, (ii) there is a k, 0 ≤ k ≤ n, such that Lk = L2, (iv) If i < j, [Li, Lj] ⊆ Li−1. As a consequence, if we consider a basis of L, {xj}(1≤j≤n), such that {xj}(1≤j≤i) is a basis of Li, we have that where i < j. [xj, xi] = i−1 Xh=1 ch ijxh, 3. The Main Result In this section, we compute the joint spectra of a nilpotent Lie algebra of linear transformations in a complex finite dimensional vector space. Besides, as we have said in the introduction, and under our assumptions, we give an elementary proof of the compactness and the non emptiness of these joint spectra. We also consider NILPOTENT LIE ALGEBRAS 5 an example which show that, in the solvable non nilpotent case, our result fails. Let us begin with our characterization of the joint spectra. Theorem 3.1. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie subalgebra of L(E)op. Then, if α is a weight of L for E, α belongs to Sp(L, E). Proof. As E is the direct sum of its weight spaces, which are L-modules of E, by the definition of the joint spectrum and elementary homological algebra, we may assume that there is only one weight of L for E. Let us denote it by α. We now observe that, as α is a character of L, we may consider the chain complex (E ⊗ ∧L, d(α)). Moreover, if {xi}(1≤i≤n) is the basis of L defined at the end of Section 2, it is easy to verify that, dn(α)ehx1 ∧ . . . ∧ xni = n Xk=1 (−1)k+1e(xk − α(xk))hx1 ∧ . . . ∧ xk ∧ . . . ∧ xni. However, as by the properties of the weights reviewed in Section 2, there exists an ordered basis of E, {ei}(1≤i≤m), where m = dim(E), such that, xk − α(xk) = (cid:18)0 ∗ 0 0(cid:19) , for all k, 1 ≤ k ≤ n. Thus, we have that, In particular, e1(xk − α(xk)) = 0. dn(α)e1hx1 ∧ . . . ∧ xni = 0, which implies that α belongs to Sp(L, E). (cid:3) We now state one of our main results. Theorem 3.2. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie subalgbra of L(E)op. Then, Sp(L, E) = {α ∈ L∗ : α is a weight of L for E}. Proof. By Theorem 3.1, it is enough to see that, Sp(L, E) ⊆ {α ∈ L∗ : α is a weight of L for E}. As in Theorem 3.1 we may suppose that there is only one weight of L for E, which we still denote by α. In order to conclude the proof, we consider β ∈ L∗, β(L2) = 0, such that β is not a weight of L for E, i. e., in our case α 6= β, and by refining an argument of [2], we prove that β does not belong to Sp(L, E). Let us consider the chain complex associated to β, (E ⊗∧L, d(β)), the sequence of ideals (Li)(1≤i≤n), and the basis {xi}(1≤i≤n) of L reviewed in Section 2. As α 6= β, there is a j, k + 1 ≤ j ≤ n, where k = dim L2 and n = dim L, such that α(xj) 6= β(xj). Now, if L′ is the ideal of codimension 1 of L generated by 6 ENRICO BOASSO {xi}(1≤i≤n,i6=j), we have that L′ ⊕hxji = L. Moreover, we may deconpose E ⊗∧pL in the following way, E ⊗ ∧pL = (E ⊗ ∧pL′) ⊕ ((E ⊗ ∧p−1L′) ∧ hxji). If β denotes the restriction of β to L′, then we may consider the complex (E ⊗ ∧L′, d( β)) and, as L′ is an ideal of codimension 1 of L, we may decompose dp(β) as follows, dp(β) : E ⊗ ∧pL′ → E ⊗ ∧p−1L′, dp(β) = dp( β), dp(β) : E ⊗ ∧p−1L′ ∧ hxji → (E ⊗ ∧p−1L′) ⊕ (E ⊗ ∧p−2L′ ∧ hxji), dp(β)(a ∧ hxji) = (−1)pLp−1(a) + (dp−1( β)(a)) ∧ hxji, where a ∈ E ⊗∧p−1L′ and Lp−1 is the linear endomorphism defined on E ⊗∧p−1L′ by Lp−1ehy1 ∧ . . . ∧ yp−1i =e(xj − β(xj))hy1 ∧ . . . ∧ yp−1i + X1≤l≤p−1 (−1)leh[xj, yl] ∧ y1 ∧ . . . ∧ yl ∧ . . . ∧ yp−1i, wheremeans deletion and {yh}(1≤h≤p−1) belong to L′. For s ∈ [[1, j − 1]], [xj, xs] = Ps−1 h=1 ch Pj−1 h=1(−ch sjxh, and for s ∈ [[j + 1, n]], [xj, xs] = js)xh. Thus, by the properties of weights considered in Section 2, it is easy to see that for each p there is a basis of E ⊗ ∧pL′ such that in this basis Lp is an upper triangular matrix with diagonal entries α(xj) − β(xj). However, as α(xj) 6= β(xj), Lp is an invertible map for each p. Thus, as in [2], we may construct an homotopy operator for the complex (E ⊗ ∧L, d(β)). However, this fact implies that H∗((E ⊗ ∧L, d(β))= 0, or equivalently, β does not belong to Sp(L, E) (cid:3) As a consequence of Theorem 3.2, we have, under our assumptions, another proof of the compactness and non emptiness of the spectrum. Theorem 3.3. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie subalgebra of L(E)op. Then, Sp(L, E) is a compact non empty subset of L∗. We now consider the joint spectra σδk(L, E) and σπk(L, E), 0 ≤ k ≤ n. We first state a lemma which we need for our characterization of these joint spectra. Lemma 3.4. Let E be a complex finite dimensional vector space and L a complex nilpotent Lie subalgebra of L(E)op such that dim L = n. Then, (i) Σn(L, E) = Sp(L, E), (ii) Σ0(L, E) = Sp(L, E). NILPOTENT LIE ALGEBRAS 7 Proof. If one looks at Theorem 3.1, a consequence of its proof is the fact that, if α is a weight of L for E, then α ∈ Pn(L, E). However, by Theorem 3.2 we have: Sp(L, E) ⊆ Σn(L, E) ⊆ Sp(L, E). In order to prove (ii), we consider the chain complex involved in the definition of Sp(L, E), (E ⊗ ∧L, d(f )), where f is a character of L, and we study d1(f ). We shall see that if α is a weight of L for E, d1(α) is not onto, equivalently, 0 6= E/R(d1(α)) = H0(E ⊗ ∧L, d(α)), i. e., α belongs to P0(L, E). Then, by Theorem 3.2, as in the proof of (i), we conclude (ii). Let us prove the previous claim. We consider α a weight of L for E. Besides, by elementary homological arguments and the properties of weights, we may suppose that E = Eα. If this is the case, let us consider the chain complex (E ⊗ ∧L, d(α)) and the map d1(α), d1(α) : E ⊗ ∧1L → E, d1(α)(ehxi) = e(x − α(x)). However, by the revision of Section 2, there is a basis of E, {ei}(1≤i≤m), where m = dim(E), such that the matrix of any x − α(x), in this basis, has the form, x − α(x) = (cid:18)0 ∗ 0 0(cid:19) . Then, em does not belong to R(d0(α)), which implies our claim. (cid:3) We now may study the relation among Sp(L, E), σδk(L, E) and σπk(L, E), 0 ≤ k ≤ n, for nilpotent Lie algebras of linear transformations in complex finite dimensional vector spaces. The following theorem states the relation among then. Theorem 3.5. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie algebra of L(E)op. Then, we have the following identity, Sp(L, E) = σδk(L, E) = σπk(L, E), where 0 ≤ k ≤ n, and dim(L) = n. Proof. The proof is a consequence of the following observation. By Lemma 3.4 and Definitions 2.1 and 2.2, we have that, Sp(L, E) = Σ0(L, E) ⊆ σδk(L, E) ⊆ Sp(L, E), Sp(L, E) = Σn(L, E) ⊆ σπk(L, E) ⊆ Sp(L, E), where 0 ≤ k ≤ n, and dim(L) = n. (cid:3) As a consequence of Theorems 3.3 and 3.5, we have the following result. Theorem 3.6. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie subalgebra of L(E)op. Then σδk(L, E) and σπk(L, E) are com- pact non empty subsets of L∗, where 0 ≤ k ≤ n, and n = dim(L). 8 ENRICO BOASSO We now give an example in order to show that, for the solvable non nilpotent case, our characterization of the joint spectra fails. Let us consider the solvable Lie algebra on two generators, L = hyi ⊕ hxi, with the bracket [x, y]op = y, i. e., [x, y] = −y. L may be viewed as a subalgebra of L(C2)op as follows, y = (cid:18) 1 −1 −1(cid:19) , 1 x = (cid:18)0 1 2 0(cid:19) . 2 1 An easy calculation shows that the weights of L for C2 are, in terms of the 2 ). However, if we still consider the previous dual basis of y and x, (0, 1 basis, Sp(L, C2) = {(0, 1 2) and (0, −1 2), (0, −3 2 )}. 4. Some Consequences In this section we shall see some consequences of the main theorems. We begin with a result which relates the weights of differents ideals. Theorem 4.1. Let E be a complex finite dimensional vector space, L a complex nilpotent Lie subalgebra of L(E)op and I an ideal of L. Then, if α is a weight of L for E, its restriction to I, α I, is a weight of I for E. Reciprocally, if β is a weight of I for E, there exists α, a weight of L for E, such that β = α I. In particular, if I is contained in L2, I has only one weight for E: β = 0. Proof. The proof is a consequence of Theorem 3.2 and Theorem 3 of [2], the projection property for ideals of the spectrum. (cid:3) From now to the end of this section, we deal with two basic module operations of the Lie algebra L, the contragradient module and the tensor product, see [5, Chapter III]. We shall compute several new joint spectra by means of these module operations. Moreover, these results give a new interpretation of these module operations in terms of the joint spectra. More precisely, we consider Propositions 3 and 4 of [5, Chapter III], and we state then as properties of the joint spectra. Let us begin with the contragradient module. If L and E are as usual, we consider the space of linear functionals on E, E∗. If x belongs to L and x∗ denotes its adjoint, then it is easy to see that the set L′ L′ = {x∗ : x ∈ L}, with the natural bracket, defines a complex nilpotent Lie subalgebra of L(E∗)op, which is isomorphic to L with the opposite bracket. We now compute the joint spectra of L′ on E∗. Theorem 4.2. Let E be a complex finite dimensional vector space and L a com- plex nilpotent Lie subalgebra of L(E)op. Let E∗ be the dual vector space of E and L′ the nilpotent Lie subalgbra of L(E∗)op defined above. Then, Sp(L′, E∗) = Sp(L, E), σδk(L′, E∗) = σδk(L, E), σπk(L′, E∗) = σπk(L, E), where 0 ≤ k ≤ n, and n = dim(L) = dim(L′). NILPOTENT LIE ALGEBRAS 9 Proof. As we may consider the inclusion as a representation of L in L(E)op, the proof is a consequence of a slight modification of the proof of [5, Chapter III, Proposition 3] and Theorems 3.2 and 3.5. (cid:3) Finally, we study the joint spectrum of a tensor product. If E1 and E2 are two complex finite dimensional vector spaces and L1 and L2 are two complex nilpotent Lie subalgebras of L(E1)op and L(E2)op, respectively, then we consider the tensor product of E1 and E2, E1 ⊗ E2, and the nilpotent Lie subalgebras of L(E1 ⊗ E2)op, L1 and L2, defined by, L1 = {x ⊗ 1 : x ∈ L1}, L2 = {1 ⊗ y : y ∈ L2}, where 1 denotes the identity of the corresponding space. As in L(E1 ⊗ E2), [ L1, L2] = 0, the set, {x ⊗ 1 + 1 ⊗ y : x ∈ L1, y ∈ L2}, may be viewed as the direct product of the Lie algebras L1 and L2, L1 × L2, with the natural bracket of L(E1 ⊗ E2)op. Besides, L1 and L2 are two ideals of the complex nilpotent Lie algebra L1 × L2. Our objective is to compute the joint spectra of L1 × L2 on E1 ⊗ E2. The following theorem describes the situation. Theorem 4.3. Let E1 and E2 be two complex finite dimensional vector spaces and L1 and L2 two complex nilpotent subalgebras of L(E1)op and L(E2)op, respectively. Then, Sp(L1 × L2, E1 ⊗ E2) = Sp(L1, E1) × Sp(L2, E2), σδk(L1 × L2, E1 ⊗ E2) = Sp(L1, E1) × Sp(L2, E2), σπk(L1 × L2, E1 ⊗ E2) = Sp(L1, E1) × Sp(L2, E2), where 0 ≤ k ≤ n + m, n = dim(L1) and m = dim(L2). Proof. Let us consider the decomposition of E1, respectively E2, as direct sum of its weight L1-submodules, respectively its weight L2-submodules, E1 = ⊕α∈ΦE1α, E2 = ⊕β∈ΨE2β, where Φ = {α ∈ L∗ L∗ 2 : β is weight of L2 for E2} = Sp(L2, E2). Then, 1 : α is weight of L1 for E1} = Sp(L1, E1) and, Ψ = {β ∈ E1 ⊗ E2 = ⊕(α,β)∈Φ×ΨE1α ⊗ E2β. An easy calculation shows that each E1α ⊗ E2β, (α, β) ∈ Φ × Ψ, is a L1 × L2- submodule of E1 ⊗ E2. Besides, as E1α 6= 0 and E2β 6= 0, E1α ⊗ E2β 6= 0. In addition, by the definition of Φ and Ψ, if (α1, β1) 6= (α2, β2), there exists an element of L1 × L2, (x, y), such that (α1, β1)(x, y) 6= (α2, β2)(x, y). Finally, in order to conclude the proof, by Theorem 3.5 and [5, Chapter II, Theorem 7], it is enough to prove that E1α ⊗ E2β is the weight module of L1 × L2 for the weight (α, β). However, this fact may be verify by a slight modification of [5, Chapter III, Proposition 4]. (cid:3) 10 ENRICO BOASSO References 1. E. Boasso, Dual properties and joint spectra for solvable Lie algebras of operators, J. Operator Theory 33 (1995), 105-116. 2. E. Boasso and A. Larotonda, A spectral theory for solvable Lie algebras of operators, Pacific J. Math. 158 (1993), 15-22. 3. N. Bourbaki, ´El´ements de Math´ematique, Groupes et Alg`ebres de Lie, Alg`ebres de Lie Fasc. XXVI, 1960. 4. M. Ch`o and M. Takaguchi, Joint spectra of matrices, Sci. Rep. Hirosaki Univ. 26 (1979), 15-19. 5. N. Jacobson, Lie Algebras, Interscience Publishers, 1962. 6. A. McIntosh, A. Pryde and W. Ricker, Comparison of joint spectra for certain classes of commuting opertors, Studia Math. 88 (1988), 23-36. 7. C. Ott, A note on a paper of E. Boasso and A. Larotonda, Pacific J.Math. 173 (1996), 173-179. 8. Z. S lodkowski, An infinite family of joint spectra, Studia Math. 61 (1973), 239-255. 9. J.L.Taylor, A joint spectra for several commuting operators, J. Funct. Anal. 6 (1970), 172- 191. Enrico Boasso E-mail address: enrico [email protected]
1108.0460	4	1108	2012-07-25T15:56:08	$\alpha$-Modulation Spaces (I)	[ "math.FA" ]	First, we consider some fundamental properties including dual spaces, complex interpolations of $\alpha$-modulation spaces $M^{s,\alpha}_{p,q}$ with $0<p,q \le \infty$. Next, necessary and sufficient conditions for the scaling property and the inclusions between $\alpha_1$-modulation and $\alpha_2$-modulation spaces are obtained. Finally, we give some criteria for $\alpha$-modulation spaces constituting multiplication algebra. As a by-product, we show that there exists an $\alpha$-modulation space which is not an interpolation space between modulation and Besov spaces.	math.FA	math	α-Modulation Spaces (I) Scaling, Embedding and Algebraic Properties Jinsheng Han Department of Mathematics, Shanghai Jiao Tong University, Shanghai 200240, P. R. China. Baoxiang Wang LMAM, School of Mathematics, Peking University and BICMR, Beijing 100871, P. R. China. ([email protected], [email protected]) Abstract. First, we consider some fundamental properties including dual spaces, com- p,q with 0 < p, q ≤ ∞. Next, necessary and suf- plex interpolations of α-modulation spaces M s,α ficient conditions for the scaling property and the inclusions between α1-modulation and α2- modulation spaces are obtained. Finally, we give some criteria for α-modulation spaces con- stituting multiplication algebra. As a by-product, we show that there exists an α-modulation space which is not an interpolation space between modulation and Besov spaces. In a sub- sequent paper, we will give some applications of α-modulation spaces to nonlinear dispersive wave equations. Key words. α-modulation space, Dual space, Multiplication algebra, Scaling property, Embedding. AMS subject classifications. 42 B35, 42 B37, 35 A23 1 Introduction and definition decomposition Rn = {ξ : ξ < 1}(cid:83)((cid:83)∞ decomposition Rn =(cid:83) Frequency localization technique plays an important role in the modern theory of function spaces. There are two kinds of basic partitions to the Euclidean space Rn, one is the dyadic j=1{ξ : ξ ∈ [2j−1, 2j)}), another is the uniform k∈Zn(k+[−1/2, 1/2)n). According to these two kinds of decompositions in frequency spaces, one can naturally introduce the dyadic decomposition operators (cid:52)j (j ∈ Z+) whose symbol ϕj is localized in {ξ : ξ ∼ 2j}, and the uniform decomposition operator (cid:3)k (k ∈ Zn) whose symbol σk is supported in k + [−1, 1]n. The difference between ϕj and σk is that the diameters of supp ϕj and supp σk are O(2j) and O(1), respectively. All tempered distributions acted on these decomposition operators with finite (cid:96)q(Lp) (quasi)- norms constitute Besov space Bs The α-modulation spaces M s,α p,q and modulation space M s p,q , introduced by Grobner [11], are proposed to be intermedi- ate function spaces to connect modulation space and Besov space with respect to parameters k (k ∈ Zn). α ∈ [0, 1], which are formulated by some new kind of α-decomposition operators (cid:3)α We denote by ηα k , whose essential characteristic is that the diameter of its support set has power growth as (cid:104)k(cid:105)α/(1−α). p,q, respectively. k the symbol of (cid:3)α Modulation spaces are special α-modulation spaces in the case α = 0, and Besov space can be regarded as the limit case of α-modulation space when α (cid:37) 1. Modulation spaces were first introduced by Feichtinger [8] in the study of time-frequency analysis to consider the decay 1 property of a function in both physical and frequency spaces. His original idea is to use the short-time Fourier transform of a tempered distribution equipping with a mixed Lq(Lp)-norm to generate modulation spaces M s p,q. Grochenig's book [12] systematically discussed the the- ory of time-frequency analysis and modulation spaces. In Grobner's doctoral thesis, he used the α-covering to the frequency space Rn and a corresponding bounded admissible partition of unity of order p (p-BAPU) to define α-modulation spaces. Some recent works have been devoted to the study of α-modulation spaces (see [1, 2, 7, 10, 14, 13] and references therein). Borup and Nielsen [1] and Fornasier [10] constructed Banach frames for α-modulation spaces in the multivariate setting, Kobayashi, Sugimoto and Tomita [14, 13] discussed the bounded- ness for a class of pseudo-differential operators with symbols in α-modulation spaces. Dahlke, Fornasier, Rauhut, Steidl and Teschke [7] established the relationship between the generalized coorbit theory and α-modulation spaces. The aim of the present paper is to describe some standard properties including the dual spaces, embeddings, scaling and algebraic structure of α-modulation spaces. Before stating the notion of α-modulation spaces, we introduce some notations frequently used in this paper. A (cid:46) B stands for A ≤ CB, and A ∼ B denote A (cid:46) B and B (cid:46) A, where C is a positive constant which can be different at different places. Let S (Rn) be the Schwartz space and S (cid:48)(Rn) be its strongly topological dual space. Suppose f ∈ S (cid:48)(Rn) and λ > 0, we write fλ(·) = f (λ·). Let X be a (quasi-)Banach space, we denote by X∗ the dual space of X. For any p ∈ [1,∞], p∗ will stand for the dual number of p, i.e., 1/p + 1/p∗ = 1. We denote by Lp = Lp(Rn) the Lebesgue space for which the norm is written by (cid:107) · (cid:107)p, and by (cid:96)p the sequence Lebesgue space. We will write (cid:104)x(cid:105) = (1 + x2)1/2. For any multi-index δ = (δ1, δ2,··· , δn), we denote Dδ = ∂δ1 n . It is convenient to divide Rn into n parts j , j = 0, 1, 2,··· , n: Rn j = {x ∈ Rn : xi (cid:54) xj, i = 1,··· , j − 1, j + 1,··· , n} . Rn 2 ··· ∂δn 1 ∂δ2 We write J = (I − ∆)s/2 and define the Sobolev space H s(Rn) =(cid:8)f ∈ S (cid:48)(Rn) : (cid:107)f(cid:107)H s = (cid:107)J sf(cid:107)2 < ∞(cid:9) (cid:13)(cid:13)(cid:13)(cid:96)q (cid:110){gk}k∈Zn : gk ∈ S (cid:48)(Rn), (cid:13)(cid:13)(cid:13)(cid:104)k(cid:105) s 1−α(cid:107)gk(cid:107)p < ∞(cid:111) . and s,α(Zn; Lp) = (cid:96)q Without additional note, we will always assume that s ∈ R, 0 < p, q (cid:54) ∞, 0 (cid:54) α < 1. k }k∈Zn satisfies Let us start with the third partition of unity on frequency space for α ∈ [0, 1) (see [1]). We suppose c < 1 and C > 1 are two positive constants, which relate to the space dimension n, if(cid:12)(cid:12)ξ − (cid:104)k(cid:105) α and a Schwartz function sequence {ηα k ⊂(cid:8)ξ :(cid:12)(cid:12)ξ − (cid:104)k(cid:105) α k (ξ) (cid:38) 1, ηα (cid:88) suppηα k (ξ)(cid:12)(cid:12) (cid:46) 1, 1−α(cid:12)(cid:12)Dδηα k∈Zn (cid:104)k(cid:105) αδ 1−α k(cid:12)(cid:12) < c(cid:104)k(cid:105) α 1−α k(cid:12)(cid:12) < C(cid:104)k(cid:105) α 1−α(cid:9); k (ξ) ≡ 1, ηα ∀ξ ∈ Rn. ∀ξ ∈ Rn; (1.1d) (1.1a) (1.1b) (1.1c) 1−α ; 2 We denote Corresponding to every sequence {ηα denoted by {(cid:3)α k}k∈Zn, and Υ =(cid:8){ηα k }k∈Zn satisfies (1.1)(cid:9) k }k∈Zn : {ηα k F . (1.2) k }k∈Zn ∈ Υ, one can construct an operator sequence k = F −1ηα (cid:3)α (cid:32) (cid:32)(cid:88) ξ − (cid:104)k(cid:105) α 1−α k C(cid:104)k(cid:105) α 1−α (cid:33) (cid:33)−1 (1.3) (1.4) ρα k (ξ) = ρ ηα k (ξ) = ρα k (ξ) ρα l (ξ) . l∈Zn Υ is nonempty. Indeed, let ρ be a smooth radial bump function supported in B(0, 2), satis- fying ρ(ξ) = 1 as ξ < 1, and ρ(ξ) = 0 as ξ (cid:62) 2. For any k ∈ Zn, we set and denote k }k∈Zn satisfies (1.1). This type of decomposition on frequency It is easy to verify that {ηα space is a generalization of the uniform decomposition and the dyadic decomposition. When 0 (cid:54) α < 1, on the basis of this decomposition, we define the α-modulation space by (cid:110) f ∈ S (cid:48)(Rn) : (cid:107)f(cid:107)M s,α s,α(Zn;Lp) < ∞(cid:111) = (cid:107){(cid:3)α k f}k∈Zn(cid:107)(cid:96)q M s,α p,q (Rn) = (1.5) Denote ϕ(ξ) = ρ(ξ) − ρ(2ξ), we may assume ϕ(ξ) = 1 if 5/8 ≤ ξ ≤ 3/2. We introduce the function sequence {ϕk}∞ p,q . k=0: (cid:26) ϕj(ξ) = ϕ(2−jξ), ϕ0(ξ) = 1 −(cid:80)∞ j ∈ N, j=1 ϕj(ξ). (1.6) (1.7) Define (cid:52)j = F −1ϕjF , j ∈ N ∪ {0}, {(cid:52)j}∞ j=0 is said to be the Littlewood -- Paley (or dyadic) decomposition operators. Denote p,q =(cid:13)(cid:13){2sj(cid:52)jf}j∈N∪{0}(cid:13)(cid:13)(cid:96)q(Lp) < ∞(cid:111) . (1.8) p,q(Rn) = Bs f ∈ S (cid:48)(Rn) : (cid:107)f(cid:107)Bs (cid:110) Strictly speaking, (1.5) cannot cover the case α = 1, however, we will denote M s,1 convenience. p,q = Bs p,q for The paper is organized as follows. In Section 2, we show some basic properties on α- modulation spaces, their dual and complex interpolation spaces are presented there. In Section 3, we discuss the scaling property. In Section 4, the inclusions between α-modulation spaces for different indices α (including Besov spaces) are obtained. In Section 5, we study the regularity conditions so that α-modulation spaces form multiplication algebra. Finally, we show the necessity for the conditions of scalings, embeddings and algebra structures by constructing several counterexamples. 3 2 Some basic properties In the sequel, we give some basic properties of M s,α Proposition 2.1 ([18], Convolution in Lp with p < 1). Let 0 < p ≤ 1. Lp Lp(Rn) : suppf ⊂ B(x0, R)}, B(x0, R) = {x : x − x0 ≤ R}. Suppose that f, g ∈ Lp then there exists a constant C > 0 which is independent of x0 and R > 0 such that p,q . We need the following B(x0,R) = {f ∈ B(x0,R), (cid:107)f ∗ g(cid:107)p ≤ CRn(1/p−1)(cid:107)f(cid:107)p(cid:107)g(cid:107)p. Proposition 2.2. ([18], Generalized Bernstein inequality) Let Ω ⊂ Rn be a compact set, 0 < r ≤ ∞. Let us denote σr = n(1/(r ∧ 1) − 1/2) and assume that s > σr. Then there exists a constant C > 0 such that (cid:107)F −1ϕF f(cid:107)r ≤ C(cid:107)ϕ(cid:107)H s(cid:107)f(cid:107)r Ω := {f ∈ Lp : supp(cid:98)f ⊂ Ω} and ϕ ∈ H s. Moreover, if r ≥ 1, then (2.1) k }k∈Zn ∈ Υ, then they genarate holds for all f ∈ Lr holds for all f ∈ Lr. Proposition 2.3 (Equivalent norm). Let {ηα equivalent quasi-norms on M s,α p,q . k }k∈Zn,{(cid:101)ηα (2.1) Proof. See [1]. Proposition 2.4 (Embedding). Let 0 < p1 (cid:54) p2 ≤ ∞, 0 < q1, q2 ≤ ∞. We have (i) if q1 (cid:54) q2 and s1 (cid:62) s2 + nα(cid:0) 1 (ii) if q1 > q2 and s1 > s2 + nα(cid:0) 1 p1 − 1 p2 − 1 p2 p1 (cid:1), then (cid:1) + n(1 − α)(cid:0) 1 p1,q1 ⊂ M s2,α M s1,α p2,q2; q2 p1,q1 ⊂ M s2,α M s1,α p2,q2. (cid:0) 1 (cid:1) (cid:46) (cid:104)k(cid:105) nα 1−α − 1 p2 p1 (cid:1), then − 1 q1 (cid:3) (2.2) (2.3) (2.4) (2.5) (cid:3) Proof. From Bernstein's inequality it follows that (cid:107)(cid:3)α k f(cid:107)p2 (cid:107)(cid:3)α k f(cid:107)p1. Then (i) follows from (cid:96)q1 ⊂ (cid:96)q2 and (2.4). For (ii), we use Holder's inequality to obtain (cid:107)f(cid:107)M s2,α p2,q2 (cid:46) (cid:107){(cid:3)α k f}k∈Zn(cid:107)(cid:96)q1 s1,α (cid:107){1}k∈Zn(cid:107) (cid:96)q1q2/(q1−q2) s2−s1+nα(1/p1−1/p2),α . For the second term in the right-hand side, we easily see that it is finite by changing the (cid:3) summation to an integration. is a quasi-Banach space, and is a Banach Proposition 2.5 (Completeness). (i) M s,α p,q space if 1 (cid:54) p (cid:54) ∞ and 1 (cid:54) q (cid:54) ∞. (ii) We have Moreover, if 0 < p, q < ∞, then S (Rn) is dense in M s,α p,q . S (Rn) ⊂ M s,α p,q (Rn) ⊂ S (cid:48)(Rn). Proof. See [3]. 4 Proposition 2.6 (Isomorphism). For any σ ∈ R, the mapping J σ : M s,α isomorphic. p,q → M s−σ,α p,q Proof. See [18] Proposition 2.7. M s,α 2,2 (Rn) = H s(Rn) with equivalent norms. Proof. Plancherel's identity implies the result, see [11]. 2.1 Duality is (cid:3) (cid:3) It is known that the dual space of Besov space Bs space of modulation space M s spaces of α-modulation spaces. Proposition 2.8. Suppose 1 (cid:54) p, q < ∞. Then we have p,q is M−s p,q is B (see [18]) and the dual (p∨1)∗,(q∨1)∗ (see [19]). In this section we study the dual −s+n(1/(p∧1)−1) (p∨1)∗,(q∨1)∗ More precisely, f ∈(cid:0)(cid:96)q (cid:96)q∗ −s,α(Zn; Lp∗ ) such that for any g = {gk}k∈Zn ∈ (cid:96)q s,α(Zn; Lp), we have is equivalent to that there exists a sequence {fk}k∈Zn ∈ = (cid:96)q∗ −s,α(Zn; Lp∗ ). (cid:0)(cid:96)q s,α(Zn; Lp)(cid:1)∗ s,α(Zn; Lp)(cid:1)∗ (cid:90) (cid:88) (cid:104)f, g(cid:105) = with (cid:107)f(cid:107)((cid:96)q s,α(Zn;Lp))∗ = (cid:107){fk}(cid:107) (cid:96)q∗ −s,α(Zn;Lp∗ . ) fk(x)gk(x)dx, k∈Zn Rn It is a direct consequence of Proposition 3.3 in [19]. Lemma 2.1. Let {gk}k∈Zn ∈ (cid:96)q s,α(Zn; Lp), and Γ be a subset of Zn, then we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:88) k∈Γ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,q (cid:3)α k gk (cid:46) (cid:107){(cid:3)α k gk}k∈Γ(cid:107)(cid:96)q s,α(Zn;Lp). (2.6) Proof. We introduce Λ(k) = {l ∈ Zn : (cid:3)α We denote the constant in (1.1b) relating to {ηα holds k l f (cid:54)= 0}. (cid:3)α (2.7) l }l∈Zn by C, thus for every l ∈ Λ(k), there (cid:104)k(cid:105) α (cid:104)k(cid:105) α 1−α (kj − C) < (cid:104)l(cid:105) α 1−α (kj + C) > (cid:104)l(cid:105) α 1−α (lj + C), 1−α (lj − C) (2.8a) (2.8b) with j = 1,··· , n. For the above l and k, we conclude that (cid:104)k(cid:105) ∼ (cid:104)l(cid:105). (2.9) If k (cid:46) 1 (or l (cid:46) 1), it is easy to see that (2.9) follows from (2.8). Thus, it suffices to show (2.9) in the case k (cid:29) 1 (or l (cid:29) 1). When k ∈ Rn j with kj > 0, from (2.8a); whereas when k ∈ Rn 1−α , and symmetrically, we j but with kj < 0, from (2.8b)×(−1), we see (cid:104)k(cid:105) 1 1−α (cid:46) (cid:104)l(cid:105) 1 5 have (cid:104)l(cid:105) 1 1−α (cid:46) (cid:104)k(cid:105) 1 1−α . Therefore, we get (2.9). Suppose both l and(cid:101)l are in Λ(k). substituting (cid:12)(cid:12)(cid:12)(cid:104)l(cid:105) α 1−α lj − (cid:104)(cid:101)l(cid:105) α 1−α(cid:101)lj l with(cid:101)l (2.8a) and (2.8b) also hold. It follows that (cid:12)(cid:12)(cid:12) (cid:46) (cid:104)k(cid:105) α Then Taylor's theorem, combined with (2.9), gives lj −(cid:101)lj (cid:46) 1. It follows that  1 (cid:88) One has that the right hand side of (2.6) is 1−α + (cid:104)(cid:101)l(cid:105) α #Λ(k) ∼ 1. 1−α + (cid:104)l(cid:105) α  1 k gk(cid:107)q (cid:3)α (cid:88) (cid:104)l(cid:105) sq 1−α (cid:104)l(cid:105) sq (cid:107)(cid:3)α 1−α . (cid:3)α (cid:3)α k gk (cid:46) p q q l l l∈Zn k∈Λ(l)∩Γ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (2.10) (2.11) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)q p l∈Zn (cid:88) (cid:88) (cid:32)(cid:88) k∈Γ k∈Γ (cid:46) (cid:46) 1−α (cid:88) (cid:88) k gk(cid:107)q p k∈Λ(l)∩Γ l∈Λ(k) (cid:107)(cid:3)α (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k gk(cid:107)q p  1 q (cid:104)l(cid:105) sq 1−α (cid:33) 1 q . We remark that in the second inequality of (2.11), by Young's inequality, or by Proposition 2.1 we see that l ; and in the third inequality of (2.11), we have applied (2.10) (cid:3) (cid:107)(cid:3)α k l f(cid:107)p (cid:46) (cid:107)(cid:3)α (cid:3)α k f(cid:107)p, which enable us to remove (cid:3)α to remove the summation on l ∈ Λ(k). Theorem 2.1. Suppose 0 < p, q < ∞, then we have −s+nα(cid:0) 1 1∧p−1(cid:1) (cid:1)∗ ⊂ M Case 1: 1 (cid:54)(cid:54)(cid:54) p, q < ∞. First, we show that(cid:0)M s,α Proof. The proof is separated into four cases. (cid:0)M s,α (1∨p)∗,(1∨q)∗ (cid:1)∗ = M p,q p,q . (2.12) −s,α p∗,q∗ . Noticing that M s,α k f} ∈ (cid:96)q p,q (cid:51) f → {2α s,α(Zn; Lp) p,q onto a subspace X = {{2α is an isometric mapping from M s,α so, any continuous functional g on M s,α X, which can be extended onto (cid:96)q g is preserved. By Proposition 2.8, there exists {gk} ∈ (cid:96)q∗ s,α(Zn; Lp), p,q can be regarded as a bounded linear functional on s,α(Zn; Lp) (the extension is written as g) and the norm of −s,α(Zn; Lp∗ k f} : f ∈ M s,α p,q } of (cid:96)q ) such that (cid:104)g,{fk}(cid:105) = gk(x)fk(x)dx (2.13) holds for all {fk} ∈ (cid:96)q isometric to X, we see that s,α(Zn; Lp). Moreover, (cid:107)g(cid:107)(M s,α p,q )∗ = (cid:107){gk}(cid:107) (cid:96)q∗ −s,α(Zn;Lp∗ . Since M s,α p,q is ) (cid:104)g, ϕ(cid:105) = (cid:104)g,{2α k ϕ}(cid:105) = 2α k gk(x)ϕ(x)dx, ϕ ∈ M s,α p,q , (2.14) (cid:90) (cid:88) k∈Zn (cid:90) (cid:88) k∈Zn 6 k∈Zn 2α Hence, g =(cid:80) which implies(cid:0)M s,α (cid:1)∗ ⊂ M S (cid:48), we show that f ∈(cid:0)M s,α p,q p,q k gk(x). In view of Lemma 2.1 and Young's inequality, (cid:107)g(cid:107)M (cid:46) (cid:107){gk}k∈Zn(cid:107) = (cid:107)g(cid:107)(M s,α −s,α p∗,q∗ (cid:1)∗ −s,α p∗,q∗ . Next, we prove the reverse inclusion. For any f ∈ M (cid:96)q∗ −s,α(Zn;Lp∗ p,q )∗, ) −s,α p∗,q∗ ⊂ k∈Zn s,α(Zn;Lp) p∧1−1(cid:1),α . (cid:32) (cid:104)(cid:3)α . p,q p,q . (cid:3)α l ϕ 1∨p,1∨q . ) l∈Λ(k) M 1∨p,1∨q  (cid:46) (cid:107)f(cid:107)M −s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q = M (1∨p)∗,(1∨q)∗ k f}k∈Zn(cid:107) (cid:1)∗ ⊃ p,q ⊂ M M s,α (cid:0)M s,α p,q (cid:104)f, ϕ(cid:105) (cid:54) (cid:107){(cid:3)α s−nα(cid:0) 1 The principle of duality implies M −s+nα(cid:0) 1 . Let ϕ ∈ S . We have This combined with the principle of duality gives Hence, only the reverse inclusion needs to be proven. (cid:96)q∗ −s,α(Zn;Lp∗ p∗,q∗(cid:107)ϕ(cid:107)M s,α −s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  (cid:88) p∗,q∗ ⊂(cid:0)M s,α (cid:1)∗ p∧1−1(cid:1),α s−nα(cid:0) 1 (cid:33)∗ p∧1−1(cid:1),α Case 2. 1 (cid:54)(cid:54)(cid:54) p < ∞, 0 < q < 1. For any f ∈(cid:0)M s,α (cid:1)∗ (cid:1)∗(cid:107)(cid:3)α (cid:54) (cid:107)f(cid:107)(cid:0)M s,α k ϕ(cid:105) 1−α(cid:107)f(cid:107)(cid:0)M s,α (cid:1)∗ k f (x) = (cid:104)f, F −1ηα = (cid:104)f, F −1ηα This implies(cid:0)M s,α Case 3. 0 < p, q < 1. For any f ∈(cid:0)M s,α (cid:46) (cid:107)f(cid:107)(cid:0)M s,α p−1(cid:1),α , take any k ∈ Zn, we have k (x − ·)(cid:105) (cid:1)∗(cid:107)F −1ηα k (· − x)(cid:105) (cid:0) 1 p−1(cid:1) (cid:107)f(cid:107)(cid:0)M s,α This implies(cid:0)M s,α Case 4. 0 < p < 1, 1 (cid:54)(cid:54)(cid:54) q < ∞. For any f ∈(cid:0)M s,α (cid:1)∗ k (· − x)(cid:107)M s,α (cid:1)∗. k ϕ(cid:107)M s,α (cid:1)∗(cid:107)ϕ(cid:107)p. −s+nα(cid:0) 1 k f, ϕ(cid:105) = (cid:104)f, (cid:3)α (cid:1)∗ ⊂ M (cid:1)∗ ⊂ M (cid:46) (cid:104)k(cid:105) s 1−α− nα 1−α (cid:46) (cid:104)k(cid:105) s −s,α p∗,∞. p,q p,q p,q ∞,∞ p,q p,q p,q . we have p,q (cid:3)α p,q p,q and every k ∈ Zn, there exists some p,q xk ∈ Rn satisfying (cid:107)(cid:3)α k f(cid:107)∞ ∼ F −1ηα k F f (xk). Let {ak} be an arbitrary (cid:96)q {(cid:101)ak}, such that (cid:101)ak = ak, and the argument of(cid:101)ak is the opposite number of the principal s−nα(1/p−1) sequence, and we construct another sequence namely 7 In the following, we discuss the left three cases. From (2.2) in Proposition 2.4, we know , take any k ∈ Zn and any ϕ ∈ S (Rn), p,q (2.15) (2.16) F f (xk). From (1.1b),(1.1d), we get (cid:107)F −1ηα k (· − xk)(cid:107)p (cid:46) (cid:104)k(cid:105)− nα 1−α argument of F −1ηα Therefore, we have(cid:88) k ak(cid:107)(cid:3)α k f(cid:107)∞ ∼ k∈Zn This implies(cid:0)M s,α p,q (cid:1)∗ ⊂ M −s+nα(cid:0) 1 ∞,q∗ (cid:43) (cid:101)akF −1ηα k (· − xk) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:88) (cid:1)∗ (cid:101)akF −1ηα (cid:1)∗(cid:107){ak}k∈Zn(cid:107)(cid:96)q k k (· − xk) . s−nα( 1 p −1) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,q k f, (cid:42) (cid:88) (cid:54) (cid:107)f(cid:107)(cid:0)M s,α (cid:46) (cid:107)f(cid:107)(cid:0)M s,α p−1(cid:1),α p,q p,q . (cid:0) 1 p−1(cid:1) . (cid:3) 2.2 Complex interpolation The complex interpolation for Besov spaces has a beautiful theory; cf. [18]. We can imitate the counterpart for the Besov space to construct the complex interpolation for α-modulation spaces. It will be repeatedly used in the following argument. Since there is little essential modification in the statement, we only provide the outline of the proof. We start with some abstract theory about complex interpolation on quasi-Banach spaces. Let S = {z : 0 < Rez < 1} be a strip in the complex plane. Its closure {z : 0 (cid:54) Rez (cid:54) 1} is denoted by S. We say that f (z) is an S (cid:48)(Rn)-analytic function in S if the following properties are satisfied: (i) for every fixed z ∈ S, f (z) ∈ S (cid:48)(Rn); (ii) for any ϕ ∈ S (Rn) with compact support, F −1ϕF f (x, z) is a uniformly continuous and bounded function in Rn × S; (iii) for any ϕ ∈ S (Rn) with compact support, F −1ϕF f (x, z) is an analytic function in S for every fixed x ∈ Rn. We denote the set of all S (cid:48)(Rn)-analytic functions in S by A(S (cid:48)(Rn)). The idea we used here is due to Calder´on [4], Calder´on and Torchinsky [5, 6] and Triebel [18]. Definition 2.1. Let A0 and A1 be quasi-Banach spaces, and 0 < θ < 1. We define F(A0, A1) = and (A0, A1)θ = (cid:111) (cid:110) ϕ(z) ∈ A(S (cid:48)(Rn)) : ϕ((cid:96) + it) ∈ A(cid:96), (cid:96) = 0, 1,∀t ∈ R, (cid:107)ϕ(z)(cid:107)F(A0,A1) (cid:110) f ∈ S (cid:48)(Rn) : ∃ϕ(z) ∈ F(A0, A1) such that f = ϕ(θ), (cid:107)f(cid:107)(A0,A1)θ (cid:107)ϕ(z)(cid:107)F(A0,A1) (cid:107)ϕ((cid:96) + it)(cid:107)A(cid:96) (cid:52) = max (cid:96)=0,1 sup t∈R (cid:111) ; , (cid:52) = inf ϕ (2.17) (2.18) where the infimum is taken over all ϕ(z) ∈ F(A0, A1) such that ϕ(θ) = f . The following two propositions are essentially known in [18] and the references therein. 8 (cid:1) (cid:0)(A0, A1)θ,(cid:107) · (cid:107)(A0,A1)θ (cid:18) (cid:107)ϕ(it)(cid:107)1−θ Proposition 2.9. Suppose all notations have the same meaning as in Definition 2.1, then we have is a quasi-Banach space. Proposition 2.10. Suppose all notations have the same meaning as in Definition 2.1, then we have (cid:107)f(cid:107)(A0,A1)θ = inf , where the infimum is taken over all ϕ(z) ∈ F(A0, A1) such that ϕ(θ) = f . sup t∈R sup t∈R A1 A0 ϕ (cid:107)ϕ(1 + it)(cid:107)θ (2.19) (cid:19) We point out the interpolation functor referred in (2.18) is an exact interpolation functor of exponent θ. For our purpose, we will use the following multi-linear case. 0 × A(2) 0 ×···× A(m) 0 Proposition 2.11. Let T be a continuous multi-linear operator from A(1) to B0 and from A(1) 1 × ··· × A(m) 1 × A(2) to B1, satisfying 1 (cid:13)(cid:13)T(cid:0)f (1), f (2),··· , f (m)(cid:1)(cid:13)(cid:13)B0 (cid:13)(cid:13)T(cid:0)f (1), f (2),··· , f (m)(cid:1)(cid:13)(cid:13)B1 (cid:1) 0 , A(1) 1 Then T is continuous from(cid:0)A(1) j=1 (cid:54) C0 m(cid:89) m(cid:89) θ×(cid:0)A(2) (cid:54) C1 j=1 0 ; (cid:13)(cid:13)f (j)(cid:13)(cid:13)A(j) (cid:13)(cid:13)f (1)(cid:13)(cid:13)A(1) (cid:1) θ×···×(cid:0)A(m) 0 , 1 0 , A(2) 1 0 Cθ 1 , provided 0 (cid:54) θ (cid:54) 1. norm at most C1−θ Proof. From Proposition 2.10, we know there exist m sequences {ϕ(j) satisfying f (j) ∈ A(j) 0 ∩ A(j) 1 . (cid:1) , A(m) 1 θ to (B0, B1)θ with k (z)}k∈N, j = 1,··· , m lim k→∞ sup t z−1 0 C m A0 (cid:107)ϕ(j) k (it)(cid:107)1−θ − z 1 T (ϕ(j) k (z)). sup t (cid:107)ϕ(j) k (1 + it)(cid:107)θ A1 = (cid:107)f (j)(cid:107)(A0,A1)θ . (2.20) It is easy to see that ψ(j) k (z) ∈ F (B0, B1) with (cid:107)ψ(j) k (1 + it)(cid:107)θ B1 (cid:107)ϕ(j) k (1 + it)(cid:107)θ B1 sup t sup t (cid:19) (cid:19) (2.21) (cid:3) − 1 (cid:96) m (cid:107)T ϕ(j) (cid:54) (cid:107)ϕ(j) k ((cid:96) + it)(cid:107)A(cid:96), (cid:96) = 0, 1. m We put ψ(j) k (z) = C 0 C−θ k (θ) = Cθ−1 ψ(j) k ((cid:96) + it)(cid:107)B(cid:96) (cid:107)ψ(j) 1 T f (j), and (cid:54) C Thus, combining Proposition 2.10, we have (cid:13)(cid:13)T(cid:0)f (1),··· , f (m)(cid:1)(cid:13)(cid:13)(B0,B1)θ = C1−θ 0 Cθ 1 k (θ)(cid:107)(B0,B1)θ (cid:107)ψ(j) k ((cid:96) + it)(cid:107)B(cid:96) m(cid:89) m(cid:89) m(cid:89) (cid:18) (cid:18) sup j=1 j=1 t sup j=1 t (cid:54) C1−θ 0 Cθ 1 (cid:54) C1−θ 0 Cθ 1 (cid:107)ψ(j) k (it)(cid:107)1−θ B0 (cid:107)ϕ(j) k (it)(cid:107)1−θ A0 The conclusion follows from (2.21),(2.20). 9 Theorem 2.2. Suppose 0 < θ < 1 and s = (1 − θ)s0 + θs1, then we have 1 p = 1 − θ p0 + θ p1 , 1 q = 1 − θ q0 + θ q1 , (cid:0)M s0,α (cid:1) p0,q0, M s1,α p1,q1 θ = M s,α p,q . Sketch of Proof. For z ∈ S, we write s(z) = (1 − z)s0 + zs1, For any f ∈ M s,α p,q , we set ϕ(x, z) = (cid:88) k∈Zn (cid:104)k(cid:105) 1 1−α 1 = p(z) (cid:2) sq q(z)−s(z)(cid:3) 1 − z p0 + z p1 , 1 q(z) = 1 − z q0 + z q1 . q(z)− p p(z) (cid:107)(cid:3)α k f(cid:107) q p ((cid:3)α k f ) p p(z) (x). (2.22) (2.23) Obviously, ϕ(z) ∈ A(S (cid:48)(Rn)) and ϕ(θ) = f . Direct calculation shows (cid:46) (cid:107)f(cid:107)M s,α p,q , (cid:96) = 0, 1. p,q ⊂(cid:0)M s0,α Conversely, for any f ∈(cid:0)M s0,α (cid:18) 1 (cid:90) (cid:107)ϕ((cid:96) + it)(cid:107)M p0,q0, M s1,α p1,q1 s(cid:96),α p(cid:96),q(cid:96) θ. p0,q0, M s1,α p1,q1 This proves that M s,α θ ∈ (0, 1), we can find two positive functions µ0(θ, t) and µ1(θ, t) in (0, 1) × R satisfying θ, if ϕ ∈ A(S (cid:48)(Rn)) such that ϕ(θ) = f , for some (cid:19) 1−θ r (cid:18) 1 (cid:19) θ (cid:90) (cid:1) (cid:1) r (cid:107)(cid:3)α k ϕ(it)(cid:107)r p0µ0(θ, t)dt (cid:107)(cid:3)α k ϕ(1 + it)(cid:107)r p1µ1(θ, t)dt , θ R R µ1(θ, t)dt = 1. Taking the (cid:96)q s,α norm of both sides leads to with 1 1−θ (cid:107)(cid:3)α k f(cid:107)p (cid:54) (cid:82) R R µ0(θ, t)dt = 1 θ 1 − θ (cid:82) k f}k∈Zn(cid:107)(cid:96)q (cid:107){(cid:3)α (cid:54) s,α (cid:104)k(cid:105) s0r 1−α(cid:107)(cid:3)α k ϕ(it)(cid:107)r p0µ0(θ, t)dt (cid:104)k(cid:105) s1r 1−α(cid:107)(cid:3)α k ϕ(1 + it)(cid:107)r p1µ1(θ, t)dt (cid:90) (cid:90) (cid:13)(cid:13)(cid:13)(cid:13) 1 (cid:13)(cid:13)(cid:13)(cid:13) 1 1 − θ × θ R R (cid:13)(cid:13)(cid:13)(cid:13) 1−θ q0 r r (cid:96) (cid:13)(cid:13)(cid:13)(cid:13) θ r (cid:96) q1 r . (2.24) (cid:3) Then, Minkowski's inequality implies that This proves(cid:0)M s0,α (cid:107)f(cid:107)M s,α (cid:1) p,q (cid:46) sup θ ⊂ M s,α p,q . t p0,q0, M s1,α p1,q1 (cid:107)ϕ(it)(cid:107)M s0,α p0,q0 (cid:107)ϕ(1 + it)(cid:107)M s1,α p1,q1 . sup t The following is a natural consequence of Proposition 2.11 and Theorem 2.2, and is frequently used later on. Corrolary 2.1. Suppose T is a continuous multi-linear mapping from M s(1) 0 ,α 0 ,q(1) p(1) p0,q0 with norm M0, and is also continuous, multi-linear from M s(1) M s(m) 1 ,α 0 p(1) 1 ,q(1) p(m) 0 M s(m) p1,q1 with norm M1. Then T is continuous and multi-linear from M s(1),α 1 p(m) 1 ··· × M s(m),α ,α ,q(m) 0 ,α ,q(m) p(m),q(m) to M s,α × ··· × ×···× p(1),q(1) × 1 , provided 0 (cid:54) θ (cid:54) 1, and to M s1,α to M s0,α 0 M θ 0 1 1 s(j) = (1 − θ)s(j) p,q with norm at most M 1−θ θ p(j) 1 1 − θ p(j) 0 1 p(j) = + 0 + θs(j) 1 , , 1 q(j) = 1 − θ q(j) 0 + θ q(j) 1 , j = 1··· m. 10 3 Scaling property For Besov space, it is well known that (cid:107)fλ(cid:107)Bs (cid:46) λ − n p (1 ∨ λs)(cid:107)f(cid:107)Bs p,q . p,q (3.1) For modulation spaces with s = 0 and 1 ≤ p, q ≤ ∞, the sharp dilation property was obtained in [15] and they showed (cid:107)fλ(cid:107)M 0 (cid:107)fλ(cid:107)M 0 (3.3) In this section, we study the scaling property of α-modulation spaces. For 0 < p, q (cid:54) ∞ , (cid:107)f(cid:107)M 0 λ < 1. λ > 1; q − 1 (3.2) − n p,q p,q , p q and (α1, α2) ∈ [0, 1] × [0, 1], we denote q p + 1 (cid:1)∨n(cid:0) 1 q −1(cid:1) (cid:1)∨n(cid:0)1− 1 (cid:1)(cid:105) (cid:107)f(cid:107)M 0 (cid:1)(cid:105) ∨(cid:104) p− 1 q − 1 p p,q p,q p λ − n p− 1 (cid:46) λ (cid:46) λ p λ0∨n(cid:0) 1 0∨n(cid:0) 1 −(cid:104) R(p, q; α1, α2) = 0 ∨(cid:104) n(α1 − α2)(cid:0) 1 (cid:110)(cid:0) 1 (cid:1) ∈ R2 (cid:110)(cid:0) 1 (cid:1) ∈ R2 S1 = q which will be frequently used in this and the next sections. Then, we divide R2 sub-domains in two ways (see Fig. 1). One way is, R2 (cid:62) 1 q − 1(cid:1)(cid:105) p + 1 n(α1 − α2)(cid:0) 1 (cid:111) + = S1 ∪ S2 ∪ S3 with (cid:54) 1 p , 1 (cid:62) 1, 1 (cid:111) p > 1 p 2 2 ; ; p + 1 q , (3.4) + into 3 Another way is, R2 + = T1 ∪ T2 ∪ T3 with If α1 (cid:62) α2, then If α1 < α2, then (cid:111) ; (cid:54) 1 2 (cid:111) ; S2 = S3 = R2 T1 = T2 = T3 = R2 + : 1 q p , 1 p , 1 + : 1 +\{S1 ∪ S2}, q q q + : 1 p p , 1 p , 1 + : 1 +\{T1 ∪ T2}. (cid:110)(cid:0) 1 (cid:1) ∈ R2 (cid:110)(cid:0) 1 (cid:1) ∈ R2 n(α1 − α2)(cid:0) 1 n(α1 − α2)(cid:0) 1 0, n(α1 − α2)(cid:0) 1 n(α1 − α2)(cid:0) 1 (cid:40) 0, (cid:62) 1 2 q , 1 p > 1 (cid:54) 1, 1 p p + 1 q (cid:1), q − 1(cid:1), q − 1 p + 1 p q − 1(cid:1), (cid:1), p + 1 q − 1 p R(p, q; α1, α2) = R(p, q; α1, α2) = 11 (cid:0) 1 (cid:0) 1 (cid:0) 1 (cid:0) 1 (cid:0) 1 (cid:0) 1 q p , 1 p , 1 p , 1 q q q p , 1 p , 1 p , 1 q q (cid:1) ∈ S1; (cid:1) ∈ S2; (cid:1) ∈ S3. (cid:1) ∈ T3; (cid:1) ∈ T2; (cid:1) ∈ T1. (3.5) (3.6) Before describing the dilation property of the α-modulation spaces, we introduce some critical powers. Let us write sp = n(1/(1 ∧ p) − 1) and sc = R(p, q; 1, α), −R(p, q; α, 1), λ > 1, λ (cid:54) 1. (3.7) Figure 1: Distribution of sc. The left-hand side figure is for λ > 1, the right-hand side figure is for λ (cid:54) 1. Theorem 3.1. Let 0 (cid:54) α < 1, λ > 0 and s + sc (cid:54)= 0. Then p(cid:2)(1 ∨ λ)sp ∨ λs+sc(cid:3)(cid:107)f(cid:107)M s,α p,q (cid:107)fλ(cid:107)M s,α p,q − n (cid:46) λ holds for all f ∈ M s,α p,q . Conversely, if . (3.8) p,q , then F (λ) (cid:38) (1 ∨ λ)sp ∨ λs+sc. k,1/λ the pseudo-differential operator with symbol (ηα k )λ. (3.9) (3.10a) (cid:107)fλ(cid:107)M s,α p,q (cid:46) λ − n p F (λ)(cid:107)f(cid:107)M s,α p,q holds for some F : (0,∞) → (0,∞) and all f ∈ M s,α Proof. (Sufficiency) We denote by (cid:3)α For every l ∈ Zn and λ > 0, we introduce Λ(l, λ) = {k ∈ Zn : (cid:3)α l f (cid:54)= 0}. (cid:3)α For any k ∈ Λ(l, λ), it follows from (1.1b) that k, l and λ satisfy 1−α (kj + C); 1−α (kj − C) 1−α (lj − C) < (cid:104)k(cid:105) α 1−α (lj + C) > (cid:104)k(cid:105) α λ(cid:104)l(cid:105) α λ(cid:104)l(cid:105) α k,1/λ (3.10b) with j = 1, 2,··· , n. In view of (3.10), one sees that k ∈ Λ(l, λ) is equivalent to l ∈ Λ(k, 1/λ). Moreover, if (3.10) holds, then (cid:104)l(cid:105) (cid:46) 1 ∨ λ−(1−α) if and only if (cid:104)k(cid:105) (cid:46) 1 ∨ λ1−α. (3.11) If (cid:104)l(cid:105) (cid:29) 1 ∨ λ−(1−α), without loss of generality, we may assume l belongs to some Rn lj > 0, from (3.10a); whereas when lj < 0, from (3.10b)×(−1), we see (cid:104)k(cid:105) 1 Conversely, for k ∈ Λ(l, λ) ∩ Rn j , when 1−α (cid:38) λ(cid:104)l(cid:105) 1 1−α . 1−α . Thus we have j , also from (3.10), we have (cid:104)k(cid:105) 1 1−α (cid:46) λ(cid:104)l(cid:105) 1 (cid:104)k(cid:105) ∼ λ1−α(cid:104)l(cid:105). 12 (3.12) When q = 1, from Lemma 2.1, we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:88) k∈Γ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,1 (cid:3)α k fλ (cid:104)k(cid:105) s 1−α(cid:107)(cid:3)α k fλ(cid:107)p k∈Γ (cid:88) (cid:88) p(cid:88) p(cid:88) k∈Γ − n k∈Γ − n = = = λ (cid:54) λ (cid:104)k(cid:105) s 1−α(cid:107)((cid:3)α k,1/λf )(λ·)(cid:107)p k,1/λf(cid:107)p 1−α(cid:107)(cid:3)α 1−α (cid:88) (cid:104)k(cid:105) s (cid:104)k(cid:105) s (cid:107)(cid:3)α k,1/λ l f(cid:107)p. (cid:3)α (3.13) (3.14) Figure 2: 9 regions for the proof of Theorem 3.1 Since the volumes of supp(ηα we see that k )λ and suppηα l are O(λ−n(cid:104)k(cid:105) nα 1−α ) and O((cid:104)l(cid:105) nα 1−α ), respectively, #Λ(l, λ) ∼ 1 ∨ λn(1−α). k∈Γ p , 1 q l∈Λ(k,1/λ) in Proposition 2.3 to remove (cid:3)α l ∈ Λ(k, 1/λ), there is 1 (cid:46) (cid:104)l(cid:105) 1 Case 1. λ (cid:54)(cid:54)(cid:54) 1,(cid:0) 1 (cid:1) ∈ I∪ II. For p = 1,∞, we apply the same technique as that appeared (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k,1/λ in (3.14). When k (cid:46) 1, from (3.11) we see that for p (cid:88) λ , which leads to (1 ∨ λs)(cid:104)l(cid:105) s p (cid:88) l f(cid:107)p (cid:46) λ 1−α(cid:107)(cid:3)α (cid:88) 1−α (cid:46) 1 l f(cid:107)p (cid:107)(cid:3)α (cid:46) λ (3.15) (cid:3)α k fλ k(cid:46)1 l∈Zn − n − n p,1 (cid:46) λ (cid:104)l(cid:105)(cid:46)λ−(1−α) − n p (1 ∨ λs)(cid:107)f(cid:107)M s,α . p,1 13 By Plancherel's identity, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,2 (cid:3)α k fλ (cid:46) λ− n 2 (cid:104)k(cid:105) 2s 1−α (cid:104)l(cid:105)(cid:46)λ−(1−α) When k (cid:29) 1, from (3.12)-(3.14), we see that − n k(cid:46)1 (cid:88)  (cid:88) p +s (cid:88) p +s(cid:88) (cid:88) (cid:88) (cid:104)l(cid:105) 2s k(cid:29)1 k(cid:29)1 − n l (cid:46) λ− n 2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,1 (cid:46) λ (cid:46) λ (cid:54) λ− n 2 +s (cid:46) λ− n 2 +s (cid:88) 1−α (cid:88) l∈Λ(k,1/λ) l k∈Λ(l,λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:3)α k,1/λ l∈Λ(k,1/λ)  1 2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2 2 (cid:3)α l f  1 2 (cid:107)(cid:3)α l f(cid:107)2 2 (cid:46) λ− n 2 (1 ∨ λs)(cid:107)f(cid:107)M s,α 2,2 (3.16) . (cid:104)l(cid:105) s 1−α(cid:107)(cid:3)α l f(cid:107)p (cid:88) 1−α (cid:88) l∈Λ(k,1/λ) (cid:104)l(cid:105) s k∈Λ(l,λ) (cid:107)(cid:3)α l f(cid:107)p (cid:46) λ − n p +s(cid:107)f(cid:107)M s,α p,1 . (3.17) 2  1  1 2 2 (cid:104)l(cid:105) 2s 1−α(cid:107)(cid:3)α k,1/λ l f(cid:107)2 (cid:3)α (cid:107)(cid:3)α k,1/λ l f(cid:107)2 (cid:3)α 2 (cid:46) λ− n 2 +s(cid:107)f(cid:107)M s,α 2,2 (3.18) . (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:29)1 (cid:3)α k fλ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:29)1 (cid:3)α k fλ In view of Plancherel's formula, q p,q = . p,q 1 p 2 , θ 2 p , 1 q (3.19) (cid:46) λ (cid:18) to the 1 Combining (3.15)-(3.18), we use complex interpolation to get − n p (1 ∨ λs)(cid:107)f(cid:107)M s,α p , 1 (cid:107)fλ(cid:107)M s,α = 1 − When k (cid:46) λ1−α, from (3.11) and (3.14), we have (cid:1), one can draw the parallel line (cid:1) 2 , 1(cid:1) and the line segment connecting (0, 0),(cid:0) 1 q -axis. We assume there exists some (θ, η) ∈ [0, 1] × [0, 1], such that the parallel line 2 , 1 Case 2. λ > 1,(cid:0) 1 (cid:1) ∈ I ∪ III∗. Through the point(cid:0) 1 cuts the line segment connecting (0, 1),(cid:0) 1 (cid:1), respectively. Assume that at(cid:0) θ 2 , 1(cid:1) and(cid:0) θ (cid:19) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 ∨ λs+n(1−α)(cid:17)(cid:107)f(cid:107)M s,α∞,1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  1  1 2 (cid:88) By the Schwartz inequality and the Plancherel identity, (3.20) (3.21) 1−α (cid:46)(cid:16) k∈Λ(l,λ) (cid:104)k(cid:105) s (cid:88)  (cid:88) 2 (1−α)(cid:17)(cid:107)f(cid:107)M s,α k∈Λ(l,λ) l(cid:46)1 1 ∨ λs+ n (cid:88) (cid:16) (cid:54)(cid:88) 1 − θ 2 l f(cid:107)2 (cid:3)α (cid:88) (cid:88) (cid:46) λ− n 2 (cid:54) λ− n (cid:104)k(cid:105) 2s 1−α (cid:107)(cid:3)α l f(cid:107)∞ k(cid:46)λ1−α k(cid:46)λ1−α k∈Λ(l,λ) (cid:107)(cid:3)α l(cid:46)1 (cid:3)α (cid:3)α k fλ k fλ k,1/λ θ 2 , 1 q η. 2,1 2 2 . 2 2 . 2,1 14 From (3.11), we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,∞ (cid:3)α k fλ (cid:104)k(cid:105) s 1−α(cid:107)(cid:3)α (cid:46) sup k(cid:46)λ1−α (cid:46) (1 ∨ λs) sup k(cid:46)λ1−α (cid:107)(cid:3)α (cid:46) (1 ∨ λs) sup l(cid:46)1 k fλ(cid:107)∞ (cid:88) l f(cid:107)∞ (cid:3)α l∈Λ(k,1/λ) l f(cid:107)∞ (cid:46) (1 ∨ λs)(cid:107)f(cid:107)M s,α∞,∞. (cid:107)(cid:3)α k,1/λ (cid:3)α k fλ (cid:46) (cid:104)k(cid:105) 2s 1−α(cid:107)(cid:3)α k fλ(cid:107)2 2 In view of Plancherel's equality, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,2  (cid:88) (cid:88) k(cid:46)λ1−α l(cid:46)1 (cid:46) λ− n 2  1 2  1 2 l f(cid:107)2 (cid:3)α 2 (cid:107)(cid:3)α k,1/λ  1 2 (cid:46) λ− n 2 (1 ∨ λs) l f(cid:107)2 2 (cid:46) λ− n 2 (1 ∨ λs)(cid:107)f(cid:107)M s,α 2,2 (3.22) (3.23) . k∈Λ(l,λ) (1 ∨ λs)2 (cid:88) (cid:88) 1−α (cid:88) (cid:107)(cid:3)α k∈Λ(l,λ) 1−α (cid:88) (cid:104)l(cid:105) s k∈Λ(l,λ) l(cid:46)1 When k (cid:29) λ1−α, from (3.12)-(3.14), we have (cid:88) (cid:3)α k fλ k(cid:29)λ1−α (cid:54) λs (cid:88) l(cid:29)1 (cid:104)l(cid:105) s (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:29)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,1 By Jensen's inequality, (cid:3)α k fλ (cid:46) λ− n 2 +s (cid:88) 2 +s (cid:88) l(cid:29)1 n (cid:46) λ l(cid:29)1 2 +s+ n (cid:46) λ− n From (3.12),(3.13), we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:3)α k fλ k(cid:29)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,∞ (cid:46) sup k (cid:104)k(cid:105) s 1−α(cid:107)(cid:3)α k,1/λf(cid:107)∞ (cid:88) (cid:46) λs sup k l∈Λ(k,1/λ) 15 (cid:107)(cid:3)α l f(cid:107)∞ (cid:46) λs+n(1−α)(cid:107)f(cid:107)M s,α∞,1 . (3.24) (cid:107)(cid:3)α l f(cid:107)2 (cid:3)α k,1/λ  (cid:88) k∈Λ(l,λ) (cid:104)l(cid:105) s 1−α [#Λ(l, λ)] 1 2 (cid:107)(cid:3)α k,1/λ l f(cid:107)2 (cid:3)α 2 2 (1−α)(cid:107)f(cid:107)M s,α . 2,1  1 2 (3.25) (cid:104)l(cid:105) s 1−α(cid:107)(cid:3)α k,1/λ l f(cid:107)∞ (cid:46) λs(cid:107)f(cid:107)M s,α∞,∞. (cid:3)α (3.26) Similar to (3.18), one has that(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Since n(1 − α)(cid:0)1 − θ k fλ (cid:3)α k(cid:29)λ1−α (cid:46) λ− n (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:1) = (1 − θ)n(1 − α) + θ n (cid:18) 1 ∨ λs+n(1−α)(cid:0)1− θ (cid:46) λ− θ 2 n (cid:107)fλ(cid:107)M s,α 2,2 2 complex interpolation yields 2 (cid:1)(cid:19) 2 +s(cid:107)f(cid:107)M s,α 2,2 2 θ ,1 . (3.27) 2 (1 − α), combining (3.20),(3.24),(3.21),(3.25), (cid:107)f(cid:107)M s,α 2 θ ,1 . (3.28) Combining (3.22),(3.26),(3.23),(3.27), complex interpolation yields . 2 θ 2 θ , 2 θ , 2 θ (3.29) (cid:46) λ− θ (cid:107)fλ(cid:107)M s,α (cid:107)fλ(cid:107)M s,α Interpolating (3.28) and (3.29), we have 2 n(1 ∨ λs)(cid:107)f(cid:107)M s,α (cid:1)(cid:19) (cid:18) 1 ∨ λs+n(1−α)(cid:0) 1 (cid:1) one can make the parallel line (cid:1) ∈ II ∪ III. Through the point(cid:0) 1 2 , 1(cid:1) and the line segment connecting (1, 0),(cid:0) 1 (cid:1) q -axis. We assume there exists some (θ, η) ∈ [0, 1] × [0, 1], such that the parallel line (cid:1), respectively. We can assume that 2 , 1 (cid:19) Case 3. λ > 1,(cid:0) 1 cuts the line segment connecting (1, 1),(cid:0) 1 at(cid:0)1 − θ 2 , 1(cid:1) and(cid:0)1 − θ (cid:107)f(cid:107)M s,α to the 1 (cid:18) (cid:46) λ (3.30) p , 1 q − 1 p , 1 2 , θ − n p,q p,q q 2 2 . q p p 1 p = 1 − θ 2 , = 1 − 1 q 1 − θ 2 η. When k (cid:46) λ1−α, similarly to (3.20) and (3.22), we have When k (cid:29) λ1−α, similarly to (3.24) and (3.26), we have 1 ∨ λs+n(1−α)(cid:17)(cid:107)f(cid:107)M s,α 1,1 (cid:46) λ−n (1 ∨ λs)(cid:107)f(cid:107)M s,α 1,∞. ; (3.31) (3.32) (cid:46) λ−n+s+n(1−α)(cid:107)f(cid:107)M s,α 1,1 (cid:46) λ−n+s(cid:107)f(cid:107)M s,α 1,∞. ; (3.33) (3.34) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:3)α k fλ k(cid:46)λ1−α (cid:88) k(cid:46)λ1−α (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k(cid:29)λ1−α (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:29)λ1−α 1,1 k fλ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:46) λ−n(cid:16) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:3)α k fλ k fλ 1,∞ 1,1 1,∞ (cid:3)α (cid:3)α 16 Combining (3.31),(3.33),(3.21),(3.25), complex interpolation yields (cid:107)fλ(cid:107)M s,α 2 2−θ ,1 (cid:46) λ − n 2 (cid:107)f(cid:107)M s,α 2 2−θ . ,1 (3.35) Combining (3.32),(3.34),(3.23),(3.27), complex interpolation yields (1 ∨ λs)(cid:107)f(cid:107)M s,α (cid:107)fλ(cid:107)M s,α (cid:46) λ − n 2 2 2−θ , 2 θ . (3.36) (cid:1)(cid:19) (cid:0)1− θ 2 2 (cid:1)(cid:18) 1 ∨ λs+n(1−α)(cid:0)1− θ (cid:1) (cid:0)1− θ (cid:18) 1 ∨ λs+n(1−α)(cid:0) 1 p + 1 2 2 2−θ , 2 θ q −1(cid:1)(cid:19) (cid:1) ∈ I∗ ∪ II∗, we have(cid:0) 1 (cid:107)f(cid:107)M s,α p,q . Interpolating (3.35) and (3.36), we have (cid:107)fλ(cid:107)M s,α − n (cid:46) λ Case 4. λ (cid:54)(cid:54)(cid:54) 1,(cid:0) 1 ∈ I∗ ∪ II∗\{(1, 0)}. We observe that, for any(cid:0) 1 (cid:1) ∈ {I∗ ∪ II∗ ∪ III ∪ III∗}\{(0, 1] × [0, 1]}; or λ > 1,(cid:0) 1 (cid:1) q∗(cid:1) ∈ I ∪ II. By p∗ , 1 (3.37) p , 1 p , 1 p , 1 p,q q q q p , from the previous several cases, we know that By the principle of duality, it follows from (3.38) and (3.39) that duality, If we denote (cid:107)fλ(cid:107)M s,α p,q λ λ λ p,q p,q p,q 1 p , 1 q −s,α p∗,q∗ −s,α p∗,q∗ . −s,α p∗,q∗ . (cid:105) (cid:54) 1 (cid:13)(cid:13)M (cid:13)(cid:13)g 1 (cid:107)fλ(cid:107)M s,α (cid:13)(cid:13)g 1 (cid:13)(cid:13)M λn(cid:104)f, g 1 λn(cid:107)f(cid:107)M s,α (cid:104)fλ, g(cid:105) = (cid:46) F (s, λ; p, q)(cid:107)f(cid:107)M s,α (cid:46) F(cid:0)−s, λ−1; p∗, q∗(cid:1)(cid:107)g(cid:107)M (cid:46) λ−nF(cid:0)−s, λ−1; p∗, q∗(cid:1)(cid:107)f(cid:107)M s,α (cid:18) 1 ∨ λs−n(1−α)(cid:0) 1 (cid:18) 1 ∨ λs−n(1−α)(cid:0)1− 1 (cid:1) ∈ {I∗ ∪ III}\{1} × [0, 1], from Case 2, (3.40) gives (cid:107)f(cid:107)M s,α (cid:1) ∈ {II∗ ∪ III∗}\{(0, 1)}, from Case 3, (3.40) gives (cid:1)(cid:19) (cid:1) ∈ I∗ ∪ II∗\{(1, 0)}, from Case 1, (3.40) gives (cid:1) ∈ {(cid:98)I ∪ (cid:98)II}\(1,∞) × (1,∞). Since (cid:96)q ⊂ (cid:96)1, we know − n p (1 ∨ λs)(cid:107)f(cid:107)M s,α (cid:107)fλ(cid:107)M s,α (cid:107)fλ(cid:107)M s,α (cid:107)fλ(cid:107)M s,α (cid:107)f(cid:107)M s,α (cid:1)(cid:19) (cid:46) λ (cid:46) λ (cid:46) λ p− 1 q − 1 p , 1 p , 1 q − n − n p,q p,q p,q p,q p,q . p,q p,q . p q . . p p q For λ (cid:54) 1 with(cid:0) 1 For λ (cid:54) 1 with(cid:0) 1 For λ > 1 with(cid:0) 1 Case 5. (cid:0) 1 p , 1 q (3.38) (3.39) (3.40) (3.41) (3.42) (3.43) (3.44)  1 q . (cid:107)(cid:3)α k fλ(cid:107)p (cid:46) λ − n p (cid:88) l∈Λ(k,1/λ) (cid:107)(cid:3)α k,1/λ l f(cid:107)p (cid:46) λ (cid:3)α − n p 17  (cid:88) l∈Λ(k,1/λ) (cid:107)(cid:3)α k,1/λ l f(cid:107)q (cid:3)α p From (3.44),(3.11), when λ (cid:54) 1 and k (cid:46) 1, we conclude that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,q (cid:3)α k fλ 1−α(cid:107)(cid:3)α k fλ(cid:107)q∞ 1−α (cid:88) l∈Λ(k,1/λ)  1 q (cid:107)(cid:3)α k,1/λ  1 q  1 q l f(cid:107)q∞ (cid:3)α (3.45) and when λ > 1 and k (cid:46) λ1−α, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)λ1−α (cid:3)α k fλ (cid:46) (cid:46) (cid:46) (cid:46) k(cid:46)1 k(cid:46)1 (cid:104)k(cid:105) sq (cid:104)k(cid:105) sq (cid:104)l(cid:105)(cid:46)λ−(1−α) (cid:88) (cid:88)  (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,q  (cid:88)  (cid:88) (cid:88) (cid:46)(cid:16)  (cid:88) (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α∞,q (cid:46) λs (cid:46) (cid:46) (cid:46) k(cid:46)λ1−α (cid:107)(cid:3)α l f(cid:107)q∞ (cid:46) (1 ∨ λs)(cid:107)f(cid:107)M s,α∞,q ;  1 q (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k fλ(cid:107)q∞  1 q (cid:104)k(cid:105) sq k(cid:46)λ1−α (cid:107)(cid:3)α l f(cid:107)q∞ l(cid:46)1 1 ∨ λs+n(1−α) 1 q (cid:107)(cid:3)α l f(cid:107)q∞  1 q (cid:104)k(cid:105) sq 1−α . k(cid:46)λ1−α l∈Λ(k,1/λ) 1−α (cid:88) (cid:88) (cid:17)(cid:107)f(cid:107)M s,α∞,q 1−α (cid:88) 1−α (cid:88) k∈Λ(l,λ) q (cid:107)f(cid:107)M s,α∞,q (cid:104)k(cid:105) sq . (cid:104)l(cid:105) sq l(cid:29)1 (cid:46) λs(1 ∨ λ)n(1−α) 1 When k (cid:29) 1 ∨ λ1−α, from (3.44), (3.12), (3.13), we have (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:3)α k fλ k(cid:29)1∨λ1−α k(cid:29)1∨λ1−α l∈Λ(k,1/λ)  1 q (cid:107)(cid:3)α l f(cid:107)q∞  1 q (cid:107)(cid:3)α l f(cid:107)∞ Therefore, combining (3.45)-(3.47), we get (cid:107)fλ(cid:107)M s,α∞,q (cid:46) 1 ∨ λs ∨ λs+n(1−α) 1 q (cid:107)f(cid:107)M s,α∞,q . The same for M s,α 1,q . Corresponding to (3.48), we get (cid:46) λ−n(cid:16) (cid:107)fλ(cid:107)M s,α 1,q (cid:17)(cid:107)f(cid:107)M s,α 1,q . 1 ∨ λs ∨ λs+n(1−α) 1 q 18 (3.46) (3.47) (3.48) (3.49) Whereas when p = 2, λ (cid:54) 1 and k (cid:46) 1, from (3.44),(3.11), we have When λ > 1,k (cid:46) λ1−α, by (3.44), (3.11) and Holder's inequality, we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,q (cid:3)α k fλ (cid:46) λ− n 2 (cid:46) λ− n 2 (cid:88) k(cid:46)λ1−α (cid:3)α k fλ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,q (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  1 q (cid:107)(cid:3)α k,1/λ l f(cid:107)q (cid:3)λ 2 (cid:46) λ− n 2 (1 ∨ λs)(cid:107)f(cid:107)M s,α 2,q (3.50) . l f(cid:107)q (cid:3)α q  1  q 2 (cid:88) 2 k(cid:46)λ1−α  2−q 2  1 q (cid:104)k(cid:105) sq 1−α· 2 2−q (cid:107)f(cid:107)M s,α 2,q . (3.51) (cid:46) λ− n 2 (cid:46) λ− n 2 (cid:46) λ− n 2 (cid:46) λ− n (cid:46) λ− n l f(cid:107)2 (cid:3)α 2 q 2 2 k,1/λ k,1/λ l(cid:46)1 l(cid:46)1 q − 1 k(cid:46)1 (cid:107)(cid:3)α k∈Λ(l,λ) k∈Λ(l,λ) (cid:104)k(cid:105) sq (cid:1)(cid:19) (cid:107)(cid:3)α l f(cid:107)q l∈Λ(k,1/λ) (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α (cid:104)l(cid:105)(cid:46)λ−(1−α)  1 1−α (cid:88) (cid:88)  (cid:88) (cid:88) (cid:88) (cid:88)  (cid:88) (cid:18) 1 ∨ λs+n(1−α)(cid:0) 1 (cid:88) (cid:88) 2 +s(1 ∨ λ)n(1−α)(cid:0) 1 (cid:18) 1 ∨ λs ∨ λs+n(1−α)(cid:0) 1 (cid:18) 1 ∨ λs ∨ λs+n(1−α)(cid:0) 1 1 ∨ λs ∨ λs+n(1−α)(cid:0) 1 1−α (cid:88) k∈Λ(l,λ) (cid:104)l(cid:105) sq (cid:18) l(cid:29)1 l(cid:29)1 q − 1 (cid:1) − n − n 2 +s 2 +s p 2 19  1 q (cid:107)(cid:3)α l k,1/λf(cid:107)q (cid:3)α 2  (cid:88) k∈Λ(l,λ) . 2,q (cid:107)f(cid:107)M s,α (cid:1)(cid:19) q − 1 p (cid:1)(cid:19) q −1(cid:1)(cid:19) Therefore, combining (3.50)-(3.52), we get q p , 1 For(cid:0) 1 while for(cid:0) 1 p , 1 q (cid:107)fλ(cid:107)M s,α (cid:46) λ− n 2 (cid:1) ∈(cid:98)I, complex interpolation between (3.48) and (3.53) yields 2,q 2 (cid:107)f(cid:107)M s,α 2,q . (3.53) (cid:107)fλ(cid:107)M s,α (cid:46) λ p,q (cid:1) ∈ (cid:98)II\(1,∞) × (1,∞), complex interpolation between (3.49) and (3.53) yields (3.54) p,q ; p q − 1 (cid:107)f(cid:107)M s,α (cid:107)fλ(cid:107)M s,α p,q (cid:46) λ p + 1 (cid:107)f(cid:107)M s,α p,q . (3.55) When k (cid:29) 1 ∨ λ1−α, in view of (3.44),(3.12),(3.13) and Holder's inequality, we have (cid:88) k(cid:29)1∨λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α 2,q (cid:3)α k fλ (cid:46) λ− n (cid:104)l(cid:105) sq 1−α [#Λ(l, λ)]1− q 2 (cid:107)(cid:3)α l k,1/λf(cid:107)2 (cid:3)α 2  q 2  1 q (3.52) Case 6. (cid:0) 1 p , 1 q (cid:1) ∈ (cid:99)III. Since lp ⊂ l1, we know (cid:107)(cid:3)α k fλ(cid:107)p = λ − n p (cid:107)(cid:3)α k,1/λf(cid:107)p (cid:54) λ − n p  (cid:88) l∈Λ(k,1/λ)  1 p . (cid:107)(cid:3)α k,1/λ l f(cid:107)p (cid:3)α p When λ (cid:54) 1 and k (cid:46) 1, from (1.1b),(3.12), we see that k )λ)(x − ·)(cid:3)α diam suppF [(F −1(ηα l f (·)] (cid:46) 1/λ, By Proposition 2.1, imitating the processes as in (??)-(??), we get From (3.56), (3.57), the embedding (cid:96)1 ⊂ (cid:96) q p , and Holder's inequality, we have (3.56) (3.57) (cid:107)(cid:3)α k,1/λ (cid:3)α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,q (cid:3)α k fλ (cid:107)(cid:3)α l f(cid:107)p (cid:46) (cid:107)(cid:3)α l f(cid:107)p. (cid:17)n(cid:0) 1 p−1(cid:1) 1−α (cid:46) λ − n p−1(cid:1)(cid:16) l f(cid:107)p (cid:46) (1/λ)n(cid:0) 1 λ(cid:14)(cid:104)k(cid:105) α (cid:88)  (cid:88) (cid:88) 1−α (cid:88)  (cid:88) (cid:88) l∈Λ(k,1/λ) (cid:104)k(cid:105) sq 1−α (cid:104)k(cid:105) sp (cid:104)k(cid:105) s k(cid:46)1 k(cid:46)1 l∈Λ(k,1/λ) − n − n (cid:46) λ (cid:46) λ p p p (cid:104)l(cid:105)(cid:46)λ−(1−α) k(cid:46)1  1 q p p l f(cid:107)p (cid:3)α  q  1  q−p q (cid:88) l f(cid:107)p (cid:3)α p p k(cid:46)1 (cid:107)(cid:3)α k,1/λ (cid:107)(cid:3)α k,1/λ 1−α· pq q−p (3.58)  p q  1 p (cid:107)(cid:3)α l f(cid:107)q p (cid:46) λ − n p (1 ∨ λs)(cid:107)f(cid:107)M s,α p,q . When λ (cid:54) 1 and k (cid:29) 1, from (1.1b), (3.12), we see that diam suppF [(F −1(ηα k )λ)(x − ·)(cid:3)α l f (·)] (cid:46) (cid:104)k(cid:105) α 1−α /λ. Similarly to (3.57), we get l f(cid:107)p (cid:46)(cid:16)(cid:104)k(cid:105) α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α (cid:46) λ − n p p,q (cid:107)(cid:3)α k,1/λ (cid:3)α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:29)1 (cid:3)α k fλ From (3.56), (3.59), (3.12), (3.13), we have p−1(cid:1) (cid:17)n(cid:0) 1 1−α (cid:107)(cid:3)α l f(cid:107)p (cid:54) (cid:107)(cid:3)α l f(cid:107)p.  q p  1 q (cid:107)(cid:3)α k,1/λ l f(cid:107)p (cid:3)α p l∈Λ(k,1/λ)  1 q q p−1 (cid:88) 1−α (cid:88) (cid:104)l(cid:105) sq l∈Λ(k,1/λ) k∈Λ(l,λ) (cid:104)l(cid:105) sq 1−α(cid:107)(cid:3)α l f(cid:107)q p  1 q (cid:107)(cid:3)α l f(cid:107)q p 1−α /λ (cid:104)k(cid:105) sq 1−α p−1(cid:1)(cid:16) (cid:17)n(cid:0) 1 λ(cid:14)(cid:104)k(cid:105) α (cid:88)  (cid:88) (cid:88) p +s−n(1−α)(cid:0) 1 p +s−n(1−α)(cid:0) 1 (cid:1)(cid:88) (cid:1) [#Λ(k, 1/λ)] k(cid:29)1 k(cid:29)1 p− 1 p− 1 p +s q q l − n − n − n (cid:46) λ (cid:46) λ (cid:46) λ (cid:107)f(cid:107)M s,α p,q . 20 (3.59) (3.60) For λ > 1 and k (cid:46) λ1−α, from (1.1b), (3.12), we know Similarly to (3.57), we get (cid:107)(cid:3)α k,1/λ From (3.56), (3.61), (3.11), we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) k(cid:46)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,q (cid:3)α k fλ diam suppF [(F −1(ηα k )λ)(x − ·)(cid:3)α l f (·)] (cid:46) 1. l f(cid:107)p. 1−α (cid:3)α (cid:107)(cid:3)α (cid:17)n(cid:0) 1 p−1(cid:1) l f(cid:107)p (cid:46)(cid:16) λ(cid:14)(cid:104)k(cid:105) α  (cid:88)  (cid:88) p−1(cid:1) (cid:88) p−1(cid:1)(cid:105) (cid:0) 1 p +n(cid:0) 1 (cid:104) s 1−α− nα 1−α p−1(cid:1)(cid:18) (cid:19) p−1(cid:1)+n(1−α) 1 1 ∨ λs−nα(cid:0) 1 p +n(cid:0) 1 l∈Λ(k,1/λ) (cid:104)k(cid:105) sq 1−α k(cid:46)λ1−α k(cid:46)λ1−α (cid:107)(cid:3)α (cid:104)k(cid:105) k,1/λ p q l f(cid:107)p (cid:3)α p  q (cid:88) p q l(cid:46)1 (cid:107)f(cid:107)M s,α p,q − n − n − n (cid:46) λ (cid:46) λ (cid:46) λ  1 q When λ > 1 and k (cid:29) λ1−α, similarly to (3.59), we get l f(cid:107)p (cid:46) (cid:107)(cid:3)α (cid:3)α From (3.56), (3.63), (3.12), and the embedding (cid:96)1 ⊂ (cid:96) (cid:107)(cid:3)α k,1/λ l f(cid:107)p. q p , we have (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) (cid:3)α k fλ k(cid:29)λ1−α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)M s,α p,q (cid:107)(cid:3)α k,1/λ l f(cid:107)p (cid:3)α p l∈Λ(k,1/λ)  1 q (cid:46) λ (cid:46) λ (cid:46) λ (cid:46) λ p p +s − n − n k(cid:29)λ1−α k(cid:29)λ1−α (cid:104)k(cid:105) sq 1−α l∈Λ(k,1/λ)  (cid:88) (cid:88) 1−α (cid:88)  (cid:88)  (cid:88) (cid:88) (cid:18) 1 ∨ λs+n(1−α)(cid:0) 1 p−1(cid:1)(cid:18) 1 ∨ λs−nα(cid:0) 1 p +n(cid:0) 1 l(cid:29)1 − n p +s+n(1−α) 1 q (cid:107)f(cid:107)M s,α k∈Λ(l,λ) (cid:104)l(cid:105) sq (cid:46) λ − n − n p +s p,q . p (cid:104)l(cid:105) sq 1−α(cid:107)(cid:3)α l f(cid:107)q p  1 q (cid:107)(cid:3)α l f(cid:107)q p p q − 1 (cid:1)(cid:19) p−1(cid:1)+n(1−α) 1 (cid:107)f(cid:107)M s,α (cid:19) p,q (3.61)  q p  1 q (3.62) (3.63) (3.64) (cid:107)(cid:3)α l f(cid:107)p p .  q p  1 q We summarize the argument in this case as: if λ (cid:54) 1, (3.58) and (3.60) give (cid:107)fλ(cid:107)M s,α p,q else if λ > 1, (3.62) and (3.64) give ; (3.65) Case 7. (cid:0) 1 p , 1 q − n (cid:46) λ (cid:107)fλ(cid:107)M s,α (cid:1) ∈ (cid:98)II ∩ (1,∞) × (1,∞). It is a natural consequence of Cases 5 and 6 by (cid:107)f(cid:107)M s,α (3.66) p,q p,q . q complex interpolation. (3.8) in the case λ (cid:62) 1 follows from (3.30), (3.54), (3.37), (3.55), (3.43), (3.66). (3.8) in (cid:3) the case λ < 1 follows from (3.19), (3.54), (3.55), (3.41), (3.66), (3.42). 21 Figure 3: The left-hand side figure is for α1 (cid:62) α2; while the right-hand side figure is for α1 < α2. Remark 3.1. If s = −sc, we have the substitution for (3.8): − n p F (λ)(cid:107)f(cid:107)M s,α (cid:107)fλ(cid:107)M s,α (cid:46) λ , p,q where F (λ) = p,q (ln λ)0∨(cid:0) 1  λn(cid:0) 1 p−1(cid:1) (cid:1)0∨(cid:0) 1 (cid:0)ln 1 λ p−1(cid:1) (cid:1)∨(cid:0) 1 (cid:1) (cid:1)∨(cid:0)1− 1 q + 1 1 q , p− 1 q , q − 1 p (ln λ) p− 1 q λ > 1, p ≥ 1; λ > 1, p ≤ 1; λ (cid:54) 1. , (3.67) (3.68) 4 Embedding between α-modulation and Besov spaces As 1 ≤ p, q ≤ ∞, some sufficient conditions for the inclusions between modulation and Besov spaces were obtained by Grobner [11], then Toft [16] improved Grobner's sufficient conditions, which were proven to be necessary by Sugimoto and Tomita [15]. Their results were generalized to the cases 0 < p, q ≤ ∞ in [19, 20]. Grobner [11] also considered the inclusions between α1-modulation and α2-modulation spaces for 1 ≤ p, q ≤ ∞ and his results are optimal in the cases (1/p, 1/q) is located in the vertices of the square [0, 1]2. We will improve Grobner's results in the cases 1 ≤ p, q ≤ ∞ and our results also cover the cases 0 < p < 1 or 0 < q < 1. 22 4.1 Embedding between α-modulation spaces Theorem 4.1. Let (α1, α2) ∈ [0, 1) × [0, 1). Then M s1,α1 holds if and only if s1 (cid:62) s2 + R(p, q; α1, α2). p,q ⊂ M s2,α2 p,q (4.1) Remark. In the first versions of the paper, we obtain the sufficiency of Theorem 4.1, soon after Toft and Wahlberg [17] independently considered the embeddings between α-modulation and Besov spaces in the cases 1 ≤ p, q ≤ ∞ and they first showed the necessity of Theorem 4.1 in the regions (1/p, 1/q) ∈ (S2 ∪ S3) ∩ [0, 1]2 (see Fig. 3). After their work we can finally show the necessity of Theorem 4.1 in all cases. Proof. (Sufficiency) For every k ∈ Zn, we introduce k f (cid:54)= 0}. (cid:3)α1 Λ(k) = {l ∈ Zn : (cid:3)α2 (4.2) l If (cid:3)α2 l k f (cid:54)= 0, then k and l satisfy (cid:3)α1 (cid:104)l(cid:105) α2 (cid:104)l(cid:105) α2 1−α2 (lj − C) < (cid:104)k(cid:105) α1 1−α2 (lj + C) > (cid:104)k(cid:105) α1 1−α1 (kj + C), 1−α1 (kj − C) (4.3) for all j = 1, 2,··· , n. If k (cid:46) 1, it is easy to see that l (cid:46) 1. If k (cid:29) 1, analogous to (3.12), we have (cid:104)l(cid:105) ∼ (cid:104)k(cid:105) 1−α2 1−α1 . Assume that p (cid:62) 1, q = 1 and s2 = 0, we have (cid:107)(cid:3)α1 (cid:54) (cid:88) (cid:88) (cid:107)f(cid:107) l f(cid:107)p (cid:46)(cid:88) (cid:3)α2 k M 0,α2 p,1 k∈Zn l∈Λ(k) k We need an estimate of #Λ(k). Similar to (3.13), we have #Λ(k) ∼ 1 ∨ (cid:104)k(cid:105) n(α1−α2) 1−α1 . #Λ(k)(cid:107)(cid:3)α1 k f(cid:107)p. (4.4) (4.5) (4.6) If p = 2, inserting (4.6) into (4.5) and noticing (4.4), in view of Jensen's inequality we get s2+0∨ n(α1−α2) 2,1 2(1−α1) ,α1 M (cid:44)→ M s2,α2 2,1 . If p = ∞ or 1, from (4.5), (4.6), (4.4), one can directly obtain that (4.7) (4.8) (4.10) 2 , 1(cid:1) and (0, 1). By (4.9) Case 1: (cid:0) 1 p , 1 q ,α1 M 1−α1 s2+0∨ n(α1−α2) ∞,1 s2+0∨ n(α1−α2) 1,1 (cid:44)→ M s2,α2∞,1 , (cid:44)→ M s2,α2 . ,α1 M 1−α1 (cid:1) ∈ I. For any θ ∈ [0, 1],(cid:0) θ 0∨(cid:0)1− θ 2 , 1(cid:1) is at the line connecting(cid:0) 1 (cid:1) n(α1−α2) 1,1 ,α1 2 1−α1 (cid:44)→ M 0,α2 2 θ ,1 . M 2 θ ,1 complex interpolation between (4.7) and (4.8), one has that 23 Since(cid:0) θ 2 , 1− θ 2 (cid:1) is at the line segment by connecting(cid:0) 1 2 , 1 2 (cid:1) and (0, 1), A complex interpolation combined with Proposition 2.7 and (4.8) yields For any(cid:0) 1 p , 1 q (cid:1) ∈ I, we may suppose that there exists some (θ, η) ∈ [0, 1] × [0, 1], such that M . 2 (4.11) (cid:44)→ M 0,α2 θ , 2 2−θ 2 0∨(1−θ) n(α1−α2) θ , 2 2−θ 1−α1 1 p = 1 p1 = θ 2 , 1 q = 1 − θη 2 . s2+(cid:0) 1 (cid:1)[0∨n(α1−α2)],α1 Therefore, a complex interpolation between (4.10) and (4.11) implies that Case 2: (cid:0) 1 p , 1 q q − 1 (cid:1) ∈ II. For any θ ∈ [0, 1],(cid:0)1− θ M p,q p (cid:44)→ M s2,α2 (cid:1) is at the line segment connecting(cid:0) 1 p,q . (4.12) (cid:1) 2 , 1 2 2 , 1− θ 2 and (1, 1). From Proposition 2.7 and (4.9), we see that Noticing that(cid:0)1− θ Noticing that for any(cid:0) 1 and (4.9), we see that M s2+(1−θ)[0∨n(α1−α2)],α1 2 , 1(cid:1) is a point at the line segment connecting(cid:0) 1 (cid:44)→ M s2,α2 2−θ , 2 2−θ 2−θ , 2 2−θ . 2 2 (cid:1)[0∨n(α1−α2)],α1 s2+(cid:0)1− θ (cid:1) ∈ II, there exists some (θ, η) ∈ [0, 1] × [0, 1] satisfying (cid:44)→ M s2,α2 2 2−θ ,1 2 2−θ ,1 M . 2 2 , 1(cid:1) and (0, 1), from (4.7) (4.13) p , 1 q 1 p = 1 − θ 2 , 1 q = 1 − θη 2 , (4.14) (4.15) on the basis of (4.13) and (4.14), we conclude that When α1 (cid:54) α2, (4.15) coincides with (4.12). Case 3: (cid:0) 1 p , 1 q q p,q M p + 1 p , 1 p∗ , 1 s2+(cid:0) 1 (cid:1) ∈ I∗ ∪ II∗. When(cid:0) 1 q −1(cid:1)[0∨n(α1−α2)],α1 (cid:1) ∈ I∗,(cid:0) 1 −s2−(cid:0) 1 (cid:1)[0∨n(α2−α1)],α1 (cid:1) ∈ II∗, by (4.15) and duality one has that (cid:1)[0∨n(α2−α1)],α1 s2+(cid:0) 1 s2+(cid:0)1− 1 −s2,α2 p∗,q∗ (cid:44)→ M p− 1 p− 1 p∗,q∗ M M p,q q q p− 1 q M p,q When(cid:0) 1 p , 1 q The duality of α-modulation space implies that When α1 > α2, (4.17) coincides with (4.16). 24 (cid:44)→ M s2,α2 . p,q q∗(cid:1) is in I. From (4.12), we know (cid:1)[0∨n(α2−α1)],α1 . (cid:44)→ M s2,α2 p,q . (4.16) (cid:44)→ M s2,α2 p,q . (4.17) η ∈ (0, 1) satisfying q p , 1 Case 4: (cid:0) 1 Notice that(cid:0) 1 If(cid:0) 1 If(cid:0) 1 p , 1 p , 1 q q (cid:1) ∈ III ∪ III∗. We may assume that for any(cid:0) 1 (cid:1) and(cid:0) 1 (cid:1) ∈ III (III∗), there exists a (cid:1) are at the boundaries of II and I∗ (II∗ and I), respectively. 1 − η p 1 − 1 p (cid:18) (cid:19) p , 1 = η 1 q + . q p p p , 1 p , 1 − 1 (cid:104) (cid:1) ∈ III, a complex interpolation between (4.15) and (4.16) yields (cid:1) ∈ III∗, a complex interpolation between (4.12) and (4.17) yields (cid:44)→ M s2,α2 s2+ p,q p− 1 p + 1 M ,α1 p,q ; q q −1(cid:1)∨n(α2−α1)(cid:0) 1 (cid:1)∨n(α2−α1)(cid:0)1− 1 n(α1−α2)(cid:0) 1 n(α1−α2)(cid:0) 1 (cid:104) (cid:1)(cid:105) (cid:1)(cid:105) q − 1 p p− 1 q ,α1 (cid:44)→ M s2,α2 s2+ p,q M (4.19) When α1 > α2, (4.18) and (4.19) coincide with (4.15) and (4.12), respectively. When α1 (cid:54) α2, (4.18) and (4.19) coincide with (4.16) and (4.17), respectively. (cid:1) ∈(cid:98)I. Imitating the proof as in the counterpart of Theorem 3.1, we can easily Case 5: (cid:0) 1 p,q . p , 1 q get s2+0∨ n(α1−α2) ∞,q 1−α1 M 1 q ,α1 s2+0∨ n(α1−α2) 2,q 1−α1 q − 1 2 ⊂ M s2,α2 2,q . (cid:0) 1 (cid:1),α1 (4.18) A complex interpolation yields ⊂ M s2,α2∞,q , M (cid:0) 1 s2+0∨ n(α1−α2) p,q 1−α1 q − 1 p (cid:1),α1 M (cid:104)k(cid:105) s2q 1−α1 + 0∨n(α2−α1) 1−α1 k 25 ⊂ M s2,α2 . (4.20) Case 6: (cid:0) 1 p , 1 q (4.20) coincides with (4.12). p,q (cid:1) ∈ (cid:99)III. From (1.1b), as well as (4.4), we see that k f (·)] (cid:46) (cid:104)k(cid:105) α1∨α2 1−α1 , diam suppF [F −1ηα2 In view of Proposition 2.1, l (x − ·)(cid:3)α1 (cid:0) 1 p−1(cid:1) 0∨n(α1−α2) 1−α1 (cid:107)(cid:3)α2 k f(cid:107)p (cid:46) (cid:104)k(cid:105) (cid:3)α1 (cid:107)(cid:3)α1 k f(cid:107)p. (4.21) Inserting (4.21), (4.4), (4.6), from the embedding (cid:96)p ⊂ (cid:96)1 and with the aid of Jensen's inequality, we have l l∈Zn (cid:88) (cid:88) (cid:32)(cid:88) l  (cid:88) k∈Λ(l) (cid:104)l(cid:105) s2q 1−α2 (cid:104)l(cid:105) s2q 1−α2 + 0∨n(α2−α1) 1−α2 q p 1−α1 (cid:107)(cid:3)α1 k f(cid:107)p  1  q (cid:104)k(cid:105)(1−p) 0∨n(α1−α2) p−1(cid:1) 0∨(α1−α2) (cid:0) q p−1(cid:1) (cid:88) (cid:104)k(cid:105)nq(cid:0) 1 q −1(cid:1) p−1(cid:1)+nq 0∨(α1−α2) (cid:0) 1 (cid:0) q k∈Λ(l) 1−α1 p + 1 p 1−α1 (cid:107)f(cid:107)M s2,α2 p,q (cid:46) (cid:46) (cid:46) (cid:107)(cid:3)α1 p k f(cid:107)q (cid:33) 1 q (cid:107)(cid:3)α1 k f(cid:107)q p .  1 q (4.22) When α1 (cid:54) α2, (4.22) gives whereas when α1 > α2, (4.22) gives M p,q s2+n(α2−α1)(cid:0) 1 s2+n(α1−α2)(cid:0) 1 p− 1 q (cid:1),α1 q −1(cid:1),α1 ⊂ M s2,α2 p,q ; (4.23) M p,q p + 1 ⊂ M s2,α2 . (4.23) and (4.24) coincide with (4.16) and (4.15), respectively. (cid:1) ∈ (cid:98)II. This is a consequence of the results in Cases 5 and 6 by complex p,q Case 7: (cid:0) 1 p , 1 interpolation. q (4.24) (cid:3) 4.2 Embedding between Besov space and α-modulation space In this section, we study the embedding between 1-modulation space and α-modulation spaces. In an analogous way to the previous subsection, we start with the embedding for the same indices p, q. Theorem 4.2. Let α ∈ [0, 1). Then Bs1 Conversely, M s1,α Proof. (sufficiency) For every j ∈ Z+, we introduce p,q holds if and only if s1 (cid:62) s2 + R(p, q; α, 1). p,q holds if and only if s1 (cid:62) s2 + R(p, q; 1, α). p,q ⊂ M s2,α p,q ⊂ Bs2 Λ(j) = {k ∈ Zn : (cid:3)α k(cid:52)jf (cid:54)= 0,∀f ∈ S (cid:48)(Rn)}; (4.25) and for every k ∈ Zn, we introduce Λ(k) = {j ∈ Z+ : (cid:3)α k(cid:52)jf (cid:54)= 0,∀f ∈ S (cid:48)(Rn)}. (4.26) To a j ∈ Z+, for any k ∈ Λ(j), it is easy to see that the quantitative relationship between k and j is (cid:104)k(cid:105) 1 1−α ∼ 2j. (cid:107)(cid:3)α k(cid:52)jf(cid:107)p = (cid:88) (cid:88) j∈Z+ k∈Λ(j) (cid:107)(cid:52)j(cid:3)α k f(cid:107)p. When p (cid:62) 1, q = 1 and s2 = 0, we have (cid:54) (cid:88) (cid:88) k∈Zn j∈Λ(k) (cid:107)f(cid:107)M 0,α p,1 For any k ∈ Zn and any j ∈ Z+, it is easy to see #Λ(k) ∼ 1, #Λ(j) ∼ 2jn(1−α). (4.27) (4.28) (4.29) Thus when p = 2, combining (4.28), (4.29), also with the aid of Jensen's inequality, we get s+ n(1−α) 2,1 2 B (cid:44)→ M s,α 2,1 . If p = 1 or ∞, combining (4.29), (4.28), we get Bs+n(1−α) 1,1 Bs+n(1−α) ∞,1 (cid:44)→ M s,α 1,1 , (cid:44)→ M s,α∞,1. 26 (4.30) (4.31) (4.32) Case1. (cid:0) 1 yields p , 1 q (cid:1) ∈ I. For any θ ∈ [0, 1], a complex interpolation between (4.30) and (4.32) s+(cid:0)1− θ (cid:1)n(1−α) 2 B 2 θ ,1 (cid:44)→ M s,α 2 θ ,1 ; while combined with Proposition 2.7 and (4.32), yields (cid:44)→ M s,α θ , 2 2−θ Bs+(1−θ)n(1−α) θ , 2 2−θ 2 2 . In analogy to (4.12), we get from (4.33), (4.34) that s+n(cid:0) 1 q − 1 p (cid:1)(1−α) B p,q (cid:44)→ M s,α p,q . for 2 (cid:54) p (cid:54) ∞, considering (4.29), we have (cid:107)(cid:3)α (cid:54) (cid:88) (cid:88) (cid:107)f(cid:107)B0 k(cid:52)jf(cid:107)p (cid:46) (cid:88) p,1 j∈Z+ k∈Λ(j) k∈Zn Conversely, when we encounter the embedding of α-modulation spaces into Besov spaces, (cid:107)(cid:3)α k f(cid:107)p = (cid:107)f(cid:107)M 0,α p,1 , (4.36) (4.33) (4.34) (4.35) (4.38) (4.39) (4.40) (4.42) (4.43) which gives Case2. (cid:0) 1 p , 1 q yields (cid:1) ∈ II. For any θ ∈ [0, 1], a complex interpolation between (4.30) and (4.31) p,q (cid:44)→ Bs p,q. M s,α (4.37) s+(cid:0)1− θ (cid:1)n(1−α) 2 B 2 2−θ ,1 (cid:44)→ M s,α 2 2−θ ,1 . From Proposition 2.7 and (4.31) it follows that Bs+(1−θ)n(1−α) 2−θ , 2 2−θ 2 (cid:44)→ M s,α . 2 2−θ , 2 2−θ Analogous to (4.15), one can conclude from (4.38) and (4.39) that p−1(cid:1)(1−α) s+n(cid:0) 1 q∗(cid:1) ∈ I ∪ II, from (4.37), we see that M (cid:1) ∈ I∗ ∪ II∗. Since(cid:0) 1 (cid:44)→ M s,α p,q . q + 1 B p,q Considering the embedding of α-modulation spaces into Besov spaces, (4.37) still holds if (1/p, 1/q) ∈ II. Case3.(cid:0) 1 p , 1 q Thus, the duality between Bs p∗,q∗, as well as between M s,α p,q and M p∗ , 1 p,q and B−s p,q (cid:44)→ M s,α Bs p,q . −s,α p∗,q∗ (cid:44)→ B−s p∗,q∗. −s,α p∗,q∗ , implies that (4.41) Conversely, if one considers the embedding of α-modulation space into Besov space, it follows from Theorem 2.1 and (4.35) that and from (4.40) it follows that M p,q s+n(α−1)(cid:0) 1 s+n(α−1)(cid:0) 1 q − 1 p (cid:1),α q −1(cid:1),α p + 1 M p,q ∈ I∗; 1 q , (cid:18) 1 (cid:18) 1 p (cid:19) (cid:19) 1 q , p ∈ II∗. ⊂ Bs p,q, ⊂ Bs p,q, 27 Case4.(cid:0) 1 p , 1 q (cid:1) ∈ III ∪ III∗. If(cid:0) 1 s+n(cid:0)1− 2 p p , 1 q B p, p p−1 By interpolation we have which coincides with (4.35). If(cid:0) 1 B p , 1 q Bs p, p p−1 ⊂ M s,α p, p p−1 (cid:1) ∈ III∗, (4.35) and (4.41) contain that (cid:1)(1−α) s+n(cid:0) 1 (cid:1) ∈ III, (4.40) and (4.41) imply that p,p (cid:44)→ M s,α p,p . (cid:44)→ M s,α p, p p−1 , Bs q − 1 p,q p (cid:44)→ M s,α p,q , (cid:1)(1−α) s+n(cid:0) 2 p−1(cid:1)(1−α) p−1(cid:1)(1−α) p,p (cid:44)→ M s,α p,q , , B s+n(cid:0) 1 p,q q + 1 ⊂ M s,α p,q . By interpolation, which coincides with (4.40). B of Theorem 2.1 and (4.44) we have while for(cid:0) 1 p , 1 q s+n(α−1)(cid:0) 1 (cid:1),α (cid:1) ∈ III∗, from (4.45), we have s+n(α−1)(cid:0) 1 q − 1 M p,q p M p,q ⊂ Bs p,q, q −1(cid:1),α p + 1 (cid:19) (cid:18) 1 , 1 q p ⊂ Bs p,q. Conversely, considering the embedding of α-modulation space into Besov space, in view ∈ III; (4.46) (4.44) (4.45) Case5 :(cid:0) 1 (4.46) and (4.47) coincide with (4.42) and (4.43), respectively. (cid:1) ∈(cid:98)I. For the embedding of Besov space into α-modulation space, imitating p , 1 q (4.47) the argument in Theorem 4.1, we get s+n(1−α) 1 ∞,q q ,1 M From them, we interpolate out (cid:1),1 q − 1 2 s+n(1−α)(cid:0) 1 (cid:1),1 2,q ⊂ M s,α p,q , ⊂ M s,α p,q , M s+n(1−α)(cid:0) 1 q − 1 p M p,q ⊂ M s,α p,q . (4.48) which coincides with (4.35). Conversely for the embedding of α-modulation space into Besov space, we have p,q ⊂ Bs p,q, M s,α Case 6. (cid:0) 1 p , 1 q (cid:1) ∈ (cid:99)III. If 2α which coincides with (4.37). k(cid:52)j (cid:54)= 0, then diam suppF [F −1ηα So, in view of Proposition 2.1 we have (cid:104)k(cid:105)− nα 1−α k(cid:52)jf(cid:107)p (cid:46) (cid:104)k(cid:105) n (cid:107)(cid:3)α 1−α (cid:0) 1 p−1(cid:1) (4.49) (4.50) (cid:107)(cid:52)jf(cid:107)p. (4.51) 1−α ∼ 2j, k (x − ·)(cid:52)jf (·)] (cid:46) (cid:104)k(cid:105) 1 (cid:0) 1 p−1(cid:1) (cid:107)(cid:52)jf(cid:107)p ∼ 2jn(1−α)(cid:0) 1 p−1(cid:1) 28 From (4.51), (4.29), (cid:96)p ⊂ (cid:96)1 and Jensen's inequality it follows that (cid:107)f(cid:107)M s,α p,q (cid:46) (cid:46) (cid:46)  q p q  1  1  1 p p q q (cid:107)(cid:52)jf(cid:107)q k∈Zn (cid:88) (cid:88) (cid:88) k p (cid:107)(cid:3)α j∈Λ(k)  (cid:88) (cid:104)k(cid:105) sq k(cid:52)jf(cid:107)p 1−α 1−α +n(cid:0) 1 p−1(cid:1)(cid:105) q (cid:88) (cid:104) s (cid:104)k(cid:105) p−1(cid:1)(cid:105) s+n(1−α)(cid:0) 1 q (cid:88) (cid:104) s+n(1−α)(cid:0) 1 j∈Λ(k) k∈Λ(j) j 2 p + 1 B p,q (cid:107)(cid:52)jf(cid:107)q (cid:46) (cid:107)f(cid:107) q −1(cid:1),α ⊂ M s,α p,q . j which implies that (cid:0) 1 s+n(1−α) p,q B p + 1 q −1 (cid:1) , (4.52) (4.53) It coincides with (4.40). Conversely, when we study the embedding of α-modulation space into Besov space, in analogy to (4.50), we see diam suppF [F −1ϕj(x − ·)(cid:3)α k f (·)] (cid:46) 2j In contrast to (4.51), we conclude Inserting (4.54), the substitution for (4.52) is (cid:107)(cid:3)α k f(cid:107)p = (cid:107)(cid:3)α k f(cid:107)p. −jn(cid:0) 1 p−1(cid:1)  q k f(cid:107)p p p  1 q (cid:107)(cid:52)j(cid:3)α (cid:107)(cid:52)j(cid:3)α 2 2jsq j∈Z+ k f(cid:107)p (cid:46) 2jn(cid:0) 1 p−1(cid:1) (cid:88)  (cid:88) (cid:88) p−1 (cid:88) (cid:32)(cid:88) (cid:1)(cid:105) s+n(1−α)(cid:0) 1 s+n(α−1)(cid:0) 1 (cid:104) (cid:104)k(cid:105) 2jsq#Λ(j) j∈Λ(j) p− 1 k j q q k∈Λ(j) q − 1 p  1 q (cid:107)(cid:52)j(cid:3)α p k f(cid:107)q (cid:33) 1 q(cid:107)(cid:3)α k f(cid:107)q p q (cid:46) (cid:107)f(cid:107) (cid:1),α ⊂ M s,1 p,q . (cid:0) 1 s+n(1−α) p,q M p − 1 q (cid:107)f(cid:107)Bs p,q (cid:46) (cid:46) (cid:46) which implies that Case7 :(cid:0) 1 p , 1 q M (cid:1) ∈ (cid:98)II. It is interpolated out from Case 5 and Case 6. p,q It coincides with (4.42). (4.54) (4.55) (cid:3) (cid:1) , 29 Figure 4: Distribution of s0 5 Multiplication algebra It is well known that Bs space, this issue is much more complicated. The regularity indices for which M s,α a multiplication algebra, are quite different from those of Besov and modulation spaces. p,q is a multiplication algebra if s > n/p, cf. [18]. But for α-modulation p,q constitutes We introduce a parameter, denoted by s0 = s0(p, q; α), to describe the regularity for which M s,α p,q with s > s0 forms a multiplication algebra. Denote (see Figure 4) (cid:111) D1 = (cid:110)(cid:0) 1 (cid:1) ∈ R2 (cid:40) nα p + n(1 − α)(cid:0)1 − 1 ∧ 1 p + n(1 − α)(cid:0)1 ∨ 1 p ∨ 1 p , 1 nα q q and s0 = , D2 = R2 + \ D1 + : 1 q (cid:62) 2 p , 1 p (cid:54) 1 2 (cid:1) + nα(1−α) (cid:0) 1 (cid:1) + nα(1−α) q − 2 (cid:1), (cid:0)1 ∨ 1 2−α p 2−α q − 1 q q − 1(cid:1), p ∨ 1 (cid:0) 1 (cid:0) 1 q p , 1 p , 1 q (cid:1) ∈ D1; (cid:1) ∈ D2. Theorem 5.1. If s > s0, then M s,α that for any f, g ∈ M s,α p,q , we have p,q is a multiplication algebra, which is equivalent to say (cid:107)f g(cid:107)M s,α (cid:46) (cid:107)f(cid:107)M s,α (cid:107)g(cid:107)M s,α (5.1) In Section 7 we will give some counterexamples to show that s0 is sharp if (1/p, 1/q) ∈ D2 ∩{p ≥ 1}. When(cid:0)1/p, 1/q(cid:1) ∈ D1, it is not very clear for us to know the sharp low bound of the index s for which M s,α consequence of Theorem 5.1, we have the following result for which M s p,q constitutes a multiplication algebra. As a straightforward p,q is an algebra. p,q p,q p,q . Corrolary 5.1. Assume that (cid:40) n(cid:0)1 − 1 ∧ 1 n(cid:0)1 ∨ 1 p ∨ 1 q (cid:1), q − 1 q (cid:1), s > (cid:0) 1 (cid:0) 1 q p , 1 p , 1 q (cid:1) ∈ D1; (cid:1) ∈ R2 +\D1. 30 Then M s p,q is a multiplication algebra, i.e., holds for all f, g ∈ M s p,q. (cid:107)f g(cid:107)M s p,q (cid:46) (cid:107)f(cid:107)M s p,q(cid:107)g(cid:107)M s p,q (5.2) A natural long standing question on modulation, α-modulation and Besov spaces is: Can we reformulate α-modulation spaces by interpolations between modulation and Besov spaces, say, M s,α p,q = (M s0,0 p,q , M s1,1 p,q )α, if s = (1 − α)s0 + αs1? (5.3) The answer is negative at least for some special cases. Indeed, we see M s1,1 algebra if s1 > n/p and s0 > 0. If (5.3) holds, then M s,α this is not true if 1 < p < 2, 0 < q < 1, see Section 7. p,q are p,q is an algebra if s > nα/p, however, p,q and M s0,0 Corrolary 5.2. Let 0 < α < 1. Then (5.3) does not hold if 1 < p < 2, 0 < q < 1 and s0 = 0+, s1 = n/p. Proof of Theorem 5.1. We start with some notations and basic conclusions. For every (k(1), k(2)) ∈ Z2n, we introduce and for every k ∈ Zn, we introduce Λ(cid:0)k(1), k(2)(cid:1) =(cid:8)k ∈ Zn : (cid:3)α (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Z2n : (cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g) (cid:54)= 0(cid:9) ; (cid:111) k(2)g) (cid:54)= 0 Λ(k) = It is worth to mention that Λ(cid:0)k(1), k(2)(cid:1) and Λ(k) are independent of f and g. From (1.1b) we see that for any k ∈ Λ(k(1), k(2)), or (k(1), k(2)) ∈ Λ(k), k(1), k(2) and k satisfy k(1)f(cid:3)α k ((cid:3)α . (cid:104)k(cid:105) α (cid:104)k(cid:105) α 1−α (kj − C) < (cid:104)k(1)(cid:105) α 1−α (kj + C) > (cid:104)k(1)(cid:105) α 1−α (k(1) 1−α (k(1) j + C) + (cid:104)k(2)(cid:105) α j − C) + (cid:104)k(2)(cid:105) α 1−α (k(2) 1−α (k(2) j + C), j − C) (5.5b) for j = 1, 2,··· , n. Let (cid:104)kmax(cid:105), (cid:104)kmin(cid:105) and (cid:104)kmed(cid:105) to be the maximal, minimal and medial ones in (cid:104)k(cid:105),(cid:104)k(1)(cid:105) and (cid:104)k(2)(cid:105), respectively. If (5.5) holds, then (cid:104)kmax(cid:105) ∼ (cid:104)kmed(cid:105), #(cid:8)kmax : kmax, kmed, kmin satisfy (5.5)(cid:9) (cid:46) 1. (5.5a) We write (5.4) (5.6) (5.7) (5.8) In order to get more precise estimates, we divide Z2n of all (k(1), k(2)) into (cid:0)k(1), k(2)(cid:1) = (cid:104)k(1)(cid:105) α K(cid:0)k(1), k(2)(cid:1) = max Kj 1(cid:54)j(cid:54)n j + (cid:104)k(2)(cid:105) α 1−α k(1) Kj(k(1), k(2)). 1−α k(2) ; j (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Z2n : (cid:104)k(1)(cid:105) ∼ (cid:104)k(2)(cid:105)(cid:111) (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Z2n : (cid:104)k(1)(cid:105) (cid:29) (cid:104)k(2)(cid:105)(cid:111) (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Z2n : (cid:104)k(1)(cid:105) (cid:28) (cid:104)k(2)(cid:105)(cid:111) , , , Ω0 = Ω1 = Ω2 = 31 and separate Ω0 into Ω0,1 = Ω0,2 = (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Ω0 : K(cid:0)k(1), k(2)(cid:1) (cid:46) (cid:104)k(1)(cid:105) α (cid:110)(cid:0)k(1), k(2)(cid:1) ∈ Ω0 : K(cid:0)k(1), k(2)(cid:1) (cid:29) (cid:104)k(1)(cid:105) α 1−α 1−α (cid:111) (cid:111) ; If (k(1), k(2)) ∈ Ω0,1, from (5.5) it is easy to see that (cid:104)k(cid:105) (cid:46) (cid:104)k(1)(cid:105)α. . (5.9) (5.10) Let (k(1), k(2)) ∈ Ω0,2 be fixed. There exists some y := y(k(1), k(2)) ∈ (α, 1] such that Ki(k(1), k(2)) = K(k(1), k(2)) ∼ (cid:104)k(1)(cid:105) y (5.11) for some i with 1 (cid:54) i (cid:54) n. We can assume that Ki(k(1), k(2)) > 0. By (5.5) and (5.11) we have 1−α , (cid:104)k(cid:105) α (cid:104)k(cid:105) α 1−α (ki − C) < Ki(k(1), k(2)) + C((cid:104)k(1)(cid:105) α 1−α (ki + C) > Ki(k(1), k(2)) − C((cid:104)k(1)(cid:105) α 1−α + (cid:104)k(2)(cid:105) α 1−α + (cid:104)k(2)(cid:105) α 1−α ) (cid:46) (cid:104)k(1)(cid:105) y 1−α ) (cid:38) (cid:104)k(1)(cid:105) y 1−α , 1−α . , we substitute(cid:101)i for j in (5.5), thus (5.5a) and (5.5b) are (5.12b) (5.12a) For every k ∈ Λ(k(1), k(2)) ∩ Rn(cid:101)i rewritten as (cid:104)k(cid:105) α (cid:104)k(cid:105) α 1−α (k(cid:101)i − C) < K(cid:101)i(k(1), k(2)) + C((cid:104)k(1)(cid:105) α 1−α (k(cid:101)i + C) > K(cid:101)i(k(1), k(2)) − C((cid:104)k(1)(cid:105) α 1−α + (cid:104)k(2)(cid:105) α 1−α + (cid:104)k(2)(cid:105) α 1−α ); 1−α ). (5.13a) (5.13b) For such k, we claim that (cid:104)k(cid:105) ∼ (cid:104)k(1)(cid:105)y. (5.14) (5.14) is obvious when k(1) (cid:46) 1 and so, it suffices to consider (5.14) in the case k(1) (cid:29) If either ki ∼ (cid:104)k(cid:105) or K(cid:101)i(k(1), k(2)) ∼ K(k(1), k(2)) exists, (5.14) follows from (5.12) 1. or (5.13) directly. Otherwise, we see that ki (cid:28) (cid:104)k(cid:105) and K(cid:101)i(k(1), k(2)) (cid:28) K(k(1), k(2)). When k(cid:101)i > 0, we let (5.13a)+(5.12a) and (5.13b)+(5.12b); whereas when k(cid:101)i < 0, we let (5.13b)×(−1)+(5.12a) and (5.12a)×(−1)+(5.12b), then we get 1−α (cid:46) (cid:104)k(1)(cid:105) y which imply (5.14). Let (cid:101)k = (k1, ..., kj−1, kj, kj+1, ..., kn) ∈ Λ(k(1), k(2)). In view of (5.5a) 1−α (cid:38) (cid:104)k(1)(cid:105) y (cid:104)k(cid:105) 1 (cid:104)k(cid:105) 1 1−α , 1−α , and (5.5b) and (k(1), k(2)) ∈ Ω0,2, we have (cid:12)(cid:12)(cid:12)(cid:104)k(cid:105) α 1−α kj − (cid:104)(cid:101)k(cid:105) α 1−α(cid:101)kj (cid:12)(cid:12)(cid:12)kj −(cid:101)kj (cid:12)(cid:12)(cid:12) (cid:46) (cid:104)k(1)(cid:105) α (cid:12)(cid:12)(cid:12) (cid:46) (cid:104)k(1)(cid:105) α(1−y) 1−α . 1−α + (cid:104)k(cid:105) α 1−α + (cid:104)(cid:101)k(cid:105) α 1−α . Thus Taylor's theorem, combined with (5.14), gives For(cid:0)k(1), k(2)(cid:1) ∈ Ω1 ∪ Ω2, in view of (5.6) we have (cid:104)k(cid:105) ∼ (cid:104)k(1)(cid:105) ∨ (cid:104)k(2)(cid:105) 32 (5.15) (5.16) and #Λ(cid:0)k(1), k(2)(cid:1) ∼ 1. (5.17) In what follows, we separate the proof into four steps. In Steps 1-3, we prove (5.1) for certain p and q. In Step 4, applying the complex interpolation together with the conclusions obtained in the previous three steps, we can get (5.1). Step 1. 1 (cid:54)(cid:54)(cid:54) p (cid:54)(cid:54)(cid:54) ∞, q (cid:54)(cid:54)(cid:54) 1. Suppose f, g ∈ M s,α embedding (cid:96)q ⊂ (cid:96)1, we have p,q , from the triangle inequality and the (cid:33) 1 q k∈Zn (cid:32)(cid:88) (cid:88) (cid:88)  2(cid:88) k k (cid:107)f g(cid:107)M s,α p,q = (cid:54) (cid:54) = (cid:104)k(cid:105) sq k (f g)(cid:107)q p (cid:104)k(cid:105) sq 1−α (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α  (cid:88) 1−α (cid:88) (cid:88) k(1),k(2) k(1),k(2) (cid:88) (cid:96)=0 (k(1),k(2))∈Ω(cid:96) k∈Λ(k(1),k(2)) q q 1  1 q (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)q p (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)q p  1 q . Applying the multiplier estimate and Holder's inequality, we see that (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p (cid:46) (cid:107)(cid:3)α k(1)f(cid:3)α (cid:88) k(2)g(cid:107)p (cid:46) (cid:107)(cid:3)α q − 1(cid:1), we have (cid:0) 1 For (k(1), k(2)) ∈ Ω0,1, when p = 1 and s (cid:62) nα + nα(1−α) 2−α k(2)g)(cid:107)q 1−α(cid:107)(cid:3)α (cid:88) (cid:46) (cid:88) k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) sqα k(1)f(cid:107)p(cid:107)(cid:3)α 1−α +nα(cid:107)(cid:3)α (k(1),k(2))∈Ω0,1 k(1)f(cid:3)α 1−α(cid:107)(cid:3)α (cid:54) (cid:107)f(cid:107)q (cid:104)k(cid:105) sq k ((cid:3)α 1(cid:104)k(2)(cid:105) nqα k(1)f(cid:107)q k(2)g(cid:107)q 1 1 k(2)g(cid:107)∞. (5.18) (5.19) (5.20) (k(1),k(2)) If p = 2 and s (cid:62) nα 2 + nα(1−α) (cid:88) 2−α (cid:104)k(cid:105) sq (k(1),k(2))∈Ω0,1 k∈Λ(k(1),k(2)) (cid:107)g(cid:107)q . M s,α 1,q M s,α 1,q q − 1(cid:1), by Plancherel and Jensen's inequality∗, we have (cid:0) 1 1−α (cid:88) k(2)g)(cid:107)q (cid:107)(cid:3)α k(1)f(cid:3) k ((cid:3)α 1−α(cid:107)(cid:3)α (cid:104)k(cid:105) sq 2 (2)g)(cid:107)q 2 k ((cid:3)α sup k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) sqα 1−α [#Λ(k(1), k(2))]1− q k∈Λ(k(1),k(2)) 2(cid:107)(cid:3)α 1−α +nα(cid:0)1− q 2 (cid:1) (cid:104)k(1)(cid:105) sqα k(1)f(cid:3)α k(2)g(cid:107)q 2 k(1)f(cid:3)α (cid:107)(cid:3)α k(1)f(cid:107)q 2(cid:104)k(2)(cid:105) nqα 2(1−α)(cid:107)(cid:3)α k(2)g(cid:107)q 2 (5.21) (cid:54) (cid:107)f(cid:107)q M s,α 2,q (cid:107)g(cid:107)q M s,α 2,q . (cid:88) (cid:46) (cid:88) (cid:46) (cid:88) (cid:46) (cid:88) (k(1),k(2)) (k(1),k(2)) (k(1),k(2)) ∗aθ 1 + ... + aθ N ≤ N 1−θ(a1 + ... + aN )θ for θ ∈ (0, 1) 33 For p = ∞ and s (cid:62) nα(1−α) 2−α (cid:88) (cid:46) (cid:88) (k(1),k(2))∈Ω0,1 (cid:88) 1 q , we have (cid:104)k(cid:105) sq k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) sqα 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)q∞ 1−α +nα(cid:107)(cid:3)α k(1)f(cid:107)q∞(cid:107)(cid:3)α k(2)g(cid:107)q∞ (cid:54) (cid:107)f(cid:107)q (cid:107)g(cid:107)q M s,α∞,q . (5.22) M s,α∞,q (k(1),k(2)) For (k(1), k(2)) ∈ Ω0,2, when s (cid:62) nα(1 + (1 − α)/q)/(2 − α), we see that 2s (cid:62) sy + nα (1 + q − y) q . (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) syq 1−α + nα 1−α (1−y)(cid:107)(cid:3)α k(1)f(cid:107)q 1 k(2)g)(cid:107)q 1(cid:104)k(2)(cid:105) nαq 1−α(cid:107)(cid:3)α k(2)g(cid:107)q 1 (cid:46) (cid:107)f(cid:107)M s,α 1,q (cid:107)g(cid:107)M s,α 1,q . (5.23) When p = ∞ and s (cid:62) nα(1−α) q(2−α) , it is suffices to get (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) syq 1−α + nα 1−α (1−y)(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(1)f(cid:107)∞(cid:107)(cid:3)α k(2)g)(cid:107)q∞ k(2)g(cid:107)∞ (cid:46) (cid:107)f(cid:107)M s,α∞,q (cid:107)g(cid:107)M s,α∞,q ; (5.24) , by Plancherel and Jensen's inequality, (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)q k ((cid:3)α 2 k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) syq 1−α [#Λ(k(1), k(2))]1−q/2(cid:107)(cid:3)α (cid:104)k(1)(cid:105) syq 1−α + nα (1−α) (1−y)(1−q/2)(cid:107)(cid:3)α 2 k(1)f(cid:3)α k(2)g(cid:107)q 2(cid:104)k(2)(cid:105) nαq k(1)f(cid:107)q 2(1−α)(cid:107)(cid:3)α k(2)g(cid:107)q 2 (5.25) (cid:88) k(1),k(2) (k(1),k(2))∈Ω0,2 It follows that(cid:88) (cid:46) (cid:88) (cid:88) (cid:46) (cid:88) (cid:88) (cid:46) (cid:88) (cid:46) (cid:88) (k(1),k(2))∈Ω0,2 (k(1),k(2))∈Ω0,2 k(1),k(2) (cid:88) (cid:88) If p = 2 and s (cid:62) nα(α/2+(1−α)/q) 2−α k(1),k(2) k(1),k(2) 2,q (cid:46) (cid:107)f(cid:107)M s,α (cid:88) (cid:46) (cid:88) . 2,q (cid:107)g(cid:107)M s,α (cid:88) From (5.16), (5.17), (2.2), if (k(1), k(2)) ∈ Ω1, when s (cid:62) nα p , k(2)g)(cid:107)q k(2)g(cid:107)q∞ (cid:46) (cid:107)f(cid:107)M s,α k(1)f(cid:3)α (cid:88) (cid:107)(cid:3)α k∈Λ(k(1),k(2)) (cid:104)k(1)(cid:105) sq (k(1),k(2))∈Ω1 k(1)f(cid:107)q 1−α(cid:107)(cid:3)α 1−α(cid:107)(cid:3)α (cid:104)k(cid:105) sq k ((cid:3)α p,q p p (cid:107)g(cid:107)M 0,α∞,q (5.26) k(2)∈Zn Similarly, if (k(1), k(2)) ∈ Ω2 and s (cid:62) nα p , k(1)∈Zn (cid:46) (cid:107)f(cid:107)M s,α (cid:88) p,q . p,q (cid:107)g(cid:107)M s,α (cid:88) (k(1),k(2))∈Ω2 k∈Λ(k(1),k(2)) 34 (cid:104)k(cid:105) sq 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)q p (cid:46) (cid:107)f(cid:107)M s,α p,q (cid:107)g(cid:107)M s,α p,q . (5.27) (cid:0) 1 q − 1(cid:1); and when 2 < p (cid:54) ∞, from (5.21) and (5.22), we get s0 = (cid:1). By complex interpolation, (5.18)-(5.27) imply that M s,α p,q is a multiplication algebra as long as s (cid:62) s0 for some s0. More precisely, when 1 (cid:54) p (cid:54) 2, from (5.20) and (5.21), we get (cid:0) 1 p + nα(1−α) s0 = nα 2−α p + nα(1−α) q − 2 nα 2−α Step 2. 0 < p (cid:54)(cid:54)(cid:54) ∞, q = ∞. First, we consider the case 1 (cid:54) p (cid:54) ∞. Suppose f, g ∈ M s,α p,∞, from the triangle inequality, we have (cid:104)k(cid:105) s (cid:104)k(cid:105) s (cid:107)f g(cid:107)M s,α (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p p p,∞ = sup k∈Zn (cid:54) sup k∈Zn (5.28) 1−α(cid:107)(cid:3)α 1−α (cid:88) k (f g)(cid:107)p 2(cid:88) (cid:88) (k(1),k(2))∈Λ(k) = sup k∈Zn (cid:96)=0 (cid:104)k(cid:105) s k ((cid:3)α 1−α(cid:107)(cid:3)α (k(1),k(2))∈Λ(k)∩Ω(cid:96) k(2)g)(cid:107)p. k(1)f(cid:3)α (cid:110) (cid:111) k(1) ∈ Zn : ∃k(2) ∈ Zn s.t.(cid:0)k(1), k(2)(cid:1) ∈ Ψ (cid:110) (cid:111) k(2) ∈ Zn : ∃k(1) ∈ Zn s.t.(cid:0)k(1), k(2)(cid:1) ∈ Ψ . ; For a Ψ ⊂ Z2n, we denote Ψ⊥ 1 = Ψ⊥ 2 = Let s > nα we easily see #Λ(−k(2), k) (cid:46) 1. Inserting (5.19) and using (2.3), we obtain 2 with every fixed k, noticing (5.6), (k(1),k(2))∈{Ω0∪Ω1}∩Λ(k) p + n(1− α). For any k(2) ∈ {{Ω0 ∪ Ω1} ∩ Λ(k)}⊥ (cid:88) k(1)∈(cid:8){Ω0∪Ω1}∩Λ(k)(cid:9)⊥ k(2)g)(cid:107)p 1−α(cid:107)(cid:3)α 1−α(cid:107)(cid:3)α k ((cid:3)α (cid:104)k(cid:105) s (cid:104)k(cid:105) s k(1)f(cid:3)α k(1)∈(cid:8){Ω0∪Ω1}∩Λ(k)(cid:9)⊥ k(1)f(cid:107)p (cid:88) (cid:88) (cid:88) k(2)∈(cid:8){Ω0∪Ω1}∩Λ(k)(cid:9)⊥ k(1)f(cid:107)p (cid:104)k(1)(cid:105) s 1−α(cid:107)(cid:3)α sup 1 1 k(1)∈Λ(−k(2),k) 2 (cid:46) (cid:46) sup k(1)∈Zn (cid:46) (cid:107)f(cid:107)M s,α (cid:46) (cid:107)f(cid:107)M s,α p,∞(cid:107)g(cid:107)M s,α p,∞. (cid:88) k(2)∈Λ(−k(1),k) (cid:107)(cid:3)α k(2)g(cid:107)∞ (cid:107)(cid:3)α k(2)g(cid:107)∞ (5.29) p,∞(cid:107)g(cid:107)M 0,α∞,1 (cid:88) For k(1) ∈ {Ω2 ∩ Λ(k)}⊥ (k(1),k(2))∈Ω2∩Λ(k) (cid:46) sup k(2)∈Zn (cid:46) (cid:107)f(cid:107)M 0,α∞,1 (cid:104)k(2)(cid:105) s 1−α(cid:107)(cid:3)α k(2)g(cid:107)p (cid:107)g(cid:107)M s,α p,∞ (cid:46) (cid:107)f(cid:107)M s,α 1 with every fixed k, symmetrically, we have (cid:104)k(cid:105) s 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α (cid:88) k(2)g)(cid:107)p (cid:88) k(1)∈{Ω2∩Λ(k)}⊥ p,∞(cid:107)g(cid:107)M s,α p,∞. p + n(1− α), M s,α 1 k(2)∈Λ(−k(1),k) Combining (5.28) -- (5.30), we know when s > nα Next, we consider the case 0 < p < 1 and q = ∞. Suppose that f, g ∈ M s,α p,∞ is a multiplication algebra. p,∞. It follows (cid:107)(cid:3)α k(1)f(cid:107)∞ (5.30) 35 from the embedding (cid:96)p ⊂ (cid:96)1 that (cid:107)f g(cid:107)M s,α p,∞ (cid:54) sup k  (cid:88) (cid:88) (k(1),k(2))∈Λ(k)  1 p (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p p (cid:104)k(cid:105) sp 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p p  1 p . (cid:104)k(cid:105) s 1−α  2(cid:88) = sup k (cid:96)=0 (k(1),k(2))∈Λ(k)∩Ω(cid:96) By Proposition 2.1, for (k(1), k(2)) ∈ Λ(k) ∩ Ω0, we have p−1)(cid:107)F −1ηα p−1)(cid:104)k(cid:105)− nα k(2)g)(cid:107)p (cid:46) (cid:104)k(1)(cid:105) nα (cid:46) (cid:104)k(1)(cid:105) nα k(1)f(cid:3)α k ((cid:3)α (cid:107)(cid:3)α (1−α) ( 1 k (cid:107)p(cid:107)(cid:3)α k(1)f(cid:3)α k(2)g(cid:107)p k(1)f(cid:3)α (1−α) ( 1 (1−α) ( 1 k(2)g(cid:107)p. p − 1), inserting (5.32) and using (2.3), we obtain that k ((cid:3)α p−1)(cid:107)(cid:3)α k(2)g)(cid:107)p k(1)f(cid:3)α p When s ≥ n (cid:88) (k(1),k(2))∈Λ(k)∩Ω0 (cid:46) p + nα(1−α) (2−α) ( 1 (cid:104)k(cid:105) sp (cid:88) 1−α(cid:107)(cid:3)α (k(1),k(2))∈Λ(k)∩Ω0 (cid:104)k(1)(cid:105) sp 1−α(cid:107)(cid:3)α 1−α (cid:104)k(1)(cid:105) spy+nα(1−y)(1−p) (cid:88) k(1)f(cid:107)p p (cid:107)(cid:3)α k(1)f(cid:107)p p(cid:107)(cid:3)α k(2)g(cid:107)p∞ (cid:88) (cid:104)k(2)(cid:105) sp(y−1)+nα(1−y)(1−p) 1−α k(2)∈{Λ(k)∩Ω0}⊥ 2 k(1)∈Λ(−k(2),k) (cid:46) sup k(1)∈Zn (cid:46) (cid:107)f(cid:107)p p,∞(cid:107)g(cid:107)p M s,α p,∞. M s,α (5.31) (5.32) (cid:107)(cid:3)α k(2)g(cid:107)p∞ (5.33) For any (k(1), k(2)) ∈ Λ(k) ∩ {Ω1 ∪ Ω2} with every fixed k, imitating the process as in (5.32) and combining (5.16), we get (cid:107)(cid:3)α (5.34) k(2)g(cid:107)p. When s > n k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p (cid:46) (cid:107)(cid:3)α k(1)f(cid:3)α p , inserting (5.34) and also using (2.3), we obtain k(2)g)(cid:107)p 1−α(cid:107)(cid:3)α 2(cid:88) (cid:104)k(cid:105) sp k ((cid:3)α p (cid:88) (k(1),k(2))∈Λ(k)∩Ω(cid:96) p,∞(cid:107)g(cid:107)p (cid:46) (cid:107)f(cid:107)p M s,α (cid:96)=1 k(1)f(cid:3)α (cid:107)g(cid:107)p (cid:46) (cid:107)f(cid:107)p Combining (5.31),(5.33) and (5.35), we conclude when s ≥ n + (cid:107)f(cid:107)p M 0,α∞,p M 0,α∞,p M s,α p,∞ M s,α p,∞(cid:107)g(cid:107)p p,∞. M s,α p + nα(1−α) (2−α) ((1∨ 1 p )− 1), M s,α p,∞ p,p . From the embedding (cid:96)p ⊂ (cid:96)1 ⊂ (cid:96)1/p it follows is a multiplication algebra. Step 3. p < 1, q = p. Suppose f, g ∈ M s,α that (5.35) (cid:107)f g(cid:107)M s,α p,p (cid:54) =  1 p (cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p p (cid:88)  2(cid:88) k (cid:96)=0 (cid:104)k(cid:105) sp 1−α (cid:88) (cid:88) k(1),k(2) (cid:88) (k(1),k(2))∈Ω(cid:96) k∈Λ(k(1),k(2)) 36 (cid:104)k(cid:105) sp 1−α(cid:107)(cid:3)α k ((cid:3)α k(1)f(cid:3)α k(2)g)(cid:107)p p  1 p . (5.36) Figure 5: For Step 4 in the proof of Theorem 5.1. For (k(1), k(2)) ∈ Ω0, if s (cid:62) nα p + nα(1−α) (2−α) ( 1 p − 1), then we see from (5.10), (5.14), (5.15) and (5.32) that(cid:88) (cid:88) (k(1),k(2))∈Ω0 k∈Λ(k(1),k(2)) (cid:46) (cid:88) (cid:46) (cid:88) (k(1),k(2))∈Ω0 (k(1),k(2))∈Ω0 (cid:107)g(cid:107)p (cid:46) (cid:107)f(cid:107)p M s,α p,p M s,α p,p . (cid:104)k(cid:105) sp 1−α(cid:107)(cid:3)α k ((cid:3)α k(2)g)(cid:107)p k(1)f(cid:3)α 1−α (1−y)+ nα p (cid:104)k(1)(cid:105) spy 1−α− nαy 1−α (1−p)+ nα 1−α (1−p)(cid:107)(cid:3)α p(cid:107)(cid:3)α k(2)g(cid:107)p∞ (cid:104)k(1)(cid:105) spy 1−α + nα 1−α (1−p)(1−y)+ nα 1−α (2−y)(cid:107)(cid:3)α k(1)f(cid:107)p k(2)g(cid:107)p p k(1)f(cid:107)p p(cid:107)(cid:3)α (5.37) p (cid:88) (cid:88) For (k(1), k(2)) ∈ Ω1 ∪ Ω2, when s (cid:62) nα M s,α p,p M 0,α∞,p k ((cid:3)α + (cid:107)f(cid:107)p (cid:104)k(cid:105) sp 1−α(cid:107)(cid:3)α k∈Λ(k(1),k(2)) (cid:107)g(cid:107)p (k(1),k(2))∈Ω1∪Ω (cid:46) (cid:107)f(cid:107)p M 0,α∞,p Combining (5.36)-(5.38), we conclude when s (cid:62) nα algebra. p , in view of (2.2), we obtain k(2)g)(cid:107)p k(1)f(cid:3)α (cid:46) (cid:107)f(cid:107)p (cid:107)g(cid:107)p (cid:0) 1 p − 1(cid:1), M s,α p + nα(1−α) q < 1(cid:9). (cid:1) ∈ D1 : 1 (cid:1) is a point at It is easy to see that (cid:0) 1 (cid:1) ∈ (cid:8)(cid:0) 1 Step 4. Let (cid:0) 1 2−α the line segment connecting (cid:0) 1 (cid:0)1 − 1 (cid:1), 1(cid:1) and (cid:0)0, 1 (cid:1), which is parallel to the line 2 , 1(cid:1) and (0, 0). At the point ( 1 ¯p , 1) := (cid:0) 1 (cid:0)1 − 1 (cid:1), 1(cid:1), in Step 1 we have connecting (cid:0) 1 ¯p,1 is a multiplication algebra if s (cid:62) nα(cid:2) 1 (cid:1). For (cid:0) 1 (cid:1)(cid:3) + nα(1−α) (cid:0)1 − 1 ¯q ) :=(cid:0)0, 1 (cid:1), complex interpolation between (0, 1) in Step 1 and (0, 0) in Step 3 shows q − 2 shown that M s,α q − 2 (0, 1 q − 2 p + 1 p + 1 p,p is a multiplication (cid:107)g(cid:107)p p + 1 2 (5.38) p , 1 q p , 1 q p , 1 q 2−α M s,α p,p M s,α p,p q q . M s,α p,p p 2 2 p p q 37 p , 1 q q + 2 p (cid:1) + nα(1−α) 2−α (cid:0) 1 q − 2 p that once s > n(1− α)(cid:0)1− 1 is a multiplication algebra. Again, using the complex interpolation and combining the result in Step 2, and we arrive at (5.1) in D1. Denote (see Fig. 5) (cid:1), the associated α-modulation space M s,α∞,¯q (cid:111)(cid:47) (cid:1), M s,α 2q,q is a multiplication algebra. For D1 ∪ D21. 2q , 1 (cid:62) 1 D21 = [0, 1] × [0, 1]\D1; + : 1 D23 = q (cid:110)(cid:0) 1 (cid:1) ∈ R2 (cid:1), we see that once s > nα 2q + n(1− α)(cid:0)1− 1 Through the point(cid:0) 1 (cid:1) ∈ D21, one can make a line segment connecting(cid:0) 1 (cid:1) and(cid:0)1, 1 (cid:1). For(cid:0) 1 (cid:1), complex interpolation between (1, 1) in Step 1 and (1, 0) in Step 3 shows that once (cid:0)1, 1 s > nα + n(1 − α)(cid:0)1 − 1 (cid:1), the associated α-modulation space is a multiplication algebra. n(1 − α)(cid:0)1 − 1 If(cid:0) 1 (cid:1). In Step 4 (cid:1). If (1/p, 1/q) ∈ D23, the result can be derived in a similar way. (cid:1) ∈ D22, then it belongs to the segment by connecting(cid:0) 1 p , 0(cid:1) and(cid:0) 1 (cid:0) 1 p − 1(cid:1), M s,α p + nα(1−α) p,∞ is a multiplication algebra; and 2−α p,p is a multiplication algebra. Then complex interpolation we see that once s (cid:62) nα p + nα(1−α) once s (cid:62) nα 2−α between them gives once s (cid:62) nα algebra. p − 1(cid:1), M s,α (cid:0) 1 (cid:1) + nα(1−α) Then we use complex interpolation to get that M s,α p + n(1− α)(cid:0) 1 (cid:0) 1 p − 1(cid:1), M s,α p + n(1 − α) 1 p,q is a multiplication algebra if s > nα p,q is a multiplication (cid:3) p − 1 2q , 1 q p + p , 1 p p q p , 1 q q q q 2−α q q p , 1 q q 6 Sharpness for the scaling and embedding properties In this section we show the necessity of Theorems 3.1, 4.1 and 4.2. Since the p-BAP U has no scaling, it is difficult to calculate the norm for a known function. However, we have the following equivalent norm on α-modulation spaces. Let ρ be a smooth radial bump function supported in B(0, 2), satisfying ρ(ξ) = 1 as ξ < 1, and ρ(ξ) = 0 as ξ (cid:62) 2. Let ρα k be as in (1.4): ρα k (ξ) = ρ ξ − (cid:104)k(cid:105)α/(1−α)k C(cid:104)k(cid:105)α/(1−α) (cid:32) (cid:32)(cid:88) k∈Zn (cid:33) . (cid:33)1/q Proposition 6.1. Let ρα k be as in the above. Then (cid:107)f(cid:107)◦ M s,α p,q = (cid:104)k(cid:105) sq 1−α(cid:107)(F −1ρα k ) ∗ f(cid:107)q p is an equivalent norm on M s,α p,q . Proof. If p ≥ 1, in view of Young's inequality, k ) ∗ f(cid:107)p ≤ (cid:88) (cid:96)∞≤C (cid:107)(F −1ρα (cid:107)(F −1ρα k+(cid:96))(cid:107)1(cid:107)2α k f(cid:107)p (cid:46) (cid:107)2α k f(cid:107)p. If p < 1, by Proposition 2.1 and the scaling argument, we have k ) ∗ f(cid:107)p ≤ (cid:104)k(cid:105)nα(1/p−1)/(1−α) (cid:88) (cid:107)(F −1ρα (cid:107)(F −1ρα k+(cid:96))(cid:107)p(cid:107)2α k f(cid:107)p (cid:46) (cid:107)2α k f(cid:107)p. (cid:96)∞≤C 38 Combining the above two cases, we have (cid:107)f(cid:107)◦ that M s,α p,q (cid:17) ≤ (cid:107)f(cid:107)M s,α (cid:16) ξ−(cid:104)k(cid:105)α/(1−α)k (cid:16) ξ−(cid:104)k+(cid:96)(cid:105)α/(1−α)(k+(cid:96)) 2C(cid:104)k(cid:105)α/(1−α) p,q (cid:17) := ρα ρ k σα k , (cid:96)∈Zn ρ C(cid:104)k+(cid:96)(cid:105)α/(1−α) (cid:80) ηα k = ρα k We have for p ≥ 1, in view of Young's inequality, . On the other hand, noticing (cid:107)2α k f(cid:107)p ≤ (cid:107)(F −1ρα k ) ∗ f(cid:107)p(cid:107)(F −1σα k )(cid:107)1 (cid:46) (cid:107)(F −1ρα k ) ∗ f(cid:107)p. If p < 1, by Proposition 2.1, the scaling argument and the generalized Bernstein inequality, we have (cid:107)2α k f(cid:107)p (cid:46) (cid:104)k(cid:105)nα(1/p−1)/(1−α)(cid:107)(F −1ρα k f(cid:107)p (cid:46) (cid:107)(F −1ρα k ) ∗ f(cid:107)p(cid:107)F −1σα k ) ∗ f(cid:107)p. k (cid:107)p(cid:107)2α (cid:3) The result follows. Proof of Theorem 3.1. (Necessity) We divide the proof into the following two cases λ (cid:29) 1 and λ (cid:28) 1, respectively. Case 1. λ (cid:29) 1. One needs to show that − n (cid:38) λ p +n(cid:0) 1 p−1(cid:1)∨(s+sc)(cid:107)f(cid:107)M s,α (cid:107)fλ(cid:107)M s,α . p,q p,q Case 1.1. We consider the case sp = n(1/p − 1) > s + sc. Our aim is to show that there exists a function f satisfying (cid:107)fλ(cid:107)M s,α p,q (cid:38) λ−n(cid:107)f(cid:107)M s,α p,q . Taking f = F −1ρ, we have (cid:107)fλ(cid:107)◦ M s,α p,q ≥ (cid:107)(F −1ρα 0 ) ∗ fλ(cid:107)p (cid:62) λ−n(cid:107)F −1ρ(cid:107)p (cid:38) λ−n(cid:107)f(cid:107)M s,α p,q . (6.1) Case 1.2. We consider the case sp < s + sc. According to the definition of sc, we separate the proof into the following three cases. Case 1.2.1. sc = n(1 − α)(1/p + 1/q − 1). Put f = F −1ρ. Since λ (cid:29) 1, we see that for some 0 < ε0 < ε1 (cid:28) 1, It follows that k ∈ [ε0λ1−α, ε1λ1−α].  1 q (cid:104)l(cid:105) sq 1−α(cid:107)F −1ρα l (cid:107)q p (F −1ρα (cid:107)fλ(cid:107)◦ M s,α p,q k ) ∗ fλ = λ−nF −1ρα k ,  (cid:88) p +s+n(1−α)(cid:0) 1 p +s+n(1−α)(cid:0) 1 (cid:38) λ−n (cid:38) λ (cid:38) λ − n − n q −1(cid:1) q −1(cid:1) p + 1 p + 1 l∈[ε0λ1−α, ε1λ1−α] (cid:107)f(cid:107)M s,α p,q . (6.2) Case 1.2.2. sc = 0. Let us take f = eix1F −1ρλ. We have fλ = λ−neix1λF −1ρ. We may assume that there exists l0 ∈ N such that (cid:104)l0(cid:105)α/(1−α)l0 ∼ λ and B((λ, 0, ..., 0), 2) ⊂ p−1(cid:1) B((cid:104)l0(cid:105)α/(1−α)(l0, 0, ..., 0), C(cid:104)l0(cid:105)α/(1−α)). It is easy to see that (cid:107)f(cid:107)p ∼ λn(cid:0) 1 and (cid:107)fλ(cid:107)M s,α p,q ∼ λ−n(cid:104)l0(cid:105)s/(1−α) ∼ λ−n+s (cid:38) λ − n p +s(cid:107)f(cid:107)p (cid:38) λ − n p +s(cid:107)f(cid:107)M s,α p,q . (6.3) 39 (cid:107)fλ(cid:107)M s,α p,q (cid:38) 1−α(cid:107)(f (l))λ(cid:107)q l∈A(λ) (cid:104)l(cid:105) sq  (cid:88)  (cid:88) p +s+n(1−α)(cid:0) 1 (cid:32) l∈A(λ) − n − n p +s q q p  1  1 p−1(cid:1) (cid:33) p . (cid:107)f (l)(cid:107)q q + 1 λξ − (cid:104)l(cid:105)α/(1−α)l c(cid:104)l(cid:105)α/(1−α) . (cid:38) λ (cid:38) λ (cid:98)f (l)(ξ) = ρ (6.6) Case 1.2.3. sc = n(1 − α)(1/q − 1/p). Put (cid:32) (cid:32) (cid:33)(cid:33)∨ λ · c(cid:104)l(cid:105) α 1−α (cid:88) (cid:96)∼λ1−α , f = f (l), (6.4) f (l) = eix(cid:104)l(cid:105) α 1−α l/λ τλ1−αl ρ then (cid:107)f (l)(cid:107)p ∼ λn(1−α)(cid:0) 1 p−1(cid:1) , and where (cid:96) ∼ λ1−α means that l ∈ [ε0λ1−α, ε1λ1−α] for some 0 < ε0 < ε1 (cid:28) 1. If l ∼ λ1−α, (f (l))λ = eix(cid:104)l(cid:105) α 1−α l τλ1−αl ρ c(cid:104)l(cid:105) α 1−α l ) ∗ (f (l))λ = (f (l))λ. Since supp ρα l overlaps at most finite many l+k(f (l))λ = 0 if k ≥ C. Let A(λ) ⊂ {l : l ∼ λ1−α} be the set so (6.5) , which follows that (F −1ρα supp ρα l+k, we see that 2α that for any l, l ∈ A(λ) (l (cid:54)= l), l − l ≥ C. We have (cid:32) (cid:32) · (cid:33)(cid:33)∨ Moreover, we easily see that It follows that supp (cid:98)f (l) is included in the unit ball. Hence, we have By Plancheral's identity, (cid:13)(cid:13)(cid:13)2 (cid:13)(cid:13)(cid:13)(cid:98)f (cid:107)f(cid:107)2 = (cid:107)f(cid:107)M s,α p,q (cid:46) (cid:90)  (cid:88) l∈A(λ) = ρ2 (cid:88) l∈A(λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p f (l) . (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:32) (cid:33) 1/2 dξ ∼ 1. λξ − (cid:104)l(cid:105)α/(1−α)l c(cid:104)l(cid:105)α/(1−α) On the other hand, in view of F −1ρ is a Schwartz function, we have F −1ρ(x) (cid:46) (cid:104)x(cid:105)−N . It follows that for N (cid:29) n, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:88) l∈A(λ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:46) λ−n(1−α) (cid:88) l∈A(λ) f (l)(x) (cid:16) 1 + λ−(1−α)x − λ1−αl(cid:17)−N (cid:46) λ−n(1−α). 40 By Holder's inequality, (cid:107)f(cid:107)M s,α p,q (cid:46) (cid:107)f(cid:107)p (cid:46) λn(1−α)(2/p−1). The result follows. Case 2. λ (cid:28) 1. It suffices to show that for some f ∈ M s,α p,q , (cid:107)fλ(cid:107)◦ M s,α p,q (cid:38) λ − n p (1 ∨ λs+sc)(cid:107)f(cid:107)M s,α p,q . Case 2.1. s + sc ≥ 0. Taking f = F −1ρ, we have (cid:107)fλ(cid:107)◦ M s,α p,q = (cid:107)(F −1ρ)(λ·)(cid:107)p ∼ λ Case 2.2. s + sc < 0. We divide the proof into the following three cases. Case 2.2.1. We can find some k0 such that λ(cid:104)k0(cid:105) 1 1−α ∼ 1. Denote (cid:32) (cid:32) p,q − n p (cid:107)f(cid:107)M s,α (cid:33)(cid:33) · (cid:33) , , c(cid:104)k0(cid:105) α 1−α f = eix(cid:104)k0(cid:105) α 1−α k0F −1 ρ (cid:32) (cid:98)fλ = λ−nρ ξ/λ − (cid:104)k0(cid:105) α c(cid:104)k0(cid:105) α 1−α 1−α k0 . (6.7) We have Therefore, (cid:107)fλ(cid:107)◦ M s,α p,q (cid:38) (cid:107)fλ(cid:107)p (cid:38) λ − n p +s(cid:104)k0(cid:105) s 1−α(cid:107)f(cid:107)p (cid:38) λ − n p +s(cid:107)f(cid:107)M s,α p,q . Case 2.2.2. Taking f = λnF −1ρλ, we have fλ = F −1ρ. It follows that (cid:107)fλ(cid:107)◦ the other hand, M s,α p,q  (cid:88)  (cid:88) k(cid:46)λα−1 k(cid:46)λα−1 (cid:107)f(cid:107)◦ M s,α p,q (cid:46) (cid:46) (cid:104)k(cid:105) sq 1−α(cid:107)(F −1ρα k ) ∗ f(cid:107)q p (cid:104)k(cid:105) sq 1−α(cid:107)F −1ρα k )(cid:107)q p(cid:107)f(cid:107)q 1 q −nα(cid:0)1− 1 p (cid:1) . (cid:46) λn−s−n(1−α) 1 (6.8) (6.9) (6.10) ∼ 1. On (6.11) q  1  1 q (cid:88) (cid:33)(cid:33) l∈A(λ) It follows that (cid:107)fλ(cid:107)◦ M s,α p,q (cid:38) λ Case 2.2.3. sc = n(1− α)(1/q − 1/p). Let A(λ) be the set so that for any l, l ∈ A(λ) (l (cid:54)= l), l − l ≥ C and l ∈ [ε0λα−1, ε1λα−1] for some 0 < ε0 < ε1 (cid:28) 1. Take . p,q − n p +s+sc(cid:107)f(cid:107)M s,α (cid:33)(cid:33) (cid:32) · c(cid:104)l(cid:105) α 1−α (cid:32) f (l) = eix(cid:104)l(cid:105) α 1−α lτC(cid:48)lF −1 ρ , f = One easily sees that (cid:32) (cid:32) ρ (f (l))λ = λ−neixλ(cid:104)l(cid:105) α 1−α lτC(cid:48)lF −1 41 · cλ(cid:104)l(cid:105) α 1−α f (l). (6.12) . (6.13) We have (cid:107)f(cid:107)◦ M s,α p,q = Since supp(cid:98)f (l) λ is contained in the unit ball, we see that (cid:107)fλ(cid:107)M s,α p,q (cid:38) q  1  1 q 1−α (1− 1 p )q 1−α(cid:107)(f (l))(cid:107)q p (cid:46) λ−s l∈A(λ) (cid:104)l(cid:105) sq (cid:104)l(cid:105) nα  (cid:88)  (cid:88) p−1(cid:1)+ n(α−1) −s+nα(cid:0) 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p = λ−n/p l∈A(λ) f (l) λ q (cid:46) λ (cid:88) l∈A(λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . (cid:88) l∈A(λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p f (l) (cid:16) (cid:17) (6.14) . We note that f (l)(x) = (cid:104)l(cid:105) nα 1−α eix(cid:104)l(cid:105) α 1−α l(F −1ρ) c(cid:104)l(cid:105) α 1−α (x − C(cid:48)l) . Due to F −1ρ is a Schwartz function, we see that F −1ρ(x) (cid:46) (cid:104)x(cid:105)−2n. We can assume that F −1ρ(0) = 1, which follows that there exists a δ > 0 such that F −1ρ(x) ≥ 1/2, x ≤ δ. (cid:0)y + C(cid:48)c l(cid:1)−2n . (cid:88) l∈Zn (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−nα 2 Hence, for any k ∈ A(λ), x ∈ B(C(cid:48)k, δ/c(cid:104)k(cid:105) α 1−α ), l∈A(λ) − Cλ−nα sup y(cid:46)δ We can take C(cid:48)c (cid:29) 1. It follows that for any k ∈ A(λ), − Cλ−nα(C(cid:48)c)−n ≥ λ−nα (cid:88) (cid:88) So, we have f (l)(x) l∈A(λ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 4 f (l)(x) (cid:88) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≥ λ−nα (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) l∈A(λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p f (l)(x) (cid:38) λ−nαλ 2nα−n p . x ∈ B(C(cid:48)k, δ/c(cid:104)k(cid:105) α 1−α ). , It follows that The result follows. (cid:107)fλ(cid:107)◦ M s,α p,q (cid:38) λ−n/p+s+n(1−α)(1/q−1/p)(cid:107)f(cid:107)M s,α p,q . (cid:3) 42 Denote Aj = {k : suppρα k ⊂ {ξ : 5 · 2j−3 ≤ ξ ≤ 3 · 2j−1}}. Proof of Theorem 4.2. (Necessity) Case 1. Let us assume that Bs1 Case 1.1. We show that s1 ≥ s2. Let k ∈ Zn, k (cid:29) 1 with B((cid:104)k(cid:105) α 5 · 2j−3 ≤ ξ ≤ 3 · 2j−1}. We see that p,q ⊂ M s2 p,q. 1−α k, C(cid:104)k(cid:105) α 1−α ) ⊂ {ξ : (cid:13)(cid:13)F −1 (ρα (cid:13)(cid:13)F −1 (ρα k (2 ·))(cid:13)(cid:13)◦ k (2 ·))(cid:13)(cid:13)Bs1 M s2,α p,q p,q (cid:62) (cid:104)k(cid:105) s2 1−α(cid:107)F −1ρα k (2 ·)(cid:107)p (cid:38) (cid:104)k(cid:105) s2 1−α + nα 1−α (cid:54) 2js1(cid:107)F −1ρα k (2 ·)(cid:107)p (cid:46) (cid:104)k(cid:105) s1 1−α + nα 1−α (cid:1) , (cid:0)1− 1 (cid:1) (cid:0)1− 1 p p . p,q ⊂ M s2,α Bs1 Case 1.2. We show that s1 ≥ s2 + n(1 − α)(1/p + 1/q − 1). One has that follows that s1 (cid:62) s2. p,q (cid:107)F −1ϕj(cid:107)Bs1 (cid:88) k∈Aj (cid:107)F −1ϕj(cid:107)◦ M s2,α p,q (cid:62) (cid:104)k(cid:105) s2q k(cid:107)q 1−α(cid:107)F −1ρα Noticing that #Aj ∼ O(2nj(1−α)), we immediately have p,q p,q M s2,α (cid:107)F −1ϕj(cid:107)◦ p,q ⊂ M s2,α Bs1 implies that s1 (cid:62) s2 + n(1 − α)(cid:0) 1 Case 1.3. We show that s1 (cid:62) s2 + n(1 − α)(cid:0) 1 (cid:1) , (cid:0)F −1ρα f (k) = τk k Noticing that #Aj ∼ O(2nj(1−α)), we have q p . p p,q (cid:38) (cid:1) (6.15) (cid:1)q (cid:104)k(cid:105) s2q  1 (cid:0)1− 1 (cid:46) 2js1+jn(cid:0)1− 1 (cid:88)  1 (cid:1)+n(1−α)j 1 (cid:38) 2s2j+nαj(cid:0)1− 1 q − 1(cid:1). (cid:1). We denote by Aj the set such that for (cid:88) p + 1 q − 1 1−α + nα 1−α (6.16) (6.17) k∈Aj q . p . p p q (6.18) f = f (k). k∈Aj every k, l ∈ Zn ∩ Aj (l (cid:54)= l), k − l ≥ C and k1/(1−α) ∈ [5· 2j−3 + C2jα, 3· 2j−1 − C2jα]. Put (cid:107)f(cid:107)◦ M s2,α p,q = On the other hand, By Plancherel's identity, (cid:1)(cid:17) . (6.19) (6.20) q +nα(cid:0)1− 1 p s2+n(1−α) 1 (cid:88) k∈Aj (cid:104)k(cid:105) s2q 1−α(cid:107)f (k)(cid:107)q p (cid:107)f(cid:107)Bs1 p,q (cid:46) 2js1 (cid:88) k∈Aj (cid:107)f(cid:107)2 (cid:46) (cid:107)ρα k (2 ·)(cid:107)2 q j (cid:16) (cid:38) 2  1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:88) 1/2 k∈Aj 2 f (k) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p . (cid:46) 2nj/2. Moreover, let us observe that f (k) = C(cid:104)k(cid:105) nα 1−α (F −1ρ)(C(cid:104)k(cid:105) nα 1−α (x − k)). 43 Using the same way as in the proof of Case 1.2.3 in Theorem 3.1, we have f (x) (cid:46) 2nαj. By Holder's inequality, that s1 (cid:62) s2 + n(1 − α)(cid:0) 1 (cid:107)f(cid:107)p (cid:46) 2jnα(1−2/p)+nj/p. q − 1 (cid:1). p,q ⊂ M s2,α p,q It follows from Bs1 Case 2. We assume that M s1,α1 Case 2.1. Let j (cid:29) 1. One can find some k ∈ Zn verifying ρα 2j ∼ (cid:104)k(cid:105)1/(1−α). Thus, p,q ⊂ Bs2 p,q. p k ϕj = ρα k . It follows that (cid:0)1− 1 p (cid:1) , (6.21) (6.22) (6.23) (cid:1) . (6.24) q +s1j+jnα(cid:0)1− 1 p (cid:1) . (cid:107)F −1ρα p . p,q p,q 1−α M s1,α p,q 1−α + nα 1−α (cid:46) (cid:104)k(cid:105) s1 (cid:0)1− 1 (cid:13)(cid:13)Bs2 (cid:62) 2js2(cid:107)F −1ρα (cid:62) 2js2(cid:107)F −1ϕ2 k(cid:107)Bs2 (cid:107)F −1ρα k(cid:107)◦ k(cid:107)p (cid:38) 2js2(cid:104)k(cid:105) nα (cid:1) j(cid:107)p (cid:38) 2js2+jn(cid:0)1− 1 (cid:13)(cid:13)F −1ϕj  1  (cid:88) (cid:1)q p,q that s1 (cid:62) s2 + n(1 − α)(cid:0)1 − 1 (cid:33)(cid:33) 1−α 2jnα(cid:0)1− 1 (cid:32)· − (cid:104)k(cid:105) α (cid:1). (cid:88) (cid:46) 2jn(1−α) 1 k1/(1−α)∼2j p − 1 (cid:104)k(cid:105) s1q (cid:32) p q p f (k) = τC(cid:48)kF −1 ρ So, we have s1 (cid:62) s2. Case 2.2. Let j (cid:29) 1. We have On the other hand, (cid:13)(cid:13)F −1ϕj (cid:13)(cid:13)M s1,α p,q (cid:46) p,q ⊂ Bs2 It follows from M s1,α Case 2.3. We denote by Aj the set such that for every k, l ∈ Zn ∩ Aj (l (cid:54)= k), k − l ≥ C and k1/(1−α) ∈ [5 · 2j−3 + C2jα, 3 · 2j−1 − C2jα]. Put q Using the same way as in the proof of Case 2.2.3 in Theorem 3.1, (cid:107)f(cid:107)◦ M s1,α p,q (cid:46) 2 j From M s1,α p,q ⊂ Bs2 (cid:107)f(cid:107)Bs2 p,q it follows that s1 (cid:62) s2 + n(1 − α)(cid:0) 1 s2+n(1−α) 1 p − 1 = 2js2 (cid:107)f(cid:107)p (cid:38) 2 p,q q j , f = f (k). (6.25) 1−α k c(cid:104)k(cid:105) α 1−α (cid:104) s1+n(1−α) 1 (cid:104) k∈Aj p q +nα(cid:0)1− 1 (cid:1)(cid:105) p +nα(cid:0)1− 1 (cid:1). , p (cid:1)(cid:105) . (6.26) (6.27) (cid:3) 44 Proof of Theorem 4.1. (Necessity) We separate the proof into the following two cases. p,q ⊂ M s2,α2 Case 1. We assume that M s1,α1 Case 1.1. We show that s1 ≥ s2. Denote p,q with α1 (cid:62) α2. Λ2(k) = {l ∈ Zn : ρα2 k ρα1 l (cid:54)= 0}. (cid:62) (cid:104)k(cid:105) s2 1−α2 (cid:107)F −1ρα2 k (cid:107)p (cid:38) (cid:104)k(cid:105) s2 1−α2 k ρα2 + nα2 1−α2 M s2,α2 We have and (cid:13)(cid:13)F −1ρα2 k (cid:13)(cid:13)◦ (cid:13)(cid:13)F −1ρα2 p,q k (cid:13)(cid:13)◦  (cid:88) l∈Λ2(k) (cid:46) M s1,α1 p,q (cid:104)l(cid:105) s1q 1−α1 (cid:107)F −1ρα2 (cid:0)1− 1 p (cid:1) , (6.28) . (6.29)  1 q Using Young's inequality and Proposition 2.1, respectively for p ≥ 1 and p < 1, we have Since #Λ2(k) is finite and k 1 1−α1 for all l ∈ λ2(k), one has that (cid:13)(cid:13)F −1ρα2 k (cid:13)(cid:13)◦ (cid:46) M s1,α1 p,q p,q ⊂ M s2,α2 it follows that s1 (cid:62) s2. From M s,α1 Case 1.2. Let k ∈ Zn with k (cid:29) 1. Denote p,q (cid:107)F −1ρα2 l (cid:107)p (cid:46) (cid:107)F −1ρα2 k (cid:107)p (cid:46) (cid:104)k(cid:105) nα2 1−α2 (cid:0)1− 1 p (cid:1)q  1 q k ρα1 1−α2 ∼ l 1 k∈Λ2(k) (cid:104)k(cid:105) sq 1−α1 (cid:104)k(cid:105) nα2 1−α2  (cid:88) Λ∗(k) =(cid:8)l ∈ Zn : ρα1 k ρα2 l = ρα2 l (cid:0)1− 1 p (cid:1) . (cid:46) (cid:104)k(cid:105) s 1−α2 + nα2 1−α2 p l (cid:107)q k ρα1 (cid:1) (cid:0)1− 1 p . (cid:9) . (cid:0)1− 1 p (cid:1) . (6.30) (6.31) (6.32) (cid:1) . (6.33) (cid:0)1− 1 p (6.34) (6.35) We have (cid:46) (cid:104)k(cid:105) s1 1−α1 + nα1 1−α1 (cid:107)F −1ρα1 and (cid:104)k(cid:105) k (cid:107)M s1,α1 1−α1 ∼ (cid:104)l(cid:105) p,q 1 1−α2 for all l ∈ Λ∗(k), one has that 1 (cid:107)F −1ρα1 k (cid:107)◦ Since #Λ∗(k) ∼ (cid:104)k(cid:105) n(α1−α2) 1−α1  1  (cid:88) implies that s1 (cid:62) s2 + n(α1 − α2)(cid:0) 1 Case 1.3. We show that s1 (cid:62) s2 + n(α1 − α2)(cid:0) 1 1−α2 (cid:107)F −1ηα2 l (cid:107)q p,q ⊂ M s2,α2 (cid:104)l(cid:105) s2q l∈Λ∗(k) M s,α1 M s2,α2 (cid:62) p,q p,q p q q − 1 p that l − l ≥ C for all l, l ∈ Λ∗ 0(k) (l (cid:54)= l). Denote f (l) = τlF −1ρα2 , l p + 1 q − 1(cid:1). (cid:1). Let Λ∗ (cid:88) f = f (l). l∈Λ∗ 0(k) (cid:38) (cid:104)k(cid:105) n(α1−α2) q(1−α1) + s2 1−α1 + nα2 1−α1 It follows that (cid:107)f(cid:107)◦ M s2,α2 p,q =  (cid:88) l∈Λ∗ 0(k) (cid:104)l(cid:105) s2q 1−α2 (cid:107)f (l)(cid:107)q p  1 q (cid:38) (cid:104)k(cid:105) s2 1−α1 + n(α1−α2) q(1−α1) + nα2 1−α1 (cid:0)1− 1 p (cid:1) . 45 0(k) be the subset of Λ∗(k) such (cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2 On the other hand, (cid:107)f(cid:107)2 = ∼ (cid:104)k(cid:105) nα2 2(1−α1) + n(α1−α2) 2(1−α1) , ρα2 l l∈Λ∗ (1−α2)(F −1)(C(cid:104)l(cid:105) α2 and noticing that f (l)(x) = (cid:104)l(cid:105) nα2 (cid:107)f(cid:107)∞ (cid:46) (cid:104)k(cid:105) nα2 0(k) (1−α1) . (1−α2) (x − l)), we have Hence, It follows that (cid:107)f(cid:107)p (cid:46) (cid:104)k(cid:105) nα2 (1−α1) (1− 1 p )+ n(α1−α2) p(1−α1) . (cid:0)1− 1 p (cid:1) . (6.36) M s1,α1 (cid:107)f(cid:107)◦ (cid:46) (cid:104)k(cid:105) s1 implies that s1 (cid:62) s2 + n(α1 − α2)(cid:0) 1 (cid:46) (cid:104)k(cid:105) s1 1−α1 q − 1 1−α1 (cid:107)f(cid:107)p p,q (cid:1). + n(α1−α2) p(1−α1) + nα2 1−α1 p,q ⊂ M s2,α2 p,q M s,α1 Case 2. We assume that M s1,α1 , α1 < α2. The idea is the same as in the proof of Theorem 4.2 and we only give a sketch proof. Case 2.1. We show that s1 ≥ s2. Let k ∈ Zn, k (cid:29) 1. One can find some l ∈ Zn such that ρα1 k ρα2 p,q ⊂ M s2,α2 l = ρα1 k , Then, p,q p (cid:107)F −1ρα1 k (cid:107)◦ (cid:107)F −1ρα1 k (cid:107)◦ implies that s1 (cid:62) s2. (cid:38) (cid:104)k(cid:105) s2 1−α1 + nα1 1−α1 M s2,α2 p,q (cid:46) (cid:104)k(cid:105) s1 1−α1 Case 2.2. We show that s1 (cid:62) s2 + n(α1 − α2)(cid:0) 1 p,q ⊂ M s2,α2 M s1,α1 M s1,α1 p,q p,q (6.37) (6.38) (6.39) (6.40) (cid:1) . (cid:1) (cid:0)1− 1 , p 0(k) be the subset of Λ∗(k) such that l − l ≥ C for all l, l ∈ Λ∗ Let Λ∗ see that #Λ∗ 0(k) ∼ (cid:104)k(cid:105) n(α2−α1) 1−α2 (6.41) 0(k) (l (cid:54)= l). It is easy to + nα1 1−α1 q − 1(cid:1). + nα2 1−α2 (cid:0)1− 1 (cid:0)1− 1 p p (cid:1) (cid:1) , . p + nα1 1−α2 (cid:0)1− 1 q − 1(cid:1). (cid:9) . (cid:88) (cid:0)1− 1 (cid:0)1− 1 0(k) p p l∈Λ∗ (cid:1) (cid:1) , . f = f (l). p,q ⊂ M s2,α2 M s1,α1 Case 2.3. Let k ∈ Zn with k (cid:29) 1, and p,q k k p,q p,q M s1,α1 (cid:13)(cid:13)◦ p + 1 (cid:38) (cid:104)k(cid:105) s2 1−α2 M s2,α2 (cid:46) (cid:104)k(cid:105) n(α2−α1) (cid:13)(cid:13)F −1ρα2 (cid:13)(cid:13)F −1ρα2 (cid:13)(cid:13)◦ implies that s1 (cid:62) s2 + n(α1 − α2)(cid:0) 1 Λ∗(k) =(cid:8)l ∈ Zn : ρα1 (cid:1) , f (l) = τClF −1(cid:0)ρα1 . Put l q(1−α2) + s1 1−α2 p + 1 l ρα2 k = ρα1 l then, p,q ⊂ M s2,α2 M s1,α1 p,q + n(α2−α1) q(1−α2) + nα1 1−α2 M s1,α1 p,q (cid:107)f(cid:107)◦ (cid:107)f(cid:107)◦ (cid:46) (cid:104)k(cid:105) s1 1−α2 (cid:38) (cid:104)k(cid:105) s2 1−α2 implies that s1 (cid:62) s2 + n(α2 − α1)(cid:0) 1 + n(α2−α1) M s2,α2 p,q p(1−α2) + nα1 1−α2 (cid:1). p − 1 q 46 (6.42) (6.43) (6.44) (cid:3) 7 Counterexamples for the algebra structure version in [9]. Let Q(a, r) :=(cid:81)n In order to show that our results are sharp, we need an α-covering which is a slightly modified i=1[ai − r, ai + r] and we consider the following covering of R: Q0 = [−1, 1], Qj = Q(jα/(1−α)j, rjjα/(1−α)), j (cid:54)= 0. Lemma 7.1. Let r > 1/2(1 − α). There exists j0 ∈ N, such that {Qj}j∈Z is an α-covering of R, where Moreover, if r < 8/15(1 − α), then rj = (cid:26) r, j > j0, j ≤ j0. suitable, Qj±1 ∩ Q(jα/(1−α)j, rjjα/(1−α)) = ∅ 7 8 (7.1) for all j ∈ Z. Proof. Let j > 100. Noticing that Qj+1 ∩ Qj (cid:54)= ∅ if and only if j + 1α/(1−α)(j + 1) − rj+1j + 1α/(1−α) < jα/(1−α)j − rjjα/(1−α). (7.2) In view of mean value theorem, we see that (7.2) is equivalent to 1 j + θα/(1−α) < rj+1j + 1α/(1−α) + rjjα/(1−α), (7.3) where θ ∈ (0, 1). Take rj+1 = rj = r. Hence, there exists j0 := j0(α) such that for any j > j0, (7.3) holds. Next, if j > j0, we have 1 − α j + 1α/(1−α)(j + 1) − rj + 1α/(1−α) > jα/(1−α)j + (7.4) which implies (7.1) for j > j0. If j ≤ j0, one can choose suitable rj so that the conclusion (cid:3) holds. Using Lemma 7.1 and the idea as in [9], we now construct a new α-covering of Rn, where the original idea goes back to Lizorkin's dyadic decomposition to Rn. Let j ∈ Z with j > j0. We may assume 8j/7r ∈ N. We divide [−j 1 1−α ] into 16j/7r equal intervals: rjα/(1−α), 1−α ,j 1 7 8 [−j 1 1−α , j 1 1−α ] = [rj,−Nj , rj,−Nj +1] ∪ ... ∪ [rj,Nj−1, rj,Nj ]. Denote We further write R = {rj,s : j ∈ N, s = −Nj,··· , Nj}. j = {k = (k1,··· , kn) : ki ∈ R, max K n 1≤i≤n ki = j 1 1−α}. For any k ∈ K n j , we write Qkj = Q(k, rj α 1−α ), j > j0. 47 From the construction of Qkj one sees that (cid:40) #{Qk(cid:48)j(cid:48) : Qk(cid:48)j(cid:48) ∩ Qkj (cid:54)= ∅} = 2n, Q(k, rj α 1−α /2) ∩ Qk(cid:48)j(cid:48) (cid:54)= ∅ ⇔ k(cid:48) = k, j(cid:48) = j. For j ≤ j0, one can choose suitable rj, and in a similar way as above to define so that 1 ≤ j ≤ j0, Q0 = Q(0, r0) Qkj = Q(k, rjj α 1−α ),  #{Qk(cid:48)j(cid:48) : Qk(cid:48)j(cid:48) ∩ Qkj (cid:54)= ∅} = 2n, Q(k, rjj α Q(0, r0/2) ∩ Qk(cid:48)j(cid:48) (cid:54)= ∅ ⇔ Qk(cid:48)j(cid:48) = Q0. 1−α /2) ∩ Qk(cid:48)j(cid:48) (cid:54)= ∅ ⇔ k(cid:48) = k, j(cid:48) = j f or j (cid:54)= 0, Lemma 7.2. {Q0} ∪ {Qkj}j∈Z\{0}, k∈Kj as in (7.5) and (7.6) is an α-covering of Rn. Let η : R → [0, 1] be a smooth bump function satisfying ξ ≤ 1/2, η(ξ) := smooth, 1/2 < ξ ≤ 1, 0, ξ ≥ 1.  1, (7.5) (7.6) (7.7) Let r and rj be as in (7.5) and (7.6), respectively. Denote for i = 1,··· , n, j (cid:54)= 0, φkj(ξi) = η k = (k1,··· , kn) ∈ Kj, φ0(ξi) = η , (cid:33) (cid:32) ξi − ki rjj α 1−α (cid:18) ξi (cid:19) r0 , φkj(ξ) = φkj(ξ1)...φkj(ξn), φ0(ξ) = φ0(ξ1)...φ0(ξn). We put ψkj(ξ) = φ0(ξ) +(cid:80) φkj(ξ) k∈Kj ,j∈Z\{0} φkj(ξ) , ψ0(ξ) = φ0(ξ) +(cid:80) φ0(ξ) k∈Kj ,j∈Z\{0} φkj(ξ) . (7.8) Lemma 7.3. {ψ0} ∪ {ψkj}j∈Z\{0}, k∈Kj as in (7.8) is a p-BAPU. On the basis of the above p-BAP U , we immediately have Proposition 7.1. Let 0 < α < 1, 0 < p, q (cid:54) ∞, then (cid:107)F −1ψ0F f(cid:107)q (cid:107)f(cid:107)M s,α p,q = (cid:88) (cid:104)j(cid:105)sq/(1−α) (cid:88) k∈Kj Lp(Rn) + j∈Z\{0} 1/q (cid:107)F −1ψkjF f(cid:107)q Lp(Rn) is an equivalent norm on α-modulation space. Theorem 7.1. Let 0 ≤ α < 1, (1/p, 1/q) ∈ D2, p > 1. If s < s0, then M s,α algebra. p,q is not a Banach 48 Proof. Step 1. p > 1, q ≤ 1. Let χA be the characteristic function on A. Now we take for J (cid:29) j0, (cid:96) = (J, J, ..., J), (cid:98)f = χA(J), (cid:98)g = χ−A(J), A(J) = Q(Jα/(1−α)(cid:96), rJα/(1−α)/2). (cid:40) 0, 2b − ξ, ξ ≥ 2b, ξ < 2b. ξi ≥ rJ α (1−α) for some i = 1, ..., n, ξi < rJ α (1−α) for all i = 1, ..., n. (7.9) (7.10) (7.11) Noticing that Hence, (cid:0)χA(J) ∗ χ−A(J) n(cid:89) i=1 Hence, One has that Let us write (χ[B−b,B+b] ∗ χ[−B−b,−B+b])(ξ) = 0, n(cid:89)  (cid:1) (ξ) = supp(cid:98)f ∗(cid:98)g = {ξ ∈ Rn : ξi ≤ rJ α (1−α) − ξi), (rJ α i=1 (1−α) , i = 1, ..., n}. (1−α) )n−1(cid:88) ξi i ξiξj + ... + ξ1...ξn (rJ α (1−α) )n + (rJ α (1−α) + ξi) = (rJ α + (rJ α (1−α) )n−2(cid:88) := (rJ α (1−α) )n + R(ξ, J). i<j Aj = {k ∈ K n j : if ξ, η ∈ suppψk,j, then ξiηi > 0 for all i = 1, ..., n}. Let 1 (cid:28) j < εJα (0 < ε (cid:28) 1) and k ∈ Aj. We may assume that ξi > 0 if ξ ∈ suppψkj. Noticing that suppψkj ⊂ Q(0, rJα/(1−α)), we have (cid:107)F −1ψkjF (f g)(cid:107)p = (cid:13)(cid:13)F −1(ψkjR(ξ, J))(cid:13)(cid:13)p (cid:46) J (n−1)α 1−α + J (n−2)α 1−α ≥ (rJ α ≥ cJ nα i=1 1−α (1− 1 (rJ α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p 1−αj nα (1−α) − ξi) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F −1ψkj n(cid:89) 1−α )n(cid:107)F −1ψkj(cid:107)p −(cid:13)(cid:13)F −1(ψkjR(ξ, J))(cid:13)(cid:13)p p ) −(cid:13)(cid:13)F −1(ψkjR(ξ, J))(cid:13)(cid:13)p . (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)p (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)F −1(ψkj n(cid:88) (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)F −1(ψkj (cid:88) ξiξj) ξi) i=1 i<j + ... +(cid:13)(cid:13)F −1(ψkjξ1...ξn)(cid:13)(cid:13)p . (7.13) (7.12) 49 For instance, we estimate (cid:107)F −1(ψkjξ1ξ2)(cid:107)p. Let k be the center of suppψkj. We have (cid:13)(cid:13)F −1(ψkjξ1ξ2)(cid:13)(cid:13)p (cid:13)(cid:13)p +(cid:13)(cid:13)F −1(ψkj(ξ1 − k1)(ξ2 − k2))(cid:13)(cid:13)p (cid:46) k1k2(cid:13)(cid:13)F −1ψkj + k1(cid:13)(cid:13)F −1(ψkj(ξ2 − k2))(cid:13)(cid:13)p + k2(cid:13)(cid:13)F −1(ψkj(ξ1 − k1))(cid:13)(cid:13)p (cid:13)(cid:13)F −1(ψkjR(ξ, J))(cid:13)(cid:13)p p ) (cid:46) ε2J 2α 1−α (1− 1 p ). 1−α (1− 1 p ). (cid:46) εJ nα 1−αj nα 1−αj nα (cid:46) j 2 1−αj nα 1−α (1− 1 So, one has that (7.14) (7.15) (7.16) (7.17) (7.18) (7.19) (7.20) (7.21) It follows from (7.12) and (7.15) that (cid:107)F −1ψkjF (f g)(cid:107)p (cid:38) J nα 1−αj nα 1−α (1− 1 p ). (7.23) yields (cid:107)f g(cid:107)M s,α p,q (cid:38)  (cid:88) 1(cid:28)j≤εJα 1−α (cid:88) j sq  (cid:88) (cid:16) s 1(cid:28)j≤εJα (cid:38) J nα 1−α (cid:38) J nα (1−α) +α (1−α) + nα 1−α (1− 1 (cid:107)F −1ψkjF (f g)(cid:107)q p k∈Aj 1−α (cid:88) j sq 1−α (1− 1 p ) j nαq (cid:17) . k∈Aj p )+ n q 1/q 1/q On the other hand, Similarly, (cid:107)f(cid:107)M s,α p,q ∼ J s (1−α)(cid:107)F −1χA(J)(cid:107)p ∼ J s (1−α) + nα 1−α (1− 1 p ). Hence, in order to M s,α p,q forms an algebra, one must has that (cid:107)g(cid:107)M s,α p,q ∼ J s (1−α) + nα 1−α (1− 1 p ). 2s (1 − α) + 2nα 1 − α (1 − 1 p ) ≥ nα (1 − α) + α (cid:19) . (1 − 1 p ) + n q (cid:18) s (cid:18) 1 (1 − α) + nα 1 − α (cid:19) − 1 . Namely, Step 2. s ≥ nα p + nα(1 − α) (2 − α) q (1/p, 1/q) ∈ [0, 1]2 ∩ D2. Let J (cid:29) 1. Put (cid:98)f (ξ) = χ[J 1/(1−α),3J 1/(1−α)]n(ξ), (cid:98)g(ξ) = χ[−3J 1/(1−α),−J 1/(1−α)]n(ξ). In view of (7.10) we have ((cid:98)f ∗(cid:98)g)(ξ) =  0, ξi ≥ 2J 1 (1−α) for some i = 1, ..., n, 1 (1−α) − ξi), (2J ξi ≤ 2J 1 (1−α) for all i = 1, ..., n. (7.22) n(cid:89) i=1 50 Hence, supp((cid:98)f ∗(cid:98)g) ⊂ {ξ : ξi ≤ 2J 1 (1−α) i = 1, ..., n}. Using the same way as in (7.23), we have for j ≤ εJ and k ∈ Aj, (cid:107)F −1ψkjF (f g)(cid:107)p (cid:38) J n 1−αj nα 1−α (1− 1 p ). (7.23) implies that (cid:107)f g(cid:107)M s,α p,q (cid:38) (cid:38) J On the other hand,  (cid:88) 1−α (cid:88)  (cid:88) 1(cid:28)j≤εJ j sq j sq k∈Aj n 1−α 1(cid:28)j≤εJ 1−α + nα 1−α (1− 1 n 1−α + s 1−α (cid:88) k∈Aj p )+ n q . (cid:38) J 1/q 1/q p (cid:107)F −1ψkjF (f g)(cid:107)q 1−α (1− 1 p ) j nαq (cid:88) 1−α (cid:88) j sq (cid:107)F −1ψkj(cid:98)f(cid:107)q p 1/q (cid:107)f(cid:107)M s,α p,q (cid:46) j∼J 1−α + nα s 1−α (1− 1 k∈Kj p )+ n q . (7.23) (7.24) (7.25) (7.26) (7.27) (cid:46) J Similarly, (cid:107)g(cid:107)M s,α p,q (cid:46) J s 1−α + nα 1−α (1− 1 p )+ n q . Hence, in order to M s,α p,q forms an algebra, one must has that (cid:18) (cid:19) . s ≥ nα p + n(1 − α) 1 − 1 q (cid:3) Acknowledgement. The first named author is supported in part by the National Sci- ence Foundation of China, grant 11026053. Part of the work was carried out while the sec- ond named author was visiting the Beijing International Center for Mathematical Research (BICMR), he would like to thank BICMR for its hospitality. References [1] L. Borup, M. Nielsen, Banach frames for multivariate α-modulation spaces, J. Math. Anal. Appl. 321 (2006), no. 2, 880 -- 895. [2] L. Borup, M. Nielsen, Boundedness for pseudodifferential operators on multivariate alpha-modulation spaces, Ark. Mat. 44 (2006) 241 -- 259. [3] L. Borup, M. Nielsen, Nonlinear approximation in α-modulation spaces. Math. Nachr. 279 (2006), no. 1-2, 101 -- 120. 51 [4] A. P. Calder´on, Intermediate spaces and interpolation, the complex method, Studia Math., 24 (1964), 113 -- 190. [5] A. P. Calder´on and A. Torchinsky, Parabolic maximal functions associated with a distribution, I, Ad- vances in Math., 16 (1975), 1 -- 64. [6] A. P. Calder´on and A. Torchinsky, Parabolic maximal functions associated with a distribution, II, Ad- vances in Math., 24 (1977), 101 -- 171. [7] S. Dahlke, M. Fornasier, H. Rauhut, G. Steidl, G. Teschke, Generalized coorbit theory, Banach frames, and the relation to α-modulation spaces, Proc. Lond. Math. Soc., 96 (2008), no. 2, 464 -- 506. [8] H. G. Feichtinger, Modulation spaces on locally compact Abelian group, Technical Report, University of Vienna, 1983. [9] H. G. Feichtinger, C. Y. Huang, B. X. Wang, Trace operators for modulation, α-modulation and Besov spaces, Appl. Comput. Harmon. Anal., 30 (2011), 110 -- 127. [10] M. Fornasier, Banach frames for α-modulation spaces, Appl. Comput. Harmon. Anal. 22 (2007), no. 2, 157 -- 175. [11] P. Grobner, Banachraume Glatter Funktionen and Zerlegungsmethoden, Doctoral thesis, University of Vienna, 1992. [12] K. Grochenig, Foundations of Time-Frequency Analysis, Birkhauser Boston, MA, 2001. [13] M. Kobayashi, M. Sugimoto, N. Tomita, On the L2-boundedness of pseudo-differential operators and their commutators with symbols in α-modulation spaces. J. Math. Anal. Appl. 350 (2009), no. 1, 157 -- 169. [14] M. Kobayashi, M. Sugimoto, N. Tomita, Trace ideals for pseudo-differential operators and their commu- tators with symbols in α-modulation spaces. J. Anal. Math. 107 (2009), 141 -- 160. [15] M. Sugimoto, N. Tomita, The dilation property of modulation space and their inclusion relation with Besov spaces, J. Funct. Anal., 248 (2007), 79 -- 106. [16] J. Toft, Continuity properties for modulation spaces, with applications to pseudo-differential calculus, I, J. Funct. Anal., 207 (2004), 399 -- 429. [17] J. Toft and P. Wahlberg, Embeddings of α-modulation spaces, Arxiv: 1110.2681. [18] H. Triebel, Theory of Function Spaces, Birkhauser-Verlag, Basel, 1983. [19] B. X. Wang and H. Hudzik, The global Cauchy problem for the NLS and NLKG with small rough data, J. Differential Equations, 232 (2007), 36 -- 73. [20] B. X. Wang and C. Y. Huang, Frequency-uniform decomposition method for the generalized BO, KdV and NLS equations, J. Differential Equations, 239 (2007), 213 -- 250. [21] B. X. Wang L. Zhao, B. Guo, Isometric decomposition operators, function spaces Eλ p,q and applications to nonlinear evolution equations. J. Funct. Anal. 233 (2006), no. 1, 1 -- 39. 52
1910.06891	2	1910	2019-10-18T12:13:26	Schur multipliers of Schatten--von Neumann classes $\boldsymbol{S_p}$	[ "math.FA", "math.CA", "math.CV", "math.SP" ]	We study in this paper properties of Schur multipliers of Schatten von Neumann classes $\boldsymbol{S}_p$. We prove that for $p\le1$, Schur multipliers of $\boldsymbol{S}_p$ are necessarily completely bounded. We also introduce for $p\le1$ a scale ${\mathscr W}_p$ of tensor products of $\ell^\infty$ and prove that matrices in ${\mathscr W}_p$ are Schur multipliers of $\boldsymbol{S}_p$. We compare this sufficient condition with the sufficient condition of membership in the $p$-tensor product of $\ell^\infty$ spaces.	math.FA	math	SCHUR MULTIPLIERS OF SCHATTEN -- VON NEUMANN CLASSES Sp A.B. ALEKSANDROV AND V.V. PELLER Abstract. We study in this paper properties of Schur multipliers of Schatten von Neumann classes Sp. We prove that for p ≤ 1, Schur multipliers of Sp are necessarily completely bounded. We also introduce for p ≤ 1 a scale Wp of tensor products of ℓ∞ and prove that matrices in Wp are Schur multipliers of Sp. We compare this sufficient condition with the sufficient condition of membership in the p-tensor product of ℓ∞ spaces. Contents Introduction 1. 2. Schur multipliers of Sp with respect to spectral measures and integral operators ℓ∞ ⊗c p 3. Completely bounded matrix Schur multipliers 4. Diagonal matrices in ℓ∞ ⊗p ℓ∞ 5. 6. Another scale of tensor products 7. Pisier's sufficient condition References ℓ∞ 6= Mp 1 5 7 8 10 13 18 20 1. Introduction In this paper we are going to study matrix Schur multipliers of Schatten -- von Neumann classes Sp and, more generally, Schur multipliers with respect to spectral measures. Recall that for matrices A = {ajk}j,k≥0 and B = {bjk}j,k≥0, their Schur -- Hadamard product A ⋆ B is defined by A ⋆ B def= {ajkbjk}j,k≥0. We are going to use the same notation A ⋆ B in the case when A = {ajk}j,k≥0 is a matrix with complex entries and B = {Bjk}j,k≥0 is a matrix with operator entries, i.e., the The research of the first author is partially supported by RFBR grant 17-01-00607. The publication was prepared with the support of the RUDN University Program 5-100. Corresponding author: V.V. Peller; email: [email protected]. 1 entries Bjk are bounded linear operators on a Hilbert space. In this case is an operator matrix. A ⋆ B def= {ajkBjk}j,k≥0 Definition. For p ∈ (0, ∞], we denote by Mp the space of matrix Schur multipliers of Sp, i.e., a matrix A = {ajk}j,k≥0, by definition, belongs to Mp if A ⋆ B ∈ Sp for every scalar matrix B in Sp. As usual, we say that a matrix belongs to Sp if it induces a linear operator on the sequence space ℓ2 of class Sp. Note that for convenience, we mean by S∞ the class of bounded matrices. The norm (p-norm for p < 1) kAkMp of a matrix Schur multiplier A is defined by kAkMp def= sup(cid:8)kA ⋆ BkSp : kBkSp ≤ 1(cid:9). (1.1) Recall that for p < 1, a functional k · k on a vector space X is called a p-norm if it has the following properties: kαxk = α · kxk, x ∈ X, α ∈ C; kxk = 0 only if kx + ykp ≤ kxkp + kykp, x = 0; x, y ∈ X. The p-norm induces the following metric on X: dist(x, y) = kx − ykp. X is called a p-Banach space if it is a complete space with respect to this metric. The fact that the right-hand side of (1.1) is finite follows easily from the closed graph theorem (it holds not only for Banach spaces but also for complete metric vector spaces, see [Ba], Ch. 1, § 3). It can easily be verified that for p < 1, the space Mp is a p-Banach space with respect to the p-norm k · kMp. It is well known (see, e.g., [Be]) that Mp = Mp′ for p ≥ 1, where p′ def= p/(p − 1). Definition. Let 0 < p ≤ ∞. A numerical matrix A is called a completely bounded Schur multiplier of Sp if A ⋆ B ∈ Sp for every operator matrix B in Sp. Obviously, a completely bounded Schur multiplier of Sp is necessarily a Schur multi- plier of Sp. The converse trivially holds for p = 2. It also holds for p = 1 and p = ∞, see [Be]. The famous problem of whether a Schur multiplier of Sp must be completely bounded for each p was posed in [Pis]. One of the main results of this paper is an affirmative answer to this question in the case p < 1. This will be established in § 3. The question for p ∈ (1, 2) ∪ (2, ∞) remains open. 2 To prove the main result of § 3, we obtain in § 2 a general result on double operator integrals that are transformers of Sp into itself. The notion of Schur multipliers can be generalized to the case of spectral measures on Hilbert space (this terminology was introduced in [Pe1]). To define Schur multipliers with respect to spectral measures, we are going to use double operator integrals. Double operator integrals are expressions of the form ZZX1×X2 Φ(x, y) dE1(x)Q dE2(y), where E1 and E2 are spectral measures, Q is a liner operator no Hilbert space and Φ is a bounded measurable function. Double operator integrals appeared first in the paper [DK]. Later Birman and Solomyak elaborated a beautiful theory of double operator integrals and found important applica- tions, see [BS1]. We refer the reader to recent surveys [Pe2] and [AP2] for definitions and basic properties of double operator integrals. Definition. The function Φ is called a Schur multiplier of Sp with respect to the spectral measures E1 and E2 if ZZ Φ(x, y) dE1(x)Q dE2(y) ∈ Sp whenever Q ∈ Sp. We use the notation Mp(E1, E2) for the class of Schur multipliers with respect to E1 and E2. Let us explain that this is a generalization of the notion of a matrix Schur multiplier. Indeed, let E1 and E2 be the spectral measures on ℓ2 defined by E1(∆)v = E2(∆)v (v, ej)ej, ∆ ⊂ Z+, v ∈ ℓ2, (1.2) def = Xj∈∆ where {ej}j≥0 is the standard orthonormal basis of ℓ2. It is easy to see that if Q is a linear operator on ℓ2 with matrix {qjk}j,k≥0 and Φ = {Φ(j, k)}j,k≥0 is a matrix with bounded entries, then the double operator integral ZZZ+×Z+ Φ(j, k) dE1(j)Q dE2(k) is a linear operator with matrix {Φ(j, k)qjk}j,k≥0. Therefore, Φ ∈ Mp(E1, E2) if and only if the matrix {Φ(j, k)}j,k≥0 ∈ Mp. To state a well-known sufficient condition for the membership in the space of Schur multipliers of Sp, p ≤ 1, we introduce the notion of p-tensor product of L∞ spaces. Definition. Let 0 < p ≤ 1 and let E1 and E2 be spectral measures defined on σ- E2 is, by definition, algebras of subsets of X1 and X2. The p-tensor product L∞ the class of functions Φ on X1 × X2 of the form E1 ⊗p L∞ Φ(x, y) =Xn≥0 ϕn(x)ψn(y), 3 (1.3) where the functions ϕn in L∞ E1 and ψn in L∞ Xn≥0 E2 satisfy 1/p L∞ kϕnkp L∞kψnkp < ∞. (1.4) By kΦkL∞⊗pL∞ we mean the infimum of the left-hand side of (1.4) over all representations of Φ in the form (1.3). Note that in the case p = 1, the space L∞ E1 ⊗1 L∞ E2 is the projective tensor product E2 of L∞ E1 and L∞ E2. L∞ E1b⊗L∞ It is well known and it is easy to see that L∞ E1 ⊗p L∞ kΦkMp(E1,E2) ≤ kΦkL∞⊗pL∞, Φ ∈ L∞ E2 ⊂ Mp(E1, E2) and E1 ⊗p L∞ E2. (1.5) Indeed, suppose that Φ is given by (1.3) and (1.4) holds and let Q ∈ Sp. We have (cid:13)(cid:13)(cid:13)(cid:13)ZZ Φ dE1Q dE2(cid:13)(cid:13)(cid:13)(cid:13) p Sp =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xn≥0(cid:18)Z ϕn dE1(cid:19) Q(cid:18)Z ψn dE2(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤Xn≥0 L∞ kψnkp L∞kQkp kϕnkp Sp . p Sp Let us consider the space L∞ E1 ⊗c p L∞ E2, which consists of functions Φ on X1 × X2, for which there exists a sequence of functions Φn in L∞ kΦnkL∞ E1 ⊗pL∞ E2 ≤ const E1 ⊗p L∞ E2 such that and Φn(x, y) → Φ(x, y) a.e. In particular, we can consider the p-tensor product ℓ∞ ⊗p ℓ∞, which corresponds to the case when E1 and E2 are the spectral measures on ℓ2 defined by (1.2). It is clear from the definition of (r, p, q)-nuclear operators (see [Pie], §18.1.1) that a matrix belongs to ℓ∞ ⊗p ℓ∞, 0 < p ≤ 1, if and only if it induces a (p, 1, 1)-nuclear operator from ℓ1 to ℓ∞. Note that (p, 1, 1)-nuclear operators and p-projective tensor products for p ∈ (0, 1] were investigated by Grothendieck [G], Chapter 2. In the case when E1 and E2 are the spectral measures on ℓ2 defined by (1.2), the space ℓ∞ ⊗c p ℓ∞ is the space of infinite matrices A such that kPnAPnkℓ∞⊗pℓ∞ ≤ const, where Pn is the standard natural projection on ℓ∞ onto span{ej : 0 ≤ j ≤ n}. It is easy to see that L∞ E1 ⊗c p L∞ E2 is still a subset of Mp(E1, E2), 0 < p ≤ 1. In particular, ℓ∞ ⊗c 1 L∞ It is also well known that L∞ It is well known that L∞ E1 ⊗c 0 < p ≤ 1. p ℓ∞ ⊂ Mp, E2 = M1(E1, E2), see [Pe1]. E1 ⊗1 L∞ E2 = L∞ E2 in general is a proper subset of M1(E1, E2). Indeed, consider the case when E1 and E2 are the spectral measures on ℓ2 defined by (1.2). Consider the identity matrix I, i.e., the infinite matrix with diagonal 1 ℓ∞. On entries equal to 1 and off diagonal entries equal to 0. Obviously, I ∈ M1 = ℓ∞ ⊗c E1b⊗L∞ (1.6) 4 the nuclear operators (we refer the reader to [Pie] for the definition of nuclear operators) the other hand, I 6∈ ℓ∞b⊗ℓ∞. Indeed, the space ℓ∞b⊗ℓ∞ can naturally be identified with from ℓ1 to ℓ∞, and so if I belonged to ℓ∞b⊗ℓ∞, it would induce a compact operator from ℓ1 to ℓ∞. However, it is not compact. We can consider the standard basis {ej}j≥0 in ℓ1. Clearly, there is no subsequence of {ej}j≥0 in ℓ1 that would converge in the norm of ℓ∞. However, we do not know whether for p < 1, E2 = L∞ L∞ E1 ⊗p L∞ It will be shown in § 4 that for p < 1, E1 ⊗c p L∞ E2. ℓ∞ ⊗c p ℓ∞ 6= Mp. This is a combination of an unpublished observation by Nigel Kalton and the description of the class of diagonal matrices in Mp obtained in [AP1]. In § 6 we introduce a scale Wp of tensor products of two ℓ∞ spaces and prove that the matrices in Wp are Schur multipliers of Sp for p < 1. We compare this sufficient condition with the sufficient condition in terms of the space ℓ∞ ⊗c p ℓ∞. We analyze in § 7.1 another sufficient condition for a matrix to be a Schur multiplier of Sp for p < 1. This sufficient condition follows from a result of Pisier [Pis2]. We compare this condition with the other sufficient conditions discussed in this paper. Let us also mention that in [AP1] we studied two special classes of matrix Schur multipliers: Toeplitz Schur multipliers and Hankel Schur multipliers. Throughout this paper we consider only separable Hilbert spaces. 2. Schur multipliers of Sp with respect to spectral measures and integral operators Suppose that (X1, B1) and (X2, B2) are measurable spaces and E1 and E2 are spectral measure on B1 and B2 that take values in the set of orthogonal projections on a Hilbert space H . Let µ1 be a positive measure that is mutually absolutely continuous with E1 and let µ2 be a positive measure that is mutually absolutely continuous with E2. For a function k in L2(µ1 ⊗µ2), we consider the integral operator Ik : L2(µ2) → L2(µ1) defined by (cid:0)Ikf(cid:1)(x) =ZX2 k(x, y)f (y) dµ2(y), f ∈ L2(µ2). (2.1) Theorem 2.1. Let 0 < p ≤ 1 and let Φ be a bounded measurable functions on X1×X2. The following are equivalent: (i) Φ ∈ Mp(E1, E2); (ii) if k is a function on X1 × X2 in L2(µ1 ⊗ µ2) such that Ik ∈ Sp, then IΦk ∈ Sp. In other words, statement (ii) of Theorem 2.1 means that ϕ is a multiplier of the space of kernel functions of integral operators of class Sp. Note that the case p = 1 is well-known, see [BS1], [Pe1]. We need an auxiliary fact known to experts. Let E be a spectral measure on a Hilbert space H and let u and v be vectors in H . Consider the positive measure σu and the 5 complex measure σu,v defined by σu(∆) = (E(∆)u, u) and the complex measure σu,v defined by σu(∆) = (E(∆)u, u), σu,v(∆) = (E(∆)u, v), where ∆ is a measurable set. Since σu,v(∆) ≤ kE(∆)uk · kvk = (E(∆)u, u)1/2kvk, it follows that σu,v is absolutely continuous with respect to σu. We denote by hv the Radon -- Nikodym density of σu,v, i.e., dσu,v = hvdσu. Lemma 2.2. For an arbitrary vector v in H , the function hv belongs to L2(σu). Moreover, (cid:8)hv : v ∈ H(cid:9) = L2(σu). Lemma 2.2 can easily be deduced from the theory of spectral multiplicities, see [BS2], Ch. 7. However, for the sake of convenience, we give a more elementary proof here. Proof. Let h ∈ L2(σu). Let A be the (not necessarily bounded) operator defined by A =R h dE whose domain D(A) is equal to D(A) =(cid:26)w ∈ H : Z h2 dσw < ∞(cid:27) , and so u ∈ D(A). Put v def= Au. It is easy to see that and so hv = h. Suppose now that v ∈ H . Let us show that hv ∈ L2(σu). Let g be a simple function, i.e., a linear combination of characteristic functions. Then (cid:0)E(∆)u, v(cid:1) =Z∆ h dσu, Z hvg dσu =(cid:18)(cid:18)Z g dE(cid:19) u, v(cid:19) (it suffices to verify this identity for characteristic functions g). Hence, (cid:12)(cid:12)(cid:12)(cid:12)Z hvg dσu(cid:12)(cid:12)(cid:12)(cid:12) ≤ kvk ·(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)Z g dE(cid:19) u(cid:13)(cid:13)(cid:13)(cid:13) ≤ kvk · kgkL2(σu). Since this inequality holds for an arbitrary simple function g, it follows that hv ∈ L2(σu). (cid:4) Remark. Before we proceed to the proof of Theorem 2.1, let us make the following observation. Suppose that µ♭ 1 is a positive measure that is mutually absolutely continuous with µ1 and µ♭ 2 is a positive measure that is mutually absolutely continuous with µ2. Then the class of multipliers of the space of kernel functions of integral operators from L2(µ♭ 1) of class Sp coincides with the class of multipliers the space of kernel functions of integral operators from L2(µ2) to L2(µ1) of class Sp. This can be seen very 2) → L2(µ2) easily by considering unitary operators U : L2(µ♭ defined by 1) → L2(µ1) and V : L2(µ♭ 2) to L2(µ♭ U f = h1/2 1 f, f ∈ L2(µ♭ 1), and 6 V f = h1/2 2 f, f ∈ L2(µ♭ 2), where dµ♭ 1 = h1 dµ1 and dµ♭ 2 = h2 dµ2. Proof of Theorem 2.1. Put Qu,v =ZZX1×X2 Φ(x, y) dE1(x)Tu,v dE2(y), where T is the rank one operator T = (·, u)v, u, v ∈ H . Since Φ is a bounded function, Q ∈ S2. Clearly, Φ is a Schur multiplier with respect to E1 and E2 is and only if Qu,v ∈ Sp and kQu,vkSp ≤ const kuk · kvk for arbitrary vectors u and v. It is easy to verify (see [BS1]) that where σw,u(∆2) (Qu,vw, z) =ZZ Φ(x, y) dσw,u(x) dλv,z(y), = (cid:0)E2(∆2)w, u(cid:1) and λv,z(∆1) def def = (cid:0)E1(∆1)v, z(cid:1), for arbitrary measurable subsets ∆1 and ∆2 of X1 and X2. Therefore, taking into account Lemma 2.2, we can conclude that statement (i) is equivalent to the fact that the integral operator IΦ(h⊗g) with kernel function (x, y) 7→ Φ(x, y)h(x)g(y) belongs to Sp for arbitrary g in L2(σv) and h in L2(σu) and (cid:13)(cid:13)IΦ(h⊗g)(cid:13)(cid:13)Sp ≤ const kgkL2(σv )khkL2(σu) for arbitrary g in L2(σv) and h in L2(σu). This exactly means that the function Φ is a multiplier of the space kernel functions of Sp integral operators from L2(σv) to L2(σu). To complete the proof, it remains to find vectors u and v such that σu is mutually absolutely continuous with E2 and σv is mutually absolutely continuous with E1. The existence of such vectors is well known, see e.g., [BS2], Ch. 7. (cid:4) 3. Completely bounded matrix Schur multipliers We prove in this section that for p ≤ 1, matrix Schur multipliers of Sp are necessarily completely bounded. Theorem 3.1. Let 0 < p ≤ 1. Suppose that a matrix A = {αjk}j,k≥0 is a Schur multiplier of Sp. Then A is a completely bounded Schur multiplier of Sp. Proof. Let K be a Hilbert space. Consider the spectral measure E on ℓ2(K ) defined by E(∆){un}n≥0 = {1I∆(n)un}n≥0 where ∆ is a subset of Z+ and 1I∆ is its characteristic function. Let ν be the counting measure on Z+, i.e., ν(∆) is the number of elements of ∆. Clearly, ν is mutually absolutely continuous with E. It is easy to see the condition that a matrix A = {αjk}j,k≥0 is a matrix Schur multiplier of Sp exactly means that it is a multiplier of the space of kernel functions of Sp integral operators on L2(ν). 7 On the other hand, it is easy to see that if T is an operator on ℓ2(K ) with block matrix {Tjk}j,k≥0, then ZZZ+×Z+ αjk dE(j)T dE(k) is the operator on ℓ2(K ) with block matrix {αjkTjk}j,k≥0. The result follow now from Theorem 2.1. (cid:4) To conclude this section we state the following problem. Problem. Can Theorem 2.1 be generalized to the case p > 1? Note that an affirmative answer to this question would imply that matrix Schur mul- tipliers of Sp with p > 1 must be completely bounded. 4. Diagonal matrices in ℓ∞ ⊗p ℓ∞ In § 1 for p ≤ 1, we defined the p-tensor product ℓ∞ ⊗p ℓ∞. In this section we are going to study diagonal matrices in ℓ∞ ⊗p ℓ∞. Recall that ℓ∞ ⊗p ℓ∞ consists of infinite matrices of the form Here and in what follows, for a matrix C, we use the notation C t for the transposed matrix. Following [AP1], we associate with p ∈ (0, 1] the number p♯, p♯ ∈ (0, ∞], defined by def= p♯ p 1 − p . The following theorem is a special case of Theorem 3.2 in [AP1]. Theorem 4.1. Let p ≤ 1 and let M be an infinite diagonal matrix with diagonal entries µj, j ≥ 0. Then M ∈ Mp if and only if {µj}j≥0 ∈ ℓp♯. Moreover, kM kMp = kµkℓp♯ = kM kSp♯ . Definition. Let p ∈ (0, 1]. Denote by Dp the space of sequences µ = {µj}j≥0 such that the diagonal matrix M with diagonal entries {µj}j≥0 belongs to ℓ∞ ⊗p ℓ∞. Put kµkDp def= kM kℓ∞⊗pℓ∞. Since ℓ∞ ⊗p ℓ∞ ⊂ Mp, it follows from Theorem 4.1 that Dp ⊂ ℓp♯. Let µ = {µj}j≥0 be a sequence whose terms tend to 0. Denote by µ∗ = {µ∗ j }j≥0 the nonincreasing rearrangement of the sequence {µj}j≥0. 8 where xn and yn are columns in ℓ∞ such that kxnkp ℓ∞kynkp ℓ∞ < ∞. xn ⊗ yn def = Xn≥0 xnyt n Xn≥0 Xn≥0 Let us remind the definition of the Lorentz space ℓq,r for r, q ∈ (0, ∞]: =  ℓq,r def µ ∈ c0 : kµkr ℓq,r (µ∗ j )r(1 + j) def = Xj≥0 ℓq,∞ def= (µ ∈ c0 : kµkℓq,∞ def= sup j≥0 (µ∗ j )(1 + j) r , r < ∞, q −1 < ∞ q < ∞) and ℓ∞,∞ def= ℓ∞. 1 Let Sq,r denote the corresponding operator ideal in the space of bounded operators on Hilbert space, i.e., where {sj(T )}j≥0 is the sequence of singular values of T . Sq,r =(cid:8)T : {sj(T )}j≥0 ∈ ℓq,r(cid:9), Theorem 4.2. Let 0 < p ≤ 1. Then ℓp♯,p ⊂ Dp, i.e., if Pj≥0 {µj}j≥0 ∈ Dp. Moreover, j )p (µ∗ (1+j)p < ∞, then kµkp Dp ≤ 2Xj≥0 j )p (µ∗ (1 + j)p < ∞. ek ⊗ ek. Then kJnkℓ∞⊗pℓ∞ = n1/p♯. e− 2πikj n ej, 0 ≤ k ≤ n − 1. k=0 Proof. Put Lemma 4.3. Let p ∈ (0, 1] and let Jn =Pn−1 n−1Xj=0 It can be easily verified that Jn = n−1Pn−1 n−1Xj=0 and yk = xk = n ej 2πikj e k=0 xk ⊗ yk. Hence, kJnkℓ∞⊗pℓ∞ ≤ n1/p♯. It remains to observe that kJnkMp ≤ kJnkℓ∞⊗pℓ∞ by (1.5) and kJnkMp = n1/p♯ by Theorem 4.1. (cid:4) Lemma 4.4. Let p ∈ (0, 1]. For matrices M = {µjk}j,k≥0 and A = {ajk}j,k≥0, the following inequality holds: where kAkℓ∞→ℓ∞ denote the operator norm on ℓ∞. kAM kℓ∞⊗pℓ∞ ≤ kAkℓ∞→ℓ∞kM kℓ∞⊗pℓ∞, Proof. Let M ∈ ℓ∞ ⊗p ℓ∞ and let ε > 0. Then M can be represented in the form where xn and yn are columns in ℓ∞ such that It follows that M =Xn≥0 xnyt n, Xn≥0 kxnkp ℓ∞ kynkp ℓ∞ < kM kp ℓ∞⊗pℓ∞ + ε. AM =Xn≥0 9 Axnyt n and kAM kp ℓ∞⊗pℓ∞ ≤Xn≥0 kAxnkp ℓ∞ kynkp ℓ∞ ≤ kAkp ℓ∞→ℓ∞(kM kp ℓ∞⊗pℓ∞ + ε). (cid:4) Corollary 4.5. Let 0 < p ≤ 1 and let A and M be infinite diagonal matrices with diagonal entries {αj}j≥0 and {µj}j≥0. Then kAM kℓ∞⊗pℓ∞ ≤(cid:16) sup j≥0 αj(cid:17)kM kℓ∞⊗pℓ∞. Remark. In the same way one can prove that kM Akℓ∞⊗pℓ∞ ≤ kAtkℓ∞→ℓ∞kM kℓ∞⊗pℓ∞ = kAkℓ∞→ℓ∞kM kℓ1⊗pℓ1, where kAkℓ1→ℓ1 denote the operator norms on ℓ1. Recall that kAkℓ1→ℓ1 = sup k Xj≥0 ajk and kAkℓ∞→ℓ∞ = sup j Xk≥0 ajk. Proof of Theorem 4.2. By Corollary 4.5, it suffices to consider the case when µj ≥ 0 for all j ≥ 0. Moreover, we may assume that {µj}≥0 is a decreasing sequence. Put Mn = M ⋆Jn, where Jn denotes the same as in Lemma 4.3. Lemma 4.3 and Corollary 4.5 imply that kM2n+1 − M2nkℓ∞→ℓ∞ ≤ 2n/p♯µ2n. It follows that ℓ∞⊗pℓ∞ ≤ µp ℓ∞⊗pℓ∞ ≤ kM1kp kM2n+1 − M2nkp 2pn/p♯µp 2n kM kp ℓ∞⊗pℓ∞ +Xn≥0 2n µp 2pn ≤ µp 2n = µp 0 +Xn≥0 0 + µp 1 + 2Xj≥1 0 +Xn≥0 (1 + j)p ≤ 2Xj≥0 µp j µp j (1 + j)p . (cid:4) 5. ℓ∞ ⊗c p ℓ∞ 6= Mp The main purpose of this section is to show that for p < 1, the space ℓ∞ ⊗c p ℓ∞ is a proper subset of Mp. We can naturally imbed the finite-dimensional space Cn to ℓ∞ and we are going to use the notation kxkℓ∞ for vectors x in Cn and kAkℓ∞⊗pℓ∞ for n × n matrices A. As usual, In stands for the n × n identity matrix. Lemma 5.1. Let p ∈ (0, 1). Suppose that n−1/p♯In =Xk≥1 λkuk ⊗ vk, where uk, vk ∈ Cn with kukkℓ∞, kvkkℓ∞ ≤ 1, {λk}k≥1 is a nonincreasing summable se- quence of nonnegative numbers. Let a and b be positive numbers such that a < 1 < b. Then k ≥ (1 − a) − bp−1kλkp λp ℓp . Xan<k≤bn 10 λkuk ⊗ vk is identical on a subspace of dimension at least (1 − a)n, λkuk ⊗ vk is at most an. Hence, the ≤ n1/p♯ Xk>bn λk(cid:17)1−p Xk>bn k>bn λk · kuk ⊗ vkkS1 λp k ≤ b−1/p♯nkλkℓp. λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p Sp . (cid:4) ≥ n(1 − a). and so Hence, p Sp 1 ≤ n Moreover, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xk>bn np Xan<k≤bn λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Sp Proof. Note that the rank of the operator Pk≤an operator n1/p♯ Pk>an n1−p(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xk>an λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)S1 ≤ n1/p♯(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xk>bn λk ≤ n1/p(cid:16) max p Xk>bn λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k ≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xan<k≤bn λkuk ⊗ vk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) −(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xk>bn ≥(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xk>an j=1 Aj with a matrix of size(cid:0)Pk j=1 nj(cid:1) ×(cid:0)Pk j=1 nj(cid:1). Lk Inj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mMj=1 ∞Xk=1 mMj=1 Lemma 5.2. Let p ∈ (0, 1) and let {nj}m ≥ np(1 − a) − bp−1npkλkp ℓp Let {Aj}k such that nj+1 ≥ 4(2m) 1−p nj. Then Proof. Let n −1/p♯ j λkuk ⊗ vk, n −1/p♯ j Inj = ≥ m 2 . p Sp p Sp p ℓ∞⊗pℓ∞ λp 1 j=1 be a sequence of matrices of size nj × nj. We can naturally identify j=1 be a finite sequence of positive integers where kukk∞, kvkk∞ ≤ 1 and {λk}k≥1 is a nonincreasing summable sequence of nonneg- ative numbers. We have to show that kλkp 4 and b = (2m) 2 . Applying Lemma 5.1 for a = 1 1−p , we get ℓp ≥ m 1 Xanj <k≤bnj λp k ≥ 3 4 − 1 2m kλkp ℓp 11 whenever 1 ≤ j ≤ m. Keeping in mind that bnj ≤ anj+1, we obtain kλkp ℓp ≥ Hence, kλkp ℓp ≥ m 2 . (cid:4) λp k ≥ ∞Xk=1 mXj=1  Xanj <k≤bnj λp k ≥ 3 4 m − 1 2 kλkp ℓp . Theorem 5.3. Let 0 < p < 1. Then Mp 6= ℓ∞ ⊗p ℓ∞. Moreover, there exists a diagonal matrix M such that M ∈ Mp but M /∈ ℓ∞ ⊗p ℓ∞. In other words, ℓp♯ 6= Dp. Proof. Assume that each diagonal matrix M in Mp belongs to the space ℓ∞ ⊗p ℓ∞. Then by the closed graph theorem, there exists a constant c such that kM kℓ∞⊗pℓ∞ ≤ ckM kMp. By Lemma 5.2, we have ≥ 2−1/pm1/p. (5.1) n −1/p♯ j = m1/p♯ (5.2) Inj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Sp♯ On the other hand, by Theorem 4.1, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mMj=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) mMj=1 n −1/p♯ j n −1/p♯ j Inj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)ℓ∞⊗pℓ∞ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Inj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Mp mMj=1 Inj(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Sp♯,r mMj=1 −1/p♯ j n (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and we get a contradiction. (cid:4) Remark. It is easy to see that in the notation of the proof of Theorem 5.3 we have ≤ const m1/r. It is clear from this remark that the proof of Theorem 5.3 allows us to get the following fact, which complements Theorem 4.2. Theorem 5.4. Let 0 < p < 1 and let r > p. Then ℓp♯,r 6⊂ Dp . Remark. As we have mentioned in § 1, the identity matrix does not belong to the space ℓ∞ ⊗1 ℓ∞. Hence, Theorem 5.3 remains true for p = 1. On the other hand, Theorem 5.4 cannot be generalized to the case p = 1 and r < ∞ because in this case ℓp♯,r ⊂ c0 = D1. For p ∈ (0, 1], we denote by Dc diagonal matrix M with diagonal entries {µj}j≥0 belongs to ℓ∞ ⊗c p the space of all sequences µ = {µj}j≥0 such that the p ⊂ ℓp♯. p ℓ∞. Clearly, Dc Theorem 5.3 allows us to prove the following fact. Theorem 5.5. Let p ∈ (0, 1). Then Dc p 6= ℓp♯. 12 Proof. Suppose that Dc p = ℓp♯. Then there exists a constant C(p) such that kµkDc C(p)kµkℓp♯ for all sequences µ = {µj}j≥0. Note that kµkDc sufficient large j. Thus, we get a contradiction with inequalities (5.2) and (5.1). (cid:4) p ≤ p = kµkDp if µj = 0 for all Theorem 5.5 immediately implies the main result of this section. Theorem 5.6. ℓ∞ ⊗c p ℓ∞ 6= Mp. To conclude the section, we state the following conjecture: Conjecture. Let p < 1. Suppose that M is a diagonal matrix. Then the following are equivalent: (i) M ∈ ℓ∞ ⊗p ℓ∞; (ii) M ∈ ℓ∞ ⊗c p ℓ∞; (iii) kM kSp♯,p < ∞. 6. Another scale of tensor products In this section we introduce a scale Wp, 0 < p ≤ 1, of tensor products of ℓ∞ spaces. The right endpoint of this scale coincides with the Haagerup tensor product of ℓ∞ spaces. We show that Wp ⊂ Mp. This together with (1.6) gives us two sufficient conditions for a matrix to belong to Mp. We show in this section that for p < 1, none of them implies the other one. For p ∈ (0, 1), we put Ξp X = {xjk}j,k≥0 such that def= ℓ2p♯(ℓ2p). In other words, Ξp is the set of matrices kXkΞp def= Xj≥0 For p = 1, put Ξ1 def= ℓ∞(ℓ2) and 1 2p♯ < ∞. p♯/p Xk≥0 xjk2p j≥0Xk≥0 xjk2 1 2 kXkΞ1 def = sup < ∞. Clearly, Ξp is a Banach space for p ∈ [1/2, 1], and Ξp is a (2p)-Banach space for Let p ∈ (0, 1]. Denote by Wp the set of matrices W representable in the form W = p ∈ (0, 1 2 ). XY t, where X, Y ∈ Ξp. Put kW kWp def= inf{kXkΞp kY kΞp : X, Y ∈ Ξp, W = XY t}. Let q ∈ (0, ∞). Clearly, that kx + ykq It is well known that W1 = M1 = M∞ and k · kW1 = k · kM1 = k · kM∞, see, e.g., [Pe1]. ℓq for every x, y ∈ ℓq such that x ⋆ y = 0 (here we consider vectors x and y as matrices). This remark implies the following fact. ℓq = kxkq ℓq + kykq 13 Lemma 6.1. Let X, Y ∈ ℓr(ℓq), where 0 < q ≤ r < ∞. Suppose that X ⋆ Y = 0. Then kX + Y kq ℓr(ℓq) ≤ kXkq ℓr (ℓq) + kY kq ℓr(ℓq). Proof. Let X = {xjk}j,k≥0 and Y = {yjk}j,k≥0. Put X (j) def = {xjk}k≥0 and Y (j) def = {yjk}k≥0. Then kX + Y kq kX (j) + Y (j)kr ℓr(ℓq) =Xj≥0 ≤Xj≥0 q/r ℓq +Xj≥0 =Xj≥0(cid:16)kX (j)kq ℓq = kXkq q/r q/r ℓq ℓq + kY (j)kq q/r ℓq(cid:17)r/q kX (j)kr kY (j)kr ℓr (ℓq) + kY kq ℓr(ℓq). (cid:4) Theorem 6.2. Let p ∈ (0, 1). Then Wp is a p-Banach space. ≤ kW kp Wp that W + R ∈ Wp and kW + Rkp Wp Proof. Let W = {wjk}j,k≥0 and R = {rjk}j,k≥0 be matrices in Wp. Let us prove . Let fix ε > 0. There exist xjlykl X, Y ∈ Ξp such that W = XY t and kW kWp > kXkΞp kY kΞp − ε. Clearly, wjk =Pl≥0 for all j, k ≥ 0. We can assume that kXkΞp = kY kΞp. Moreover, we can assume in addition that xjs = yks = 0 for all j, k ≥ 0 and all odd s ≥ 0. In a similar way we can represent R in the form R = U V t in such a way that kRkWp > kU kΞp kV kΞp − ε, ujlvkl for all j, k ≥ 0 and ujl = vkl = 0 for all j, k ≥ 0 and all + kRkp Wp kU kΞp = kV kΞp, rjk = Pl≥0 even l ≥ 0. Clearly, U ⋆ Y = 0 and X ⋆ V = 0. Hence, U Y t = 0 and XV t = 0, whence W + R = (X + U )(Y + V )t. Applying Lemma 6.1 for q = 2p and r = p♯ we obtain kW + Rk2p Wp ≤ kX + U k2p Ξp kY + V k2p Ξp + kU k2p Ξp + kU k2p Ξp ≤ (kXk2p + kV k2p ) Ξp Ξp )2 ≤ ((kW kWp + ε)p + (kRkWp + ε)p)2 )(kY k2p Ξp = (kXk2p Ξp Passing to the limit as ε → 0, we get kW + Rkp Wp ≤ kW kp Wp + kRkp Wp . It remains to prove that the space Wp is complete. It suffices to prove that the series Wn converges in Wp if kWnkWp < 4−n for all n ≥ 1. We can take two sequences {Xn}n≥0 and {Yn}n≥0 in Ξp such that Wn = XnY t n and kXnkΞp = kYnkΞp < 2−n. Moreover, it is easy to see that sequences {Xn}n≥0 and {Yn}n≥0 can be chosen in such Wn = Pn≥1 a way that in addition Xm ⋆ Yn = 0 for all m, n ≥ 0 such that m 6= n. Then Pn≥0 (cid:16) Pn≥0 . Since the space Ξp is complete, it follows that the series Pn≥0 and Pn≥0 Xn(cid:17)(cid:16) Pn≥0 Yn converge in Ξp. Hence, the series Pn≥0 Wn converges in Wp. (cid:4) Yn(cid:17)t Xn Theorem 6.3. Let p ∈ (0, 1]. Then Wp ⊂ Mp and k · kMp ≤ k · kWp . Theorem 6.3 can be reformulated in the following way: 14 Let p ∈ (0, 1) and let {ajn}j,n≥0 and {bnk}n,k≥0 be matrices such that the right-hand side of (6.1) is finite. If ϕjk =Xn≥0 ajnbkn, j, k ≥ 0, and Φ = {ϕjk}j,k≥0, then Φ ∈ Mp and kΦkMp ≤Xj≥0 Xn≥0 ajn2p p♯/p 1 2p♯ Xk≥0 Xn≥0 bkn2p p♯/p 1 2p♯ . (6.1) Proof. It suffices to verify that kXY tkMp ≤ kXkΞp kY kΞp for every X, Y ∈ Ξp. In other words we have to prove that kAXY tBkSp ≤ kXkΞp kY kΞpkAkS 2kBkS2 for every diagonal matrices A, B ∈ S2. Let {aj}∞ diagonal entries of A and B. Let X = {xjk}j,k≥0 and Y = {yjk}j,k≥0. We have j=0 and {bk}∞ k=0 be the sequences of 15 kAXY tBkp Sp p Sp 1/2 p Sp p/2 Xl≥0 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ajxjlyklbkj,k≥0 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤Xl≥0(cid:13)(cid:13)(cid:13)najxjlyklbkoj,k≥0(cid:13)(cid:13)(cid:13) Xj≥0 aj2xjl2 p/2Xk≥0 bk2ykl2 =Xl≥0 ≤Xl≥0 Xj≥0 aj2xjl2 p 1/2Xl≥0 Xk≥0 bk2ykl2 p ≤Xl≥0Xj≥0 aj2pxjl2p 1/2Xl≥0Xk≥0 bk2pykl2p =Xj≥0 xjl2p ykl2p(cid:17)1/2 (cid:16)Xk≥0 aj2pXl≥0 bk2pXl≥0 1−p 2Xj≥0 Xl≥0 xil2p aj2 ≤Xj≥0 1−p 2Xk≥0 Xl≥0 ykl2p ×Xk≥0 bk2 1/2 1/2 1−p 1−p p p 1 1 2 2 = kXkp ΞpkY kp ΞpkAkp S 2 kBkp S 2 . (cid:4) Theorem 6.4. Let p ∈ (0, 1] and let W be a diagonal matrix. Then kW kWp = kW kMp = kW kSp♯ . Proof. The inequality kW kWp ≥ kW kMp follows from Theorem 6.3. To prove the opposite inequality we may take the diagonal matrix X such that X 2 = W , i.e., W = XX t. Then kW kWp ≤ kXk2 = kW kMp by Theorem 4.1. (cid:4) Ξp = kXk2 = kW kSp♯ S 2p♯ Theorem 6.5. Let W = {wjk}j,k≥0 ∈ Wp with p ∈ (0, 1). Then there exist sequences α = {αj}j≥0 and β = {βk}k≥0 in ℓ2p♯ with nonnegative terms such that wjk ≤ αjβk for all j, k ≥ 0 and kαkℓ2p♯ kβkℓ2p♯ = kW kWp . 16 Proof. For each ε > 0, there exist matrices X = {xjk}j,k≥0 and Y = {yjk}j,k≥0 such that X, Y ∈ Ξp, kXk2 Ξp = kY k2 Ξp < kW kWp + ε and W = XY t. Put 1/2p . ykl2p < kW kWp + ε. 1/2 ≤ αj(ε)βk(ε). Then We have 1/2p 1/2p♯ αj(ε) =Xl≥0 xjl2p (αj(ε))2p♯ Xj≥0 wjk =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) xjlykl(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xl≥0 and βk(ε) =Xl≥0 (βk(ε))2p♯ =Xk≥0 ≤Xl≥0 xjl2 1/2Xl≥0 ykl2 1/2p♯ Since the sequences {αj(ε)}j≥0 and {βk(ε)}k≥0 are uniformly bounded in ℓ∞, one can select a sequence {ει}ι≥0 such that the sequences {αj(ει)}j≥0 and {βk(ει)}k≥0 converge in the weak-star topology of ℓ∞ to sequences α = {αj}j≥0 and β = {βk}k≥0. It is easy to verify that α and β satisfy the requirements. (cid:4) Corollary 6.6. Let p ∈ (0, 1). Then Wp ⊂ ℓ2p♯(Z2 +) and kW kℓ2p♯ ≤ kW kWp . Corollary 6.7. Let W = {wjk}j,k≥0 be a matrix in Wp, where p ∈ (0, 1]. Assume that wjk ≥ 1 for all (j, k) ∈ Λ, where Λ is a finite subset Z2 Λ denotes the number of elements of the set Λ. +. Then kW kWp ≥ Λ 1 2p♯ , where The following result shows that Theorem 6.5 is sharp. Theorem 6.8. Let p ∈ (0, 1) and let W = {wj}j≥0 be a matrix of rank one. Then W ∈ Wp if and only if W ∈ ℓ2p♯ . Moreover, kW kWp = kW kℓ2p♯ . Proof. Let wjk = ujvk, j, k ≥ 0, for sequences {uj}j≥0 and {vk}k≥0, and suppose that kW kℓ2p♯ = kukℓ2p♯ kukℓ2p♯ . Let X = {ujak}j,k≥0 and Y = {vjak}j,k≥0, where a0 = 1 and aj = 0 for all j ≥ 1. Clearly, X, Y ∈ Ξp, kXkΞp = kukℓ2p♯ and kY kΞp = kvkℓ2p♯ . We have XY t = {ujvk}j,k≥0. Hence, kW kWp ≤ kukℓ2p♯ kvkℓ2p♯ = kW kℓ2p♯ . The opposite inequality follows from Theorem 6.5. (cid:4) Theorem 6.9. Let p ∈ (0, 1). Then ℓ∞ ⊗p ℓ∞ 6⊂ Wp and Wp 6⊂ ℓ∞ ⊗c p ℓ∞. Proof. To prove the first statement, it suffices to apply Theorem 6.8 and observe that under the assumption rank W = 1, the matrix W belongs to ℓ∞ ⊗p ℓ∞ if and only if W ∈ ℓ∞(Z2 +). The second statement follows from Theorems 5.5 and 6.4. (cid:4) The next example shows that in a sense Corollary 6.7 is sharp. Example. Let W = {wjk}j,k≥0, where wjk = 1IS×T (j, k) for finite subsets S and T of 1 1 2p♯ . To prove Z+. Then kW kWp = S ×T the opposite inequality, we can consider the matrices X = {xjk}j,k≥1 and Y = {yjk}j,k≥1, 2p♯ . By Corollary 6.7, we have kW kWp ≥ S ×T 17 where xjk = 1IS×{1}(j, k) and yjk = 1IT ×{1}(j, k). It remains to observe that W = XY t and kXkΞpkY kΞp = S × T 2p♯ . Note that kW kMp = kW kℓ∞⊗pℓ∞ = 1 if S × T 6= ∅. 1 7. Pisier's sufficient condition In this section we analyze another sufficient condition for a matrix to be a Schur multiplier of Sp for p < 1. It is a consequence of a result of Pisier [Pis2]. We compare this sufficient condition with the sufficient condition given in § 6. For q ∈ (0, ∞], we define the space Yq of scalar matrices by and we put Yq def = ℓq(ℓ∞) +(cid:0)ℓq(ℓ∞)(cid:1)t kZkYq def = inf{kXkℓq (ℓ∞) + kY kℓq(ℓ∞) : Z = X + Y t}, q ∈ [1, +∞], kZkq Yq def= inf{kXkq ℓq (ℓ∞) + kY kq ℓq(ℓ∞) : Z = X + Y t}, q ∈ (0, 1). Clearly, Yq coincides with the set of matrices Z = {zjk}j,k≥0 such that zjk ≤ αj + βk, j, k ≥ 0, for some nonnegative sequences α = {αj }j≥0 and β = {βk}k≥0 in ℓq. We get the same space Yq if we require in addition that α = β. It is easy to see that Y∞ = ℓ∞(Z2) and k · kY∞ = k · kℓ∞(Z2). Let us also observe that if W is a self-adjoint matrix of rank one, then W ∈ Yq if and only if W ∈ ℓ2q(Z2). Indeed, suppose that W = {ujuk}j,k≥0 for a sequence {uj}j≥0. Let {αj}j≥0 be a nonnegative sequence. Then ujuk ≤ αj + αk for all j, k ≥ 0 if and only if αj ≥ 1 2 uj2 for all j ≥ 0. Recall that in [AP1], with a given number p in (0, 2], we associated the number def = p♭ 2p 2 − p . Let M(S2, Sp) be the space of matrices A = {ajk}j,k≥0 such that A ⋆ B ∈ Sp whenever B is a scalar matrix in S2. The following result was obtained by Pisier, see [Pis2], Theorem 5.1. for all p ∈ (0, 2]. We need only the easy part of Pisier's result, the inclusion M(S2, Sp) = Yp♭ Yp♭ ⊂ M(S2, Sp), 0 < p ≤ 2, (7.1) We give here a proof of this inclusion for the reader's convenience. Proof of (7.1). The result is trivial if p = 2. Suppose that p < 2. Clearly, it suffices to prove that kW ⋆ ZkSp ≤ kW kℓp♭ (ℓ∞)kZkS2 for all W ∈ ℓp♭(ℓ∞) and all Z ∈ S2. Let W = {wjk}j,k≥0 ∈ ℓp♭(ℓ∞). Put αj = supk≥0 wjk and α = {αj}j≥0. Then kW kℓp♭ (ℓ∞) = 18 kW ⋆ Zkp p 2 αp jXk≥0 Xk≥0 wjkzjk2 ≤Xj≥0 Sp ≤Xj≥0 j  ≤Xj≥0 2 Xj≥0Xk≥0 zjk2 2p 2−p 2−p α p 2 p 2 zjk2 = kW kp ℓp♭ (ℓ∞)kZkp S2 . (cid:4) kαkℓp♭ . Let Z = {zjk}j,k≥0 ∈ S2. Using the inequality kAkSp ≤ kAkℓp(ℓ2) for p ≤ 2, we obtain It is easy to see that (7.1) gives us a sufficient condition for a matrix to belong to Mp. Corollary 7.1. Let 0 < p ≤ 2. Then Yp♭ ⊂ Mp. Let us prove that this sufficient condition does not cover the sufficient condition Wp ⊂ Mp, nor it is covered by the condition Wp ⊂ Mp. Theorem 7.2. Let p ∈ (0, 1]. Then Wp 6⊂ Yp♭ and Yp♭ 6⊂ Wp. Proof. Clearly, kW kYq = kW kSq for any diagonal matrix W . This remark and Theorem 6.4 implies that Wp 6⊂ Yp♭ for p ∈ (0, 1]. Let us show that Yp♭ 6⊂ Wp. Let W = {wjk}j.k≥0 with wjk = αj 6= 0 for all j, k ≥ 0, where α ∈ ℓp♭. Then W ∈ Yp♭. However, W 6∈ Wp by Theorem 6.8. (cid:4) Theorem 7.3. Let p ∈ (0, 2]. Then Yq ⊂ Mp if and only if 0 < q ≤ p♭. Lemma 7.4. There exists an N × N matrix Z = {zjk}0≤j,k≤N −1 such that zjk = 1 and kZkMp = N 1 p − 1 2 for all p ∈ (0, 2]. Proof. It is well known that there exists a unitary N ×N matrix U = {ujk}0≤j,k≤N −1 such that ujk = N −1/2. Put Z = N 1/2U . Denote by IN be N × N matrix with all entries equal to 1. Then Z = Z ⋆ IN , whence kZkSp ≤ kZkMpkIN kSp. Clearly, kZkSp = N 2 . To prove the opposite inequality it suffices to observe that for every N × N matrix A, we have p and kIN kSp = N . Hence, kZkMp ≥ N 2 kU kSp = N p − 1 2 + 1 1 1 1 kZ ⋆ AkSp ≤ N 1 p − 1 2 kZ ⋆ AkS2 ≤ N 1 p − 1 2 kAkS 2 ≤ N 1 p − 1 2 kAkSp. (cid:4) Proof of Theorem 7.3. If q ≤ p♭, the Yq ⊂ Yp♭ ⊂ Mp. It remains to prove that if Yq ⊂ Mp, then q ≤ p♭. Suppose that Yq ⊂ Mp for some q > p♭. Then there exists a constant c > 0 such that kW kMp ≤ ckW kYq . Let us apply this inequality to the matrix W = {wjk}j,k≥0 such that def= ( zjk, 0, wjk if max(j, k) ≤ N − 1, if max(j, k) ≥ N, where Z = {zjk}0≤j,k≤N −1 is the matrix constructed in the proof of Lemma 7.4. Then 2 ≤ cN 1/q for all N ≥ 1, and so kW kMp = N p−1 ♭ ≤ q−1, i.e. q ≤ p♭. (cid:4) 2 and kW kYq ≤ N 1/q. Hence, N p − 1 p − 1 1 1 19 Remark 1. It is easy to see that ℓ∞ ⊗p ℓ∞ 6⊂ Yp♭ for p ∈ (0, 1]. Indeed, the infinite matrix I∞ with entries identically equal to 1 belongs to ℓ∞ ⊗p ℓ∞ and does not belong to Yp♭. In the case p ∈ (0, 1), the authors do not know whether Yp♭ ⊂ ℓ∞ ⊗p ℓ∞ or not. We conclude this section with a few observations. Let A = {ajk}j,k≥0 be a Schur multiplier of Sp. It is said to be an absolute Schur multiplier of Sp if every matrix B = {bjk}j,k≥0 satisfying bjk ≤ ajk, j, k ≥ 0, is also a Schur multiplier of Sp. Denote by (Mp)C the space of absolute Schur multipliers of Sp. We can use a similar same notation in a more general situation. Let X be linear subset of ℓ∞(Z2 +). Put XC def = {X ∈ X : X ⋆ ℓ∞(Z2 +) ⊂ X }. Clearly, A ∈ (Mp)C if and only if A is a Schur multiplier from Sp into (Sp)C. It is well known that (Sp)C = ℓp(ℓ2) +(cid:0)ℓp(ℓ2)(cid:1)t for p ∈ (0, 2], see [Pis2] in the case p < 1, the cases p ∈ (1, 2) and p = 1 were considered in [LP] and [LPP]. It is easy to see that for p ∈ (0, 2], every matrix A in M(S2, Sp) is an absolute Schur multiplier of Sp, i.e. M(S2, Sp) ⊂ (Mp)C for p ∈ (0, 2]. Hence, in Corollary 7.1 and Theorem 7.3 the space Mp can be replaced with the space (Mp)C. Remark 2. Let p ∈ (0, 2). Then M(S2, Sp) 6= (Mp)C. Indeed, it is easy to see that a diagonal matrix belongs to Mp if and only if it belongs to (Mp)C. It remains to observe that a diagonal matrix W belongs to M(S2, Sp) if and only if W ∈ Sp♭ and apply Theorem 4.1. This remark implies Yp♭ 6= (Mp)C for p ∈ (0, 2) which complements Corollary 7.1. Remark 3. Let p ∈ (0, 2). Lemma 7.4 allows us to prove that there exist a matrix A = {ajk}j,k≥0 and two sequences α = {αj}j≥0 and β = {βk}k≥0 in ℓ2p♯ with nonnegative terms such that ajk ≤ αjβk for all j, k ≥ 0 but A 6∈ Wp, see the proof of Theorem 7.3. This remark implies that Wp 6⊂ (Mp)C for p ∈ (0, 1). This is another way to see that Wp 6⊂ Yp♭ for p ∈ (0, 1), see Theorem 7.2. References [AP1] A B. Aleksandrov and V. V. Peller, Hankel and Toeplitz-Schur multipliers, Math. Ann., 324 (2002), no. 2, 277 -- 327. [AP2] A B. Aleksandrov and V. V. Peller, Operator Lipschitz functions, Uspekhi Matem. Nauk 71:4, 3 -- 106. [Ba] S. Banach, Th´eorie des op´erations lin´eaires (French), ´Editions Jacques Gabay, Sceaux, 1993. [Be] G. Bennett, Schur multipliers, Duke Math. J. 44 (1977), 603 -- 639. [BS1] M.S. Birman and M.Z. Solomyak, Double Stieltjes operator integrals, Problems of Math. Phys., Leningrad. Univ. 1 (1966), 33 -- 67 (Russian). English transl.: Topics Math. Physics 1 (1967), 25 -- 54, Consultants Bureau Plenum Publishing Corporation, New York. [BS2] M.S. Birman and M.Z. Solomyak, Spectral theory of selfadjoint operators in Hilbert space, Math- ematics and its Applications (Soviet Series), D. Reidel Publishing Co., Dordrecht, 1987. [DK] Yu.L. Daletskii and S.G. Krein, Integration and differentiation of functions of Hermitian opera- tors and application to the theory of perturbations (Russian), Trudy Sem. Functsion. Anal., Voronezh. Gos. Univ. 1 (1956), 81 -- 105. 20 [G] A. Grothendieck Produits tensoriels topologiques et espaces nucl´eaires, Mem. Amer. Math. Soc. 16, Providence 1955. [LP] F. Lust-Piquard, In´egalit´es de Khintchine dans Cp (1 < p < ∞), C. R. Acad. Sci. Paris S´er. I Math. 303 (7) (1986) 289292. [LPP] F. Lust-Piquard, G. Pisier, Noncommutative Khintchine and Paley inequalities, Ark. Mat. 29 (2) (1991) 241260. [Pe1] V.V. Peller, Hankel operators in the theory of perturbations of unitary and self-adjoint operators, Funktsional. Anal. i Prilozhen. 19:2 (1985), 37 -- 51 (Russian). English transl.: Funct. Anal. Appl. 19 (1985), 111 -- 123. [Pe2] V.V. Peller, Multiple operator integrals in perturbation theory, Bull. Math. Sci. 6 (2016), 15 -- 88. [Pie] A. Pietsch, Operator ideals Mathematische Monographien [Mathematical Monographs], 16. VEB Deutscher Verlag der Wissenschaften, Berlin, 1978. 451 pp. [Pis] G. Pisier, Non-commutative vector valued Lp-spaces and completely p-summing maps, Ast´erisque 247, 1998. [Pis2] G. Pisier, Remarks on the non-commutative Khintchine inequalities for 0 < p < 2, Funct. Anal. 256 (2009), no. 12, 4128 -- 4161. A.B. Aleksandrov St.Petersburg Branch Steklov Institute of Mathematics Fontanka 27, 191023 St.Petersburg Russia email: [email protected] V.V. Peller Department of Mathematics Michigan State University East Lansing, Michigan 48824 USA and Peoples' Friendship University of Russia (RUDN University) 6 Miklukho-Maklaya St., Moscow, 117198, Russian Federation email: [email protected] 21
1611.03339	2	1611	2016-12-05T14:11:44	Limits of some weighted Cesaro averages	[ "math.FA" ]	We investigate the existence of the limit of some high order weighted Cesaro averages.	math.FA	math	LIMITS OF SOME WEIGHTED CESARO AVERAGES VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU Abstract. We investigate the existence of the limit of some high order weighted Cesaro averages. Mathematics Subject Classification: 40G05, 40B05, 11B99. Key words: Sequences; High Order Cesaro Averages; Ergodic Averages; Multi-indices Sequences. 1. introduction Motivated by potential applications to several branches of the math- ematics, we study the possible convergence of high order weighted Ce- saro means of the type (1.1) 1 np n Xk=1 bkf (k/n) , where p > 0 and f : (0, 1] → R, provided (bk)k∈N ⊂ C is a p-mean convergent sequence: lim n 1 np n Xk=1 bk = b ∈ C . Averages like those in (1.1) naturally appear in Ergodic Theory. They also play a role in Probability, for example in the investigation of the central limit (see e.g. [10]), as well as in Infinite Dimensional Analysis in managing the so-called L´evy Laplacian (cf. [12]) and exotic, i.e. high order ones, see e.g. [3] and the references cited therein. Cesaro aver- ages as above might find natural applications also in Harmonic Anal- ysis, Linear Algebra and Matrix Theory, Numerical Analysis, Number Theory and in other sectors of pure and applied mathematics. The convergence of the mean in (1.1) depends on the conditions imposed on the function f , which are listed in our main result in Section 2. For example, we get convergence for the simple cases f (x) = xq , f (x) = (1 − x)q , q > 0 , Date: August 27, 2018. 1 2 VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU which leads to the results in Section 3 concerning averages of multi- indices sequences. The weighted averages of multi-indices sequences appear in manag- ing some quantum central limit theorems, when the sequence of mean covariances is not constant but at least convergent, and an order struc- ture on some indices affects the value of the so-called mixed moments. Indeed, Propositions 3.1 and 3.2, which quite surprisingly lead to re- sults which cannot be reflected, may be naturally exploited in Anti- Monotone and Monotone cases (see e.g [6, 7, 13]). In order to get a flavour of the several kinds of mixed moments naturally emerging in Quantum Probability and the associated problem of their computation, the reader is referred to [1, 2, 5, 8] and the references cited therein. The last section is devoted to counterexamples which explain that all the conditions imposed on our results are essentially optimal. We end by noticing that particular cases of averages considered here appear in Section 5 of [4] (see also [3]), where also several continuous versions of averages are investigated. 2. limits of weighted cesaro means In the present note we suppose that the set of natural numbers does not contain 0: N := {1, 2, . . . , n, . . .} . We start with some elementary notations by denoting for each func- tion f : (0, 1] → R, a sequence b := (bn)n∈N ⊂ C, and finally p ∈ (0, +∞), Mb,f ;p(n) := bkf (k/n) 1 np n Xk=1 some useful high order weighted Cesaro means. For any sequence b, by b we denote the sequence (bn)n∈N. A sequence b is said to be p-mean convergent if the sequence (Mb,1;p(n))n∈N of its Cesaro p-averages is convergent, where 1 stands for the constant function f = 1 identically. When p = 1, we recover the usual setting concerning the arithmetic means. It is easy to show that, if b is p-mean convergent then bn = o(np) for n → +∞. Let f : (0, 1] → R be a monotone function. Define on (0, 1] the possible infinite Borel measure df induced by the Stieltjes integral with respect to f if it is increasing, of by −f if f is decreasing, see e.g. [14], Section 12.3. The following result is useful in the sequel: WEIGHTED CESARO AVERAGES 3 Lemma 2.1. Let a and b be convergent and p-mean convergent se- quences with limn an = a and limn Mb,1;p(n) = b, respectively. Suppose that (2.1) Mb,1;p(n) ≤ B , n ∈ N , then the product sequence ab is p-mean convergent with lim n Mab,1;p(n) = ab . Proof. Fix ε > 0 and choose l0 such that n > l0 implies an − a < ε. We get for n > l0, +(cid:12)(cid:12)(cid:12)(cid:12) aMb,1;p(l0)(cid:12)(cid:12)(cid:12)(cid:12) (cid:19) + εB . Mab,1;p(n) − ab(cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:18)l0 (cid:12)(cid:12)(cid:12)(cid:12) Mab,1;p(l0)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:19)p(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) a(cid:18)Mb,1;p(n) − b(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Mab,1;p(n) − ab(cid:12)(cid:12)(cid:12)(cid:12) ≤ εB , (cid:12)(cid:12)(cid:12)(cid:12) lim sup n We then have which leads to the assertion being ε arbitrary. (cid:3) Here, there is our main result: Theorem 2.2. Fix a p-mean convergent sequence b with limn Mb,1;p(n) = b, and a monotone function f : (0, 1] → R such that f ∈ L1((0, 1], xp−1dx) and xp ∈ L1((0, 1],df). Then Mb,f ;p(n) = bpZ 1 0 lim n xp−1f (x)dx . Proof. We can suppose, without loosing generality, that f is decreasing by passing possibly to the opposite function, and positive by possibly adding a constant. Under the last hypotheses, for each ε > 0 there exists n0 such that, if n > n0 ] [ n n0 Xk=1 0 ≤ f(cid:18) k n(cid:19)(cid:20)(cid:18) k n(cid:19)p −(cid:18)k − 1 n (cid:19)p(cid:21) ≤ pZ 1/n0 0 xp−1f (x)dx ≤ ε , 4 VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU where [x] is the unique integer such that [x] ≤ x < [x] + 1 for any arbitrary real x. We then argue that xp−1f (x)dx − xp−1f (x)dx + xp−1f (x)dx − 0 0 0 ≤pZ 1 ≤pZ 1/n0 +(cid:26)pZ 1 ≤2ε +(cid:26)pZ 1 →2ε 1/n0 xp−1f (x)dx − 1/n0 ] n [ n n0 n(cid:19)(cid:20)(cid:18) k n(cid:19)p n (cid:19)p(cid:21) −(cid:18)k − 1 f(cid:18) k Xk=1 f(cid:18) k n (cid:19)p(cid:21) n(cid:19)(cid:20)(cid:18) k n(cid:19)p −(cid:18)k − 1 Xk=1 f(cid:18) k n(cid:19)(cid:20)(cid:18) k n(cid:19)p −(cid:18) k − 1 Xk=[ n f(cid:18) k n(cid:19)(cid:20)(cid:18) k n(cid:19)p ]+1 n0 n n Xk=[ n n0 ]+1 n (cid:19)p(cid:21)(cid:27) −(cid:18)k − 1 n (cid:19)p(cid:21)(cid:27) for n → +∞, since one recognises the last term as the Riemann- Stieltjes sum of Z 1 0 f (x)dxp = pZ 1 0 f (x)xp−1dx . As ε > 0 is arbitrary, we conclude that lim n n Xk=1 f(cid:18) k n(cid:19)(cid:20)(cid:18) k n(cid:19)p −(cid:18)k − 1 n (cid:19)p(cid:21) = pZ 1 0 xp−1f (x)dx . cn := Mb,1;p(n) − b , n ∈ N , (2.2) With we get Mb,f ;p(n) = cnf (1) + b n f(cid:18) k n(cid:19)(cid:20)(cid:18) k n(cid:19)p Xk=1 n(cid:19)(cid:21) . n (cid:19) − f(cid:18) k n (cid:19)p(cid:20)f(cid:18)k − 1 −(cid:18) k − 1 n (cid:19)p(cid:21) + n Xk=2 ck−1(cid:18)k − 1 WEIGHTED CESARO AVERAGES 5 For each ε > 0, let n0 such that n > n0 implies cn < ε. Then for every n sufficiently big, ≤ + n n0+1 ck−1(cid:18)k − 1 ck−1(cid:18)k − 1 ck−1(cid:18)k − 1 (cid:12)(cid:12)(cid:12)(cid:12) Xk=2 Xk=2 Xk=n0+2 n cnZ n0+1 n n(cid:19)(cid:21)(cid:12)(cid:12)(cid:12)(cid:12) n (cid:19) − f(cid:18) k n(cid:19)(cid:21) n (cid:19) − f(cid:18) k n(cid:19)(cid:21) n (cid:19) − f(cid:18) k n (cid:19)p(cid:20)f(cid:18)k − 1 n (cid:19)p(cid:20)f(cid:18)k − 1 n (cid:19)p(cid:20)f(cid:18)k − 1 xpdf (x) + εZ 1 xpdf (x) , 0 < sup n 0 which goes to 0 as n → +∞, because ε > 0 is arbitrary. Collecting the last computation with (2.2), we get the result. (cid:3) 3. some multi-dimensional cases The present section is devoted to the investigation of some ergodic limits of multi-dimensional Cesaro averages which may appear in the study of Quantum Central Limit Theorems as those considered in [7]. Proposition 3.1. Let b be a p-mean convergent sequence satisfying (2.1) with limn Mb,1;p(n) = b, and (ak1,...,km)k1,...,km∈N ⊂ C a multi- indices sequence such that for q > 0, lim n 1 nq X1≤k1,...,km≤n ak1,...,km = a . Then lim n 1 np+q n Xk=1 bk X1≤k1,...,km≤k ak1,...,km = abp p + q . Proof. Notice that where 1 np+q n Xk=1 bk X1≤k1,...,km≤k ak1,...,km = Mab,xq;p(n) , ak := 1 kq X1≤k1,...,km≤k ak1,...,km , k ∈ N , defines the sequence a which is supposed to be convergent. The proof now follows from Lemma 2.1 and Theorem 2.2. (cid:3) 6 VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU Recall that the Euler's Beta and Gamma functions are defined re- spectively as 0 ( ¯ z, t) := Z 1 Γ(z) := Z +∞ xz−1(1 − x)t−1dx , Re(z), Re(t) > 0 , z ∈ C\{0,−1,−2, . . .} . Such special functions are related by the celebrated identity xz−1e−xdx , 0 (3.1) see e.g. [9]. ( z, t) = ¯ Γ(z)Γ(t) Γ(z + t) , The functions above appear in the following result concerning the tail-average. Proposition 3.2. Let (ak1,...,km)k1,...,km∈N and b be a multi-indices se- quence and a sequence respectively, satisfying all the hypotheses of Proposition 3.1. If in addition, (3.2) ak1−h,...,km−h = ak1,...,km for any k1, . . . , km ∈ N and h < min{k1, . . . , km}, then (3.3) lim n 1 np+q n Xk=1 bk Xk+1≤k1,...,km≤n ak1,...,km = ab Γ(p + 1)Γ(q + 1) Γ(p + q + 1) . Proof. Notice that (3.2) gives Xk+1≤k1,...,km≤n ak1,...,km = X1≤k1,...,km≤n−k ak1,...,km and, consequently, 1 np+q 1 np+q a np+q = + n n Xk=1 Xk=1 Xk=1 n bk Xk+1≤k1,...,km≤n bk(n − k)q(cid:20) bk(n − k)q . ak1,...,km 1 (n − k)q X1≤k1,...,km≤n−k ak1,...,km − a(cid:21) WEIGHTED CESARO AVERAGES 7 From Proposition 2.2, one has lim n 1 np+q n Xk=1 bk(n − k)q = bpZ 1 0 xp−1(1 − x)qdx = b Γ(p + 1)Γ(q + 1) Γ(p + q + 1) , the last equality coming from (3.1) and Γ(z + 1) = zΓ(z). The thesis then follows once one shows 1 np+q n Xk=1 bk(n − k)q(cid:20) 1 (n − k)q X1≤k1,...,km≤n−k ak1,...,km − a(cid:21) is infinitesimal for n → ∞. Indeed, since for any ε > 0, there is l0 ∈ N such that for any h ≥ l0 1 hq X1≤k1,...,km≤h one has for each k = 1, . . . , n − l0, (cid:12)(cid:12)(cid:12)(cid:12) (n − k)q X1≤k1,...,km≤n−k 1 ak1,...,km − a(cid:12)(cid:12)(cid:12)(cid:12) ≤ ε, ak1,...,km − a(cid:12)(cid:12)(cid:12)(cid:12) ≤ ε. Thus, denoting by M > 0 a uniform bound for the sequence of the multiple of Cesaro means of (ak1,...,km), by (2.1) one finds n 1 (cid:12)(cid:12)(cid:12)(cid:12) np+q (cid:12)(cid:12)(cid:12)(cid:12) bk(n − k)q(cid:20) Xk=1 n (cid:19)q bk(cid:18)n − k Xk=1 +(cid:12)(cid:12)(cid:12)(cid:12) bk(cid:18)n − k Xk=n−l0+1 n(cid:19)q(cid:21)B . ≤(cid:20)ε + 2M(cid:18) l0 ε np 1 np ≤ n−l0 n 1 (n − k)q X1≤k1,...,km≤n−k ak1,...,km − a(cid:21)(cid:12)(cid:12)(cid:12)(cid:12) n (cid:19)q(cid:20) 1 (n − k)q X1≤k1,...,km≤n−k ak1,...,km − a(cid:21)(cid:12)(cid:12)(cid:12)(cid:12) The proof is achieved as ε is arbitrary. (cid:3) 4. some counterexamples We end the present note by showing some counterexamples concern- ing the average-convergence of sequences. We start by noticing that in Theorem 2.2, the case with b identically equal to 1 and p = 1 corresponds simply to ask whether the sequence 8 VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU of the Riemann sums of a L1-function f , made partitioning the interval [0, 1] in n subintervals of uniform length 1/n and Riemann integrable on all the subintervals [ε, 1], converges to the integral of f . The following simple counterexample (which can be easily modified to achieve the continuous case) f = +∞ Xn=1 n2χ{1/n} tells us that it is not always the case, even if one imposes mild natural conditions on f . Now we pass to see that the convergence of 1 imply that b is mean-convergent. Let b be the sequence defined as k=1 bk does not 23 n Pn { hn := 4n+1 − 1 . z −1, . . . ,−1 , . . . . } Define, for each integer n, b := 20 21 22 1, 1, 1, 1 , z } { z } { −1,−1 , z}{1 , mn := 2 · 4n − 1 , On one hand, it is easy to check that 1 n n Xk=1 bk = 1. On the other hand, for the subsequences indexed by mn and hn respec- tively, one finds Mb,1;1(mn) = and Mb,1;1(hn) = 1 m n(cid:18) n Xk=0 h n(cid:18) n Xk=0 1 22k − 1 2 22k − 1 2 n+1 Xk=1 n 22k(cid:19) = Xk=1 22k(cid:19) = − 1 3 , 1 3 . What follows is a simple counterexample for the general failure of Lemma 2.1 if condition (2.1) is not satisfied. Let b = (bk)k∈N and a = (ak)k∈N be defined as follows: b2n−1 := − a2n−1 := − √2n , 1 √2n , b2n := 1 + √2n , n ∈ N , , n ∈ N . 1 √2n a2n := Then a = limn an = 0, and b = limn Mb,1;1(n) = 1 n → +∞, first 2 . Furthermore, as 1 2n 2n Xk=1 bk = 1 2 + 1 n n Xk=1 √2k → +∞ , WEIGHTED CESARO AVERAGES 9 and second Mab,1;1(2n) = 1 + 1 2n n Xk=1 1 √2k → 1 > 0 = ab . Finally, one can wonder if (3.3) holds true under all the assumptions of Proposition 3.1 but (3.2). The answer is negative as the following example shows for the case p = 1, m = 2, and q = m. Indeed, take bk = 1 , ak1,k2 = (pk1 −pk1 − 1)pk2 , Then b = 1 and a = 2 3 as k, k1, k2 ∈ N . lim n = lim n 1 n2 X1≤k1,k2≤n √n n 1 n2 ak1,k2 = lim n Xk2=1pk2 = lim n 1 n 1 n2 X1≤k1,k2≤n r k2 Xk2=1 n n (pk1 −pk1 − 1)pk2 = Z 1 2 dx = 2 3 x . 1 0 Computing the left hand side of (3.3), we get ak1,k2 = lim n bk Xk+1≤k1,k2≤n (cid:18)1 −r k n(cid:19) n Xk2=k+1 rk2 n n n √k) 1 n3 (√n − Xk=1 = Z 1 dx(1 − √x)Z 1 Xk2=k+1pk2 dy√y x 0 n n 1 n3 Xk=1 Xk=1 ab 6= 1 n2 ab 3 . lim n = lim n = 4 15 Note added in proof The authors are grateful to O. Kouba who has drawn their attention to Theorem 1 in his note [11] while the present article was in press. The statement of such a theorem is the same as our Theorem 2.2, provided that the involved function f and the sequence (bn)n∈N are uniformly continuous on (0, 1] and positive, respectively. By using Weierstrass' Density Theorem as in [11], the former is a corollary of the latter, and can be extended to general p-mean convergent complex- valued sequences (bn)n∈N, provided that the sequence of their moduli (bn)n∈N satisfies (2.1). Acknowledgements The authors have been partially supported by Italian INDAM -- GNAMPA. They kindly acknowledge R. Peirone for some fruitful suggestions. 10 VITONOFRIO CRISMALE, FRANCESCO FIDALEO, AND YUN GANG LU References [1] Accardi L., Crismale V., Lu Y.G. Constructive universal central limit the- orems based on interacting Fock spaces, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 8 (2005), no. 4, 631650. [2] Accardi L., Hashimoto Y., Obata N. Notions of independence related to the free group, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 1 (1998), 201 -- 220. [3] Accardi L., Ji U. C., Saito K. The exotic (higher order L´evy) Laplacians gen- erate the Markov processes given by distribution derivatives of white noise, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 16 (2013), 1350020 (26 pages). [4] Accardi L., Ji U. C., Saito K. Higher order multi-dimensional extensions of Ces`aro theorem, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 18 (2015), 1550030 (14 pages). [5] Bozejko M., Speicher R. Completely positive maps on Coxeter groups, de- formed commutation relations and operator spaces, Math. Ann. 300 (1994), 97-120. [6] Crismale V., Fidaleo F., Lu Y.G. Ergodic theorems in quantum probability: an application to the monotone stochastic processes, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5), to appear, doi: 10.2422/2036-2145.201506 009, available at arXiv:1505.04688. [7] Crismale V., Fidaleo F., Lu Y.G. From discrete to continuous monotone C ∗-algebras via quantum central limit theorems, preprint (2016). [8] Crismale V., Lu Y.G. Rotation invariant interacting Fock spaces, Infin. Di- mens. Anal. Quantum Probab. Relat. Top. 10 (2007), 211 -- 235. [9] Erd´elyi A., Magnus W., Oberhettinger F., Tricomi F. G. Higher transcen- dental functions Vol. I, based on notes left by Harry Bateman (Reprint of the 1953 original). Robert E. Krieger Publishing Co., Inc., Fla., Melbourne 1981. [10] Feller W. An introduction to probability theory and its applications. Vol. II, second edition John Wiley & Sons, Inc., New York-London-Sydney 1971. [11] Kouba O. A generalization of Riemann sums, in Mathematical Reflections: two more years (2010-2011) (ed. Titu Andreescu), XYZ Press (2014), 379- 384, available at arXiv:1407.4679. [12] L´evy P. Le¸cons d'analyse fonctionnelle, Gauthier-Villars, Paris, 1922. [13] Muraki N. Monotonic independence, monotonic central limit theorem and monotonic law of small numbers, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 4 (2001), 39-58. [14] Royden H. L. Real analysis, third edition. Macmillan Publishing Company, New York 1988. WEIGHTED CESARO AVERAGES 11 Vitonofrio Crismale, Dipartimento di Matematica, Universit`a degli studi di Bari, Via E. Orabona, 4, 70125 Bari, Italy E-mail address: [email protected] Francesco Fidaleo, Dipartimento di Matematica, Universit`a degli studi di Roma Tor Vergata, Via della Ricerca Scientifica 1, Roma 00133, Italy E-mail address: [email protected] Yun Gang Lu, Dipartimento di Matematica, Universit`a degli studi di Bari, Via E. Orabona, 4, 70125 Bari, Italy E-mail address: [email protected]
1902.04545	1	1902	2019-02-12T18:43:54	Asymptotic expansion for the eigenvalues of a perturbed anharmonic oscillator	[ "math.FA", "math.AP" ]	In this article, we study the spectral properties of the perturbation of the generalized anharmonic oscillator. We consider a piecewise H\"older continuous perturbation and investigate how the H\"older constant can affect the eigenvalues. More precisely, we derive several first terms in the asymptotic expansion for the eigenvalues.	math.FA	math	ASYMPTOTIC EXPANSION FOR THE EIGENVALUES OF A PERTURBED ANHARMONIC OSCILLATOR KSENIA FEDOSOVA AND MEDET NURSULTANOV Abstract. In this article, we study the spectral properties of the perturba- tion of the generalized anharmonic oscillator. We consider a piecewise Hölder continuous perturbation and investigate how the Hölder constant can affect the eigenvalues. More precisely, we derive several first terms in the asymptotic expansion for the eigenvalues. 1. Introduction Spectral properties of the Sturm-Liouville operators have been studied for more than a century due to numerous applications in mathematics, mechanics, physics and other branches of natural sciences. One of the special cases is the anharmonic oscillator (AHO), which is defined as H ≡ − d2 dx2 + q(x), q(x) = m Xj=1 ajx2j , x ∈ R. where am > 0. To mention few of many applications, this class of operators provides an equivalent approach to λφ4-field theory [BW69], AHOs are used in vibrational spectroscopy as a model for diatomic molecules [Sat15], and besides, AHOs describe the thermal expansion of crystals [Vis57]. The most famous type of AHO is the harmonic oscillator (HO) that is q(x) = x2. Due to the importance in physics and being a simple and an elegant model, HO is well understood. In particular, its eigenvalues equal λn = 2n − 1, n ∈ N and the corresponding normalized eigenfunctions are explicitly expressed in terms of the Chebyshev-Hermite polynomials; see for instance [LS75]. Moreover, there are a lot of publications about perturbed HO, q(x) = x2 + V (x); we mention works concerning the spectral asymptotics[PS06, Akh08, Gur89, Kos70, KKP05, Sak81]. A HO perturbed by a smooth compactly supported perturbation V (x) is consid- ered in [PS06], where for the corresponding eigenvalues, a complete asymptotic expansion and trace formulas are obtained in terms of the heat invariants. In [KKP05], the authors study a perturbation of HO by a bounded complex function with bounded derivative and bounded indefinite integral. As a main result they ob- tain the asymptotics for the eigenvalues. The same question, but for a non-smooth 2010 Mathematics Subject Classification. Primary: 34E10, 34L20; Secondary: 81Q15, 34B24. Key words and phrases. anharmonic oscillator, harmonic oscillator, perturbation, asymptotic expansion, heat trace expansion. 1 2 K. FEDOSOVA AND M. NURSULTANOV perturbation is considered in [Akh08]. For a real-valued measurable perturbation V (x) with certain decay, the asymptotic formula for eigenvalues is established. In contrast to the HO, the model of AHO, in general settings, cannot be solved analytically, and thus one has to resort to approximation methods for its solution. Several approaches have been used for the numerical evaluation of the eigenvalue problem, see [GSS15] and references therein. The spectral properties of AHO were studied in [CDR18, HR82, Cam16, Fuc18, CR15]. In [HR82], the authors studied, in particular, the operators of the form −d2k/dx2k + x2l + p(x), for k, l ∈ N and with p(x) being a polynomial of degree less than 2l. They established the asymptotic formula which describes the behavior of the eigenvalues. In higher dimension, the Laplace operator, perturbed by a smooth radially symmetric polynomial potential on unbounded domain, is considered in [Fuc18]. The author obtains the asymptotic expansion of the heat kernel trace. In [MA79], the authors obtain the eigenvalue asymptotic of H in L2[0,∞) with the Dirichlet boundary condition and q(x) = xα + V (x), where α > 0 and V (x) is a real-valued, compactly supported, twice differentiable function on [0,∞). In most works concerning the spectral properties of perturbed AHO, the perturba- tions are smooth. For an actual real-world potential smoothness is not necessarily guaranteed. For this reason, we want to reduce the smoothness and explore how this will affect the eigenvalues, we require V (x) to be only piecewise Hölder-continuous. Theorem 1.1. Let H be the self-adjoint operator in L2(R), generated by (1) d2 dx2 + xα + V (x), − where α > 0, and V (x) is a bounded, real-valued, compactly supported, piecewise1 Hölder continuous function with an exponent τ > 0. Then the sequence of eigen- values {λn}∞ n=1 of H satisfies the following asymptotic formula2 − 2α α+2 λn = C 1 2α α+2 2α , C0 = V (s)ds, C2 = α − 1 12π(2 + α) cot(cid:16) π α(cid:17) C−1 1 , and Γ(·) is the gamma function. To the best knowledge of the authors, these types of perturbations of anharmonic oscillators have not been previously treated in the physics literature. There has, 1We mean that there is a finite number of pieces such that V is Hölder continuous on each piece. 2Note that the constants C0,C1, and C2 are defined slightly differently than in [MA79]. (2n − 1) − α+4 α+2 C0C 1 − α+4 α+2 C 1 1 4π − α+6 α+2 C2C 1 + (2) + + α + 2 2α α + 2 2α α + 2 where C1 = 4Γ(cid:0) 3 απΓ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) α(cid:1) 2 + 1 α+2 (2n − 1)− 2 (2n − 1)− 2 (2n − 1)− 4 π Z ∞ −∞ 1 V (s) cos(cid:16)2C −∞ α+2 Z ∞ α+2 + O(cid:0)n−1(cid:1) , − α 1 α+2 (2n − 1) α α+2 s(cid:17) ds 3 however, been research on perturbations of an harmonic oscillator by Gaussian noise: [BGJ18], [Git05]. Moreover, Hölder continuous potentials in Sturm-Liouville operators (that are not AHO) appear in [Ike60] and [Sch70]. Theorem 1.1 shows that the perturbation, V (x), does not affect the first term. However, it appears in the second term, while the regularity, the parameter τ , affects only the third term. Indeed, in case V (x) being smooth and compactly supported, the third term would decay rapidly. When V (x) is Hölder continuous with an exponent τ > 0, we can say only that the third term is O(n− ατ +2 α+2 ). In order to demonstrate more explicitly the effect of the smoothness, we construct an example. There we consider the operator H from Theorem 1.1 for α = 2 and V (x) being the Weierstrass function defined as in (53). Then we find the subsequence of the eigenvalues {λnk}∞ k=1 such that λnk = 2nk − 1 + n − 1 2 k 1 4√2Z π −π V (s)ds + n − 1+τ 2 k 2− 5+3τ 2 + O(n−1 k ). To some extent, the proof of Theorem 1.1 generalizes to more general potentials. However, the results are not so explicit: Theorem 1.2. Assume that {aj}N j=1 are sets of real numbers such that 0 < α1 < ... < αN and aN > 0. Let H be a self-adjoint operator in L2(R), generated by the expression j=1 and {αj}N (3) d2 dx2 + q(x), − q(x) = N Xj=1 ajxαj + V (x), where V (x) is a bounded, real-valued, piecewise Hölder continuous function with an exponent τ > 0, compactly supported in (−b, b) for some b > 0, which we consider to be fixed. Then the sequence of eigenvalues {λn}∞ 4√λn Z b (2n − 1) =Q(b, λn) + bpλn − q(b) − (q(s) − q(b))ds n=1 of H satisfies π 4 (4) −b 1 − where α := αN , 1 4√λn Z b −b V (s) cos(2pλns)ds + O(cid:16)λ − α+2 2α n (cid:17) + O(cid:0)λ−1 n (cid:1) . Q(x, λ) :=Z a(λ) x pλ − q(t)dt, and a(λ) is the turning point, that is q(a(λ)) = λ for sufficiently large λ > 0. We note that, in some cases, it is easy to express the function Q(b, λ) in terms {λs}, so that (4) can be written more explicitly. In the last section, we give some examples, including a quartic AHO; see also Remark 4.6. In Corollary 4.7, we consider the sequences of the eigenvalues of the operators in L2[0,∞) generated by (1) and the Dirichlet and Neumann boundary conditions. We derive their asymptotic formulas, which show that they are interlacing at infinity. 4 K. FEDOSOVA AND M. NURSULTANOV 1.1. Strategy and structure of the paper. The proof of the theorem uses the idea of [MA79] and goes as follows: for a sufficiently large b > 0, we construct solutions, f+(x, λ) and f−(x, λ), of Hy = λy in [0, b] and [−b, 0], respectively, satisfying the following boundary conditions f ′ +(0, λ) = f ′ f+(0, λ) = f−(0, λ) = cos φ(λ), −(0, λ) =pλ − q(b) sin φ(λ) for some φ(λ) ∈ [0, 2π). Then we construct (with a different method) a solution, y(x, λ), of Hy = λy, that is square-integrable on [b,∞). A square-integrable so- lution on (−∞,−b] can be obtained by a flip x 7→ −x, as the potential, q(x), is symmetric outside the support of the perturbation. Note that f+(x, λ) can be extended to an L2-solution of H on [0,∞), if and only if the vector (f+(b, λ), f ′ +(b, λ)) is linearly dependent with (y(b, λ), y′(b, λ)). The linear dependency at points +b and −b gives a system of two equations depending on λ and φ(λ). These two equations imply that as the spectral parameter, λ, goes to infinity, φ(λ) would be forced to tend either to 0, or to π 2 . Note that in the case of the symmetric perturbation, V (x), we would obtain either the equality φ = 0 or π/2 straight away, that would correspond to the case of either Dirichlet or Neumann boundary conditions at zero, or, that is the same, would force the solution to be even or odd. So, heuristically we can say that the condition of the potential, q(x), being symmetric at infinity turns out to be strong enough in order to get an asymptotical evenness or oddness of the eigenfunction, as the spectral parameter λ tends to infinity. We split these two cases and obtain an equality on λ, which holds asymptotically as λ tends to infinity, would allows us to obtain the asymptotics of the counting function and the asymptotic behavior of eigenvalues. It turns out that the aforementioned eigenvalues of "almost odd" and "almost even" (that is, corresponding to the cases φ is approximately 0 or π/2) eigenfunctions are interlacing. In order to be sure that the asymptotic equalities allow us to take care of all the eigenvalues of the problem, we give a rough estimate on the asymptotic expansion of the counting function of eigenvalues. 1.2. Acknowledgements. The first author is grateful to the supervisor of her master's thesis, Vladimir Podolskii, as the article was inspired by the aforemen- tioned master's thesis. The second author was partially supported by the Ministry of Education Science of the Republic of Kazakhstan under the grant AP05132071. We would like to thank Julie Rowlett for reading and commenting upon prelimi- nary version of this manuscript. We are also grateful to Grigori Rozenblum for the attention and useful comments and Simone Murro for helpful discussions. 2. Preliminaries 5 In this section, we first prove that H, defined as in (3), is a self-adjoint operator. Next, we construct the solutions of (5) N − y′′(x, λ) + Xj=1  ajxαj + V (x)  y(x, λ) = λy(x, λ) in [0, b] with certain boundary conditions at x = 0, and in [b,∞) under the condition that the solution is square-integrable. Above, b > 0 is such that the perturbation, V , is compactly supported in (−b, b). Finally, we study the asymptotic behavior of these solutions at a point b as the spectral parameter, λ, tends to infinity. Let {aj}N j=1 be sets of positive numbers and V (x) be a bounded, real valued function supported inside (−b, b). We also require V (x) to be a piecewise Hölder continuous function with an exponent τ > 0, that is there exist C > 0 and −∞ = x0 < x1... < xm = +∞, m ∈ N, such that j=1, {αj}N V (x) − V (y) ≤ Cx − yτ , for all x, y ∈ (xj, xj+1) and j = 0, ..., m. Consider the differential expression (3). Note that the operator of multiplication by V (x) is symmetric and bounded, whilst the operator associated with the non perturbed AHO in C∞ c (R) is essentially self- adjoint; see [LS75]. The Rellich-Kato theorem [Kat95, Theorem 4.4] implies that the operator defined by (3) in C∞ c (R) is essentially self-adjoint as well and hence has a unique self-adjoint extension, which we denote by H. 2.1. Construction of solutions outside the support of the perturbation. We start by considering (5) in [b,∞). Note that in this interval the perturbation, V (x), equals zero and hence (5) becomes (6) − y′′(x, λ) + N Xj=1 ajxαj y(x, λ) = λy(x, λ). Its solutions have already been studied in [MA79], so that we recall some of their results. Let y(x, λ) satisfy (6) and Q(x, λ) be the function defined as in Theorem 1.2. Define the function According to [MA79, Lemma 2], it follows that η(x, λ) := λ − q(x)1/4y(x, λ). η(b, λ) = d1(λ) sin(cid:16)Q(b, λ) + π 4(cid:17) − d2(λ) cos(cid:16)Q(b, λ) + π α (cid:17) , 4(cid:17) + O(cid:16)λ− α+2 (7) (8) η′(b,λ) = ∂ ∂x η(x, λ)(cid:12)(cid:12)(cid:12)(cid:12)x=b = −pµ(λ)hd1(λ) cos(cid:16)Q(b, λ) + π 4(cid:17) + d2(λ) sin(cid:16)Q(b, λ) + π 2α (cid:17) 4(cid:17)i + O(cid:16)λ− α+4 6 K. FEDOSOVA AND M. NURSULTANOV as λ → ∞, where µ(λ) := λ − q(b) and d1(λ), d2(λ) are given explicitly in terms of Airy functions. Moreover, by [MA79, (36)], as λ → ∞, (9) d1(λ) = 1 + O(λ− α+2 d2(λ) = O(λ− α+2 2α ). 2α ), 2.2. Construction of solutions inside the support of the perturbation. Next we study the differential equation (5) in [0, b] with the boundary conditions y(0) = c1 = c1(λ), y′(0) = c2 = c2(λ). Note its solutions are in one-to-one correspondence with the solutions of the fol- lowing integral equation:3 f+(x, λ) =c1 cospµ(λ)x + c2 pµ(λ)Z x + 1 0 sinpµ(λ)x pµ(λ) sin(pµ(λ)(x − s))[q(s) − q(b)]f+(s, λ)ds. (10) (11) This is a Volterra equation, and thus has a unique solution, f+(x, λ); see [Eva10, page 398]. Differentiating it, we obtain f ′ +(x, λ) = − c1pµ(λ) sinpµ(λ)x + c2 cospµ(λ)x cos(pµ(λ)(x − s))[q(s) − q(b)]f+(s, λ)ds. +Z x 0 We write (10) and (11) in the following way (12) 0 1 sin(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds# f+(b, λ) = cos(pµ(λ)b)"c1 − + sin(pµ(λ)b)" c2 pµ(λ) +(b, λ) = cos(pµ(λ)b)"c2 +Z b pµ(λ)Z b pµ(λ)Z b cos(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds# cos(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds# . and (13) f ′ + 1 0 0 + sin(pµ(λ)b)"−c1pµ(λ) +Z b 0 sin(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds# . Define k+ 0 1 (λ) :=Z b 2 (λ) :=Z b 0 k+ cos(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds, sin(pµ(λ)s)[q(s) − q(b)]f+(s, λ)ds. 3We have chosen the notation in such a way that f+ corresponds to a segment of a positive half-line, [0, b], and f− corresponds to a segment of a negative half-line, [−b, 0]. Hence (12) and (13) can be rewritten as (14) 7 k+ 2 (λ) f+(b, λ) = cos(pµ(λ)b) c1 − +(b, λ) = cos(pµ(λ)b)(cid:0)c2 + k+ pµ(λ)! + sin(pµ(λ)b) c2 pµ(λ) 1 (λ)(cid:1) + sin(pµ(λ)b)(cid:16)−c1pµ(λ) + k+ pµ(λ)! , 2 (λ)(cid:17) . k+ 1 (λ) f ′ + Next, we consider (5) in [−b, 0]. This is equivalent to considering a "reflected" equation (15) − y′′(x) + q(−x)y(x) = λy(x), x ∈ [0, b], with boundary conditions y(0) = c1 = c1(λ), y′(0) = −c2 = c2(λ). Similarly to (10), we construct the solution of f−(x, λ) = c1 cospµ(λ)x − c2 sinpµ(λ)x pµ(λ) and derive 0 1 + pµ(λ)Z x f−(b, λ) = cos(pµ(λ)b) c1 − −(b, λ) = cos(pµ(λ)b)(cid:0)−c2 + k− sin(pµ(λ)(x − s))[q(−s) − q(−b)]f−(s, λ)ds pµ(λ)! + sin(pµ(λ)b) − 1 (λ)(cid:1) + sin(pµ(λ)b)(cid:16)−c1pµ(λ) + k− pµ(λ) k− 2 (λ) f ′ c2 + where k− 1 (λ) pµ(λ)! , 2 (λ)(cid:17) , k− 0 1 (λ) :=Z b 2 (λ) :=Z b 0 k− cos(pµ(λ)s)[q(−s) − q(−b)]f−(s, λ)ds, sin(pµ(λ)s)[q(−s) − q(−b)]f−(s, λ)ds. In the following two lemmas we investigate the asymptotic behavior of the functions f±(s, λ), k± Lemma 2.1. The solutions f±(s, λ) satisfy, as λ → +∞, 2 (λ) as λ → +∞. 1 (λ) and k± f±(x, λ) = c1(cospµ(λ)x) ± c2 Proof. Denote by Hλ the operator sin(pµ(λ)x) pµ(λ) + c1O(cid:16)λ− 1 2(cid:17) + c2O(cid:0)λ−1(cid:1) . Hλ : u 7→ µ(λ)−1/2Z x 0 in C[0, b]. Then [q(s) − q(b)] sin(pµ(λ)(x − s))u(s)ds (16) f+(x, λ) = c1 cos(pµ(λ)x) + c2 sin(pµ(λ)x) pµ(λ) + (Hλf+)(x). 8 K. FEDOSOVA AND M. NURSULTANOV By the triangle inequality (17) kf+(·, λ)k∞ ≤ c1 + c2 pµ(λ) + kHλkop · kf+(·, λ)k∞, where k·k∞ is the uniform norm and k·kop is the operator norm. By [MA79, (30)], kHλkop = O(λ− 1 2 ) as λ → ∞, hence for sufficiently large λ, 1 − kHλkop > 0 and therefore (18) kf+(·, λ)k∞ ≤ Moreover, (18) implies that c1 + c2√µ(λ) 1 − kHλkop . kf+(·, λ)k∞ = c1O(1) + which together with (16) gives O(1), λ → ∞, c2 pµ(λ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) f+(·, λ) − c1 cos(pµ(λ)(·)) − c2 = c1O(cid:16)λ− 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ sin(pµ(λ)(·)) pµ(λ) 2(cid:17) + c2O(cid:0)λ−1(cid:1) . ≤ kHλkop · f+(·, λ)∞ (cid:3) Next we derive asymptotic formulas for the functions k± 1 (λ) and k± 2 (λ). 1 (λ) and k± 2 (λ) satisfy the following asymptotic for- O(cid:0)λ− τ 2(cid:1) , O(cid:0)λ− τ 2(cid:1) . c2 2(cid:17) + c2O(cid:0)λ−1(cid:1) . Lemma 2.2. The functions k± mulas, as λ → ∞, k± (19) 1 (λ) = c1 Proof. Lemma 2.1 implies that 0 0 k+ 1 (λ) = c1 0 0 c2 [q(s) − q(b)]ds + (q(±s) − q(±b)) ds + c1O(cid:0)λ− τ 2(cid:1) + pµ(λ) (q(±s) − q(±b)) ds + c1O(cid:0)λ− τ 2(cid:1) + pµ(λ) 2 Z b 2pµ(λ)Z b 2 (λ) = ±c2 k± 2 Z b 2pµ(λ)Z b Z b [q(s) − q(b)] cos(2pµ(λ)s)ds = O(cid:0)λ− τ 2(cid:1) , Z b [q(s) − q(b)] sin(2pµ(λ)s)ds = O(cid:0)λ− τ 2(cid:1) [q(s) − q(b)] sin(2pµ(λ)s)ds + c1O(cid:16)λ− 1 [q(s) − q(b)] cos(2pµ(λ)s)ds c2 c1 2 Z b 0 0 0 + (20) (21) (22) Since q(x) is a piecewise Hölder continuous function with an exponent τ , by [SR03, page 92], we obtain as λ → ∞. Combining (22) with (21), we derive (19). Similarly, one proves (20). (cid:3) We also will need sharper asymptotics for k+ (23) 1 + k− 1 and k+ 2 − k− 2 . By (21), we derive 9 k+ 1 (λ) + k− 1 (λ) = Similarly, we obtain (24) k+ 2 (λ) − k− 2 (λ) = −b c2 c1 2 Z b 2pµ(λ)Z b 2pµ(λ)Z b 2pµ(λ)Z b c2 c2 −b + + c1 2 Z b (q(s) − q(b))ds + (q(s) − q(b) sin(2pµ(λ)s)ds + c1O(cid:16)λ− 1 (q(s) − q(b)) cos(2pµ(λ)s)ds 2(cid:17) + c2O(cid:0)λ−1(cid:1) . −b −b c1 2 Z b −b (q(s) − q(b))ds + (q(s) − q(b)) cos(2pµ(λ)s)ds + c1O(cid:16)λ− 1 (q(s) − q(b)) sin(2pµ(λ)s)ds 2(cid:17) + c2O(cid:0)λ−1(cid:1) . −b 3. Proof of Theorems 1.1 and 1.2 In this section we use the notations introduced in the previous sections. Moreover, we skip writing arguments for the following functions, where j = 1, 2, Q := Q(b, λ), µ := µ(λ), dj := dj(λ), k± j := k± j (λ). We start by proving Theorem 1.2. Proof. For fixed c1, c2 ∈ R, the solutions f+(x, λ) and y(x, λ) are linearly dependent, provided that λ is an eigenvalue of H in L2(R). This gives us the following equation for the eigenvalues or f ′ +(b, λ)y(b, λ) − f+(b, λ)y′(b, λ) = 0, f ′ +(b, λ)η(b, λ) − f+(b, λ)η′(b, λ) = (4µ)−1q′(b)η(b, λ)f+(b, λ). By inserting (7), (8) and (14) into the previous equation and using Lemma 2.1, we obtain π π π h cos(√µb)(cid:0)c2 + k+ ×hd1 sin(cid:16)Q + +h cos(√µb)(cid:0)c1√µ − k+ ×hd1 cos(cid:16)Q + = cos(√µb) sin(cid:16)Q + + sin(√µb) cos(cid:16)Q + + cos(√µb) cos(cid:16)Q + + sin(√µb) sin(cid:16)Q + 1(cid:1) + sin(√µb)(cid:0)−c1√µ + k+ 2(cid:1)i 4(cid:17) − d2 cos(cid:16)Q + 4(cid:17) + O(cid:16)λ− α+2 α (cid:17)i 2(cid:1) + sin(√µb)(cid:0)c2 + k+ 1(cid:1)i α )i 4(cid:17) + O(λ− α+2 4(cid:17) + d2 sin(cid:16)Q + 1(cid:1) + d2(cid:0)c1√µ − k+ 2(cid:1)(cid:17) 4(cid:17)(cid:16)d1(cid:0)c2 + k+ 1(cid:1) + d2(cid:0)c1√µ − k+ 2(cid:1)(cid:17) 4(cid:17)(cid:16)d1(cid:0)c2 + k+ 4(cid:17)(cid:16)d1(cid:0)c1√µ − k+ 1(cid:1)(cid:17) 2(cid:1) − d2(cid:0)c2 + k+ 4(cid:17)(cid:16) − d1(cid:0)c1√µ − k+ 1(cid:1)(cid:17) 2(cid:1) + d2(cid:0)c2 + k+ π π π π π 10 K. FEDOSOVA AND M. NURSULTANOV π 4 π 4 α (cid:17) + c2O(cid:16)λ− α+2 α (cid:17) + c1√µO(cid:16)λ− α+2 = c1O(cid:0)λ−1(cid:1) + c2O(cid:16)λ− 3 2(cid:17) . + √µb(cid:17)(cid:16)d1(cid:0)c2 + k+ sin(cid:16)Q + + √µb(cid:17)(cid:16)d1(cid:0)c1√µ − k+ + cos(cid:16)Q + = c1√µO(cid:16)λ− α+2 α (cid:17) + c2O(cid:16)λ− α+2 + √µb(cid:17)(cid:16)d1(cid:0)−c2 + k− + √µb(cid:17)(cid:16)d1(cid:0)c1√µ − k− α (cid:17) + c2O(cid:16)λ− α+2 sin(cid:16)Q + + cos(cid:16)Q + = c1√µO(cid:16)λ− α+2 π 4 π 4 1(cid:1) + d2(cid:0)c1√µ − k+ 2(cid:1)(cid:17) 1(cid:1)(cid:17) 2(cid:1) − d2(cid:0)c2 + k+ α (cid:17) + c1O(cid:0)λ−1(cid:1) + c2O(cid:16)λ− 3 2(cid:17) . 1 (cid:1) + d2(cid:0)c1√µ − k− 2 (cid:1)(cid:17) 1 (cid:1)(cid:17) 2 (cid:1) − d2(cid:0)−c2 + k− α (cid:17) + c1O(cid:0)λ−1(cid:1) + c2O(cid:16)λ− 3 2(cid:17) . By the same arguments, but for the solutions f−(−x, λ) and y(−x, λ), we obtain Now we have two equations (25) and (26) and variables λ, c1 = c1(λ) and c2 = c2(λ). Without lost of generality, we assume that the eigenfunction, corresponding to λ, is normalized in the sence that c1 = cos φ(λ) c2 = √µ sin φ(λ) for some φ(λ) ∈ [0, 2π). Subtracting (26) from (25) and adding (26) to (25) give Hence (25) (26) (27) and (28) π 4 π 4 sin(cid:16)Q + − cos(cid:16)Q + = O(cid:16)λ− α+4 sin(cid:16)Q + + cos(cid:16)Q + = O(cid:16)λ− α+4 + √µb(cid:17)(cid:16)2d1c2 + d1(cid:0)k+ 1 − k− + √µb(cid:17)(cid:16)2d2c2 + d1(cid:0)k+ 2α (cid:17) + O(cid:0)λ−1(cid:1) + √µb(cid:17)(cid:16)2d2√µc1 + d1(cid:0)k+ + √µb(cid:17)(cid:16)2d1√µc1 − d1(cid:0)k+ 2α (cid:17) + O(cid:0)λ−1(cid:1) . π 4 π 4 2(cid:1)(cid:17) 1 (cid:1) − d2(cid:0)k+ 2 − k− 1 (cid:1)(cid:17) 2 − k− 2 (cid:1) + d2(cid:0)k+ 1 − k− 1 + k− 1 (cid:1) − d2(cid:0)k+ 2 (cid:1) − d2(cid:0)k+ 2 + k− 2 + k− 2 (cid:1)(cid:17) 1 (cid:1)(cid:17) 1 + k− We will distinguish two types of solutions, ΛD and ΛN , of the system of equations (27)-(28): we say that solution λ ∈ ΛD if sin φ(λ) ≥ cos φ(λ) and λ ∈ ΛN otherwise. Let {λk}∞ k=1 be the increasing sequences of solutions of types ΛD and ΛN , respectively. We consider two cases: 1) λ ∈ ΛD; 2) λ ∈ ΛN . Case 1. Assume λ ∈ ΛD, so that sin φ(λ) ≥ 1/√2. Therefore (27), together with k=1 and {λk}∞ Lemma 2.2 and (9), implies sin(cid:16)Q + π 4 + √µb(cid:17)√µ sin φ(λ) + O(1) = O(λ− α+4 2α ) + O(λ−1). 11 Since sin φ(λ) ≥ 1/√2, (29) Therefore, solutions {λk}∞ (30) Q(b, λn) + sin(cid:16)Q + π 4 + √µb(cid:17) = O(cid:16)λ− 1 2(cid:17) . k=1 of type ΛD satisfy π 4 + bqλn − q(b) = nπ + β(λn), for some function β(λ) tending to zero as λ → ∞. To investigate (30) further, we need a better estimate on β(λ). Combining (27) and (30) we obtain, for λ = λn, 2 (cid:1)(cid:17) 1 (cid:1) − d2(cid:0)k+ 2 − k− 1 (cid:1)(cid:17) 2 − k− 2 (cid:1) + d2(cid:0)k+ 1 − k− (31) (32) Then, for λ = λn, Lemma 2.2 and (9) imply that sin β(λ)(cid:16)2d1√µ sin φ(λ) + d1(cid:0)k+ 1 − k− − cos β(λ)(cid:16)2d2√µ sin φ(λ) + d1(cid:0)k+ 2α (cid:17) + O(cid:0)λ−1(cid:1) . = O(cid:16)λ− α+4 sin β(λ)(cid:16)2√µ sin φ(λ) + k+ − cos β(λ)(cid:0)k+ 2 − k− 1 − k− 1 + O(λ− 1 α )(cid:17) 2 + 2d2√µ sin φ(λ)(cid:1) = O(cid:0)λ−1(cid:1) + O(cid:16)λ− α+2 2α (cid:17) . Now, let us investigate (28). By using (9), (29) and Lemma 2.2, we conclude, from (28), that √µc1 = O(1), so that (33) c1 = cos φ(λ) = O(λ− 1 2 ), and consequently k± 1 = O(λ− 1 2 ). Therefore (32) gives sin β(λ) = (34) and hence (35) β(λ) = 2 ) + O(λ− 1 α )(cid:16)k+ cos β(λ) 2√µ sin φ(λ) + O(λ− 1 + O(cid:16)λ− 2α+2 2α (cid:17) + O(cid:16)λ− 3 2(cid:17) . 2 ) + d2 + O(cid:16)λ− 2α+2 (k+ 2 − k− 2√µ sin φ(λ) 1 Thus, from (24), it follows 2 − k− 2 + 2d2√µ sin φ(λ)(cid:17) 2α (cid:17) + O(cid:16)λ− 3 2(cid:17) . β(λ) = (36) (q(s) − q(b))ds + 1 −b 4√µZ b 4√µ sin φ(λ)Z b cos φ(λ) + −b so that the Riemann-Lebesgue lemma and (33) imply 1 −b 4√µZ b (q(s) − q(b)) cos(2√µs)ds (q(s) − q(b)) sin(2√µs)ds + d2 + O(cid:0)λ−1(cid:1) , 4√µZ b V (s) cos(2√µs)ds + d2 + O(cid:0)λ−1(cid:1) −b 1 β(λ) = 1 4√µZ b −b (q(s) − q(b))ds + 12 K. FEDOSOVA AND M. NURSULTANOV and consequently (30) gives nπ = Q(b, λn) + (37) − 1 4pλn Z b −b 1 π 4 + bqλn − q(b) − V (s) cos(2qλns)ds − d2(λn) + O(cid:16)λ−1 n (cid:17) . 4pλn Z b −b (q(s) − q(b))ds Case 2. Assume λ ∈ ΛN , so that cos φ(λ) ≥ 1/√2. Therefore (28), together with Lemma 2.2 and (9), implies π 4 (38) cos(cid:16)Q + Therefore, the solutions {λk}∞ k=1 of type ΛN satisfy π (39) 4 Q(b, λn) + + bqλn − q(b) = nπ − + √µb(cid:17) = O(cid:16)λ− 1 2(cid:17) . π 2 + γ(λn) By combining (28), (39) and then using Lemma 2.2 together with (9), for λ = λn, we obtain (40) 1 + k− cos γ(λ)(cid:0)k+ − sin γ(λ)(cid:16)2√µ cos φ(λ) − k+ 1 + 2d2√µ cos φ(λ)(cid:1) 2 + k− Let us investigate (27). By using (9), (38) and Lemma 2.2, we conclude, from (27), that c2 = O(1), so that 2 + O(λ− 1 2α (cid:17) . α )(cid:17) = O(cid:0)λ−1(cid:1) + O(cid:16)λ− α+2 (41) sin φ(λ) = O(λ− 1 2 ), and consequently k± 2 = O(λ− 1 2 ). Therefore (40) gives sin γ(λ) = (42) and hence 1 + k− 2 ) + O(λ− 1 α )(cid:0)k+ cos γ(λ) 2√µ cos φ(λ) + O(λ− 1 + O(cid:16)λ− 2α+2 2α (cid:17) + O(cid:16)λ− 3 2(cid:17) 1 ) + d2 + O(cid:16)λ− 2α+2 1 + k− (k+ 1 γ(λ) = 2√µ cos φ(λ) Therefore, from (23), it follows 1 + +2d2√µ cos φ(λ)(cid:1) 2α (cid:17) + O(cid:16)λ− 3 2(cid:17) . γ(λ) = (43) (q(s) − q(b))ds + 1 −b 4√µZ b 4√µ cos φ(λ) Z b sin φ(λ) + −b 1 4√µZ b (q(s) − q(b)) cos(2√µs)ds (q(s) − q(b)) sin(2√µs)ds + d2 + O(λ−1), −b so that the Riemann-Lebesgue lemma and (41) imply γ(λ) = 1 4√µZ b −b (q(s) − q(b))ds + 1 4√µZ b −b V (s) cos(2√µs)ds + d2 + O(cid:0)λ−1(cid:1) . This and (39) give 13 (44) nπ =Q(b, λn) + 3π 4 − 1 4pλn Z b −b 1 + bqλn − q(b) − (q(s) − q(b))ds V (s) cos(2qλns)ds − d2(λn) + O(cid:0)λ−1 n (cid:1) . 4pλn Z b −b From (37) and (44), we notice that solutions of types ΛD and λN are interlacing at infinity, that is (45) for sufficiently large n ∈ N. Denote ν2n−1 = λn and ν2n = λn. Then {νj}∞ j=1 are solutions to (25), (26), and νn+1 > νn for sufficiently large n ∈ N. Moreover, the asymptotic formulas (37) and (44) give λn ≤ λn ≤ λn+1 ≤ λn+1 ≤ ··· π 4 (46) (2n − 1) =Q(b, νn) + bpνn − q(b) − 1 4√νn Z b (q(s) − q(b))ds V (s) cos(2√νns)ds − d2(νn) + O(cid:0)ν−1 n (cid:1) −b − 1 4√νn Z b −b as n → ∞. As we mentioned before, the eigenvalues of H are solutions to (25), (26). Therefore, to prove the theorem, it is sufficient to show (47) N (λ, H) = #{νj ≤ λ} + O(1), λ > 0, where # means the cardinality of a set, and N (λ, H) is the counting function of H, that is N (λ, H) := #{λ′ ≤ λ, λ′ is an eigenvalue of H} = #{λj ≤ λ}. In order to prove (47), let us consider the operator H0 in L2(R), generated by the expression d2 dx2 + q0(x), − q0(x) = N Xj=1 cjxαj . Since the operator of multiplication by V (x) is bounded and symmetric in L2(R), [Kat95, Theorem 4.10] implies (48) N (λ, H) = N (λ, H0) + O(1), On the other hand, [Tit62, formula (7.7.4)] gives λ → ∞. (49) N (λ, H0) = 2 π Z a(λ) 0 pλ − q0(t)dt + O(1), λ → ∞. Finally, since q0(x) = q(x) for x > b, we estimate Z a(λ) 0 pλ − q0(t)dt − Q(b, λ) − bpλ − q(b) =Z a(λ) 0 pλ − q0(t)dt −Z a(λ) =Z b 0 pλ − q0(t)dt − bpλ − q(b) =Z b 0 b pλ − q(t)dt − bpλ − q(b) −q0(t) + q(b) pλ − q0(t) +pλ − q(b) dt = O(λ− 1 2 ). 14 K. FEDOSOVA AND M. NURSULTANOV Therefore (48) and (49) imply N (λ, H) = 2 π Q(b, λ) + 2 π bpλ − q(b) + O(1). This and (46) give (47), which consequently implies that eigenvalues {λn}∞ the asymptotic formula (46), and (9) completes the proof. n=1 satisfy (cid:3) Next we derive Theorem 1.1 from Theorem 1.2. Proof. According to [MA79, formula (38)], the function Q(x, λ), corresponding to the potential q(x) = xα + V (x), has the following asymptotic behavior, as λ → ∞, 2α − b√λ + 1 2√λ α+2 (50) Q(b, λ) = 2(cid:17) bα+1(α + 1)−1 + O(cid:16)λ− 3 for fixed b > 0 such that V (x) = 0 for x ≥ b. Inserting this into (46) gives + bpλn − q(b) (2n − 1) = bα+1 α + 1 π 4 α+2 1 2α λ n − bpλn + bα+1 2√λn − + bα+1 α + 1 2√λn · 4√λn Z b 1 −b V (s)ds λ 2 + 1 Γ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) αΓ(cid:0) 3 α(cid:1) Γ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) αΓ(cid:0) 3 α(cid:1) 2√λn · − 4√λn Z b 2 + 1 1 − 1 −b V (s) cos(2pλns)ds − d2(λn) + O(cid:0)λ−1 n (cid:1) . (51) Note that pλn − q(b) −pλn + bα 2√λn = = −bα 2√λn −pλn − q(b) − √λn 2√λn(cid:16)pλn − q(b) + √λn(cid:17) bα 2√λn −q(b) + pλn − q(b) + √λn = O(cid:16)λ n (cid:17) − 3 2 and, see [MA79, (49)], Therefore we can rewrite (51) in the following way π 4 (52) if α ≤ 2, if α > 2. λ− α+2 2α α α ) cot π α ) 2 )Γ( 1 O(λ−1) α(α−1)Γ( 3 2 + 1 48(2+α)Γ( 3 d2(λ) =  Γ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) n − λ αΓ(cid:0) 3 α(cid:1) 2 + 1 α(α − 1)Γ(cid:0) 3 − 48(2 + α)Γ(cid:0) 3 4√λn Z b 2 + 1 − −b α+2 1 2α (2n − 1) = V (s)ds 1 −∞ 4√λn Z +∞ α(cid:1) cot π 2(cid:1) Γ(cid:0) 1 α(cid:1) α − α+2 λ 2α n V (s) cos(2pλns)ds + O(λ−1 n ). as n → ∞. This implies (2). (cid:3) 4. Remarks and consequences 15 We start with the following remark. Remark 4.1. According to [SR03, page 92], the third and fourth terms are O(n− ατ +2 α+2 ) and O(n− 4 α+2 ), respectively, so that they may change their order, depending on V (x) and the relation between α and τ . Moreover, in case α ≤ 2, the fourth term will be absorbed by O(n−1). However, the third term, in general, is not absorbed by O(n−1); see Example 4.2. Next, we give an example demonstrating the effect of the parameter τ on the eigenvalues. Example 4.2. Let H be the self-adjoint operator in L2(R) generated by expression where, for 1 > τ > 0, d2 dx2 + x2 + V (x), − (53) V (x) =(P∞ 0 j=1 2−jτ cos(2jx) if x ≤ π, otherwise . By [Zyg77, Theorem 4.9], V (x) is a Hölder continuous function with exponent τ > 0 in [−π, π]. Therefore Theorem 1.1 shows that eigenvalues of H satisfy λn = 2n − 1 + 1 4√2nZ π −π V (s)ds + 1 4π√2nZ π −π V (s) cos(2√2ns)ds + O(n−1). Let us consider the subsequence of the eigenvalues {λnk}∞ k=2 with nk = 22k−3. Since Z π −π V (s) cos(2√2nks)ds =Z π 2−jτ cos(2js) cos(2ks)ds = π2−kτ , ∞ −π Xj=1 4√2Z π −π 1 we derive the asymptotic for the subsecuence of the eigenvalues of H, {λnk}∞ k=2, λnk = 2nk − 1 + n − 1 2 k V (s)ds + n − 1+τ 2 k 2− 5+3τ 2 + O(n−1 k ). Next, we give two examples for which the asymptotic formulas (4) can be written more explicitly. Example 4.3. Assume that V (x) satisfies conditions of Theorem 1.2 and α ∈ N, c ∈ R. Let H be the self-adjoint operator in L2(R) generated by the expression d2 dx2 + (x + c)α + V (x). − By Theorem 1.2, the eigenvalues of H satisfy (4). By [MA79, (38)], we compute: 1 b Q(b, λ) =Z a(λ) (λ − (t + c)α) Γ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) αΓ(cid:0) 3 α(cid:1) 2 + 1 = α+2 λ α+2 2 dt = λ 2α Z 1 2α − (b + c)√λ + 1 2√λ b+c λ1/α 1 2 dt (1 − tα) (b + c)α+1(α + 1)−1 + O(cid:16)λ− 3 2(cid:17) 16 K. FEDOSOVA AND M. NURSULTANOV Therefore (4) gives π 4 (2n − 1) = − 2α 1 λ α+2 2 + 1 Γ(cid:0) 3 2(cid:1) Γ(cid:0) 1 α(cid:1) n − cpλn − αΓ(cid:0) 3 α(cid:1) 4√λn Z b V (s) cos(2pλns)ds + O(cid:16)λ 4√λn (cid:18) 2cα+1 α + 1 −Z +∞ V (s)ds(cid:19) (cid:17) + O(cid:0)λ−1 n (cid:1) . − α+2 2α n −∞ −b 1 This gives the first two terms of the heat trace. Indeed, this implies that N (λ) = 1 2 C1λ α+2 2α − 2c π 1 2 + O(1) λ By applying the Watson's Lemma [PS06, Lemma 2.2], we obtain as t → +0. Therefore, as t → +0, e−tλN (λ)dλ = 0 Z ∞ e−tλn =Z ∞ 0 ∞ Xn=1 1 2 C1Γ(cid:18)1 + e−tλdN (λ) = tZ ∞ 1 √π = 0 α + 2 2α (cid:19) t− α+2 2α −1 − c √π t− 3 2 + O(t−1) e−tλN (λ)dλ Γ(cid:18) α + 1 α (cid:19) t− α+2 2α − c √π t− 1 2 + O (1) . By the same way, one can derive the first term of the heat trace corresponding to the operator defined in Theorem 1.1 and verify that the perturbation does not affect the first term: Corollary 4.4. Under the conditions of Theorem 1.1, the heat trace satisfies e−tλn = 1 √π ∞ Xn=1 Γ(cid:18) α + 1 α (cid:19) t− α+2 2α + O (1) , as t → +0. The next example is a quartic AHO, for which Q(b, λ) expands in terms of {λ− k Consequently, the asymptotic formula can be written more explicitly. 4 }+∞ k=−3. Example 4.5 (Quartic AHO). Assume that function V (x) satisfies conditions of Theorem 1.2 and c ∈ R. Let H be the self-adjoint operator in L2(R) generated by d2 dx2 + (x2 + c)2 + V (x), − c ∈ R. Then, by Theorem 1.2, the eigenvalues of H satisfy (4). Let us investigate Q(b, λ): b Q(b, λ) =Z a(λ) 4 Z 1 = λ 3 (λ [λ − (x2 + c)2] 1 2 −c) 1 2 1 2 dx =Z (λ b "1 −(cid:18)(cid:18)1 − c λ 1 2(cid:19) y2 + b 1 2 −c) 1 2 1 2 dx 1 [λ − (x2 + c)2] 2(cid:19)2# 2 (cid:18)1 − c λ 1 c λ 1 2(cid:19) 1 2 dy. Let λ = 1/r4 and g(r) :=Z 1 br (1−cr2 ) 2 (cid:2)1 − ((1 − cr2)y2 + cr2)2(cid:3) 1 1 2 (1 − cr2) 1 2 dy. By computing the Taylor expansion of g(r) at r = 0, one can express Q(b, λ) in terms of {λ− n 4 }. Therefore, by (4), we conclude 17 π 4 (2n − 1) = 6 Xk=0 akλ 3−k 4 − 1 4√λn Z b −b V (s) cos(2pλns)ds + O(λ−1), were ak = g(k)(0)/k! for k 6= 1, 5 and 1 a1 = g(1)(0) − 1, a5 = g(5)(0)/5! − 4Z b Remark 4.6. Similarly, one can investigate AHO with potential q(x) = (xα + c)n with α > 0, m ∈ N, and derive more explicit forms of (4). We end this section by considering the operators in L2[0,∞) generated by the expression (1) with Dirichlet and Neumann boundary conditions, respectively. q(s)ds. −b Corollary 4.7. Assume that V (x) satisfies the conditions of Theorem 1.1. Let {λD j=1 be the eigenvalues of the operators in L2[0,∞) generated by (1) with Dirichlet and Neumann boundary conditions, respectively. Then j=1 and {λN j }∞ j }∞ (54) (55) − 2α λN α+2 n = C 1 (4n − 1) 1 4π − α+4 α+2 C 1 − α+6 α+2 C2C 1 + + 2α α + 2 2α α + 2 (4n − 3) 1 − α+4 α+2 C 1 4π − α+6 α+2 C2C 1 − 2α λN α+2 n = C 1 + + 2α α + 2 2α α + 2 π R ∞ 2α α+2 + 2α α + 2 − α+1 α+2 C0C 1 (4n − 1)− 2 (4n − 1)− 4 0 α+2 Z ∞ α+2 + O(cid:0)n−1(cid:1) , α+2 (4n − 3)− 2 V (s) cos(cid:16)2C − α 1 α+2 2α α+2 + 2α α + 2 − α+4 α+2 C0C 1 (4n − 3)− 2 (4n − 3)− 4 0 α+2 Z ∞ α+2 + O(cid:0)n−1(cid:1) , α+2 (4n − 3)− 2 V (s) cos(cid:16)2C − α 1 α+2 (4n − 1) α α+2 s(cid:17) ds (4n − 3) α α+2 s(cid:17) ds where C0 = 1 0 V (x)dx and C1, C2 are the constants defined in Theorem 1.1. Proof. Consider the solution of (5) in [0, b] with boundary conditions y(0) = 1, y′(0) = 0, and the solution of (5) in [b,∞). By gluing them together, we obtain (25) (with c1 = 1 and c2 = 0) (56) π 4 sin(cid:16)Q + + cos(cid:16)Q + + √µb(cid:17)(cid:16)d1k+ 1 + d2√µ − d2k+ 2(cid:17) + √µb(cid:17)(cid:16)d1√µ − d1k+ 2 − d2k+ π 4 1(cid:17) = O(cid:16)λ− α+4 2α (cid:17) + O(cid:0)λ−1(cid:1) . The same analysis we did for (28)(in order to investigate λn) yields formula (44), which consequently implies (55). Similarly, one can prove (54). (cid:3) Remark 4.8. Corollary 4.7 for Dirichlet boundary condition is obtained in [MA79]. 18 K. FEDOSOVA AND M. NURSULTANOV References [Akh08] E.F. Akhmerova, The asymptotics of the spectrum of nonsmooth perturbations of a harmonic oscillator, Sibirsk. Mat. Zh. 49 (2008), no. 6, 1216 -- 1234. MR 2499095 [BGJ18] C. Bernardin, P. Gonçalves, and M. Jara, Weakly harmonic oscillators perturbed by a conservative noise, Ann. Appl. Probab. 28 (2018), no. 3, 1315 -- 1355. MR 3809465 [BW69] C. M. Bender and T. T. Wu, Anharmonic oscillator, Phys. Rev. (2) 184 (1969), 1231 -- 1260. MR 0260323 [Cam16] B. Camus, Lower bounds for non-classical eigenvalue problems , Integral Equations Op- erator Theory 85 (2016), no. 1, 25 -- 36. MR 3503177 [CDR18] M. Chatzakou, J. Delgado, and M. Ruzhansky, On a class of anharmonic oscillators, arXiv:1811.12566v2 (2018). [CR15] B. Camus and N. Rautenberg, Higher dimensional nonclassical eigenvalue asymptotics, J. Math. Phys. 56 (2015), no. 2, 021506, 14. MR 3390858 [Eva10] G.C. Evans, Volterra's integral equation of the second kind, with discontinuous kernel , Trans. Amer. Math. Soc. 11 (1910), no. 4, 393 -- 413. MR 1500871 [Fuc18] G. Fucci, Asymptotic expansion of the heat kernel trace of Laplacians with polynomial potentials, Lett. Math. Phys. 108 (2018), no. 11, 2453 -- 2478. MR 3861383 [Git05] M. Gitterman, Classical harmonic oscillator with multiplicative noise, Physica A: Sta- tistical Mechanics and its Applications. 352 (2005), 309 -- 334. [GSS15] P.J. Gaudreau, R.M. Slevinsky, and H. Safouhi, Computing energy eigenvalues of anharmonic oscillators using the double exponential sinc collocation method , Ann. Physics 360 (2015), 520 -- 538. MR 3367544 [Gur89] D. Gurarie, Asymptotic inverse spectral problem for anharmonic oscillators with odd potentials, Inverse Problems 5 (1989), no. 3, 293 -- 306. MR 999064 [HR82] B. Helffer and D. Robert, Asymptotique des niveaux d'énergie pour des hamiltoniens à un degré de liberté, Duke Math. J. 49 (1982), no. 4, 853 -- 868. MR 683006 [Ike60] T. Ikebe, Eigenfunction expansions associated with the Schroedinger operators and their applications to scattering theory, Arch. Rational Mech. Anal. 5 (1960), 1 -- 34 (1960). MR 0128355 [Kat95] T. Kato, Perturbation theory for linear operators , Classics in Mathematics, Springer- Verlag, Berlin, 1995, Reprint of the 1980 edition. MR 1335452 [KKP05] Spectral asymptotics of the harmonic oscillator perturbed by bounded potentials, Ann. Henri Poincaré 6 (2005), no. 4, 747 -- 789. MR 2211839 [Kos70] N.M. Kostenko, The asymptotics of the eigenvalues of the anharmonic oscillator, Mat. [LS75] Sb. (N.S.) 81 (123) (1970), 163 -- 175. MR 0259253 B.M. Levitan and I.S. Sargsyan, Introduction to spectral theory: selfadjoint ordinary differential operators, American Mathematical Society, Providence, R.I., 1975, Trans- lated from the Russian by Amiel Feinstein, Translations of Mathematical Monographs, Vol. 39. MR 0369797 [MA79] H.H. Murtazin and T.G. Amangildin, Asymptotic behavior of the spectrum of the Sturm-Liouville operator, Mat. Sb. (N.S.) 110(152) (1979), no. 1, 135 -- 149, 160. MR 548522 [PS06] A. Pushnitski and I. Sorrell, High energy asymptotics and trace formulas for the perturbed harmonic oscillator, Ann. Henri Poincaré 7 (2006), no. 2, 381 -- 396. MR 2210237 [Sak81] L.A. Sakhnovich, The asymptotic behavior of the spectrum of an anharmonic oscillator, Teoret. Mat. Fiz. 47 (1981), no. 2, 266 -- 276. MR 626987 [Sat15] D. N. Sathyanarayana, Vibrational spectroscopy: theory and applications, New Age International, 2015. [Sch70] E. J. P. G. Schmidt, On the representation of the potential scattering operator in [SR03] quantum mechanics, J. Differential Equations 7 (1970), 389 -- 394. MR 0256006 E.M. Stein and S. Rami, Fourier analysis, Princeton Lectures in Analysis, vol. 1, Prince- ton University Press, Princeton, NJ, 2003, An introduction. MR 1970295 [Tit62] E.C. Titchmarsh, Eigenfunction expansions associated with second-order differential equations. Part I, Second Edition, Clarendon Press, Oxford, 1962. MR 0176151 19 [Vis57] K. S. Viswanathan, The theory of the anharmonic oscillator, Proc. Indian Acad. Sci. Sect. A. 46 (1957), 203 -- 217. MR 0092356 [Zyg77] A. Zygmund, Trigonometric series. Vol. I, II, Cambridge University Press, Cambridge- New York-Melbourne, 1977, Reprinting of the 1968 version of the second edition with Volumes I and II bound together. MR 0617944 KF: Albert-Ludwigs-Universität Freiburg, Mathematisches Institut, Ernst-Zermelo- Str. 1, 79104 Freiburg im Breisgau, Germany E-mail address: [email protected] MN: Chalmers University of Technology and University of Gothenburg, SE-412 96, Gothenburg, Sweden E-mail address: [email protected]
math/0601556	4	0601	2010-10-31T20:37:18	Games in Banach spaces	[ "math.FA", "math.LO" ]	The notion of Aronszajn-null sets generalizes the notion of Lebesgue measure zero on the real line and the Euclidean space to infinite dimensional Banach spaces. We present a game-theoretic approach to Aronszajn-null sets, establish its basic properties, and discuss the ensuing open problems.	math.FA	math	NULL SETS AND GAMES IN BANACH SPACES JAKUB DUDA AND BOAZ TSABAN Abstract. The notion of Aronszajn-null sets generalizes the no- tion of Lebesgue measure zero in the Euclidean space to infinite dimensional Banach spaces. We present a game-theoretic approach to Aronszajn-null sets, establish its basic properties, and discuss some ensuing open problems. 1. Motivation Aronszajn null sets were introduced by Aronszajn in the context of studying almost-everywhere differentiability of Lipschitz mappings between Banach spaces. Christensen, Phelps and Mankiewicz studied the same problem independently and used Haar null, Gaussian null and cube null sets, respectively. More information about the history is available in the monograph [1]. Csornyei [3] proved that (Borel) Aronszajn null, Gaussian null, and cube null sets coincide. It is well known that Haar null sets form a strictly larger family than Aronszajn null sets (see [1]). One of the questions in differentiability theory is to understand the structure of the sets of points of Gateaux nondifferentiability of Lip- schitz mappings defined on separable Banach spaces. The strongest result in this context is due to Preiss and Zaj´ıcek [4]. Let A denote the family of Borel Aronszajn-null sets (to be defined in the sequel). Preiss and Zaj´ıcek introduced a Borel σ-ideal A such that A ⊆ A. It follows from a recent result of Preiss that A = A in R2, and it is unknown for 2 < dim X < ∞. In infinite dimensions, the inclusion A ⊆ A is strict. It is also unknown whether, according to the definitions of [4], A = C and C ⊆ A∗. Understanding the structure of the sets of points of non-differentiabi- lity could possibly also be helpful in answering the longstanding open problem whether two separable Lipschitz isomorphic spaces are actu- ally linearly isomorphic. This is known for some special Banach spaces, but is open for example for ℓ1 and L1. 2000 Mathematics Subject Classification. 46G99, 91A44. Key words and phrases. Aronszajn null, selection principles, topological games. 1 2 JAKUB DUDA AND BOAZ TSABAN We introduce a game-theoretic approach to Aronszajn null sets. One idea behind this approach is that the Aronszajn null sets are defined as sets for which there exists a certain decomposition for each complete sequence of directions, whereas in the game setting, such a decompo- sition is being constructed while we are only given one direction at a time. Games have often been used in Banach spaces (see, e.g., the sur- vey paper [2]), and it would be interesting to see whether this new perspective can yield interesting results which do not involve the new notions. 2. The Aronszajn-null game Let X be a separable Banach space (over R). The following defini- tions are classical: (1) For a nonzero x ∈ X, A(x) denotes the collection of all Borel sets A ⊆ X such that for each y ∈ X, A∩(Rx+y) has Lebesgue (one dimensional) measure zero. (2) A Borel set A ⊆ X is Aronszajn-null if for each dense sequence {xn}n∈N ⊆ X, there exist elements An ∈ A(xn), n ∈ N, such that A ⊆Sn An. A is a Borel σ-ideal. (3) A denotes the collection of Aronszajn-null sets. Remark 2.1. Replacing "dense" by "complete" in item 2 of the above definition of Aronszajn-null sets (i.e., requiring just that the linear span of {xn : n ∈ N} is dense in X), one gets an equivalent definition [1, Corollary 6.30]. The definition of Aronszajn-null sets motivates the following. Definition 2.2. The Aronszajn-null game AG for a Borel set A ⊆ X is a game between two players, I and II, who play an inning per each natural number. In the nth inning, I picks xn ∈ X, and II responds by picking An ∈ A(xn). This is illustrated in the following figure. I: x1 ∈ X ց x2 ∈ X ր ց . . . II: A1 ∈ A(x1) A2 ∈ A(x2) . . . I is required to play such that {xn}n∈N is dense in X. II wins the game if A ⊆Sn An; otherwise I wins. NULL SETS AND GAMES IN BANACH SPACES 3 For a game G, the notation I ↑ G is a shorthand for "I has a winning strategy in the game G", and I9G stands for "I does not have a win- ning strategy in the game G". Define II ↑ G and II9G similarly. The following is easy to see. Lemma 2.3. If I9AG for A, then A is Aronszajn-null. (cid:3) The converse is open. Conjecture 2.4. If A is Aronszajn-null, then I9AG for A. Lemma 2.5. The property II ↑ AG is preserved under taking Borel sub- sets and countable unions, i.e., it defines a Borel σ-ideal. Proof. It is obvious that II ↑ AG is preserved under taking Borel sub- sets. To see the remaining assertion, assume that B1, B2, . . . all satisfy II ↑ AG, and for each k let Fk be a winning strategy for II in the game AG played on Bk. Define a strategy F for II in the game AG played on In the nth inning we have (x1, A1, x2, A2, . . . , xn) given, where xn is the nth move of I. For each k let Ak,n = Fk(x1, Ak,1, x2, Ak,2, . . . , xn), and Sk Bk as follows. Assume that I played x1 ∈ X in the first inning. For each k let Ak,1 = Fk(x1), and set A1 = Sk Ak,1 ∈ A(x1). II plays A1. set An =Sn Ak,n ∈ A(xn). II plays An. . . . ) is a play according to the strategy Fk, and therefore Bk ⊆Sn Ak,n. Consider the play (x1, A1, x2, A2, . . . ). For each k, (x1, Ak,1, x2, Ak,2, Consequently, An, (cid:3) B = [k∈N Bk ⊆ [k∈N [n∈N Ak,n = [n∈N[k∈N Ak,n = [n∈N thus II won the play. A Borel set A ⊆ X is directionally-porous if there exist λ > 0 and a nonzero v ∈ X such that for each a ∈ A and each positive ǫ, there is x ∈ Rv + a such that kx − ak < ǫ and A ∩ B(x, λkx − ak) = ∅. If A is directionally-porous, then so is A. A is σ-directionally-porous if it is a countable union of directionally-porous sets. Proposition 2.6. For each σ-directionally-porous set, II ↑ AG. Proof. By Lemma 2.5, it suffices to consider the case where A ⊆ X is directionally-porous. Let λ > 0 and v ∈ X be witnesses for that. In this case, the function F (x1, A1, x2, A2, . . . , xn) =(A kxn − vk < λ/2 otherwise ∅ is a winning strategy for II in the game AG. (cid:3) 4 JAKUB DUDA AND BOAZ TSABAN For a nonzero x ∈ X and a positive ǫ, let A(x, ǫ) denote the collection of all Borel sets A ⊆ X such that for each v ∈ X with kv − xk < ǫ, A ∈ A(v). C∗ is the collection of all countable unions of sets An such that each An ∈ A(xn, ǫn) for some xn, ǫn. C∗ is a Borel σ-ideal. The proof of Proposition 2.6 actually establishes the following. Proposition 2.7. For each A ∈ C∗, II ↑ AG. (cid:3) The following diagram summarizes our knowledge thus far: σ-directionally-porous =⇒ C∗ =⇒ II ↑ AG =⇒ I9AG =⇒ A. The open problems concerning this diagram are whether any of the last three arrows can be reversed (i.e., turned into an equivalence) and therefore produce a characterization. The first arrow is not reversible [4]. We conjecture that II ↑ AG is strictly stronger than A. For brevity, we introduce the following. Definition 2.8. For Y ⊆ X, A ∈ A(∧ Y ) means: For each y ∈ Y , A ∈ A(y). In other words, A(∧ Y ) = \y∈Y A(y). Thus, A(x, ǫ) = A(∧ B(x, ǫ)). Using this notation, we can see that the property II ↑ AG implies something quite close to A∗, see Corollary 2.10. Recall that for a topological space X, a pseudo-base is a family U of open subsets of X, such that each open subset of X contains some element of U as a subset. Clearly, every base is a pseudo-base. Theorem 2.9. Assume that II ↑ AG holds for A. Then: For each countable dense D ⊆ X and each pseudo-base {Un}n∈N for the topology of X, there exist elements An ∈ A(∧ D ∩ Un), n ∈ N, such that A ⊆Sn An. Proof. Assume that D ⊆ X is countable and dense, and {Un}n∈N is a pseudo-base for the topology of X. For each n, fix an enumeration {xn,m : m ∈ N} of D ∩ Un. Let F be a winning strategy for II in the game AG. To each finite sequence η of natural numbers we associate a Borel set Aη and an element yη ∈ D ∩ Un where n is the length of the sequence. This is done by induction on n. n = 1: For each k, set Ak = F (x1,k). NULL SETS AND GAMES IN BANACH SPACES 5 n = m + 1: For each η ∈ Nm and each k, define Aηk = F (x1,η1, Aη1, x2,η2, Aη2, . . . , xm,ηm, Aη, xm+1,k), where for each i, ηi is the ith element of η and ηi is the sequence (η1, . . . , ηi). a /∈ A(k1,k2), etc. Then the play (x1,k1, Ak1, x2,k2, A(k1,k2), . . . ) is accord- ing to the strategy F and lost by II, a contradiction. Consequently, Next, for each η, define Bη =Tk Aηk. Assume that A 6⊆Sη Bη, and let a ∈ A\Sη Bη. Choose inductively k1 such that a /∈ Ak1, k2 such that A ⊆Sη Bη. For each m and each η ∈ Nm, Bη = Tk Aηk ∈ A(∧ D ∩ Um). Consequently, Cm = Sη∈Nm Bη ∈ A(∧ D ∩ Um) too, and A ⊆ Sm Cm as required. (cid:3) Corollary 2.10. Assume that II ↑ AG holds for A. Then: For each countable dense D ⊆ X, there exist elements An ∈ A(∧ D ∩ B(xn, ǫn)), where each xn ∈ X and each ǫn > 0, such that A ⊆Sn An. Problem 2.11. Is the property in Corollary 2.10 equivalent to II ↑ AG, or does it at least imply I9AG? (cid:3) 3. Selection hypotheses Definition 3.1. A is the collection of Borel sets A ⊆ X such that: For each sequence {Dn}n∈N of dense subsets of X, there exist elements xn ∈ Dn and An ∈ A(xn), n ∈ N, such that A ⊆ Sn An. AG is the corresponding game, played as follows: I: D1 ⊆ X ց II: x1 ∈ D1, A1 ∈ A(x1) D2 ⊆ X ր ց . . . . . . x2 ∈ D2, A2 ∈ A(x2) where each Dn is dense in X, and II wins the game if A ⊆ Sn An; otherwise I wins. The appealing property in the game AG is that, unlike the case in the game AG, there is no commitment of I which has to be verified "at the end" of the play. Proposition 3.2. A = A. 6 JAKUB DUDA AND BOAZ TSABAN Proof. (⊆) Assume that A ∈ A, and let D = {xn}n∈N be dense in X. For each n, take Dn = D and apply A. Then there are yn ∈ D and An ∈ A(yn), n ∈ N, such that A ⊆ Sn An. As each A(xn) is σ-additive, we may assume that no xn appears more than once in the sequence {yn}n∈N. Thus, A ∈ A. (⊇) Assume that A ∈ A, and let {Dn}n∈N be a sequence of dense subsets of X. For each n choose xn ∈ Dn such that D = {xn}n∈N is dense in X (to do that, fix a countable base {Un}n∈N for the topology of X, and for each n pick xn ∈ Un ∩ Dn). By A, there exist sets (cid:3) An ∈ A(xn) such that A ⊆Sn An. This shows that A ∈ A. A simple modification of the last proof gives the following. Theorem 3.3. I ↑ AG if, and only if, I ↑ AG. Proof. (⇒) Let F be a winning strategy for I in the game I ↑ AG on A. Define a strategy for I in the game I ↑ AG as follows. Fix a countable base {Un}n∈N for the topology of X. In the first inning, I plays any x1 ∈ U1 ∩ D1 where D1 is I's first move according to the strategy F . Assume that the first n moves where (x1, A1, . . . , xn−1, An−1). Let Dn = F (D1, (x1, A1), . . . , Dn−1, (xn−1, An−1)). Then I plays any xn ∈ Un ∩ Dn. For each play (x1, A1, x2, A2, . . . ) according to this strategy, {xn}n∈N is dense in X, and since (D1, (x1, A1), D2, (x2, A2), . . . ) is a play in the game AG according to the strategy F , A 6⊆Sn An. (⇐) Let F be a winning strategy for I in the game I ↑ AG on A. Define a strategy for I in the game I ↑ AG as follows. I's first move is D1, the set of all points x which are possible moves of I at some inning according to its strategy F . Obviously, D1 is dense. In the nth inning, we are given (D1, (x1, A1), . . . , Dn−1, (xn−1, An−1)), such that there is a sequence of moves (y1, B1, y2, B2, . . . , ykn) according to the strategy F , with ykn = xn−1. Then I plays Dn, the set of all points x which are possible moves of I at some future inning, in a play according to the strategy F whose first moves are (y1, B1, y2, B2, . . . , ykn = xn−1, An−1). (y1, B1, y2, B2, . . . ) is a play according to the strategy F , and therefore (cid:3) A 6⊆Sn Bn ⊇Sn An, so that A 6⊆Sn An. The following is immediate. Lemma 3.4. If II ↑ AG, then II ↑ AG. (cid:3) Problem 3.5. Is it true that II ↑ AG if, and only if, II ↑ AG? By Theorem 3.3 and Proposition 3.2, Conjecture 2.4 can be refor- mulated as follows. Conjecture 3.6. For a Borel set A ⊆ X: I 9AG if, and only if, A ∈ A. NULL SETS AND GAMES IN BANACH SPACES 7 Acknowledgments. The second author was partially supported by the Koshland Center for Basic Research. We thank Ori Gurel-Gurevich for presenting our results at the Weizmann Institute's seminar on Geo- metric Functional Analysis and Probability, and for making useful com- ments. References [1] Y. Benyamini and J. Lindenstrauss, Geometric Nonlinear Functional Analy- sis, Volume 1, Colloquium Publications 48, American Mathematical Society, Providence, 2000. [2] J. Cao and W. Morrs, A survey on topological games and their applications in analysis, Revista de la Real Academia de Ciencias Exactas, F´ısicas y Naturales, Serie A, Matem´aticas, 100 (2006), 39 -- 49. [3] M. Csornyei, Aronszajn null and Gaussian null sets coincide, Israel Journal of Mathematics 111 (1999), 191 -- 201. [4] D. Preiss and L. Zaj´ıcek, Directional derivatives of Lipschitz functions, Israel Journal of Mathematics 125 (2001), 1 -- 27. (Jakub Duda) Department of Mathematics, Weizmann Institute of Science, Rehovot 76100, Israel. Current address: CEZ, a.s., Duhov´a 2/1444, 140 53 Praha 4, Czech Republic E-mail address: [email protected] (Boaz Tsaban) Department of Mathematics, Bar-Ilan University, Ramat- Gan 52900, Israel; and Department of Mathematics, Weizmann Insti- tute of Science, Rehovot 76100, Israel. E-mail address: [email protected]
1007.3110	1	1007	2010-07-19T10:17:18	Arens Regularity And Factorization Property	[ "math.FA" ]	In this paper, we will study some Arens regularity properties of module actions. Let $B$ be a Banach $A-bimodule$ and let ${Z}^\ell_{B^{}}(A^{})$ and ${Z}^\ell_{A^{}}(B^{})$ be the topological centers of the left module action $\pi_\ell:~A\times B\rightarrow B$ and the right module action $\pi_r:~B\times A\rightarrow B$, respectively. In this paper, we will extend some problems from topological center of second dual of Banach algebra $A$, $Z_1(A^{})$, into spaces ${Z}^\ell_{B^{}}(A^{})$ and ${Z}^\ell_{A^{}}(B^{})$. We investigate some relationships between ${Z}_1({A^{}})$ and topological centers of module actions. For an unital Banach $A-module$ $B$ we show that ${Z}^\ell_{A^{}}(B^{}){Z}_1({A^{}})={Z}^\ell_{A^{}}(B^{})$ and as results in group algebras, for locally compact group $G$, we have ${Z}^\ell_{{L^1(G)}^{}}(M(G)^{})M(G)={Z}^\ell_{{L^1(G)}^{}}(M(G)^{})$ and ${Z}^\ell_{M(G)^{}}({L^1(G)}^{})M(G)={Z}^\ell_{M(G)^{}}({L^1(G)}^{*})$. For Banach $A-bimodule$ $B$, if we assume that $B^B^{*}\subseteq A^$, then $~B^{}{Z}_1(A^{})\subseteq {Z}^\ell_{A^{}}(B^{})$ and moreover if $B$ is an unital as Banach $A-module$, then we conclude that $B^{}{Z}_1({A^{}})={Z}^\ell_{A^{}}(B^{})$. Let ${Z}^\ell_{A^{}}(B^{})A\subseteq B$ and suppose that $B$ is $WSC$, so we conclude that ${Z}^\ell_{A^{}}(B^{})=B$. If $\overline{B^{}A}\neq B^$ and $ B^{}$ has a left unit $A^{}-module$, then $Z^\ell_{B^{}}(A^{})\neq A^{**}$. We will also establish some relationships of Arens regularity of Banach algebras $A$, $B$ and Arens regularity of projective tensor product $A\hat{\otimes}B$.	math.FA	math	DERIVATIONS AND COHOMOLOGICAL GROUPS OF BANACH ALGEBRAS KAZEM HAGHNEJAD AZAR Abstract. Let B be a Banach A − bimodule and let n ≥ 0. We investigate the relationships between some cohomological groups of A, that is, if the topo- logical center of the left module action πℓ : A × B → B of A(2n) on B(2n) is B(2n) and H 1(A(2n+2), B(2n+2)) = 0, then we have H 1(A, B(2n)) = 0, and we find the relationships between cohomological groups such as H 1(A, B(n+2)) and H 1(A, B(n)), spacial H 1(A, B∗) and H 1(A, B(2n+1)). We obtain some results in Connes-amenability of Banach algebras, and so for every compact group G, we conclude that H 1 w∗ (L∞(G)∗, L∞(G)∗∗) = 0. Let G be an amenable locally com- pact group. Then there is a Banach L1(G) − bimodule such as (L∞(G), .) such that Z 1(L1(G), L∞(G)) = {Lf : f ∈ L∞(G)}. We also obtain some conclusions in the Arens regularity of module actions and weak amenability of Banach alge- bras. We introduce some new concepts as lef t − weak∗ − to − weak convergence property [= Lw∗wc−property] and right − weak∗ − to − weak convergence prop- erty [= Rw∗wc−property] with respect to A and we show that if A∗ and A∗∗, respectively, have Rw∗wc−property and Lw∗wc−property and A∗∗ is weakly amenable, then A is weakly amenable. We also show to relations between a derivation D : A → A∗ and this new concepts. 1. Preliminaries and Introduction Let B be a Banach A − bimodule. A derivation from A into B is a bounded linear mapping D : A → B such that D(xy) = xD(y) + D(x)y f or all x, y ∈ A. The space of continuous derivations from A into B is denoted by Z 1(A, B). Easy example of derivations are the inner derivations, which are given for each b ∈ B by δb(a) = ab − ba f or all a ∈ A. The space of inner derivations from A into B is denoted by N 1(A, B). The Banach algebra A is said to be a amenable, when for every Banach A − bimodule B, the inner derivations are only derivations existing from A into B∗. It is clear that A is amenable if and only if H 1(A, B∗) = Z 1(A, B∗)/N 1(A, B∗) = {0}. The concept of amenability for a Banach algebra A, introduced by Johnson in 1972, has proved to be of enormous importance in Banach algebra theory, see [13]. A Banach algebra A is said to be a weakly amenable, if every derivation from A into A∗ is inner. Similarly, A is weakly amenable if and only if H 1(A, A∗) = Z 1(A, A∗)/N 1(A, A∗) = {0}. The 2000 Mathematics Subject Classification. 46L06; 46L07; 46L10; 47L25. Key words and phrases. Amenability, weak amenability, n-weak amenability, cohomology groups, derivation, Connes-amenability, super-amenability, Arens regularity, topological centers, module ac- tions, lef t − weak∗ − to − weak convergence, n-th dual . 1 2 concept of weak amenability was first introduced by Bade, Curtis and Dales in [2] for commutative Banach algebras, and was extended to the noncommutative case by Johnson in [14]. For Banach A−bimodule B, the quotient space H 1(A, B) of all continuous derivations from A into B modulo the subspace of inner derivations is called the first cohomology group of A with coefficients in B. In this paper, by using the Arens regularity of module actions, for Banach algebra A, we find some relations between cohomology groups A and A(2n) with some appli- cations in the n − weak amenability of Banach algebras that introduced by Dales, Ghahramani, Grønbaek in [6]. So for this aim, we extended some propositions from [6, 7, 10] into general situations. We investigated to relationships between cohomology groups of A ⊕ B and A, B where A and B are Banach algebras and on the other hand, we establish some relationships between the Connes-amenability and supper- amenability of Banach algebras with some results in group algebras. We have also some conclusions in the Arens regularity of module actions. In last section, for Banach A−module B, we introduce a new definitions as lef t−weak∗ −to−weak convergence property [ = Lw∗wc−property] and right − weak∗ − to − weak convergence property [ = Rw∗wc−property] with respect to A and we show that if A∗ and A∗∗, respectively, have Rw∗wc−property and Lw∗wc−property and A∗∗ is weakly amenable, then A is weakly amenable. We introduce some notations and definitions that we used throughout this paper. Let A be a Banach algebra and A∗, A∗∗, respectively, are the first and second dual of 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∗. 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]. 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 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 : 3 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 [5, 20]. 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 [4, 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, 17, 20]. The mapping m is left strongly Arens irregular if Z1(m) = X and m is right strongly Arens irregular if Z2(m) = Y . Regarding A as a Banach A − bimodule, the operation π : A × A → A extends to π∗∗∗ and πt∗∗∗t defined on A∗∗ × A∗∗. These extensions are known, respectively, as the first (left) and the second (right) Arens products, and with each of them, the second dual space A∗∗ becomes a Banach algebra. In this situation, we shall also simplify our notations. So the first (left) Arens product of a′′, b′′ ∈ A∗∗ shall be simply indicated by a′′b′′ and defined by the three steps: < a′a, b >=< a′, ab >, < a′′a′, a >=< a′′, a′a >, < a′′b′′, a′ >=< a′′, b′′a′ > . for every a, b ∈ A and a′ ∈ A∗. Similarly, the second (right) Arens product of a′′, b′′ ∈ A∗∗ shall be indicated by a′′ob′′ and defined by : < aoa′, b >=< a′, ba >, < a′oa′′, a >=< a′′, aoa′ >, < a′′ob′′, a′ >=< b′′, a′ob′′ > . for all a, b ∈ A and a′ ∈ A∗. 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 4 A∗∗, the second dual of A. By Goldstine,s Theorem [4, 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′, π(aα, bβ) >=< a′′b′′, a′ > lim β lim α < a′, π(aα, bβ) >=< a′′ob′′, a′ >, where a′′b′′ and a′′ob′′ are the first and second Arens products of A∗∗, respectively, see [6, 17, 20]. We find the usual first and second topological center of A∗∗, which are Z1(A∗∗) = Z ℓ Z2(A∗∗) = Z r 1(A∗∗) = {a′′ ∈ A∗∗ : b′′ → a′′b′′ is weak∗ − to − weak∗ continuous}, 2 (A∗∗) = {a′′ ∈ A∗∗ : a′′ → a′′ob′′ is weak∗ − to − weak∗ continuous}. 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 now B be a Banach A − bimodule, and let πℓ : A × B → B and πr : B × A → B. be the right and left 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∗∗. 2. Cohomological groups of Banach algebras Let B be a Banach A − bimodule and let n ≥ 1. Suppose that B(n) is an n − th dual of B. Then B(n) is also Banach A − bimodule, that is, for every a ∈ A, b(n) ∈ B(n) and b(n−1) ∈ B(n−1), we define and < b(n)a, b(n−1) >=< b(n), ab(n−1) >, < ab(n), b(n−1) >=< b(n), b(n−1)a > . In the following theorem, we extend Theorem 1.9 from [6] into general situation. Theorem 2-1. Let B be a Banach A−bimodule and let n ≥ 1. If H 1(A, B(n+2)) = 0, then H 1(A, B(n)) = 0. Proof. Let D ∈ Z 1(A, B(n)) and suppose that i : B(n) → B(n+2) is the canonical linear mapping as A − bimodule homomorphism. Take eD = ioD. Then we can be viewed eD as an element of Z 1(A, B(n+2)). Since H 1(A, B(n+2)) = 0, there exist b(n+2) ∈ B(n+2) such that for every a ∈ A, we have 5 eD(a) = ab(n+2) − b(n+2)a. Set a A − linear mapping P from B(n+2) into B(n) such that P oi = IB(n) . Then we have P oeD = (P oi)oD = D, and so for every a ∈ A, we conclude that D(a) = P oeD(a) = aP (b(n+2)) − P (b(n+2))a. It follows that D ∈ N 1(A, B(n)). Consequently we have H 1(A, B(n)) = 0. (cid:3) Suppose that A is a Banach algebra and B is a Banach A − bimodule. According to [6, pp.27 and 28], B∗∗ is a Banach A∗∗ − bimodule, where A∗∗ is equipped with the first Arens product. So we can define the topological centers of module actions. Thus, for a Banach A − bimodule B, we define the topological centers of the left and right module actions of A∗∗ on B∗∗ as follows: Z ℓ A∗∗(B∗∗) = Z(πr) = {b′′ ∈ B∗∗ : the map a′′ → π∗∗∗ r (b′′, a′′) : A∗∗ → B∗∗ is weak∗ − to − weak∗ continuous} Z ℓ B∗∗(A∗∗) = Z(πℓ) = {a′′ ∈ A∗∗ : the map b′′ → π∗∗∗ ℓ (a′′, b′′) : B∗∗ → B∗∗ is weak∗ − to − weak∗ continuous} Z r A∗∗(B∗∗) = Z(πt ℓ) = {b′′ ∈ B∗∗ : the map a′′ → πt∗∗∗ ℓ (b′′, a′′) : A∗∗ → B∗∗ is weak∗ − to − weak∗ continuous} Z r B∗∗(A∗∗) = Z(πt r) = {a′′ ∈ A∗∗ : the map b′′ → πt∗∗∗ r (a′′, b′′) : B∗∗ → B∗∗ is weak∗ − to − weak∗ continuous}. In every parts of this paper, for left or right Banach A − module B, we take πℓ(a, b) = ab and πr(b, a) = ba, for each a ∈ A and b ∈ B. Let A(n) and B(n) be n − th dual of A and B, respectively. By [25, page 4132-4134], if n ≥ 0 is an even number, then B(n) is a Banach A(n) − bimodule. Then for n ≥ 2, we define B(n)B(n−1) as a subspace of A(n−1), that is, for all b(n) ∈ B(n), b(n−1) ∈ B(n−1) and a(n−2) ∈ A(n−2) we define < b(n)b(n−1), a(n−2) >=< b(n), b(n−1)a(n−2) > . If n is odd number, then for n ≥ 1, we define B(n)B(n−1) as a subspace of A(n), that is, for all b(n) ∈ B(n), b(n−1) ∈ B(n−1) and a(n−1) ∈ A(n−1) we define < b(n)b(n−1), a(n−1) >=< b(n), b(n−1)a(n−1) > . and if n = 0, we take A(0) = A and B(0) = B. We also define the topological centers of module actions of A(n) on B(n) as follows A(n) (B(n)) = {b(n) ∈ B(n) : the map a(n) → b(n)a(n) Z ℓ is weak∗ − to − weak∗ continuous} : A(n) → B(n) 6 B(n) (A(n)) = {a(n) ∈ A(n) : the map b(n) → a(n)b(n) Z ℓ is weak∗ − to − weak∗ continuous}. : B(n) → B(n) In the following, for every n ≥ 1, using of the Arens regularity of left module action of A(2n) on B(2n), we extend every derivation D : A → B(2n) into derivation eD from A(2n) into B(2n) such that eD(a) = D(a) for every a ∈ A and we have some conclusion in cohomological groups. Theorem 2-2. Let B be a Banach A − bimodule and D : A → B(2n) be a continuous A(2n)(B(2n)) = B(2n). Then there is a continuous derivation derivation. Assume that Z ℓ eD : A(2n) → B(2n) such that eD(a) = D(a) for every a ∈ A. Proof. By using [6, Proposition 1.7], the linear mapping D′′ : A∗∗ → B(2n+2) is a continuous derivation. Take X = B(2n−2). Since ZA(2n) (X ∗∗) = ZA(2n)(B(2n)) = B(2n) = X ∗∗, by [6, Proposition 1.8], the canonical projection P : X (4) → X ∗∗ is a A∗∗ − bimodule morphism. Set eD = P oD′′. Then eD is a continuous derivation from A∗∗ into B(2n). Now by replacing A∗∗ by A and repeating of the proof, we obtain the result. (cid:3) Corollary 2-3. Let B be a Banach A−bimodule and n ≥ 0. If Z ℓ and H 1(A(2n+2), B(2n+2)) = 0, then H 1(A, B(2n)) = 0. A(2n) (B(2n)) = B(2n) Proof. By using [6, Proposition 1.7] and proceeding theorem the result is hold. (cid:3) Let A be a Banach algebra and n ≥ 0. Then A is n − weakly amenable if H 1(A, A(n)) = 0. A is permanently weakly amenable if A is n − weakly amenable for each n ≥ 0. Corollary 2-4 [6]. Let A be a Banach algebra such that A(2n) is Arens regular and H 1(A(2n+2)), A(2n+2)) = 0 for each n ≥ 0. Then A is 2n − weakly amenable. Assume that A is Banach algebra and n ≥ 0. We define A[n] as a subset of A as follows We write An the linear span of A[n] as a subalgebra of A. A[n] = {a1a2...an : a1, a2, ...an ∈ A}. Theorem 2-5. Let A be a Banach algebra and n ≥ 0. Let A[2n] dense in A and suppose that B is a Banach A − bimodule. Assume that AB∗∗ and B∗∗A are subsets of B. If H 1(A, B∗) = 0, then H 1(A, B(2n+1)) = 0. Proof. For n = 0 the result is clear. Let B⊥ be the space of functionals in A(2n+1) which annihilate i(B) where i : B → B(2n) is a natural canonical mapping. Then, by using [26, lemma 1], we have the following equality. B(2n+1) = i(B∗) ⊕ B⊥, as Banach A − bimodules, and so H 1(A, B(2n+1)) = H 1(A, i(B∗)) ⊕ H 1(A, B⊥). 7 Without lose generality, we replace i(B∗) by B∗. Since H 1(A, B∗) = 0, it is enough to show that H 1(A, B⊥) = 0. Now, take the linear mappings La and Ra from B into itself by La(b) = ab and a (b′′) = ab′′ Ra(b) = ba for every a ∈ A. Since AB∗∗ ⊆ B and B∗∗A ⊆ B, L∗∗ and R∗∗ a (b′′) = b′′a for every a ∈ A, respectively. Consequently, La and Ra from B into itself are weakly compact. It follows that for each a ∈ A the linear mappings L(2n) and R(2n) from B(n) into B(n) are weakly compact and for every b(2n) ∈ B(2n), we have L(2n) (b(2n)) = b(2n)a ∈ B(2n−2). Set a1, a2, ..., an ∈ A and b(2n) ∈ B(2n). Then a1a2...anb(2n) and b(2n)a1a2...an are belong to B. Suppose that D ∈ Z 1(A, B⊥) and let a, x ∈ A[n]. Then for every b(2n) ∈ B(2n), since xb(2n), b(2n)a ∈ B, we have the following equality (b(2n)) = ab(2n) ∈ B(2n−2) and R(2n) a a a a < D(ax), b(2n) >=< aD(x), b(2n) > + < D(a)x, b(2n) > =< D(x), b(2n)a > + < D(a), xb(2n) >= 0. It follows that D A[2n] . Since A[2n] dense in A, D = 0. Hence H 1(A, B⊥) = 0 and result follows. (cid:3) Corollary 2-6. i) Let A be a Banach algebra with bounded left approximate identity [=LBAI], and let B be a Banach A − bimodule. Suppose that AB∗∗ and B∗∗A are subset of B. Then if H 1(A, B∗) = 0, it follows that H 1(A, B(2n+1)) = 0. ii) Let A be an amenable Banach algebra and B be a Banach A − bimodule. If AB∗∗ and B∗∗A are subset of B, then H 1(A, B(2n+1)) = 0. Example 2-7. Assume that G is a compact group. Then i) we know that L1(G) is M (G) − bimodule and L1(G) is an ideal in the second dual of M (G), M (G)∗∗. By using [18, corollary 1.2], we have H 1(L1(G), M (G)) = 0. Then for every n ≥ 1, by using proceeding corollary, we conclude that H 1(L1(G), M (G)(2n+1)) = 0. ii) we have L1(G) is an ideal in its second dual , L1(G)∗∗. By using [15], we know that L1(G) is a weakly amenable. Then by proceeding corollary, L1(G) is (2n+1)−weakly amenable. Corollary 2-8. Let A be a Banach algebra and let A[2n] be dense in A. Suppose that AB∗∗ and B∗∗A are subset of B. Then the following are equivalent. 8 (1) H 1(A, B∗) = 0. (2) H 1(A, B(2n+1)) = 0 for some n ≥ 0. (3) H 1(A, B(2n+1)) = 0 for each n ≥ 0. Corollary 2-9 [6]. Let A be a weakly amenable Banach algebra such that A is an ideal in A∗∗. Then A is (2n + 1) − weakly amenable for each n ≥ 0. Proof. By using Proposition 1.3 from [6] and proceeding theorem, result is hold. (cid:3) Let B be a dual Banach algebra and dimB < ∞. Then by using Proposition 2.6.24 from [5], we know that N (B), the collection of all operators from B into B, is an ideal in N (B)∗∗. By using Corollary 4.3.6 from [21], N (B) is amenable, and so it also is weakly amenable. Consequently, by using the proceeding corollary, N (B) is (2n + 1) − weakly amenable for every n ≥ 0. We know that every von Neumann algebra A is an weakly amenable Banach algebra, see [21]. Now if A is an ideal in its second dual, A∗∗, then by using proceeding corol- lary, A is (2n + 1) − weakly amenable Banach algebra for each n ≥ 0. Assume that A and B are Banach algebra. Then A ⊕ B , with norm k (a, b) k=k a k + k b k, and product (a1, b1)(a2, b2) = (a1a2, b1b2) is a Banach algebra. It is clear that if X is a Banach A and B − bimodule, then X is a Banach A ⊕ B − bimodule. In the following, we investigated the relationships between the cohomological group of A ⊕ B and cohomological groups of A and B. Theorem 2-10. Suppose that A and B are Banach algebras. Let X be a Banach A and B −bimodule. Then, H 1(A⊕B, X) = 0 if and only if H 1(A, X) = H 1(B, X) = 0. Proof. Suppose that H 1(A ⊕ B, X) = 0. Assume that D1 ∈ Z 1(A, X) and D2 ∈ Z 1(B, X). Take D = (D1, D2). Then for every a1, a2 ∈ A and b1, b2 ∈ B, we have D((a1, b1)(a2, b2)) = D(a1a2, b1b2) = (D1(a1a2), D2(b1b2)) = (a1D1(a2) + D1(a1)a2, b1D2(b2) + D2(b1)b2) = (a1D1(a2), b1D2(b2)) + (D1(a1)a2 + D2(b1)b2) = (a1, b1)(D1(a2), D2(b2)) + (D1(a1), D2(b1))(a2, b2) = (a1, b1)D(a2, b2) + D(a1, b1)(a2, b2). It follows that D ∈ Z 1(A ⊕ B, X). Since H 1(A ⊕ B, X) = 0, there is x ∈ X such that D = δx where δx ∈ N 1(A ⊕ B, X). Since δx = (δ1 x ∈ N 1(B, X), we have D1 = δ1 x. Thus we have H 1(A, X) = H 1(B, X) = 0. For the converse, take A as an ideal in A ⊕ B, and so by using Proposition 2.8.66 from [5], proof hold. x ∈ N 1(A, X) and δ2 x and D2 = δ2 x, δ2 x) where δ1 (cid:3) 9 Let G be a locally compact group and X be a Banach L1(G) − bimodule. Then by [6, pp.27 and 28], X ∗∗ is a Banach L1(G)∗∗ − bimodule. Since L1(G)∗∗ = LU C(G)∗ ⊕ LU C(G)⊥, by using proceeding theorem, we have H 1(L1(G)∗∗, X ∗∗) = 0 if and only if H 1(LU C(G)∗, X ∗∗) = H 1(LU C(G)⊥, X ∗∗) = 0. On the other hand, we know that L1(G)∗∗ = L1(G) ⊕ C0(G)⊥. By [15], we know that, H 1(L1(G), L∞(G)) = 0. By using proceeding theorem, if and only if H 1(C0(G)⊥, L∞(G)) = 0. H 1(L1(G)∗∗, L∞(G)) = 0, Corollary 2-11. i) Suppose that A and B are Banach algebras. Let X be a Banach A and B−bimodule. Then A ⊕ B is an amenable Banach algebra if and only if A and B are amenable Ba- nach algebras. ii) Let A be a Banach algebra and n ≥ 1. Then H 1(⊕n is weakly amenable. i=1A, A∗) = 0 if and only if A Assume that B is a Banach A − bimodule. In the following, we will study the relationships between two cohomological groups H 1(A, B∗) and H 1(A∗∗, B∗∗∗). If H 1(A, B∗) = 0, we want to know that dose H 1(A∗∗, B∗∗∗) = 0 and when its con- verse hold? If we give an answer to these questions, then we want to establish the results of them in the weak amenability of Banach algebras. As some applications of these discussion in the group algebras, for a compact group G, we show that every weak∗ − to − weak continuous derivation from L1(G)∗∗ into M (G) is inner. Theorem 2-12. Let B be a Banach A − bimodule and suppose that every derivation from A into B∗ is weakly compact. If Z ℓ A∗∗(B∗∗) = B∗∗ and H 1(A∗∗, B∗∗∗) = 0, then H 1(A, B∗) = 0. Proof. Suppose that D ∈ Z 1(A, B∗). Since D : A → B∗ is weakly compact D′′(A∗∗) ⊆ w∗ → x′′ B∗. Assume that a′′, x′′ ∈ A∗∗ and (aα)α, (xβ)β ⊆ A such that aα in A∗∗. Then, since Z ℓ A∗∗(B∗∗) = B∗∗, for every b′′ ∈ B∗∗, we have w∗ → a′′ and xβ lim α < aαD′′(x′′), b′′ >= lim α < D′′(x′′), b′′aα >=< b′′a′′, D′′(x′′) > =< b′′, a′′D′′(x′′) >=< a′′D′′(x′′), b′′ > . Consequently, we have D′′(a′′x′′) = lim α lim β D(aαxβ ) = lim α (D(aα)xβ + aαD(xβ )) lim β = D′′(a′′)x′′ + a′′D′′(x′′). It follows that D′′ ∈ Z 1(A∗∗, B∗∗∗). Since H 1(A∗∗, B∗∗∗) = 0, there is b′′′ ∈ B∗∗∗ such that D′′ = δb′′′ . Then we have D = D′′ A= δb′′′ A. So for every a ∈ A, we have D(a) = δb′′′ (a) = ab′′′ − b′′′a. If we take b′ = b′′′ B, then D(a) = ab′ − b′a = δb′(a). 10 We conclude that H 1(A, B∗) = 0. (cid:3) Let A be a Banach algebra, and let B be a Banach A − bimodule. A continuous derivation D : A → B is approximately inner [resp. weakly approximately inner], if there exists a bounded net (bα)α ⊆ B such that D(a) = limα(abα − bαa) in B [resp. D(a) = w − limα(abα − bαa) in B]. w∗ Assume that (b′ → b′′′ in B∗∗∗. Then by assumptions of pro- ceeding theorem, for every a ∈ A and for every derivation from A into B∗, we have D(a) = D′′ A (a) = δb′′′ A (a) = ab′′′ − b′′′a = w∗ − limα(ab′ αa) in B∗∗∗. Since αa) in B∗, and so D is weakly approximately D(a) ∈ B∗, D(a) = w − limα(ab′ inner in B. α)α ⊆ B∗ such that b′ α α − b′ α − b′ Theorem 2-13. Let B be a Banach A − bimodule and suppose that D′′(A∗∗) ⊆ B∗. If Z ℓ A∗∗(B∗∗) = B∗∗ and H 1(A∗∗, B∗∗∗) = 0, then H 1(A, B∗) = 0. Proof. Proof is similar to proceeding theorem. (cid:3) A functional a′ in A∗ is said to be wap (weakly almost periodic) on A if the map- In [20], Pym showed that this ping a → a′a from A into A∗ is weakly compact. definition to the equivalent following condition 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 denoted by wap(A). Also we have a′ ∈ wap(A) if and only if < a′′b′′, a′ >=< a′′ob′′, a′ > for every a′′, b′′ ∈ A∗∗. Let B be a Banach left A − module. Then, b′ ∈ B∗ is said to be left weakly almost periodic functional if the set {π∗ ℓ (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 [20], b′ ∈ wapℓ(B) is equivalent to the following (a′′, b′′), b′ >=< πt∗∗∗t (a′′, b′′), b′ > < π∗∗∗ ℓ ℓ 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∗∗}. Corollary 2-14. Let B be a Banach A − bimodule and let every derivation D : A → B∗, satisfies D′′(A∗∗) ⊆ wapℓ(B). If H 1(A∗∗, B∗∗∗) = 0, then 11 H 1(A, B∗) = 0. In Corollary 2-14, if we take B = A, then we obtain Theorem 2.1 from [10]. Corollary 2-15. Let B be a Banach A − bimodule and suppose that for every derivation D : A → B∗, we have D′′(A∗∗)B∗∗ ⊆ A∗. If H 1(A∗∗, B∗∗∗) = 0, then H 1(A, B∗) = 0. Theorem 2-16. Let B be a Banach A − bimodule and suppose that AA∗∗ ⊆ A and B∗∗A = B∗∗. If H 1(A∗∗, B∗∗∗) = 0, then H 1(A, B∗) = 0. α)α ⊆ A∗∗ such that a′′ α w∗ → a′′ in A∗∗. Assume that b′′ ∈ B∗∗. Since Proof. Let (a′′ B∗∗A = B∗∗, there are a ∈ A and y′′ ∈ B∗∗ such that b′′ = y′′a. We know that α)α ⊆ A and aa′′ ∈ A, we a ∈ A ⊆ Z ℓ have aa′′ α B∗∗(A∗∗), and so aa′′ w→ aa′′ in A. Then for every a′′ ∈ A∗∗, we have the following equalities w∗ → aa′′ in A∗∗. Since (aa′′ α < D′′(x′′)b′′, a′′ α >→< D′′(x′′)y′′, aa′′ > α >=< D′′(x′′)y′′a, a′′ α >=< D′′(x′′)y′′, aa′′ =< D′′(x′′)b′′, a′′ > . It follows that D′′(x′′)b′′ ∈ (A∗∗, weak∗)∗ = A∗. So, by using Corollary 2-15, proof is hold. (cid:3) Theorem 2-17. Let B be a Banach A−bimodule and let every derivation D : A∗∗ → B∗ is weak∗ − to − weak∗ continuous. Then if Z ℓ B∗∗(A∗∗) = A∗∗ and H 1(A, B∗) = 0, it follows that H 1(A∗∗, B∗) = 0. Proof. Let D : A∗∗ → B∗ be a derivation. Then D A: A → B∗ is a derivation. Since H 1(A, B∗) = 0, there is b′ ∈ B∗ such that D A= δb′ . Suppose that a′′ ∈ A∗∗ and (aα)α ⊆ A such that aα D(a′′) = w∗ − lim α w∗ → a′′ in A∗∗. Then D A (aα) = w∗ − lim α δb′(aα) = w∗ − lim α (aαb′ − b′aα) = a′′b′ − b′a′′. Now, we show that b′a′′ ∈ B∗. Assume that (b′′ Then since Z ℓ B∗∗(A∗∗) = A∗∗, we have β)β ∈ B∗∗ such that b′′ = w∗ − limβ b′′ β. < b′a′′, b′′ β >=< a′′b′′ β, b′ >→< a′′b′′, b′ >=< b′a′′, b′′ > . Thus, we conclude that b′a′′ ∈ (B∗∗, weak∗)∗ = B∗, and so H 1(A∗∗, B∗) = 0. (cid:3) Corollary 2-18. Let A be a Arens regular Banach algebra and let every derivation D : A∗∗ → A∗ is weak∗ − to − weak∗ continuous. Then if A is weakly amenable, it follows that H 1(A∗∗, A∗) = 0. 12 Let B be a dual Banach space, with predual X and suppose that X ⊥ = {x′′′ : x′′′ X = 0 where x′′′ ∈ X ∗∗∗} = {b′′ : b′′ X = 0 where b′′ ∈ B∗∗}. Then the canonical projection P : X ∗∗∗ → X ∗ gives a continuous linear map P : B∗∗ → B. Thus, we can write the following equality B∗∗ = X ∗∗∗ = X ∗ ⊕ kerP = B ⊕ X ⊥, as a direct sum of Banach A − bimodule. Theorem 2-19. Let B be a Banach A − bimodule and every derivation A∗∗ into B is weak∗ − to − weak continuous. Let A∗∗B, BA∗∗ ⊆ B. Then the following assertions are hold. (1) If H 1(A, B) = 0, then H 1(A∗∗, B) = 0. (2) Suppose that A has a LBAI. Let B has a predual X and let AB∗, B∗A ⊆ X. Then, if H 1(A, B) = 0, it follows that H 1(A∗∗, B∗∗) = 0. Proof. (1) Proof is similar to Theorem 2-17. (2) Set B∗∗ = B ⊕ X ⊥. Then we have H 1(A∗∗, B∗∗) = H 1(A∗∗, B) ⊕ H 1(A∗∗, X ⊥). It is clear that D ∈ Z 1(A∗∗, X ⊥). Assume that w∗ → x′′ on A∗∗. a′′, x′′ ∈ A∗∗ and (aα)α, (xβ)β ⊆ A such that aα Since AB∗, B∗A ⊆ X, for every b′ ∈ B∗, by using the weak∗ − to − weak Since H 1(A, B) = 0, by part (1), H 1(A∗∗, B) = 0. Now let eD ∈ Z 1(A∗∗, X ⊥) and we take D = eD A. continuity of eD, we have w∗ → a′′ and xβ < D(aαxβ), b′ > < eD(a′′x′′), b′ >= lim α lim β lim β lim β = lim α < D(aα)xβ, b′ > + lim α < D(aα), xβ b′ > + lim α lim β = lim α = lim α < (D(aα)xβ + aαD(xβ )), b′ > < aαD(xβ ), b′ > lim β < D(xβ )), b′aα > lim β Since A has a LBAI, A∗∗ has a left unit e′′ with respect to the first Arens product. Then D(x′′) = D(e′′x′′) = 0, and so D = 0. = 0. Assume that G is a compact group. Then we know that L1(G) is M (G)− bimodule and L1(G) is an ideal in the second dual of M (G), M (G)∗∗. By using [18, Corollary 1.2], we have H 1(L1(G), M (G)) = 0. Then by using proceeding corollary, every weak∗ − to − weak continuous derivation from L1(G)∗∗ into M (G) is inner. Consider the algebra c0 = (c0, .) is the collection of all sequences of scalars that convergence to 0, with the some vector space operations and norm as ℓ∞. We know that c0 is a C∗ − algebra and Since enery C∗ − algebra is weakly amenable, it follows (cid:3) 13 that c0 is weakly amenable. Then by using proceeding theorem, every weak∗ − to − weak continuous derivation from ℓ∞ into ℓ1 is inner. Now let B be a Banach A − bimodule and for every derivation D : A∗∗ → B, we have D(A∗∗) ⊆ D A (A) w w . Assume that D ∈ Z 1(A∗∗, B). Take eD = D A. Then If H 1(A, B) = 0, then there is a b ∈ B such that eD = δb. Since , for every a′′ ∈ A∗∗, there is (aα)α ⊆ A such that eD(aα) w→ α eD(aα) = w − lim δb(aα) = w − lim α (aαb − baα) in B. w α Thus, for every derivation D : A∗∗ → B, by assumption D(A∗∗) ⊆ D A (A) H 1(A, B) = 0, we can write D(a′′) = limα(aαb − baα) in B. and eD ∈ Z 1(A, B). D(A∗∗) ⊆ D A (A) D(a′′). Thus we have the following equality. D(a′′) = w − lim Theorem 2-20. Let B be a Banach A − bimodule and A has a LBAI. Suppose that AB∗∗, B∗∗A ⊆ B and every derivation A∗∗ into B∗ is weak∗ − to − weak∗ continuous. Then if H 1(A, B∗) = 0, it follows that H 1(A∗∗, B∗∗∗) = 0. Proof. Take B∗∗∗ = B∗ ⊕ B⊥ where B⊥ = {b′′′ ∈ B∗∗∗ : b′′′ B= 0}. Then we have H 1(A∗∗, B∗∗∗) = H 1(A∗∗, B∗) ⊕ H 1(A∗∗, B⊥). Since H 1(A, B∗) = 0, it is similar to Theorem 2.17 that H 1(A∗∗, B∗) = 0. It suffices for the result to show that H 1(A∗∗, B⊥) = 0. Without loss generality, let (eα)α ⊆ A w∗ → e′′ in A∗∗ where e′′ is a left unit for A∗∗ with be a LBAI for A such that eα respect to the first Arens product. Let a′′ ∈ A∗∗ and suppose that (aβ)β ⊆ A such w∗ → a′′ in A∗∗. Then, if we take D ∈ Z 1(A∗∗, B⊥), for every b′′ ∈ B∗∗, by that aβ using the weak∗ − to − weak∗ continuity of D, we have < D(a′′), b′′ >=< D(e′′a′′), b′′ >= lim α lim β < (D(eαaβ), b′′ > = lim α lim β < (D(eα)aβ + eαD(aβ)), b′′ >= lim α < D(eα), aβb′′ > < eαD(aβ), b′′ >= lim α lim β lim β lim β + lim α < D(eα)aβ, b′′ > It follows that D = 0, and so the result is hold. + lim α lim β < D(aβ), b′′eα >= 0. Corollary 2-21. Assume that A is a Banach algebra with LBAI. Suppose that A is two-sided ideal in A∗∗ and every derivation D : A∗∗ → A∗∗∗ is weak∗ − to − weak∗ continuous. Then if A is weakly amenable, it follows that A∗∗ is weakly amenable. (cid:3) Let n ≥ 0 and suppose that A is (2n + 1) − weakly amenable. By condations of the proceeding corollary and by using [6, Corollary 1.14], we conclude that A∗∗ is weakly amenable. Assume that G is a locally compact group. We know that L1(G) is weakly amenable Banach algebra, see [15]. So by using proceeding corollary, every weak∗ − to − weak∗ 14 continuous derivation from L1(G)∗∗ into L1(G)∗∗∗ is inner. w∗ (A, B) = 0. Let A be a Banach Algebra. A dual Banach A − bimodule B is called normal if, for each b ∈ B the map a → ax and a → xa from A into B is weak∗ − to − weak∗ continuous. A dual Banach algebra A is Connes-amenable if, for every normal, dual Banach A − bimodule B, every weak∗ − to − weak∗ continuous derivation D ∈ Z 1(A, B) is inner. Then we write H 1 A Banach algebra A is called super-amenable if H 1(A, B) = 0 for every Banach A − bimodule B. It is clear that if A is super-amenable, then A is amenable. If B is a Banach algebra and it is consisting a closed subalgebra A such that A is weakly dense in B. Then, it is easy to show that A is a super-amenable if and only if B is a super-amenable. In the following, we will study some problems on the Connes-amenability of Banach al- gebras. In the Theorem 2-19, with some conditions, we showed that H 1(A∗∗, B∗∗) = 0. Now, in the following theorem, by using some cocitions and super-amenability of Ba- nach algebra A, we conclude that H 1 By Theorem 4.4.8 from [22], we know that if A is Arens regular Banach algebra and A is an ideal in A∗∗, then A is amenable if and only if A∗∗ is Connes-amenable. Now in the following by using of super-amenability of Banach algebra A, we show that A∗∗ is Connes-amenable. In the Theorem 2-19, for a Banach A − bimodule B, with some conditions, we showed that H 1(A∗∗, B∗∗) = 0. In the following, by using some new conditions, we show that H 1 w∗(A∗∗, B∗∗) = 0 and as results in group algebra, for a amenable compact group G, we show that H 1 w∗ (A∗∗, B∗∗) = 0. w∗(L1(G)∗∗, M (G)∗∗∗) = 0. Theorem 2-22. Suppose that A is supper-amenable Banach algebra. Then we have the following statements. (1) A∗∗ is Connes-amenable Banach algebra with respect to the first Arens prod- uct. (2) If A has a LBAI and for every Banach A − bimodule B with predual X, we have AB∗, B∗A ⊆ X, then H 1 w∗(A∗∗, B∗∗) = 0. Proof. (1) Let B be a normal dual Banach A − bimodule B and let be D ∈ Z 1(A∗∗, B∗∗) be weak∗ − to − weak∗ continuous derivation. Suppose that eD = D A. Then we have eD ∈ Z 1(A, B). Since A is super-amenable, there is element b ∈ B such that eD = δb. Suppose that a′′ ∈ A∗∗ and (aα)α ⊆ A such → a′′ in A∗∗. Since eD is weak∗ − to − weak∗ continuous derivation, that aα we have w∗ D(a′′) = D(w∗ − lim α aα) = w∗ − lim α D(aα) = w∗ − lim = w∗ − lim α It follows that H 1 w∗ (A∗∗, B) = 0. δb(aα) = w∗ − lim α α eD(aα) (aαb − baα) = a′′b − ba′′. (2) By using part (1), proof is similar to Theorem 2-19 (2). 15 (cid:3) Theorem 2-23. Suppose that A is an amenable Banach algebra with LBAI. If for every Banach A − bimodule B, we have AB∗∗, B∗∗A ⊆ B, then H 1 w∗(A∗∗, B∗∗∗) = 0. Proof. Proof is similar to the proceeding theorem. (cid:3) Suppose that G is a compact group. Then L1(G) is an ideal in M (G)∗∗. If G is amenable, then by using proceeding theorem, we conclude the following equality. H 1 w∗(L1(G)∗∗, M (G)∗∗∗) = 0. Corollary 2-24. Assume that A is a weakly amenable Banach algebra with LBAI. Then if A is an ideal in A∗∗, it follows that H 1 w∗(A∗∗, A∗∗∗) = 0. Example 2-25. Assume that G is a compact group. Then we know that L1(G) has a bounded approximate identity [=BAI] and L1(G) is two-sided ideal in L1(G)∗∗. We know that L1(G) is weakly amenable. Hence it is clear H 1 w∗ (L∞(G)∗, L∞(G)∗∗) = 0. Let B be a Banach A − bimodule. Then for every b′′ ∈ B∗∗, we define Lb′′ (a′′) = b′′a′′ and Rb′′ (a′′) = a′′b′′, for every a′′ ∈ A∗∗. These are the operation of left and right multiplication by b′′ on A∗∗. Let B be a Banach A − bimodule. We say that B is a left [resp. right] factors with respect to A, if BA = B [resp. AB = B]. In the following, for super-amenable Banach algebra A and a Banach A − bimodule C, we have represented for every derivation D from A into C, and as a result in group algebras, for amenable locally compact group G, we show that every derivation D from L1(G) into L∞(G) is in the form D = Lf where f ∈ L∞(G). Theorem 2-26. Assume that A is a super-amenable Banach algebra and B is a Banach A − bimodule such that B factors on the left [resp. right]. Then there is a Banach A − bimodule (C, .) such that C = B and Z 1(A, C) = {LD′′(e′′) : D ∈ Z 1(A, C)} [resp. Z 1(A, C) = {RD′′(e′′) : D ∈ Z 1(A, C)}] where e′′ is a right [resp. left] identity for A∗∗. 16 Proof. Certainly, every super-amenable Banach algebra is amenable. So, by using [22, Propostion 2.2.1], A has a BAI such as (eα)α. Since BA = B, for every b ∈ B, there are y ∈ B and a ∈ A such that b = ya. Then we have lim α beα = lim α (ya)eα = lim α y(aeα) = ya = b. It follows that B has RBAI as (eα)α ⊆ A. Without loss generality, let e′′ be a right unit for A∗∗ such that eα Take C = B and for every a ∈ A and x ∈ C, we define a.x = 0 and x.a = xa. It is clear that (C, .) is a Banach A − bimodule. Suppose that D ∈ Z 1(A, C). Then there is a c ∈ C such that D = δc. Then for every a ∈ A, we have w∗ → e′′ in A∗∗. D(a) = δc(a) = a.c − c.a = −ca. Suppose that x ∈ C and x′ ∈ C∗. Since C.A = C, there is t ∈ C and s ∈ A such that x = t.s = ts. Then we have < x′, xeα >=< x′, tseα >=< x′t, seα >→< x′t, s >=< x′, x > . It follows that xeα we have w→ x in C. Since D′′ is a weak∗ − to− weak∗ continuous derivation, D′′(e′′) = D′′(w∗ − lim α eα) = w∗ − lim α D′′(eα) = w − lim α D(eα) Thus we conclude that D(a) = D′′(e′′)a for all a ∈ A. It follows that = w − lim α (ceα) = −c. D = LD′′(e′′). On the other hand, since for every derivation D ∈ Z 1(A, C), we have LD′′(e′′) ∈ Z 1(A, C), the result is hold. (cid:3) Corollary 2-27. Suppose that A is an amenable Banach algebra and B is a Ba- nach A − bimodule such that AB∗ = B∗ [resp. B∗A = B∗]. Then there is a Banach A − bimodule (C, .) such that C = B and Z 1(A, C∗) = {LD′′(e′′) : D ∈ Z 1(A, C∗)} [resp. Z 1(A, C∗) = {RD′′(e′′) : D ∈ Z 1(A, C∗)}] where e′′ is a right [resp. left] identity for A∗∗. Corollary 2-28. Suppose that A is a super-amenable Banach algebra and B is a Banach A − bimodule. Then there is a Banach A − bimodule (C, .) such that C = B and Z 1(A, C) = {Lc : c ∈ C} [resp. Z 1(A, C) = {Rc : c ∈ C}]. Example 2-29. (1) Let G be an amenable locally compact group. Then there is a Banach L1(G)− f ∈ bimodule such as (L∞(G), .) such that Z 1(L1(G), L∞(G)) = {Lf : L∞(G)}. (2) Let G be locally compact group and H 1(M (G), L1(G)) = 0. Then there is a Banach M (G) − bimodule (L1(G), .) such that Z 1(M (G), L1(G)) = {LD′′(e′′) : D ∈ Z 1(M (G), L1(G))} [or Z 1(M (G), L1(G)) = {RD′′(e′′) : D ∈ Z 1(M (G), L1(G))}] where e′′ is a left [or right] identity for L1(G)∗∗. 17 Theorem 2-30. Let B be a Banach A − bimodule and suppose that D : A → B∗ is a derivation. If D′′ : A∗∗ → B∗∗∗ is a derivation and B∗ ⊆ D′′(A∗∗), then Z ℓ A∗∗(B∗∗) = B∗∗. Proof. Since D′′ : A∗∗ → B∗∗∗ is a derivation, by [19, Theorem 4.2], D′′(A∗∗)B∗∗ ⊆ α)α ⊆ A∗∗ such that B∗. Due to B∗ ⊆ D′′(A∗∗), we have B∗B∗∗ ⊆ B∗. Let (a′′ w∗ → a′′ in A∗∗. Assume that b′′ ∈ B∗∗. Then for every b′ ∈ B∗, since b′b′′ ∈ B∗, we a′′ α have < b′′a′′ α, b′ >=< a′′ αb′, b′′ >→< a′′, b′b′′ >=< b′′a′′, b′ > . Thus b′′a′′ α w∗ → b′′a′′ in B∗∗, and so b′′ ∈ Z ℓ A∗∗(B∗∗). (cid:3) Corollary 2-31. Assume that A is a Banach algebra and D : A → A∗ is a derivation such that A∗ ⊆ D′′(A∗∗). Then if the linear mapping D′′ : A∗∗ → A∗∗∗ is a derivation, it follows that A is Arens regular. Lemma 2-32. Let B be a Banach left A − module and B∗∗ has a LBAI with respect to A∗∗. Then B∗∗ has a left unit with respect to A∗∗. Proof. Assume that (e′′ suppose that there is e′′ ∈ A∗∗ such that e′′ α and b′ ∈ B∗, we have α)α ⊆ A∗∗ is a LBAI for B∗∗. By passing to a subnet, we may w∗ → e′′ in A∗∗. Then for every b′′ ∈ B∗∗ < π∗∗∗ ℓ (e′′, b′′), b′ >=< e′′, π∗∗ ℓ (b′′, b′) >= lim α < e′′ α, π∗∗ ℓ (b′′, b′) > = lim α < π∗∗∗ ℓ (e′′ α, b′′), b′ >=< b′′, b′ > . It follows that π∗∗∗ ℓ (e′′, b′′) = b′′. (cid:3) Corollary 2-33. Let A be a Banach algebra and A∗∗ has a LBAI. Then A∗∗ has a left unit with respect to the first Arens product. Theorem 2-34. Let A be a left strongly Arens irregular and suppose that A∗∗ is an amenable Banach algebra. Then we have the following assertions. (1) A has an identity. (2) If A is a dual Banach algebra, then A is reflexive. 18 Proof. (1) Since A∗∗ is amenable, it has a BAI. By using the proceeding corol- lary, A∗∗ has an identity e′′. So, the mapping x′′ → e′′x′′ = x′′ is weak∗ − to − weak∗ continuous from A∗∗ into A∗∗. It follows that e′′ ∈ Z1(A∗∗) = A. Consequently, A has an identity. (2) Assume that E is predual Banach algebra for A. Then we have A∗∗ = A⊕E⊥. Since A∗∗ is amenable, by using Theorem , A is amenable, and so E⊥ is amenable. Thus E⊥ has a BAI such as (e′′ α)α ⊆ E⊥. Since E⊥ is a closed and weak∗ − closed subspace of A∗∗, without loss generality, there is e′′ ∈ E⊥ such that e′′ α w∗ → e′′ and e′′ α k k → e′′ Then e′′ is a left identity for E⊥. On the other hand, for every x′′ ∈ E⊥, since E⊥ is an ideal in A∗∗, we have x′′e′′ ∈ E⊥. Thus for every a′ ∈ A∗, we have < x′′e′′, a′ >= lim α < (x′′e′′)e′′ α, a′ >= lim α < x′′(e′′e′′ α), a′ > = lim α < x′′e′′ α, a′ >=< x′′, a′ > . It follows that x′′e′′ = x′′, and so e′′ is a right identity for E⊥. Consequently, e′′ is a two-sided identity for E⊥. Now, let a′′ ∈ A∗∗. Then we have the following equalities. e′′a′′ = (e′′a′′)e′′ = e′′(a′′e′′) = a′′e′′. Thus we have e′′ ∈ Z1(A∗∗) = A. It follows that e′′ = 0, and so E⊥ = 0. Consequently, we have A∗∗ = A. Let G be a locally compact group. Then if L1(G)∗∗ or M (G)∗∗ are amenable, by using proceeding theorem part (1) and (2), respectively, we conclude that G is finite group. (cid:3) 19 3. Weak amenability of Banach algebras In this section, for Banach A − module B, we introduce some new concepts as lef t − weak∗ − to − weak convergence property [ = Lw∗wc−property] and right − weak∗ − to − weak convergence property [ = Rw∗wc−property] with respect to A and we show that if A∗ and A∗∗, respectively, have Rw∗wc−property and Lw∗wc−property and A∗∗ is weakly amenable, then A is weakly amenable. We also show the relations between a derivation D : A → A∗ and this new concepts. Now in the following, for left and right Banach A − module B, we define, respectively, Lw∗wc−property and Rw∗wc−property concepts with some examples. Definition 3-1. Assume that B is a left Banach A − bimodule. We say that b′ ∈ B∗ has lef t−weak∗ −to−weak convergence property [ = Lw∗wc−property] with respect to A, if for every bounded net (aα)α ⊆ A, b′aα When every b′ ∈ B∗ has Lw∗wc-property with respect to A, we say that B∗ has Lw∗wc−property with respect to A. The definition of right−weak∗−to−weak convergence property [= Rw∗wc−property] with respect to A is similar and if b′ ∈ B∗ has lef t − weak∗ − to − weak convergence property and right−weak∗−to−weak convergence property, then we say that b′ ∈ B∗ has weak∗ − to − weak convergence property [= w∗wc−property]. w∗ → 0 implies b′aα w→ 0. Example 3-2 . (1) Every reflexive Banach A − module has w∗wc−property. (2) Let Ω be a compact group and suppose that A = C(Ω) and B = M (Ω). Let w∗ → 0, then for each a ∈ A, we (aα)α ⊆ A and µ ∈ B. Suppose that µ ∗ aα have < µ ∗ aα, a >=< µ, aα ∗ a >=ZΩ (aα ∗ a)dµ → 0. We set a = 1Ω . Then µ(aα) → 0. Now let b′ ∈ B∗. Then < b′, µ ∗ aα >=< aαb′, µ >=ZΩ aαb′dµ ≤k b′ k ZΩ aαdµ =k b′ k µ(aα) → 0. It follows that µ ∗ aα to A. w→ 0, and so that µ has Rw∗wc−property with respect Theorem 3-3. Let A be a Banach algebra and suppose that A∗ and A∗∗, respectively, have Rw∗wc−property and Lw∗wc−property with respect to A. If A∗∗ is weakly amenable, then A is weakly amenable. Proof. Assume that a′′ ∈ A∗∗ and (aα)α ⊆ A such that aα a′ ∈ A∗, we have aαa′ w∗ A, aαa′ w→ a′′a′ in A∗. Then for every x′′ ∈ A∗∗, we have w∗ → a′′. Then for each → a′′a′ in A∗. Since A∗ has Rw∗wc−property with respect to < x′′aα, a′ >=< x′′, aαa′ >→< x′′, a′′a′ >=< x′′a′′, a′ > . 20 w∗ → x′′a′′. Since A∗∗ has Lw∗wc−property with respect to A, It follows that x′′aα w→ x′′a′′. If D : A → A∗ is a bounded derivation, we extend it to a bounded linear x′′aα mapping D′′ from A∗∗ into A∗∗∗. Suppose that a′′, b′′ ∈ A∗∗ and (aα)α, (bβ)β ⊆ A such that aα w→ x′′a′′ for every x′′ ∈ A∗∗, we have w∗ → b′′. Since x′′aα w∗ → a′′ and bβ < D′′(b′′), x′′aα >=< D′′(b′′), x′′a′′ > . lim α In the following we take limit on the weak∗ topologies. Thus we have lim α lim β D(aα)bβ = D′′(a′′)b′′. Consequently, it follows that D′′(a′′b′′) = lim α lim β D(aαbβ) = lim α lim β D(aα)bβ + lim α lim β aαD(bβ) = D′′(a′′)b′′ + a′′D′′(b′′). Since A∗∗ is weakly amenable, there is a′′′ ∈ A∗∗∗ such that D′′ = δa′′′ . Then we have D = D′′ A∗= δa′′′ A∗ . Hence for each x′ ∈ A∗, we have D = x′a′′′ − a′′′x′. Take a = a′′′ A∗ . It follows that H 1(A, A∗) = 0. (cid:3) Theorem 3-4. Let A be a Banach algebra and suppose that D : A → A∗ is a surjective derivation. If D′′ is a derivation, then we have the following assertions. (1) A∗ and A∗∗, respectively, have w∗wc−property and Lw∗wc−property with respect to A. (2) For every a′′ ∈ A∗∗, the mapping x′′ → a′′x′′ from A∗∗ into A∗∗ is weak∗ − to − weak continuous. (3) A is Arens regular. (4) If A has LBAI, then A is reflexive. Proof. (1) Since D is surjective, D′′ is surjective, and so by using [19, Theorem 2.2], we have A∗∗∗A∗∗ ⊆ D′′(A∗∗)A∗∗ ⊆ A∗. Suppose that a′′ ∈ A∗∗ and w∗ w∗ → x′a′′. → a′′. Then for each x′ ∈ A∗, we have x′aα (aα)α ⊆ A such that aα Since A∗∗∗A∗∗ ⊆ A∗, x′a′′ ∈ A∗. Then for every x′′ ∈ A∗∗, we have < x′′, x′aα >=< x′′x′, aα >→< a′′, x′′x′ >=< x′a′′, x′′ >=< x′′, x′a′′ > . w→ x′a′′ in A∗. Thus by easy calculation, we conclude It follows that x′aα that x′ has Lw∗wc−property with respect to A. It is to similar that x′ has Rw∗wc−property with respect to A, and so A∗ has w∗wc−property. For next part, suppose that x′′′ ∈ A∗∗∗. Since A∗∗∗A∗∗ ⊆ A∗, x′′aα for each x′′ ∈ A∗∗. Then w∗ → x′′a′′ < x′′′, x′′aα >=< x′′′x′′, aα >→< x′′′x′′, a′′ >=< x′′′, x′′a′′ > . It follows that x′′aα has Lw∗wc−property with respect to A. w→ x′′a′′. Thus by easy calculation, we conclude that x′′ (2) Suppose that (a′′ α)α ⊆ A∗∗ and a′′ α w∗ → a′′. Let x′′ ∈ A∗∗. Then for every x′′′ ∈ A∗∗∗, since A∗∗∗A∗∗ ⊆ A∗, we have < x′′′, x′′a′′ α >=< x′′′x′′, a′′ α >→< x′′′x′′, a′′ >=< x′′′, x′′a′′ > . 21 It follows from (2). (3) (4) Let (eα)α ⊆ A be a BLAI for A. Then without loss generality, let e′′ be a w∗ → a′′. left unit for A∗∗ such that eα Then for every a′′′ ∈ A∗∗∗, since A∗∗∗A∗∗ ⊆ A∗, we have w∗ → e′′. Suppose that (a′′ α)α ⊆ A∗∗ and a′′ α < a′′′, a′′ α >=< a′′′, e′′a′′ α >=< a′′′e′′, a′′ α >→< a′′′e′′, a′′ >=< a′′′, a′′ > . It follows that a′′ α w→ a′′. Consequently A is reflexive. Corollary 3-5. Let A be a Banach algebra and suppose that D : A → A∗ is a surjective derivation. The following statement are equivalent. (1) A∗ and A∗∗, respectively, have Rw∗wc−property and Lw∗wc−property with respect to A. (2) For every a′′ ∈ A∗∗, the mapping x′′ → a′′x′′ from A∗∗ into A∗∗ is weak∗ − to − weak continuous. (cid:3) Problems. i) By notice to Theorem 2-19, dose the following assertions hold. (1) For compact group G, we have H 1(L1(G)∗∗, M (G)) = 0? (2) If c0 is weakly amenable, H 1(ℓ∞, ℓ1) = 0? ii) Suppose that S is a compact semigroup. Dose L1(S)∗ and M (S)∗ have Lw∗wc−property or Rw∗wc−property? References 1. R. E. Arens, The adjoint of a bilinear operation, Proc. Amer. Math. Soc. 2 (1951), 839-848. 2. W. G. Bade, P.C. Curtis and H.G. Dales, Amenability and weak amenability for Beurling and Lipschitz algebra, Proc. Lodon Math. Soc. 137, no.656, 1999. MR 99g: 46059. 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, Banach algebra and automatic continuity, Oxford 2000. 6. H. G. Dales, F. Ghahramani, N. Grønbaek Derivation into iterated duals of Banach algebras Studia Math. 128 1 (1998), 19-53. 7. H. G. Dales, A. Rodrigues-Palacios, M.V. Velasco, The second transpose of a derivation, J. London. Math. Soc. 2 64 (2001) 707-721. 8. N. Dunford, J. T. Schwartz, Linear operators.I, Wiley, New york 1958. 9. M. Eshaghi Gordji, M. Filali, Arens regularity of module actions, Studia Math. 181 3 (2007), 237-254. 10. M. Eshaghi Gordji, M. Filali, Weak amenability of the second dual of a Banach algebra, Studia Math. 182 3 (2007), 205-213. 11. F. Gourdeau, Amenability of Lipschits algebra, Math. Proc. Cambridge. Philos. Soc. 112 (1992), 581-588. 12. E. Hewitt, K. A. Ross, Abstract harmonic analysis, Springer, Berlin, Vol I 1963. 13. B.E. Johoson, Cohomology in Banach algebra, Mem. Amer. Math. Soc. 127, 1972. 22 14. B. E. Johoson, Derivation from L1(G) into L1(G) and L∞(G), Harmonic analysis. Luxembourg 1987, 191-198 Lecture Note in Math., 1359, Springer, Berlin, 1988. MR 90a:46122. 15. B. E. Johoson, Weak amenability of group algebra, Bull. Lodon. Math. Soc. 23(1991), 281-284. 16. A. T. Lau, V. Losert, On the second Conjugate Algebra of locally compact groups, J. London Math. Soc. 37 (2)(1988), 464-480. 17. A. T. Lau, A. Ulger, Topological center of certain dual algebras, Trans. Amer. Math. Soc. 348 (1996), 1191-1212. 18. V. Losert, The derivation problem for group algebra, Annals of Mathematics, 168 (2008), 221- 246. 19. 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. 20. J. S. Pym, The convolution of functionals on spaces of bounded functions, Proc. London Math Soc. 15 (1965), 84-104. 21. V. Runde, Lectures on the amenability, springer-verlag Berlin Heideberg NewYork. 22. A. Ulger, Some stability properties of Arens regular bilinear operators, Proc. Amer. Math. Soc. (1991) 34, 443-454. 23. A. Ulger, Arens regularity of weakly sequentialy compact Banach algebras, Proc. Amer. Math. Soc. 127 (11) (1999), 3221-3227. 24. P. K. Wong, The second conjugate algebras of Banach algebras, J. Math. Sci. 17 (1) (1994), 15-18. 25. Y. Zhing, Weak amenability of module extentions of Banach algebras, Trans. Amer. Math. Soc. 354 (10) (2002), 4131-4151. 26. Y. Zhing, Weak amenability of a class of Banach algebra, Cand. Math. Bull. 44 (4) (2001) 504-508.
1712.10065	1	1712	2017-12-28T21:51:46	Pointwise amenable, non-amenable Banach algebras	[ "math.FA" ]	It is shown that a pointwise amenable Banach algebra need not be amenable. This positively answer a question raised by Dales, Ghahramani and Loy.	math.FA	math	POINTWISE AMENABLE, NON-AMENABLE BANACH ALGEBRAS SARA BEHNAMIAN AND AMIN MAHMOODI Abstract. It is shown that a pointwise amenable Banach algebra need not be amenable. This positively answer a question raised by Dales, Ghahramani and Loy. Keywords: amenable, pointwise amenable, Connes amenable, character amenable. MSC 2010: Primary: 46H25; Secondary: 43A07. 1. Introduction Let A be a Banach algebra, and let E be a Banach A-bimodule. A derivation is a bounded linear map D : A −→ E satisfying D(ab) = Da · b + a · Db (a, b ∈ A). A Banach algebra A is amenable if for any Banach A-bimodule E, any derivation D : A −→ E ∗ is inner, that is, there exists η ∈ E ∗ with D(a) = adη(a) = a · η − η · a, a ∈ A. This powerful notion introduced by B. E. Johnson [5]. The pointwise variant of amenability introduced by H. G. Dales, F. Ghahramani and R. J. Loy, and appeared formally in [2]. A Banach algebra A is pointwise amenable at a ∈ A if for any Banach A-bimodule E, any derivation D : A −→ E ∗ is pointwise inner at a, that is, there exists η ∈ E ∗ with D(a) = adη(a). Further, A is pointwise amenable if it is pointwise amenable at each a ∈ A. For a Banach algebra A, recall that the projective tensor product Ab⊗A is a Banach A-bimodule in a natural way and the bounded linear map π : Ab⊗A −→ A defined by π(a ⊗ b) = ab, (a, b ∈ A) is a Banach A-bimodule homomorphism. Obviously, every amenable Banach algebra is pointwise amenable. For the con- verse, however, it has been open so far if there is a pointwise amenable Banach algebra which is not already amenable [2, Problem 5]. In this note, we give an illuminating example to show that the problem has an affirmative answer. 2. The results Let V be a Banach space, and let f ∈ V ∗ (the dual space of V) be a non-zero element such that f ≤ 1. Then V equipped with the product defined by ab := 1 2 SARA BEHNAMIAN AND AMIN MAHMOODI In general, Vf is a non- f (a)b for a, b ∈ V, is a Banach algebra denoted by Vf . commutative, non-unital Banach algebra without right approximate identity. One may see [3, 8] for more details and properties on this type of algebras. A small variation of standard argument in [6] gives the following characterization of pointwise amenability. Proposition 2.1. A Banach algebra A is pointwise amenable if and only if for each a ∈ A there exists a net (mα)α ⊆ Ab⊗A (depending on a) such that a·mα−mα·a −→ 0 and aπ(mα) −→ a. Theorem 2.2. Let V be a Banach space, and let f ∈ V ∗ be a non-zero element such that f ≤ 1. Then Vf is pointwise amenable, but not amenable. Proof. For a ∈ Vf , we consider m := P∞ and cn := a for all n. Then n=1 bn ⊗ cn ∈ Vf b⊗Vf with P∞ n=1 f (bn) = 1 f (a) a · m = ∞X abn ⊗ cn = f (a) ∞X bn ⊗ cn = f (a)m n=1 n=1 and similarly m · a = f (a)m, so that a · m = m · a. Next, π(m) = P∞ (P∞ n=1 f (bn))a = 1 f (a) a. Hence n=1 bncn = aπ(m) = 1 f (a) a2 = 1 f (a) f (a)a = a . Therefore, Vf is pointwise amenable by Proposition 2.1. Because of the lack of a bounded approximate identity, Vf is not amenable. (cid:3) Remark 2.3. Similar to [2], we may define pointwise contractible Banach algebras (see for instance [1] for the definition of contractible Banach algebras). Then The- orem 2.2 says that Vf is pointwise contractible. Notice that Vf is not contractible, because it has no identity. Let A be a Banach algebra. A Banach A-bimodule E is dual if there is a closed submodule E∗ of E ∗ such that E = (E∗)∗. A Banach algebra A is dual if it is a dual Banach space such that multiplication is separately continuous in the w∗- topology. For a dual Banach algebra A, a dual Banach A-bimodule E is normal if the module actions of A on E are w∗-continuous. The notion of Connes amenability for dual Banach algebras, which is another modification of the notion of amenability systematically introduced by V. Runde [9]. A dual Banach algebra A is Connes amenable if every w∗-continuous derivation from A into a normal, dual Banach A- bimodule is inner. The pointwise variant of Connes amenability introduced in [11]. A dual Banach algebra A is pointwise Connes amenable at a ∈ A if for every normal, dual Banach A-bimodule E, every w∗-continuous derivation D : A → E is POINTWISE AMENABLE, NON-AMENABLE BANACH ALGEBRAS 3 pointwise inner at a. Moreover, A is pointwise Connes amenable if it is pointwise Connes amenable at each a ∈ A. Theorem 2.4. Let V be a dual Banach space, and let f ∈ V ∗ be a w∗-continuous non-zero element such that f ≤ 1. Then Vf is a pointwise Connes amenable, non-Connes amenable dual Banach algebra. Proof. Since f is w∗-continuous, the multiplication on Vf is separately w∗-continuous, and so Vf is a dual Banach algebra. As Vf is pointwise amenable by Theorem 2.2, it is automatically pointwise Connes amenable. Since Vf does not have an identity, it is not Connes amenable by [9, Proposition 4.1]. (cid:3) The following is [11, Theorem 2.6]. Theorem 2.5. Every commutative, pointwise Connes amenable dual Banach alge- bra has an identity. Remark 2.6. The fact that Vf is non-commutative pointwise Connes amenable but non-unital, shows that the commutativity in Theorem 2.5 can not be dropped. Proposition 2.7. Let V be a Banach space, and let f ∈ V ∗ be a non-zero element such that f ≤ 1. Then Vf is not approximately amenable. Proof. As Vf has no right approximate identity, it is not approximately amenable by [4, Lemma 2.2]. (cid:3) The following is a result of F. Ghahramani, see [2, Theorem 1.5.4]. Theorem 2.8. Every pointwise amenable, commutative Banach algebra is approx- imately amenable. Remark 2.9. By Theorem 2.2 and Proposition 2.7, Vf is a pointwise amenable, non-approximately amenable Banach algebra which is not commutative. So, without commutativity Theorem 2.8 is not true. We conclude by looking at character amenability of Vf . For a Banach algebra A, we write ∆(A) for the set of all continuous homomorphisms from A onto C. From [7], we recall that a Banach algebra A is ϕ-amenable, ϕ ∈ ∆(A), if there exists a bounded linear functional m on A∗ satisfying m(ϕ) = 1 and m(g · a) = ϕ(a)m(g), (a ∈ A, g ∈ A∗). Further, A is character amenable if it is ϕ-amenable for all ϕ ∈ ∆(A) [10]. Proposition 2.10. Let V be a Banach space, and let f ∈ V ∗ be a non-zero element such that f ≤ 1. Then Vf is character amenable. 4 SARA BEHNAMIAN AND AMIN MAHMOODI Proof. Clearly f is a continuous homomorphisms from Vf onto C. Choose u ∈ Vf with f (u) = 1, and set uα := u for each α. Then auα − f (a)uα = 0 for all a ∈ Vf . Hence, Vf is f -amenable by [7, Theorem 1.4]. It follows from ∆(Vf ) = {f } that Vf is character amenable. (cid:3) References [1] H. G. Dales, Banach algebras and automatic continuity, London Mathematical Society Mono- graphs 24, Clarendon Press, Oxford, 2000. [2] H. G. Dales and R. J. Loy, Approximate amenability of semigroup algebras and Segal algebras, Dissertations Math. 474 (2010). [3] E. Desquith, Banach algebras associated to bounded module maps, Available in http://streaming.ictp.trieste.it/preprints/p/98/194.pdf. [4] F. Ghahramani and R. J. Loy, Generalized notions of amenability, J. Funct. Anal. 208 (2004), 229-260. [5] B. E. Johnson, Cohomology in Banach algebras, Mem. Amer. Math. Soc. 127, 1972. [6] B. E. Johnson, Approximate diagonals and cohomology of certain annihilator Banach algebras, Amer. J. Math. 94, (1972), 685-698. [7] E. Kaniuth, A. T. Lau and J. Pym, On ϕ-amenability of Banach algebras, Math. Proc. Camb. Phil. Soc. 144 (2008), 85-96. [8] A. R. Khoddami and H. R. Ebrahimi Vishki, The higher duals of a Banach algebra induced by a bounded linear functional, Bull. Math. Anal. Appl. 3 (2011), 118-122. [9] V. Runde, Amenability for dual Banach algebras, Studia Math. 148 (2001), 47-66. [10] M. S. Monfared, Character amenability of Banach algebras, Math. Proc. Camb. Phil. Soc. 144 (2008), 697-706. [11] M. Shakeri and A. Mahmoodi, Pointwise amenability for dual Banach algebras, to appear. Department of Mathematics, Science and research Branch, Islamic Azad Univer- sity, Tehran, Iran, e-mail: [email protected] Department of Mathematics, Central Tehran Branch, Islamic Azad University, Tehran, Iran, e-mail: a [email protected]
1610.09879	2	1610	2017-08-23T19:08:58	Ultradistributional boundary values of harmonic functions on the sphere	[ "math.FA", "math.CV" ]	We present a theory of ultradistributional boundary values for harmonic functions defined on the Euclidean unit ball. We also give a characterization of ultradifferentiable functions and ultradistributions on the sphere in terms of their spherical harmonic expansions. To this end, we obtain explicit estimates for partial derivatives of spherical harmonics, which are of independent interest and refine earlier estimates by Calderon and Zygmund. We apply our results to characterize the support of ultradistributions on the sphere via Abel summability of their spherical harmonic expansions.	math.FA	math	ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS ON THE SPHERE ¯DOR¯DE VU CKOVI ´C AND JASSON VINDAS Abstract. We present a theory of ultradistributional boundary values for har- monic functions defined on the Euclidean unit ball. We also give a characteriza- tion of ultradifferentiable functions and ultradistributions on the sphere in terms of their spherical harmonic expansions. To this end, we obtain explicit estimates for partial derivatives of spherical harmonics, which are of independent interest and refine earlier estimates by Calder´on and Zygmund. We apply our results to characterize the support of ultradistributions on the sphere via Abel summability of their spherical harmonic expansions. 1. Introduction The study of boundary values of harmonic and analytic functions is a classical and important subject in distribution and ultradistribution theory. There is a vast literature dealing with boundary values on Rn, see e.g. [1, 5, 6, 10, 12, 14, 20] and references therein. In the case of the unit sphere Sn−1, the characterization of harmonic functions in the Euclidean unit ball of Rn having distributional boundary values on Sn−1 was given by Estrada and Kanwal in [11]. In a recent article [13], Gonz´alez Vieli has used the Poisson transform to obtain a very useful description of the support of a Schwartz distribution on the sphere (cf. [27] for support char- acterizations on Rn). Representations of analytic functionals on the sphere [17] as initial values of solutions to the heat equation were studied by Morimoto and Suwa [18]. In this article we generalize the results from [11] to the framework of ultra- distributions [15, 16] and supply a theory of ultradistributional boundary values of harmonic functions on Sn−1. Our goal is to characterize all those harmonic functions U, defined in the unit ball, that admit boundary values limr→1− U(rω) in an ultradistribution space E∗′(Sn−1). Our considerations apply to both non- quasianalytic and quasianalytic ultradistributions, and, in particular, to analytic functionals. As an application, we also obtain a characterization of the support of a non-quasianalytic ultradistribution in terms of Abel summability of its spherical harmonic series expansion. Since Schwartz distributions are naturally embedded into the spaces of ultradistributions in a support preserving fashion, our support characterization contains as a particular instance that of Gonz´alez Vieli quoted above. 2010 Mathematics Subject Classification. Primary 31B05, 46F20; Secondary 42C10, 46F15. Key words and phrases. Harmonic functions on the unit ball; boundary values on the sphere; partial derivatives of spherical harmonics; support of ultradistributions; ultradifferentiable func- tions; Abel summabiity. The authors gratefully acknowledge support by Ghent University, through the BOF-grant 01N01014. 1 2 ¯D.VU CKOVI ´C AND J. VINDAS In Section 4 we study spaces of ultradifferentiable functions and ultradistributions through spherical harmonics. Our main results there are descriptions of these spaces in terms of the decay or growth rate of the norms of the projections of a function or an ultradistribution onto the spaces of spherical harmonics. We also establish the convergence of the spherical harmonic series in the corresponding space. Note that eigenfunction expansions of ultradistributions on compact analytic manifolds have recently been investigated in [8, 9] with the aid of pseudodifferential calculus (cf. [28] for the Euclidean global setting). However, our approach here is quite different and is rather based on explicit estimates for partial derivatives of solid harmonics and spherical harmonics that are obtained in Section 3. Such estimates are of independent interest and refine earlier bounds by Calder´on and Zygmund from [4]. Harmonic functions with ultradistributional boundary values are characterized in Section 5. The characterization is in terms of the growth order of the harmonic function near the boundary Sn−1; we also show in Section 5 that a harmonic func- tion satisfying such growth conditions must necessarily be the Poisson transform of an ultradistribution. In the special case of analytic functionals, our result yields as a corollary: any harmonic function on the unit ball arises as the Poisson trans- form of some analytic functional on the sphere. Finally, Section 6 deals with the characterization of the support of non-quasianalytic ultradistributions on Sn−1. 2. Preliminaries We employ the notation Bn for the open unit ball of Rn. We work in dimension n ≥ 2. 2.1. Spherical harmonics. The theory of spherical harmonics is a classical sub- ject in analysis and it is very well explained in several textbooks (see e.g. [2, 3]). The space of solid harmonics of degree j will be denoted by Hj(Rn), its elements are the harmonic homogeneous polynomials of degree j on Rn. A spherical harmonic of degree j is the restriction to Sn−1 of a solid harmonic of degree j and we write Hj(Sn−1) for space of all spherical harmonics of degree j. Its dimension, denoted as dj = dimHj(Sn−1), is (cf. [3] or [26, Thm. 2, p. 117]) dj = (2j + n − 2)(n + j − 3)! j!(n − 2)! ∼ 2jn−2 (n − 2)! . From this exact formula, it is not hard to see that dj satisfies the bounds 2 (2.1) (n − 2)! It is well known [3] that jn−2 < dj ≤ njn−2, for all j ≥ 1. L2(Sn−1) = Hj(Sn−1), ∞ Mj=0 where the L2-inner product is taken with respect to the surface measure of Sn−1. The orthogonal projection of f ∈ L2(Sn−1) onto Hj(Sn−1) will always be denoted as fj; it is explicitly given by fj(ω) = 1 Sn−1ZSn−1 f (ξ)Zj(ω, ξ)dξ, ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 3 where Zj(ω, x) is the zonal harmonic of degree j with pole ω [3]. We then have that Sn−1−1Zj(ξ, x) is the reproducing kernel of Hj(Sn−1), namely, (2.2) Yj(ξ)Zj(ω, ξ)dξ, Yj(ω) = 1 for every Yj ∈ Hj(Sn−1). Sn−1ZSn−1 2.2. Homogeneous extensions and differential operators on Sn−1. We write E(Ω) = C∞(Ω), where Ω is an open subset of Rn or Sn−1. Given a function ϕ on Sn−1, its homogeneous extension (of order 0) is the function ϕ↾ defined as ϕ↾(x) = ϕ(x/x) on Rn\{0}. It is easy to see that ϕ ∈ E(Sn−1) if and only if ϕ↾ ∈ E(Rn\{0}). Furthermore, we define the differential operators ∂α Sn−1 : E(Sn−1) → E(Sn−1) via Sn−1ϕ)(ω) = (∂αϕ↾)(ω), ω ∈ Sn−1. (∂α We can then consider L(∂Sn−1) for any differential operator L(∂) defined on Rn\{0}. In particular, ∆Sn−1 stands for the Laplace-Beltrami operator of the sphere. Finally, if F is a function on Rn, we simply write kFkLq(Sn−1) for the Lq(Sn−1)- norm of its restriction to Sn−1. 3. Estimates for partial derivatives of spherical harmonics Calder´on and Zygmund showed [4, Eq. (4), p. 904] the following estimates for the partial derivatives of a spherical harmonic Yj ∈ Hj(Sn−1), Sn−1YjkL∞(Sn−1) ≤ Cα,n jαkYjkL∞(Sn−1), (3.1) k∂α where the constants Cα,n depend on the order of differentiation and the dimension on an unspecified way. The same topic is treated in Seeley's article [26]. The goal of this section is to refine (3.1) by exhibiting explicit constants Cα,n. We also give explicit bounds for the partial derivatives of spherical harmonics in spherical coordinates. Such estimates in spherical coordinates play an important role in the next section. We consider here p(θ) = (p1(θ), . . . , pn(θ)), p(θ) = (cos θ1, sin θ1 cos θ2, . . . , sin θ1 · · · sin θn−2 cos θn−1, sin θ1 . . . sin θn−2 sin θn−1), where θ ∈ Rn−1. Naturally, the estimate (3.3) below also holds if we choose the north pole to be located at a point other than (1, 0, . . . , 0). Theorem 3.1. We have the bounds: (3.2) (3.3) (a) For every solid harmonic Qj ∈ Hj(Rn) and all α 6= 0, k∂αQjkL∞(Sn−1) ≤ e 2√n 2 n 4 − 1 α 2 jα+ n 2 −1kQjkL∞(Sn−1). (b) For all spherical harmonic Yj ∈ Hj(Sn−1) and all α 6= 0, 2 jα+ n 4 − 1 α n k∂α θ (Yj ◦ p)kL∞(Rn−1) ≤ e 2√n(cid:0)(n + 1)α − 1(cid:1) 2 (c) For all spherical harmonic Yj ∈ Hj(Sn−1), all α 6= 0, and any ε > 0, (3.4) Sn−1YjkL∞(Sn−1) ≤ en(cid:16) 1 2 −1α!kYjkL∞(Sn−1). k∂α Proof. (a) For (3.2), we assume that α ≤ j, otherwise the result trivially holds. Our starting point is the same as the inductive step in the proof of [26, Thm. 4, p. 120], namely, the inequality 4 +√2+3√2+4/ε(cid:17)− 1 (2 + ε)αjα+ n 2 n α+1 2 2 −1kYjkL∞(Sn−1). (3.5) ZSn−1 ∂αQj(ω)2dω ≤ (j − α + 1)(n + 2j − 2α)ZSn−1 ∂βQj(ω)2dω, 4 ¯D.VU CKOVI ´C AND J. VINDAS valid for all multi-index β with β = α − 1 and β ≤ α. Successive application of (3.5) leads to ZSn−1 ∂αQj(ω)2dω ≤ α−1 Yi=0 (j − i) · α (n + 2j − 2i)ZSn−1 Qj(ω)2dω. Yi=1 The coefficient in this bound can be estimated as follows, α−1 Yi=0 (j − i) · α Yi=1 (n + 2j − 2i) ≤ j2α2α α j (cid:19) n/2 − i (cid:19)α Yi=1(cid:18)1 + ≤ 2αj2α(cid:18)1 + ≤ 2αj2αe L2(Sn−1) = dj−αSn−1 for each ω ∈ Sn−1 n/2 − 1 2 −1. j n Now, ∂αQj ∈ Hj−α(Rn) and kZj−α(ω, · )k2 (cf. [3, pp. 79 -- 80]). Thus, we obtain (cf. (2.2)), for all ω ∈ Sn−1, Sn−1k∂αQjkL2(Sn−1)kZj−α(ω, · )kL2(Sn−1) =s dj−α 2r n Sn−1 2 −1kQjkL2(Sn−1), ∂αQj(ω) ≤ 2 jα+ n ≤ e 4 − 1 1 2 α n Sn−1 k∂αQjkL2(Sn−1) where we have used dj−α ≤ njn−2 (see (2.1)). This shows (3.2). (b) Our proof of (3.3) is based on the multivariate Fa`a di Bruno formula for the partial derivatives of the composition of functions. Let m = α. Specializing [7, Eq. (2.4)] to h = f ◦ p, where f is a function on Rn, we obtain θ pj]kj n ∂α θ h = X1≤λ≤m (∂λ x f ) ◦ p X(k,l)∈p(α,λ) α! Yj=1 [∂lj (kj!)[lj!]kj , where the set of multi-indices p(α, λ) ⊂ N2n is as described in [7, p. 506]. We also employ the identity [7, Cor. 2.9] n α! Xλ=k Xp(α,λ) Yj=1 1 (kj!)[lj!]kj = nkS(m, k), where S(m, k) are the Stirling numbers of the second kind. For such numbers [22, Thm. 3] we have the estimates1 k(cid:19)km−k, S(m, k) ≤(cid:18)m Since obviously ∂lj pj(θ) ≤ 1, we obtain Xk=1(cid:18)m (3.6) θ (f ◦ p)kL∞(Ω) ≤ k∂α m 1 ≤ k ≤ m. k(cid:19)km−knk max λ=k k∂λ x fkL∞(p(Ω)), 1Actually, S(m, k) ≤ 1 2(cid:18)m k(cid:19)km−k holds for 1 ≤ k ≤ m − 1 if m ≥ 2, and S(m, m) = 1. ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 5 for any Ω ⊆ Rn−1 and the corresponding set p(Ω) ⊆ Sn−1. We now apply this inequality to estimate ∂α θ (Yj ◦ p). Let Qj ∈ Hj(Rn) be the solid harmonic corre- sponding to Yj, clearly Qj ◦ p = Yj ◦ p and kQjkL∞(Sn−1) = kYjkL∞(Sn−1). Using (3.6) with f = Qj, the bound (3.2), and the fact that ∂λ x Qj = 0 if λ > j, we conclude that k∂α θ (Yj ◦ p)kL∞(Rn−1) ≤ e = e m n n m 2 j 2 −1kYjkL∞(Sn−1) 2√n2 4 − 1 2√n ((n + 1)m − 1) 2 4 − 1 m n k(cid:19)nkjm−kjk Xk=1(cid:18)m 2 kYjkL∞(Sn−1). 2 jm+ n−2 (c) We need to estimate the partial derivatives of Y ↾ j = Qj◦F , where Qj ∈ Hj(Rn) and F (x) = x/x, x ∈ Rn\{0}. Instead of using the Fa`a di Bruno formula to handle directly the partial derivatives of this composition, we will adapt Hormander's proof of [14, Prop. 8.4.1, p. 281] to our problem. Let 0 < r < 1/2 and ω ∈ Sn−1. Note that if z − ω ≤ r and we write z = x + iy, then ℜe (z2 1 + · · · + z2 n) = x2 − y2 ≥ 1 − 2r > 0. So, F is holomorphic on this region of Cn. For m ≥ 1, we define the sequence of functions Gm(z) = Xβ≤m(cid:0)∂βQj(cid:1) (F (ω)) (F (z) − F (ω))β . β! Each Gm is holomorphic when z − ω ≤ r and the derivatives of Gm of order m at z = ω are the same as those of Y ↾ j (x) at x = ω. We keep z − ω ≤ r. We have the bound = Cr, 3 2√1 − 2r 2 −1 (Cr√2)β β! jβ+ n (Cr√2)β β! F (z) − F (ω) ≤ 1 + and hence, by (3.2), z 1 + · · · + z2 n) pℜe (z2 < 1 + Gm(z) ≤ e n 4 − 1 2√n kYjkL∞(Sn−1) Xβ≤min{m,j} 2√n jm+ n 2 −1kYjkL∞(Sn−1) Xβ∈Nn 2√n jm+ n 4 +√2Cr)− 1 n 4 − 1 ≤ e = en( 1 2 −1kYjkL∞(Sn−1). The Cauchy inequality applied in the polydisc zj − ωj ≤ r/√n yields Sn−1Yj(ω) = ∂αGα(ω) ≤ en( 1 ∂α 4 +√2Cr)− 1 2 n One obtains (3.4) upon setting r = 1/(2 + ε). α+1 2 r−αjα+ n 2 −1α!kYjkL∞(Sn−1). (cid:3) 4. Spherical harmonic characterization of ultradifferentiable functions and ultradistributions In this section we characterize ultradifferentiability properties of a function on the sphere in terms of its spherical harmonic expansion. We also obtain a spherical harmonic characterization of ultradistributions on the sphere. 6 ¯D.VU CKOVI ´C AND J. VINDAS We start by introducing ultradifferentiable functions on Sn−1. A weight sequence is simply a positive sequence (Mp)p∈N of real numbers with M0 = 1. Throughout the rest of the article, we always impose the following assumptions on weight sequences, (M.0) p! ≤ A0H p (M.1) M 2 (M.2)′ Mp+1 ≤ AH pMp, p ∈ N, for some A, H > 1. function of the sequence Mp is defined as The meaning of these standard conditions is explained in [15]. The associated 0 Mp, p ∈ N, for some A0 > 0 and H0 > 1, p ≤ Mp−1Mp+1, p ≥ 1, M(t) = sup p∈N log tp Mp , t > 0, and M(0) = 0. See also [15] for its properties and the translation of (M.1) and (M.2)′ into properties of M. In particular, we shall often make use of the inequality for all t > 0, tηe−M (H ηt) ≤ Aηe−M (t), (4.1) for any η > 0, implied by (M.1) and (M.2)′ [15, Eq. (3.13), p. 50]. We also point out that, under (M.1), the condition (M.0) becomes equivalent to the bound M(t) = O(t) [15, Lemma 3.8]. As a typical example, we mention Mp = (p!)s with s ≥ 1, whose associated function has growth order M(t) ≍ t1/s. (or class {Mp}) as the space of all smooth functions ϕ ∈ E(Sn−1) such that (4.2) We define the space E{Mp}(Sn−1) of ultradifferentiable functions of Roumieu type Sn−1ϕkL∞(Sn−1) hαk∂α sup α∈N Mα < ∞, for some h > 0. Note that if Mp = (p!)s with s ≥ 1, one recovers the spaces of Gevrey differentiable functions on the sphere. In the special but very important case Mp = p!, we also write A(Sn−1) = E{p!}(Sn−1); this is in fact the space of real analytic functions on Sn−1 [17]. The space E (Mp)(Sn−1) of ultradifferentiable functions of Beurling type (class (Mp)) is defined by requiring that (4.2) holds for every h > 0. Whenever we consider the Beurling case, we suppose that Mp satisfies the ensuing stronger as- sumption than (M.0), (NA) For each L > 0 there is AL > 0 such that p! ≤ ALLpMp, p ∈ N. Notice (M.1) implies that (NA) is equivalent to M(t) = o(t) as t → ∞ [15, Lemma 3.10]. As customary, we write ∗ = {Mp} or (Mp) when considering both cases simulta- neously. It should be noticed that the condition (M.0) (the condition (NA)) implies that A(Sn−1) is the smallest among all spaces of ultradifferentiable functions that we consider here, that is, one always has the inclusion A(Sn−1) ⊆ E∗(Sn−1). A word about the definition of E∗(Sn−1) that we have adopted here. Since we have used the differential operators ∂α Sn−1 in (4.2), ∗-ultradifferentiability of ϕ on Sn−1 is the same as ∗-ultradifferentiability of its homogeneous extension (of order 0) on Rn \ {0}, namely, ϕ ∈ E∗(Sn−1) if and only if ϕ↾ ∈ E∗(Rn \ {0}), with the spaces of ultradifferentiable functions on an open subset of Rn defined in the usual way [15]. Moreover, in view of the analyticity of the mapping x → x/x and the fact that the pullbacks by analytic functions induce mappings between spaces of ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 7 ∗-ultradifferentiable functions under the assumptions (M.0) ((NA) in the Beurling case), (M.1) and (M.2)′ (cf. [14, Prop. 8.4.1], [16, p. 626], [23]), our definition of E∗(Sn−1) coincides with that of ∗-ultradifferentiable functions on compact analytic manifolds via local analytic coordinates. We are ready to characterize E∗(Sn−1) in terms the norm decay of projections onto the spaces of spherical harmonics. Recall our convention is to write ϕj for the projection of ϕ onto Hj(Sn−1). Theorem 4.1. Let ϕ ∈ L2(Sn−1) and let 1 ≤ q ≤ ∞. The following statements are equivalent: (i) ϕ belongs to E{Mp}(Sn−1) (to E (Mp)(Sn−1)) . (ii) ∆p Sn−1ϕ ∈ L2(Sn−1) for all p ∈ N and there are h, C > 0 (for every h > 0 there is C = Ch > 0) such that k∆p Sn−1ϕkL2(Sn−1) ≤ Ch−2pM2p. (iii) There are C, h > 0 (for every h > 0 there is C = Ch > 0) such that (4.4) Proof. (i)⇒(ii). The proof of this implication is simple. Indeed, suppose that kϕjkLq(Sn−1) ≤ Ce−M (hj). ∂αϕ↾(x) ≤ Ch−αMα, for all x ∈ Rn \ {0}. (4.3) Since and ∂x2 1 ∂2 ∂x2 (cid:18) ∂2 + · · · + n(cid:19)p Xα1+···+αn=p Sn−1ϕkL2(Sn−1) ≤ k∆p k∆p the condition (M.1) gives = Xα1+···+αn=p p! α1!α2! . . . αn! ∂2α1 ∂x2α1 1 ∂2αn ∂x2αn n · · · p! α1!α2! . . . αn! = np, (h/√n)−2pMp. ≤ C Sn−1 1 2 2 Sn−1ϕkL∞(Sn−1) Sn−1 1 Sn−1ZSn−1 1 (ii)⇒(iii). Suppose (4.3) holds. The projection of ϕ onto Hj(Sn−1) is (4.5) ϕj(ω) = ϕ(ξ)Zj(ω, ξ)dξ. We first assume that j ≥ 1. The Laplace-Beltrami operator is self-adjoint [26, Lemma 1] and each spherical harmonic of degree j, such as Zj(ω, ξ), is an eigenfunc- tion of ∆Sn−1 with eigenvalue −j(j + n− 2). Also, kZj(ω, · )kL2(Sn−1) =pdjSn−1 ≤ 2 −1pnSn−1 (see (2.1) and [3, pp. 79 -- 80]); therefore, 1 j n ϕj(ω) = jp(j + n − 2)pSn−1(cid:12)(cid:12)(cid:12)(cid:12) ZSn−1 2 −1h−2pM2p. j−2p+ n (∆pϕ)(ξ)Zj(ω, ξ)dξ(cid:12)(cid:12)(cid:12)(cid:12) C√n Sn−1 1 2 ≤ Taking supremum over ω and infimum over p, we conclude that 2√n j q − 1 n 1 1 qkϕjkL∞(Sn−1) ≤ CSn−1 kϕjkLq(Sn−1) ≤ Sn−1 2 −1e−M (hj). 8 ¯D.VU CKOVI ´C AND J. VINDAS Taking η = n/2 − 1 in (4.1), we obtain kϕjkL2(Sn−1) ≤ CSn−1 1 q − 1 2√n(A/h) n 2 −1H ( n 2 −1)2 e−M(jhH 1−n/2), j ≥ 1. For j = 0, using (4.5), we have kϕjkLq(Sn−1) ≤ CSn−1 kϕjkLq(Sn−1) ≤ ChCe−M(jhH 1−n/2), 2 max{1,√n(A/h) q − 1 2 −1)2}. 2 −1H ( n n 1 with Ch = Sn−1 (iii)⇒(i). Assume now (4.4). In view of (4.5) and (4.1), we may also assume that q = ∞. We estimate the partial derivatives of ϕ in spherical coordinates. Write ϕ = ϕ ◦ p and ϕj = ϕj ◦ p. Let α 6= 0. Let r be an integer larger than n/2 + 1. If we combine the estimate (3.3) with (4.4), we obtain 1 q − 1 2 , thus j ≥ 0. k∂α θ ϕjkL∞(Rn−1) ≤ e ≤ C n 4 − 1 2√n(cid:16)√2(n + 1)(cid:17)α jα+ n h−r j2 e n 4 − 1 2√n(cid:16)√2(n + 1)/h(cid:17)α Mα+r, 2 −1kϕjkL∞(Sn−1) j ≥ 1. Calling Ch = e 2 h−r√nπ2/6, we conclude that n 4 − 1 k∂α θ ϕkL∞(Rn−1) ≤ ∞ Xj=1 θ ϕjkL∞(Rn−1) ≤ Ch(cid:16)√2(n + 1)/h(cid:17)α Mα+r. k∂α The assumption (M.2)′ implies Mp+r ≤ H rpArH (4.6) θ ϕkL∞(Sn−1) ≤ CChArH k∂α r(r−1) r(r−1) 2 Mp, so 2 (cid:16)H r√2(n + 1)/h(cid:17)α Mα. Setting the north pole at different points of the sphere induces an analytic atlas of Sn−1 and x → x/x is analytic on Rn. As previously mentioned, the conditions (M.0) ((NA) in the Beurling case), (M.1), and (M.2)′ ensure that pullbacks by an- alytic functions preserve ∗-ultradifferentiability. So, ϕ ∈ E∗(Sn−1). The inequality (4.6) and the proof of [14, Prop. 8.1.4] give actually a more accurate result: There are constants C′h and ℓ, depending also on the sequence Mp and the dimension n but not on ϕ, such that k∂α Sn−1ϕkL∞(Sn−1) ≤ CC′h(ℓh)−αMα. (cid:3) The proof of Theorem 4.1 actually yields stronger information than what has been stated. The canonical topology of E∗(Sn−1) is defined as follows. For each h > 0, consider the Banach space E{Mp},h(Sn−1) of all smooth functions ϕ on Sn−1 such that the norm (4.7) kϕkh = sup α∈N hαk∂α Sn−1ϕkL∞(Sn−1) Mα is finite. As locally convex spaces, we obtain the (DF S)-space and (F S)-space E{Mp},h(Sn−1). E{Mp}(Sn−1) = lim−→h→0+ E{Mp},h(Sn−1) and E (Mp)(Sn−1) = lim←−h→∞ ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 9 What we have shown is that the family of norms (4.7) is tamely equivalent to the norms (4.8) kϕk′h = sup j∈N eM (hj)kϕjkLq(Sn−1), h > 0 (1 ≤ q ≤ ∞), in the sense that there are positive constants ℓ and L, only depending on the dimension n, the parameter q, and the weight sequence, such that one can find Ch > 0 and ch > 0 with chk · k′ℓh ≤ k · kh ≤ Chk · k′Lh, for all h > 0. Working with the family of norms (4.8) is more convenient than (4.7) when dealing with assertions about spherical harmonic expansions. Proposition 4.2. Let ϕ ∈ E∗(Sn−1). Then its spherical harmonic series expansion ϕ =P∞j=0 ϕj converges in (the strong topology of ) E∗(Sn−1). Proof. Let h > 0. Invoking (4.1) with η = 1, kϕ − k Xj=0 ϕjk′h = sup j>k eM (hj)kϕjk ≤ A kh kϕk′Hh, for each k ≥ 1. (cid:3) If we specialize our results to the space of real analytic functions and use the fact that the associated function of p! is M(t) ≍ t, we obtain the following characteri- zation of A(Sn−1) = E{p!}(Sn−1). Corollary 4.3. A sequence of spherical harmonics with ϕj ∈ Hj(Sn−1) gives rise to a real analytic function ϕ =P∞j=0 ϕj on Sn−1 if and only if 1 j < 1. lim sup j→∞ (cid:0)kϕjkLq(Sn−1)(cid:1) Here is another application of the norms (4.8). The space of ultradistributions E∗′(Sn−1) (of class ∗) on Sn−1 is the strong dual of E∗(Sn−1). When ∗ = {p!}, one obtains the space of analytic functionals A′(Sn−1) [17]. Given f ∈ E∗′(Sn−1), we can also define its projection onto Hj(Sn−1) as fj(ω) = 1 Sn−1hf (ξ), Zj(ω, ξ)i, where the ultradistributional evaluation in the dual pairing is naturally with respect to the variable ξ. Note that, clearly, (4.9) hfj, ϕi =ZSn−1 fj(ω)ϕ(ω)dω = hf, ϕji, for each ϕ ∈ E∗(Sn−1). Theorem 4.4. Every ultradistribution f ∈ E{Mp}′(Sn−1) (f ∈ E (Mp)′(Sn−1)) has spherical harmonic expansion (4.10) ∞ fj, f = Xj=0 e−M (hj)kfjkLq(Sn−1) < ∞ where its spherical harmonic projections fj satisfy (4.11) sup j∈R (1 ≤ q ≤ ∞), 10 ¯D.VU CKOVI ´C AND J. VINDAS for all h > 0 (for some h > 0). Conversely, a series (4.10) converges in the strong topology of E{Mp}′(Sn−1) (of E (Mp)′(Sn−1)) if the Lq(Sn−1)-norms of fj have the stated growth properties. Proof. Since E∗(Sn−1) are Montel spaces, the strong convergence of (4.10) follows from its weak convergence, and the latter is a consequence of Proposition 4.2 and (4.9). For the bound (4.11), the continuity of f implies that for each h > 0 (for some h > 0) there is a constant Ch such that hf, ϕi ≤ Chkϕk′h, for all ϕ ∈ A(Sn−1). We may assume that j ≥ 1. Considering the case q = 2 of (4.8), taking ϕ(ξ) = Sn−1−1Zj(ω, ξ), and using the inequalities (2.1) and (4.1), one has kfjkLq(Sn−1) ≤ Sn−1 ≤ Sn−1 1 qkfjkL∞(Sn−1) ≤ Sn−1 q − 1 q − 1 2 −1√neM (jhH 2 Ch(A/h) n 1 1 n 2 −1eM (hj) 2 Ch√nj 2 −1). n (cid:3) For analytic functionals we have, 1 lim sup j→∞ (cid:0)kfjkLq(Sn−1)(cid:1) Corollary 4.5. A sequence fj ∈ Hj(Sn−1) gives rise to an analytic functional f =P∞j=0 fj on Sn−1 if and only if We mention that the strong topologies of the (F S)-space E{Mp}′(Sn−1) and the (DF S)-space E (Mp)′(Sn−1) can also be induced via the family of norms (4.11) as P∞j=0 fj satisfying (4.11). For each j ∈ N select an orthonormal basis of real spherical harmonics {Yk,j}dj of Hj(Sn−1). It is then clear that every ultradistribution f ∈ E∗′(Sn−1) and every ϕ ∈ E∗(Sn−1) can be expanded as the projective and inductive limits of the Banach spaces of ultradistributions f = j ≤ 1. k=1 (4.12) and (4.13) f = ϕ(ω) = where the coefficients satisfy ∞ Xj=0 dj Xk=1 ∞ Xj=0 dj Xk=1 ck,jYk,j ak,jYk,j(ω), k,j ck,je−M (hj) < ∞ sup (for each h > 0 in the Roumieu case and for some h > 0 in the Beurling case), and k,j ak,jeM (hj) < ∞ sup (for some h > 0 or for each h > 0, respectively). Conversely, any series (4.12) and (4.13) converge in E∗′(Sn−1) and E∗(Sn−1), respectively, if the coefficients have the stated growth properties. We have used here (4.4), (4.11), and (4.1). ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 11 From here one easily derives that E∗(Sn−1) (and hence E∗′(Sn−1)) is a nuclear space. We also obtain that {Yk,j} is an absolute Schauder basis [25, p. 340] for both E∗(Sn−1) and E∗′(Sn−1). We end this section with a remark concerning Theorem 4.1. Remark 4.6. It is very important to emphasize that Theorem 4.1 is no longer true without the assumption (M.0). To see that it is imperative to assume (M.0), we give an example in which the implication (ii)⇒(i) fails without it. In fact, let Mp be any weight sequence for which (M.1) and (M.2)′ hold but limp→∞ (Mp/p!) p = 0. (For example, the sequence Mp = p!s with 0 < s < 1.) We consider ϕ(ω) = Y1(ω1, . . . , ωn) = ω1. This function is a spherical harmonic of degree 1, and thus it is an eigenfunction for the Laplace- Beltrami operator corresponding to the eigenvalue −(n − 2). Thus, 1 k∆p Sn−1ϕkL2(Sn−1) ≤ np Sn−1 1 2 and in particular (4.3) is satisfied for Mp. If there would be an h > 0 such that (4.2) holds with Mp = p!s, we would have for the function f (t) = 1 pt2 + 1/2 that kf (p)kL∞(R) = √2 sup t∈R ∂p x2ϕ↾(√2/2, t, 0, . . . , 0) ≤ C′h−pMp, for all p ∈ N, for some C′ > 0. But then f would be analytically continuable to the whole C as an entire function, which is impossible because f has branch singularities at t = ±i√2/2. On the other hand, note that in establishing the implications (i)⇒(ii)⇒(iii) the condition (M.0) plays no role because we have only made use there of (M.1) and (M.2)′. 5. Boundary values of harmonic functions We now generalize the results from [11] to ultradistributions. We shall charac- terize all those harmonic functions on the open unit ball Bn that admit ultradistri- butional boundary values on Sn−1 in terms of their growth near the boundary. Our characterization applies for sequences satisfying the additional conditions discussed below. Let us fix some notation and terminology. We write H(Bn) for the space of all harmonic functions on Bn. We say that U ∈ H(Bn) has ultradistribution boundary values in the space E∗′(Sn−1) if there is f ∈ E∗′(Sn−1) such that (5.1) U(rω) = f (ω) in E∗′(Sn−1). lim r→1− Since E∗′(Sn−1) is Montel, the converge of (5.1) in the strong topology is equivalent to weak convergence, i.e., (5.2) lim r→1−hU(rω), ϕ(ω)i = lim r→1−ZSn−1 U(rω)ϕ(ω)dω = hf, ϕi, for each ϕ ∈ E∗(Sn−1). 12 ¯D.VU CKOVI ´C AND J. VINDAS We first show that (5.1) holds with U being the Poisson transform of f . For this, our assumptions are the same as in the previous section, i.e., (M.1), (M.2)′ and (M.0) ((NA) in the Beurling case). The Poisson kernel of Sn−1 is [3] (5.3) P (x, ξ) = 1 Sn−1 1 − x2 x − ξn = 1 Sn−1 ∞ Xj=0 xjZj(cid:18) x x , ξ(cid:19) , ξ ∈ Sn−1, x ∈ Bn. Since P is real analytic with respect to ξ, we can define the Poisson transform of f ∈ E∗′(Sn−1) as (5.4) P [f ](x) = hf (ξ), P (x, ξ)i, x ∈ Bn. Clearly, P [f ] ∈ H(Bn) and, by (5.3), P [f ](rω) =P∞j=0 rjfj(ω). Proposition 5.1. For each f ∈ E∗′(Sn−1) and ϕ ∈ E∗(Sn−1), we have in E∗′(Sn−1) (5.5) P [f ](rω) = f (ω) lim r→1− and (5.6) P [ϕ](rω) = ϕ(ω) lim r→1− in E∗(Sn−1). Proof. Due to the Montel property of these spaces (which also implies they are reflexive), it is enough to verify weak convergence of the Poisson transform in both cases in order to prove strong convergence of (5.5) and (5.6). By Theorem 4.4 (or series then yields Theorem 4.1), we have that hf, ϕi = P∞j=0hfj, ϕji; Abel's limit theorem on power r→1−ZSn−1 rjhfj, ϕji = hf, ϕi. r→1−hf (ω), P [ϕ](rω)i = lim r→1− P [f ](rω)ϕ(ω)dω = lim lim ∞ Xj=0 We now deal with the characterization of harmonic functions U that satisfy (5.1). This characterization is in terms of the associated function of Mp/p!, which we denote by M∗ as in [15], i.e., the function (cid:3) M∗(t) = sup p∈N log(cid:18)p!tp Mp(cid:19) for t > 0 and M∗(0) = 0. We need two extra assumptions on the sequence, namely, (M.1)∗ Mp/p! satisfies (M.1), (M.2) Mp+q ≤ AH p+qMqMq, p, q ∈ N, for some A, H ≥ 1. Naturally, (M.1)∗ implies (M.0) and (M.1) while (M.2) is stronger than (M.2)′. 1p! ≤ Mp ≤ C2Lp Note that (M.1)∗ delivers essentially two cases. Either (NA) holds or there are constants such that C1Lp 2p!. In the latter case we may assume that Mp = p! as for any such a sequence E{Mp}(Sn−1) = A(Sn−1). When (NA) holds M∗(t) is finite for all t ∈ [0,∞), whereas Mp = p! gives M∗(t) = 0 for 0 ≤ t ≤ 1 and M∗(t) = ∞ for t > 1. In the (NA) case we also have M∗(t) = 0 for t ∈ [0, M1]. The importance of the assumptions (M.1)∗ and (M.2) lies in the ensuing lemma of Petzsche and Vogt: ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 13 Lemma 5.2 ([20]). Suppose that Mp satisfies (M.1)∗ and (M.2). Then, there are constants L, ℓ > 0 such that inf y>0 (M∗(1/y) + ty) ≤ M(ℓt) + log L, for all t > 0. We then have, Theorem 5.3. Assume Mp satisfies (M.1)∗ and (M.2). Then, a harmonic function U ∈ H(Bn) admits boundary values in E{Mp}′(Sn−1) (in E (Mp)′(Sn−1)) if and only if for each h > 0 there is C = Ch > 0 (there are h > 0 and C > 0) such that (5.7) U(x) ≤ CeM ∗( h 1−x ) for all x ∈ Bn. In such a case U = P [f ], where f is its boundary ultradistribution given by (5.1). Proof. Suppose U(x) = P [f ](x) with f ∈ E∗′(Sn−1). Then, U(rω) = rjfj(ω). ∞ Xj=0 If kfjkL∞(Sn−1) ≤ CeM (hj), for a fixed h > 0, the inequality (4.1) gives U(rω) ≤ ∞ Xj=0 ≤ C(cid:18)1 + fj(ω)rj = C + 6h2 (cid:19) sup A2π2 j∈N A2 h2 ∞ Xj=1 rjeM (hH 2j). 1 j2fj(ω)e−M (hj)eM (hH 2j)rj Now, and rjjp ≤ sup j∈N Therefore, by (M.2)′, sup j∈N rjeM (hH 2j) = sup p∈N (H 2h)p Mp rjjp sup j∈N ∞ Xj=0 rjjp ≤ ∞ Xj=0 (j + p)! j! rj =(cid:18) 1 1 − r(cid:19)(p) = p! (1 − r)p+1 < (p + 1)! (1 − r)p+1 . U(rω) ≤ C A H 3h(cid:18)1 + A2π2 6h2 (cid:19) sup p (p + 1)!(H 3h)p+1 Mp+1(1 − r)p+1 ≤ CCheM ∗(cid:16) H3h 1−r (cid:17). Assume now that (5.7) holds for each h > 0 (for some h > 0). Every harmonic function on Bn can be written as U(rω) = rjfj(ω), ∞ Xj=0 with each fj a spherical harmonic of degree j. By Proposition 5.1, it is enough to check that f =P∞j=0 fj ∈ E∗′(Sn−1), because in this case U = P [f ] and f would be the boundary ultradistribution of U. By Theorem 4.4, it is then suffices to verify that the sequence fj satisfies the bounds (4.11) for each h > 0 (for some h > 0). Here we use q = ∞. Fix h > 0 and assume that (5.7) holds. One clearly has fj(ω) = U(rξ)Zj(ω, ξ)dξ. 1 rjSn−1ZSn−1 14 ¯D.VU CKOVI ´C AND J. VINDAS When j = 0, we obtain f0 = U(0) and so kf0kL∞(Sn−1) ≤ CeM ∗(h). Keep now j ≥ 1. Since the zonal harmonic satisfies kZj(·, ξ)kL∞(Sn−1) = dj ≤ njn−2 [3, p. 80], we obtain, for all j ≥ 1, kfjkL∞(Sn−1) ≤ Cnjn−2 inf 0<r<1 r−jeM ∗( h 1−r ). Performing the substitution r = e−y, and using Lemma 5.2 and M∗(t) = 0 for t ≤ M1, kfjkL∞(Sn−1) ≤ CChnjn−2 exp(cid:18) inf M∗ (2h/y) + jy(cid:19) ≤ CChLnjn−2eM (2ℓhj). 0<y<∞ Finally, using the estimate (4.1), we conclude that there is C′h such that kfjkL∞(Sn−1) ≤ CC′heM (2ℓhj), for each j ∈ N. (cid:3) When Mp = p!, the bound (5.7) holds for any arbitrary harmonic function since M∗(t) = ∞ for t > 1. Hence, Corollary 5.4. Any harmonic function U ∈ H(Bn) can be written as the Poisson transform U = P [f ] of an analytic functional f on Sn−1. Suppose that Mp satisfies (NA). Consider the family of Banach spaces HMp,h(Bn) = {U ∈ H(Bn) : kUkHMp ,h(Bn) = sup x∈Bn U(x)e−M ∗( h 1−x ) < ∞}. We define the Fr´echet and (LB) spaces of harmonic functions H{Mp}(Bn) = lim←−h→0+ H{Mp},h(Bn) and H(Mp)(Bn) = lim−→h→∞ H{Mp},h(Bn). This definition still makes sense for {p!} because for h < 1 we have sup x∈Bn U(x)e−M ∗( h 1−x ) = sup x≤1−hU(x). In this case we obtain the space of all harmonic functions H(Bn) = H{p!}(Bn) with the canonical topology of uniform convergence on compact subsets of Bn. By Theorem 5.3, the mapping bv(U) = f , with f given by (5.1), provides a linear isomorphism from H∗(Bn) onto E∗′(Sn−1) if Mp satisfies (M.1)∗ and (M.2). Our proof given above actually yields a topological result: Theorem 5.5. Suppose Mp satisfies (M.1)∗ and (M.2). The boundary value map- ping is a topological vector space isomorphism with the Poisson transform bv : H∗(Bn) → E∗′(Sn−1) as inverse. P : E∗′(Sn−1) → H∗(Bn) Remark 5.6. Suppose Mp satisfies (M.0) ((NA) in the Beurling case). Theorem 5.5 is valid if one replaces (M.1)∗ by the condition (M.4) Mp ≤ Lp+1p!M∗p , p ∈ N, for some L ≥ 1. ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 15 Here M∗p is the convex regularization of Mp/p!, namely, the sequence M∗p = sup t>0 tp eM ∗(t) . In fact, p!M∗p satisfies (M.1)∗ and, under (M.4), gives rise to the same ultradistri- bution spaces as Mp. We mention that strong non-quasianalyticity (i.e., Komatsu's condition (M.3) [15]) automatically yields (M.4), as was shown by Petzsche [19, Prop. 1.1]. Furthermore, Petzsche and Vogt [20, Sect. 5] proved under the assump- tion (M.2) that (M.4) is equivalent to the so-called Rudin condition: (M.4)′′ max 1 q q≤p (cid:18)Mq q! (cid:19) ≤ A(cid:18)Mp p! (cid:19) 1 p , p ∈ N, for some A > 0, which is itself equivalent to the property that E∗(Sn−1) is inverse closed (cf. [21, 24]). 6. The support of ultradistributions on the sphere This section is devoted to characterizing the support of non-quasianalytic ul- tradistributions in terms of (uniform) Abel-Poisson summability of their spherical harmonic expansions. Our assumptions on the weight sequence are (M.1), (M.2)′, and non-quasianalyticity [15], that is, (M.3)′ Mp−1 Mp < ∞. ∞ Xp=1 Note that (NA) is automatically fulfilled because of (M.3)′ [15, Lemma 4.1, p. 56]. To emphasize we are assuming (M.3)′, we write D∗′(Sn−1) = E∗′(Sn−1). By the Denjoy-Carleman theorem [15], the support of an ultradistribution f ∈ D∗′(Sn−1) can be defined in the usual way. Since the natural inclusion D′(Sn−1) ⊂ D∗′(Sn−1) is support preserving, Theorem 6.2 below contains Gonz´alez Vieli's characterization of the support of Schwartz distributions on the sphere [13]. The key to the proof of our generalization is the ensuing lemma about the Poisson kernel. Given a non- empty closed set K ⊂ Sn−1 and a weight sequence Np, we consider the family of seminorms kϕkE {Np},h(K) = sup α∈N hαk∂α Sn−1ϕkL∞(K) Nα . Lemma 6.1. Let K1 and K2 be two disjoint non-empty closed subsets of Sn−1. Write Prω(ξ) = P (rω, ξ), regarded as a function in the variable ξ ∈ Sn−1. Then, there are two positive constants ℓ and C, only depending on K1 and K2, such that kPrωkE {p!},ℓ(K2) ≤ C(1 − r), for all ω ∈ K1 and 1 2 ≤ r < 1. Proof. For the sake of convenience, we introduce the spherical type distance d(ω, ξ) = 1 − ω · ξ. Let V ⊂ Sn−1 be open such that K1 ∩ V = ∅ and K2 ⊂ V . Set ρ = d(K1, V ). Note that if ω ∈ K1 and ξ ∈ V , the term in the denominator of the Poisson kernel, P (rω, ξ) = 1 1 − r2 , n 2 Sn−1 (1 − 2rω · ξ + r2) 16 ¯D.VU CKOVI ´C AND J. VINDAS can be estimated by using the lower bound 1 − 2rω · ξ + r2 = (1 − r)2 + 2r(1 − ω · ξ) > 2rρ. We estimate the derivatives of the Poisson kernel in spherical coordinates p(θ) where the north pole is chosen to be located at an arbitrary point of the sphere. Keep ω ∈ K1 and 1/2 ≤ r < 1 arbitrary. Let V ′ ⊂ Rn−1 be such that V = p(V ′). Call m = α. Using the estimate (3.6) and the obvious inequality mm ≤ em−1m! , we obtain k∂α θ (Prω ◦ p)kL∞(V ′) ≤ < < m 1 − r2 Sn−1 m Xk=1(cid:18)m 1 − r2 2 Sn−1 2 −2 3e 2 Sn−1 k(cid:19)km−knk Γ(cid:0) n 2 + k(cid:1) Γ(cid:0) n 2(cid:1) k(cid:19)(cid:18)1 + Xk=1(cid:18)m (1 − r)m!(cid:18)e(cid:18)n (2r)kωk sup (1 − 2rω · ξ) + r2)k+ n ξ∈V k (cid:19)k(cid:18)n ρ(cid:19)k 2 − 1 + 1(cid:19)(cid:19)m = C1(1 − r)ℓ−α α!. mm ρ n 1 n (2rρ) n n 2ρ 2 Varying the north poles, we can cover K2 by a finite number of open subsets of V , each of which parametrized by a system of invertible spherical coordinates. Inverting the polar coordinates on each of the open sets of this covering with the aid of [14, Prop. 8.1.4], we deduce that there are ℓ, C > 0, depending only on V , such that ℓαk∂α Sn−1PrωkL∞(K2) α! ≤ C(1 − r), for all α ∈ Nn. This completes the proof of the lemma. (cid:3) We are ready to state and prove our last result: Theorem 6.2. Let f = P∞j=0 fj ∈ D∗′(Sn−1) and let Ω be an open subset of Sn−1. If holds uniformly for ω on compact subsets of Ω, then Ω ⊆ Sn−1 \ supp f . Conversely, (6.1) holds uniformly on any compact subset of Sn−1 \ supp f . Proof. The first part follows immediately from Proposition 5.1. E∗(Sn−1) be an arbitrary test function such that supp ϕ ⊂ Ω. Then, Indeed, let ϕ ∈ hf, ϕi = lim r→1−ZSn−1 which gives that f vanishes on Ω. r→1−Zsupp ϕ P [f ](rω)ϕ(ω)dω = lim P [f ](rω)ϕ(ω)dω = 0, Conversely, since we have the dense and continuous embeddings E (Mp)(Sn−1) ֒→ E{Mp}(Sn−1) (by Proposition 4.2 the linear span of the spherical harmonics is dense in both spaces), we have the natural inclusion E{Mp}′(Sn−1) → E (Mp)′(Sn−1) which is obviously support preserving. Thus, we may just deal with the case f ∈ E (Mp)′(Sn−1). Let K1 be closed such that K1 ∩ supp f = ∅. Select a closed subset of the sphere K2 such that K1 ∩ K2 = ∅ and supp f ⊂ int K2. There are then C1 and h > 0 such that hf, ϕi ≤ C1kϕkE {Mp},h(K2), for all ϕ ∈ E (Mp)(Sn−1). (6.1) lim r→1− rjfj(ω) = lim r→1 P [f ](rω) = 0 ∞ Xj=1 ULTRADISTRIBUTIONAL BOUNDARY VALUES OF HARMONIC FUNCTIONS 17 The sequence Mp satisfies (NA), hence, given ℓ, one can find C2 > 0, depending only on h and ℓ, such that kϕkE {Mp},h(K2) ≤ C2kϕkE {p!},ℓ(K2) for all ϕ ∈ A(Sn−1). Using this with ϕ = Prω and employing Lemma 6.1, P [f ](rω) = hf, Prωi ≤ C1C2C(1 − r), for all ω ∈ K1 and whence (6.1) holds uniformly for ω ∈ K1. References 1 2 ≤ r < 1, (cid:3) [1] J. Alvarez, M. Guzm´an-Partida, S. P´erez-Esteva, Harmonic extensions of distributions, Math. Nachr. 280 (2007), 1443 -- 1466. [2] K. Atkinson, W. Han, Spherical harmonics and approximations on the unit sphere: An in- troduction, Springer, Berlin-Heidelberg, 2012. [3] S. Axler, P. Bourdon, W. Ramey, Harmonic function theory, Springer-Verlag, New York, 1992. [4] A. P. Calder´on, A. Zygmund, Singular integral operators and differential equations, Amer. J. Math. 79 (1957), 901 -- 921. [5] R. Carmichael, A. Kami´nski, S. Pilipovi´c, Boundary values and convolution in ultradistribu- tion spaces, World Scientific Publishing Co. Pte. Ltd., Hackensack, New Jersey, 2007. [6] R. Carmichael, D. Mitrovi´c, Distributions and analytic functions, Pitman Research Notes in Mathematics Series, 206, Longman Scientific & Technical, Harlow; John Wiley & Sons, Inc., New York, 1989. [7] G. M. Constantine, T. H. Savits, A multivariate Fa`a di Bruno formula with applications, Trans. Amer. Math. Soc. 348 (1996), 503 -- 520. [8] A. Dasgupta, M. Ruzhansky, Gevrey functions and ultradistributions on compact Lie groups and homogeneous spaces, Bull. Sci. Math. 138 (2014), 756 -- 782. [9] A. Dasgupta, M. Ruzhansky, Eigenfunction expansions of ultradifferentiable functions and ultradistributions, Trans. Amer. Math. Soc. 368 (2016), 8481 -- 8498. [10] P. Dimovski, S. Pilipovi´c, J. Vindas, Boundary values of holomorphic functions in translation- invariant distribution spaces, Complex Var. Elliptic Equ. 60 (2015), 1169 -- 1189. [11] R. Estrada, R. P. Kanwal, Distributional boundary values of harmonic and analytic functions, J. Math. Anal. Appl. 89 (1982), 262 -- 289. [12] C. Fern´andez, A. Galbis, M. C. G´omez-Collado, (Ultra)distributions of Lp-growth as boundary values of holomorphic functions, Rev. R. Acad. Cienc. Exactas F´ıs. Nat. Ser. A Mat. 97 (2003), 243 -- 255. [13] F. J. Gonz´alez Vieli, Abel means for orthogonal expansions of distributions on spheres, balls and simplices, J. Math. Anal. Appl. 433 (2016), 496 -- 508. [14] L. Hormander, The analysis of linear partial differential operators. I. Distribution theory and Fourier analysis, Springer-Verlag, Berlin, 1990. [15] H. Komatsu, Ultradistributions I: Structure theorems and a characterization, J. Fac. Sci. Tokyo Sect. IA Math. 20 (1973), 25 -- 105. [16] H. Komatsu, Ultradistributions II: The kernel theorem and ultradistributions with support in a submanifold, J. Fac. Sci. Tokyo Sect. IA Math. 24 (1977), 607 -- 628. [17] M. Morimoto, Analytic functionals on the sphere, American Mathematical Society, Provi- dence, RI, 1998. [18] M. Morimoto, M. Suwa, A characterization of analytic functionals on the sphere. II, in: Proceedings of the Second ISAAC Congress, Vol. 2 (Fukuoka, 1999), pp. 799 -- 807, Int. Soc. Anal. Appl. Comput., 8, Kluwer Acad. Publ., Dordrecht, 2000. [19] H.-J. Petzsche, On E. Borel's theorem, Math. Ann. 282 (1988), 299 -- 313. [20] H.-J. Petzsche, D. Vogt, Almost analytic extension of ultradifferentiable functions and the boundary values of holomorphic functions, Math. Ann. 267 (1984), 17 -- 35. [21] A. Rainer, G. Schindl, Composition in ultradifferentiable classes, Studia Math. 224 (2014), 97 -- 131. [22] B. C. Rennie, A. J. Dobson, On Stirling numbers of the second kind, J. Combinatorial Theory 7 (1969), 116 -- 121. 18 ¯D.VU CKOVI ´C AND J. VINDAS [23] C. Roumieu, Ultra-distributions d´efinies sur Rn et sur certaines classes de vari´et´es diff´erentiables, J. Analyse Math. 10 (1962/1963), 153 -- 192. [24] W. Rudin, Division in algebras of infinitely differentiable functions, J. Math. Mech. 11 (1962), 797 -- 809. [25] H. H. Schaefer, Topological vector spaces, Springer, New York, 1971. [26] R. T. Seeley, Spherical harmonics, Amer. Math. Monthly 73 (1966), 115 -- 121. [27] J. Vindas, R. Estrada, On the support of tempered distributions, Proc. Edinb. Math. Soc. 53 (2010), 255 -- 270. [28] ¯D. Vuckovi´c, J. Vindas, Eigenfunction expansions of ultradifferentiable functions and ultra- distributions in Rn, J. Pseudo-Differ. Oper. Appl. 7 (2016), 519 -- 531. Department of Mathematics, Ghent University, Krijgslaan 281, 9000 Ghent, Bel- gium E-mail address: [email protected] Department of Mathematics, Ghent University, Krijgslaan 281, 9000 Ghent, Bel- gium E-mail address: [email protected]
1202.5841	2	1202	2012-07-23T13:55:33	An Inverse Problem for Localization Operators	[ "math.FA", "math-ph", "math.CV", "math-ph" ]	A classical result of time-frequency analysis, obtained by I. Daubechies in 1988, states that the eigenfunctions of a time-frequency localization operator with circular localization domain and Gaussian analysis window are the Hermite functions. In this contribution, a converse of Daubechies' theorem is proved. More precisely, it is shown that, for simply connected localization domains, if one of the eigenfunctions of a time-frequency localization operator with Gaussian window is a Hermite function, then its localization domain is a disc. The general problem of obtaining, from some knowledge of its eigenfunctions, information about the symbol of a time-frequency localization operator, is denoted as the inverse problem, and the problem studied by Daubechies as the direct problem of time-frequency analysis. Here, we also solve the corresponding problem for wavelet localization, providing the inverse problem analogue of the direct problem studied by Daubechies and Paul.	math.FA	math	AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS LU´IS DANIEL ABREU AND MONIKA D ORFLER. Abstract. A classical result of time-frequency analysis, obtained by I. Daubechies in 1988, states that the eigenfunctions of a time-frequency localization operator with circular localization domain and Gaussian analysis window are the Hermite functions. In this contribution, a converse of Daubechies' theorem is proved. More precisely, it is shown that, for simply connected localization domains, if one of the eigenfunctions of a time-frequency localization operator with Gaussian window is a Hermite function, then its localization domain is a disc. The general problem of obtaining, from some knowledge of its eigenfunctions, information about the symbol of a time-frequency localization operator, is denoted as the inverse problem, and the problem studied by Daubechies as the direct problem of time-frequency analysis. Here, we also solve the corresponding problem for wavelet localization, providing the inverse problem analogue of the direct problem studied by Daubechies and Paul. October 10, 2018 1. Introduction [...] The fundamental things apply As time goes by. (Herman Hupfeld) As time goes by, most real-life signals of interest change their frequency properties. Therefore, a signal description by means of time-frequency analysis is often preferable to the signal's Fourier transform, which reliably yields frequency information, but without any localization in time. The core purpose of time-frequency analysis is to represent a given signal as a function in the time-frequency or in the time-scale plane. However, in real world applications like optics and wireless communications, one can only "sense" a signal within a certain region of those planes. This means that, in practice, the part of the signal outside the region of interest is neglected and only its "localized" version is observed. Localization operators turn this observation process into rigorous mathematical terms. They transform a given signal into one that is localized in a given region by reducing the signal energy outside that region to a negligible amount. The first approach to time-frequency localization, introduced in 1961, consists in separately selecting time- and frequency-content, and is described in a famous series Daniel Abreu was supported by the ESF activity "Harmonic and Complex Analysis and its Applications", by FCT (Portugal) project PTDC/MAT/114394/2009 and CMUC through COM- PETE/FEDER. Monika Dorfler was supported by the Austrian Science Fund (FWF):[T384-N13] Locatif. 1 2 LU´IS DANIEL ABREU AND MONIKA D ORFLER. of papers known as the "Bell labs papers". We refer to Slepian's review [20] for an account of this beautiful body of work. In 1988, Daubechies added a new perspec- tive by introducing operators, that localize directly in the time-frequency plane [4] and, together with Paul [5], extended the analysis to the time-scale plane. The time- frequency plane is associated to the short-time Fourier transform and the time-scale plane is associated to the wavelet Transform. We begin our presentation by defin- ing the short-time Fourier transform, which leads to the concept of time-frequency localization operators. respect to a window function g ∈ L2(R) is defined to be, for z = (x, ξ) ∈ R2: The short-time Fourier (or Gabor) transform of a function or distribution f with Vgf (z) = Vgf (x, ξ) = f (t)g(t − x)e−2πiξtdt (1) where the overline denotes complex conjugation. We let π(z)g(t) = g(t− x)e2πiξt and observe that f can be resynthesized from Vgf as R (cid:90) (cid:90) Given a symbol σ ∈ L1(R2), time-frequency localization operators Hσ,g are defined by f = 1 (cid:107)g(cid:107)2 2 R2 Vgf (z)π(z)g dz. (cid:90) Hσ,gf = R2 σ(z)Vgf (z)π(z)g dz = V∗ g σVgf. In signal processing it is very common to modify a signal f by acting on its time- frequency coefficients Vgf , for example in order to achieve noise reduction [14]; the cor- responding localization operators have been the object of research in time-frequency 4 e−πt2, the analysis, [7, 3]. In [4], Daubechies considered the window g(t) = ϕ(t) = 2 1 symbol σ(z) = χΩ(z), i.e. the indicator function of a set Ω ⊂ R2, and investigated the eigenvalue problem (2) HΩf := HχΩ,ϕf = λf for the case where Ω is a disc centered at zero. She concluded, that in this situation, the eigenfunctions of HχΩ,ϕ are the Hermite functions. Consequently, since, HΩ2\Ω1 = HΩ2 − HΩ1 for two sets Ω1 ⊂ Ω2, the Hermite functions are also eigenfunctions with respect to domains in the form of an annulus centered at zero and for any union of annuli. The problem (2) is important in time-frequency analysis, because its solutions are the functions with best concentration in the subregion Ω of the time-frequency plane, where we consider the time-frequency concentration of a function f in Ω ⊂ R2 defined as (3) CΩ(f ) = (cid:82) Ω Vϕf (z)2dz (cid:107)f(cid:107)2 2 . In this paper we will be concerned with the inverse situation of the one considered by Daubechies. This leads us to the following question: AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 3 • Given a localization operator with unknown localization domain Ω, can we recover the shape of Ω from information about its eigenfunctions? This is a new type of inverse problem, and we will call it the "inverse problem of time-frequency localization". We solve the problem in the case where explicit computations can be made, which is the set-up of [4]. Our main contribution is the following. Theorem 1. Let Ω ⊂ R2 be simply connected. If one of the eigenfunctions of the localization operator HΩ is a Hermite function, then Ω must be a disk centered at 0. Let us briefly discuss some motivations for our studies and consequences of our result. Hermite functions have been proposed as modulating pulses in pulse-shape modulation for ultra-wideband (UWB) communication, mainly due to their maximal joint concentration in time and frequency, cf. [19, 15, 9] and references therein. The receiver of communication system often applies a filter with the modulation pulses as eigenfunctions corresponding to large eigenvalues, in order to suppress random noise that accumulates during transmission. In this situation, Theorem 1 shows that the filter must be designed with a circular localization domain. Furthermore, if the filter on the receiver side is known or designed to have one single Hermite function as an eigenfunction, it is possible to guarantee the location of the time-frequency plane that the filter is sensing. This is particularly important in UWB communication, where the permitted spectrum is officially prescribed, cp. [6]. This last remark also hints at an additional possible application, which is system identification. Identification of linear time-variant systems is a notoriously difficult task in general, cp. [13, 2, 17]. While a linear time-invariant system is straight- forwardly identified by sending an impulse to the system and retrieving the impulse response, [10, Sec. 4.2], no similar method exists for general linear time-variant sys- tems. By our result, and for a system that is known to be a localization operator of the form HΩ for some time-frequency region Ω, we may send any Hermite function to the system and judge from the response, whether Ω can be a disk. In the positive case, one can then evaluate the size of the disk by subsequently sending additional Hermite functions and evaluating the resulting scaling factors. Obviously, this ap- proach should be extended to other shapes and its feasibility will be the topic of further research. In analogy to Theorem 1, we will also consider an inverse problem for wavelet lo- calization operators. Here, we show that the domain of localization of the localization operators investigated by Daubechies and Paul [5] is a pseudohyperbolic disc in the upper half plane whenever one of the operator's eigenfunctions is the Fourier trans- form of a Laguerre function. We will essentially use methods from complex analysis and our techniques are strongly influenced by the ideas contained in [1] and [18]. This paper is organized as follows. Section 2.1 collects some properties of the eigenfunctions of localization operators with respect to radially weighted measures and Section 2.2 deduces the geometry of localization domains under the assumption of orthogonality of any single monomial to almost all monomials. The corresponding 4 LU´IS DANIEL ABREU AND MONIKA D ORFLER. inverse problem for Gabor localization is studied in Section 3 and Section 4 is devoted to the investigation of the inverse problem for wavelet localization. 2. Double orthogonality and reproducing kernel Hilbert spaces This section is devoted to the properties of complex monomials, namely their double orthogonality with respect to any radially weighted measures and the consequences of this property. 2.1. Eigenfunctions of Localization Operators. Let Da denote a disk of radius a, 0 < a ≤ ∞, dµ(z) = µ(z)dz a radially weighted measure and dz Lebesgue measure on C. In the sequel, we will denote by Ha = L2(Da, dµ(z)) the Hilbert space of analytic functions F on C, such that (cid:90) (cid:107)F(cid:107)H = F (z)2dµ(z) Da is finite. In Proposition 1 we collect the most important facts about the "direct problem" studied in [4] [5] when transfered to the complex domain. This point of view is essentially contained in [18], but we have observed that both problems can be understood as special cases of a more general formulation with general radial measures on complex domains. This viewpoint is later reflected in our derivation of the results about the inverse problems. Proposition 1. Consider all radial measures on disks DR with radius R in the com- plex plane, i.e. the measures constituted by the weighted measure dµ(z) = µ(z)dz, defined on DR, whose weight µ(z) depends only on r = z. The following statements are true: (a): The monomials are orthogonal on any disk DR centered at zero with radius R in the complex plane and with respect to all concentric measures. Consequently, the monomials are also orthogonal on any annulus centered at zero. (b): Assume 0 < cn,a < ∞ for all moments cn,a of µ(z)dz. Then, the nor- malized monomials en,a = zn/(cid:112)(c2n+1,a) constitute an orthonormal basis for (c): If, in addition, (cid:80) n≥0(c2n+1,a)−1z2n is finite for all z ∈ Da, then Ha is a Ha. reproducing kernel Hilbert space with reproducing kernel (cid:88) n≥0 K(z, w) = (c2n+1,a)−1znwn. (d): The functions F (z) = en,a are eigenfunctions of the problem (4) F (z)K(z, w)dµ(z) = λF (w). (cid:90) DR (cid:90) DR (5) with cn,R = 2π(cid:82) R power series(cid:80) AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 5 Proof. (a) Orthogonality can directly be seen by (cid:90) R (cid:90) 2π znzmdµ(z) = r(n+m+1) 0 0 ei(n−m)θdθµ(r)dr = c2n+1,Rδn,m, 0 rnµ(r)dr. (b) Consider a domain Da, R < a ≤ ∞ such that limr→a dµ(r) = 0. Since the n≥0 anzn of an analytic function F on C converges uniformly on every DR, we may interchange integral and summation in the following equations: suppose that (cid:104)F, en,a(cid:105) = 0 for all n ∈ Z, then (cid:88) (cid:90) n≥0 an (cid:90) (cid:88) DR n≥0 lim R→a lim R→a lim R→a amc2m+1,R 0 = = = 1√ c2m+1,a 1√ c2m+1,a 1√ c2m+1,a anznzmµ(z)dz znzmµ(z)dz DR n≥0 (cid:90) which implies am = 0 for all m and hence F ≡ 0, which proves completeness of the functions {en,a} in Ha. (c) We need to show that point evaluations of F ∈ Ha are bounded. Expanding F in terms of {en,a}, we observe that (cid:104)F, en,a(cid:105) F (z) = (cid:88) ≤ (cid:107)F(cid:107)Ha · ( (cid:88) z2n) 2 . 1 1 zn√ c2n+1,a c2n+1,a n≥0 Thus, by the assumption on the growth of the moments, Ha is a reproducing kernel Hilbert space. (d) Write U for the operator which multiplies a function F ∈ H by the characteristic function of the circle DR and P for the orthogonal projection onto Ha, given by the reproducing kernel. Since P ( , we note that ) = zn√ zn√ c2n+1,a c2n+1,a 0 = en,aen,aµ(z)dz = en,aP U (en,a)dµ(z). DR Da and completeness of en,a implies that P U en,a = en,a. Denoting by K(z, w) the repro- ducing kernel of Ha, the functions F (z) = en,a are eigenfunctions of problem (4). (cid:3) Using appropriate unitary operators (the so-called Bargmann and Bergman trans- form, to be defined later in this paper), the solution to the general problem just described can be shown to be equivalent to the solution of the "direct" problems considered in [4] and [5]. Indeed, the dµ(z) = e−πz2dz case can be translated to the Gabor localization problem studied by Daubechies and the case dµ(z) = (1−z2)αdz to the wavelet localization studied by Daubechies and Paul. Details will be given in Section 3 and Section 4. (cid:90) 6 LU´IS DANIEL ABREU AND MONIKA D ORFLER. 2.2. The localization domain of monomials. We now turn to the general prob- lem, given by (4). The following, central proposition states that orthogonality of any monomial to almost all other monomials with respect to a bounded, simply connected domain Ω ⊂ C forces Ω to be a disk centered at zero. We also consider more general domains as described in Theorem 1(b). Note that we identify R2 with C for the geo- metric description. The proof is based on an idea of Zalcman [22] and is essentially similar to the proof given in [1], but in a more general setting, namely generalizing from area measure to general concentric measures. To adapt the original argument, we rely on Proposition 1. Proposition 2. Let dµ(z) be a positive, concentric measure on Da ⊆ C and consider a simply connected set Ω ⊂ Da. Assume, for some m and k ≥ 0 that (6) Then Ω must be a disk centered at zero. Ω Proof. Since z2m zkdµ(z) = λδk,0. zw z − w = − z 1 − z ∞(cid:88) = − n=1 zn wn−1 , we have for every z ∈ Ω and w such that w > sup{z ; z ∈ Ω}, the following expansion: (cid:19) . (cid:90) (cid:90) w (cid:18) (cid:90) z2m zw z − w = −z2m z + + z2 w z3 w2 + ... Integrating term wise and using (6) yields z2m zw z − w (7) Ω dµ(z) = 0, hence (8) (9) z2m z2 − zw z − w2 dµ(z) = Ω dµ(z) (cid:90) Ω z2m z z − w z2m zw z − w Ω = 1 w dµ(z) = 0. The left expression in (8) is continuous as a function of w since the integrand is locally integrable in z. Therefore, (9) holds on Ωc. We next show that 0 is inside Ω. Begin by observing that, for w > sup{z ; z ∈ Ω}, we can expand and integrate term wise so that z2m 1 z − w dµ(z) = (10) Let C > sup{z ; z ∈ Ω}. We let d(w, Ω) denote for the Euclidean distance between w and Ω, i.e. d(w, Ω) = inf w(cid:48)∈Ω w − w(cid:48). Then the following pointwise estimate in w ∈ Ωc holds: dµ(z) = λ. Ω 1 w 1 w Ω z2m w z − w (cid:90) (cid:12)(cid:12)(cid:12)(cid:12) dµ(z) ≤ C 2m . d(w, Ω) (cid:90) Ω (cid:12)(cid:12)(cid:12)(cid:12)z2m 1 z − w (cid:90) (cid:90) AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 7 (cid:90) (cid:90) (cid:90) Ω (cid:90) This allows to extend (10) by analytic continuation to Ωc. Suppose now that 0 ∈ Ωc. Then we can find a sequence of points {wn} contained in Ωc such that wn → 0. This would give 1 z2m lim n→∞ z − wn dµ(z) = lim n→∞ λm = ∞. 1 wn On the other hand, because of the continuity of the left expression in w, lim n→∞ Ω z2m 1 z − wn dµ(z) = z2m 1 z Ω dµ(z), and the integral on the right is bounded for every m ≥ 0, since we are assuming that 0 /∈ Ω. This is a contradiction and we must have 0 ∈ Ω. Finally, we can consider DR, the largest disc centered at zero and contained in Ω. Using Proposition 1(a), we can repeat the steps leading to (10) with DR instead of Ω. Pick a point w0 ∈ ∂DR ∩ ∂Ω. Then z2m z2 − Re zw0 z − w02 Ω\DR dµ(z) = 0. Since, for z ∈ Ω\DR, Re zw0 ≤ zw0 ≤ z2, the integrand is positive on Ω\DR. This forces Ω\DR to be of area measure zero, which implies Ω = DR. (cid:3) Figure 1. The situation described in Corollary 1. Ω is an annulus. 8 LU´IS DANIEL ABREU AND MONIKA D ORFLER. For the next statement, we consider a more general situation. Let γj, j = 1, . . . , n be a family of non-intersecting Jordan curves with interiors I γj such that I γj−1 ⊂ I γj for all j > 1. If n is even, set K = n If n is odd, set K = n+1 2 and let Ωk = I γ2k \ I γ2k−1 for k = 1, . . . , K. 2 and let Ω1 = I γ1 and Ωk = I γ2k−1 \ I γ2k−2 for k = 2, . . . , K. k=1 Ωk and consider the corresponding localization operator. The next corollary shows that under the double orthogonality condition (6), all curves must contain 0 in their interior. Furthermore, for n = 2, if one of the two curves is a circle, Ω must be a annulus. For the situation just described, we set Ω =(cid:83)K (a) Let (6) hold for Ω = (cid:83)K k=1 Ωk defined by a family of nested Corollary 1. Jordan curves as described above. Then all curves γj must contain zero. (b) If n = 2 and γj is a circle centered at 0 for j = 1 or j = 2, then Ω is an annulus, see Figure 1. Proof. (a) We will show by induction, that 0 must be inside all curves γj, j = 1, . . . , n. Case n = 1. Then Ω is the interior of γ1, therefore simply connected, and it follows from the proof of Proposition 2, that 0 ∈ Ω. Case n=2. Then Ω = I γ2 \ I γ1 and I γ1 is simply connected. We apply, by assuming that 0 ∈ (I γ2)c. the argument used in the first paragraph of Case n = 1 to show that 0 ∈ (Ω ∪ I γ1). Then, either 0 ∈ Ω or 0 ∈ I γ1. In the first case we consider again DR, the largest disc centered at zero contained in Ω and argue as in Case n=1 to show that Ω = DR, which contradicts the assumption that n = 2. Therefore, 0 ∈ I γ1 Arbitrary n ∈ N. Assume that, for n − 1 curves, 0 is inside all curves. For n curves, we first show that 0 ∈ I γn, assume that 0 ∈ ΩK and use, as before, the argument from Case n=1 to show that this leads to n = 1. Consequently, 0 must be inside the remaining n − 1 curves and, by induction hypothesis, inside all curves γj, j = 1 . . . n. (b) First assume that Ω is a disk, centered at zero, with a hole, in other words, that γ2 is a circle. Then, I γ2 is a disk centered and zero, such that (6) holds for I γ2 an therefore also for I γ1. Since the latter is simply connected, it must be a disk centered at 0. Now let I γ1 enclose a disk centered at 0. We then consider the largest annulus Π contained in Ω, it is given by Π = DR \ I γ1 where DR is the largest disk centered at zero and contained in I γ2. Due to Proposition 1(a), condition (6) holds on Π and we obtain (10) with Π instead of Ω. Pick a point w0 ∈ ∂DR ∩ γ2. Then (cid:90) z2m z2 − Re zw0 z − w02 dµ(z) = 0. Ω\Π and Re zw0 ≤ zw0 ≤ z2 and the integrand is positive on z ∈ Ω\Π, which implies (cid:3) Ω = Π. AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 9 3. An inverse problem for Gabor localization (cid:90) In this section we prove Theorem 1 and derive the complete solution of the classical eigenvalue problem (2) from the assumption, that any single solution is a Hermite function. In the sequel, we identify (x, ξ) with z = x + iξ and we recall that π(z)ϕ(t) = π(x, ξ)ϕ(t) = ϕ(t − x)e2πiξt. 3.1. Bargmann transform. In the Gabor case, the choice of the Gaussian function 4 e−πt2 allows the translation of the time-frequency localization operator ϕ(t) = 2 1 HχΩ,ϕ to the complex analysis set-up via the Bargmann transform B. Bf is defined for functions of a real variable as Bf (z) = z2 2 Vϕf (x,−ξ). f (t)e2πtz−πt2− π 2 z2dt = e−iπxξ+π R (11) B maps L2(R) unitarily onto F 2(C), the Bargmann-Fock space of analytic functions with the inner product obtained by choosing the measure dµ(z) = e−πz2dz. 3.2. The Hermite functions. The normalized monomials en = (πn/n!) · zn = Bhn(z) = e−iπxξ+π 2 Vϕhn(z) form an orthonormal basis for F 2(C). Here, hn(t) = dt)n(e−2πt2) are the Hermite functions, which, by appropriate choice of cn, pro- cneπt2( d vide an orthonormal basis of L2(R). As a direct consequence of the unitarity of B and Vϕ, the set {Vϕhn, n ∈ N} is orthogonal over all discs DR. 3.3. Proof of Theorem 1. We first deduce the equivalent formulation of the eigen- value problem (2) in the Bargmann domain. Since the Bargmann transform is unitary, (2) is equivalent to z2 Vϕf (z)B(π(z)ϕ)(w) dz = λBf (w) Now, since B(π(z)ϕ)(w) = e−πixξe−πz2/2eπwz, we write the previous equation as Vϕf (z)e−πixξe−πz2/2eπwzdz = λBf (w). Ω Thus, by (11), we have Bf (z)eπzw−πz2 dz = λBf (w). (cid:90) Ω (cid:90) (cid:90) (cid:90) Ω By the unitarity of the Bargmann transform we conclude that the eigenvalue problem (2) on L2(R) is equivalent to (12) on F 2 (C). We may now expand the kernel eπzw in its power series which transforms the eigenvalue equation to dz = λF (w), Ω F (z)eπzw−πz2 (cid:90) ∞(cid:88) n=0 πn n! wn Ω F (z)zne−πz2 dz (13) λF (w) = By the identity theorem for analytic functions, this implies (cid:90) πn n! wn Ω zmzne−πz2 dz. ∞(cid:88) n=0 λmwm = (cid:90) znzme−πz2 dz = λm m! πm δn,m. Ω 10 LU´IS DANIEL ABREU AND MONIKA D ORFLER. Now we use the assumption that zm solves (13) for λ = λm, in other words, that any of the solutions of (2) is a Hermite function. Setting F (z) = zm then gives In particular, setting n = m + k, (cid:90) (14) z2m zke−πz2 dz = λδk,0, for all k ≥ 1. Ω Now Proposition 2 can be applied and we conclude that Ω must be the union of n 2 annuli centered at 0 for even n and the union of a disk and n−1 2 annuli centered at 0 for odd n. In particular, for simply connected Ω, we obtain a disk centered at zero. 3.4. Consequences of Theorem 1. Corollary 2. Let Ω be simply connected. If the Gabor transform of one of the eigen- functions of the localization operator HΩ has Gaussian growth, O(e−πz2 ), then Ω must be a disk. The same conclusion holds, if some eigenfunction has Gaussian growth in both the time and the frequency domains. Proof. This is a consequence of the version of Hardy's uncertainty principle for the Gabor transform proved by Grochenig and Zimmermann [11]. They showed that, if Vgf (z) = O(e−πz2 ), then both f and g must be time-frequency shifts of a Gaussian function. Therefore, under the hypotheses of the corollary, the Gaussian (which is the first Hermite function) is an eigenfunction of the localization operator HΩ and by Theorem 1, Ω must be a disk. The second statement follows in a similar fashion from (cid:3) the classical Hardy uncertainty principle [12]. The result of Theorem 1 immediately implies that the complete solution of (2) is given by the orthonormal basis of Hermite functions. Corollary 3. Assume that an orthonormal basis of L2(R) has doubly orthogonal Gabor transform with respect to the Gaussian window ϕ and some domain Ω: (15) Vϕϕj(z)Vϕϕj(cid:48)(z)dz = cjδj,j(cid:48). Let Ω be simply connected or of the form stated in Corollary 1(b). If, for any j0, ϕj0 = hj0 is a Hermite function, then for every j ≥ 0, Ω Proof. Note that an orthonormal basis of L2(R) satisfies (15) if and only if it consists of eigenfunctions of the localization operator HΩ. Hence, we are in the situation of Theorem 1, and Ω must be disk centered at zero, the union of a disk and a finite ϕj = hj. (cid:90) AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 11 number of annuli centered at zero or an annulus centered at zero, respectively. This, (cid:3) in turn, implies that all eigenfunctions are Hermite functions. Remark 1. Note the following consequence of Theorem 1: if the localization domain Ω is not a disk, then the function of optimal concentration inside Ω, in the sense of (3), cannot be a Gaussian window. On the other hand, it is well-known that Gaussian windows uniquely minimize the Heisenberg uncertainty relation. In this sense, disks seem to be the optimal domain for measuring time-frequency concentration. Gaussian windows ϕ (and also higher order Hermite functions) are a popular choice for the basic atom in the generation of Gabor frames, whose members are given as π(λ)ϕ, λ ∈ Λ for some discrete subgroup Λ ⊂ R2. A popular choice of Λ is Λ = aZ × bZ, i.e. a rectangular lattice. In this case, the fundamental domain of Λ in R2 is rectangular and thus, according to Theorem 1, no Hermite function can be maximally concentrated inside the fundamental domain. This observation suggests that Gaussian or Hermitian windows are not an ideal choice for generating Gabor frames along a rectangular lattice. Although no proof for a precise statement exists, this observation has been made before and it is consistently confirmed in numerical experiments. In particular, in [21] it is mentioned that Gabor frames generated by time-frequency shifting Gaussian pulses over a hexagonal lattice have better condition number than frames obtained via a corresponding rectangular lattice. It is well-known and was shown by Gauss in 1840 that hexagonal lattices provide the densest packing of circles in the plane 1. On the other hand, it is known that a Gabor frame with a Gaussian basic window is never tight [8]. Motivated by our observations we formulate the following conjecture: Given a fixed redundancy red > 1, the condition number of a Gabor frame with Gaussian window ϕ is optimal for a hexagonal lattice. 3.5. Remark (Due to Karlheinz Grochenig). Since Daubechies' results also ex- tend to localization operators with symbols σ other than indicator functions, stating that any radial symbol equally leads to localization operators diagonalized by the Hermite functions, one may ask the obvious question, whether a similar inverse state- ment to Theorem 1 can be expected for more general symbol classes than indicator functions. The following example shows that this is not true. Let Hσ be a time-frequency localization operator. Then, for every N ∈ N0 there exist non-negative, non-radial symbols σ, such that HσhN = λhN . To construct σ, we proceed as in Section 3.3 and consider the equivalent operator on F 2 (C), i.e. We then claim that Tσ(zN ) = λzN for some non-radial σ. We fix N ∈ N0 and let σ(z) = σ(re2πit) = σ0(r) + σ1(r) · (e2πi(N +1)t + e−2πi(N +1)t), 1Weisstein, Eric W. "Circle Packing." From MathWorld -- A Wolfram Web Resource. http:// mathworld.wolfram.com/CirclePacking.html (cid:90) C TσF (w) = σ(z)F (z)eπzw−πz2 dz. LU´IS DANIEL ABREU AND MONIKA D ORFLER. 12 where σ0(r) ≥ 2σ1(r) ∀r ≥ 0 and (cid:82) ∞ 0 σ1(r)r3N +1e−πr2r dr = 0. Observe that σ1 can be chosen to be bounded, compactly supported and real-valued. Then we have σ(z) ≥ σ0(r)− 2σ1(r) ≥ 0. Since σ0 is radial, we have Tσ0(zN ) = λN zN with λN > 0. Therefore, it is enough to show that Tσ1(zN ) = 0. However, this is easy to see by considering σ+ = σ1(r)e2π(N +1)it and σ− = σ1(r)e−2πi(N +1)t separately and noting dwl Tσ±(zN )w=0 = 0 for that, since Tσ± is entire, the task is reduced to showing that dl all l ∈ N0. A straightforward calculation shows that, setting F (z) = zN and writing (Tσ±(f ))(l) = dl dwl Tσ±(f ), we have (Tσ±(F ))(l)(0) = σ±(z)zN (πz)le−πz2dz. (cid:90) C (cid:90) ∞ 0 (cid:90) 1 0 We finally substitute polar coordinates z = re2πit to obtain, for σ+: (Tσ±(F ))(l)(0) = πl σ1(z)rN +le−πr2r dr e2πi(2N +1−l)tdt. The integral over t is zero for l (cid:54)= 2N + 1 by orthogonality of the Fourier basis and the integral over r is zero for l = 2N + 1 by assumption. The argument for σ− is similar. 4. An inverse problem for wavelet localization By replacing "Gabor transform" by "wavelet transform" in the formulation of the inverse problem for time-frequency localization, we may define a completely analogous inverse problem for wavelet localization. The corresponding direct problem has been treated by Daubechies and Paul in [5] and by Seip in [18]. This section is related to the previous one in the same way as the direct problem studied in [5] is related to the problem studied in [4]. It is quite remarkable that, after appropriate reformulation of the eigenvalue problem, we can apply Proposition 2 to wavelet localization operators. Since our arguments depend on the connection to complex variables, it is essential to consider the Hardy space of the upper-half plane as the domain of the wavelet transform. Then, we choose a certain analyzing wavelet which plays the role of the Gaussian and the localization problem can be reformulated in certain weighted Bergman spaces. This basic strategy follows the lines which lead to the Bargmann- Fock space formulation in the Gabor case. One relevant difference between the wavelet and the Gabor case stems from the hyperbolic geometry of the upper-half plane. Since the set-up of Proposition 1 is not visible in the spaces defined on the half-plane, we will translate the problem to a conformally equivalent hyperbolic region: the unit disc. There, the problem finds a natural formulation and Proposition 2 applies. This point of view is suggested by Seip's approach in [18]. In short, while in the Gabor case the Bargmann transform maps L2(R) to the Bargmann-Fock space, where the monomials are orthogonal, B : L2(R) → F 2(C) we now need to further transform the images of the so called Bergman transform (Berα) to a space defined in the unit disc. This transformation is given by a Cayley AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 13 transform Tα, as defined in Section 4.2: (16) H 2(C+) Berα→ Aα(C+) Tα→ Aα(D). The role of the Hermite functions is taken over by special functions, whose Fourier transforms are the Laguerre functions. This is possible, since the Laguerre functions constitute an orthogonal basis for L2(0,∞) and the Fourier transform provides a unitary isomorphism H 2(C+) → L2(0,∞). 4.1. The wavelet transform. Since analyticity will play a fundamental role, in this section we restrict ourselves to functions in a subspace of L2(R), namely to f ∈ H 2(C+), the Hardy space in the upper half plane. H 2(C+) is constituted by analytic functions f such that (cid:90) ∞ −∞ sup 0<s<∞ f (x + is)2 dx < ∞. The functions in the space H 2(C+) may be considered as being of "positive frequency" since a well known Paley-Wiener theorem says that F(H 2(C+)) = L2(0,∞). For this reason it is common to study H 2(C+) on the "frequency side", where many calculations become easier. For convenience we will use a different normalization of the Fourier transform in this section, namely (Ff )(ξ) = (2π) −∞ e−iξtf (t)dt. Now consider C+ = {z ∈ C : Re(z) > 0}. For every x ∈ R and s ∈ R+, let z = x + is ∈ C+ and define 2(cid:82) ∞ − 1 Fix a function g (cid:54)= 0 such that πzg(t) = s− 1 2 g(s−1(t − x)). 0 < (cid:107)Fg(cid:107)2 L2(R+,t−1) = Cg < ∞. Such functions are called admissible and the constant Cg is the admissibility constant. Then the continuous wavelet transform of a function f with respect to a wavelet g is defined, for everyz = x + is ∈ C as (17) Let dµ+(z) denote the standard normalized area measure in C+. The orthogonal relations for the wavelet transform Wgf (z) = (cid:104)f, πzg(cid:105)H 2(C+) . C+ Wg1f1(x, s)Wg2f2(x, s)s−2dµ+(z) = (cid:104)Fg1,Fg2(cid:105)L2(R+,t−1) (cid:104)f1, f2(cid:105)H 2(C+) , (18) are valid for all f1, f2 ∈ H 2 (C+) and g1, g2 ∈ H 2 (C+) admissible. As a result, the continuous wavelet transform provides an isometric inclusion Wg : H 2(cid:0)C+(cid:1) → L2(C+,s−2dxds), (cid:90) which is an isometry for Cg = 1. (Tαf )(w) = 2 α 2 +1 (1 − w)α+2 f KαC+(z, w) = (cid:19) , i(w − 1) (cid:18) w + 1 (cid:19)α+2 . (cid:18) 1 w − z 14 LU´IS DANIEL ABREU AND MONIKA D ORFLER. 4.2. Bergman spaces. Let α > −1 and dµ+(z) denote the standard normalized area measure in C+. The Bergman space in the upper half plane, Aα(C+), is constituted by the analytic functions in C+ such that f (z)2 sαdµ+(z) < ∞. (19) Now consider D = {z ∈ C : z < 1} with normalized area measure dA(w). The Bergman space in the unit disc, denoted by Aα(D), is constituted by the analytic functions in D such that C+ (cid:90) (cid:90) (20) The map Tα : Aα(C+) → Aα(D), defined as D f (w)2 (1 − w)αdA(w) < ∞. provides a unitary isomorphism between the two spaces. The reproducing kernel of Aα(C+) is Now observe that, letting Tα act on the reproducing kernel of Aα(C+), first as a function of w and then as a function of z, we are led to the reproducing kernel of Aα(D), (21) KαD(z, w) = 1 (1 − wz)α+2 . 4.3. The Bergman transform. Let Γ denote the Gamma function and set (cid:90) ∞ 0 c2 α = t2α−1e−2tdt = 22α−1Γ(2α), We can relate the wavelet transform to Bergman spaces of analytic functions by choosing the window ψα as (22) Fψα(t) = 1[0,∞]tαe−t. 1 cα The choice of cα implies Cψα = 1 and the corresponding wavelet transform is isometric. The Bergman transform of order α is the unitary map Berα : H(C+) → Aα(C+) given by (23) Berα f (z) = s− α 2 −1Wψ α+1 2 f (−x, s) = cα α+1 2 (Ff )(t)eiztdt. t (cid:90) ∞ 0 AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 15 4.4. The Laguerre and other related systems of functions. We define the Laguerre functions in terms of the Laguerre polynomials (24) Lα n(x) = exx−α n! dn dxn n(x) = 1[0,∞](x)e−x/2xα/2Lα lα n(x). (cid:2)e−xxα+n(cid:3) = n(cid:88) k=0 (−1)k (cid:18)n + α n − k (cid:19) xk k! . By repeated integration by parts, one sees that the polynomials Lα n(x) are orthogonal on (0,∞) with respect to the weight function e−xxα . Thus, for α ≥ 0, the Laguerre n constitute an orthogonal basis for the space L2(0,∞). We will use a functions lα related system of functions ψα n defined as (cid:18) (Fψα n ) (t) = (−1)nn! (cid:19) 1 2 lα+1 n (2t). Now consider the monomials eα n(w) = 22α+2n+1Γ(n + 2 + α)Γ(2 + α) (cid:18)Γ(n + 2 + α) (cid:19) 1 2 n!Γ(2 + α) wn. We can apply Proposition 1 with µ(z) = (1 − w2)α. We conclude that {eα n}∞ forms an orthonormal basis for Aα(D) and that they are orthogonal on every disk Dr ⊂ D: for every r > 0, n=0 (25) The normalization constant C(r, m) depends on r and m and satisfies limr→1− C(r, m) = 1. Now, the functions Ψα eα n(w)eα m(w)(1 − w2)αdA(w) = C(r, m)δnm. (cid:19)α+2 n , for every n ≥ 0 and α > −1, z ∈ C+, (cid:19) 1 2(cid:18) z − i (cid:19)n(cid:18) 1 (cid:18)Γ(n + 2 + α) , n!Γ(2 + α) z + i z + i Ψα n(z) = 1 4α+ 1 2 (cid:90) Dr (cid:90) (cid:12)(cid:12)(cid:12) z1−z2 (cid:90) ∞ z1−z2 0 are conveniently choosen such that (26) Thus, a change of variables w = z−i n )(w) = eα (TαΨ2α n(w). z+i in (25) leads to Ψα (z,i)<r (27) n(z)Ψα n(z)sαdµ+(z) = C(r, m)δnm, (cid:12)(cid:12)(cid:12) is the pseudohyperbolic metric on C+. Moreover, the unitar- where (z1, z2) = ity of the operator Tα translates the basis property of {eα In other words, (26) shows that {Ψα we observe that (23) together with the special function formula n=0 in Aα(D) to Aα(C+). n=0 is an orthogonal basis of Aα(C+). Finally n(z)}∞ n}∞ (28) xαLα n(x)e−xsdx = Γ(α + n + 1) n! s−α−n−1(s − 1)n 16 gives LU´IS DANIEL ABREU AND MONIKA D ORFLER. For an intuitive grasp of this section, keep in mind that to the composition of trans- forms (16) one associates the transformations of the basis functions: Berα ψα n = Ψα n. n ∈ H(C+) Berα→ Ψα ψα n ∈ Aα(C+) Tα→ eα n ∈ Aα(D). 4.5. The inverse problem. We now consider the wavelet localization operator P∆,g defined as (cid:90) P∆,αf = Wψ α+1 2 f (z)πzψ α+1 2 (t)dµ+(z) ∆ and set up the corresponding eigenvalue problem (29) P∆,αf = λf n}, then ∆ must be a pseudohyperbolic disc centered at i. Theorem 2. If one of the eigenfunctions of the localization operator P∆,α belongs to the family {ψα Proof. We first rewrite the eigenvalue problem (29). A simple change of variables on the "Fourier" side of the wavelet representation gives Berα(πzψ α+1 2 )(w) = mαs α+2 2 α+2 2 KαC+(z, w), = s (cid:90) ∆ (cid:90) where mα = α+1 2α . Now apply the Bergman transform and use (23) to rewrite (29) as Berα f (z)KαC+(z, w)sαdµ+(z) = λBerα f (w). By the unitarity Berα : H(C+) → Aα(C+) we conclude that our eigenvalue problem is equivalent to F (z)KαC+(z, w)sαdµ+(z) = λF (w), with F ∈ Aα(C+). Making the change of variables ∆ (cid:19)α+2 (cid:18) 1 z − w zD = , wD = z − i z + i w − i w + i , ∞(cid:88) (cid:90) (cid:90) we move the eigenvalue problem to the unit disc (TαF )(zD) Ω=Tα∆ (1 − zD)α (1 − wDzD)2+α dAD(zD) = λ(TαF )(wD), where TαF (wD) ∈ Aα(D). We simplify the notation writing zD = z, wD = z. Now, using the uniformly convergent expansion of the reproducing kernel, n=0 we can transform the eigenvalue equation into 1 (1 − wz)2+α = eα n(w)eα n(z), (30) λ(TαF )(w) = Γ(n + 2 + α) n!Γ(2 + α) wn Ω (TαF )(z)zn(1 − z)αdAD(z) ∞(cid:88) n=0 AN INVERSE PROBLEM FOR LOCALIZATION OPERATORS 17 If one of the eigenfunctions of the localization operator P∆,α belongs to the family {ψα n}, then Tα(Berαψα n )(z) = Tα(Ψα n)(z) = eα n(z) solves (30) for λ = λn. Setting (TαF )(z) = eα m(z) gives ∞(cid:88) n=0 (cid:90) Ω λmwm = Γ(n + 2 + α) n!Γ(2 + α) wn zmzn(1 − z)αdAD(z). (cid:90) Ω (cid:90) Comparison of coefficients yields znzme−πz2 (1 − z)αdAD(z) = λm n!Γ(2 + α) Γ(n + 2 + α) δn,m and further, with n = m + k, the condition of Proposition 2: (31) z2m zk(1 − z)αdAD(z) = λδk,0, for all k ≥ 1. Ω Hence Proposition 2 can be applied and Ω must be a disk centered at zero. Now we can go back to the upper half plane by the change of variables u = i z + 1 1 − z ∈ C+, which maps 0∈ D to i ∈ C+ and leaves the pseudohyperbolic metric invariant. Finally, notice that the condition z < r, can be written in terms of the pseudohyperbolic metric of the disc as D(z, 0) < r. Hence, the disc Ω centered at zero is mapped to the pseudohyperbolic disc ∆ = {C+(u, i) < r} centered at u = i. (cid:3) Remark 2. We can draw a conclusion similar to the one in Remark 1 after Theorem 1. Indeed, it follows from Theorem 2 that, if the localization domain ∆ is not a pseudo- hyperbolic disc, then the function f providing optimal concentration in the sense of maximizing (32) C∆(f ) = (cid:82) ∆ Wψαf (z)2dz (cid:107)f(cid:107)2 H 2(C+) cannot be the function ψα in (22) - the so called Cauchy wavelet. On the other hand it is known that the functions ψα minimize the affine uncertainty principle as first mentioned in [16]. In this sense, pseudohyperbolic discs seem to be optimal domains to measure wavelet localization. We would like to thank Saptarshi Das for his precious advice on UWB technology. Acknowledgments 18 LU´IS DANIEL ABREU AND MONIKA D ORFLER. References [1] M. E. Andersson. An inverse problem connected to double orthogonality in Bergman spaces. Math. Proc. Cambridge Philos. Soc., 128(3):535 -- 538, 2000. [2] P. A. Bello. Measurement of random time-variant linear channels. IEEE Trans. Inform. Theory, 15(4):469 -- 475, 1969. [3] E. Cordero and K. Grochenig. Time-frequency analysis of localization operators. J. Funct. Anal., 205(1):107 -- 131, 2003. [4] I. Daubechies. Time-frequency localization operators: a geometric phase space approach. IEEE Trans. Inform. Theory, 34(4):605 -- 612, July 1988. [5] I. Daubechies and T. Paul. Time-frequency localisation operators -- a geometric phase space approach. II. The use of dilations. Inverse Problems, 4(3):661 -- 680, 1988. [6] FCC, revision of part 15 of the commissions rules regarding ultra wideband transmission sys- tems, et docket 98-153, fcc 02-48, feb. 14, 2002. [7] H. G. Feichtinger and K. Nowak. A Szego-type theorem for Gabor-Toeplitz localization opera- tors. Michigan Math. J., 49(1):13 -- 21, 2001. [8] J. P. Gabardo. Weighted irregular Gabor tight frames and dual systems using windows in the Schwartz class. J. Funct. Anal., 256(3):635 -- 672, 2009. [9] M. Ghavami, L. B. Michael, S. Haruyama, and R. Kohno. A novel UWB pulse shape modulation system. Wireless Personal Communications, 23:105 -- 120, 2002. 10.1023/A:1020953424161. [10] K. Grochenig. Foundations of Time-Frequency Analysis. Appl. Numer. Harmon. Anal. Birkhauser Boston, 2001. [11] K. Grochenig and G. Zimmermann. Hardy's theorem and the short-time Fourier transform of Schwartz functions. J. London Math. Soc., 63(1):205 -- 214, February 2001. [12] G. H. Hardy. A theorem concerning Fourier transforms. Journal L. M. S., 8:227 -- 231, 1933. [13] T. Kailath. Time-variant communication channels. IEEE Trans. Inform. Theory, 9(4):233 -- 237, 1963. [14] S. Mallat. A Wavelet Tour of Signal Processing: The Sparse Way. Academic Press, 2009. [15] B. Parr, B. Cho, K. Wallace, and Z. Ding. A novel ultra-wideband pulse design algorithm. Communications Letters, IEEE, 7(5):219 -- 221, may 2003. [16] T. Paul. Affine Coherent States and the Radial Schrdinger Equation I: Radial Harmonic Os- cillator and Hydrogen Atom. Technical Report CPT-84/P-1710, Centre de Physique Thorique, Marseille, 1984. [17] G. E. Pfander and D. F. Walnut. Measurement of time-variant channels. IEEE Trans. Inform. Theory, 52(11):4808 -- 4820, November 2006. [18] K. Seip. Reproducing formulas and double orthogonality in Bargmann and Bergman spaces. SIAM J. Math. Anal., 22(3):856 -- 876, 1991. [19] J. Silva and M. Campos. Spectrally efficient UWB pulse shaping with application in orthogonal PSM. IEEE Trans. on Communications, 55/2:313 -- 322, 2007. [20] D. Slepian. Some comments on Fourier Analysis and Uncertainty and Modelling. SIAM Rev., 25/Nr.3:379 -- 393, 1983. [21] T. Strohmer and S. Beaver. Optimal OFDM system design for time-frequency dispersive chan- nels. IEEE Trans. Comm., 51(7):1111 -- 1122, July 2003. [22] L. Zalcman. Some inverse problems of potential theory. Contemporary Mathematics., 63:337 -- 350, 1987. Institut fur Mathematik, Universitat Wien, Alserbachstrasse 23 A-1090 Wien, on leave from Department of Mathematics of University of Coimbra, Portugal. E-mail address: [email protected] Institut fur Mathematik, Universitat Wien, Alserbachstrasse 23. E-mail address: [email protected]
1109.3806	1	1109	2011-09-17T17:27:07	On greedy algorithms with respect to generalized Walsh system	[ "math.FA" ]	In this paper we proof that there exists a function f(x) belongs to L^1[0,1] such that a greedy algorithm with regard to generalized Walsh system does not converge to f(x) in L^1[0,1] norm, i.e. the generalized Walsh system is not a quasi-greedy basis in its linear span L^1[0,1].	math.FA	math	On greedy algorithms with respect to generalized Walsh system S. A. Episkoposian e-mail: [email protected] In this paper we proof that there exists a function f (x) belongs to L1[0, 1] such that a greedy algorithm with regard to generalized Walsh system does not converge to f (x) in L1[0, 1] norm, i.e. the generalized Walsh system is not a quasi-greedy basis in its linear span L1[0, 1]. Keywords: greedy algorithms, generalized Walsh system, quasi-greedy basis. 2000 Mathematics Subject Classification: Primary 42C10; Secondary 46E30. 1 Introduction In this paper we consider a question of convergence of greedy algorithm with regard to generalized Walsh system in L1[0, 1] norm. {φk}∞ Let X be a Banach space with a norm · = · X and a basis Φ = k=1, φkX = 1, k = 1, 2, .. . Denote by Σm the collection of all functions in X which can be expressed as a linear combination of at most m- functions of Φ. Thus each function g ∈ Σm can be written in the form asφs, #Λ ≤ m. g = Xs∈Λ 1 For a function f ∈ X we define its approximate error by σm(f, φ) = inf g∈Σm f − gX , m = 1, 2, ... and we consider the expansion f = ∞ Xk=1 ak(f )φk . Definition 1. Let an element f ∈ X be given. Then the m-th greedy approximant of the function f with regard to the basis Φ is given by Gm(f, φ) = Xk∈Λ ak(f )φk, where Λ ⊂ {1, 2, ...} is a set of cardinality m such that an(f ) ≥ ak(f ), n ∈ Λ, k /∈ Λ, (1) (2) We'll say that the greedy approximant of f (t) ∈ Lp [0,1], p ≥ 0 converges with regard to the basis Φ , if the sequence Gm(x, f ) converges to f (t) in Lp norm. This new and very important direction invaded many mathematician's attention (see [1]-[10]). Definition 2. We call a basis Φ greedy basis if for every f ∈ X there exists a subset Λ ⊂ {1, 2, ...} of cardinality m, such that f − Gm(f, Φ)X ≤ C · σm(f, Φ) where a constant C = C(X, Φ) independent of f and m. In [4] it is proved that each basis Φ which is Lp -equivalent to the Haar basis H is Greedy basis for Lp(0, 1), 1 < p < ∞ . Definition 3. We say that a basis Φ is Quasi-Greedy basis if there exists a constant C such that for every f ∈ X and any finite set of indices Λ, having the property we have min k∈Λ ak(f ) ≥ max n /∈Λ ak(f ) SΛ(f, Φ)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) Xk∈Λ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)X ak(f )φk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 ≤ C · f X. In [5] it is proved that a basis Ψ is quasi-greedy if and only if the sequence {Gm(f )} converges to f , for all f ∈ X. Not that the trigonometric and Walsh system are not a quasi-greedy basis for Lp if 1 < p < ∞ (see [7] and [8] ). 2 Definition and properties of generalized Walsh system Let a denote a fixed integer, a ≥ 2 and put ωa = e 2πi a . Now we will give the definitions of Rademacher and generalized Walsh systems (see [12] ). Definition 4. The Rademacher system of order a is defined by ϕ0(x) = ωk a if x ∈ " k a , k + 1 a ! , k = 0, 1, ..., a − 1, and for n ≥ 0 ϕn(x + 1) = ϕn(x) = ϕ0(anx). (3) Definition 5. The generalized Walsh system of order a is defined by ψ0(x) = 1, and if n = αn1an1 + ... + αnsans where n1 > ... > ns, then ψn(x) = ϕαn1 n1 (x) · ... · ϕαns ns (x). (4) Let Ψa = {ψn(x)}∞ that Ψ2 is the classical Walsh system. n=0 denote the generalized Walsh system of order a. Not Remark. The generalized Walsh system Ψa, a ≥ 2 is a complete or- thonormal system in L2[0, 1) (see [12]). The basic properties of the generalized Walsh system of order a are ob- tained by R. Paley , H.E.Chrestenson, J. Fine, N. Vilenkin and others (see [11]- [16]). 3 Define In,k = In,k(a) = " k an , k + 1 an ! , k = 0, ..., an − 1, n = 1, 2, .... If ϕn(x) is the nth Rademacher function of order a, then from Definition 4 it follows ϕn(x) = ωk a = e 2πi·k a , x ∈ In+1,k. Note some properties of generalized Walsh system: Property 1. From definition 5 we have ψak+j(x) = ϕk(x) · ψj(x), if 0 ≤ j ≤ ak − 1. Denote by Dn(t) = n−1 Xk=0 ψk(t), (5) (6) (7) the Dirichlet kernel by generalized Walsh system. Property 2. The Dirichlet kernel has the following properties (see [12] ) Dan(t) =   an, x ∈ In,0 = [0, 1 an ); 0, x ∈ [ 1 an , 1). (8) Property 3. If n = ak + m, 0 ≤ m < k and consequently by (6) - (8) we have Dn(t) = Dak(t) + ϕk(t) · Dm(t), t ∈ [0, 1). (9) Property 4. For any natural number m and any t ∈ (0, 1) the following is true Dm(t) ≤ m. (10) 4 3 A Basic Lemma Denote by Lk = Z 1 0 Dk(t)dt the kth Lebesgue constant of the generalized Walsh system {Ψa}. In [12] it is proved that the Lebesgue constant satisfy Lk = O(loga k) where O depends upon a. Next Lemma shows that there exists a sequence of natural numbers {nk} so that the sequence Lnk has the same order of growth as loga nk. Namely the following is true: Lemma . There exists a sequence of natural numbers {nk}∞ k=1 of the form such that n2s = s Xi=0 a2i : n2s+1 = s Xi=0 a2i+1, s = 0, 1, 2, ..., (11) ak ≤ nk < ak+1, Lnk = Z 1 0 Dnk(t)dt > 1 a · k 2 + 1! > 1 2a · loga nk, k ≥ 1. (12) Proof. Note that n2s = a2s+2 − 1 a2 − 1 < a2 a2 − 1 · a2s, n2s+1 = a2s+3 − a a2 − 1 = a · a2s+2 − 1 a2 − 1 < a2 a2 − 1 · a2s+1 i.e. nk < a2 a2 − 1 · ak < ak+1, k = 0, 1, 2, ... (13) From this and (10) we have Dnk(t) < a2 a2 − 1 · ak, t ∈ [0, 1), k = 0, 1, 2, ... (14) First we'll prove that Z 1 1 ak+2 Dnk(t)dt > 1 a · k 2 + 1! (15) 5 Let k = 2s, then we have to prove Z 1 1 a2s+2 Dn2s(t)dt > 1 a · (s + 1) , s = 0, 1, 2, ... (16) By Definition 5 and (7) for s = 0 we have Z 1 1 a2 D1(t)dt = Z 1 1 a2 ψ0(t)dt = 1 − 1 a2 > 1 a . Now assume that for some s − 1 the inequality (16) holds, i.e. Z 1 1 a2s Dn2(s−1)(t)dt > 1 a · s. By (8) and (14) we get Da2s(t) = a2s, if t ∈ I2s,0. Dn2(s−1)(t) < a2 a2 − 1 · a2(s−1) = a2s a2 − 1 . From (11) it follows that n2s = a2s + n2(s−1) and consequently by (9), (17) and (18) for t ∈ I2s,0 we have Dn2s(t) = Da2s(t) + ϕ2s(t) · Dn2(s−1)(t) ≥ Da2s(t) − Dn2(s−1)(t) > a2 − 2 a2 − 1 · a2s. (17) (18) (19) (20) Hence, taking into account that (cid:16) 1 a2s+2 , 1 a2s(cid:17), we obtain a2s(cid:17) ⊂ I2s,0 = h0, 1 · a2s ·(cid:18) 1 a2s − 1 a2s+2(cid:19) = a2 − 2 a2 − 1 1 a2s 1 a2s+2 Z Dn2s(t)dt > a2 − 2 a2 − 1 · a2 − 1 a2s+2 = a2 − 2 a2 ≥ 1 a . (21) From (8), (9), (17) and (20) follows Z 1 1 a2s Dn2s(t)dt > Z 1 1 a2s Dn2(s−1)(t)dt > 1 a · s. 6 Hence and from (21) we conclude 1 a2s+2 Z 1 Dn2s(t)dt +Z 1 1 a2s Dn2s(t)dt = Dn2s(t)dt > · s > 1 a · (s + 1) , s = 0, 1, 2, ... (22) 1 a2s 1 a2s+2 + 1 a Z 1 a In a case k = 2s + 1 (s = 0, 1, ...) we have to prove Z 1 1 a2s+3 Dn2s+1(t)dt > 1 a ·(cid:18)s + 1 2 + 1(cid:19) . For s = 0 this inequality holds because in this case n2s+1 = n1 = a and Z 1 1 a3 1 a 1 a3 Da(t)dt = Z ·(cid:18)a − 1 a a adt = a ·(cid:18) 1 a(cid:19) ≥ 1 a 3 2 1 · . − 1 a3(cid:19) = The next reasonings are similar to a case when k = 2s. Since ak ≤ nk < ak+1 then loga nk < k + 1 and consequently + 1! > Dnk(t)dt > Z 1 Lnk = Z 1 Dnk(t)dt > · k 1 a ak+2 2 0 1 1 2a · k 2 + 1! > 1 2a · loga nk. Completing the proof. 7 4 The Main Theorem and It's Proof. In [10] we proved the following theorem: Theorem 1. Let a sequence {Mn}∞ n=1 be given so that lim k→∞ (M2k − M2k−1) = +∞. Then the Walsh subsystem {Wnk(x)}∞ k=1 = {Wm(x) : M2s−1 ≤ m ≤ M2s, s = 1, 2, ...} (1) is not a quasi-greedy basis in its linear span in L1[0, 1]. Let Ψa = {ψn(x)}∞ From Corollary 2.3 (see [8]) it follows that generalized Walsh system is n=0 denote the generalized Walsh system of order a. not a quasi-greedy basis for Lp[0, 1] if 1 < p < ∞. In this paper we prove the following theorem. Theorem 2. There exists a function f (x) belongs to L1[0, 1) such that the approximate Gn(f, Ψa) with regard to the generalized Walsh system does not converge to f (x) by L1[0, 1) norm, i.e. the generalized Walsh system is not a quasi-greedy basis in its linear span in L1. Proof. Let a ≥ 2 denote a fixed integer. For any natural k we set ak2 fk(x) = −1 Xi=a(k−1)2 (cid:18) 1 k2 + 2−i(cid:19) · ψi(x). (23) It is easy to see that the Fourier coefficients by generalized Walsh system of the function fk(x) are defined as follows C (k) i = 1 k2 + 2−i if a(k−1)2 ≤ i < ak2 . (24) Now we consider the following function f (x) = ∞ Xi=1 Ciψi(x) = fk(x) = ∞ Xk=1 8 = ∞ Xk=1   where ak2 −1 Xi=a(k−1)2 (cid:18) 1 k2 + 2−i(cid:19) · ψi(x)  Ci = C (k) i if a(k−1)2 ≤ i < ak2 . , (25) (26) Now we will show that f (x) ∈ L1[0, 1]. For this we represent the function f (x) in the following way: f (x) = g(x) + h(x), (27) where g(x) = ∞ Xk=1 1 k2   Xk=1 ∞ ak2 −1 Xi=a(k−1)2 ∞ 1 = ψi(x) Xk=1  2−i · ψi(x)  −1 ak2  Xi=a(k−1)2  h(x) = k2 hDak2 (x) − Da(k−1)2 (x)i , = ∞ Xj=1 2−j · ψj(x). For the function g(x) and from (8) and definition 5 we have Z 1 0 g(x)dx ≤ 2 · 1 k2 < ∞ ∞ Xk=1 which means g(x) ∈ L1[0, 1]. Analogously Z 1 0 h(x)dx ≤ 1 2j < ∞ ∞ Xj=1 i.e. h(x) ∈ L1[0, 1]. Hence and from (27) it follows that f (x) ∈ L1[0, 1]. For any natural k we choose numbers i, j so that a(k−1)2 ≤ i < ak2 ≤ j < a(k+1)2 . Then and from (24) we have C (k+1) 1 (k + 1)2 + 2−j < (f ) < C (k) (f ). j i 1 k2 + 2−i, 9 Analogously for any number i, a(k−1)2 ≤ i < ak2, we have 1 k2 + 2−(i+1) < 1 k2 + 2−i, (f ). i.e. C (k) i+1(f ) < C (k) i Thus for any natural numbers i we get Ci+1(f ) < Ci(f ). On the other hand if i → ∞ then k → ∞ (see (24)). Then from (24) and (26) we get Ci(f ) ց 0. For any numbers mk so that a(k−1)2 + mk < ak2 , (28) by (24) - (26) and Definition 1 we have Ga(k−1)2 +mk (f, Ψa) − Gak2 (f, Ψa) = a(k−1)2 +mk −1 Xi=a(k−1)2 C (k) i · ψi(x) = 1 k2 a(k−1)2 +mk−1 Xi=a(k−1)2 ψi(x)+ a(k−1)2 +mk−1 Xi=a(k−1)2 1 2i · ψi(x) = J1 + J2. (29) Taking into account (6) and (7) we get J1 = 1 k2 a(k−1)2 +mk −1 Xi=a(k−1)2 ψi(x) = 1 k2 mk −1 Xi=0 ψa(k−1)2 +i(x) = 1 k2 · ψa(k−1)2 (x) · mk−1 Xi=0 ψi(x) = 1 k2 · ψa(k−1)2 (x) · Dmk(x). J2 ≤ a(k−1)2 +mk−1 Xi=a(k−1)2 1 2i ψi(x) ≤ ∞ Xi=a(k−1)2 1 2i ≤ 2−a(k−1)2 +1. From this and (29) we obtain 10 1 (f, Ψa) − Gak2 (f, Ψa) ≥ Ga(k−1)2 +mk k2 · ψa(k−1)2 (x) Dmk(x) −2−a(k−1)2 +1. k2 · Dmk(x) −2−a(k−1)2 1 +1 = (30) Now we take the sequence of natural numbers mν defined by Lemma (see (11) (12))such that a(k−1)2 ≤ mν < a(k−1)2+1. Then from (30) we have Ga(k−1)2 +mk (f, Ψa) − Gak2 (f, Ψa) dx > Z 1 0 1 k2 ·Z 1 Dmk (x)dx − 2−a(k−1)2 +1 ≥ 0 (k − 1)2 2a · k2 − 2−a(k−1)2 +1 ≥ 1 4a 1 2a · k2 · loga mk − 2−a(k−1)2 − 2−a(k−1)2 +1 ≥ C1, k ≥ 4. +1 ≥ Thus the sequence {Gn(f, Ψa)} does not converge by L1[0, 1] norm, i.e. the generalized Walsh system Ψa is not a quasi-greedy basis in its linear span in L1[0, 1]. Completing the proof. 11 R E F E R E N C E S [1] DeVore R. A., Temlyakov V. N., Some remarks on Greedy Algorithms, Adv. Comput. Math., 1995, v.5, p.173-187. [2] DeVore R. A., Some remarks on greedy algorithms. Adv Comput. Math., 1996, v.5, p. 173-187. [3] Davis G., Mallat S. and Avalaneda M., Adaptive greedy approximations. Constr, Approx., 1997, v.13, p. 57-98. [4] Temlyakov V. N., The best m - term approximation and Greedy Algo- rithms, Advances in Comput. Math., 1998, v.8, p.249-265. [5] Wojtaszcyk P., Greedy Algorithm for General Biorthogonal Systems, Journal of Approximation Theory, 2000, v.107, p. 293-314. [6] Konyagin S. V., Temlyakov V. N., A remark on Greedy Approximation in Banach spaces, East Journal on Approximation, 1999, v.5, p. 493-499. [7] Temlyakov V. N., Greedy Algorithm and m - term trigonometric approx- imation, Constructive Approx., 1998, v.14, p. 569-587. [8] Gribonval R., Nielsen M., On the quasi-greedy property and uniformly bounded orthonormal systems , http://www.math.auc.dk/research/reports/R-2003-09.pdf. [9] Grigorian M.G., "On the convergence of Greedy algorithm", International Conference, Mathematics in Armenia, Advances and Perspectives,Yerevan, Armenia, Abstract, 2003, p. 44 - 45. [10] Episkoposian S.A., On the divergence of Greedy algorithms with respect to Walsh subsystems in L1 , Nonlinear Analysis: Theory, Methods & Appli- cations, 2007, v. 66, Issue 8, p. 1782-1787. 12 [11] Pely R., A remarkable systems of orthogonal functions, Proc. London Math. Soc., 1932, v.34, p. 241-279. [12] Chrestenson H.E., A class of generalized Walsh functions, Pacific J. Math., 1955, v.45, p. 17-31. [13] Fine J., The generalized Walsh functions Trans. Amer. Math. Soc., 1950, v.69, p. 66-77. [14] Vilenkin N., On a class of complete orthonormal systems, AMS Transl., 1963, v.28, p. 1-35. [15] Selfridge R., Generalized Walsh transform, Pacific J. Math., 1955, v.5, p. 451-480. [16] Watari C., On generalized Walsh-Fourier series, Tohoku Math. J., 1958, v.10, p. 211-241. Department of Mathematics, State Engineering University of Armenia Yerevan, Terian st. 105, e-mail: [email protected] 13
1708.09306	2	1708	2017-09-18T13:01:39	Sharp Hardy and Rellich type inequalities on Cartan--Hadamard manifolds and their improvements	[ "math.FA" ]	In this paper, we prove several new Hardy type inequalities (such as the weighted Hardy inequality, weighted Rellich inequality, critical Hardy inequality and critical Rellich inequality) for radial derivations (i.e., the derivation along the geodesic curve) on Cartan--Hadamard manifolds. By Gauss lemma, our new Hardy inequality are stronger than the classical one. We also established the improvements of these inequalities in terms of sectional curvature of underlying manifolds which illustrate the effect of curvature to these inequalities. Furthermore, we obtain some improvements of Hardy and Rellich inequality in hyperbolic space $\mathbb H^n$. Especially, we show that our new Rellich inequality is indeed stronger the classical one in hyperbolic space $\mathbb H^n$.	math.FA	math	Sharp Hardy and Rellich type inequalities on Cartan–Hadamard manifolds and their improvements Van Hoang Nguyen∗ September 19, 2017 Abstract In this paper, we prove several new Hardy type inequalities (such as the weighted Hardy inequality, weighted Rellich inequality, critical Hardy inequality and critical Rellich inequal- ity) for radial derivations (i.e., the derivation along the geodesic curve) on Cartan–Hadamard manifolds. By Gauss lemma, our new Hardy inequality are stronger than the classical one. We also established the improvements of these inequalities in terms of sectional curvature of underlying manifolds which illustrate the effect of curvature to these inequalities. Further- more, we obtain some improvements of Hardy and Rellich inequality in hyperbolic space Hn. Especially, we show that our new Rellich inequality is indeed stronger the classical one in hyperbolic space Hn. 7 1 0 2 p e S 8 1 ] . A F h t a m [ 2 v 6 0 3 9 0 . 8 0 7 1 : v i X r a 1 Introduction The motivation of this paper is to prove some new Hardy and Rellich type inequalities on Cartan– Hadamard manifolds (i.e., complete, simply connected Riemannian manifolds with non-positive sectional curvature). All our obtained inequalities are in sharp forms. They are stronger than and generalize several classical inequalities in the Euclidean space Rn. Let n ≥ 2 and p ∈ (1, n), the classical Hardy inequality in Rn states that (cid:18) n − p p (cid:19)pZRn fp xp dx ≤ZRn ∇fpdx, f ∈ C∞0 (Rn). (1.1) The constant (n − p)p/pp in (1.1) is sharp. A similar inequality with the same best constant also holds if Rn is replaced by any domain Ω containing the origin. The Hardy inequality (1.1) plays important role in several branches of mathematics such as the partial differential equations, spectral theory, geometry etc. We refer the reader to [3, 4, 11] for reviews of this subject and to [5, 16] for the improvements of this inequality when Rn is replaced by the bounded domains containing the origin. In the critical case p = n, the inequality (1.1) fails for any constant. However, the following inequality holds in the unit ball B with center at origin of Rn ZB ∇fndx ≥ (n − 1)n nn ZB fn xn(cid:16)1 + ln 1 x(cid:17)n dx, f ∈ C∞0 (B), (1.2) ∗Institut de Math´ematiques de Toulouse, Universit´e Paul Sabatier, 118 Route de Narbonne, 31062 Toulouse c´edex 09, France. Email: [email protected], [email protected] 2010 Mathematics Subject Classification: 26D10, 31C12, 53C20, 53C21. Key words and phrases: Hardy inequality, Rellich inequality, critical Hardy inequality, critical Rellich inequality, Cartan–Hadamard manifolds, sharp constant 1 for any n ≥ 2 (see, e.g., [14]). The constant (n − 1)n/nn is the best constant in (1.2). It also was shown in [14] that (1.2) is equivalent to the critical case of the Sobolev–Lorentz inequality. It is remarked that (1.2) is not invariant under the scaling as (1.1). A scaling invariant version of (1.2) (nowaday called the critical Hardy inequality) was recently established by Ioku and Ishiwata [25], (n − 1)n nn ZB fn xn(cid:16)ln 1 x(cid:17)n dx ≤ZB(cid:12)(cid:12)(cid:12)(cid:12) x x · ∇f (x)(cid:12)(cid:12)(cid:12)(cid:12) n dx, f ∈ C∞0 (B). (1.3) Again, the constant (n − 1)n/nn is sharp. The inequality (1.3) in bounded domains was also discussed in [25]. It is surprise that the critical Hardy inequality (1.3) is equivalent to the Hardy inequality (1.1) in larger dimension spaces (see [45]). We refer the readers to [26] for a global scaling invariant version of (1.3) which we do not mention here. Recently, there is an enourmous work to generalize the Hardy inequality to many different settings. For examples, the fractional Hardy inequality was established in [17–19, 38, 47] and references therein. The Hardy inequality also was proved on group structure, e.g., on Heisenberg groups [8,10,21,40], on polarisable groups [22], on Carnot groups [28,33], on stratified groups [7,41], and on more general homogeneous groups [42–44]. The Hardy inequality on Riemannian manifold (M, g) was studied by Carron [6] in the weighted L2−form under some geometric assumption on the weighted function ρ. More precisely, he proved the following inequality (γ − 1)2 4 ZM u2 ρ2 dVg ≤ZM ∇gu2 gdVg, u ∈ C∞0 (M ), where ρ is a nonnegative function on M such that ∇gρg = 1, ∆gρ ≥ γ/ρ, and dVg,∇g, ∆g and ·g denote the volume element, gradient, Laplace–Beltrami operator and the length of a vector field with respect to the Riemannian metric g on M respectively, and the set ρ−1(0) is a compact set of zero capacity. Under the same hypotheses on the function ρ, Kombe and Ozaydin [29] extended the result of Carron to the case p 6= 2 and they presented an application to the hyperbolic space Hn with ρ being the distance function from the origin point 0. We refer the reader to [9] and the references therein for more results in this direction. The sharp Hardy inequality on Cartan–Hadamard manifolds (M, g) was recently obtained by Yang, Su and Kong [48] ZM fp ρp+β dVg ≤(cid:18) p n − p − β(cid:19)pZM ∇gfp g ρβ dVg, f ∈ C∞0 (M ), (1.4) where n ≥ 3, p ∈ (1, n), β < n − p and ρ(x) = d(x, P ) for a fix point P in M (we refer the reader to Section 2 for more about on Cartan–Hadamard manifolds). Furthermore, the constant pp/(n − p − β)p appeared in (1.4) is sharp. Our first aim in this paper is to establish an stronger version of (1.4) as follows: let ∂ρ denote the derivation along the geodesic curves starting from P , then our first result is as follows ZM fp ρp+β dVg ≤(cid:18) p n − p − β(cid:19)pZM ∂ρfp ρβ dVg, f ∈ C∞0 (M ) with the constant pp/(n− p− β)p being sharp. Obviously, our result is stronger than (1.4) because of Gauss's lemma ∂ρf ≤ ∇gfg. This inequality also extends a result of Ioku et al. [27] (see also [35]) on Rn to the Cartan–Hadamard manifolds. We also study the Hardy inequality in the critical case p = n − β and obtain the following inequality (cid:18) p − 1 p (cid:19)pZB1(P ) fp ρn(cid:16)ln 1 ρ(cid:17)p dVg ≤ZB1(P ) ∂ρfp ρn−p dVg, f ∈ C∞0 (B1(P )) with (p − 1)p/pp being the sharp constant and B1(P ) denoting the geodesic unit ball around P . The case p = n gives us an extension of the inequality (1.3) to the Cartan–Hadamard manifolds. 2 In my knowledge, the critical Hardy inequality on Cartan–Hadamard manifolds seems to be new in literature. Rellich inequalities are the higher order derivative version of Hardy inequality. Let us denote (n − 2p − β − 2ip)(n(p − 1) + β + 2ip)! for l ≥ 1, p ∈ (1, n/(2l)) and −n(p − 1) < β < n − 2lp. It was proved in [12] that c(n, 2l, β, p) = l−1 Yi=0 p2 p ZRn fp xβ+2lp dx ≤ c(n, 2l, β, p)ZRn ∆lfp xβ dx, f ∈ C∞0 (Rn) (1.5) (1.6) if l ≥ 1, p ∈ (1, n/(2l)) and −2(p − 1) < β < n − 2lp, and (n − p − β)p c(n, 2l, p + β, p)ZRn xβ+(2l+1)p dx ≤ ZRn fp pp ∇∆lfp xβ dx, f ∈ C∞0 (Rn) (1.7) if l ≥ 1, p ∈ (1, n/(2l + 1)) and 2 − 3p < β < n − (2l + 1)p. Again, these inequalities (1.6) and (1.7) are shown to be sharp in [12]. To prove (1.6) and (1.7), Davies and Hinz firstly proved (1.6) for l = 1 and then iterating this inequality together with the weighted Hardy inequality. The iterative approach is now a standard method to study the inequality of this type. The other approaches are given by Mitidieri [37] based on the divergence theorem and the Rellich–Pohozaev type identities [36]. For the improvements of (1.6) and (1.7) on bounded domains and p = 2, the reader may consult the paper [46]. The Rellich inequality was generalized to the Riemannian manifolds in the work of Kombe and Ozaydin [29] for p = 2 ∆gu2 ρα dVg ≥ ZM (C − α − 3)2(C + α + 1)2 16 u2 ρα+4 dVg, ZM u ∈ C∞0 (M ) (1.8) where ρ is a nonnegative weight function such that ∇gρg = 1, ∆gρ ≥ C/ρ with α > −2 and α < C − 3. In the case of Cartan–Hadamard manifold (M, g), the inequality (1.8) was extended by Yang, Su and Kong [48] to p 6= 2 and to higher order derivatives with ρ(x) = d(x, P ) the geodesic distance from x to a fix point P ∈ M . The results of Yang, Su and Kong [48] read as follows: ZM fp ρβ+2lp dVg ≤ c(n, 2l, β, p)ZM gfp ∆l ρβ dVg, f ∈ C∞0 (M ), (1.9) with l ≥ 1, p ∈ (1, n/(2l)) and −2(p − 1) < β < n − 2lp, and (n − p − β)p c(n, 2l, β + p, p)ZM ρβ+(2l+1)p dVg ≤ ZM fp pp ∇g∆l gfp ρβ g dVg, f ∈ C∞0 (M ), (1.10) with l ≥ 1, p ∈ (1, n/(2l + 1)) and 2− 3p < β < n− (2l + 1)p. Yang, Su and Kong also proved that these inequalities are sharp on M . Our next results show that the inequalities (1.9) and (1.10) also true with the same best constants if we replace the Laplace–Beltrami operator by the radial Laplace ∆g,ρ taked along the geodesic curve starting from P (see Section 2 below for the precise definition of ∆g,ρ). The following sharp inequalities will be proved in this paper ZM fp ρ2lp+β dVg ≤ c(n, 2l, β, p)ZM g,ρfp ∆l ρβ dVg, f ∈ C∞0 (M ), (1.11) if l ≥ 1, p ∈ (1, n/(2l)) and n(1 − p) < β < n − 2lp, and (n − p − β)p c(n, 2l, p + β, p)ZM ρ(2l+1)p+β dVg ≤ ZM fp pp g,ρfp ∂ρ∆l ρβ dVg, f ∈ C∞0 (M ), (1.12) 3 if l ≥ 1, p ∈ (1, n/(2l + 1)) and n − (n + 1)p < β < n − (2l + 1)p. It is emphasized here that the domain of β is extended in your inequalities comparing with (1.9) and (1.10). In the critical β = n − 2lp or n = (2l + 1)p, we will prove the following critical Rellich type inequalities which we believe to be new in Cartan–Hadamard manifolds p 1 n − 2i − 2! l−1 Yi=0 ZB1(P ) g,ρfp ∆l ρ(x)n−2lp dVg, p 1 n − 2i − 2! ZB1(P ) ∂ρ∆l ρ(x)n−(2l+1)p dVg, g,ρfp (1.13) (1.14) (l − 1)! ZB1(P ) fp ρ(x)n(cid:16)ln 1 with p′ = p/(p − 1), and fp ρ(x)n(cid:16)ln 1 ρ(x)(cid:17)p dVg ≤ p′ 21−l ρ(x)(cid:17)p dVg ≤ p′ 1 ZB1(P ) 2ll! l−1 Yi=0 for any f ∈ C∞0 (B1(P ). We also show that these inequalities (1.13) and (1.14) are sharp. It is worthy to note that our inequalities (1.13) and (1.14) contain the critical Rellich inequalities on the Euclidean space Rn due to Adimurthi and Santra [2]. In fact, Adimurthi and Santra proved (1.13) and (1.14) for radial functions supported in the unit ball centerd at origin in Rn and for radial function f , ∆g,ρf is exactly its Laplace. We should mention here that in the setting of homogeneous groups, the inequalities of the type (1.11), (1.12), (1.13) and (1.14) was recently established by the author in [39] which extend the result of Ruzhansky and Suragan [42–44] to the higher order derivatives. The last remark is that if the sectional curvature KM of (M, g) satisfies KM ≤ −b ≤ 0 along each plane section at each point of M , then all our obtained inequalities in this paper will be strengthened. In this situation, we will prove the quantitative versions of Hardy, Rellich, critical Hardy and critical Rellich type inequalities on Cartan–Hadamard manifolds in terms of the upper bounded of KM (see Sections 3 and 4 below). Especially, in the case of hyperbolic space Hn with KHn ≡ −1, we show that our obtained Rellich inequality is stronger than the classical one in [29] (see Section 5 below). This is an immediate consequence of the following Machihara–Ozawa– Wadade type result in hyperbolic space (see Theorem 5.2) ZHn ∆g,ρf2dVg ≤ZHn ∆gf2dVg, f ∈ C∞0 (Hn). In Euclidean space, the previous inequality was prove by Machihara, Ozawa and Wadade [35]. We also prove several improved Hardy and Rellich type inequalities in hyperbolic space Hn (see Theorems 5.1, 5.3 and 5.4 below). The rest of this paper is organized as follows. In Section 2, we recall some basic notions and properties of Riemannian manifolds (especially, of Cartan–Hadamard manifolds) which we frequently use in this paper. In Sectiona 3, we prove Hardy type inequalities on Cartan–Hadamard manifolds such as weighted Hardy inequality, critical Hardy inequality. We also establish the quantitative improvements for these Hardy type inequalities. Section 4, we prove Rellich type inequalities (both in the critical case and subcritical case) and their quantitative version on Cartan– Hadamard manifolds. In the last section, we establish some improvements of Hardy and Rellich inequalities in hyperbolic spaces. 2 Preliminaries In this section, we list some useful properties of Riemannian manifolds which will be used in this paper. We refer the reader to the book of Helgason [23] and the book of Gallot, Hulin, and Lafontaine [20] for standard references. Let (M, g) be an n−dimensional complete Riemannian manifold. In a local coordinate system {xi}n Xi,j=1 i=1, we can write gijdxidx. g = n 4 n ∆g = Xi,j=1 1 ∂ ∂xi (cid:18)pggij ∂ ∂xj(cid:19) , pg In such a local coordinate system, the Laplace-Beltrami operator ∆g with respect to the metric g is of the form where g = det(gij) and (gij) = (gij)−1. Let us denote by ∇g the corresponding gradient. Then h∇gf,∇ghi = gij ∂f ∂xi ∂h ∂xj . n Xi,j=1 For simplicity, we shall use the notation ∇gfg =ph∇gf,∇gfi. Let KM be the sectional curvature on (M, g). A Riemannian manifold (M, g) is called Cartan– Hadamard manifolds if it is complete, simply connected and with non-positive sectional curvature, i.e., KM ≤ 0 along each plane section at each point of M . If (M, g) is Cartan–Hadamard manifold, then for each point P ∈ M , M contains no points conjugate to P . Moreover, the exponential map ExpP : TP M → M is a diffeomorphism, where TP M is the tangent space to M at P (see, e.g., [23, Chapter I]) Let (M, g) be a Cartan–Hadamard manifold. Fix a point P ∈ M and denote by ρ(x) = d(x, P ) for all x ∈ M , where d denotes the geodesic distance on M . Note that ρ(x) is smooth on M \{P} and satisfies Moreover, since ExpP is a diffeomorphism, then the function ∇gρ(x)g = 1, x ∈ M \ {P}. ρ(x)2 = kExp−1 P (x)k2 ∈ C∞(M ). The radial derivation ∂ρ = ∂ on M by ∂ρ along geodesic curve starting from P is defined for any function f ∂ρf (x) = d(f ◦ ExpP ) dr (Exp−1 P (x)), here d dr denotes the radial derivation on TP M , i.e., d dr F (u) =(cid:28) u u ,∇F (u)(cid:29) , u ∈ TP M \ {0}. For any δ > 0, denote by Bδ(P ) = {x ∈ M : ρ(x) < δ} the geodesic ball in M with center at P and radius δ. We introduce the density function J(u, t) of the volume form in normal coordinates as follows (see, e.g., [20, pp. 166 − 167]). Choose an orthonormal basis {u, e2, . . . , en} on TP M and let c(t) = ExpP (tu) be a geodesic curve. The Jacobian fields {Yi(t)}n i=2 satisfy Yi(0) = 0, Y ′i (0) = ei, so that the density function can be given by J(u, t) = t1−nqdet(hYi(t), Yj (t)i), t > 0. We note that J(u, t) does not depend on {e2, . . . , en} and J(u, t) ∈ C∞(TP M \ {0}) by the definition of J(u, t). Moreover, if we set J(u, 0) ≡ 1 then J(u, t) ∈ C(TP M ) and has the following asymptotic expansion (2.1) J(u, t) = 1 + O(t2) as t → 0+ since Yi(t) has the asymptotic expansion (see, e.g., [20, p. 169]) Yi(t) = tei − t3 6 R(c′(t), ei)c′(t) + o(t3), as t → 0+, where R(·,·) is the curvature tensor on M . 5 From the definition of J(u, t), we have the following polar coordinate on M ZM f dVg =Z ∞ 0 ZSn−1 f (ExpP (tu))J(u, t)tn−1dtdu, (2.2) where du denotes the canonical measure of the unit sphere of TP M . Moreover, the Laplacian of the distance function ρ(x) has the following expansion via the function J(u, t) (see, e.g., [20, 4.B.2]) ∆gρ = n − 1 ρ + J′(u, ρ) J(u, ρ) , ρ > 0, where J′(u, ρ) = ∂J(u,ρ) ∂ρ . Therefore, for any radial function f (ρ) on M , we have ∆gf (ρ) = f′′(ρ) +(cid:18) n − 1 ρ + J′(u, ρ) J(u, ρ)(cid:19) f′(ρ). Let us introduce the radial Laplacian of a function f by (2.3) (2.4) ∆g,ρf (x) = ∂2 ρf (x) +(cid:18) n − 1 ρ + J′(u, ρ) J(u, ρ)(cid:19) ∂ρf (x), x = ExpP (tu). (2.5) Note that if KM is constant then J(u, t) depends only on t. We denote by Jb(t) the corresponding density function if KM ≡ −b for some b ≥ 0, i.e., (cid:16) sinh(√bt) √bt (cid:17)n−1 For b ≥ 0, we consider the function ctb : (0,∞) → R defined by if b = 0 if b > 0, √b coth(√bt) t 1 Jb(t) =  ctb(t) =( 1 and the function Db : [0,∞) → R defined by Db(t) =(0 if b = 0 if b > 0. if t = 0 if t > 0. tctb(t) − 1 Clearly, we have Db ≥ 0. If the section curvature KM on M satisfies KM ≤ −b then the Bishop– Gunther comparison theorem (see, e.g., [20, p. 172] for its proof) says that J′(u, t) J(u, t) ≥ J′b(t) Jb(t) = n − 1 t Db(t), t > 0. (2.6) 3 Hardy type inequalities This section is devoted to proved the Hardy type inequalities on Cartan–Hadamard manifolds (M, g). By (2.6), the function t 7→ ρ(u, t) is non-decreasing monotone on (0,∞) for any u ∈ Sn−1. We first have the following weighted Hardy inequalities Theorem 3.1. Let (M, g) be a n−dimensional Cartan–Hadamard manifolds. Suppose that n ≥ 2, p ∈ (1, n) and β < n − p. There holds for any f ∈ C∞0 (M ) n − p − β(cid:19)pZM ∂ρf (x)p ρ(x)β dVg. ZM (3.1) p f (x)p ρ(x)p+β dVg ≤(cid:18) n−p−β )p is the best constant in (3.1). p Furthermore, the constant ( 6 By Gauss's lemma (see, e.g., [20]), we have ∂ρf ≤ ∇gfg, f ∈ C1(M ). g p dVg, ρ(x)β Hence, the inequality (3.1) is stronger than the weighted Hardy inequality obtained in [29] (see also [48]) ∇gf (x)p f ∈ C∞0 (M ) n − p − β(cid:19)pZM f (x)p ρ(x)p+β dVg ≤(cid:18) ZM for any β < n − p. The inequality (3.1) was proved by Ioku et al. [27] and by Machihara et al. [34] when M = Rn. Furthermore, they proved inequality (3.1) with an extra sharp remainder term. In the proof of the inequality (3.1) below, we also get the inequality (3.1) with extra remainder terms: one comes from the density function J(u, t) and the other is as in the one of Ioku, Ishiwata and Ozawa [27] and Machihara, Ozawa and Wadade [34]. In order to prove Theorem 3.1, let us introduce the quantity Rp(ξ, η) for p > 1 and ξ, η ∈ TP M by 1 pηp + p − 1 p ξp − ξp−2hξ, ηi. Rp(ξ, η) = (3.2) This quantity closely relates to the remainder term in (3.1) (see, also, [27, 34]). By the convexity of ξ → ξp we see that Rp(ξ, η) ≥ 0 with equality if and only if ξ = η. Furthermore, we can see that Rp(ξ, η) = (p − 1)Z 1 0 tξ + (1 − t)ηp−2tdtξ − η2. Proof of Theorem 3.1. Let f ∈ C∞0 (M ), then the function F (y) = f (ExpP (y)) ∈ C∞0 (TP M ). Using the polar coordinate (2.2) and the assumption β < n − p, we get F (ρu)pJ(u, ρ)ρn−p−β−1dρdu ρ(x)p+β dVg =ZSn−1Z ∞ f (x)p ZM 0 F (ρu)pJ(u, ρ)dρn−p−βdu Fp−2F ∂ρF J(u, ρ)ρn−p−βdρdu FpJ′(u, ρ)ρn−p−βdρdu 0 1 1 p = n − p − β ZSn−1Z ∞ n − p − β ZSn−1Z ∞ = − n − p − β ZSn−1Z ∞ fp−2f n − p − β ZM n − p − β ZSn−1Z ∞ = − ρ(x) − − p−1 p 1 0 0 0 p (p+β) ∂ρf β p ρ(x) dVg FpJ′(u, ρ)ρn−p−βdρdu. Using the definition of Rp, we then have − p n − p − β ZM ρ(x) = p−1 p (p+β) fp−2f ZM Rp p − 1 −ZM p ∂ρf dVg β p ρ(x) fp ρ(x)p+β dVg + f ρ(x)1+ β p ,− 7 1 p p(cid:18) n − p − β n − p − β(cid:19)pZM p ! dVg. ρ(x) ∂ρf p β ∂ρfp ρ(x)β dVg (3.3) (3.4) Combining (3.3) and (3.4) together, we get ZM f (x)p ρ(x)p+β dVg =(cid:18) − p n − p − β(cid:19)pZM n − p − β ZM p ∂ρfp ρ(x)β dVg − pZM fp ρ(x)p+β J(ux, ρ(x)) J′(ux, ρ(x))ρ(x) Rp(cid:16) f ρ(x) ,− p ρ(x)β n−p−β ∂ρf(cid:17) dVg dVg, (3.5) here for x ∈ M\{P}, we denote by ux the unique unit vector in TP M such that x = ExpP (ρ(x)ux). The inequality (3.1) is now an immediate consequence of (3.5) by dropping the nonnegative remainder terms. It remains to check the sharpness of constant. To do this, we approximate the function ρ−(n−p−β)/p as follows. Let φ ∈ C∞0 (R) such that 0 ≤ φ ≤ 1, φ(t) = 1 if t ≤ 1 and φ(t) = 0 if t ≥ 2. For ǫ > 0, define fǫ(x) = φ(ρ(x))(1 − φ(ǫ−1ρ(x)))ρ(x)− n−p−β p . A straightforward computation shows that ZM fǫ(x)p ρ(x)p+β dVg =ZSn−1Z ∞ ≥ZSn−1Z 1 2ǫ 0 φ(t)p(1 − φ(ǫ−1t))pJ(u, t)t−1dt t−1dtdu ≥ −Sn−1 ln(2ǫ), here we use the increasing monotonicity of J(u, t). Consequently In the other hand, we have lim ǫ→0+ZM fǫ(x)p ρ(x)p+β dVg = ∞. ∂ρfǫ = φ′(ρ(x))(1 − φ(ǫ−1ρ(x)))ρ(x)− n−p−β p − ǫ−1φ′(ǫ−1ρ(x))φ(ρ(x))ρ(x)− n−p−β p n − p − β p − φ(ρ(x))(1 − φ(ǫ−1ρ(x)))ρ(x)− n−β p . Easy computations shows that p ρβ φ′(ρ(x))(1 − φ(ǫ−1ρ(x)))ρ(x)− n−p−β ǫ−1φ′(ǫ−1ρ(x))φ(ρ(x))ρ(x)− n−p−β ZM ZM φ(ρ(x))(1 − φ(ǫ−1ρ(x)))ρ(x)− n−β p p ρβ p ρ(x)β p dVg = O(1), p dVg = O(1), dx =ZM fǫ(x)p ρ(x)p+β dVg. and ZM These estimates prove the sharpness of (3.1). It is worthy to note that the proof above of Theorem 3.1 also gives a quantitative Hardy type inequality on Cartan–Hadamard manifolds. Theorem 3.2. Suppose the assumptions in statement of Theorem 3.1 and suppose that KM ≤ −b ≤ 0. Then for any f ∈ C∞0 (M ), we have ∂ρfp ρβ dVg ≥(cid:18) n − p − β Db(ρ)(cid:19) dVg. fp ρp+β (cid:18)1 + (n − 1)p n − p − β (cid:19)pZM ZM (3.6) p 8 Consequently, the following quantitative Hardy inequalities hold g ZM p ρβ ∇gfp dVg ≥(cid:18) n − p − β ZM ∂ρfp ρβ dVg ≥(cid:18) n − p − β (cid:19)pZM fp ρp+β (cid:18)1 + (cid:19)pZM fp ρp+β dVg + 3b(n − 1)(cid:18) n − p − β (cid:19)p−1ZM p p (n − 1)p n − p − β Db(ρ)(cid:19) dVg, (3.7) fp ρp+β−2(π2 + bρ2) dVg. (3.8) and ZM ∇gfp g ρβ p dVg ≥(cid:18) n − p − β (cid:19)pZM fp ρp+β dVg + 3b(n − 1)(cid:18) n − p − β p (cid:19)p−1ZM fp ρp+β−2(π2 + bρ2) dVg. (3.9) The inequalities (3.7) and (3.9) was proved by Krist´aly for p = 2 and β = 0 in [32, Theorems 4.1 and 4.2] on Cartan–Hadamard manifolds. (3.7) was extended to Finsler–Hadamard manifolds again for p = 2 by Krist´aly and Repovs in [31, Lemma 3.1] (see also [15] for a particular form). Obviously, (3.7) and (3.9) provide the improvements of the Hardy inequality on Cartan–Hadamard manifolds due to Yang, Su and Kong [48, Theorem 3.1] if b > 0. Proof. The inequality (3.6) is followed from (3.5) and (2.6). The inequality (3.7) is consequence of (3.6) and Gauss lemma. The inequality (3.8) is derived from (3.6) and the simple fact (see the proof of Theorem 1.4 in [32]) t coth(t) − 1 ≥ 3t2 π2 + t2 , t > 0. The inequality (3.9) is followed from (3.8) and Gauss lemma (or from (3.7) and the simple fact above). We next consider the critical case β = n − p. In this case, the inequality (3.1) fails for any In order to establish the inequalities in this case, we need add an extra logarithmic constant. term. The next result of this section reads as follows. Theorem 3.3. Let n ≥ 2 and let M be an n-dimensional Cartan–Hadamard manifold. Then, for any p ∈ (1,∞), there holds (cid:18) p − 1 p (cid:19)pZB1(P ) ∂ρfp ρ(x)n−p dVg, (3.10) fp ρ(x)n(cid:16)ln 1 ρ(x)(cid:17)p dVg ≤ZB1(P ) p )p is the best constant in (3.10). for any function f ∈ C∞0 (B1(P )). Furthermore, the constant ( p−1 Proof. Suppose that f ∈ C∞0 (B1(P )) then where B1 denotes the unit ball in TP M with center at origin. Applying polar coordinate (2.2), we F (y) = f (ExpP (y)) ∈ C∞0 (B1), 9 have ZB1(P ) fp ρ(x)n(cid:16)ln 1 ρ(x)(cid:17)p dVg =ZSn−1Z 1 0 F (ρu)pJ(u, ρ)ρ−1 (− ln ρ)−p dρdu 0 1 1 p = p − 1ZSn−1Z 1 p − 1ZSn−1Z 1 = − p − 1ZSn−1Z 1 p − 1ZB1(P ) p − 1ZB1(P ) = − − − p 1 0 F (ρu)pJ(u, ρ)d (− ln ρ)1−p du Fp−2F ∂ρF J(u, ρ) (− ln ρ)p−1 dρdu 0 F (ρu)pJ′(u, ρ) (− ln ρ)1−p dρdu fp−2f ρ(x)n p−1 p (− ln ρ)p−1 fp J′(ux, ρ) ρ(cid:17)p J(ux, ρ) ρn(cid:16)ln 1 ∂ρf ρ(x) n p −1 dVg 1 ρ dVg. ρ ln From the definition of Rp, we get fp−2f ρ(x)n p−1 p − p p − 1ZB1(P ) p − 1 = p ZB1(P ) Rp(cid:18) −ZB1(P ) ∂ρf ρ(x) n p −1 dVg 1 ρ(x)(cid:17)p dVg + (− ln ρ)p−1 fp ρ(x)n(cid:16)ln 1 f ρ(x) n p (− ln ρ(x)) ,− p − 1 p − 1(cid:19)pZB1(P ) p(cid:18) p p −1(cid:19) dVg. ρ(x) ∂ρf n p ∂ρfp ρ(x)n−p dVg Combining (3.11) and (3.12) together implies (3.11) (3.12) dVg (3.13) ZB1(P ) fp ρn(cid:16)ln 1 ρ(cid:17)p dVg =(cid:18) p p − 1(cid:19)pZB1(P ) p − 1ZB1(P ) p ∂ρfp ρn−p dVg − pZB1(P ) fp ρ(cid:17)p ρn(cid:16)ln 1 J′(ux, ρ) J(ux, ρ) ρ ln − Rp(cid:16) f ρ ln 1 ρ p−1 ∂ρf(cid:17) ,− p ρn−p 1 ρ dVg. The inequality (3.10) is now followed from (3.13) by dropping the nonnegative remainder terms. It remains to check the sharpness of (3.10). Let ϕ be a cut-off function in (−1, 1), i.e., ϕ ∈ C∞0 ((−1, 1)) such that 0 ≤ ϕ ≤ 1, ϕ(t) = 1 if t ≤ 1/2. For δ > 0 small enough, define fδ(x) =(cid:18)ln 1 ρ(x)(cid:19) p−1 p −δ ϕ(ρ(x)). Firstly, it follows from (2.2) and the increasing monotonicity of J(u, ρ) that ZB1(P ) fδp ρn(cid:16)ln 1 ρ(cid:17)p dVg =ZSn−1Z 1 0 (cid:18)ln ≥ Sn−1Z 1/2 0 (cid:18)ln = Sn−1 1 pδ (ln 2)−pδ . 1 ρ ρ(x)(cid:19)−1−pδ ϕ(ρ)p ρ(x)(cid:19)−1−pδ 1 ρ 1 dρ J(u, ρ)dρdu 10 In the other hand, the straightforward computations show that We thus have ∂ρfδ = − (cid:16)ln 1 p−1−pδ p ZB1(P ) and Consequently, we obtain lim δ→0+ZB1(P ) p − 1 − pδ p (cid:18)ln ϕ(ρ)p ρn−p p ρ(cid:17)− 1+pδ (cid:12)(cid:12)(cid:12)(cid:12)(cid:16)ln 1 ρ(cid:17) ZB1(P ) δ→0+ RB1(P ) RB1(P ) lim 1 fδp ρ(cid:17)p dVg = ∞. ρn(cid:16)ln 1 ρ(cid:19)− 1+pδ ϕ(ρ) +(cid:18)ln ρ(cid:19) (cid:19)pZB1(P ) dVg =(cid:18) p − 1 − pδ p 1 p p p−1−pδ p−1−pδ p ρn−p p dVg = O(1). ϕ′(ρ)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:19)p =(cid:18) p − 1 . ∂ρfδp ρn−p dVg fδp ρ )p dVg ρn(ln 1 ϕ′(ρ), fδp ρn(cid:16)ln 1 ρ(cid:17)p dVg, This proves the sharpness of (3.10). Theorem 3.3 together with Gauss's lemma yields the following critical Hardy type inequalities for full gradient on M , g (3.14) ∇gfp ρ(x)n−p dVg, (cid:18) p − 1 p (cid:19)pZB1(P ) ρ(x)(cid:17)p dVg ≤ZB1(P ) fp ρ(x)n(cid:16)ln 1 for any function f ∈ C∞0 (B1(P )). Using again the test functions in the proof of Theorem 3.3, we see that the constant (p − 1)p/pp in (3.14) is sharp. Similar to the subcritical case, we also obtain from the proof of Theorem 3.3 the following quantitative critical Hardy inequalities whose proof is completely similar with the one of Theorem 3.2. Theorem 3.4. Suppose the assumptions in statement of Theorem 3.3 and suppose KM ≤ −b ≤ 0. Then the following inequalities hold for any function f ∈ C∞0 (B1(P )) (n − 1)p p − 1 (n − 1)p p − 1 fp ρ )p (cid:18)1 + ρn(ln 1 fp ρ )p (cid:18)1 + ρn(ln 1 ρ(cid:19) dVg, ρ(cid:19) dVg, Db(ρ) ln Db(ρ) ln (3.15) (3.16) 1 1 ZB1(P ) ZB1(P ) ZB1(P ) ∂ρfp ρn−p dVg ≥(cid:18) p − 1 ∇gfp ρn−p dVg ≥(cid:18) p − 1 ∂ρfp ρn−p dVg ≥(cid:18) p − 1 g p (cid:19)pZB1(P ) p (cid:19)pZB1(P ) p (cid:19)pZB1(P ) + 3b(n − 1)(cid:18) p − 1 p (cid:19)pZB1(P ) + 3b(n − 1)(cid:18) p − 1 fp ρn(ln 1 ρ )p dVg p (cid:19)p−1ZB1(P ) fp ρn(ln 1 ρ )p dVg p (cid:19)p−1ZB1(P ) 11 and ZB1(P ) g ∇gfp ρn−p dVg ≥(cid:18) p − 1 fp ρ )p−1(π2 + bρ2) ρn−2(ln 1 dVg, (3.17) fp ρ )p−1(π2 + bρ2) ρn−2(ln 1 dVg. (3.18) Especially, in the case p = n, we obtain from Theorems 3.3 and 3.4 the following critical Hardy inequalities and their quantitative versions. Corollary 3.5. Suppose the assumptions in the statement of Theorem 3.3. Then, there holds for any function f ∈ C∞0 (B1(P )) (cid:18) n − 1 n (cid:19)nZB1(P ) (3.19) fn ρ(x)n(cid:16)ln 1 fn ρ(x)n(cid:16)ln 1 ρ(x)(cid:17)n dVg ≤ZB1(P ) ∂ρfndVg, ρ(x)(cid:17)n dVg ≤ZB1(P ) ∇gfn g dVg. (3.20) and (cid:18) n − 1 n (cid:19)nZB1(P ) Furthermore, the constant ( n−1 n )n is the best constant in (3.19) and (3.20). Suppose, in addition, KM ≤ −b ≤ 0, then we have n (cid:19)nZB1(P ) (n − 1)n nn−1 ZB1(P ) ZB1(P ) ∂ρfndVg ≥(cid:18) n − 1 + 3b fn ρn(ln 1 ρ )n dVg fn ρ )n−1(π2 + bρ2) ρn−2(ln 1 dVg, (3.21) and ZB1(P ) ∇gfn g dVg ≥(cid:18) n − 1 fn ρn(ln 1 ρ )n dVg n (cid:19)nZB1(P ) (n − 1)n nn−1 ZB1(P ) + 3b fn ρ )n−1(π2 + bρ2) ρn−2(ln 1 dVg (3.22) In Euclidean space (i.e., M = Rn), the inequality (3.19) was proved by Ioku and Ishiwata [25], and then was extended for any p > 1 by Ruzhansky and Suragan [43] (i.e, the inequality (3.10) in the Rn). More plus, Ruzhansky and Suragan was generalized the inequality (3.10) to any homogeneous groups and any homogeneous quasi-norm with the same best constant. 4 The Rellich type inequalities In this section, we study the Rellich type inequalities on Cartan–Hadamard manifolds (M, g). The following result will play an important role in our analysis below. Lemma 4.1. Let (M, g) be an n−dimensional Cartan–Hadamard manifold. Suppose that n ≥ 2, p ∈ (1, n) and −n(p − 1) < β < n − p. There holds for any f ∈ C∞0 (M ) ZM fp ρ(x)p+β dVg ≤(cid:18) p n(p − 1) + β(cid:19)pZM ∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) J(ux,ρ(x) )fp ρ(x)β dVg. (4.1) Furthermore, the constant ( p n(p−1)+β )p is the best constant in (4.1). Proof. Suppose f ∈ C∞0 (M ), then F (y) = f (ExpP (y)) ∈ C∞0 (TP M ). 12 It follows from the polar coordinate (2.2) and integration by parts that ZM ρ(x)p+β dVg =ZSn−1Z ∞ f (x)p FpJ(u, ρ)ρn−p−β−1dρdu Fp−2F ∂ρF J(u, ρ)ρn−p−βdρdu FpJ′(u, ρ)ρn−p−βdρdu 0 p = − − = − − 0 1 n − p − β ZSn−1Z ∞ n − p − β ZSn−1Z ∞ fp−2f n − p − β ZM n − p − β ZM ρ(x) fp ρ(x)p+β p−1 p 1 0 p (p+β) ∂ρf β p ρ(x) dVg J′(ux, ρ(x))ρ(x) J(ux, ρ(x)) dVg, (4.2) here we use β < n − p. In the other hand, we have ZM fp−2f ρ p−1 p (p+β) ∂ρf β p ρ dVg =ZM ∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) J(ux,ρ(x)) )f β p ρ(x) dVg p (p+β) p−1 ρ(x) fp−2f − (n − 1)ZM fp ρp+β dVg −ZM fp ρ(x)p+β J′(ux, ρ(x))ρ(x) J(ux, ρ(x)) dVg. (4.3) Plugging (4.3) into (4.2), we obtain ∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) J(ux,ρ(x)) )f dVg ZM p p−1 = ρ(x) fp−2f fp ρ(x)p+β dVg n(p − 1) + β ZM fp p − 1 n(p − 1) + β ZM − ρ(x)p+β n(p − 1) + β(cid:19)pZM p(cid:18) Rp(cid:16) f −ZM p − 1 n(p − 1) + β ZM − fp ρ(x)p+β ρ(x) , = p 1 p p (p+β) ρ(x) J′(ux, ρ(x))ρ(x) β p dVg J(ux, ρ(x)) ∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) ρ(x)β n(p−1)+β (∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) ρ(x)β J′(ux, ρ(x))ρ(x) J(ux, ρ(x)) dVg, dVg + p − 1 p ZM fp ρ(x)p+β dVg J(ux,ρ(x)) )fp J(ux,ρ(x)) )f )(cid:17) dVg which is equivalent to ZM fp ρ(x)p+β dVg =(cid:18) p n(p − 1) + β(cid:19)pZM Rp(cid:16) f − pZM p(p − 1) n(p − 1) + β ZM − ρ(x) , ∂ρf + ( n−1 ρ(x) + J ′(ux,ρ(x)) J(ux,ρ(x)) )fp ρ(x)β ρ(x) + J ′(ux,ρ(x)) n(p−1)+β (∂ρf + ( n−1 p dVg J(ux,ρ(x)) )f )(cid:17) dVg ρ(x)β J′(ux, ρ(x))ρ(x) J(ux, ρ(x)) fp ρ(x)p+β dVg. (4.4) Since β > −n(p − 1), Rp ≥ 0 and J′(u, ρ) ≥ 0, then the inequality (4.1) is an immediate conse- quence of (4.4). It remains to check the sharpness of (4.1). For 0 < δ < 1/2, define fδ(x) = ϕ(ρ(x))(1 − ϕ(δ−1ρ(x)))ρ(x)− n−p−β p , 13 where ϕ is cut-off function in (−1, 1). An easy computation shows that fδ(x)p ρ(x)p+δ dVg =ZSn−1Z 1 δ 2 ZM Hence Obviously, we have ϕ(ρ)p(1 − ϕ(δ−1ρ))pJ(u, ρ)ρ−1dρdu ≥ Sn−1 ln(2δ)−1. fδ(x)p ρ(x)p+δ dVg = ∞. lim δ→0 ∂ρf +(cid:18) n − 1 ρ(x) + J′(ux, ρ(x)) J(ux, ρ(x))(cid:19) f = ϕ′(ρ)ρ− n−p−β p − n(p − 1) + β 1 δ + ϕ(ρ)(1 − ϕ(δ−1ρ))ρ− n−β p . p ϕ′(δ−1ρ)ρ− n−p−β p + J′ J ϕ(ρ)(1 − ϕ(δ−1ρ))ρ− n−p−β p We can readily check that dVg = O(1), p p ρ(x)β ϕ′(ρ)ρ− n−p−β ZM δ ϕ′(δ−1ρ)ρ− n−p−β 1 ZM J ϕ(ρ)(1 − ϕ(δ−1ρ))ρ− n−p−β J ′ ρ(x)β p p p ZM ϕ(ρ)(1 − ϕ(δ−1ρ))ρ− n−β p p ρ(x)β ρ(x)β dVg = O(1), p dVg = O(1), dVg =ZM fδ(x)p ρ(x)p+δ dVg. dVg =(cid:18) n(p − 1) + β p (cid:19)p . and Therefore, ZM δ→0+ RM lim This finishes our proof. ∂ρf +(cid:16) n−1 ρ(x) + J ′(ux ,ρ(x)) J(ux ,ρ(x)) (cid:17)fp ρ(x)p+β fδ (x)p ρ(x)p+δ dVg RM If KM ≤ −b ≤ 0, then the identity (4.4) implies a quantitative version of (4.1) as follows ZM ρ(x) + J ′(ux,ρ(x)) ∂ρf + ( n−1 dVg ρ(x)β J(ux,ρ(x) )fp fp ≥(cid:18) n(p − 1) + β ρ(x)p+β dVg + 3b(n − 1)(p − 1)(cid:18) n(p − 1) + β (cid:19)pZM p p (cid:19)p−1ZM fp ρp+β−2(π2 + bρ2) dVg. (4.5) Replacing f by ∂ρf in Lemma 4.1 and (4.5), we obtain the following Rellich type inequality which connects first to second order derivatives. Theorem 4.2. Let (M, g) be an n−dimensional Cartan–Hadamard manifold. Suppose that n ≥ 2, p ∈ (1, n) and −n(p − 1) < β < n − p. There holds for any f ∈ C∞0 (M ) ∂ρfp ρ(x)p+β dVg. ∆g,ρfp ρ(x)β dVg ≥(cid:18) n(p − 1) + β (cid:19)pZM ZM (4.6) p 14 Furthermore, the constant ( n(p−1)+β )p is the best constant in (4.6). p If KM ≤ −b ≤ 0, then we have ZM ∆g,ρfp ρ(x)β dVg ≥(cid:18) n(p − 1) + β p (cid:19)pZM ∂ρfp ρ(x)p+β dVg + 3b(n − 1)(p − 1)(cid:18) n(p − 1) + β p (cid:19)p−1ZM ∂ρfp ρp+β−2(π2 + bρ2) dVg. (4.7) Combining (4.6), (4.7), (3.1) and (3.8), we get the following weighted Rellich inequalities on M . Theorem 4.3. Let M be an n−dimensional Cartan–Hadamard manifold. Suppose that n ≥ 3, p ∈ (1, n/2) and −n(p − 1) < β < n − 2p. There holds for any f ∈ C∞0 (M ) fp ∆g,ρfp ρ(x)β dVg ≥(cid:18) (n(p − 1) + β)(n − 2p − β) ρ(x)2p+β dVg. (cid:19)pZM ZM (4.8) p2 Furthermore, the constant ( (n(p−1)+β)(n−2p−β) )p is the best constant in (4.8). p2 If KM ≤ −b ≤ 0, then we have ZM ∆g,ρfp ρ(x)β dVg (cid:19)pZM ≥(cid:18) (n(p − 1) + β)(n − 2p − β) + 3b(n − 1)(p − 1)(cid:18) n(p − 1) + β + 3b(n − 1)(cid:18) n − 2p − β fp ρ(x)2p+β dVg (cid:19)p−1ZM (cid:19)p−1(cid:18) n(p − 1) + β p2 p p p ∂ρfp (cid:19)pZM ρp+β−2(π2 + bρ2) dVg fp ρ2p+β−2(π2 + bρ2) dVg. (4.9) Proof. Since −n(p − 1) < p + β < n − p, then by the weighted Hardy inequality (3.1), we have ZM fp ρ2p+β dVg ≤(cid:18) p n − 2p − β(cid:19)pZM ∂ρf2 ρp+β dVg. In the other hand, by (4.6) we get ZM ∂ρf2 ρp+β dVg ≤(cid:18) p n(p − 1) + β(cid:19)pZM ∆g,ρfp ρ(x)β dVg. Combining these two estimates, we obtain (4.8). To check the sharpness of (4.8), we use the approximation of ρ−(n−2p−β)/p as follows fδ(x) = ϕ(ρ(x))(1 − ϕ(δ−1ρ(x)))ρ(x)− n−2p−β p where ϕ is cut-off function in C∞0 ((−1, 1)) and 0 < δ < 1/2. Using the same argument as in the proof of Lemma 4.1 by making the straightforward (but tedious) compuations, we can show that ∆g,ρfδp ρ(x)β dVg fp ρ2p+β dVg =(cid:18) (n − 2p − β)(n(p − 1) + β) p2 (cid:19)p lim δ→0+ RM RM which implies the sharpness of (4.8). The proof of (4.9) is completely similar by iterating (3.8) and (4.7). We next consider the critical case β = n−2p. In this case, we obtain a critical Rellich inequality which generalizes the inequality (3.10) to order two. 15 Theorem 4.4. Let (M, g) be an n−dimensional Cartan–Hadamard manifold. Suppose that n ≥ 3, p ∈ (1, n). There holds for any f ∈ C∞0 (B1(P )) dVg ≤(cid:18) ∆g,ρfp ρ(x)n−2p dVg. f (x)p ρ(x)n(ln 1 ZB1(P ) ρ(x) )p (4.10) (p − 1)(n − 2)(cid:19)pZB1(P ) (n−2)(p−1) )p is the best constant in (4.10). Furthermore, the constant ( p p If KM ≤ −b ≤ 0, then we have ZB1(P ) ∆g,ρfp ρn−2p dVg dVg ρ(x) )p p (cid:19)pZB1(P ) f (x)p ≥(cid:18) (p − 1)(n − 2) ρ(x)n(ln 1 + 3b(n − 1)(p − 1)(n − 2)p−1ZB1(P ) p (cid:19)p−1ZB1(P ) + 3b(n − 1)(n − 2)p(cid:18) p − 1 ∂ρfp dVg ρn−p−2(π2 + bρ2) fp ρ )p−1(π2 + bρ2) ρn−2(ln 1 dVg. (4.11) Proof. The inequality (4.10) is consequence of (3.10) and (4.6) with β = n − 2p. Note that the condition −n(p − 1) < β < n − p holds true since n ≥ 3. To check the sharpness of (4.10), we use the following sequence of test functions fδ(x) =(cid:18)ln 1 ρ(x)(cid:19) p−1 p −δ ϕ(ρ(x)), where ϕ is cut-off function in (−1, 1). Making the computations as in the proof of Theorem 3.3, we obtain the desire result. The inequality (4.11) is followed from (3.17) and (4.7). Iterating the weighted Hardy and Rellich inequalities (both in the subcritical and critical cases), we obtain the following weighted Rellich inequality for higher order derivatives (both in the subcritical and critical cases respectively) on M . The detail proof is left to the readers. Let us denote c(n, 2l, β, p) = l−1 Yi=0 (n − 2p − β − 2ip)(n(p − 1) + β + 2ip)! p2 p for l ≥ 1, p ∈ (1, n/(2l)) and −n(p − 1) < β < n − 2lp. Theorem 4.5. Let M be an n−dimensional Cartan–Hadamard manifold and let k be a positive integer. Suppose that n ≥ 3, and p ∈ (1, n/k). Then for any function f ∈ C∞0 (M ) the following inequalities hold true. (i) If k = 2l, l ≥ 1 and n(1 − p) < β < n − 2lp, then we have fp ρ(x)2lp+β dVg ≤ c(n, 2l, β, p)ZM ZM and if KM ≤ −b ≤ 0 then ∆l g,ρfp ρ(x)β dVg, (4.12) ∆l g,ρfp ρ(x)β dVg ≥ ZM 1 fp c(n, 2l, β, p)ZM ρ(x)2lp+β dVg (n − 2lp − β)c(n, 2l, β, p)ZM 3b(n − 1)p + fp ρ2lp+β−2(π2 + bρ2) dVg. (4.13) 16 (ii) If k = 2l + 1, l ≥ 1 and n − n(p + 1) < β < n − (2l + 1)p then we have ∂ρ∆l pp fp g,ρfp (n − p − β)p c(n, 2l, p + β, p)ZM ρ(x)(2l+1)p+β dx ≤ ZM dVg. (4.14) ρ(x)β and if KM ≤ −b ≤ 0 then pp (n − p − β)p c(n, 2l, p + β, p)ZM ≥ZM ρ(2l+1)p+β dVg + fp ρ(x)β 3b(n − 1)p (n − (2l + 1)p − β)ZM ∂ρ∆l g,ρfp dVg fp ρ(2l+1)p+β−2(π2 + bρ2) dVg. (4.15) Furthermore, the inequalities (4.12) and (4.14) are sharp. For the critical case β = n− kp, we have the following critical Rellich inequalities on M which generalize Theorem 3.3 and Theorem 4.4 to higher order derivatives. Theorem 4.6. Let (M, g) be an n−dimensional Cartan–Hadamard manifold and let k be a positive integer. Suppose that n ≥ 3 and p ∈ (1, n/k). Then for any function f ∈ C∞0 (B1(P )) the following inequalities hold true. (i) If k = 2l, l ≥ 1 then we have p ZB1(P ) ∆l g,ρfp ρn−2lp dVg, (4.16) g,ρfp ∆l ρn−2lp dVg ZB1(P ) fp ρ )p−1(π2 + bρ2) ρn−2(ln 1 dVg. (4.17) p 1 n − 2i − 2! ZB1(P ) g,ρfp ∂ρ∆l ρn−(2l+1)p dVg, (4.18) (ii) If k = 2l + 1, l ≥ 1 then we have ZB1(P ) n − 2i − 2! here p′ = p/(p − 1), and if KM ≤ −b ≤ 0 then we have fp ρn(cid:16)ln 1 (l − 1)! Yi=0 l−1 1 p 1 l−1 p′ 21−l (l − 1)! ≥ZB1(P ) ρ(cid:17)p dVg ≤ p′ 21−l n − 2i − 2! ZB1(P ) Yi=0 fp 3b(n − 1)p ρ(cid:17)p dVg + p − 1 ρn(cid:16)ln 1 ρ(cid:17)p dx ≤ p′ 1 and if KM ≤ −b ≤ 0 then we have n − 2i − 2! ZB1(P ) fp ρ(cid:17)p dx + ρn(cid:16)ln 1 p′ 1 Yi=0 ≥ZB1(P ) fp ρn(cid:16)ln 1 ZB1(P ) 3b(n − 1)p Yi=0 p − 1 l−1 l−1 2ll! 2ll! 1 p g,ρfp ∂ρ∆l ρn−(2l+1)p dVg ZB1(P ) fp ρ )p−1(π2 + bρ2) ρn−2(ln 1 dVg. (4.19) Furthermore, the inequalities (4.16) and (4.18) are sharp. We emphasize here that in the Euclidean space M = Rn, Theorems 4.2, 4.3, 4.4, 4.5 and 4.6 was recently proved by the author [39] (The same inequalities on radial functions was previously proved by Adimurthi et al. [1] and by Adimurthi and Santra [2]). More precisely, in [39], the author also proved the generalizations of these inequalities on Rn to more general class of homogeneous groups equipped with any homogeneous quasi-norm with the same best constants. 17 5 Hardy and Rellich inequalities in hyperbolic spaces We conclude this section by giving some concrete examples on the n−dimensional hyperbolic spaces Hn. We will use the Poincar´e conformal disc model for hyperbolic spaces Hn, i.e., the underlying space which we consider is the unit ball Bn = {x = (x1, . . . , xn) ∈ Rn : x =qx2 1 + ··· + x2 n < 1} equipped with metric The volume element on Bn is given by dV = (cid:16) 2 operator is given by g(x) =(cid:18) dx. 2 1 − x2(cid:19)n 1−x2(cid:17)n dx and the associated Laplace–Beltrami ∆g = (1 − x2)2 4 n Xi=1 ∂2 ∂x2 i + 2(n − 2) n Xi=1 xi 1 − x2 ∂ ∂xi! and the corresponding gradient is ∇g =(cid:18) 1 − x2 2 (cid:19)2(cid:18) ∂ ∂x1 , . . . , ∂ ∂xn(cid:19) . The geodesic distance from x to 0 is ρ(x) = ln 1+x 1−x . Finally, recall that KBn ≡ −1. Our main results in this section give us several quantitative Hardy type inequalities on the hyperbolic spaces Bn as follows. ZBn ∇gfp Theorem 5.1. Suppose n ≥ 2, p ∈ (1, n) and β < n − p. Then there exists c > 0 such that the following inequality holds for any f ∈ C∞0 (Bn) dVg ≥(cid:18) n − p − β (cid:19)p−1ZBn Let B1(0) denote the geodesic unit ball with center at 0 in Bn. There exists a constant c > 0 such that the following inequality holds for any p > 1 and f ∈ C∞0 (B1(0)) fp ρp+β−2 dx. fp ρp+β dVg + 3c(n − 1)(cid:18) n − p − β (cid:19)pZBn (5.1) ρβ p p g ZB1(0) ∇gfp ρn−p dVg ≥(cid:18) p − 1 p (cid:19)pZB1(0) + 3C(n − 1)(cid:18) p − 1 Proof. From the formula for ρ(x), we have x = eρ−1 4eρ (cid:19)n =(cid:18) (1 + eρ)2 1 − x2(cid:19)n (cid:18) 1 dVg ρ )p fp ρn(ln 1 p (cid:19)p−1ZB1(0) eρ+1 . Hence ≥ C(π2 + ρ2) fp ρn−2(ln 1 ρ )p−1 dx. (5.2) for some C > 0. Thus Theorem 5.1 follows from (3.9), (3.18) and the previous inequality. Note that (5.1) extends a result of Kombe and Ozaydin (see [29, Theorem 3.1]) in the case p = 2 to any p ∈ (1, n). 18 We next consider the Rellich inequalities. Denote by ∂r = x x · ∇ the radial derivative in Rn, and denote by ∆r = ∂2 r + n − 1 r ∂r the radial Laplace in Rn. It was proved by Machihara, Ozawa and Wadade [35] that ZRn ∆ru2dx ≤ZRn ∆u2dx for any u ∈ C∞0 (Rn). Our next results shows that such a result also holds true on hyperbolic space Hn (even in the weighted form). Theorem 5.2. Let n ≥ 3 and −2 < β ≤ n − 4. It holds ZBn ∆g,ρf2ρ−βdVg ≤ZBn ∆gf2ρ−βdVg, f ∈ C∞0 (Bn). (5.3) Furthermore, equality holds in (5.3) if and only if f is radial function. Proof. Denote r = x = eρ−1 eρ+1 , then ∆g = (1 − r2)2 4 ∆ + (n − 2) 1 − r2 2 r∂r, ∆g,ρ = (1 − r2)2 4 ∆r + (n − 2) 1 − r2 2 r∂r. For any function f ∈ C∞0 (Bn), we decompose it into spherical harmonic as f (x) = ∞ Xk=0 fk(r)φk(σ), σ ∈ Sn−1, x = rσ, (5.4) where φk is orthonormal eigenfunction of the Laplace–Beltrami operator on the sphere Sn−1 with respect to eigenvalue ck = k(n + k − 2) with k = 0, 1, 2, . . .. The function fk belongs to C∞0 (Bn) and satifies fk(r) = O(rk), f′k(r) = O(rk−1) as r ↓ 0. In particular, we have φ0 ≡ 1 and f0(r) = 1 nωn RSn−1 f (rσ)dσ. From the decomposition of f , we have ∆gfk(r) φk(σ), ∞ ∆g,ρf (x) = Xk=0 Xk=0(cid:18)∆gfk(r) − ck ∞ and ∆gf (x) = (1 − r2)2 4 fk(r) r2 (cid:19) φk(σ). Thus, to prove (5.3), it's enough to verify that ckZBn f 2 k 2r (cid:19)4 ρβ (cid:18) 1 − r2 k − 2∇gfk2 dVg − 2ZBn Note that 2fk∆gfk = ∆gf 2 to (fk∆gfk) 1 2r (cid:19)2 ρβ (cid:18) 1 − r2 dVg ≥ 0, k ≥ 1. (5.5) g. Hence, by using integration by parts, (5.5) is equivalent ckZBn f 2 k 2r (cid:19)4 ρβ (cid:18) 1 − r2 dVg + 2ZBn ∇gfk2 g 1 2r (cid:19)2 ρβ (cid:18) 1 − r2 −ZBn 19 dVg f 2 k ∆g 1 ρβ (cid:18) 1 − r2 2r (cid:19)2! dVg ≥ 0, (5.6) for any k ≥ 1. Remark that r = eρ−1 eρ+1 and hence 1 2r (cid:19)2 ρβ (cid:18) 1 − r2 = 1 ρβ sinh2 ρ =: k(ρ). From (2.4), we have ∆gk(ρ) = k′′(ρ) + (n − 1) cosh ρ sinh ρ k′(ρ) = −k(ρ)(cid:18)2 + 2(n − 4) cosh2 ρ sinh2 ρ − β(β + 1) 1 ρ2 + β(n − 5) cosh ρ ρ sinh ρ(cid:19) . (5.7) We next show that 1 2r (cid:19)4 ρβ (cid:18) 1 − r2 for any radial function u ∈ C∞0 (Bn). Define the function F on [0,∞) by 2r (cid:19)2 ρβ (cid:18) 1 − r2 ZBn ∇gu2 (n − β − 4)2 dVg ≥ ZBn u2 4 g dVg, (5.8) F (ρ) = u(r), r = eρ − 1 eρ + 1 . Then, (5.8) is equivalent to Z ∞ 0 (F ′(ρ))2ρn−β−3J1(ρ) n−3 n−1 dρ ≥ (n − β − 4)2 4 Z ∞ 0 F (ρ)2ρn−β−5J1(ρ) n−5 n−1 dρ, (5.9) Recall that J1(ρ) = ( sinh ρ ρ )n−1. Indeed, using integration by parts and β < n − 4 we get Z ∞ 0 F (ρ)2ρn−β−5J1(ρ) n−5 n−1 dρ = − − 2 n − β − 4Z ∞ n − β − 4Z ∞ 1 0 0 F (ρ)F ′(ρ)ρn−β−4J1(ρ) n−5 n−1 dρ F (ρ)2ρn−β−4J′1(ρ)J1(ρ)− 4 n−1 dρ. Applying Holder inequality and using J′1 ≥ 0, J1 ≥ 1, we get (n − 4 − β)2 Z ∞ (n − 4 − β)2 Z ∞ F (ρ)2ρn−β−5J1(ρ) n−1 dρ ≤ Z ∞ ≤ n−5 4 4 0 0 0 (F ′(ρ))2ρn−β−3J1(ρ) (F ′(ρ))2ρn−β−3J1(ρ) n−5 n−1 dρ n−3 n−1 dρ which implies (5.9). Consequently, the left hand side of (5.6) is at least (cid:18)ck + 2 (n − β − 4)2 +ZBn f 2 k dVg 2r (cid:19)4 ρβ (cid:18) 1 − r2 (cid:19)ZBn 2r (cid:19)4 ρβ (cid:18) 1 − r2 (cid:16)2 sinh2 ρ + 2(n − 4) cosh2 ρ− f 2 k Hence, to prove (5.6), it is enough to show − β(β + 1) sinh2 ρ ρ2 + β(n − 5) sinh(2ρ) 2ρ (cid:17)dVg. ck + (n − β − 4)2 2 + 2 sinh2 ρ + 2(n − 4) cosh2 ρ − β(β + 1) sinh2 ρ ρ2 + β(n − 5) sinh(2ρ) 2ρ ≥ 0, (5.10) 20 for ρ > 0. It is suffice to check (5.10) for k = 1. Expanding the exponent function in series form, the left hand side of (5.10) is equal to (n − 2 + k)2 + k2 − (β + 2)2 2 + ∞ Xl=1(cid:18)(n − 3) − β(β + 1) (l + 1)(2l + 1) + 2l + 1 (cid:19) (2ρ)2l β(n − 5) (2l)! . Since k ≥ 1 and −2 < β ≤ n − 4, we can easily check that (n − 2 + k)2 + k2 − (β + 2)2 2 > 0, and (n − 3) − β(β + 1) (l + 1)(2l + 1) + β(n − 5) 2l + 1 ≥ 0, (5.11) for any l ≥ 1. This finishes the proof of (5.3). Suppose that equality holds true in (5.3) for some function f . Expanding f in spherical harmonic expression as in (5.4). By (5.11), we must have ck = 0 for any k ≥ 1. This shows that f is radial function. By Theorem 5.2, we see that in the hyperbolic space Hn, the Rellich inequality (4.8) with p = 2 is stronger than the inequality of Kombe and Ozaydin [29]: suppose −2 < β < n − 4 f ∈ C∞0 (Bn). (n + β)2(n − 4 − β)2 dVg, ∆gf2 ρβ 16 ZBn f2 ρβ+4 dVg ≤ZBn (5.12) Similarly, (4.9) implies an improvements of (5.12) ∆gf2 ρβ dVg ≥ ZBn (n + β)2(n − 4 − β)2 16 (n − 4 − β)(n + β)2 8 ZBn f2 ρβ+4 dVg + 3(n − 1) f2 ρβ+2(π2 + ρ2) + 3(n − 1) ZBn n + β 2 ZBn ∂ρf2 ρβ(π2 + ρ2) dVg dVg. (5.13) It is easy to prove that ZBn ∂ρf2 ρβ(π2 + ρ2) dVg ≥ (n − β − 4)2 4 f2 ρβ+2(π2 + ρ2) dVg. ZBn (5.14) Combining (5.13) and (5.14) yields ZBn ∆gf2 ρβ dVg ≥ (n + β)2(n − 4 − β)2 16 f2 ρβ+4 dVg ZBn (n − 1)(n − 2)(n + β)(n − 4 − β) + 3 4 f2 ρβ+2(π2 + ρ2) ZBn dVg. (5.15) Using the simple inequality in the proof of Theorem 5.1, we prove the following improved Rellich inequality in Hn. Theorem 5.3. Suppose that n ≥ 4 and −2 < β < n− 4. Then there exists a constant C > 0 such that ZBn ∆gf2 ρβ dVg ≥ (n + β)2(n − 4 − β)2 16 f2 ρβ+4 dVg ZBn (n − 1)(n − 2)(n + β)(n − 4 − β) 4 f2 ρβ+2 dx, ZBn (5.16) + 3C for any f ∈ C∞0 (Bn). By iterating the inequalities (5.12), (5.16) and (5.1), we obtain the following improved Rellich type inequalities of higher order derivatives in Hn. 21 Theorem 5.4. Suppose n ≥ 3 and k ∈ (1, n/2) be an integer, and −2 < β < n − 2k. Then there exists a constant C > 0 such that the following inequalities hold for any f ∈ C∞0 (Bn) (i) If k = 2l, l ≥ 1 then we have ∆l gf2 ρ(x)β dVg ≥ f2 1 ZBn c(n, 2l, β, 2)ZBn ρ(x)β+4l dVg 6C(n − 1) (n − 4l − β)c(n, 2l, β, 2)ZBn + f2 ρ4l+β−2 dx. (5.17) (ii) If k = 2l + 1, l ≥ 1 then we have 4 (n − 2 − β)2 c(n, 2l, 2 + β, 2)ZBn ≥ZBn ρβ+2(2l+1) dVg + f2 dVg gf2 g ∇g∆l ρ(x)β 6C(n − 1) (n − 2(2l + 1) − β)ZBn f2 ρβ+4l dx. (5.18) By the same way, we can obtain the critical Rellich type inequalities in Hn for ∆l g and ∇g∆l g and their improvements. The details are left for interest readers. Acknowledgments The author would like to thank Professor Alexandru Krist´aly for drawing our attentions to his works [15, 31, 32]. This work was supported by the CIMI's postdoctoral research fellowship. References [1] Adimurthi, M. Grossi, and S. Santra, Optimal Hardy-Rellich inequalities, maximum prin- ciple and related eigenvalue problem, J. Funct. Anal., 240 (2006) 36–83. [2] Adimurthi, and S. Santra, Generalized Hardy-Rellich inequalities in critical dimension and its applications, Commun. Comtemp. Math., 11 (2009) 367–394. [3] A. A. Balinsky, W. D. Evans, and R. T. Levis, The analysis and geometry of Hardy's inequality, Universitext, Springer, Cham, 2015. [4] H. Br´ezis, and M. Marcus, Hardy's inequalities revisited, Ann. Scuola Norm. Sup. Pisa Cl. Sci., 25 (1997) 217–237. [5] H. Br´ezis, and J. L. V´azquez, Blowup solutions of some nonlinear elliptic problems, Rev. Mat. Univ. Complut. Madrid, 10 (1997) 443–469. [6] G. Carron, In´egalit´es de Hardy sur les vari´et´es riemmanniennes non-compactes, J. Math. Pures Appl., 76 (1997) 443–469. [7] P. Ciatti, M. G. Cowling, and F. Ricci, Hardy and uncertainty inequalities on stratified Lie groups, Adv. Math., 277 (2015) 365–387. [8] L. D'Ambrosio, Hardy-type inequalities related to degenerate elliptic differential operators, Ann. Sc. Norm. Super. Pisa Cl. Sci., 4 (2005) 449–475. [9] L. D'Ambrosio, and S. Dipierro, Hardy inequalities on Riemannian manifolds and applications, Ann. Inst. H. Poincar´e Anal. Non Lin´eaire, 31 (2014) 449–475. [10] D. Danielli, N. Garofalo, and N. C. Phuc, Hardy–Sobolev type inequalities with sharp con- stants in Carnot–Carath´eodory spaces, Potential Anal., 34 (2011) 223–242. 22 [11] E. B. Davies, A review of Hardy inequalities, In The Maz'ya anniversary collection, Vol. 2 (Rostock, 1998), volume 110 of Oper. Theory Adv. Appl., pp. 55–67, Birkhauser, Basel, 1999. [12] E. B. Davies, and A. M. Hinz, Explicit constants for Rellich inequalities in Lp(Ω), Math. Z., 227 (1998) 511–523. [13] Y. Di, L. Jiang, S. Shen, and Y. Jin, A note on a class of Hardy–Rellich type inequalities, J. Inequal. Appl., 84 (2013) 1–6. [14] D. E. Edmunds, and H. Triebel, Sharp Sobolev embedding and related Hardy inequalities: the critical case, Math. Nachr., 207 (1999) 79–92. [15] C. Farkas, A. Krist´aly, and C. Varga, Singular Poisson equations on Finsler–Hadamard manifolds, Calc. Var. Differential Equations, 54 (2015) 1219–1241. [16] S. Filippas, and A. Tertikas, Optimizing improved Hardy inequalities, J. Funct. Anal., 192 (2002) 186–233. [17] S. Filippas, L. Moschini, and A. Tertikas, Sharp trace Hardy–Sobolev–Maz'ya inequalities and the fractional Laplacian, Arch. Ration. Mech. Anal., 208 (2013) 109–161. [18] R.L. Frank, E.H. Lieb, and R. Seiringer, Hardy–Lieb–Thirring inequalities for fractional Schrdinger operators, J. Amer. Math. Soc., 21 (2008) 925–950. [19] R.L. Frank, and R. Seiringer, Non–linear ground state representations and sharp Hardy inequalities, J. Funct. Anal., 255 (2008) 3407–3430. [20] S. Gallot, D. Hulin and J. Lafontaine, Riemannian Geometry, 3rd edn. (Springer-Verlag, Berlin, 2004). [21] N. Garofalo, and E. Lanconelli, Frequency functions on the Heisenberg group, the uncer- tainty principle and unique continuation, Ann. Inst. Fourier (Grenoble), 40 (1990) 313–356. [22] J. Goldstein, and I. Kombe, The Hardy inequality and nonlinear parabolic equations on Carnot groups, Nonlinear Anal., 69 (2008) 4643–4653. [23] S. Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces, Graduate Studies in Mathematics, Vol. 34 (American Mathematical Society, Providence, RI, 2001). [24] I. W. Herbst, Spectral theory of the operator (p2 + m2)1/2 − Ze2/r, Comm. Math. Phys., 53 (1977) 285–294. [25] N. Ioku, and M. Ishiwata, A scale invariant form of a critical Hardy inequality, Int. Math. Res. Not., 18 (2015) 8830–8846. [26] N. Ioku, M. Ishiwata and T. Ozawa, Sharp remainder of a critical Hardy inequality, J. Inequal. Appl., 5 (2017) 7pp. [27] N. Ioku, M. Ishiwata and T. Ozawa, Hardy type inequalities in Lp with sharp remainders, Arch. Math. (Basel), 106 (2016) 65–71. [28] Y. Jin, and S. Shen, Weighted Hardy and Rellich inequality on Carnot groups, Arch. Math. (Basel), 96 (2011) 263–271. [29] I. Kombe, and M. Ozaydin, Improved Hardy and Rellich inequalities on Riemannian manifolds, Trans. Amer. Math. Soc., 361 (2009) 6191–6203. [30] I. Kombe, and M. Ozaydin, Hardy-Poincar, Rellich and uncertainty principle inequalities on Riemannian manifolds, Trans. Amer. Math. Soc., 365 (2013) 5035–5050. 23 [31] A. Krist´aly, and D. Repovs, Quantitative Rellich inequalities on Finsler–Hadamard manifolds, Commun. Contemp. Math., 18 (2016) 1650020 (17 pages). [32] A. Krist´aly, Sharp uncertainty principles on Riemannian manifolds: curvature, to appear in Journal de Math´ematiques Pures et Appliqu´ees. the influence of [33] B. Lian, Some sharp Rellich type inequalities on nilpotent groups and application, Acta. Math. Sci. Ser. B Engl. Ed., 33 (2013) 59–74. [34] S. Machihara, T. Ozawa, and H. Wadade, Remarks on the Hardy type inequalities with remainder terms in the framework of equalities, to appear in Adv. Studies Pure Math., arXiv:1611.03580v1. [35] S. Machihara, T. Ozawa, and H. Wadade, Remarks on the Rellich inequality, Math. Z., 286 (2017) 1367–1373. [36] E. Mitidieri, A Rellich type identity and applications, Comm. Partial Differential Equations, 18 (1993) 125–151. [37] E. Mitidieri, A simple approach to Hardy's inequalities, Math. Notes, 67 (2000) 479–486. [38] V. H. Nguyen, Some trace Hardy type inequalities and trace Hardy–Sobolev–Maz'ya type inequalities, J. Funct. Anal., 270 (2016), no. 11, 4117–4151. [39] V. H. Nguyen, The sharp higher order Hardy–Rellich inequalities on stratified Lie groups, preprint. [40] P. Niu, H. Zhang, Y. Wang, Hardy type and Rellich type inequalities on the Heisenberg group, Proc. Amer. Math. Soc., 129 (2001) 3623–3630. [41] M. Ruzhansky, and D. Suragan, On horizontal Hardy, Rellich, Caffarelli-Kohn-Nirenberg and p-sub-Laplacian inequalities on stratified groups, J. Differential Equations, 262 (2017) 1799–1821. [42] M. Ruzhansky, and D. Suragan, Hardy and Rellich inequalities, identities and sharp re- mainders on homogeneous groups, Adv. Math., 317 (2017) 799–822. [43] M. Ruzhansky, and D. Suragan, Critical Hardy inequalities, arXiv:1602:04809. [44] M. Ruzhansky, and D. Suragan, Anisotropic L2−weighted Hardy and L2−Caffarelli–Kohn– Nirenberg inequalities, Commun. Comtemp. Math., 19 (2017) 1750014 (12 pages). [45] M. Sano, and F. Takahashi, Scale invariance structures of the critical and the subcritical Hardy inequalities and their improvements, to appear in Calc. Var. Partial Differential Equations. [46] A. Tertikas, and N. B. Zographopoulos, Best constants in the Hardy–Rellich inequalities and related improvements, Adv. Math., 209 (2007) 407–459. [47] D. Yafaev, Sharp constants in the Hardy–Rellich inequalities, J. Funct. Anal., 168 (1999) 121–144. [48] Q. Yang, D. Su, and Y. Kong, Hardy inequalities on Riemannian manifolds with negative curvature, Commun. Contemp. Math., 16 (2014) 24 pages. 24
1303.4569	2	1303	2013-03-22T23:33:00	Local triple derivations on C-algebras and JB-triples	[ "math.FA", "math.OA" ]	In a first result we prove that every continuous local triple derivation on a JB$^$-triple is a triple derivation. We also give an automatic continuity result, that is, we show that local triple derivations on a JB$^$-triple are continuous even if not assumed a priori to be so. In particular every local triple derivation on a C$^$-algebra is a triple derivation. We also explore the connections between (bounded local) triple derivations and generalised (Jordan) derivations on a C$^$-algebra.	math.FA	math	LOCAL TRIPLE DERIVATIONS ON C∗-ALGEBRAS AND JB∗-TRIPLES MARIA BURGOS, FRANCISCO J. FERN ´ANDEZ-POLO, AND ANTONIO M. PERALTA Abstract. In a first result we prove that every continuous local triple deriva- tion on a JB∗-triple is a triple derivation. We also give an automatic continuity result, that is, we show that local triple derivations on a JB∗-triple are con- tinuous even if not assumed a priori to be so. In particular every local triple derivation on a C∗-algebra is a triple derivation. We also explore the con- nections between (bounded local) triple derivations and generalised (Jordan) derivations on a C∗-algebra. 1. Introduction An (associative) derivation from a C∗-algebra A into a Banach A-bimodule X is a linear mapping D : A → X satisfying D(ab) = D(a)b + aD(b), for every a, b in A. A linear mapping T : A → X is a local derivation if for each a in A there is a derivation Da from A into X with Da(a) = T (a). Local derivations were introduced by R.V. Kadison [27] who proved that every continuous local derivation from a von Neumann algebra M (i.e. a C∗-algebra which is also a dual Banach space) into a dual Banach M -bimodule is a derivation. Kadison's theorem motivated a flourishing line of research which culminates in 2001 with a definite contribution by B.E. Johnson [26], who shows that every bounded local derivation from a C∗-algebra A into a Banach A-bimodule is a derivation. In the just quoted paper, Johnson showed that the continuity hypothesis is, in fact, superfluous by proving that every local derivation from a C∗-algebra A into a Banach A-bimodule is continuous. Besides of holomorphic motivations coming from the classification of bounded symmetric domains in arbitrary complex Banach spaces (cf. [28]), there are addi- tional reasons, from the point of view of Functional Analysis, to study a strictly wider class of complex Banach spaces which comprises the categories of C∗-algebras, JB∗-algebras and ternary rings of operators, and is known as the category of JB∗-triples. For example, contrary to what happens for C∗-algebras, the cat- egory of JB∗-triples is stable under contractive projections (cf. [29] and [20]). A JB∗-triple is a complex Banach space E equipped with a triple prod- uct {., ., .} : E × E × E → E which is conjugate linear in the middle variable and [36], Date: February, 2013. 2000 Mathematics Subject Classification. Primary 47B47, 46L57; Secondary 17C65, 47C15, 46L05, 46L08. Key words and phrases. triple derivation; (continuous) local triple derivation; generalised derivation; generalised Jordan derivation; C∗-algebra; real C∗-algebra; JB∗-triple; real JB∗-triple; range tripotent; compact tripotent; non-commutative Urysohn's lemma. Authors partially supported by the Spanish Ministry of Economy and Competitiveness, D.G.I. project no. MTM2011-23843, and Junta de Andaluc´ıa grants FQM0199 and FQM3737. 2 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA symmetric and linear in the outer variables satisfying certain algebraic-analytic ax- ioms (see Subsection 1.1 for more details). Every C∗-algebra is a JB∗-triple when equipped with the triple product given by {x, y, z} = 1 2 (xy∗z + zy∗x). A triple derivation on a JB∗-triple E is a linear mapping δ : E → E satisfying that δ {a, b, c} = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} , for every a, b, c ∈ E. A local triple derivation on E is a linear map T : E → E such that for each a in E there exists a triple derivation δa on E satisfying T (a) = δa(a). Inspired by the results proved by R.V. Kadison and B.E. Johnson it is natural to consider the following problems: Problem 1.1. Is every bounded local triple derivation on a JB∗-triple E a triple derivation? Problem 1.2. Is every local triple derivation on a JB∗-triple E a triple derivation? Problem 1.3. Is every local triple derivation on a JB∗-triple E continuous? We note that the above problems also make sense in the setting of C∗-algebras and are interesting questions in its own right. Problem 1.4. Is every (bounded) local triple derivation on a C∗-algebra B a triple derivation? M. Mackey gave in [31, Theorem 5.11] a positive answer to the above Problem (1.1) under the additional hypothesis of E being a JBW∗-triple (i.e. a JB∗-triple which is also a dual Banach space). Problems (1.1), (1.2) and (1.3) already appear in [31] and in [14]. Mackey's theorem can be considered an appropriate version of Kadison's theorem in the setting of JB∗-triples. Problem 1.4 was solved by J. Garc´es and the authors of this note for bounded local triple derivations on unital C∗-algebras in [14]. The above problems have remained open in their full generality until now. In this paper we completely solve Problems (1.1) and (1.3) (and consequently Problems (1.2) and (1.4)). In Section 2 we prove that every continuous local triple derivation on a JB∗-triple is a triple derivation (see Theorem 2.4). Our strategy consists in studying the behavior of the bitranspose, T ∗∗, on compact and range tripotents in E∗∗. It follows from our first main result that Problems (1.3) and (1.2) are equivalent. In a subsequent result, we explore an automatic continuity result for local triple derivations. T. Barton and Y. Friedman proved in [8] that every triple derivation on a JB∗-triple is automatically continuous, we shall adapt their arguments to show that the same statement remains valid for local triple derivations (see Theorem 2.8). In Section 3, we particularize our results to the setting of C∗-algebras and explore the connections between (local) triple derivations and generalised derivations on a C∗-algebra. Among the consequences, we establish appropriate generalizations, to general C∗-algebras, of results due to V. Shu'lman [35] and J. Li and Z. Pan [30] in the context of unital C∗-algebras. LOCAL TRIPLE DERIVATIONS 3 1.1. Preliminaries. A JB∗-triple is a complex Banach space E together with a continuous triple product {., ., .} : E × E × E → E, which is conjugate linear in the middle variable and symmetric bilinear in the outer variables satisfying the following axioms: (a) (Jordan Identity) (1) {a, b, {x, y, z}} = {{a, b, x} , y, z} − {x, {b, a, y} , z} + {x, y, {a, b, z}} , for all a, b, x, y, z in E; (b) If L(a, b) denotes the operator on E given by L(a, b)x = {a, b, x} , the mapping L(a, a) is an hermitian operator with non-negative spectrum; (c) k {a, a, a} k = kak3, for every a ∈ E. Given a, b ∈ E, the symbol Q(a, b) will denote the conjugate linear operator defined by Q(a, b)(x) = {a, x, b}. We shall write Q(a) instead of Q(a, a). Every C∗-algebra is a JB∗-triple via the triple product given by (2) and every JB∗-algebra is a JB∗-triple under the triple product 2 {x, y, z} = xy∗z + zy∗x, (3) {x, y, z} = (x ◦ y∗) ◦ z + (z ◦ y∗) ◦ x − (x ◦ z) ◦ y∗. Every element e in a JB∗-triple E satisfying {e, e, e} = e is called tripotent. When a C∗-algebra, A, is regarded as a JB∗-triple, the set of tripotents of A is precisely the set of all partial isometries in A. Associated with each tripotent e in a JB∗-triple E, there is a decomposition of E (called Peirce decomposition) in the form: E = E0(e) ⊕ E1(e) ⊕ E2(e), where Ek(e) = {x ∈ E : L(e, e)x = k 2 x} for k = 0, 1, 2. The Peirce rules are that {Ei(e), Ej(e), Ek(e)} ⊆ Ei−j+k(e) if i − j + k ∈ {0, 1, 2}, and {Ei(e), Ej(e), Ek(e)} = {0} otherwise. Moreover, {E2(e), E0(e), E} = {E0(e), E2(e), E} = {0}. The Peirce space E2(e) is a unital JB∗-algebra with unit e, product x ◦e y := {x, e, y} and involution x∗e := {e, x, e}, respectively. The corresponding Peirce projections, Pi(e) : E → Ei(e), (i = 0, 1, 2) are given by P2(e) = Q(e)2, P1(e) = 2L(e, e) − 2Q(e)2, and P0(e) = Id − 2L(e, e) + Q(e)2, where Id is the identity map on E. A JBW∗-triple is a JB∗-triple which is also a dual Banach space (with a unique isometric predual [9]). Since the triple product of every JBW∗-triple is separately weak∗ continuous (cf. [9]), it can be easily checked that Peirce projections associated with a tripotent e in a JBW∗-triple are weak∗ continuous. The second dual, E∗∗, of a JB∗-triple E is a JBW∗-triple [17]. Following standard notation, for each element a in a JB∗-triple E we denote a[1] = a and a[2n+1] := (cid:8)a, a[2n−1], a(cid:9) (∀n ∈ N). It follows from Jordan identity that JB∗-triples are power associative, that is, (cid:8)a[k], a[l], a[m](cid:9) = a[k+l+m]. The JB∗-subtriple of E generated by the element a will be denoted by the symbol Ea, and coincides with the closed linear span of the elements a[2n−1] with n ∈ N. The local Gelfand theory for JB∗-triples assures that Ea is JB∗-triple isomorphic (and hence isometric) to C0(L) for some locally compact Hausdorff space L ⊆ (0, kak], 4 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA such that L ∪ {0} is compact and kak ∈ L. It is further known that there exists a triple isomorphism Ψ from Ea onto C0(L), satisfying Ψ(a)(t) = t (t ∈ L) (compare [28, Lemma 1.14]). This result provides us with an spectral resolution for every element in a JB∗-triple E. The local Gelfand theory for JB∗-triples is a powerful tool; among its conse- quences, it follows that for an element a in a JB∗-triple E and each natural n, there exists (a unique) element a[1/(2n−1)] in Ea satisfying (a[1/(2n−1)])[2n−1] = a. When a is a norm one element, the sequence (a[1/(2n−1)]) converges in the weak∗ topology of E∗∗ to a tripotent denoted by r(a) and called the range tripotent of a. The tripotent r(a) is the smallest tripotent e in E∗∗ satisfying that a is positive in 2 (e). It is also known that the sequence (a[2n−1]) converges the JBW∗-algebra E∗∗ in the weak∗ topology of E∗∗ to a tripotent (called the support tripotent of a) s(a) in E∗∗, which satisfies s(a) ≤ a ≤ r(a) in A∗∗ 2 (r(a)) (compare [18, Lemma 3.3]; the reader should be noted that in [19], r(a) is called the support tripotent of a). Two elements a, b in a JB∗-triple E are said to be orthogonal (written a ⊥ b) if L(a, b) = 0. It is known that a ⊥ b if, and only if, one of the following statements holds: b ⊥ a; {a, a, b} = 0; a ⊥ r(b); r(a) ⊥ r(b); E∗∗ 2 (r(a)) ⊥ E∗∗ 2 (r(b)); r(a) ∈ E∗∗ 0 (r(b)); a ∈ E∗∗ 0 (r(b)); b ∈ E∗∗ 0 (r(a)); Ea ⊥ Eb, (see, for example, [13, Lemma 1]). The natural partial order in the set of tripotents of a JB∗-triple E is defined as follows: given two tripotents e and u in E we say that u ≤ e if e − u is a tripotent in E and e − u ⊥ u. Range and support tripotent are examples of open and compact tripotents, re- spectively. In attempt to generalize the studies of C.A. Akemann, L.G. Brown, and G.K. Pedersen on open and compact projections in the bidual of a C∗-algebra ([1, 2, 3, 12, 4]), C.M. Edwards and G.T. Ruttimann introduce in [19] the notions of open and compact tripotents in the bidual of a JB∗-triple. We recall that a tripotent u in the bidual of a JB∗-triple E is said to be open when E∗∗ 2 (u) ∩ E is 2 (u). A tripotent e in E∗∗ is said to be compact-Gδ (relative to weak∗ dense in E∗∗ E) if there exists a norm one element a in E such that e coincides with s(a), the support tripotent of a. A tripotent e in E∗∗ is said to be compact (relative to E) if there exists a decreasing net (eλ) of tripotents in E∗∗ which are compact-Gδ with infimum e, or if e is zero. Closed and bounded tripotents in E∗∗ were introduced and studied by the second and third authors of this note in [21] and [22]. A tripo- tent e in E∗∗ such that E∗∗ 0 (e) is called closed relative to E. When there exists a norm one element a in E such that a = e + P0(e)(a), the [21]). tripotent e is called bounded (relative to E) and we shall write e ≤T a (c.f. The relation ≤T is consistent with the natural partial order on the set of tripotents, that is, for any two tripotents e and u we have e ≤ u if and only if e ≤T u. An useful result established in [21, Theorem 2.6] (see also [23, Theorem 3.2]) asserts that a tripotent e in E∗∗ is compact if, and only if, e is closed and bounded. 0 (e) ∩ E is weak∗ dense in E∗∗ Given a JBW∗-triple W, a norm one functional ϕ in W∗ and a norm one element z in W with ϕ(z) = 1, Proposition 1.2 in [8] assures that the mapping (x, y) 7→ ϕ {x, y, z} LOCAL TRIPLE DERIVATIONS 5 is a positive sesquilinear form on W. Furthermore, for every norm one element w in W satisfying ϕ(w) = 1, we have ϕ {x, y, z} = ϕ {x, y, w} , for all x, y ∈ W . 2 , is a prehilbertian seminorm on Thus, the mapping x 7→ kxkϕ := (ϕ {x, x, z}) W . The strong-topology is the topology on W generated by the family {k · kϕ : ϕ ∈ W∗, kϕk = 1} (c.f. [8]). For later purposes, we recall that the triple product of a JBW∗-triple is jointly strong∗-continuous on bounded sets (see [33, Theorem 9]). Since the strong-topology of a JBW∗-triple W is compatible with the duality (W, W∗) (cf. [33, Corollary 9]), it follows from the bipolar theorem that for each convex C ⊆ W we have 1 σ(W,W∗) C = C S ∗(W,W∗) . We shall also make use of the following fact due to L. Bunce (see [11]): Let F be a JBW∗-subtriple (i.e. a weak∗ closed JB∗-subtriple) of a JBW∗-triple W , then the strong∗-topology of F coincides with the restriction to F of the strong∗-topology of W , that is, S∗(F, F∗) = S∗(W, W∗)F . Given a Banach space X, we habitually regard X as being contained in X ∗∗ and we identify the weak∗ closure, in X ∗∗, of a closed subspace Y of X with Y ∗∗. Let F be a JB∗-subtriple of a JB∗-triple E and let e be a tripotent in F ∗∗ ≡ F ⊆ E∗∗. Corollary 2.9 in [21] and [23, Proposition 3.3] prove that u is compact in F ∗∗ if, and only if, u is compact in E∗∗. σ(E ∗∗,E ∗) The following lemma, whose statement is required later, can be directly deduced from the joint strong∗-continuity of the triple product of every JBW∗-triple on bounded sets and the definition of the corresponding Peirce projection. Lemma 1.5. Let W be a JBW∗-triple. Suppose that (eλ) is a net (or a sequence) of tripotents in W converging, in the strong∗-topology of W to a tripotent e in W . Let (xµ) be a net (or a sequence) in W , converging to some x ∈ W in the strong∗-topology. Then, for each i ∈ {0, 1, 2}, the net (sequence) Pi(eλ)(xµ) tends to Pi(e)(x). (cid:3) 2. Local triple derivations on JB∗-triples are triple derivations A triple derivation on a JB∗-triple E is a linear mapping δ : E → E satisfying that δ {a, b, c} = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} , for every a, b, c ∈ E. The Jordan identity implies that, for each a, b in E, the mapping δ(a, b) = L(a, b)−L(b, a) is a triple derivation on E. Every triple derivation on a JB∗-triple is continuous (cf. [8, Corollary 2.2] and [34, Corollary 10]). The separate weak∗ continuity of the triple product of E∗∗ together with Goldstine's theorem, imply that (4) δ∗∗ is a triple derivation on E∗∗ whenever δ is a triple derivation on E. This section contains the main result of the paper, which asserts that every continuous local triple derivation T on a JB∗-triple E is a derivation. Our strategy will consist in studying the behavior of T ∗∗ on compact and range tripotents in E∗∗. We start with a technical lemma borrowed from [14]. Lemma 2.1. [14, Lemma 4] Let F be a JB∗-subtriple of a JB∗-triple E, where the latter is regarded as a Jordan Banach triple F -module with respect to its natural 6 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA triple product. Let T : F → E be a local triple derivation. Then the products of the form {a, T (b), c} vanish for every a, b, c in F with a, c ⊥ b. (cid:3) We shall survey now some of the properties of triple derivations. Let δ : E → E be a triple derivation on a JB∗-triple. Suppose that e is a tripotent in E. In such a case, δ(e) = δ {e, e, e} = 2 {δ(e), e, e}+{e, δ(e), e} = 2P2(e)(δ(e))+P1(e)(δ(e))+Q(e)(δ(e)) = 2P2(e)(δ(e)) + P1(e)(δ(e)) +(cid:16)P2(e)(δ(e))(cid:17)∗e , which implies that (5) (6) P0(e)(δ(e)) = 0 and P2(e)(δ(e)) = −(cid:16)P2(e)(δ(e))(cid:17)∗e = −Q(e)(δ(e)). If T : E → E is merely a local triple derivation on E, we can find a triple derivation δe : E → E such that T (e) = δe(e), which gives P0(e)(T (e)) = 0 and P2(e)(T (e)) = −(cid:16)P2(e)(T (e))(cid:17)∗e = −Q(e)(T (e)). Though a JB∗-triple E need not contain, in general, tripotent elements, its bid- ual has a rich set of tripotents. The next proposition explains the behavior of a continuous local triple derivation on a JB∗-triple E on compact tripotents in E∗∗. Proposition 2.2. Let T : E → E be a bounded local triple derivation on a JB∗- triple. Suppose e is a compact tripotent in E∗∗. Then the following statements hold: (a) P0(e)T ∗∗(e) = 0; (b) If a is a norm one element in E whose support tripotent is e (that is, e is a compact-Gδ tripotent), then Q(e)T (a) = Q(e)T ∗∗(e); (c) P2(e)T ∗∗(e) = −Q(e)(T ∗∗(e)). Proof. (a) Let us assume that e is a compact-Gδ tripotent. So, there exists a norm one element a in E such that s(a) = e. Let Ea denote the JB∗-subtriple of E generated by a. We have already mentioned that there exists a subset L ⊆ (0, 1] with 1 ∈ {0} ∪ L compact and a triple isomorphism Ψ from Ea onto C0(L) such that Ψ(a)(t) = t, ∀t ∈ L (see page 3). Pick a sequence of norm one elements (bn) in Ea such that bn = e + P0(e)(bn), {bn, bn+1, bn} = {bn, bn, bn+1} = bn+1, (bn) → e, in the strong∗-topology of E∗∗ (take, for example, bn(t) :=  0, n(n + 1)(t − 1 + 1 1, n ), if t ∈ L ∩ [0, 1 − 1 if t ∈ L ∩ [1 − 1 if t ∈ L ∩ [1 − 1 n ]; n+1 ]; n , 1 − 1 n+1 , 1].   . Fix a natural n. Since, the support tripotent of bn, s(bn), is a compact tripo- 0 (s(bn)) we can find (bounded) nets (cµ) and (dν ) in 0 (s(bn)) ∩ E converging to z and w in the strong∗-topology of E∗∗, respectively. tent in E∗∗, given z, w ∈ E∗∗ E∗∗ Clearly, cµ, dν ⊥ bn+1 for every µ and ν, and hence, by Lemma 2.1, Taking strong∗-limits in µ and ν we have {cµ, T (bn+1), dν } = 0 (∀µ, ν). {z, T (bn+1), w} = 0, for every n ∈ N, z, w ∈ E∗∗ 0 (s(bn)), equivalently, {P ∗∗ 0 (s(bn))(x), T (bn+1), P ∗∗ 0 (s(bn))(y)} = 0, for every n ∈ N, x, y ∈ E∗∗. LOCAL TRIPLE DERIVATIONS 7 Now, since the triple product of E∗∗ is jointly strong∗-continuous and T ∗∗ is strong∗-continuous, we can take strong∗-limit in the above expression, and by Lemma 1.5, we have {P ∗∗ 0 (e)(y)} = 0, for every x, y ∈ E∗∗. It follows, for example, from Peirce arithmetic and the third axiom in the definition of JB∗-triples, that P ∗∗ 0 (e)(x), T ∗∗(e), P ∗∗ 0 (e)T ∗∗(e) = 0. Let us consider a compact tripotent e ∈ E∗∗. By definition, there exists a decreas- ing net (eλ) of compact-Gδ tripotents in E∗∗ converging to e in the strong∗-topology of E∗∗. From the above argument, P0(eλ)T ∗∗(eλ) = 0 (∀λ), and by Lemma 1.5, P0(e)T ∗∗(e) = 0, as we desired. (b) Again, let a be a norm one element in E such that s(a) = e. Let us denote a0 = P0(e)(a). Adapting the previous arguments, we consider the JB∗-subtriple, Ea, generated by a, and pick two sequences (an) and (bn) in the closed unit ball of Ea defined by an(t) :=  and bn(t) :=(cid:26) 0, t, −n(n + 1)(t − 1 + 1 0, n+1 ), if t ∈ L ∩ [0, 1 − 1 if t ∈ L ∩ [1 − 1 if t ∈ L ∩ [1 − 1 n ]; n+1 ]; n , 1 − 1 n+1 , 1] , (n + 1)(t − 1 + 1 n+1 ), if t ∈ L ∩ [0, 1 − 1 if t ∈ L ∩ [1 − 1 n+1 ]; n+1 , 1]. Clearly, an ⊥ bn (∀n), (an) → a0 and (bn) → e in the strong∗-topology of E∗∗. Lemma 2.1 assures that {bn, T (an), bn} = 0 (∀n ∈ N). Taking strong∗ limits in the above expression we have {e, T ∗∗(a0), e} = 0, and hence {e, T ∗∗(a), e} = {e, T ∗∗(e), e} . (c) Assume, one more time, that e is a compact-Gδ tripotent and a is a norm one element in E such that s(a) = e. Since T is a local triple derivation, we can find a triple derivation δa : E → E such that T (a) = δa(a). We notice that δ∗∗ a : E∗∗ → E∗∗ is a triple derivation on E∗∗ (see (4)). Since δa is triple derivation, the identity in (b) also holds whenever we replace T with δa. Therefore, P2(e)T ∗∗(e) = P2(e)T (a) = P2(e)δa(a) = P2(e)δ∗∗ a (e) (by (5)) = −Q(e)(cid:16)δ∗∗ a (e)(cid:17) = −Q(e)(cid:16)δa(a)(cid:17) = −Q(e)(cid:16)T (a)(cid:17) = −Q(e)(cid:16)T ∗∗(e)(cid:17), which proves statement (c) for compact-Gδ tripotents in E∗∗. Let us consider a decreasing net (eλ) of compact-Gδ tripotents in E∗∗ con- verging to e in the strong∗-topology of E∗∗. Since, for each λ, P2(eλ)T ∗∗(eλ) = −Q(eλ)(T ∗∗(eλ)), Lemma 1.5, assures the desired identity for e. (cid:3) In the hypothesis of the above proposition, let a be a norm one element in E and let Ea denote the JB∗-subtriple of E generated by a. By the Gelfand theory for commutative JB∗-triples, there exists a subset L ⊆ (0, 1] with 1 ∈ {0} ∪ L compact and a triple isomorphism Ψ from Ea onto C0(L) such that Ψ(a)(t) = t (∀t ∈ L). Clearly, the range tripotent of a can be approximated, in the strong∗ topology of E∗∗, by a sequence (en) of compact-Gδ tripotents in E∗∗, that is, (en) → r(a) in the strong∗-topology. Since, by the above Proposition 2.2, P0(en)T ∗∗(en) = 0, 8 and BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA taking strong∗-limit in n we deduce, by Lemma 1.5, that P2(en)T ∗∗(en) = −Q(en)(T ∗∗(en)), (7) P0(r(a))T ∗∗(r(a)) = 0, and, P2(r(a))T ∗∗(r(a)) = −Q(r(a))(T ∗∗(r(a))). Let e be a compact or a range tripotent in E∗∗. It follows from Proposition 2.2 and (7) that T ∗∗ {e, e, e} = T ∗∗(e) = P2(e)T ∗∗(e) + P1(e)T ∗∗(e), 2 {e, e, T ∗∗(e)} = 2P2(e)T ∗∗(e) + P1(e)T ∗∗(e), and which assures that {e, T ∗∗(e), e} = Q(e)T ∗∗(e) = −P2(e)T ∗∗(e), T ∗∗ {e, e, e} = 2 {e, e, T ∗∗(e)} + {e, T ∗∗(e), e} . (8) Corollary 2.3. Let T : E → E be a continuous local triple derivation on a JB∗- triple. Suppose that e and u are two orthogonal compact tripotents in E∗∗, r1 and r2 are two orthogonal range tripotent in E∗∗ and e ⊥ r2. Then the following identities hold: T ∗∗ {e ± u, e ± u, e ± u} = 2 {e ± u, e ± u, T ∗∗(e ± u)} + {e ± u, T ∗∗(e ± u), e ± u} , T ∗∗ {r1 ± r2, r1 ± r2, r1 ± r2} = 2 {r1 ± r2, r1 ± r2, T ∗∗(r1 ± r2)} + {r1 ± r2, T ∗∗(r1 ± r2), r1 ± r2} ; and T ∗∗ {e ± r2, e ± r2, e ± r2} = 2 {e ± r2, e ± r2, T ∗∗(e ± r2)} + {e ± r2, T ∗∗(e ± r2), e ± r2} . Proof. Since e and u are two orthogonal compact tripotents in E∗∗, it follows from Proposition 3.7 in [23] that e ± u is a compact tripotent in E∗∗. It is easy to see that the sum and the difference of two orthogonal range tripotents in E∗∗ is again a range tripotent in E∗∗. Thus, the first and the second identity have been proved in (8). To see the last identity, we recall that since e ⊥ r2 and r2 is a range projection, we can find a sequence of compact tripotents (en) in E∗∗ such that en ≤ r2, and hence en ⊥ e for every n, and (en) → r2 in the strong∗-topology of E∗∗. The first identity shows that T ∗∗ {e ± en, e ± en, e ± en} = 2 {e ± en, e ± en, T ∗∗(e ± en)} + {e ± en, T ∗∗(e ± en), e ± en} , for every n. The desired equality follows by taking strong∗-limit in n. (cid:3) Let T : E → E be a bounded local triple derivation on a JB∗-triple. Another application of the local Gelfand structure of JB∗-triples allows us to see that each element a in a JB∗-triple E can be approximated in norm by a finite real linear a ⊆ E∗∗. combination of mutually orthogonal range and compact tripotents in E∗∗ Corollary 2.3 will imply that T ∗∗ behaves as a triple derivation on those elements which are finite real linear combinations of mutually orthogonal range and compact tripotents in E∗∗, and consequently, T (and hence T ∗∗) is a triple derivation. We can state now our main result, which solves a problem conjectured by M. Mackey in [31, Conjecture (C3)] (and also posed in [14, Problem 1]). LOCAL TRIPLE DERIVATIONS 9 Theorem 2.4. Every bounded local triple derivation on a JB∗-triple is a triple derivation. Proof. Let T : E → E be a bounded local triple derivation on a JB∗-triple E. Let e1, . . . , em be a family of mutually orthogonal range or compact tripotents in E∗∗. Let us pick i, j, k ∈ {1, . . . , m} with i, k 6= j. By Proposition 2.2 and (7) we have P0(ej)T ∗∗(ej) = 0. By assumptions, ei, ek ∈ E∗∗ 0 (ej) and hence (9) {ei, T ∗∗(ej), ek} = 0, by Peirce arithmetic. Now, fix i 6= j in {1, . . . , m}. Since ei and ej are compact or range tripotents in E∗∗, Corollary 2.3 implies that (10) and T ∗∗ {ei, ei, ei} = 2 {ei, ei, T ∗∗(ei)} + {ei, T ∗∗(ei), ei} ; T ∗∗ {ei ± ej, ei ± ej, ei ± ej} = 2 {ei ± ej, ei ± ej, T ∗∗(ei ± ej)} + {ei ± ej, T ∗∗(ei ± ej), ei ± ej} . Combining the last two identities we get ±2 {ei, ei, T ∗∗(ej)} + 2 {ej, ej, T ∗∗(ei)} ± {ei, T ∗∗(ej), ei)} + {ej, T ∗∗(ei), ej} ±2 {ei, T ∗∗(ei), ej} + 2 {ei, T ∗∗(ej), ej} = 0, and consequently, +4 {ej, ej, T ∗∗(ei)} + 2 {ej, T ∗∗(ei), ej} + 4 {ei, T ∗∗(ej), ej} = 0. Applying (9) we obtain (11) {ej, ej, T ∗∗(ei)} + {ei, T ∗∗(ej), ej} = 0. Consider now an element b = λiei, where e1, . . . , em are as above. Having in mind that e1, . . . , em are mutually orthogonal, we compute λ3 i {ei, ei, T ∗∗(ei)} + 2 λ2 i λj {ei, ei, T ∗∗(ej)} ; {b, T ∗∗(b), b} = (by (9)) = mXi=1 λ3 i {ei, T ∗∗(ei), ei}+2 mXi,j=1,i6=j λ2 i λj {ei, T ∗∗(ei), ej} . Finally, by (10) and (11) we have T ∗∗ {b, b, b} = 2 {T ∗∗(b), b, b} + {b, T ∗∗(b), b} . mXi=1 T ∗∗({b, b, b}) = 2 {T ∗∗(b), b, b} = 2 λ3 i T ∗∗({ei, ei, ei}); λ2 i λj {ei, ei, T ∗∗(ej)} mXi=1 mXi,j=1 mXi,j=1,i6=j (12) (13) (14) = 2 mXi=1 10 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA Since every element a ∈ E can be approximated in norm, by elements of the form b = λiei, we conclude that T {a, a, a} = 2 {T ∗∗(a), a, a} + {a, T ∗∗(a), a}, for mXi=1 every a ∈ E. Finally, an standard polarisation argument, via formula (15) {x, y, z} = 8−1 3Xk=0 2Xj=1 (−1)jik(cid:16)x + ik y + (−1)jz(cid:17)[3] (x, y, z ∈ E), assures that T is a triple derivation. (cid:3) In 1997, J.M. Isidro, W. Kaup and A. Rodr´ıguez Palacios introduce a strictly wider class of Jordan triples over the real field and called the elements of this new category real JB∗-triples. A real JB∗-triple is a norm closed real subtriple of a JB∗-triple (cf. [25]). When restricted to real scalar multiplication, every (complex) JB∗-triple is also a real JB∗-triple. The class of real JB∗-triples also includes real C∗-algebras, JB-algebras, JC-algebras, real JB∗-algebras, operator spaces between real, complex and quaternionic Hilbert spaces and real Hilbert spaces. Every real JB∗-triple can be complexified to become a JB∗-triple. Furthermore, every real JB∗-triple A is a real form of a complex JB∗-triple, that is, there exist a (complex) JB∗-triple B ∼= A ⊕ iA and a period 2 conjugate linear isometry τ : B → B (which is also a JB∗-triple homomorphism) such that A = {b ∈ B : τ (b) = b} (see [25]). Let us consider eτ : B∗ → B∗ defined by eτ (φ)(z) = φ(τ (z)). The mapping eτ is a conjugation on B∗, and the restriction mapping (B∗)eτ −→ (Bτ )∗ (∼= A∗) φ 7→ φA is an isometric bijection, where (B∗)eτ := {φ ∈ B∗ : eτ (φ) = φ}. The second transpose τ ∗∗ : B∗∗ → B∗∗ is a period 2 conjugate linear isometry satisfying (B∗∗)τ ∗∗ = A∗∗. A real JBW∗-triple is a real JB∗-triple which is also a dual Banach space. Every real JBW∗-triple has a unique (isometric) predual and separately weak∗ continuous triple product (see [32]). The bidual of every real JB∗-triple is a JBW∗-triple (compare [25]). The notions of range, support or compact-Gδ, open, closed and compact tripo- tents also make sense in the bidual of every real JB∗-triple. When real JB∗-triples are regarded as real forms of complex JB∗-triples, the generalised Urysohn's lem- mas proved by the second and third author of this note in [23] remain valid for real JB∗-triples. Furthermore, an appropriate local Gelfand theory for single- generated real JB∗-subtriples is also available in the real setting (cf. [15, §3]). Thus, the arguments given above to prove Theorem 2.4 can be applied to show that every bounded local triple derivation T on a real JB∗-triple A satisfies that T {a, a, a} = 2 {T (a), a, a} + {a, T (a), a}, for every a ∈ A. Unfortunately, the polar- isation formula (15) employed at the end of the proof of Theorem 2.4 is not valid for real JB∗-triples, so we cannot obtain the conclusion of that theorem in the real setting. Corollary 2.5. Let T be a continuous local triple derivation on a real JB∗-triple A. Then T {a, a, a} = 2 {T (a), a, a} + {a, T (a), a}, for every a ∈ A, that is, T is a LOCAL TRIPLE DERIVATIONS 11 triple derivation of the symmetrized Jordan triple product 3 < a, b, c >:= {a, b, c} + {c, a, b} + {b, c, a} . (cid:3) In the light of the above corollary, it seems natural to consider the following problem: Problem 2.6. Is every bounded local triple derivation on a real JB∗-triple a triple derivation? In a recent paper (see [34]), B. Russo and the third author of this note initiated the study of triple module-valued triple derivations on (real and complex) JB∗- triples. Let E be a complex (resp. real) Jordan triple. We recall that a Jordan triple E-module (also called triple E-module) is a vector space X equipped with three mappings {., ., .}1 : X × E × E → X, {., ., .}2 : E × X × E → X satisfying the following axioms: and {., ., .}3 : E × E × X → X (J T M 1) {x, a, b}1 is linear in a and x and conjugate linear in b (resp., trilinear), {a, b, x}3 is linear in b and x and conjugate linear in a (resp., trilinear) and {a, x, b}2 is conjugate linear in a, b, x (resp., trilinear) (J T M 2) {x, b, a}1 = {a, b, x}3, and {a, x, b}2 = {b, x, a}2 for every a, b ∈ E and x ∈ X. (J T M 3) Denoting by {., ., .} any of the products {., ., .}1, {., ., .}2 and {., ., .}3, the Jordan identity (1) holds whenever one of the elements is in X and the rest are in E. When the products {., ., .}1, {., ., .}2 and {., ., .}3 are (jointly) continuous we shall say that X is a Banach (Jordan) triple E-module. Henceforth, the triple products {·, ·, ·}j, j = 1, 2, 3, which occur in the definition of Jordan triple module will be denoted simply by {·, ·, ·} whenever the meaning is clear from the context. It is obvious that every real or complex Jordan triple E is a real triple E-module. The dual space, E∗, of a complex (resp., real) Jordan Banach triple E is a complex (resp., real) triple E-module with respect to the products: {a, b, ϕ} (x) = {ϕ, b, a} (x) := ϕ {b, a, x} , and, {a, ϕ, b} (x) := ϕ {a, x, b}, for all ϕ ∈ E∗, a, b, x ∈ E. Let E be a complex (respectively, real) JB∗-triple and let X be a triple E-module. We recall that a conjugate linear (respectively, linear) mapping δ : E → X is said to be a triple or ternary derivation if δ{a, b, c} = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} . A conjugate linear (respectively, linear) mapping T : E → X will be called a local triple derivation if for each a in E there exists a triple derivation δa : E → X such that T (a) = δa(a). Problem 2.7. Is every continuous local triple derivation from a real or complex JB∗-triple E into a Banach triple E-module a triple derivation? 12 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA 2.1. Automatic continuity. We establish now an automatic continuity result for local triple derivations, giving a positive answer to Problem (1.3). We shall review the arguments given by T. Barton and Y. Friedman to show that a triple derivation on a JB∗-triple is a triple derivation. Let X be a complex Banach space. We recall that a linear mapping T : X → X is dissipative if for each x ∈ X and each functional φ ∈ X ∗ with kxk = kφk = φ(x) = 1 we have ℜeφ(T (x)) ≤ 0. It is known that T is continuous whenever it is dissipative (compare [10, Proposition 3.1.15]). In [8, Theorem 2.1], Barton and Friedman prove that every derivation on a JB∗-triple E is dissipative and hence continuous. Let us review some aspect in their proof. Let x be an element in E, let φ a functional in E∗ with kxk = kφk = φ(x) = 1 and let δ : E → E a triple derivation. Applying Peirce arithmetic and support tripotents, the arguments in the proof of Theorem 2.1 in [8] show that φδ({x, x, x}−x) = 0 and hence ℜeφ(δ(x)) = 0. It is also justified in the same result that φ {a, x, x} = φ(a) and φ {x, a, x} = φ(a), for every a in E. Let T : E → E be a local triple derivation, not assumed a priori to be continuous, on E. Pick a triple derivation δx satisfying δx(x) = T (x). In this case, we have 2φT (x) + φT (x) = 2φ {T (x), x, x} + φ {x, T (x), x} = 2φ {δx(x), x, x} + φ {x, δx(x), x} = φδx {x, x, x} . It follows from the above that 2ℜeφT (x) = φδx {x, x, x} −φT (x) = φδx {x, x, x} −φδx(x) = φδx({x, x, x} −x) = 0, because δx is a triple derivation on E. This shows that T is dissipative. Theorem 2.8. Every local triple derivation on a (real or complex) JB∗-triple is continuous. Consequently, every local triple derivation on a JB∗-triple is a triple derivation. (cid:3) 3. Local triple derivations and generalised derivations on C∗-algebras We begin this section with the following corollary which solves Problem 2 in [14]. Theorem 3.1. Every local triple derivation on a C∗-algebra is a triple derivation. (cid:3) In [14], it is established that every local triple derivation on a unital C∗-algebra is a triple derivation. The strategy to prove this result relies on the connections between (local) triple derivations and generalised derivations on unital C∗-algebras in the sense introduced by J. Li and Z. Pan in [30]. We recall that a linear mapping D from a unital C∗-algebra B to a (unital) Banach B-bimodule X is called a generalised derivation whenever the identity D(ab) = D(a)b + aD(b) − aD(1)b holds for every a, b in B. A generalised Jordan derivation from B to X is a linear mapping D : B → X satisfying D(a ◦ b) = D(a) ◦ b + a ◦ D(b) − Ua,bD(1), for every a, b in B, where the Jordan product is given by a ◦ b := 1 2 (ab + ba) and Ua,b(x) := 1 2 (axb+bxa) (cf. [14]). Every generalised (Jordan) derivation D : B → X with D(1) = 0 is a (Jordan) derivation. As remarked in [14, Remark 8], generalised Jordan derivations from B to X and a generalised derivations from B to X coincide. A linear mapping T : B → X is a local generalised (Jordan) derivation if for LOCAL TRIPLE DERIVATIONS 13 each a ∈ B, there exists a generalised (Jordan) derivation Da : B → X with T (a) = Da(a). In the setting of unital C∗-algebras, J. Alaminos, M. Bresar, J. Extremera, and A. Villena [5, Corollary 3.2] and J. Li and Z. Pan [30, Corollary 2.9] established the following interesting result: Theorem 3.2. ([5, Corollary 3.2], [30, Corollary 2.9]) Suppose that B is a unital C∗-algebra and M is a unital Banach B-bimodule. Then for any bounded linear map T from B to M , the following are equivalent: (i) T is a generalised (Jordan) derivation from B to M ; (ii) T is a local generalised (Jordan) derivation from B to M ; (iii) aT (b)c = 0, whenever a, b, c ∈ B with ab = bc = 0. (cid:3) Unfortunately, the above definitions of generalised (Jordan) derivations and local generalised (Jordan) derivation make sense only when the domain is a unital C∗- algebra. However, statement (iii) in the above theorem makes sense in the non- unital setting too. Our next goal is to generalise the above theorem to the setting of not necessarily unital C∗-algebras. In a first step we should consider a consequence derived from Theorem 4.5 in [6]. First we recall a definition taken from [6, §4]: a generalised derivation from a Banach algebra A to a Banach A-bimodule X is a linear operator D : A → X for which there exists ξ ∈ X ∗∗ satisfying D(ab) = D(a)b + aD(b) − aξb (a, b ∈ A). Theorem 3.3. [6, Theorem 4.5] Suppose that A is a C∗-algebra and X is an essential Banach A-bimodule. Let T : A → X be a continuous linear operator satisfying a, b, c ∈ A, ab = bc = 0 ⇒ aT (b)c = 0. Then T is a generalized derivation. Given a C∗-algebra, A, the multiplier algebra of A, M (A), is the set of all elements x ∈ A∗∗ such that, for each element a ∈ A, xa and ax both lie in A. We notice that M (A) is a C∗-algebra and contains the unit element of A∗∗. Clearly, A = M (A) whenever A is unital. For later purposes, we recall some basic results on Arens regularity (cf. [7]). Let X, Y and Z be Banach spaces and let m : X × Y → Z be a bounded bilinear mapping. Defining m∗(z′, x)(y) := z′(m(x, y)) (x ∈ X, y ∈ Y, z′ ∈ Z ∗), we obtain a bounded bilinear mapping m∗ : Z ∗ × X → Y ∗. Iterating the process, we define a mapping m∗∗∗ : X ∗∗ × Y ∗∗ → Z ∗∗. The mapping x′′ 7→ m∗∗∗(x′′, y′′) is weak∗ to weak∗ continuous whenever we fix y′′ ∈ Y ∗∗, and the mapping y′′ 7→ m∗∗∗(x, y′′) is weak∗ to weak∗ continuous for every x ∈ X. One can consider the transposed mapping mt : Y × X → Z, mt(y, x) = m(x, y) and the extended mapping mt∗∗∗t : X ∗∗ × Y ∗∗ → Z ∗∗. In this case, the mapping x′′ 7→ mt∗∗∗t(x′′, y) is weak∗ to weak∗ continuous whenever we fix y ∈ Y , and the mapping y′′ 7→ mt∗∗∗t(x′′, y′′) is weak∗ to weak∗ continuous for every x′′ ∈ X ∗∗. In general, the mappings mt∗∗∗t and m∗∗∗ do not coincide (cf. [7]). When mt∗∗∗t = m∗∗∗, we say that m is Arens regular. It is well known that the product of every C∗-algebra is Arens regular and the unique Arens extension of the product of A to A∗∗ ×A∗∗ coincides with the product of its enveloping von Neumann algebra (cf. [16, Corollary 3.2.37]). 14 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA Let X be a Banach A-bimodule over a C∗-algebra A. Let us denote by π1 : A × X → X and π2 : X × A → X the corresponding module operations given by π1(a, x) = ax and π2(x, a) = xa, respectively. By an abuse of notation, given a ∈ A∗∗ and z ∈ X ∗∗, we shall frequently write az = π∗∗∗ (z, a). It is known that X ∗∗ is a Banach A∗∗-bimodule for the just defined operations ([16, Theorem 2.6.15(iii)]). It is also known that whenever (aλ) and (xµ) are nets in A and X, respectively, such that aλ → a ∈ A∗∗ in the weak∗ topology of A∗∗ and xµ → x ∈ X ∗∗ in the weak∗ topology of X ∗∗, then (a, z) and za = π∗∗∗ 1 2 (16) ax = π∗∗∗ 1 (a, x) = lim λ lim µ aλxµ and xa = π∗∗∗ 2 (x, a) = lim µ lim λ xµaλ in the weak∗ topology of X ∗∗ (cf. [16, 2.6.26]). Our next proposition completes the whole picture. Proposition 3.4. Let X be an essential Banach A-bimodule over a C∗-algebra A and let T : A → X be a bounded linear operator. The following are equivalent: (a) T ∗∗M(A) : M (A) → X ∗∗ is a generalised (Jordan) derivation; (b) T ∗∗ : A∗∗ → X ∗∗ is a generalised (Jordan) derivation; (c) T ∗∗ = d + MT ∗∗(1), where d : A∗∗ → X ∗∗ is a derivation and MT ∗∗(1)(a) = T ∗∗(1) ◦ a = 1 2 (aT ∗∗(1) + T ∗∗(1)a); (d) T ∗∗ : A∗∗ → X ∗∗ is a local generalised (Jordan) derivation; (e) T ∗∗M(A) : M (A) → X ∗∗ is a local generalised (Jordan) derivation; (f ) T is a generalised derivation; (g) aT (b)c = 0, whenever ab = bc = 0 in A. Proof. The implications (b) ⇔ (c) ⇒ (d) ⇒ (e) ⇒ (g) are clear. The equivalence (f ) ⇔ (g) was established in Theorem 3.3 ([6, Theorem 4.5]). To see (a) ⇒ (b) suppose that T ∗∗M(A) is a generalised derivation, i.e., T (ab) = T (a)b + aT (b) − aT ∗∗(1)b, for every a, b in M (A). Since, by Goldstine's theorem, the closed unit ball of A is weak∗ dense in the closed unit ball of A∗∗, we deduce from (16) that the above equality also holds when a and b are in A∗∗ and T is replaced with T ∗∗. Thus, T ∗∗ is a generalised derivation. We shall finally prove the implication (g) ⇒ (a). Let a, b, c be elements in M (A) with ab = bc = 0. We may assume that a, b and c lie in the closed unit ball of M (A). We observe that a[2n−1]b = 0 for every natural n. Therefore, αb = 0, for every α in the JB∗-subtriple, M (A)a, of M (A) generated by a. We can similarly show that (17) αβ = βγ = 0 for every α ∈ M (A)a, β ∈ M (A)b and γ ∈ M (A)c. Since M (A) is a C∗-subalgebra of A∗∗, by Goldstine's theorem, we can find nets (xλ), (yµ) and (zν) in the closed unit ball of A, converging in the weak∗ topology of A∗∗ to a[ 1 3 ] and c[ 1 3 ], b[ 1 3 ]x∗ 3 ]y∗ λa[ 1 µb[ 1 3 ](cid:17), and (cid:16)c[ 1 3 ]z∗ ν c[ 1 3 ](cid:17) lie in A, and by (17), we have 3 ], respectively. The nets (cid:16)a[ 1 3 ](cid:17)(cid:16)b[ 1 3 ](cid:17) = 0 =(cid:16)b[ 1 µb[ 1 µb[ 1 (cid:16)a[ 1 λa[ 1 3 ]x∗ 3 ]y∗ 3 ]y∗ 3 ](cid:17)(cid:16)c[ 1 3 ](cid:17), (cid:16)b[ 1 3 ](cid:17) , ν c[ 1 3 ]z∗ LOCAL TRIPLE DERIVATIONS 15 for every λ, µ and ν. Our hypothesis assures that (cid:16)a[ 1 3 ]x∗ λa[ 1 3 ](cid:17) T(cid:16)b[ 1 3 ]y∗ µb[ 1 3 ](cid:17)(cid:16)c[ 1 3 ]z∗ ν c[ 1 3 ](cid:17) = 0, for every λ, µ and ν. Taking weak∗ limit in ν, it follows from the properties of π∗∗∗ that 2 (cid:16)a[ 1 3 ]x∗ λa[ 1 3 ](cid:17) T(cid:16)b[ 1 3 ]y∗ µb[ 1 3 ](cid:17) c = 0, for every λ, and µ. Finally, taking weak∗ limits first in µ and later in λ we have aT ∗∗(b)c = 0. We have therefore shown that aT ∗∗(b)c = 0 whenever ab = bc = 0 in M (A). Since X is an essential A-bimodule, and A∗∗ is unital, it can be easily checked that X ∗∗ is a unital M (A)-bimodule. Thus, the statement (a) follows from Theorem 3.2 above. (cid:3) It is worth to notice that the proof of the above theorem makes use of the local structure of a C∗-algebra when it is regarded as a JB∗-triple. Jordan techniques are also employed in the next remark. Remark 3.5. In the hypothesis above, every statement in Proposition 3.4 is equiv- alent to: (h) For each right closed ideal R of A and each left closed ideal L of A, we have T (R ∩ L) ⊆ XL + RX. Indeed, suppose that T ∗∗ : A∗∗ → X ∗∗ is a generalised derivation. Let a be an element in the intersection R ∩ L. The element a[3] = aa∗a lies in R and in L. By induction on n, we proved that a[2n−1] belongs to L ∩ R for every natural n. Therefore, the JB∗-subtriple, Aa, of A generated by a is contained in R ∩ L. Let us take b ∈ Aa satisfying b[3] = a. Thus, T (a) = T (b[3]) = T (bb∗b) = T (b)b∗b + bT (b∗b) − bT ∗∗(1)b∗b ∈ XL + RX, which proves (b) ⇒ (h). Assume now that T (R ∩ L) ⊆ XL + RX whenever R is a right closed ideal of A and L is a left closed ideal of A. Pick a, b and c in A with ab = 0 = bc. Let R = {x ∈ A : ax = 0} and L = {x ∈ A : xc = 0}. In this case, R is a right closed ideal, L is a left closed ideal and b ∈ L ∩R, then, by assumptions, T (b) ∈ XL + RX. Therefore, aT (b)c ∈ aXLc + aRXc = {0}, which gives the desired equivalence. When, in the hypothesis of the above Remark 3.5, X coincides with A, we have XL+RX = L+R. Under the additional assumption of A being unital, V. Shul'man established in [35, Theorem 1] that a bounded linear mapping T : A → A satisfies statement (h) in Remark 3.5 if, and only if, T = D + La, where D : A → A is a derivation and La is the left multiplication operator by the element a = T (1) (if and only if T is a generalised derivation). So, our Proposition 3.4 and the equivalence with Remark 3.5(h) provides us a generalisation of the result proved by Shul'man. We culminate this section exploring the connections with (local) triple deriva- tions on a C∗-algebra. Let δ : A → A be a triple derivation on a C∗-algebra. We have already commented that δ∗∗ : A∗∗ → A∗∗ is a triple derivation (see (4)). By Lemma 1 in [24] δ∗∗(1)∗ = −δ∗∗(1). It can be easily seen that δ∗∗(a ◦ b) = δ∗∗({a, 1, b}) = {δ∗∗(a), 1, b} + {a, δ∗∗(1), b} + {a, 1, δ∗∗(b)} = δ∗∗(a) ◦ b + a ◦ δ∗∗(b) − Ua,bδ∗∗(1), 16 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA which shows that δ∗∗ is a generalised Jordan derivation and hence a generalised derivation (compare [14, Comments before Proposition 5 and Remark 8]). More- over, δ∗∗(a∗) = δ∗∗ {1, a, 1} = 2 {δ∗∗(1), a, 1} + {1, δ∗∗(a), 1} = 2δ∗∗(1) ◦ a∗ + δ∗∗(a)∗. In particular, every triple derivation δ with δ∗∗(1) = 0 is a symmetric derivation or a ∗-derivation (i.e. δ is a derivation with δ(a∗) = δ(a)∗, for every a ∈ A). In particular, for every triple derivation δ, δ∗∗ − δ( 1 2 δ∗∗(1), 1) is a ∗-derivation on A. Corollary 3.6. Let T : A → A be a linear operator on a C∗-algebra. The following are equivalent: (a) T is a local triple derivation; (b) T is a triple derivation; (c) T ∗∗ is a bounded generalised derivation with T ∗∗(1) = −T ∗∗(1)∗ and T ∗∗ − LT ∗∗(1) is a symmetric operator; (d) T ∗∗ − LT ∗∗(1) is a bounded ∗-derivation and T ∗∗(1) = −T ∗∗(1)∗; (e) T ∗∗ is a bounded generalised derivation with T ∗∗(1) = −T ∗∗(1)∗ and T ∗∗ − δ( 1 2 T ∗∗(1), 1) is a symmetric operator; (f ) T ∗∗ − δ( 1 (g) T ∗∗(1) = −T ∗∗(1)∗, T ∗∗ − LT ∗∗(1) is a symmetric operator, T is bounded, and 2 δ∗∗(1), 1) is a bounded ∗-derivation and T ∗∗(1) = −T ∗∗(1)∗; aT (b)c = 0, whenever ab = bc = 0 in A. (h) T ∗∗(1) = −T ∗∗(1)∗, T ∗∗ − δ( 1 2 δ∗∗(1), 1) is a symmetric operator, T is bounded, and aT (b)c = 0, whenever ab = bc = 0 in A. (cid:3) References [1] C.A. Akemann, The general Stone-Weierstrass problem, J. Funct. Anal. 4, 277-294 (1969). [2] C.A. Akemann, Left ideal structure of C-algebras, J. Funct. Anal. 6, 305-317 (1970). [3] C.A. Akemann, A Gelfand representation theory for C-algebras, Pacific J. Math. 39, 1-11 (1971). [4] C.A. Akemann, G.K. Pedersen, Facial structure in operator algebra theory. Proc. London Math. Soc. 64, 418-448 (1992). [5] J. Alaminos, M. Bresar, J. Extremera, A. Villena, Characterizing homomorphisms and deriva- tions on C∗-algebras, Proc. Roy. Soc. Edinb. A 137 1-7 (2007). [6] J. Alaminos, M. Bresar, J. Extremera, A. Villena, Maps preserving zero products, Studia Math. 193, no. 2, 131-159 (2009). [7] R. Arens, The adjoint of a bilinear operation, Proc. Amer. Math. Soc. 2, 839-848 (1951). [8] T.J. Barton, Y. Friedman, Bounded derivations of JB∗-triples, Quart. J. Math. Oxford 41, 255-268 (1990). [9] T.J. Barton, R.M. Timoney, Weak∗-continuity of Jordan triple products and its applications. Math. Scand. 59, 177-191 (1986). [10] O. Bratteli, D.W. Robinson, Operator algebras and quantum statistical mechanics, Springer- Verlag, 2002. [11] L.J. Bunce, Norm preserving extensions in JBW∗-triple, Quart. J. Math. Oxford 52, No.2, 133-136 (2001). [12] L.G. Brown, Semicontinuity and multipliers of C-algebras, Canad. J. Math. 40, no. 4, 865- 988 (1988). [13] M. Burgos, F.J. Fern´andez-Polo, J. Garc´es, J. Mart´ınez, A.M. Peralta, Orthogonality pre- servers in C∗-algebras, JB∗-algebras and JB∗-triples, J. Math. Anal. Appl. 348, 220-233 (2008). [14] M. Burgos, F.J. Fern´andez-Polo, J.J. Garc´es, A.M. Peralta, Local triple derivations on C∗- algebras, to appear in Communications in Algebra. LOCAL TRIPLE DERIVATIONS 17 [15] M. Burgos, A.M. Peralta, M. Ram´ırez, M.E. Ruiz Morillas, von Neumann regularity in Jordan-Banach triples, to appear in Proceedings of Jordan structures in Algebra and Analysis Meeting. Tribute to El Amin Kaidi for his 60th birthday. Almer´ıa, 20, 21 y 22 de Mayo de 2009, Edited by J. Carmona et al., Universidad de Almer´ıa, 2010. [16] H.G. Dales, Banach algebras and automatic continuity. London Mathematical Society Mono- graphs (New Series), Volume 24, Oxford Science Publications. The Clarendon Press, Oxford University Press, New York, 2000. [17] S. Dineen, The second dual of a JB∗-triple system, n: J. Mujica (Ed.), Complex analysis, Functional Analysis and Approximation Theory, North-Holland, Amsterdam, 1986. [18] C.M. Edwards, G.T. Ruttimann, On the facial structure of the unit balls in a JBW∗-triple and its predual. J. London Math. Soc. 38, 317-322 (1988). [19] C.M. Edwards, G.T. Ruttimann, Compact tripotents in bi-dual JB∗-triples, Math. Proc. Camb. Phil. Soc. 120, 155-173 (1996). [20] Y. Friedman, B. Russo, Solution of the contractive projection problem, J. Funct. Analysis, 60, 56-79 (1986). [21] F.J. Fern´andez-Polo, A.M. Peralta, Closed tripotents and weak compactness in the dual space of a JB∗-triple, J. London Math. Soc. 74, 75-92 (2006). [22] F.J. Fern´andez-Polo, A.M. Peralta, Compact tripotents and the Stone-Weierstrass Theorem for C∗-algebras and JB∗-triples, J. Operator Theory, 58, no. 1, 157-173 (2007). [23] F.J. Fern´andez-Polo, A.M. Peralta, Non-commutative generalisations of Urysohn's lemma and hereditary inner ideals, J. Funct. Anal. 259, 343-358 (2010). [24] T. Ho, J. Martinez-Moreno, A.M. Peralta, B. Russo, Derivations on real and complex JB∗- triples, J. London Math. Soc. (2) 65, no. 1, 85-102 (2002). [25] J.M. Isidro, W. Kaup, A. Rodr´ıguez, On real forms of JB∗-triples, Manuscripta Math. 86, 311-335 (1995). [26] B.E. Johnson, Local derivations on C∗-algebras are derivations, Trans. Amer. Math. Soc. 353, 313-325 (2001). [27] R. Kadison, Local derivations, J. Algebra 130, no. 2, 494-509 (1990). [28] W. Kaup, A Riemann Mapping Theorem for bounded symmentric domains in complex Ba- nach spaces, Math. Z. 183, 503-529 (1983). [29] W. Kaup, Contractive projections on Jordan C∗-algebras and generalizations, Math. Scand. 54, 95-100 (1984). [30] J. Li, Z. Pan, Annihilator-preserving maps, multipliers, and derivations, Linear Algebra Appl. 423, 5-13 (2010). [31] M. Mackey, Local derivations on Jordan triples, preprint 2012 (arXiv:1207.5394v1). [32] J. Mart´ınez, A.M. Peralta, Separate weak-continuity of the triple product in dual real JB∗- triples, Math. Z. 234, 635-646 (2000). [33] A.M. Peralta, A. Rodr´ıguez Palacios, Grothendieck's inequalities for real and complex JBW*- triples, Proc. London Math. Soc. (3) 83, no. 3, 605-625 (2001). [34] A.M. Peralta, B. Russo, Automatic continuity of derivations on C∗-algebras and JB∗-triples, preprint 2010. [35] V. Shul'man, Operators preserving ideals in C∗-algebras, Studia Math. 109, 67-72 (1994). [36] L.L. Stacho, A projection principle concerning biholomorphic automorphisms, Acta Sci. Math. 44, 99-124 (1982). Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain. e-mail: [email protected] E-mail address: [email protected] 18 BURGOS, FERN ´ANDEZ-POLO, AND A.M. PERALTA Departamento de Matematicas, Facultad de Ciencias Sociales y de la Educacion, Universidad de C´adiz, 11405, Jerez de la Frontera, Spain. E-mail address: [email protected] Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain. E-mail address: [email protected] Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain.
1510.07289	2	1510	2017-10-20T23:04:49	On Dvoretzky's theorem for subspaces of $L_p$	[ "math.FA", "math.MG" ]	We prove that for any $2<p<\infty$ and for every $n$-dimensional subspace $X$ of $L_p$, represented on $\mathbb R^n$, whose unit ball $B_X$ is in Lewis' position one has the following two-level Gaussian concentration inequality: \[ \mathbb P\left( \big\| \\|Z\\| - \mathbb E\\|Z\\| \big\| > \varepsilon \mathbb E\\|Z\\| \right) \leq C \exp \left (- c \min \left\{ \alpha_p \varepsilon^2 n, (\varepsilon n)^{2/p} \right\} \right), \quad 0<\varepsilon<1 , \] where $Z$ is a standard $n$-dimensional Gaussian vectors, $\alpha_p>0$ is a constant depending only on $p$ and $C,c>0$ are absolute constants. As a consequence we show optimal lower bound for the dimension of almost spherical sections for these spaces. In particular, for any $2<p<\infty$ and every $n$-dimensional subspace $X$ of $L_p$, the Euclidean space $\ell_2^k$ can be $(1+\varepsilon)$-embedded into $X$ with $k\geq c_p \min\{ \varepsilon^2 n , (\varepsilon n)^{2/p}\}$, where $c_p>0$ is a constant depending only on $p$. This improves upon the previously known estimate due to Figiel, Lindenstrauss and V. Milman.	math.FA	math	On Dvoretzky's theorem for subspaces of Lp Grigoris Paouris∗ Petros Valettas† September 24, 2018 Abstract We prove that for any 2 < p < ∞ and for every n-dimensional subspace X of L p, represented on Rn, whose unit ball BX is in Lewis' position one has the following two-level Gaussian concentration inequality: > εEkZk(cid:17) ≤ C exp(cid:16)−c minnαpε2n, (εn)2/po(cid:17) , 0 < ε < 1, P(cid:16)(cid:12)(cid:12)(cid:12)kZk − EkZk(cid:12)(cid:12)(cid:12) where Z is a standard n-dimensional Gaussian vectors, αp > 0 is a constant depending only on p and C, c > 0 are absolute constants. As a consequence we show optimal lower bound for the dimension of almost spherical sections In particular, for any 2 < p < ∞ and every n-dimensional for these spaces. subspace X of L p, the Euclidean space ℓk 2 can be (1 + ε)-embedded into X with k ≥ cp min{ε2n, (εn)2/p}, where cp > 0 is a constant depending only on p. This improves upon the previously known estimate due to Figiel, Lindenstrauss and V. Milman. 1 Introduction In the present note we study the classical result of Dvoretzky [7] on almost spher- ical sections of normed spaces in the case of subspaces of Lp. Grothendieck in [14] motivated by the well known Dvoretzky-Rogers lemma from [8] to ask if every finite-dimensional normed space has lower dimensional subspaces which are almost Euclidean and their dimension grows with respect to the dimension of the ambient space. Dvoretzky in [7] gave an affirmative answer in the above question by proving that for any positive integer k and every ε ∈ (0, 1) there exists N = N(k, ε) with the following property: For every n ≥ N and any n-dimensional normed space X there ex- ists k-dimensional subspace E which (1 + ε)-isomorphic to the Euclidean space ℓk 2. In ∗Supported by the NSF CAREER-1151711 grant; †Supported by the NSF grant DMS-1612936. 2010 Mathematics Subject Classification. Primary: 46B06, 46B07, 46B09 Keywords and phrases. Dvoretzky's theorem, Almost Euclidean subspaces, L p spaces, Concentration of measure, Gaussian analytic inequalities, isotropic measures on S n−1 1 modern functional analytic language this means that every infinite-dimensional Ba- nach space contains ℓn 2's uniformly. Dvoretzky's proof in [7, Theorem 1] provides the quantitative estimate N(k, ε) ≥ exp(cε−2k2 log2 k) (see [35] for a related discussion), for some absolute constant c > 01 However, the aforementioned estimate is not optimal. The optimal dependence with respect to the dimension was proved later by V. Mil- man in his groundbreaking work [22] where he obtained N(k, ε) ≥ exp(ckε−2 log 1 ε ) (an alternative approach which yields the same estimate was presented by Szankowski in [35]). Equivalently, this states that for any ε ∈ (0, 1) there exists a function c(ε) > 0 with the following property: for every n-dimensional normed space X there exists k ≥ c(ε) log n and a linear map T : ℓk 2 → X with kxk2 ≤ kT xkX ≤ (1 + ε)kxk2 for all x ∈ ℓk 2 can be (1 + ε)-embedded into X or that X has a k-dimensional subspace which is (1 + ε)-Euclidean and we write ℓk 2 2. In that case we say that ℓk The example of X = ℓn ∞ shows that this result is best possible with respect to n (see [22] or [10, Proposition 3.2] for the details). The approach of [22] is probabilistic in nature and provides that the vast majority of subspaces (in terms of the Haar probability measure on the Grassmannian manifold Gn,k) are (1 + ε)-spherical, as long as k ≤ c(ε)k(X), where k(X) is the critical dimension of X (see below for the definition). Nowadays this is customary addressed as the randomized Dvoretzky theorem or random version of Dvoretzky's theorem. V. Milman in this work revealed the significance of the concentration of measure as a basic tool for the understanding of the high-dimensional structures. That was the starting point for many applications of the concentration of measure method in high-dimensional phenomena. Since then, this tool has found numerous applications in various fields such as quantum information [3], combinatorics [6], random matrices [38], compressed sensing [11], theoretical computer science [21], geometry of high-dimensional probability measures [9] and more. ֒→ X. 1+ε Another remarkable fact of V. Milman's approach is that the critical quantity k(X) can be described in terms of the global parameters of the space. In particular, X /b2(X) where Z is a standard Gaussian random vector in X and b(X) = k(X) ≃ EkZk2 maxkθk2 =1 kθkX . Then, one can find a good position of the unit ball of X for which k(X) is large enough with respect to n (see [23] for further details). It has been proved in [24] that this formulation is optimal with respect to the dimension k(X) in the sense that the k-dimensional subspaces which are 4-Euclidean with probability greater than n n+k cannot exceed Ck(X) (see [15] for a recent development on this fact). The proof of [22] gave the estimate c(ε) ≥ cε2/ log 1 ε and this was improved to c(ε) ≥ cε2 by Gordon in [13] and later, adopting the methods of V. Milman, by Schechtman in [27]. This dependence is known to be optimal in the setting of the randomized Dvoretzky theorem; see [30]). The works of Schechtman in [29] and Tikhomirov in [36] established that the dependence on ε in the randomized Dvoretzky for ℓn ε and this is best possible. Optimal bounds on c(ε) in ∞ the randomized Dvoretzky for ℓn is of the order ε/ log 1 p, 1 ≤ p ≤ ∞ have recently been studied in [25]. As far as the dependence on ε in the "existential version" of Dvoretzky's theorem 2 in any is concerned, Schechtman proved in [28] that one can always (1 + ε)-embed ℓk 1Here and elsewhere in this paper c and C denote positive absolute constants. 2 n-dimensional normed space X with k ≥ cε log n/(log 1 ε )2. Tikhomirov in [36] proved that for 1-symmetric space X we may have k ≥ c log n/ log 1 ε and this was subsequently extended by Fresen in [12] for permutation invariant spaces with bounded basis constant. For more detailed information on the subject, explicit statements and historical remarks the reader is referred to the recent monograph [2]. The purpose of this note is to study the dependence on ε and dimension in Dvoretzky's theorem for finite-dimensional subspaces of Lq, 2 < q < ∞. The case of subspaces of Lq, 1 ≤ q < ∞ have been previously studied in the classical paper [10] by Figiel, Lindenstrauss and V. Milman. The approach in [10] is based on V. Milman's asymptotic formula and the fact that the Lp spaces enjoy the cotype property. Let us recall that for 2 ≤ q < ∞ the q-cotype constant of a normed space X in n vectors, denoted by Cq(X, n), is defined as the smallest constant C > 0 which satisfies: n  Xi=1 kzikq X 1/q , n ≤ CE(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 εizi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X for any n vectors z1, . . . , zn ∈ X. Then, the q-cotype constant of X is defined as Cq(X) := supn Cq(X, n). Following the terminology of G. Pisier, the notion of cotype is a super-property, that is it depends only on the finite dimensional subspaces of the space. It is also isomorphic invariant and the spaces Lp, 1 ≤ p < ∞ are of cotype q = max{2, p} with Cq(Lp) = O(q1/2) (see [1] for a proof). Therefore, for any finite- dimensional X of Lq, 2 < q < ∞ we have Cq(X) ≤ C √q. The authors in [10], using the the classical Dvoretzky-Rogers lemma, show that any n-dimensional normed space X of cotype q whose unit ball is in John's position (see e.g. [10] for the related definition) satisfies k(X) ≥ cC−2 It follows that if X is an n-dimensional subspace of Lq, 2 < q < ∞ whose unit ball is in John's position then k(X) ≥ cq−1n2/q and the standard concentration techniques yield (1 + ε)-spherical sections of X of dimension k ≥ cq−1ε2n2/q (see [10] for the details). Moreover, the same argument provides k(X) ≥ cn for any n-dimensional subspace X of Lq with 1 ≤ q < 2 in John's 2 can be (1 + ε)-embedded into X with k ≥ cε2n which is best position, and thus ℓk possible. In the present note we show that for the range 2 < q < ∞ the estimate can be considerably improved. Our result reads as follows. q (X)n2/q. Theorem 1.1. For any 2 < p < ∞ there exists a constant c(p) > 0 with the following property: for any n-dimensional subspace X of Lp and for any ε ∈ (0, 1) there exists k ≥ c(p) min{ε2n, (εn)2/p} so that ℓk 2 can be (1 + ε)-embedded into X. Our approach is different and depends on a Gaussian functional analytic in- equality rather than the spherical isoperimetric inequality that is used in the classical framework. Thus, our proof still depends on random methods, but the main tool is a variant of an inequality due to Pisier from [26]. To prove the above theorem, we have to bypass Milman's asymptotic formula, which involves the Lipschitz constant of the norm. As several examples show this parameter is insufficient to describe efficiently phenomena in the almost isometric scale. Our argument outclasses the latter one since it takes into account the order 3 of magnitude of the length of the gradient of the norm. The idea of sufficiently estimating averages of the Euclidean norm of the gradient of a function in order to get sharp concentration results seems to be only recently applied and was also successfully exploited in [25]. Moreover, the selection of the position of the unit ball of the space is different. Instead of using John's position we employ Lewis' position (see Section 2 for details) for the unit ball of finite-dimensional subspaces of Lp. This permits us to express the norm in an integral form, with respect to some isotropic measure on the sphere, and therefore to use the aforementioned inequality. In fact we derive Theorem 1.1 from the randomized Dvoretzky theorem for those spaces in Lewis' position. For this end we prove that the norm of the underlying subspace in this position exhibits two-level Gaussian concentration and minimal fluctuations. Theorem 1.2. Let 2 < p < ∞ and let X be an n-dimensional subspace of Lp represented on Rn whose unit ball BX is in Lewis' position. Then, P(cid:16)(cid:12)(cid:12)(cid:12)kZk − EkZk(cid:12)(cid:12)(cid:12) In particular, we have > εEkZk(cid:17) ≤ C exp(cid:16)−c minnαpε2n, (εn)2/po(cid:17) , 0 < ε < 1. VarkZk ≤ C pn2/p−1, where αp, C p > 0 are constants depending only on p and Z is a standard n-dimensional Gaussian vector. It is worth mentioning that the Gaussian concentration and the variance estimate obtained for these spaces is best possible (up to constants of p) as the example of ℓp norms shows (see [25] for the exact formulation). Consequently, the random version of Dvoretzky's theorem we prove for this position (or for this type of norms) is sharp in the sense that in the case of ℓn p spaces the corresponding critical dimension is optimal (see [25]). In other words the ℓn p space occurs as the approximately extremal structure in this study or is the worst subspace of Lp with respect to the local almost Euclidean structure. The novelty of the work is not only observed in the techniques used but also in the content of the results. For our analysis is crucial the perspective of differently selecting the position of the unit ball of the underlying space and this is reflected in the improved estimates we obtain. To the best of our knowledge the concentration estimates we derive in Theorem 1.2 are new and it is also clear that the dimension k(n, p, ε) ≃p min{ε2n, (εn)2/p}, that one can find almost Euclidean subspaces, is always better than the previously known ε2n2/p due to Figiel, Lindestrauss and V. Milman. In addition, the improved estimate for k(n, p, ε) yields "new dimensions" of almost Euclidean sections in the following sense: The previous setting was only permitting almost isometric embeddings of distortion 1 + ε with ε ≫ n−1/p in order to achieve non-trivial dimensions. Now this phenomenon admits an improvement and one can find (1 + ε)-linear embeddings with ε ≫ n−1/2. It is worth mentioning that the dimension k(n, p, ε) that one finds almost Euclidean sections for these spaces is given implicitly as function of ε and n rather than as function of separated variables as V. Milman's formula suggests. This phenomenon had not been observed prior to this work and [25]. 4 The rest of the paper is organized as follows: In Section 2 we introduce the notation, some background material on isotropic measures on the (n− 1)-dimensional Euclidean sphere and finally we give the proof of the aforementioned Gaussian inequality. In Section 3 we prove concentration results for the family of the Lq- bodies associated with an isotropic measure µ on the (n − 1)-dimensional Euclidean sphere. In Section 4 we provide the proof of our main result. Finally, in Section 5 we conclude with some further remarks. 2 Background material and auxiliary results i=1 xip)1/p for x = (x1, . . . , xn) ∈ Rn. We set ℓn We work in Rn equipped with the standard Euclidean structure h·, ·i. The (n − 1)- dimensional Euclidean sphere is defined as S n−1 := {x ∈ Rn : hx, xi = 1}. The ℓp norm is defined as kxkp := (Pn = (Rn, k·kp) and let Bn p its unit ball. More generally, for any centrally symmetric convex body K on Rn we write k · kK for the norm induced by K. The n-dimensional Lebesgue measure (volume) of a body A is denoted by A. The space Lp(Ω, E, µ), 1 ≤ p < ∞ consists of all E-measurable functions f : Ω → R so that RΩ fp dµ < ∞, equipped with the norm k fkLp (µ) := (RΩ fp dµ)1/p. The n-dimensional (standard) Gaussian measure is denoted by γn and its density is p dγn(x) := (2π)−n/2e−kxk2 2/2dx. More generally, let dγn,σ(x) := (2πσ2)−n/2e−kxk2 2/(2σ2 )dx for σ > 0. Random vectors, usually distributed according to γn, are denoted by Z, W . . . while the random vari- ables by gi, ξ, . . .. The notation E(·) is used for the expectation. The moments with respect to γn of norms whose unit ball is the body K are denoted by Ir(γn, K) :=(cid:0)EkZkr K(cid:1)1/r = ZRn kzkr K dγn(z)!1/r and more generally, for an arbitrary probability measure ν as Ir(ν, K). Recall the pth moment σp of a standard Gaussian random variable g (2.1) σp p := Egp = 2p/2 √π Γ p + 1 2 ! ∼ p2/e p + 1 e !p/2 , p → ∞, where f ∼ h means f (t)/h(t) → 1 as t → ∞. We write f . h when there exists absolute constant C > 0 such that f ≤ Ch. We write f ≃ h if f . h and h . f , whereas the notation f .p h means that the involved constant depends only on p. The letters C, c, C1, c0, . . . are frequently used throughout the text in order to denote absolute constants which may differ from line to line. The random version of Dvoretzky's theorem due to V. Milman from [22] (for the optimal dependence on ε see [13] and [27]) reads as follows. 5 Theorem 2.1. Let X = (Rn, k · k) be a normed space. Define the critical dimension of X as the quantity k(X) := EkZk2 b2(X) , Z ∼ N(0, In) where b(X) := maxθ∈S n−1 kθk. Then, for every ε ∈ (0, 1) and for any k ≤ cε2k(X) the random (with respect to the Haar measure on the Grassmannian Gn,k) k-dimensional subspace F of X is (1 + ε)-spherical, i.e. 1 − ε M BF ⊆ BX ∩ F ⊆ 1 + ε M BF , with probability greater than 1 − e−ck(X), where M = M(X) =RS n−1 kθk dσ(θ) and σ is the uniform probability measure on S n−1. 2.1 Logarithmic Sobolev inequality Let ν be a Borel probability measure on Rn which satisfies a log-Sobolev inequality with constant ρ > 0 Entν( f 2) :=Z f 2 log f 2 dν −Z f 2 dν log Z f 2 dν! ≤ 2 ρZRn k∇ fk2 2 dν, : Rn → R. The n-dimensional for all smooth (or locally Lipschitz) functions f Gaussian measure satisfies the log-Sobolev inequality with ρ = 1 (see [18]). The next lemma is essentially from [34] (see also [25]). We provide a sketch of proof for reader's convenience. Lemma 2.2. Let ν be a Borel probability measure on Rn which satisfies a log-Sobolev inequality with constant ρ. Then, for any smooth function f : Rn → R we have (2.2) k fk2 Lq (ν) − k fk2 Lp (ν) ≤ 2 Ls(ν) for all 2 ≤ p ≤ q. Moreover, if f is Lipschitz continuous, then we have 1 ρZ q p (cid:13)(cid:13)(cid:13) k∇ fk2(cid:13)(cid:13)(cid:13) ds, k fk2 Lq (ν) − k fk2 Lp (ν) ≤ k fk2 Lip ρ (q − p). In particular, we obtain k fkLq (ν) k fkL2 (ν) ≤ s1 + L2 (ν)/k fk2 Lip. for q ≥ 2, where k( f ) := k fk2 q − 2 ρk( f ) , 6 Sketch of Proof. For p ≥ 2 we define I(p) := k fkLp . Differentiation with respect to p yields dI d p = Entν( fp) p2I(p)p−1 . Applying the log-Sobolev inequality for g = fp/2 we obtain dI d p ≤ 1 2ρI(p)p−1 ZRn fp−2k∇ fk2 2 dν ≤ 1 2ρI(p)p−1 by Hölder's inequality. This shows that (I(p)2)′ ≤ 1 the interval [p, q] proves (2.2). 2 , Lp (ν) I(p)p−2(cid:13)(cid:13)(cid:13)k∇ fk2(cid:13)(cid:13)(cid:13) 2 Lp(ν) , thus integration over (cid:3) ρ(cid:13)(cid:13)(cid:13)k∇ fk2(cid:13)(cid:13)(cid:13) 2.2 Lewis' position Given any finite Borel measure µ on S n−1 (which is not supported in any hyperplane) and any 1 ≤ p < ∞ we can equip Rn with the norm . kxkµ,p := ZS n−1 hx, θip dµ(θ)!1/p It's clear that the space X = (Rn, k · kµ,p) can be naturally embedded into Lp(S n−1, µ) via the linear isometry U : X → Lp(S n−1, µ) with U x := hx, ·i. Lewis' fundamental result from [19], states that the previous situation can always be realized for finite-dimensional subspaces of Lp(ν) after a suitable change of the density ν (see also [31] for an alternative proof which extends to the whole range 0 < p < ∞ and arises as a solution of an optimization problem). The formulation we use here follows the exposition from [20]. Theorem 2.3 (Lewis). Let 1 ≤ p < ∞ and let X be an n-dimensional subspace of Lp. Then, there exists an even Borel measure µ on S n−1 which satisfies (2.3) kxk2 2 =ZS n−1 hx, θi2dµ(θ), for all x ∈ Rn and the normed space (Rn, k · kµ,p) is isometric to X. It is clear that taking into account this representation for any finite-dimensional subspace of Lp, the problem of embedding ℓk 2 in subspaces of Lp reduces to spaces (Rn, k·kµ,p) with µ satisfying the condition (2.3). Hence, the next paragraph is devoted to the study of these measures. 2.3 Isotropic measures on the sphere An even Borel measure µ on S n−1 is said to be isotropic if it satisfies the following condition: kxk2 2 =ZS n−1 hx, θi2 dµ(θ), 7 for all x ∈ Rn. Equivalently, for all linear transformations T : Rn → Rn we have trace(T ) =ZS n−1hθ, T θi dµ(θ). For any such measure we may define the following family of centrally symmetric convex bodies Bp(µ) associated with µ and corresponding norms: x 7→ kxkBp(µ) := khx, ·ikLp (µ) = ZS n−1 hx, zip dµ(z)!1/p , 1 ≤ p < ∞. The corresponding spaces, whose unit ball is Bp(µ), will be denoted by Xp(µ). Under this terminology and notation, Lewis' theorem reads as follows: Theorem 2.4 (Lewis). Let 1 ≤ p < ∞ and let X be an n-dimensional subspace of Lp. Then, there exists an isotropic Borel measure µ on S n−1 and a linear isometry U : Xp(µ) → X. The next simple lemma collects several properties for the bodies Bp(µ). Lemma 2.5. Let µ be a Borel isotropic measure on S n−1 and let Z be an n-dimensional standard Gaussian vector. Then, we have the following properties: Bq(µ) = σq qµ(S n−1), for 0 < q < ∞. i. EkZkq ii. µ(S n−1) = n. iii. For p ≥ 2 we have kxkBp(µ) ≤ kxk2 and for 1 ≤ p < q < ∞ we have kxkBp(µ) ≤ n1/p−1/qkxkBq(µ), for all x ∈ Rn. (K. Ball) For every 1 ≤ p < ∞ we have Bp(µ) ≤ Bn p. iv. v. For the body Bq(µ), q ≥ 1 we have k(Bq(µ)) ≥ cnmin{1,2/q}. vi. There exists an absolute constant c > 0 such that for all 2 ≤ q ≤ c log n, one has Bq(µ))1/2 ≃ q1/2n1/q. In particular, for those q's one has k(Bq(µ)) ≥ cqn2/q. (EkZk2 (i) We use Fubini's theorem and the rotation invariance of the Gaussian Proof. measure to write EkZkq Bq(µ) =ZRn kxkq Bq(µ) dγn(x) =ZS n−1ZRn hx, θiq dγn(x) dµ(θ) = σq qµ(S n−1). (ii) It follows from the above formula applied for q = 2 and by employing the isotropic condition. (iii) Let p ≥ 2. Note that for all u ∈ S n−1 we have kukp Bp(µ) =ZS n−1 hu, θip dµ(θ) ≤ZS n−1 hu, θi2 dµ(θ) = 1. 8 For 1 ≤ p ≤ q we apply Hölder's inequality kxkBp(µ) = ZS n−1 hx, θip dµ(θ)!1/p ≤ µ(S n−1) 1 p − 1 q ZS n−1 hx, θiq dµ(θ)!1/q . (iv) This result was essentially proved by K. Ball in [4]. A sketch of his very elegant proof is reproduced below for the sake of completeness. Without loss of generality i=1 ciδui, for some vectors (ui) in S n−1 and i=1 ciui ⊗ ui. Now we use the formula, which holds true for any centrally symmetric convex body K on Rn, we may assume that µ is discrete, i.e. µ =Pm positive numbers (ci) with I = Pm K = (Γ(1 + n/p))−1ZRn e−kzkp K dz, to get Bp(µ) = 1 Γ(1 + n p )ZRn m Yi=1 fi(hz, uii)ci dz, where fi(t) = exp(−tp). The result follows by the Brascamp-Lieb inequality. (v) First consider the case 2 < q < ∞. For the critical dimension of the space Bq(µ)/b2(Bq(µ)) ≥ n2/q by the third Xq(µ) = (Rn, k · kBq(µ)), note that k(Xq(µ)) = EkZk2 assertion. Now we turn in the range 1 ≤ q ≤ 2. Using Hölder's inequality we may write (cid:16)EkZk2 Bq(µ)(cid:17)1/2 ≥ n1/2 Bn 2 Bq(µ)!1/n q!1/n ≥ n1/2 Bn 2 Bn ≃ n1/q, where in the last step we have used Ball's volumetric estimate (iv). The result follows once we recall that b(Bq(µ)) ≤ n1/q−1/2 for 1 ≤ q ≤ 2. (vi) We define the parameter q0 ≡ q0(µ) := maxnq ∈ [2, n] : k(Bp(µ)) ≥ p, ∀p ∈ [2, q]o . By the continuity of the map p 7→ k(Bp(µ)) and the fact that k(Bq(µ)) ≤ n for all q ≥ 2, Bq0 (µ)(cid:19)1/q0 while k(B2(µ)) = n we get q0 = k(Bq0 (µ)). Lemma 2.2 shows that (cid:18)EkZkq0 ≤ c1(cid:18)EkZk2 Bq0 (µ)(cid:19)1/2 , so we may write q0 = k(Bq0(µ)) = EkZk2 b2(Bq0 (µ)) ≥ c−2 Bq0 (µ) 1 (EkZkq0 Bq0 (µ))2/q0 = c−2 1 σ2 q0 n2/q0 =⇒ q0 ≥ c2 log n. Therefore, by the definition of q0 we have k(Bq(µ)) ≥ q for all 2 ≤ q ≤ q0 and by Lemma 2.2 again, we get σqn1/q =(cid:16)EkZkq Bq(µ)(cid:17)1/q ≤ c1(cid:16)Ekgk2 Bq(µ)(cid:17)1/2 . 9 Moreover, we have Bq(µ) EkZk2 b2(Bq(µ)) ≥ c−2 k(Bq(µ)) = 1 (EkZkq This can be interpreted as k(Bq(µ)) ≥ ck(ℓn absolute constant c > 0. For a proof of the fact that k(ℓn the reader is referred to [32]. Bq(µ))2/q = c−2 q), provided that 2 ≤ q ≤ c log n for some q) ≃ qn2/q when 2 ≤ q ≤ c log n (cid:3) 1 σ2 qn2/q ≥ c3qn2/q. Lemma 2.6. Let µ be a Borel isotropic measure on S n−1. For q ≥ 2 and for all r ≥ 1 we have Irq(γn, Bq(µ))/Iq(γn, Bq(µ)) ≤ s1 + qn2/q ≤ r1 + q(r − 1) σ2 c(r − 1) n2/q , where c > 0 is an absolute constant. Hence, if we use Lemma 2.2 we obtain Proof. Note that Lemma 2.5 (iii) implies (cid:12)(cid:12)(cid:12)kxkBq(µ) − kykBq(µ)(cid:12)(cid:12)(cid:12) ≤ kx− yk2 for all x, y ∈ Rn. Iq !2 Irq ≤ 1 + q(r − 1) I2 q = 1 + q(r − 1) σ2 qn2/q , where the last estimate follows from Lemma 2.5. Finally, using the fact that σq ≃ √q we conclude the second estimate. (cid:3) 2.4 A Gaussian inequality The next inequality is due to Pisier (for a proof see [26]). Theorem 2.7. Let φ : R → R be a convex function and let f : Rn → R be C1-smooth. Then, if Z, W are independent copies of a Gaussian random vector, we have Eφ ( f (Z) − f (W)) ≤ Eφ(cid:18) π 2h∇ f (Z), Wi(cid:19) . Here we prove a generalization of this inequality in the context of Gaussian processes generated by the action of a random matrix with i.i.d standard Gaussian entries on a fixed vector in S n−1. The next inequality was stated in [25] without a proof. Below we give the details for reader's convenience. Theorem 2.8. Let φ : R → R be a convex function and let f : Rn → R be C1-smooth. If G = (gi j)n,k i, j=1 is a Gaussian matrix and a, b ∈ S k−1, then we have 2ka − bk2h∇ f (Z), Wi(cid:19) , Eφ ( f (Ga) − f (Gb)) ≤ Eφ(cid:18) π where Z, W are independent copies of a standard Gaussian n-dimensional random vector. 10 Proof. If a = b then, there is nothing to prove. If a = −b then, by setting F(z) = f (z) − f (−z) we may write: Eφ ( f (Ga) − f (Gb)) = Eφ(F(Z)) ≤ Eφ(F(Z) − F(W)), for Z, W independent copies of a standard Gaussian random vector, where we have used the fact EF(Z) = 0 and Jensen's inequality. Then, a direct application of Theorem 2.7 yields: Eφ(F(Z) − F(W)) ≤ Eφ πh∇ f (Z), Wi + πh∇ f (−Z), Wi ! φ(πh∇ f (Z), Wi) + φ(πh∇ f (−Z), Wi) 2 ≤ E = Eφ(πh∇ f (Z), Wi), 2 by the convexity of φ. In the general case, fix a, b ∈ S k−1 with a , ±b and define p := a+b 2 . Note that since kak2 = kbk2 we have that the vector u := a − p is perpendicular to p. Set W := G(u) and Z := G(p) and note that W, Z are independent random vectors in Rn with W ∼ N(0, kuk2 2In). Since G(a) = Z + W and G(b) = Z − W, we may write: 2In), Z ∼ N(0, kpk2 Eφ ( f (Ga) − f (Gb)) = EZ EW φ ( f (Z + W) − f (Z − W)) . Denote F(w, z) := f (z + w) − f (z − w). Then, we may write: Eφ( f (Ga) − f (Gb)) = " φ(F(w, z)) dγn,σ1(w) dγn,σ2 (z), where σ1 = kuk2 > 0, σ2 = kpk2 > 0. For fixed z, we may apply Theorem 2.7 to the function w 7→ F(w, z) (note that R F(w, z) dγn,σ1(w) = 0) to get: Z φ(F(w, z)) dγn,σ1(w) ≤ " φ(cid:18) π 2h∇wF(w, z), yi(cid:19) dγn,σ1 (w) dγn,σ1 (y) ≤ " φ(πh∇ f (w + z), yi) + φ(πh∇ f (z − w), yi) = " φ (πh∇ f (w + z), yi) dγn,σ1 (w) dγn,σ1 (y), 2 dγn,σ1 (w) dγn,σ1 (y) by the convexity of φ. Integration with respect to γn,σ2 over z provides: " φ(F(w, z)) dγn,σ1(w)dγn,σ2 (z) ≤Z "" φ (πh∇ f (w + z), yi) dγn,σ1 (w)dγn,σ2 (z)# dγn,σ1 (y) =Z "Z φ (πh∇ f (x), yi) d(γn,σ1 ∗ γn,σ2 )(x)# dγn,σ1 (y) = " φ (πσ1h∇ f (x), yi) dγn(x) dγn(y), 2 ≡ γn, since σ2 = kak2 = γn,σ2 + σ2 2 2 +σ2 1 1 = 1. (cid:3) where we have used the fact that γn,σ1 ∗ γn,σ2 The result follows. 11 Remark 2.9. 1. Applying this for φ(t) = tr, r ≥ 1 and taking into account the invariance of the Gaussian measure under orthogonal transformations we derive the next (r, r)-Poincaré inequalities: (E f (Ga) − f (Gb)r)1/r ≤ C √rka − bk2(cid:0)Ek∇ f (Z)kr 2(cid:1)1/r , (2.4) for a, b ∈ S k−1, where Z is a standard Gaussian random vector in Rn. 2. Assuming further that eλt, λ > 0 to get: f is L-Lipschitz we may apply Theorem 2.8 for φ(t) = (2.5) E exp (λ ( f (Ga) − f (Gb))) ≤ E exp λ2 π2 2 ka − bk2 2k∇ f (Z)k2 2! ≤ exp λ2 π2 2 ka − bk2 2L2! . Then Markov's inequality yields Schechtman's distributional inequality from [27]: (2.6) P ( f (Ga) − f (Gb) > t) ≤ C exp(cid:16)−ct2/(ka − bk2L2)(cid:17) , for all t > 0, where a, b ∈ S k−1. Let us note that (2.5) for f being a norm, has also appeared in [33]. 3. For a, b ∈ S k−1 with ha, bi = 0 the matrix G generates the vectors Z = Ga and W = Gb which are independent copies of a standard n-dimensional Gaussian random vector. For example, inequality (2.6) reduces to the classical concentration inequality: (2.7) for all t > 0. P ( f (Z) − f (W) > t) ≤ C exp(cid:16)−ct2/L2(cid:17) , 3 Gaussian concentration for Bp(µ) norms P(cid:16)(cid:12)(cid:12)(cid:12)kZkBp(µ) − I1(cid:12)(cid:12)(cid:12) > tI1(cid:17) ≤ C exp(−ct2I2 A direct application of the Gaussian concentration inequality (2.7) for the norms k · kBp(µ), 2 < p < ∞ implies: (3.1) 1 ) ≤ C exp(−ct2n2/p), for all t > 0, where I1 ≡ I1(γn, Bp(µ)). It is known (see [25]) that the large deviation estimate (t ≥ 1) the inequality (3.1) provides is sharp (up to constants). In this paragraph we prove that for 2 < p < ∞ and µ isotropic Borel measure on S n−1, the bodies Bp(µ) exhibit better concentration (0 < t < 1) than the one implied by the Gaussian concentration inequality on Rn in terms of the Lipschitz constant. Later, this will be used to prove the announced dependence on ε and n in Dvoretzky's theorem for any n-dimensional subspace of Lp. Our main tool is the probabilistic inequality proved in Theorem 2.8 and as was formulated further in Remark 2.9.1. We apply inequality (2.4) for f (x) = kxkp = R hx, θip dµ(θ). To this end we have to compute the gradient. Note that k∇ f (x)k2 2 = p2 n ZS n−1 Xi=1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2 θihx, θip−1 sgn(hx, θi) dµ(θ)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . 12 We also have the following: Claim. For almost every x ∈ Rn we have k∇ f (x)k2 2 ≤ p2kxk2p−2 B2p−2(µ). may write Proof of Claim. Let bi ≡ bi(x) := RS n−1 hx, zip−1 sgn(hx, zi)zi dµ(z). Using duality we Xi=1 ZS n−1 hx, zip−1 sgn(hx, zi)zi dµ(z)!2 2 n n = max θ∈S n−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) biθi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xi=1 ZS n−1 hx, zip−1 sgn(hx, zi)hz, θi dµ(z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) θ∈S n−1(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ZS n−1 hx, zi2p−2 dµ(z), = max 2 where we have used the Cauchy-Schwarz inequality and the isotropic condition. Therefore, using the Claim and the inequality (2.4) we get for every a, b ∈ S k−1 (cid:3) , B2p−2(µ)(cid:17)1/r p−1 2 , p−1 2 1 p−1 (r − 2)(p − 1) σ2 2p−2n  p−1 2p−2n1/2  r p/2(p − 1) σp−1 p−1 2 for all r ≥ 1. By employing Lemma 2.6 we find 1 + (E f (Ga) − f (Gb)r)1/r ≤ C pr1/2ka − bk2(cid:16)EkZkr(p−1) B2p−2(µ)(cid:17)1/2 (E f (Ga) − f (Gb)r)1/r ≤ C pr1/2ka − bk2(cid:16)EkZk2p−2 2p−2n1/2 2 max 2p−2n1/2  α(n, p, r) := max < C pr1/2ka − bk2σp−1 < C pka − bk2σp−1 r p/2(p − 1) σp−1 2p−2n1/22 r1/2, r1/2, 1 + p−1 2 p−1 , r(p − 1) σ2 2p−2n 1 r > 0 for all r ≥ 2. We define (3.2) and we summarize the above discussion to the following: Proposition 3.1. Let 2 < p < ∞ and let µ be a Borel isotropic measure on S n−1. If G = (gi j)n,k i, j=1 is a Gaussian matrix and a, b ∈ S k−1, then we have 2p−2n1/22 ≤ C pka − bk2σp−1 Bp(µ) − kGbkp p 2 α(n, p, r), (cid:16)E(cid:12)(cid:12)(cid:12)kGakp r(cid:17)1/r Bp(µ)(cid:12)(cid:12)(cid:12) for all r ≥ 2, where α(n, p, ·) is defined in (3.2) 13 We are now ready to prove the main result of this section. Theorem 3.2. Let 2 < p < ∞ and let µ be a Borel isotropic measure on S n−1 with n > ep. Then, we have P(cid:16)(cid:12)(cid:12)(cid:12)kZkBp(µ) − (EkZkp Bp(µ))1/p(cid:12)(cid:12)(cid:12) ≥ ε(EkZkp for every ε > 0, where ψ(n, p, ·) is defined by ψ(n, p, t) := min( t2n Bp(µ))1/p(cid:17) ≤ C exp (−cψ(n, p, ε)) , , (tn)2/p) , t > 0, (3.3) p4p and C, c > 0 are absolute constants. Proof. Using Proposition 3.1 for a, b ∈ S k−1 with ha, bi = 0 and applying Jensen's inequality we obtain ≤ C pσp−1 for all r ≥ 2. Therefore Markov's inequality yields C pσp−1 Bp(µ) − EkZkp Bp(µ) − EkZkp r(cid:17)1/r (cid:16)E(cid:12)(cid:12)(cid:12)kZkp Bp(µ)(cid:12)(cid:12)(cid:12) > ε(cid:17) ≤ P(cid:16)(cid:12)(cid:12)(cid:12)kZkp Bp(µ)(cid:12)(cid:12)(cid:12) α−1(n, p, s) = min s2, Note that the inverse of the map r 7→ α(n, p, r) is given by , s > 0, 2p−2 p 2p−2 p−1 p s2/pn1/pσ (p − 1)  2p−2n1/22p/2α(n, p, r), 2p−2n1/22p/2α(n, p, r) ε . r  thus we may choose rε ≥ 2 such that α(n, p, rε) = 2p−2 n1/22p/2 , as long as the range of ε > 0 satisfies α(n, p, rε) ≥ α(n, p, 2). Otherwise α(n, p, rε) < α(n, p, 2) ≃ max{1, (ep/n)1/2} ≃ 1 provided that n is large enough with respect to p. Hence, we get ε eC pσp−1 which implies n, p, ε eC pσp−1 ε2 np22pσ2p−2 2p−2 , 2p−2n1/22p/2  p  ε2/p , , ε2 np22pσ2p−2 2p−2 , , ε2/p p   for all ε > 0. We may check that n, p, ε eC pσp−1 Bp(µ) − EkZkp P(cid:16)(cid:12)(cid:12)(cid:12)kZkp α−1 P(cid:16)(cid:12)(cid:12)(cid:12)kZkp > ε(cid:17) ≤ C1 exp −α−1 Bp(µ)(cid:12)(cid:12)(cid:12) ≃ min 2p−2n1/22p/2 −c1 min > ε(cid:17) ≤ C1 exp Bp(µ) − EkZkp Bp(µ)(cid:12)(cid:12)(cid:12) 14 for every ε > 0. It follows that P(cid:16)(cid:12)(cid:12)(cid:12)kZkp Bp(µ) − EkZkp Bp(µ)(cid:12)(cid:12)(cid:12) > εEkZkp Bp(µ)(cid:17) ≤ C1 exp −c1 min ε2nσ2p p p22pσ2p−2 2p−2 , for every ε > 0. The asymptotic estimate (2.1) yields σ2p thus we conclude p /σ2p−2 2p−2 ≃ p2−p and σp ≃ p1/2, ,  p (εn)2/pσ2 p  , (εn)2/p)! , Bp(µ)(cid:17) ≤ C1 exp −c′1 min( ε2n p4p > εEkZkp (3.4) for all ε > 0. This further implies that P(cid:16)(cid:12)(cid:12)(cid:12)kZkp P(cid:18)(cid:12)(cid:12)(cid:12)kZkBp(µ) −(cid:16)EkZkp Bp(µ) − EkZkp Bp(µ)(cid:12)(cid:12)(cid:12) Bp(µ)(cid:17)1/p(cid:12)(cid:12)(cid:12) > ε(cid:16)EkZkp for all ε > 0. In order to verify the latter we may write Bp(µ)(cid:17)1/p(cid:19) ≤ 2C1 exp −c′1 min( ε2n p4p , (εn)2/p)! , P(cid:18)kZkBp(µ) > (1 + ε)(cid:16)EkZkp Bp(µ)(cid:17)1/p(cid:19) ≤ P(cid:16)kZkp Bp(µ) > (1 + ε)EkZkp ≤ C1 exp −c′1 min( ε2n p4p Bp(µ)(cid:17) , (εn)2/p)! , for all ε > 0 by the estimate (3.4). We argue similarly for the other case. (cid:3) Remark 3.3. By the well known symmetrization argument for any random variable ξ P(ξ − med(ξ) > t) ≤ 4 inf α∈R P(ξ − α > t/2), Bp(µ))1/p by a median of x 7→ kxkBp(µ) we may replace (EkZkp EkXkBp(µ)) with respect to the Gaussian measure γn (see also [23, Appendix V]). (or the expected value t > 0, We shall also need the next variant of Theorem 3.2. Theorem 3.4. Let 2 < p < ∞ and let µ be a Borel isotropic probability measure on S n−1 with n > ep. If G = (gi j)n,k i, j=1 is a Gaussian matrix and a, b ∈ S k−1, then ka − bk2!! , Bp(µ)(cid:17) ≤ C exp −cψ n, p, > tEkZkp t P(cid:16)(cid:12)(cid:12)(cid:12)kGakp Bp(µ) − kGbkp Bp(µ)(cid:12)(cid:12)(cid:12) for all t > 0, where ψ(n, p, ·) is defined in (3.3). Proof. The proof is similar to the proof of Theorem 3.2. We omit the details. (cid:3) The following estimate on the fluctuations of the norm x 7→ kxkBp(µ) is immediate: 15 Corollary 3.5 (Gaussian variance for Bp(µ)). Let µ be an isotropic Borel measure on S n−1 and let 1 ≤ p < ∞. Then, VarkZkBp(µ) ≤ ecpn2/p−1. In particular, we have VarkZkBp(µ) EkZk2 where Z is a standard Gaussian vector. Bp(µ) ≤ ecp n , Proof. We distinguish two cases. For 1 ≤ p ≤ 2 we may bound as follows: VarkZkBp(µ) . b2(Bp(µ)) . n2/p−1, where we have used Lemma 2.5. For 2 < p < ∞ consider Z′ an independet copy of Z to write Bp(µ) − kZ′kp kZkp min{kZkp−1 Bp(µ), kZ′kp−1 E 2 , Bp(µ) Bp(µ)} 2VarkZkBp(µ) = E(kZkBp(µ) − kZ′kBp(µ))2 ≤ 1 p2 where we have used the numerical inequality ap − bp ≥ pa − b min{ap−1, bp−1} for a, b > 0 and p > 1. The Cauchy-Schwarz inequality implies VarkZkBp(µ) ≤ 1 p2(cid:18)E(cid:12)(cid:12)(cid:12)(cid:12)kZkp Bp(µ) − kZ′kp I2(p−1) −4(p−1)(γn, Bp(µ)) 4(cid:19)1/2 Bp(µ)(cid:12)(cid:12)(cid:12)(cid:12) . The numerator is directly estimated by Proposition 3.1. For the denominator we employ the main result of [16] along with the fact k(Bp(µ)) ≥ c1 pn2/p for n ≥ eC1 p (proved in Lemma 2.5.vi) to obtain I−4(p−1)(γn, Bp(µ)) ≥ c2Ip(γn, Bp(µ)) = σpn1/p, by Lemma 2.5.i. Combining all the above we arrive at the desired estimate. The details are left to the reader. (cid:3) 4 Embedding ℓk 2 into subspaces of Lp for 2 < p < ∞ In this paragraph we prove the improved estimate on Dvoretzky's theorem for the subspaces of Lp, 2 < p < ∞. Theorem 4.1. Let 2 < p < ∞. Then for every n-dimensional subspace X of Lp and any 0 < ε < 1 there exists k ≥ cpψ(n, p, ε) and linear map T : ℓk 2 → X such that kxk2 ≤ kT xkX ≤ (1 + ε)kxk2 for all x ∈ ℓk 2, where cp > 0 is constant depending only on p and ψ(n, p, ·) is given by (3.3). 16 Clearly, the above theorem follows from the random version of Dvoretzky's theo- rem for subspaces of Lp whose unit ball is in Lewis' position, or equivalently for the bodies Bp(µ) with µ isotropic measure on S n−1. More precisely, we have the following: Theorem 4.2. Let 2 < p < ∞ and let X be an n-dimensional subspace of Lp represented on Rn whose unit ball BX is in Lewis' position. Then, for any ε ∈ (0, 1) there exists k ≥ cpψ(n, p, ε) such that the random k-dimensional subspace F of X is (1 + ε)-spherical with probability greater than 1−e−cpψ(n,p,ε), where cp > 0 depends only on p and ψ(n, p, ·) is defined in (3.3). Let us note that once we have established the concentration estimate of Theorem 3.2, then a standard net argument yields the result with an extra log(1/ε) term. Indeed; fix k ≤ n and let G = (gi j)n,k i, j=1 be a Gaussian matrix with independent standard entries. Let D be a δ-net on S k−1 with cardinality D ≤ (3/δ)k (see [23, Lemma 2.6] for the details). Then using the union bound, Theorem 3.2 and the fact that Gu is equidistributed to Z ∼ N(0, In) we obtain P(cid:16)∃ u ∈ D :(cid:12)(cid:12)(cid:12)kGukBp(µ) − EkZkBp(µ)(cid:12)(cid:12)(cid:12) ≥ εEkZkBp(µ)(cid:17) ≤ (3/δ)kC exp (−cψ(n, p, ε)) . Choosing δ ≃ ε we find that with probability greater than 1 − e−c′ψ(n,p,ε) the random operator G satisfies for all u ∈ D, as long as k . (log 1 to the whole sphere S k−1 at cost of an oscillation at most 2εEkZkBp(µ), see e.g. Lemma 4.1]. ε )−1ψ(n, p, ε). It's routine to check that we may pass [23, (cid:12)(cid:12)(cid:12)kGukBp(µ) − EkZkBp(µ)(cid:12)(cid:12)(cid:12) ≤ εEkZkBp(µ), Theorem 3.4 serves exactly the purpose of removing this term. Then we use this inequality along with a chaining method to conclude the logarithmic-free dependence on ε in our main result. This approach has been inspired by [27]. However, the method from [27] is not directly applicable here, since it lies in estimates involving the Lipschitz constant. As we have already explained such estimates would only yield suboptimal bounds and one has to keep track of the higher moments of the length of the gradient until the very last step. This forces us to establish the inequality in Bp(µ) − Theorem 2.8. In probabilistic terms Theorem 3.4 says that the process (kGθkp I p p )θ∈S k−1 has two-level tail behavior described by ψ(n, p, ·). Now we turn to proving the main result. Let 2 < p < ∞ and let X be an n-dimensional subspace Proof of Theorem 4.2. of Lp whose unit ball is in Lewis' position. Then, Lewis' theorem (Theorem 2.4) yields the existence of an isotropic Borel measure µ on S n−1 and a linear isometry S : Xp(µ) → X, hence we may identify X with Xp(µ). We have to show that the ball Bp(µ) has random almost spherical k-dimensional sections with k as large as possible. Let {gi j(ω)}n,k i, j=1 be i.i.d. standard normals in some probability space (Ω, P) and consider the random Gaussian operator Gω = (gi j(ω))n,k 2 → Xp(µ). We will prove that with overwhelming probability the operator G is (1 + ε)-isomorphic embedding when k is sufficiently large. Toward this end, we employ Theorem 3.4 i, j=1 : ℓk 17 and the chaining argument from [27]. For each j = 1, 2, . . . consider δ j-nets N j on S k−1 with cardinality N j ≤ (3/δ j)k (see [23, Lemma 2.6]). Note that for any θ ∈ S k−1 and for all j there exist u j ∈ N j with kθ − u jk2 ≤ δ j and by the triangle inequality it follows that ku j − u j−1k2 ≤ δ j + δ j−1. Moreover, if we assume that δ j → 0 as j → ∞ and (t j) is a sequence of numbers with t j ≥ 0 and P j t j ≤ 1 then, for any ε > 0 we have the next claim. Claim. If we define the following sets: > εI p > t1εI p p(cid:12)(cid:12)(cid:12) p(cid:12)(cid:12)(cid:12) and for j ≥ 2 Bp(µ) − I p Bp(µ) − I p A :=nω ∃θ ∈ S k−1 :(cid:12)(cid:12)(cid:12)kGω(θ)kp A1 :=nω ∃u1 ∈ N1 :(cid:12)(cid:12)(cid:12)kGω(u1)kp A j :=(cid:26)ω ∃u j ∈ N j, u j−1 ∈ N j−1 :(cid:12)(cid:12)(cid:12)(cid:12)kGω(u j)kp po , po Bp(µ)(cid:12)(cid:12)(cid:12)(cid:12) Bp(µ) − kGω(u j−1)kp where Ip ≡ Ip(γn, Bp(µ)), then the inclusion A ⊆S∞j=1 A j holds. Proof of Claim. If ω <S∞j=1 A j then for any j and any u j ∈ N j we have Bp(µ)(cid:12)(cid:12)(cid:12)(cid:12) ≤ εt jI p (cid:12)(cid:12)(cid:12)kGω(u1)kp Bp(µ) − kGω(u j−1)kp (cid:12)(cid:12)(cid:12)(cid:12)kGω(u j)kp p(cid:12)(cid:12)(cid:12) ≤ εt1I p Bp(µ) − I p and p For any θ there exist u j ∈ N j such that kθ − u jk2 < δ j for j = 1, 2, . . .. Hence, for any N ≥ 2 we may write > t jεI p p(cid:27) , p , j = 2, 3, . . . . Bp(µ) − I p p − kGω(u1)kp I p (cid:12)(cid:12)(cid:12)kGω(θ)kp N + Bp(µ)(cid:12)(cid:12)(cid:12) p(cid:12)(cid:12)(cid:12) ≤(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)kGω(uN)kp Xj=1 εt jI p ≤ N p Xj=2(cid:12)(cid:12)(cid:12)(cid:12)kGω(u j−1)kp Bp(µ)(cid:12)(cid:12)(cid:12) 2→Xp (µ), Bp(µ) − kGω(θ)kp + p · δN · kGωkp Bp(µ) − kGω(u j)kp + Bp(µ)(cid:12)(cid:12)(cid:12)(cid:12) which proves the assertion, since N is arbitrary. Fix 0 < ε < 1. Choose δ j = e− j, t j = jp/2e− j/ap with ap := P∞j=1 jp/2e− j (thus, 18 ∞ ∞ ∞ P(A) ≤ CN1 exp(−c1ψ(n, p, εt1)) + C P j t j ≤ 1). Then, if we employ the previous claim and Theorem 3.4 we may write Xj=2 N j−1 · N j exp(−c1ψ(n, p, εt je j/4)) (3e j)2k exp(cid:16)−c′1ψ(cid:16)n, p, ε jp/2a−1 p (cid:17)(cid:17) p ψ(cid:16)n, p, ε jp/2(cid:17)(cid:17) exp(cid:16)c2 jk − c′2a−2 exp(cid:16)c2 jk − c′2a−2 p jψ(n, p, ε)(cid:17) exp(cid:16)−c3a−2 p jψ(n, p, ε)(cid:17) ≤ C′ exp(cid:16)−c′3a−2 Xj=1 Xj=1 Xj=1 Xj=1 p ψ(n, p, ε)(cid:17) , ≤ C ≤ C ∞ ∞ ≤ C ≤ C as long as k . a−2 than 1 − e−ca−2 p ψ(n, p, ε) . (2e/p)pψ(n, p, ε). Therefore, with probability greater p ψ(n,p,ε) the random operator G satisfies (1 − ε)1/pIpkxk2 ≤ kG(x)kBp(µ) ≤ (1 + ε)1/pIpkxk2, for all x ∈ Rk. To conclude we have to recall that ImG = F is Haar-distributed on Gn,k (see [29] for the details). (cid:3) Note 4.3. Let us mention that if n− p−2 2(p−1) .p ε < 1 then we get k ≤ c(εn)2/p/p by taking into account the form of ψ(n, p, ·). Indeed; for n large enough, i.e. n ≥ eC p log p, if we consider (cp)p/2n− p−2 2(p−1) < ε < 1 in the previous series of inequalities we obtain ∞ P(A) ≤ C Xj=1 (3e j)2k exp(cid:16)−c′1ψ(cid:16)n, p, ε jp/2a−1 ∞ (cid:17) p ∞ Xj=1 Xj=1 exp(cid:16)c2 jk − c4 j(εn)2/pa−2/p exp(cid:16)c2 jk − c5 jp−1(εn)2/p(cid:17) p (cid:17)(cid:17) ≤ C ≤ C ≤ C′ exp(cid:16)−c′5 p−1(εn)2/p(cid:17) , provided that k ≤ c′ p−1(εn)2/p. 5 Further remarks 1. Optimality of the result. If the isotropic measure µ on S n−1 is the one supported on ±ei's i.e. Xp(µ) ≡ ℓn p, then Theorem 3.2 is optimal (up to constants depending on p) as was proved in [25]. Moreover, Theorem 4.2 is optimal, in the sense that if the p is (1 + ε)-spherical, then k ≤ C p(εn)2/p for some typical k-dimensional subspace of ℓn 19 absolute constant C > 0 (see [25]). We should mention that it is known, that for concrete values of p one can embed ℓk p even isometrically (see [17] for details). However, this is not a typical subspace. 2 into ℓn 2. Selection of randomness. Embeddings of ℓk 2 into Lq, 2 < q < ∞ under different randomness have appeared in the literature in [5]. The authors there consider large random matrices with independent Rademacher entries in order to K(q)-embed ℓk 2 q with N ≃ kq/2, where K(q) > 0 depends only on q. Then, they use this into ℓN result in order to prove that for any 1 < p < 2 there exists uncomplemented subspace of Lp which is isomorphic to Hilbert space. It is worth mentioning, that one can prove a concentration result similar to that of Theorem 3.2 using other randomness than Gaussian. In particular, if ν is an isotropic2 Borel probability measure on Rn which satisfies a log-Sobolev inequality with constant ρ > 0 then we may prove the following: Theorem 5.1. Let 2 < p < ∞, let µ be a Borel isotropic measure on S n−1 and let ν be an isotropic Borel probability measure on Rn which satisfies a log-Sobolev inequality with constant ρ > 0. Then, we have " (cid:12)(cid:12)(cid:12)kxkp p (ν, Bp(µ)) max((cid:18) r n(cid:19)1/2 for all r ≥ 2, where C(p, ρ) > 0 is constant depending only on p and ρ. Bp(µ) − kykp Bp(µ)(cid:12)(cid:12)(cid:12) dν(x)dν(y)!1/r ≤ C(p, ρ)I p r , r p/2 n ) , direct application of Lemma 2.2 yields Having proved Theorem 5.1, we apply Markov's inequality as in Section 3 to get the corresponding concentration inequality. For the proof of Theorem 5.1 we argue as follows: Consider the function f (x) = kxkp Bp(µ) and define F = f − Eν f . Then, a ρZ r 2 (cid:13)(cid:13)(cid:13)k∇ fk2(cid:13)(cid:13)(cid:13) for all r ≥ 2. Recall the known fact (e.g. see [18]) that if a measure ν satisfies a log- Sobolev inequality with constant ρ, also satisfies a Poincaré inequality with constant ρ, that is Lr (ν) ≤ kFk2 kFk2 (5.1) ds, L2 (ν) Ls(ν) + 1 2 kh − Eνhk2 L2(ν) ≤ 1 ρZRn k∇hk2 2 dν = for any smooth function h. Therefore, (5.1) becomes kFk2 Lr (ν) ≤ 2 2 ρZ r 2 (cid:13)(cid:13)(cid:13)k∇ fk2(cid:13)(cid:13)(cid:13) Ls(ν) ds ≤ , L2(ν) 2 , Lr (ν) 1 2 ρ(cid:13)(cid:13)(cid:13)k∇hk2(cid:13)(cid:13)(cid:13) ρ (cid:13)(cid:13)(cid:13)k∇ fk2(cid:13)(cid:13)(cid:13) 2r 2The isotropic measures on Rn are defined similarly: A Borel probability measure ν on Rn is said to be isotropic if RRnhx, θi2 dν(x) = 1 for all θ ∈ S n−1. 20 for all r ≥ 3, where we have used the fact that s 7→ khkLs is non-decreasing function. Taking into account the Claim in Section 3 we get (5.2) kFk2 Lr (ν) ≤ 2p2r ρ ZRn kxkr(p−1) B2p−2(µ) dν(x)!2/r , r ≥ 3. Again, Lemma 2.2 implies that ZRn kxkr(p−1) B2p−2(µ) dν(x)!2/r for r ≥ 2. Plug this back in (5.2) we obtain ρ !1/2 kFkLr (ν) < 2p2 for all r ≥ 3. Finally, we have n ≤ I p p−1 , p−1 2 , r1/2I p−1 1 + 1 + ρI2 ≤ I2p−2 (r − 2)(p − 1) 2p−2 (ν, B2p−2(µ)) 2p−2(ν, B2p−2(µ)) p (ν, Bp(µ)) ≤ 1 + 2p−2(ν, B2p−2(µ)) 2p−2(ν, B2p−2(µ)) ρ !p/2 p − 2 (r − 2)(p − 1) ρI2 n, for all p ≥ 2. Indeed; we may write I p p (ν, Bp(µ)) =ZS n−1ZRn hx, θip dν(x) dµ(θ) ≥ZS n−1 ZRnhx, θi2 ν(x)!p/2 where we have used Hölder's inequality and the isotropicity of ν. For the right-hand side, we fix θ ∈ S n−1 and we apply Lemma 2.2 for x 7→ hx, θi to get dµ(θ) = µ(S n−1), ZRn hx, θip dν(x)!2/p ≤ZRnhx, θi2 dν(x) + p − 2 ρ = 1 + p − 2 ρ , where we have used the isotropicity again. Finally, integration with respect to µ yields: I p p (ν, Bp(µ)) =ZS n−1ZRn hx, θip dν(x) dµ(θ) ≤ZS n−1 1 + ρ !p/2 p − 2 dµ(θ), as asserted. Taking into account these estimates, we argue as in Section 3 to complete the proof. The details are left to the reader. 3. Minimal Gaussian variance for subspaces of Lp. Let us point out that our method also provides upper estimate for the variance of the norm of any finite dimensional subspace of Lp in Lewis' position. We should mention that the following estimate turns out to be optimal (up to constants of p) since they agree with the ℓn p case (see [25, Section 3] for details). 21 Theorem 5.2. Let 1 ≤ p < ∞. Then, for any n-dimensional subspace X of Lp repre- sented on Rn, there exists a position B of its unit ball BX such that VarkZk B ≤ C pn 2 p−1, where C > 0 is an absolute constant. minimal possible order (up to constants of p) In particular, the normalized variance is of VarkZk B EkZk2 B C p n . ≤ Sketch of Proof. If B is a Lewis' position of BX, we may identify X with Xp(µ) for some Borel isotropic measure µ on S n−1. Then, the result follows from Corollary 3.5. On the other hand note that the normalized variance is minimal since for every norm k · k on Rn one has VarkZk & EkZk2/n. The latter may be easily checked by using integration in polar coordinates and the Cauchy-Schwarz inequality. (cid:3) 4. A Johnson-Lindenstrauss type result. Note that using Lewis' lemma (Theorem 2.4) and the two level concentration for the Bp(µ) bodies (Theorem 3.2) we can conclude the following low-dimensional embedding of Hilbertian sets to any subspace of Lp, 2 < p < ∞. This can be viewed as a Johnson-Lindenstrauss type result for target spaces which sit in Lp. , (log N)p/2 Theorem 5.3. Let 2 < p < ∞. For any ε ∈ (0, 1) and for any N ≥ 1, there exists m . (cid:27) with the following property: For any subset S = {x1, . . . , xN} ⊂ max(cid:26)p4p log N ℓ2 and for any subspace X of Lp with dim X = m, there exists a linear mapping T : S → X which satisfies ε2 ε (1 − ε)kxi − x jk2 ≤ kT xi − T x jkX ≤ (1 + ε)kxi − x jk2, for all i, j = 1, . . . , N. References [1] F. Albiac and N. Kalton, Topics in Banach Space Theory, GTM, Springer-Verlag, (2006). [2] S. Artstein-Avidan, A. Giannopoulos and V. D. Milman, Asymptotic Geometric Analysis, Part I, AMS-Mathematical Surveys and Monographs 202 (2015). [3] G. Aubrun, S. Szarek and E. Werner, Hastings's additivity counterexample via Dvoretzky's theorem, Comm. Math. Phys. 305 (2011), no. 1, 85 -- 97. [4] K. Ball, Volume ratios and a reverse isoperimetric inequality, J. London Math. Soc. (2) 44, (1991), 351 -- 359. [5] G. Bennett, L. E. Dor, V. Goodman, W. B. Johnson and C. M. Newman, On uncomple- mented subspaces of L p, 1 < p < 2, Israel J. Math. 26, no. 2 (1977), 178 -- 187. 22 [6] S. Boucheron, G. Lugosi, and P. Massart, A sharp concentration inequality with appli- cations, Random Structures Algorithms 16 (2000), 277 -- 292. [7] A. Dvoretzky, Some results on convex bodies and Banach spaces, Proc. Int. Symp. on linear spaces, Jerusalem (1961), 123 -- 160. [8] A. Dvoretzky and C. A. Rogers, Absolute and unconditional convergence in normed linear spaces, Proc. Nat. Acad. Sci., U.S.A 36 (1950), 192 -- 197. [9] R. Eldan and B. Klartag, Pointwise estimates for marginals of convex bodies, J. Funct. Anal. 254 (2008), no. 8, 2275 -- 2293. [10] T. Figiel, J. Lindenstrauss and V. D. Milman, The dimension of almost spherical sections of convex bodies, Acta Math. 139 (1977), 53 -- 94. [11] S. Foucart and H. Rauhut, A mathematical introduction to compressive sensing, Applied and Numeric Harmonic Analysis, Birkhäuser/Springer, New York, 2013. [12] D. Fresen, Explicit Euclidean embeddings in permutation invariant normed spaces, Adv. Math. 266, 1 -- 16 (2014). [13] Y. Gordon, Some inequalities for Gaussian processes and applications, Israel J. Math. 50 (1985), 265 -- 289. [14] A. Grothendieck, Sur certaines classes de suites dans les espaces de Banach et le théorème de Dvoretzky-Rogers, (French) Bol. Soc. Mat. São Paulo 8 1953, 81 -- 110, (1956). [15] H. Huang and F. Wei, Upper bound for the Dvoretzky dimension in Milman-Schechtman theorem, In Geometric aspects of functional analysis, Lecture Notes in Math. 2169, 181 -- 186, Springer (2017). [16] B. Klartag and R. Vershynin, Small ball probability and Dvoretzky theorem, Israel J. Math., Vol. 157, no. 1 (2007), 193 -- 207. [17] H. König, Isometric imbeddings of Euclidean spaces into finite-dimensional ℓp-spaces, Panoramas of Mathematics (Colloquia 93-94), 79 -- 87, Banach Center Publ. 34, Polish Acad. Sci., Warszawa, 1995. [18] M. Ledoux, The concentration of measure phenomenon, Mathematical Surveys and Monographs 89, American Mathematical Society, Providence, RI, (2001). [19] D. R. Lewis, Finite dimensional subspaces of L p, Studia Math. 63 (1978), 207 -- 212. [20] E. Lutwak, D. Yang and G. Zhang, L p John ellipsoids, Proceedings of the London Math- ematical Society, Vol. 90, no. 2, (2005), 497 -- 520. [21] A. Magen, Dimensionality reductions that preserve volumes and distance to affine spaces, and their algorithmic applications, Randomization and approximation techniques in com- puter science, 239 -- 253, Lecture Notes in Comput. Sci.,2483, Springer, Berlin, 2002. [22] V. D. Milman, New proof of the theorem of A. Dvoretzky on sections of convex bodies, (Russian), Funkcional. Anal. i Prilozen. 5 (1971) 28 -- 37. [23] V. D. Milman and G. Schechtman, Asymptotic theory of finite dimensional normed spaces, Lecture Notes in Math. 1200 (1986), Springer, Berlin. [24] V. D. Milman and G. Schechtman, Global versus Local asymptotic theories of finite- dimensional normed spaces, Duke Math. Journal 90 (1997), 73 -- 93. 23 [25] G. Paouris, P. Valettas and J. Zinn, Random version of Dvoretzky's theorem in ℓn p, Stochastic Process. Appl. 127 (10) (2017), 3187 -- 3227. [26] G. Pisier, Probabilistic Methods in the Geometry of Banach Spaces, Lecture Notes in Mathematics 1206, Springer (1986), 167 -- 241. [27] G. Schechtman, A remark concerning the dependence on ε in Dvoretzky's theorem, Geo- metric Aspects of Functional Analysis (1987-88), 274 -- 277, Lecture Notes in Mathematics 1376, Springer, Berlin (1989). [28] G. Schechtman, Two observations regarding embedding subsets of Euclidean spaces in normed spaces, Advances in Mathematics 200 (2006), 125 -- 135. [29] G. Schechtman, The random version of Dvoretzky's theorem in ℓn ∞ 2005, 265 -- 270, Lecture Notes in Math., 1910, Springer-Verlag (2007). , GAFA Seminar 2004- [30] G. Schechtman, Euclidean sections of convex bodies, Asymptotic geometric analysis, 271 -- 288, Fields Inst. Commun. 68, Springer, New York, 2013. [31] G. Schechtman and A. Zvavitch, Embedding subspaces of L p into ℓN p , 0 < p < 1, Math. Nachr. 227 (2001), 133 -- 142. [32] G. Schechtman and J. Zinn, On the volume of the intersection of two Ln p balls, Proc. Amer. Math. Soc. 110, No.1, (1990), 217 -- 224. [33] M. Schmuckenschläger, On the dependence on ε in a theorem of J. Bourgain, J. Linden- strauss and V. D. Milman, Geometric Aspects of Functional Analysis, Israel Seminar GAFA 1989-1990, 166-173, Lecture Notes in Mathematics 1469, (1991), Springer. [34] P. Stavrakakis and P. Valettas, On the geometry of log-concave probability measures with bounded log-Sobolev constant, Proceedings of the Asymptotic Geometric Analysis Programme, Fields Institute Communications 68 (2013), 359 -- 380. [35] A. Szankowski, On Dvoretzky's theorem on almost spherical sections of convex bodies, Israel J. Math. 17 (1974), 325 -- 338. [36] K. E. Tikhomirov, The Randomized Dvoretzky's theorem in ℓn ∞ and the χ-distribution, Geometric Aspects of Functional Analysis, Israel Seminar GAFA 2011 -- 2013, 455 -- 463 (eds. B. Klartag and E. Milman), Lecture Notes in Mathematics 2116, (2013) Springer. [37] K. E. Tikhomirov, Almost Euclidean sections in symmetric spaces and concentration of order statistics, J. Funct. Anal. 265 (9), 2074 -- 2088, (2013). [38] R. Vershynin, Introduction to the non-asymptotic analysis of random matrices, In: Com- pressed sensing, 210 -- 268. Cambridge Univ. Press, Cambridge (2012). 24 Grigoris Paouris: [email protected] Department of Mathematics, Mailstop 3368 Texas A&M University College Station, TX 77843-3368 Petros Valettas: [email protected] Mathematics Department University of Missouri Columbia, MO 65211 25
1102.1739	1	1102	2011-02-08T21:54:18	Self-Adjoint Extension of Symmetric Maps	[ "math.FA" ]	A densely-defined symmetric linear map from/to a real Hilbert space extends to a self-adjoint map. Extension is expressed via Riesz representation. For a case including Friedrichs extension of a strongly monotone map, self-adjoint extension is unique, and equals closure of the given map.	math.FA	math	Self-Adjoint Extension of Symmetric Maps H. N. Friedel November 13, 2018 Abstract A densely-defined symmetric linear map from/to a real Hilbert space extends to a self-adjoint map. Extension is expressed via Riesz representation. For a case including Friedrichs extension of a strongly monotone map, self-adjoint extension is unique, and equals closure of the given map. Let {A : X ⊇ Do(A) → X} be a densely-defined symmetric linear map. Recall that if Hilbert- space X is complex, then A may lack self-adjoint extension (see e.g. [R]). In contrast, self-adjoint extension must exist if our Hilbert-space is real, as will be shown here. To prepare, we express well-known material in a form convenient for the present purpose. For x ∈ X, let (xA) denote the linear function {Do(A) ∋ y 7→ (xAy)}; we use the convention that scalar-product is linear in the second entry, conjugate-linear in the first. Observe the adjoint domain Do(A∗) equals {x ∈ X : (xA) continuous}. Recall: Do(A) ⊆ Do(A∗) ; A is self-adjoint iff Do(A) = Do(A∗) . Let J denote the duality-map on X, which maps x to function (x·) in dual-space X ∗ ; write J −1 = R, Riesz-representation. Extend Riesz-map R so as to act on densely-defined (continuous linear) functions, such as (xA) if x ∈ Do(A∗) . Note. Let A have symmetric extension B. Then (i ) Do(A) ⊆ Do(B) ⊆ Do(B ∗) ⊆ Do(A∗) . (ii ) R(xB) = R(xA), if x ∈ Do(B ∗) ⊆ Do(A∗) . (iii ) Bx = R(xA) if x ∈ Do(B) . Proof. (i ) is known. For y ∈ Do(A), see (R(xB) y) = (xBy) = (xAy) = (R(xA) y) ; density of Do(A) gives (ii ). For x ∈ Do(B) and y ∈ Do(A) , see (xA) is continuous, and (Bxy) = (xBy) = (xAy) = ( R(xA) y) ; density of Do(A) gives (iii ). Done. Denote by Λ the linear map {Do(A∗) ∋ x 7→ R(xA)}. Note(iii) (above) says A has at-most-one 1 symmetric extension to a given subspace Y , with Do(A) ⊆ Y ⊆ Do(A∗) ; if such extension exists, then it equals the restriction Λ(cid:12)(cid:12)Y . Theorem. Every symmetric map from/to a real Hilbert space has self-adjoint extension. Proof. Let E denote the order-set of linear subspaces Y , with Do(A) ⊆ Y ⊆ Do(A∗) , for which restriction Λ(cid:12)(cid:12)Y by the union of subspaces in C; so Zorn's lemma ensures E has a maximal member, Z. Λ(cid:12)(cid:12)Z is symmetric; order by inclusion. (E ∋ Do(A).) A chain C in E is bound above is a maximal symmetric extension of A. = M . We claim Do(M ) = Do(M ∗); if true, then M would be self-adjoint, concluding Write Λ(cid:12)(cid:12)Z the proof. It is enough to show Do(M ∗) ⊆ Do(M ); suppose not, seek a contradiction. Fix p ∈ Do(M ∗)(cid:15)Do(M ). On the subspace Do(M ) ⊕ R p , define a map T : T (x + a p) = M x + a R(pM ) if x ∈ Do(M ), a ∈ R. See T is linear, and T properly extends M . To show symmetry of T , let {x, y} ⊂ Do(M ) and {a, b} ⊂ R; note (x (cid:12)(cid:12) R(pM )) = (pM x), (R(pM ) (cid:12)(cid:12) y) = (pM y) ; compute: (cid:0) T (x + ap) (cid:12)(cid:12) y + bp(cid:1) = (cid:0) M x + aR(pM ) (cid:12)(cid:12) y + bp(cid:1) = (M xy) + b(M xp) + a(cid:0) R(pM ) (cid:12) (cid:12) y(cid:1) + ab(cid:0) R(pM ) (cid:12) (xM y) + b(cid:0)x (cid:12) (cid:0)x + ap (cid:12) M has symmetric proper extension T , so M is not a maximal symmetric extension of A; contra. (cid:12) R(pM ) (cid:1) + a(pM y) + ab(cid:0)p (cid:12) (cid:12) M y + bR(pM )(cid:1) = (cid:0)x + ap (cid:12) (cid:12) p(cid:1) = (cid:12) R(pM ) (cid:1) = (cid:12) T (y + bp)(cid:1) . Done. So, self-adjoint extension exists; now treat uniqueness. Fortunately, extension is unique for some cases of interest; sometimes we may even express extension simply, as closure of the given map. To prepare to show this, recall A has symmetric closure ¯A ⊆ M . Here, as before, {A : X ⊇ Do(A) → X} is symmetric, with self-adjoint extension M , from/to a Hilbert space X, now assumed real. We also need the following two facts. Note 1. If A has dense image and continuous inverse, then ¯A is the unique self-adjoint extension of A; M = ¯A. ¯A maps onto X, and has continuous self-adjoint inverse. ¯A has dense image (since A does); recall a symmetric map ( ¯A) with dense image has Proof. ¯A−1 equals closure of a continuous map symmetric inverse; ¯A−1 is also closed, since ¯A is so. (A−1), hence ¯A−1 is continuous. Since ¯A−1 is closed, continuous, and has dense domain (including Im(A) ), we have Do(cid:0) ¯A−1(cid:1) = X. A continuous symmetric map (cid:0) ¯A−1(cid:1) on the whole Hilbert space is self-adjoint. Recall a self-adjoint map (cid:0) ¯A−1(cid:1) with dense image (including Do(A) ) has self-adjoint inverse ( ¯A). Hence { ¯A, M } are self-adjoint extensions of A, with ¯A ⊆ M ; this forces ¯A = M , because a self-adjoint map is maximal-symmetric. Done. Note 2. A (densely-defined) closed 1 : 1 symmetric map has dense image. 2 Proof. It is enough to show p = 0, if p ∈ Im ⊥(A) (orthogonal complement of image). Since Do(A) is dense, it has a sequence {un} converging to p. If x ∈ Do(A) , then 0 = (pAx) = lim(unAx) = lim(Aunx). Density of Do(A) forces lim Aun = 0. A is closed; (lim un = p) and (lim Aun = 0); hence p ∈ Do(A), Ap = 0. Since A is 1 : 1, we have p = 0. Done. Recall (e.g. [Z]) that if our map A is strongly monotone, then it has Friedrichs extension, which is self-adjoint, 1 : 1, onto, with continuous self-adjoint inverse. Theorem. If A is strongly monotone, then closure ¯A is the unique self-adjoint extension of A; ¯A equals Friedrichs extension. Proof. Let A denote Friedrichs extension; A ⊇ ¯A . Since A is 1 : 1 with continuous inverse, so is its restriction ¯A. By Note 2, closed symmetric 1 : 1 map ¯A has dense image; then Note 1 makes ¯A the unique self-adjoint extension of itself, and of A. A is a self-adjoint extension of A, hence A = ¯A. Done. Construction of the Friedrichs extension is complicated; how nice to express it simply (as closure), and to know it is the only self-adjoint extension. References [R] Rudin, W. Functional Analysis. McGraw-Hill, 1991. [Z] Zeidler, E. Applied Functional Analysis: Applications to Mathematical Physics. Springer, 1995. 3
1905.09206	1	1905	2019-05-22T15:54:54	On additive property of finitely additive measures	[ "math.FA", "math.GN" ]	By the additive property, we mean a condition under which $L^p$ spaces over finitely additive measures are complete. Basile and Rao gives a necessary and sufficient condition that a finite sum of finitely additive measures has the additive property. We generalize this result to the case of a countable sum of finitely additive meaures. An application of this result to density measures are also presented.	math.FA	math	ON ADDITIVE PROPERTY OF FINITELY ADDITIVE MEASURES RYOICHI KUNISADA Abstract. By the additive property, we mean a condition under which Lp spaces over finitely additive measures are complete. Basile and Rao gives a necessary and sufficient condition that a finite sum of finitely additive measures has the additive property. We generalize this result to the case of a countable sum of finitely additive meaures. An application of this result to density measures are also presented. 1. Introduction Let X be a set and F be a field of subsets of X. A function µ on F is called a finitely additive measure or a charge if the following conditions are satisfied. (1) µ(∅) = 0, (2) µ(A ∪ B) = µ(A) + µ(B) if A, B ∈ F, A ∩ B = ∅. The triple (X, F, µ) is called a finitely additive measure space or charge space. In what follows, we use the term charge exclusively. Obviously, charges are a generalization of measures obtained by replacing countable additivity of measures with finite additivity. The theory of charges was developed systematically in [2]. In this book, various notions and results in measure theory are generalized and transferred to charge spaces. Among these, the notion of Lp spaces over charges is of particular importance for applications of charge theory. Recall that one of the important conclusions of measure theory is the completeness of Lp spaces over measures. Unfortunately however, Lp spaces over charges are not complete in general. The condition under which Lp spaces over charges are complete is given by a certain additivity property which is intermediate between countable additivity and finite additivity (see [1]): Let (X, F, µ) be a charge space. We say that µ has the additive property if for any ε > 0 and increasing sequence A1 ⊆ A2 · · · An ⊆ · · · in F, there exists a set B ∈ F such that (1) µ(B) ≤ limi→∞ µ(Ai) + ε, (2) µ(Ai \ B) = 0 The study of the additive property for concrete examples of charges were developed, for every i = 1, 2, . . .. for instance, in [3], [4], [5]. In this paper, we deal with the additive property of sums of charges. Note that if charges µ and ν on (X, F) have the additive property, µ + ν does not necessarily have the additive property. In [ ], the necessary and sufficient condition that finite sums of This paper is a part of the outcome of research performed under a Waseda University Grant for Special Research Projects (Project number: 2018K-152). 1 charges have the additive property has been obtained. One of the main aims of this paper is to generalize the result to the case of countable sums of charges. Further, we give an application of this result to the additive property of density measures, which gives a simpler proof of the main result of [4]. This can be seen as another main result of this paper. The paper is organized as follows. In Section 2, we introduce some notions and results which will be used throughout the paper. In particular, the notions of singularity and strong singularity of charges and a Borel measure on the stone space of an algebra of sets play an important role in studying the additive proeperty. In Section 3, we chronicle equivalent conditions to the additive property, which illustrate the importance of the additive property in the theory of charges. In fact, these theorems asserts that some of the main theorems in meaure theory, including the completeness of Lp spaces and the Radon-Nikodym theorem, are also valid for charges having the additive proeperty. Section 4 deals with one of the main results of this paper, i.e., we prove a necessary and sufficient condition that a countable sum of charges has the additive property. This is done by using one of the equivalent formulations of the additive property introduced in Section 3. Section 5 is devoted to an application of the result of Section 4. 2. Preliminaries In what follows, if not otherwise stated, charges will always be nonnegative and bounded, i.e., any given charge space (X, F, µ), we have µ(A) ≥ 0 for every A ∈ F and µ(X) < ∞. First, for a charge space (X, F, µ) we introduce an extension of µ to a regular Borel measure of the stone space of F. This method plays an important role for formulating various notions concerning charges. Regarding F as a Boolean algebra, by the Stone representation theorem, there exists a totally disconnected compact space F and a natural Boolean isomorphism φ : F → C, where C is the algebra of the clopen subsets of F . Now define a charge µ on C by µ(φ(A)) = µ(A) and we get a charge space (F, C, µ). Since any infinite union of clopen subsets can not be a clopen subset, µ is countably additive on C and thus by the E. Hopf extension theorem, we can extend it to a countable additive measure on the σ-algebra generated by C, i.e., a Baire measure on F . This can also be extended to a regular Borel measure on F in a unique way. We still denote it by µ and thus we obtain a measure space (F, B(F ), µ), where B(F ) is the Borel sets of F . We denote by supp µ the support of µ in F . Following [2], we consider the notions of absolute continuity and singularity for charges. Definition 2.1. Let µ and ν be charges on (X, F). 2 (1) We say that ν is absolutely continuous with respect to µ if for any ε > 0, there exists δ > 0 such that ν(A) < ε whenever µ(A) < δ, where A ∈ F. In this case, we write ν ≪ µ. (2) We say that µ and ν are singular if for every ε > 0, there exists a set D ∈ F such that µ(D) < ε and ν(Dc) < ε. In this case, we write µ ⊥ ν. Next we define the notions of weakly absolute continuity and strong singularity. Definition 2.2. Let µ and ν be charges on (X, F). (1) We say that ν is weakly absolutely continuous with respect to µ if ν(A) = 0 whenever µ(A) = 0, where A ∈ F. In this case, we write ν ≺ µ. (2) We say that µ and ν are strongly singular if there exists a set D ∈ F such that µ(D) = 0 and ν(Dc) = 0. In this case, we write µ (cid:15) ν. Obviously, these are ordinary notions of absolute continuity and singularity for mea- sures in case that F is a σ-algebra and µ and ν are measures on it. In fact, as the following theorems show, absolute continuity and weakly absolute continuity, and sin- gularity and strong singularity coincide respectively in this case. See [2] for the proofs. Theorem 2.1. Let µ and ν be measures on (X, F). Then we have the following results. (1) If ν is a bounded measure, then ν ≪ µ if and only if ν ≺ µ. (2) ν ⊥ µ if and only if µ (cid:15) ν. Now we give formulations of these notions by using the extended measure space. The following is essentially due to [1]. Theorem 2.2. Let µ and ν be charges on (X, F) and µ and ν be the extended measures on (F, B(F )), respectively. Then the following statements hold: (1) ν ≪ µ if and only if ν ≺ µ. (2) ν ≺ µ if and only if supp ν ⊆ supp µ. (3) ν ⊥ µ if and only if ν (cid:15) µ. (4) µ (cid:15) ν if and only if supp ν ∩ supp µ = ∅. Proof . (1) Let us assume that ν ≪ µ. Given ε > 0, take δ > 0 as above. Let E ∈ B(F ) with µ(E) = 0. Since µ is regular, for any δ > δ′ > 0, one can choose A ∈ F such that µ(φ(A)△E) < δ′ and ν(φ(A)△E) < δ′. Thus we have µ(A) = µ(φ(A)) ≤ µ(φ(A)△E) + µ(E) = µ(φ(A)△E) < δ′. On the other hand, ν(E) ≤ ν(φ(A)△E) + ν(φ(A)) ≤ δ′ + ε. Since δ′ and ε can be arbitrary small, we have ν(E) = 0. Thus ν ≺ µ. Conversely, suppose that ν ≺ µ. Since B(F ) is a σ-algebra and µ and ν are measures, ν ≺ µ if and only if ν ≪ µ by Theorem 2.1. Thus through φ−1 we have ν ≪ µ. µ(Dε1,i) < ε1 (2) This is obvious. (3) Let us assume that ν ⊥ µ. Let ε1 > 0 and Dε1,i ∈ F, i ≥ 1 be such that 2i for every i ≥ 1. Hence 2i = ε1. Now we choose a decreasing sequence {εj}j≥1 of positive numbers such that limj→∞ εj = 0. we have ν(∪i≥1φ(Dε1,i)) = 1. On the other hand, µ(∪i≥1φ(Dε1,i)) ≤ P∞ 2i and ν(Dc ε1,i) < ε1 2i . Notice that ν(Dε1,i) > 1 − ε1 ε1 i=1 3 Then we put E = ∩j≥1 ∪i≥1 φ(Dεj ,i) and we get ν(Ec) = 0 and µ(E) = 0, which means that µ (cid:15) ν. Now suppose that µ (cid:15) ν. Then there exists a set D in B(F ) such that µ(D) = 0 and ν(Dc) = 0. By the regularity of µ and ν, there exists a set C in F such that µ(φ(C)△D) < ε and ν(φ(C)△D) < ε. Thus we have µ(C) = µ(φ(C)) ≤ µ(φ(C)△D)+ µ(D) < ε and ν(C c) = ν(φ(C c)) ≤ ν(φ(C c)△Dc) + ν(Dc) = ν(φ(C)△D) + ν(Dc) < ε. This shows that ν ⊥ µ. (4) This is obvious. Concerning these notions, we give a generalization of the Lebesgue decomposition theorem to charges (Refer to [2] for details). Theorem 2.3. For given charges µ and ν on (X, F), there exists charges ν1 and ν2 on (X, F) such that (1) ν = ν1 + ν2. (2) ν1 ≪ µ. (3) ν2 ⊥ µ. Furthermore, a decomposition of ν satisfying (2) and (3) is unique. 3. Equivalent conditions to additive property In this section, we introduce some equivalent assertions to the additive property. As we have mentioned in Section 1, one can generalize some of the main theorems in measure theory to charges having the additive property. In fact, conversely, the validity of these theorems are also sufficient conditions for charges to have the additive property. We begin with the completeness of Lp spaces over charges, which is the original motive of introducing the notion of the additive property. See [1] for the proofs of the following results. Theorem 3.1. For a charge µ (not necessarily bounded) on (X, F), µ has the additive property if and only if Lp(µ) is complete. The next result is a generalization of the Radon-Nikodym theorem to charges. Theorem 3.2. For a charge µ on (X, F), µ has the additive property if and only if for every charge ν (not necessarily nonnegative) on (X, F) with ν ≪ µ, there exists some f ∈ L1(µ) such that ν(A) = RA f dµ holds for every A ∈ F. The Hahn decomposition theorem can be generalized to charges as follows. Theorem 3.3. For a charge µ on (X, F) where F is a σ-algebra, µ has the additive property if and only if for every charge ν (not necessarily nonnegative) on (X, F) with ν ≪ µ, there exists some A ∈ F satisfying the following property; for each B ∈ F with B ⊆ A, ν(B) ≥ 0 holds, and each B ∈ F with B ⊆ Ac, ν(B) ≤ 0 holds. Finally, we give the following formulation of the additive property using extended measure spaces, which plays an important role in proving the main theorem in the 4 following section. As we have seen in Section 2, we extend a charge space (X, F, µ) to a measure space (F, B(F ), µ). Theorem 3.4. A charge µ on (X, F) has the additive property if and only if µ(U) = µ(U) for every open sets U of supp µ, where U denotes the closure of U in supp µ. 4. Main result In this section, we consider a necessary and sufficient condition that charges which are expressed by sums of charges have the additive property. Generally, as shown in [1], if charges µ and ν on (X, F) have the additive property, the sum µ + ν need not have the additive property. First we have the following result. Theorem 4.1. Let µ, ν be charges on (X, F) such that ν ≪ µ. If µ has the additive property, then ν has the additive property. From this result together with the Lebesgue decomposition theorem, it is sufficient to consider the condition for pairs of charges µ, ν which are mutually singular. It is given by the following, in which a slightly general result of the additive property of finite sums of charges are treated. Theorem 4.2. Let µ1, µ2, . . . , µn be mutually singular charges one another on (X, F) where F is a σ-algebra. Then µ1 + µ2 + . . . + µn has the additive property if and only if every µi, 1 ≤ i ≤ n, has the additive property and they are mutually strongly singular. Note that a more general result was proved in [1], namely, the case in which F is not necessarily a σ-algebra. It is obtained by replacing in the above assertion strong singularity with separability, in between singularity and strong singularity. The notion of separability is rather complicated than strong singularity and thus we will confine ourselves to the case of σ-algebra. Now we consider an extension of Theorem 4.2 to the case of countable sums of charges, which is one of the main results of this paper. Theorem 4.3. Let {µi}i≥1 be a countable family of charges on (X, F) where F is a σ-algebra such that they are mutually singular and µ = Pi≥1 µi exists. Let Si be the support of µi and S be the support of µ. Then µ has the additive property if and only if each µi has the additive property and they are mutually strongly singular and (cid:18)lim sup i Si(cid:19) \ [i≥1 Si = ∅ holds, where lim supi Si = ∩i≥1∪j≥iSj. Proof . (Sufficiency) We prove the condition in Theorem 3.4. By the definition of µ, it holds that µ = Pi≥1 µi and thus µ is on ∪i≥1Si. Also since µi are mutually strongly singular, µ(A) = µ(A ∩ ∪i≥1Si) = Pi≥1 µ(A ∩ Si) holds. In what follows, for any Borel set X of F we denote the closure of B ⊆ X in In the case of X = F , we omit the superscript. Note that the closure X by B X . 5 of B ⊆ X in X is X ∩ B. Take any A ∈ B(F ). By the assumption, note that S = ∪i≥1Si ∩ lim supi Si and lim supi Si ∩ ∪i≥1Si = ∅. Since each µi has the additive property, from Theorem 3.4 and the fact that charges µi are mutually strongly singular, we have µ(A ∩ Si) = µi(A ∩ Si) = µi(A ∩ Si Si) = µi(A ∩ Si) = µ(A ∩ Si). On the other hand, it holds that A ∩ S = A ∩ (∪i≥1Si ∪ lim supi Si) = ∪i≥1(A ∩ Si)∪ A ∩ lim supi Si = ∪i≥1A ∩ Si∪A ∩ lim supi Si. Together with the fact that µ(lim supi Si) = 0, we have µ(A ∩ S S ) = µ(A ∩ S) = µ(∪i≥1A ∩ Si) + µ(A ∩ lim sup Si) = Xi≥1 µ(A ∩ Si) = Xi≥1 i µ(A ∩ Si) = µ(A ∩ S). Hence by Theorem 3.4 we see that µ has the additive property. (Necessity) Suppose that {µi}i≥1 are singular and µ = Pi≥1 µi has the additive property. Let µn and µm be any pair of distinct charges. Put µ′ = Pi≥1,i6=n µi and since µ′ ⊥ µn and µ = µ′ + µn, we have that by Theorem 4.2 µ′ and µn have the additive property and they are strongly singular. Hence we conclude that each µn has the additive property and they are strongly singular one another. Next we show that lim supi Si ∩ ∪i≥1Si = ∅. Assume to the contrary that lim supi Si ∩ ∪i≥1Si 6= ∅. Fix some n ≥ 1 with lim supi Si ∩ Sn 6= ∅ and consider the charge µ′ = Pi≥1,i6=n µi. Since the support S′ of µ′ is ∪i≥1,i6=nSi ∪ lim supi Si, S′ ∩ Sn 6= ∅ holds and thus µ′ and µn are not strongly singular. But this contradicts Theorem 4.2 by the same arguments above, which completes the proof. 5. Application to density measures We consider the asymptotic density on natural numbers N, which is one of the most famous finitely additive set functions on a countable space. Let P(N) be the set of all subsets of N. For A ∈ P(N) the asymptotoic density d(A) of A is defined by d(A) = lim n A ∩ [1, n] n provided the limit exists, where X denotes the number of elements of X ∈ P(N). Let D be the set of all subsets A ∈ P(N) having the asymptotic density. Then d is obviously a finitely additive set function defined on D. However, unfortunaltely, the triple (N, D, d) is not a charge space since the class D is not an algebra of sets, i.e., D does not closed under union or intersection. Though (N, D, d) itself is not a charge space, one can construct a charge space from the asymptotic density by extending d to the whole P(N). A charge ν defined on P(N) is called a density measure if it extends the asymptotic density. An example of density meausres can be constructed simply by using the limit along an ultrafilter in place of the usual limit in the definition of the asymptotic density. 6 Recall that an ultrafilter on N is a filter on N which is not properly contained in any other filters. In particular, an ultrafilter is said to be free if the intersection of its elements is empty. Let U be an ultrafilter on N and f be in l∞ of the set of all real-valued bounded functions on N. Then there exists a unique number α such that {n ∈ N : f (n) − α < ε} ∈ U holds for any ε > 0. The number α is called the limit of f along U and denoted by U- limn f (n) = α. Now one can define a density meaure through the limit along an ultrafilter. Let U be an free ultrafilter on N, let us define a density measure νU by νU (A) = U- lim n A ∩ [1, n] n , A ∈ P(N). Let C be the set of all such density measures. In what follows, we show a necessary and sufficient condition for a density measure in C to have the additive property. Note that there exists distinct ultrafilters U and U ′ such that νU = νU ′. Thus, introducing an equivalence relation defined by U ∼ U ′ if and only if νU = νU ′ on the set of all free ultrafilter on N, C can be regarded as the quotient space by ∼. In fact, there exists a seciton such that each representative has a form which is well suited for investigating the corresponding density measure. Now we have to make some preparation before stating the precise assertion. Let βN be the Stone- Cech compactification of N. Note that βN can be characterized by the following property: for any mapping of N into a compact space X, there exists a continuous extension to βN. Recall that βN can be identified with the set of all ultrafilters on N in which a basis of open sets are those subsets A = {U : A ∈ U} for every A ∈ P(N). We denote by N∗ := βN \ N the set of all free ultrafilters on N. Then N∗ is also a compact space by the relative topology of βN and obviously a basis of open sets of N∗ are the sets of the form A∗ = A ∩ N∗ for every A ∈ P(N). We now introduce a topological dynamical system on N∗. Let us consider the trans- lation on N: τ0 : N → N, n 7→ n + 1. Embedding the range N into βN, one can regard τ0 as a mapping of N into a compact space βN. Thus it can be extended to the continuous function τ of βN into itself: τ : βN → βN. Though τ is not a homeomorphism, the restriction of τ to N∗ is a homeomorphism of N∗ onto itself. We still denote it by the same symbol τ and then the pair (N∗, τ ) is a topological dynamical system. Let us define the negative semi-orbit of U ∈ N∗ by o−(U) := {τ −nU : n = 0, 1, 2, . . . .}. o−(U) is said to be recurrent if for any neighborhood U of U and any natural number N ≥ 1, there exists some n ≥ N such that τ −nU ∈ U holds. The set of all such points in N∗ is denoted by Rd,−. Next we extend this discrete flow to a continuous flow which is the suspension of (N∗, τ ). Let us consider the product space N∗ × [0, 1] and define an equivalence relation on it by (U, 1) ∼ (τ U, 0) for every U ∈ N∗. Let Ω∗ be the quotient space of N∗ × [0, 1] 7 by ∼. Then we define a family of homeomorphisms τ s : Ω∗ → Ω∗ by τ s(U, t) = (τ [t+s]U, t + s − [t + s]), for each s ∈ R, where [x] denotes the largest integer not exceeding x for a real number x. Then it is readily verified that the pair (Ω∗, {τ s}s∈R) is a continuous flow. Similarly as above, we define the negative semi-orbit of ω ∈ Ω∗ by o−(ω) := {τ −sω : s ≥ 0}. Also, we say that o−(ω) is recurrent if for any neighborhood U of ω and positive real number R > 0, there exists some s ≥ R such that τ −sω ∈ U holds. The set of all recurrent points in Ω∗ is denoted by R−. Now the following theorem asserts that each element of C is expressed in a special form, which is convenient to investigate its measure theoretic properties. See [4], [5] for the proof and related results. Theorem 5.1. Let us define the mapping Φ : Ω∗ → C as follows: Let ω = (U, t) ∈ Ω∗ and θ = 2t, and let us denote the image Φ(ω) of ω by νω, then we define νω(A) = U- lim n A ∩ [1, [θ · 2n]] θ · 2n , A ∈ P(N). Then Φ is a one to one and onto mapping of Ω∗ to C. Before proving the main theorem, we introduce the following auxiliary charges; for ω = (U, t) ∈ Ω∗ and m = 0, 1, . . ., we define a charge νω,n : P(N) → [0, 1] by νω,m(A) = U- lim n A ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 , A ∈ P(N). Then it is obvious that νω = ∞ Xm=0 1 2m+1 νω,m. Lemma 5.1. For any ω = (U, t) ∈ Ω∗ and m = 0, 1, 2, . . ., νω,m has the additive property. Proof . Given an increasing sequence A1 ⊆ A2 ⊆ . . . ⊆ Ai ⊆ . . . of P(N). Set limi→∞ νω,m(Ai) = α. We take a decreasing sequence {Xi}i≥1 of U such that Ai ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 (cid:12)(cid:12)(cid:12)(cid:12) whenever n ∈ Xi. Then we define B ⊆ N as B ∩ ([θ · 2n−m−1], [θ · 2n−m]] = Ai ∩ ([θ · 2n−m−1], [θ · 2n−m]] if n ∈ Xi \ Xi+1 and B ∩ ([θ · 2n−m−1], [θ · 2n−m]] = ∅ otherwise. First we show that νω,m(B) = α. For any ε > 0, take i ∈ N with ε > 1 i and α − νω,m(Ai) < ε. 8 − νω,m(Ai)(cid:12)(cid:12)(cid:12)(cid:12) < 1 i Then for any n ∈ Xi, there exists some jn ≥ i such that n ∈ Xjn \ Xjn+1. Hence νω,m(B) − α = U- lim n (cid:12)(cid:12)(cid:12)(cid:12) B ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 B ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 Ajn ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 1 i + ε < 2ε. (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) ≤ lim sup n∈Xi = lim sup n∈Xi 1 jn ≤ + ε ≤ − α(cid:12)(cid:12)(cid:12)(cid:12) − νω,m(Ajn)(cid:12)(cid:12)(cid:12)(cid:12) − νω,m(Ajn)(cid:12)(cid:12)(cid:12)(cid:12) + νω,m(Ai) − α + νω,m(Ai) − α (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) Since ε > 0 is arbitrary, we have νω,m(B) = α. Next we show that νω,m(Ai \ B) = 0 for every i ≥ 1. For any n ∈ Xi, there exists some jn ≥ i such that n ∈ Xjn \ Xjn+1. Hence we have νω,m(Ai \ B) ≤ lim sup n∈Xi (Ai \ B) ∩ ([θ · 2n−m−1], [θ · 2n−m]] θ · 2n−m−1 = 0 since B ∩ ([θ · 2n−m−1], [θ · 2n−m]] = Ajn ∩ ([θ · 2n−m−1], [θ · 2n−m]] and Ai ⊆ Ajn. So we obtain the result. Based on the above lemma, now we prove the main result: Theorem 5.2. νω has the additive property if and only if ω 6∈ R−. Proof. Let Sm = supp νω,m and S = supp νω. Then by Theorem 4.3 and the fact mentioned above it is equivalent to show that (cid:18)lim sup m Sm(cid:19)\[i≥1 Si = ∅ if and only if ω 6∈ R−. Assume that ω = (U, t) 6∈ R− and thus U 6∈ Rd,−. This means that there exists some X ∈ U such that X ∗ ∩ o−(U) \ {U} = ∅. Put I0 = [n∈X ([θ · 2n−1], [θ · 2n]] and it is obvious that νω,0(I0) = 1, which means that S0 ⊆ I ∗ borhood of S0. Now we show that I ∗ such that τ −mXm ∩ X = ∅ and put 0 is a neigh- 0 ∩ Sm = ∅ for every m ≥ 1. In fact, take Xm ∈ U 0 , that is, I ∗ Im = [n∈τ −mXm ([θ · 2n−1], [θ · 2n]]. Then we have νω,m(Im) = 1 and thus Sm ⊆ I ∗ m. Since we have assumed that τ −mXm ∩ X = ∅, we have I ∗ 0 of S0 we have that I ∗ 0 ∩Sm = ∅ for all m ≥ 1 and thus lim supm Sm ∩S0 = ∅. In the same way, we can show that lim supm Sm ∩ Si = ∅ for every i ≥ 1 and thus we have lim supm Sm ∩ ∪i≥1Si = ∅. m = ∅. Hence for the neighborhood I ∗ 0 ∩ I ∗ 9 On the other hand, assume that ω ∈ R−, i.e., U is in the closure of o−(U). Then notice that for any X ∈ U and positive integer N > 0 there exists some n ≥ N such that τ −nU ∈ X ∗. We show that for any neighborhood I ∗ 0 of S0 where I0 ∈ P(N) and positive integer N, there is some n ≥ N such that I ∗ 0 ∩ Sn 6= ∅, which immediately implies that lim supm Sm ∩ ∪i≥0Si 6= ∅. Let f ∈ l∞ be the function defined by f (m) = I0 ∩ ([θ · 2m−1], [θ · 2m]]/θ · 2m−1, m = 1, 2, . . . . Then νω,0(I0) = f (U) holds by the definition of νω,0. Since f (U) = νω,0(N) = 1 and f is continuous on N∗, there exists a neighborhood X ∗ of U such that U ′ ∈ X ∗ implies f (U ′) > 0. Let n ≥ N be any integer such that τ −nU ∈ X. Then νω,n(I0) = 2n · f (τ −nU) > 0, which means that Sn ∩ I ∗ 0 6= ∅. We completes the proof. References [1] A. Basile, K. P. S. Bhaskara Rao, Completeness of L p-Spaces in the Finitely Additive Setting and Related Stories, J. Math. Anal. Appl. 248, (2000) 588-624. [2] K.P.S. Bhaskara Rao, M. Bhaskara Rao, Theory of charges, Academic Press (1983). [3] A. Blass, R. Frankiewicz, G. Plebanek, C. Ryll-Nardzewski, A note on extensions of asymptotic density, Proc. Amer. Math. Soc. 129 (2001) 3313-3320. [4] R. Kunisada, Density measures and additive property, J. Number Theory. 176 (2017), 184-203. [5] R. Kunisada, On a relation between density measures and a certain flow, Proc. Amer. Math. Soc. 147 (2019) 1941-1951. Faculty of Education and Integrated Arts and Science, Waseda University, Shinjuku- ku, Tokyo 169-8050, Japan E-mail address: [email protected] 10
1506.00634	1	1506	2015-06-01T12:52:35	Polynomial as a new variable - a Banach algebra with a functional calculus	[ "math.FA" ]	Given any square matrix or a bounded operator $A$ in a Hilbert space such that $p(A)$ is normal (or similar to normal), we construct a Banach algebra, depending on the polynomial $p$, for which a simple functional calculus holds. When the polynomial is of degree $d$, then the algebra deals with continuous $\mathbb C^d$-valued functions, defined on the spectrum of $p(A)$. In particular, the calculus provides a natural approach to deal with nontrivial Jordan blocks and one does not need differentiability at such eigenvalues.	math.FA	math	Polynomial as a new variable - a Banach algebra with a functional calculus Olavi Nevanlinna April 10, 2018 Aalto University Department of Mathematics and Systems Analysis email: [email protected] Abstract Given any square matrix or a bounded operator A in a Hilbert space such that p(A) is normal (or similar to normal), we construct a Banach algebra, depending on the polynomial p, for which a simple functional calculus holds. When the poly- nomial is of degree d, then the algebra deals with continuous Cd-valued functions, defined on the spectrum of p(A). In particular, the calculus provides a natural ap- proach to deal with nontrivial Jordan blocks and one does not need differentiability at such eigenvalues. 1 Introduction There are many situations in which it would be desirable to be able to treat polynomials as new global variables. For example, by Hilbert's lemniscate theorem (see e.g [12]) polynomials can be used to map complicated sets of the complex plane onto discs. As polynomials are not one-to-one we represent scalar functions in the original variable by a vector valued function in the polynomial. This leads to multicentric holomorphic calculus [9]. In [10] we applied it to generalize the von Neumann theorem on contrac- tions in Hilbert spaces. In such applications one would, given a bounded operator A, search for a polynomial p such that p(A) has a small norm - thus mapping a potentially complicated spectrum into a small disc. In this paper we study multicentric calculus without assuming the functions to be analytic. As an application we consider situations in which p(A) is diagonalizable or similar to normal. Thus, the aim is to remove the Jordan blocks by moving from A to p(A). To illustrate the goal consider finite dimensional matrices. If D = diag{αj} is a diagonal matrix and ϕ is a continuous function, then any reasonable functional calculus satisfies ϕ(D) = diag{ϕ(αj )}. Further, if A is diagonalizable so that with a similarity T we have A = T DT −1, then we of course set ϕ(A) = T ϕ(D)T −1. (1.1) However, if A has an eigenvalue with a nontrivial Jordan block, then the customary ap- proach is to assume that ϕ is smooth enough at the eigenvalues so that the off-diagonal 1 elements can be represented by derivatives of ϕ. For example, if J =   α 1 α 1 α   then ϕ(J) =   ϕ(α) ϕ′(α) ϕ(α) 1 2 ϕ′′(α) ϕ′(α) ϕ(α)   . (1.2) (1.3) A collection of different ways to define ϕ(A) for matrices can be found from [11], where Higham, following Gantmacher [6], says that a function ϕ is defined at the spectrum σ(A) = {αj} if the values ϕ(k)(αj) are known for 0 ≤ k ≤ nj, where nj + 1 are the powers in the minimal polynomial. This has two obvious drawbacks. First, since it is based on the Jordan form the functional calculus is discontinuous: for diagonalizable matrices it is given for all con- tinuous functions while it requires existence of derivatives at eigenvalues with nontriv- ial Jordan blocks. Second, the approach cannot conveniently be extended to infinite dimensional spaces. Recall that there is a natural functional calculus for normal op- erators which easily extends to operators which are similar to normal. If, however, an eigenvalue with a nontrivial Jordan block would exist in the middle of a cluster of other eigenvalues, then one would need to have a way to treat function classes which are continuous and additionally have derivatives at that particular eigenvalue. We shall show in this paper that there is a simple way to parametrize continuous functions which slow down at those places where some extra smoothness is needed. And it turns out that this allows a functional calculus which agrees with the holomor- phic functional calculus if applied to holomorphic functions but is defined for functions which do not need to be differentiable at any point. The starting point for the calculus is taking w = p(z) as a new variable. Since such a change of variable is only locally injective we compensate this by replacing the scalar function by a vector valued function ϕ : z 7→ ϕ(z) ∈ C f : w 7→ f (w) ∈ Cd where d is the degree of the polynomial p. The multicentric representation of ϕ is then of the form d ϕ(z) = Xj=1 δj(z)fj(p(z)), (1.4) where δj's are the Lagrange interpolation polynomials such that δj(λj ) = 1 while δj(λk) = 0 when k 6= j, [9]. If now p(A) is diagonalizable, one can apply the known functional calculus to represent fj(p(A)). But since δj's are polynomials, δj(A) is well defined and differ- entiability of ϕ is not needed. The paper is organized as follows. We first consider the Banach space of continuous functions f : M → Cd and associate with it a product , "polyproduct" ⊚, such that it becomes a Banach algebra, which we denote by CΛ(M ). Here Λ denotes the set of zeros of the polynomial p. Then the functions ϕ in (1.4) can be viewed as Gelfand transformations f of functions f ∈ CΛ(M ). Towards the end of the paper we discuss 2 the functional calculus for operators in Hilbert spaces H such that p(A) is similar to a normal operator. In particular we study the mapping χA which associates to f a bounded operator χA(f ) ∈ B(H) χA(f ) = d Xj=1 δj(A)fj(p(A)) and show that we get a homomorphism χA(f ⊚ g) = χA(f )χA(g) which, in an appro- priate quotient algebra, satisfies a spectral mapping theorem. 2 Construction of the Banach algebra 2.1 Multicentric representation of functions We assume given a polynomial p(z) = (z − λ1)··· (z − λd) with distinct zeros Λ = j=1 mapping the z-plane onto w-plane: w = p(z). In addition we denote by {λj}d Λ1 = {z : p′(z) = 0} the set of critical points of p. We call the points of Λ as the local centers of the multicentric calculus. Recall that by the Gauss-Lucas theorem Λ1 is in the convex hull of Λ. Suppose δj(z) are the Lagrange interpolation polynomials with interpolation points in Λ so that δj(z) = p(z) p′(λj )(z − λj ) = Yk6=j z − λk λj − λk . Assume then that we are given a function f mapping a compact M ⊂ C into Cd. It determines a unique function ϕ on K = p−1(M ) if we set ϕ(z) = d Xj=1 δj(z)fj(p(z)) for z ∈ K. We say that ϕ is given on K by a multicentric representation and denote it in short ϕ = Lf. In the reverse direction, suppose we are given a scalar function ϕ on a set K0. Then a necessary condition for f to be determined uniquely is that K0 is balanced w.r.t. Λ in the following sense: K0 = p−1(p(K0)). We shall assume throughout that K0 ⊂ K = p−1(M ) is such that p(K0) = M . Assuming that K is balanced and contains no critical points, then the function f is pointwisely uniquely determined by the values of ϕ. In order to write down a formula we agree about some additional notation. Denote the roots of p(z) − w = 0 by zj = zj(w). Away from critical values these are analytic and we assume a fixed numbering so that zj(w) → λj if z1(w) → λ1 (when w → 0). In the inversion we essentially exchange interpolation and evaluation points. To that end let δj(ζ; w) denote the interpolation polynomial, with w fixed, which takes the value 1 at ζ = zj(w) while vanishing at other zk(w)'s: δj(ζ; w) = p(ζ) − w p′(zj(w))(ζ − zj(w)) , (2.1) so that in particular δj(ζ; 0) = δj(ζ). 3 Proposition 2.1. Suppose K is a balanced compact set with respect to local centers Λ. Assume that ϕ is given pointwisely in K. Then f is uniquely defined for all noncritical values w ∈ M \ p(Λ1) by fk(w) = d Xj=1 δj(λk; w)ϕ(zj (w)). (2.2) The functions fk inherit the smoothness of ϕ, and additionally, if λc ∈ Λ1 is an interior point of K and ϕ is at that point analytic, then the singularities of each fk at the critical value p(λc) are removable. Proof. See the discussions in [9] and [10]. So, we could use the expression f = L−1ϕ at least when the components of f are determined by (2.2) for noncritical values w provided ϕ is given in a balanced set. In particular this is natural when ϕ is analytic in a balanced domain. However, the topic of this paper is in functions which are perhaps given only on discrete sets, such as the set of eigenvalues of a matrix and then some extra care is needed in considering the possible lack of injectivity of L. We shall therefore build a Banach algebra and view L as performing the Gelfand transformation f = Lf . We then get many general properties of Gelfand transform to be transported into our situation with relatively small amount of work. 2.2 Multiplication of the vector functions: polyproduct Consider now continuous functions f mapping M into Cd. We are aiming to define a Banach algebra structure into C(M )d. Denoting by ⊚ the multiplication in C(M )d we then want that L takes the vector functions into scalar functions in such a way that L becomes an algebra homomorphism L(f ⊚ g) = (Lf )(Lg) Since Pd where the multiplication of scalar functions Lf is pointwise. j=1 δj(z) = 1 the constant vector 1 = (1, . . . , 1)t ∈ Cd serves as the unit in the algebra. In order to define f ⊚ g we hence need to code the differences between components of f . Definition 2.2. For a ∈ Cd we set (cid:3) : a 7→ (cid:3)a =  0  a2 − a1 ad − a1 ad − ad−1 a1 − ad a2 − ad . . . 0   . . . a1 − a2 0 . . . . . . . . . . . . . . . and call it boxing the vector a. In order to define the product we still need to introduce two "scaling" entities, matrix L and vector ℓ. The matrix L has zero diagonal and Lij = 1/(λi − λj ) for i 6= j, while the vector ℓ ∈ Cd has components ℓj = 1/p′(λj). Now, denoting by ◦ the Hadamard (or Schur, elementwise) product we can define the product as follows. 4 Definition 2.3. Let f and g be pointwisely defined functions from M ⊂ C into Cd. Then their "polyproduct" f ⊚ g is a function defined on M , taking values in Cd such that (f ⊚ g)(w) = (f ◦ g)(w) − w (L ◦ (cid:3)f (w) ◦ (cid:3)g(w) )ℓ. Remark 2.4. We shall write this in short, with slight abuse of notation, as f ⊚ g = f ◦ g − w (L ◦ (cid:3)f ◦ (cid:3)g )ℓ. Further, for the powers we write f n = f ⊚ f n−1 and the inverse in particular as f −1 whenever it exists: f ⊚ f −1 = 1. Proposition 2.5. The vector space of functions f : M ⊂ C → Cd equipped with the product ⊚ becomes a complex commutative algebra with 1 as the unit. Proof. In addition to the obvious properties of scalar multiplication and summation of vectors we observe that the vector product is commutative f ⊚ g = g ⊚ f and since (cid:3)1 = 0 we have 1 ⊚ f = f. Further, since (cid:3)(αf + βg) = α(cid:3)f + β(cid:3)g, we get (αf + βg) ⊚ h = α(f ⊚ h) + β(g ⊚ h). These are enough for the structure to be an algebra. Theorem 2.6. Let f and g be defined in M and K = p−1(M ). Then if ϕ and ψ are functions defined on K by ϕ = Lf and ψ = Lg, then ϕψ is given by Proof. When we multilply ϕ and ψ products δiδj appear. For writing the expressions in a simple form we introduce ϕψ = L(f ⊚ g). σij = 1 1 p′(λj ) λi − λj . Lemma 2.7. We have with w = p(z) δ2 i (z) = δi(z) − wXj6=i [σij δi(z) + σjiδj(z)] while for i 6= j δi(z)δj(z) = w [σij δi(z) + σjiδj(z)]. (2.3) (2.4) (2.5) Proof of the lemma. Let first i 6= j. Since δi = and p(z) = w we can write p p′(λi)(z − λi) δiδj = w δj p′(λi) z − λi . 5 But δj/(z − λi) is a polynomial of degree d − 2 and can thus be written as a linear combination in these basis polynomials. This gives δj z − λi = 1 λj − λi δj + p′(λi) p′(λj )(λi − λj) δi which then yields (2.5). Consider then (2.4). Since Pj δj = 1 we can write δi = 1 −Pj6=i δj to get δ2 i = δi −Xj6=i δiδj, which, with the help of (2.5), yields the claim and completes the proof of the lemma. We can now multiply the expressions for ϕ and ψ. δifiδjgj δ2 ϕψ = Xi,j = Xi = Xi + wXi Xj6=i i figi +Xi Xj6=i δifigi − wXi Xj6=i [σij δi(z) + σjiδj(z)]figj. δiδjfigj [σij δi(z) + σjiδj(z)]figi (2.6) (2.7) (2.8) (2.9) Here the term multiplying δk appears in the form δkfkgk − wXj6=k σkj (fk − fj)(gk − gj) and hence the whole expression reads ϕψ = Xi δi [figi − wXj6=i σij (fi − fj)(gi − gj)]. This is easily seen to be of the form ϕψ = L(f ⊚ g) which completes the proof of the theorem. 2.3 The norm in the algebra We shall be considering continuous functions f from a compact M ⊂ C into Cd and begin with the uniform norm fM = maxw∈M f (w)∞ where a∞ = max1≤j≤d aj. The definition of polyproduct makes this into an algebra, but · M is not an algebra norm in general, so we need to move into the "operator norm". Definition 2.8. For f ∈ C(M )d we set kfk = sup gM ≤1f ⊚ gM . This is clearly a norm in C(M )d and it is in fact equivalent with · M . 6 Proposition 2.9. There is a constant C, only depending on M and on Λ such that kf ⊚ gk ≤ kfkkgk fM ≤ kfk ≤ CfM . (2.10) (2.11) Proof. In fact On the other hand, from the definition of the polyproduct it is clear that there exists a constant C such that fM = f ⊚ 1M ≤ kfk. f ⊚ gM ≤ CfM gM . kfk = sup gM ≤1f ⊚ gM ≤ CfM . f ⊚ g ⊚ hM ≤ kfk g ⊚ hM ≤ kfk??kgk?hM But then Finally, implies (2.10). Since the polyproduct ⊚ is uniquely determined by Λ, we shall denote the algebra in short as CΛ(M ). Definition 2.10. The vector space C(M )d of continuous functions f from a compact M ⊂ C into Cd, with the operator norm kfk and product ⊚ is denoted by CΛ(M ). The discussion can be summarized as follows. Theorem 2.11. The Banach space C(M )d equipped with polyproduct ⊚, and denoted by CΛ(M ), is a commutative unital Banach algebra. The algebra-norm k·k is equiva- lent with ·M and functions with components given by polynomials p(w, w) are dense in CΛ(M ). Proof. Recall that polynomials p(w, w) are dense in the sup-norm on a compact M ⊂ C among continuous functions by Stone-Weierstrass theorem. Applying this on each component of functions f ∈ CΛ(M ) gives the result. 2.4 Characters of CΛ(M) In order to be able to apply the Gelfand theory we need to know all characters in the algebra CΛ(M ). Definition 2.12. A nontrivial linear functional χ : CΛ(M ) → C is called a character if it is additionally multiplicative: χ(f ⊚ g) = χ(f )χ(g). The set of all characters is the character space, which we denote here by X . Remark 2.13. In commutative unital Banach algebras all characters - complex ho- momorphisms - are automatically bounded and of norm 1. Since maximal ideals are kernels of characters, the focus is sometimes on the maximal ideals rather than on the characters, [1], [2], [3], [13]. 7 Because the polyproduct ⊚ is constructed to yield L(f ⊚ g) = LfLg, we conclude immediately that for each fixed z0 ∈ p−1(M ) the functional χz0 : f 7→ d Xj=1 δj(z0)fj(p(z0)) (2.12) is a character. We show next that there are no others. Theorem 2.14. The character space X is X = {χz : z ∈ p−1(M )} where χz is given in (2.12). Proof. We need to show that all characters are of the form (2.12). Let χ ∈ X be given and apply it within the subalgebra consisting of elements of the form f = α1, where α is a scalar function α ∈ C(M ). Now, it is well known that all multiplicative functionals in C(M ) are given by evaluations at some w0 ∈ M : α 7→ α(w0); hence χ(α1) = α(w0) for some w0 ∈ M . Next, take an arbitrary g ∈ CΛ(M ). Then we conclude from χ(α1 ⊚ g) = α(w0)χ(g) that χ(g) depends on g(w0), only. In fact χ(α1 ⊚ g) = χ(αg) = α(w0)χ(g) and if α(w0) = 1 we have χ(g) = χ(αg) + χ((1 − α)g) so that χ((1 − α)g) = 0. We assume next that w0 is chosen and χ is a character f 7→ χ(f ) such that the value only depends on f (w0). We may therefore view χ as an arbitrary linear functional in Cd which is multiplicative with respect to the polyproduct ⊚ at w0. In fact, setting for a, b ∈ Cd ab = (a ⊚ b)(w0) makes Cd into a Banach algebra, for each fixed w0. Let a, b ∈ Cd, then χ is of the form χ(a) = d Xj=1 ηjaj and we require χ((a ⊚ b)(w0)) = χ(a)χ(b). First observe that χ(1) = 1 gives Pd and using the notation in the proof of it, we see that we must have j=1 ηj = 1. Then, comparing with Lemma 2.7 while for j 6= i η2 i = ηi − w0Xj6=i (σij ηi + σjiηj) ηiηj = w0(σij ηi + σjiηj). 8 (2.13) (2.14) ηj = 1. But for w0 = p(λi) = 0 and thus ηj = δj(λj) = 1 and there are exactly d different solutions. Let first w0 = 0. We have then ηi ∈ {0, 1} and, since Pi ηi = 1, exactly one For w0 6= 0 we have from (2.14) that ηi 6= 0 for all i. We take η1 as an unknown so that for j > 1 ηj = w0σ1j η1 η1 − w0σj1 . Substituting these into (2.13) and dividing with η1 6= 0 yields σj1σ1j η1 = 1 − w0Xj6=1 σ1j − w2 0 Xj6=1 . η1 − w0σj1 This has, counting multiplicities, exactly d solutions for η1. However, we already know d solutions, namely, δ1(zk(w0)), for k = 1,··· , d where p(zk(w0)) = w0, which completes the proof. 2.5 Gelfand transform and the spectrum When ϕ is holomorphic it is natural to think ϕ as the "primary" function which is represented or parametrized by the vector function f . However, when dealing with functions with less smoothness it is easier to think their roles to be reversed. This is because f can be taken as any continuous vector function while the behavior of ϕ is in general complicated near critical points. We take CΛ(M ) as the defining algebra while the functions ϕ appear as Gelfand transforms. Before applying this machinery we recall some basic properties of Gelfand theory. Let A be a commutative unital Banach algebra with unit e and denote by h a character: h(ab) = h(a)h(b) for all a, b ∈ A. Let us denote by ΣA the character space of A. Then every h ∈ ΣA has norm 1 and ΣA is compact in the Gelfand topology: one gives ΣA the relative weak∗-topology it has as a subset of the dual of A. Then the Gelfand transform of a ∈ A is a : ΣA → C where a(h) = h(a). The function a is then always continuous in the Gelfand topology and this allows one to study the algebra A by studying continuous functions on ΣA. Since every maximal ideal of A is of the form Nh = {a ∈ A : h(a) = 0}, the character space is sometimes called the maximal ideal space of A. We collect here basic facts on the Gelfand theory, and here we treat a ∈ C(ΣA) as a continuous function with the sup-norm. Recall that the spectrum σ(a) of an element a ∈ A consists of those λ ∈ C for which λe − a does not have an inverse in A. We denote by ρ(a) the spectral radius of a: ρ(a) = max{λ : λ ∈ σ(a)}. Theorem 2.15. (Gelfand representation theorem) Let A be a commutative unital Ba- nach algebra. Then for all a ∈ A (i) σ(a) = a(ΣA) = {a(h) : h ∈ ΣA}; 9 (ii) ρ(a) = kak∞ = limn→∞ kank1/n ≤ kak; (iii) a ∈ A has an inverse if and only if a(h) 6= 0 for all h ∈ ΣA; (iv) rad A = {a ∈ A : a(h) = 0 for all h ∈ ΣA}. (See any text book treating Banach algebras, e.g. [1], [2], [3], [13]). We shall now consider CΛ(M ). In what follows we write f ⊚ f n−1 = f n and in particular f −1 for the inverse of f . Recall that we denote by X the character space of CΛ(M ) where X = {χz : z ∈ p−1(M )} χz(f ) = d Xj=1 δj(z)fj(p(z)). This allows us to identify χz with z and consequently X with p−1(M ). Hence we shall view the Gelfand transform f as a function of z ∈ p−1(M ). Definition 2.16. Given f ∈ CΛ(M ) we set f : p−1(M ) → C f : z 7→ f (z) = d Xj=1 δj(z)fj(p(z)). Thus, we can view the multicentric representation operator L as performing the Gelfand transformation L : f 7→ f . We denote this Gelfand transformation by L to remind that for constant vectors a ∈ Cd the transformation a is just the Lagrange interpolation polynomial (restricted into p−1(M )). We denote fK = supz∈K f (z). We specify now the general Gelfand representation theorem for the algebra CΛ(M ). Theorem 2.17. (Multicentric representation as Gelfand transform) For f ∈ CΛ(M ) the following hold with K = p−1(M ): (i) σ(f ) = { f (z) : z ∈ K}; (ii) ρ(f ) = fK = limn→∞ kf nk1/n ≤ kfk; (iii) f has an inverse if and only if f (z) 6= 0 for all z ∈ K; (iv) rad CΛ(M ) = {f ∈ CΛ(M ) : f (z) = 0 for all z ∈ K}. Recall, that an algebra A is called semi-simple if rad A = {0}. Theorem 2.18. CΛ(M ) is semi-simple if and only if M contains no isolated critical values of p. 10 Proof. Take f ∈ rad(CΛ(M )) so that f (z) = 0 for all z ∈ K. Since f determines f uniquely outside critical values, we have f (w) = 0 away from the critical values. If every critical value is an accumulation point of M then by continuity f vanishes everywhere. On the other hand, if w0 ∈ M is an isolated critical value, take critical points zi ∈ p−1({w0}). By assumption, they are isolated and all we need to do is to find 0 6= a ∈ Cd such that Pj δj(zi)aj = 0 for all i. However, since w0 is a critical value at least two of the roots zi coincide and hence the matrix (δj(zi))ij is not of full rank. So, we conclude that nontrivial solutions f exist and CΛ(M ) is not semi-simple. Remark 2.19. If sA is a simplifying polynomial of minimal degree for an n× n matrix A (see Definition 3.1), then all critical values of sA are isolated and inside σ(sA(A)). 2.6 Invertible elements of CΛ(M) From Theorem 2.17 conclude that if ϕ is given by multicentric representation ϕ = Lf where f is continuous and bounded, then 1/ϕ = Lg with a bounded and continuous g if and only if ϕ(z) 6= 0 for z ∈ p−1(M ). We shall now derive a quantitative version of this. Theorem 2.20. There exists a constant C depending on M and Λ such that the follow- ing holds. If f ∈ CΛ(M ) is such that for all z ∈ p−1(M ) Lf (z) ≥ η > 0, then there exists g ∈ CΛ(M ) such that f ⊚ g = 1 and kgk ≤ C kfkd−1 ηd . (2.15) Before turning to prove this we look at an instructive example. Example 2.21. We shall first consider the degree two case with w = z2 − 1. If we put ϕ(z) = Lf (z), then the inverse g = f −1 is given simply as follows: (cid:18)g1(w) g2(w)(cid:19) = 1 ϕ(z)ϕ(−z) (cid:18)f2(w) f1(w)(cid:19) . g ⊚ f = g ◦ f + w 4 (g1 − g2)(f1 − f2)1 (2.16) In fact, since we have g ⊚ f = 1 ϕ(z)ϕ(−z) and then expanding ϕ(z)ϕ(−z) we obtain (f1f2 − w 4 (f1 − f2)2)1 [ 1 + z 2 f1 + 1 − z 2 f2][ 1 − z 2 f1 + 1 + z 2 f2] = f1f2 − w 4 (f1 − f2)2. In particular, the constant C in (2.15) equals 1 in this case. 11 Proof. The example suggests to look at the inverse in the following way. Denote by zj = zj(w) the roots of p(ζ) − w = 0 and when needed, we put z1(p(z)) = z. With ϕ = Lf and ψ = 1/ϕ = Lg we then have 1 d ψ(z1) = ϕ(zj ) Φ(w) Yj=2 where Φ(w) = Qd j=1 ϕ(zj). We need the following lemma. Lemma 2.22. Suppose that f has analytic components in M . Then Φ : w 7→ d Yj=1 ϕ(zj(w)) (2.17) is analytic in M . Proof of lemma. All roots zj(w) are analytic except possibly at critical values. Since ϕ(zj(w)) is given by ϕ(zj(w)) = d Xk=1 δk(zj(w))fk(w) with fk's analytic, we may as well assume that fk's are constants as the only source for lack of analyticity at the critical values would come from products of δk's. But if fk's are constants, we may put q(ζ) = Pd k=1 fkδk(ζ). However, then p(ζ1,··· , ζd) = d Yj=1 q(ζj) is a symmetric polynomial and it can be expressed uniquely by elementary polynomials si by Newton's theorem. If we now substitute ζj = zj(w), where zj(w)'s are the roots of p(ζ)−w = 0, we observe that all elementary polynomials si except sd are constants. For example, s1 = −Pd j=1 λj, while sd(w) = (−1)d(p(0) − w). Thus we arrive at a polynomial in w, which completes the proof of the lemma. j=1 zj(w) = −Pd Next we concentrate on Φ(w)ψ(z) = d d Yj=2 Xk=1 δk(zj)fk(w) and organize this as a sum of the form δ1(z2)··· δ1(zd) f1(w)d−1 + ··· = Xα=d−1 qα(z)Fα(w), where α = (α1,··· , αd) and Fα(w) = Qd k=1 fk(w)αk while qα(z) is a rather com- plicated sum of products of different δk's evaluated at zj's with j > 1. Treating zj = zj(p(z)) as functions of z, qα(z) are clearly analytic away from the critical points z ∈ Λ1. Individual products of δk(zj)'s within qα(z) may have branch points at these critical points while the sum qα(z) itself is however a polynomial. To see this, 12 let fk(w) = xk be constants and denote x = (x1,··· , xd)t ∈ Cd. Then Fα(w) = xα and if we put P (z, x) = Xα=d−1 qα(z)xα, then we can view P (z, x) as a polynomial in C × Cd. In fact, Φ(p(z)) is a polynomial and ϕ(z) divides it so P (z, x) must be a polynomial in z. But then, for example by differentiating P (z, x) with ∂α = Q( ∂ )αk gives ∂αP (z, x) = α!qα(z) showing that each qα is a polynomial in z. ∂xk Finally write qα(z) = Pd j=1 δj(z)Qα,j(w) so that Φ(w)ψ(z) = d Xj=1 δj(z) Xα=d−1 Fα(w)Qα,j(w) and, written in CΛ(M ), g = Pα=d−1 FαQα/Φ. Since Φ(w) ≥ ηd and Fα(w) ≤ kfkd−1 the claim follows with C = Pα=d−1 kQαk. 2.7 Characteristic function, resolvent estimates and nilpotent ele- ments From the previous discussion we see that f is invertible in the algebra if and only if Φ does not vanish. This suggests to introduce a characteristic function for f . This gives still another view to the algebra. Denoting again by zj(w) the roots of p(z) − w = 0 we have λ ∈ σ(f ) = { f (z) : j=1(λ − f (zj(w))) = 0 at some w ∈ M . z ∈ K = p−1(M )} if and only if Qd Expanding the product as a polynomial in λ takes the form πf (λ, w) = d Yj=1 (λ − f (zj(w))) = λd − Φ1(w)λd−1 + ··· + (−1)dΦd(w), (2.18) since the coefficient functions Φj are again functions of w by the same argument as in Lemma 2.22; notice that Φd equals the Φ in (2.17). Definition 2.23. Given f ∈ CΛ(M ) we call πf (λ, w) the characteristic function of f . Further, we call (λ1 − f )−1 the resolvent element whenever it exists. This allows us to formulate a different version of the estimate for the inversion. To that end we denote by · ∞ the max-norm in Cd. Theorem 2.24. There exists a constant C, depending on M and on Λ, such that for w ∈ M (λ1 − f )−1(w)∞ ≤ C (λ + kfk)d−1 . πf (λ, w) Proof. This follows in an obvious way from Theorem 2.20 and from the definitions. Remark 2.25. From ρ(f ) = fK ≤ kfk we have the lower bound λ − f K ≤ k(λ1 − f )−1k. dist(λ, σ(f )) = 1 1 (2.19) 13 This in particular implies that if f 6= λ1 and λ ∈ ∂σ(f ), there exists gn ∈ CΛ(M ) of unit length such that (λ1 − f ) ⊚ gn → 0. In other words, λ1 − f is a topological divisor of zero. We noted earlier that f ∈ rad CΛ(M ) if and only if f vanishes identically, or , which is the same thing, σ(f ) = {0}. Proposition 2.26. If σ(f ) = {0}, then f is nilpotent and there exists n ≤ d such that f n = 0. Proof. In other words, we need to show that all quasinilpotent elements are actually nilpotent. It is clear from Theorem 2.18 that nontrivial quasinilpotent elements exist when M contains an isolated critical value, say w0. We can proceed now as follows. We view the multiplication as an operator in C(M, Cd) and hence for each w ∈ M there is a matrix Bf (w) such that f : g 7→ f ⊚ g (f ⊚ g)(w) = Bf (w)g(w). If f is quasinilpotent, it means that each Bf (w) must be for fixed w quasinilpotent. However, a d × d-matrix is quasinilpotent only when it is nilpotent, from which the claim follows. Example 2.27. Consider w = z2 − 1. Put h = w Bf = (cid:18)f1 + h h 4 (f1 − f2). Then f2 − h(cid:19) −h which for fixed w has the eigenvalues 1 2 (f1(w) + f2(w)) ± √1 + w 2 (f1(w) − f2(w)). That is, the eigenvalues are simply ϕ(z) and ϕ(−z). Denote δ2(−z)(cid:19) 2z (cid:18)z + 1 z − 1 z − 1 z + 1(cid:19) . E(z) = (cid:18) δ1(z) δ1(−z) so that for z 6= 0 E(z)−1 = δ2(z) 1 Finally, (cid:18)ϕ(z) 0 0 ϕ(−z)(cid:19) = E(z)Bf (w)E(z)−1 and we see that the eigenvectors are independent of the function f . At z = 0 the eigenvalues agree, and E(0) is no longer invertible. Put f (−1) = (1,−1)t so that ϕ(0) = 0. Then Bf (−1) = 1 1 2 (cid:18) 1 −1 −1(cid:19) is similar to (cid:18)0 0 1 0(cid:19) . 14 2.8 Quotient algebra CΛ(M)/IK0 When we apply the functional calculus, discussed in the next section, the natural re- quirement for ϕ is that it is well defined at the spectrum σ(A) of the operator A, which means that f representing ϕ must be well defined on a set which includes p(σ(A)). However, p−1(p(σ(A)) is likely to be properly larger than σ(A) which in practice shows up in lack of uniqueness in representing ϕ. assume here that the inclusion K0 ⊂ K is proper. Let K0 ⊂ C be compact, put p(K0) = M and denote as before K = p−1(M ). We Let IK0 be the closed ideal in CΛ(M ) IK0 = {f ∈ CΛ(M ) : f (z) = 0? for z ∈ K0}. Then the set of elements we are dealing with can be identified with the cosets [f ] : This is a unital Banach algebra with norm defined as CΛ(M )/IK0 = {[f ] : [f ] = f + IK0}. k? [f ] k = inf g∈IK0 kf + gk. We need to identify the character space of this quotient algebra. Definition 2.28. Given a closed ideal J ⊂ A the hull of the ideal is the set of all characters which vanish at every element in the ideal. Lemma 2.29. (Theorem 6.2 in [5]) Given a closed ideal J in a commutative Banach algebra A, the character space of the quotient algebra A/J is the hull of J. Corollary 2.30. The quotient algebra CΛ(M )/IK0 is a Banach algebra with unit and the character space can be identified with K0, so that the Gelfand transformation becomes [f ] 7→ fK0. 2.9 Additional remarks on LCΛ(M) Here we make some observations on the range of the Gelfand transformation. Denoting LCΛ(M ) = {ϕ ∈ C(K) : ∃f ∈ CΛ(M ) such that ϕ = f} we clearly have a normed subalgebra of C(K) with the sup-norm on K = p−1(M ) but the algebra need not be closed. Example 2.31. Let Λ = {−1, 1} so that p(x) = x2−1, and M = [−1, 0] = {x : −1 ≤ x ≤ 0} so that K = p−1(M ) = [−1, 1]. Then LCΛ(M ) contains all polynomials as any polynomial Q(x) can uniquely be written as Q(x) = d Xj=1 δj(x)Qj(p(x)) where Qj's are polynomials. Now, polynomials are dense in C(K) and we conclude that the closure of LCΛ(M ) equals C(K) in this case. However, if we take ϕ ∈ C(K) such that ϕ(x) = max{xα, 0} then for 0 < α < 1 we have ϕ ∈ C(K) \ LCΛ(M ). In fact, for x 6= 0 we have ϕ(x) = Lf (x) with f (x2 − 1) becoming unbounded as x tends to 0. Note, that in this example the Gelfand transformation is injective. 15 Example 2.32. Let Λ = {−1, 1} but K = Λ1 = {0}. Then CΛ({−1}) is a two- dimensional complex algebra, with nontrivial radical consisting of vectors f such that f1(−1) + f2(−1) = 0. On the other hand LCΛ({−1}) is one-dimensional, closed and isomorphic with the complex field. Example 2.33. Let Λ = {−1, 1} and K = {z : ε ≤ z ≤ 2} with some small positive ε. Then the critical point, the origin, is not in K and the following hold with some constant C However, if Lf (x + iy) = xα with 0 < α < 1 for x > 0 and vanishing on the left half plane , then kLfk∞ ≤ kfk ≤ C kLfk∞. kfk ∼ Const/ε1−α. It is natural to ask whether ϕ ∈ LCΛ(M ) shall be differentiable at the interior critical points. After all, we shall be able to apply the functional calculus in such a case for matrices which do have a nontrivial Jordan block and we usually assume that the value on the off-diagonal would be the derivative of ϕ at the eigenvalue in question. Example 2.34. Let again p(z) = z2 − 1 but K such that it contains the critical point in the interior: K = {z : p(z) ≤ 2}. Then M likewise contains a neighborhood of -1. We have ϕ(z) = [f1(w) + f2(w)] + If ϕ ∈ LCΛ(M ), then fi ∈ C(M ) and we have 1 2 z 2 [f1(w) − f2(w)]. 1 2z and hence the limit [ϕ(z) − ϕ(−z)] = [f1(z2 − 1) − f2(z2 − 1)] 1 2 lim z→0 1 2z [ϕ(z) − ϕ(−z)] = 1 2 [f1(−1) − f2(−1)] always exists. However, it does not imply that ϕ would be differentiable at the origin. In fact, we have 1 [ϕ(z) − ϕ(0)] z 1 2z{[f1(z2 − 1) + f2(z2 − 1)] − [f1(−1) + f2(−1)] 1 [f1(z2 − 1) − f2(z2 − 1)]. 2 = + (2.20) (2.21) (2.22) Here the last term is continuos as z tends to origin. Thus the derivative exists depending on the behavior of f1 + f2 near w = −1. In particular, if f1 + f2 is Hölder continuous with exponent >1/2, then ϕ is differentiable. 3 Functional calculi 3.1 Functional calculus for matrices We discuss first the functional calculus related to CΛ(M ) for matrices. Denote by Mn complex n × n-matrices with the norm kAk = sup x2=1Ax2. 16 Further, we denote by σ(A) = {αk} the eigenvalues of A and by mA the minimal polynomial of A, that is, the monic polynomial q of smallest degree such that q(A) = 0: mA(z) = m Yk=1 (z − αk)nk+1. As mentioned in the introduction, the usual way to formulate the class of functions ϕ for which ϕ(A) is well defined, asks the following to be known at every eigenvalue αk ϕ(αk),··· , ϕ(nk)(αk), [6], [11]. Based on this information one can then construct an Hermite interpolation polynomial p and set ϕ(A) = p(A). As we saw in Example 2.34 the functions in our algebra do not need to be differ- entiable - but of course when they are the resulting functional calculus yields the same matrices ϕ(A). Definition 3.1. Given A ∈ Mn we call all monic polynomials p such that p(A) is similar to a diagonal matrix as simplifying polynomials for A. If K denotes those indices k for which nk > 0 in the minimal polynomial, then setting sA(z) = Z z 0 Yk∈K (ζ − αk)nk dζ + c we have a polynomial of minimal degree such that s(j) A (αk) = 0 for j = 1,··· , nk and k ∈ K. Clearly then sA(A) is similar to the diagonal matrix diag(sA(αk)). Since we can add an arbitrary constant to sA we may assume as well that sA has distinct roots. Let now p be a simplifying polynomial for A with distinct roots and assume ϕ is given on σ(A) as ϕ(z) = d Xj=1 δj(z)fj(p(z)). Denoting B = p(A) we could then define for fj ∈ C(σ(B)) the matrix function fj(B) either by Lagrange interpolation at p(αk) or by assuming the similarity transformation to the diagonal form B = T DT −1 be given and setting fj(B) = T fj(D)T −1, both yielding the same matrix fj(B) which commute with A. Then the following matrix is well defined: d ϕ(A) = Xj=1 δj(A)fj (B). It follows immediately that if we have two functions f, g ∈ CΛ(σ(B)), and we denote ϕ = Lf , ψ = Lg and ϕψ = L(f ⊚ g), then this definition yields (ϕψ)(A) = ϕ(A)ψ(A). However, we formulate the exact statement using a different notation to underline the fact that knowing the values of ϕ at the spectrum of A need not determine f uniquely, and hence not ϕ(A), either. 17 Definition 3.2. Assume p is a simplifyng polynomial for A ∈ Mn with disting roots Λ. Then we denote by χA the mapping CΛ(p(σ(A))) → Mn given by f 7→ χA(f ) = d Xj=1 δj(A)fj(B). (3.1) Theorem 3.3. The mapping χA is a continuous homomorphism CΛ(p(σ(A))) → Mn. Proof. That χA is a homomorphism is build in the construction and in particular we have χA(f ⊚ g) = χA(f ) χA(g). The continuity of χA is seen from kχA(f )k ≤ d Xj=1 kδj(A)k kfj(B)k combined with kfj(B)k ≤ κ(T )fσ(p(A)) and with fσ(p(A)) ≤ f, see Proposition 2.9. Here κ(T ) = kTk kT −1k denotes the condition number of the diagonalizing similarity transformation. We can now also conclude that we can formulate a spectral mapping theorem. Let M = p(σ(A)) and, as it would likely to be the case, σ(A) is a proper subset of p−1(M ). Then it follows from Corollary 2.30 that the spectrum of [f ] in CΛ(M )/Iσ(A) is σ([f ]) = { f (z) : z ∈ σ(A)}. Theorem 3.4. We have for [f ] ∈ CΛ(p(σ(A)))/Iσ(A) and χA(f ) ∈ Mn σ(χA(f )) = σ([f ]). Proof. Even so the statement may look rather complicated the proof here can be re- duced to the standard spectral mapping theorem for polynomials. However, the state- ment holds as such in more general setting and then in particular the present simple proof is not available. Consider fj(B) where B = p(A) and denote by βi the eigenvalues of B. There are in general s ≤ m different eigenvalues of B. Let qj be the polynomial of degree s − 1 such that (3.2) qj(βi) = fj(βi) for i = 1,··· , s. Then we set fj(B) = qj(B). Thus we have χA(f ) = P (A) (3.3) if we set P (z) = Pd j=1 δj(z)qj(p(z)). The conclusion follows as P is a polynomial. Remark 3.5. There are two different steps to be taken when consructing χA(f ). (i) Given A ∈ Mn one could for example compute the Schur decomposition of A. From there one must decide what diagonal elements are to be considered as the same and based on that one chooses a simplifying polynomial p such that it has simple roots. Notice in particular that then the eigenvalues αk for which nk > 0, are distinct from the roots λj of p. 18 (ii) Given f one then computes the Lagrange interpolating polynomials qj(w) sat- isfying (3.2) for each j. Then χA(f ) is given by (3.3). Remark 3.6. It is natural to ask how this approach is different from the definition based on Hermite interpolation on the spectrum of A. Consider the minimal polynomial mA as simplifying polynomial. In the Hermite interpolation one interpolates at the eigenvalues while we add a constant c so that he polynomial p(z) = mA(z) + c has simple roots. The effect on the differentiability requirement on f and/or ϕ is then removed and replaced by a balanced limiting behavior of the roots of p near its critical points - and this happens automatically, independent of the function f as long as it is continuous. To illustrate this, suppose f is holomorphic and ϕ = Lf so that ϕ′(z) = d Xj=1 [δ′ j(z)fj(p(z)) + δj(z)f ′ j(p(z))p′(z)]. However, at critical points zc we have ϕ′(zc) = Pd j(zc)fj(p(zc)) so this value does not depend on whether f is differentiable at critical values or not. See also Exam- ple 2.34. j=1 δ′ 3.2 Polynomially normal operators in Hilbert spaces We shall now consider bounded operators A in complex Hilbert spaces H. The operator norm of A ∈ B(H) is denoted by kAk. Definition 3.7. We call A ∈ B(H) polynomially normal, if there exists a noncon- stant monic polynomial p such that p(A) is normal. The polynomial p is then called a simplifying polynomial for A. Polynomially normal operators have been discussed in [4], [7], as operator valued roots for polynomial equations p(z)− N = 0 with N normal. We formulate a structure result (see Theorem 3.1, in [7], also Theorem 2 in [8] ). Theorem 3.8. Let H be separable and A ∈ B(H) such that p(A) is normal for some nonconstant polynomial p. Then there exist reducing subspaces {Hn}∞ n=0 for A, such that H = ⊕∞ n=0Hn and AH0 is algebraic while AHn are for n ≥ 1 similar to normal. We could take use of this structure result but proceed independently of it. We start by assuming that p(A) is normal and then comment the straightforward extension to the case where p(A) is similar to normal. Let N = p(A) be normal, and as before, we may assume that p has simple roots. Then the first task is to define fj(N ) in a consistent way. Recall the following two results, see e.g. [2]. Lemma 3.9. Let M ⊂ C be compact. Then the closure of polynomials of the form q(w, w) in the uniform norm over M equals C(M ). Since N commutes with N ∗ the operator q(N, N ∗) is well defined and the follow- ing holds. Lemma 3.10. If N ∈ B(H) is normal, then kq(N, N ∗)?k = max w∈σ(N )q(w, w). 19 Given now a normal operator N and a continuous function fj on σ(N ) one approx- imates fj by a sequence {qj,n} such that fj − qj,n∞ = max w∈σ(N )fj(w) − qj,n(w, w) → 0 and sets fj(N ) = lim n→∞ qj,n(N, N ∗). (3.4) Then fj(N ) ∈ B(H) is normal, with kfj(N )?k = fj∞ ≤ kfk. Definition 3.11. Assume p is a simplifying polynomial for A ∈ B(H) with distinct roots Λ, so that N = p(A) is normal. Then we denote by χA the mapping CΛ(p(σ(A))) → B(H) given by d f 7→ χA(f ) = Xj=1 δj(A)fj(N ). (3.5) Note that here δj(A) and fj(N ) commute. In fact, A commutes with N = p(A) and since N commutes with N ∗ the operator A commutes with N ∗ as well, by Fu- glede's theorem, [2]. We combine the construction into the following theorem. Theorem 3.12. Let A ∈ B(H) and a simplifying polynomial p be given as in Definition 3.11. Then the mapping χA is a continuous homomorphism from CΛ(p(σ(A))) to B(H). In particular, χA(f ⊚ g) = χA(f ) χA(g) and kχA(f )k ≤ Ckfk with C = d Xj=1 kδj(A)k. Remark 3.13. The case of p(A) similar to normal. We can extend the construction above to operators which are similar to polynomially normal ones. In short, assume that A ∈ B(H) is such that there exists a polynomial p and a bounded T with bounded inverse, such that N = T −1p(A)T is normal. Denote V = T −1AT so that N = p(V ) and B = p(A). Then we can define fj(B) = T fj(N )T −1 and again A commutes with fj(B) as Afj(B) = T [V fj(p(V ))]T −1. This allows us to define χA(f ) = T χV (f )T −1 (3.6) and the extension shares all the natural properties. Remark 3.14. Spectral measure. Recall that if N is normal, then there exists (see e.g. Section 12 in [14]) a spectral measure E from the σ-algebra of all Borel sets of σ(N ) into B(H) such that if ϕ is an essentially bounded Borel-measurable function on σ(N ) then ϕ(N ) = Zσ(N ) ϕ dE. This could in an obvious way be used in defining fj(p(A)), thus extending the func- tional calculus even further. 20 3.3 Spectral mapping theorem for operators If A ∈ B(H) is such that p(A) is similar to normal, then we have χA(f ) = T χV (f )T −1 and therefore χA(f ) and χV (f ) have the same spectrum. Therefore we may as well assume that A is polynomially normal. Theorem 3.15. Suppose p has simple zeros and A ∈ B(H) is such that p(A) is normal. Then for all [f ] ∈ CΛ(p(σ(A)))/Iσ(A) we have σ(χA(f )) = σ([f ]). Proof. Recall that σ([f ]) = { f (z) : z ∈ σ(A)}. Consider first the inclusion f (z) ∈ σ(χA(f )) for all z ∈ σ(A). (3.7) where f is of the form fj(w) = qj (w, w). (3.8) We take a λ ∈ σ(A) and need to show that f (λ) ∈ σ(χA(f )). The discussion splits (3.9) into two as to whether or, if that is not the case, then necessarily, λ ∈ σap(A), Since p(A) is normal we have in both cases p(λ) ∈ σap(p(A)). Assuming (3.9) there exists a sequence of unit vectors xn such that λ ∈ σp(A∗). which by writing p(A) − p(λ) = q(A, λ)(A − λ) implies immediately that (A − λ)xn → 0 (p(A) − p(λ))xn → 0. But then also In fact, if N is normal and N yn → 0, then (p(A) − p(λ))∗xn → 0. (N yn, N yn) = (N ∗yn, N ∗yn) → 0. Denoting p(A) = N and p(λ) = ν we have d χA(f ) − f (λ) = Xj=1 Xj=1 δj(A)[Qj(N, N ∗) − Qj(ν, ν)]+ [δj(A) − δj(λ)]Qj (ν, ν). Operating with these at xn we have [Qj(N, N ∗) − Qj(ν, ν)]xn → 0 21 (3.10) (3.11) (3.12) (3.13) (3.14) since both (N − ν)xn and (N ∗ − ν)xn tend to 0. In fact, there are polynomials R, S of three variables such that we can write Q(N, N ∗) − Q(ν, ν) = [Q(N, N ∗) − Q(ν, N ∗)] + [Q(ν, N ∗) − Q(ν, ν)] = R(N, ν, N ∗)(N − ν) + S(ν, N ∗, ν)(N ∗ − ν) Likewise, by (3.11), [δj(A) − δj(λ)]Qj(ν, ν)xn → 0, and so f (λ) ∈ σap(χA(f )). Next, assume that λ ∈ σp(A∗) and suppose x is an eigenvector such that A∗x = λx. Then clearly However, we also have {[δj(A) − δj(λ)]Qj (ν, ν)}∗x = 0. Qj(N, N ∗)∗x = Qj(ν, ν)x since from A∗x = λx we conclude p(A)∗x = p(λ)x and so N = p(A) being normal this implies N x = νx as well. Substituting these into χA(f )∗ − f (λ) gives [χA(f )∗ − f (λ)]x = 0. Hence f (λ) ∈ σp(χA(f )∗) and so f (λ) ∈ σ(χA(f )). fn approximates f uniformly in σ(A) where fn is of the special form (3.8). We still need to show (3.7) when f is not of the form (3.8). To that end assume that Take µ ∈ f (σ(A)) and we need to show that µ ∈ σ(χA(f )). For some λ ∈ σ(A) we thus have µ = f (λ). Let { fn} be an approximative sequence of the special form (3.8) such that in particular and hence also sup z∈σ(A) f (z) − fn(z) → 0 χA(f ) = lim n χA(fn). Fix an arbitrary open set V such that σ(χA(f )) ⊂ V . We show that µ ∈ V which completes the argument. Fix an open set U such that σ(χA(f )) ⊂ U ⊂ cl(U ) ⊂ V. Since the spectrum is upper semicontinuous (e.g. Theorem 3.4.2 in [1]) there exists an ε > 0 such that σ(B) ⊂ U whenever kχA(f ) − Bk < ε. Let nε be such that kχA(f ) − χA(fn)k < ε for all n ≥ nε. Then σ(χA(fn)) ⊂ U . But for fn we then have Finally, from fn(λ) → f (λ) we conclude that fn(λ) ∈ σ(χA(fn)) ⊂ U. µ = f (λ) ∈ cl(U ) ⊂ V. 22 Consider now the other direction. Here the conclusion follows easily already from Corollary 2.30 with K0 = σ(A). In fact, suppose f (z) 6= 0 for z ∈ σ(A). Then there exists g ∈ CΛ(σ(p(A))) such that f (z)g(z) = 1 for z ∈ σ(A). By Theorem 2.17 we then know that [g] is the inverse of [f ] and since χA is a homomorphism from Cλ(σ(p(A)))/Iσ(A) to B(H), we have χA(f )χA(g) = I and 0 /∈ σ(χA(f )). Thus, if µ ∈ σ(χA(f )), then there must exist λ ∈ σ(A) such that f (λ) − µ = 0. But this simply means that σ(χA(f )) ⊂ f (σ(A)). References [1] B.Aupetit, A Primer on Spectral Theory, Springer 1991 [2] John B. Conway, A Course in Functional Analysis, Second Edition, Springer (1990) [3] Eberhard Kaniuth, A Course in Commutative Banach Algebras, Springer, 2009 [4] S. Foguel, Algebraic functions of normal operators, Israel J. Math., 6 (1968), 199-201 [5] Theodore W. Gamelin, Uniform Algebras, Prentice-Hall, Inc,1969 [6] F.R. Gantmacher, The Theory of Matrices, Volume one. AMS, 1959 [7] Frank Gilfeather, Operator valued roots of Abelian analytic functions, Pacific J. of Mathematics, Vol. 55, No. 1, 1974, 127- 148 [8] Fuad Kittaneh, On the structure of polynomially normal operators, Bull. Austr. Math. Soc. Vol. 30 (1984), 11-18 [9] O. Nevanlinna, Multicentric Holomorphic Calculus, Computational Methods and Function Theory, June 2012, Volume 12, Issue 1, pp 45-65 [10] O.Nevanlinna, Lemniscates and K-spectral sets, J. Funct. Anal. 262 (2012), 1728-1741 [11] Nicholas J. Higham, Functions of matrices: theory and computation. SIAM 2008 [12] T. Ransford, Potentail Theory in the Complex Plane, London Math. Soc. Student Texts 28, Cambridge Univ. Press, 1995 [13] Ch.E.Rickart, General Theory of Banach Algebras, D.Van Nostrand Company, inc. 1960 [14] Walter Rudin, Functional Analysis, McGraw-Hill, 1973 Acknowledgement Much of this work was written during 2014 while the author was a Visiting Fellow at Clare Hall, University of Cambridge. The visit was partially supported by The Osk. Huttunen Foundation. 23
1910.03411	1	1910	2019-10-08T14:17:01	Nonlinear generalised functions on manifolds	[ "math.FA", "gr-qc", "math.DG" ]	This paper lays the foundations for a nonlinear theory of differential geometry that is developed in a subsequent paper which is based on Colombeau algebras of tensor distributions on manifolds. We adopt a new approach and construct a global theory of algebras of generalised functions on manifolds based on the concept of smoothing operators. This produces a generalisation of previous theories in a form which is suitable for applications to differential geometry. The generalised Lie derivative is introduced and shown to commute with the embedding of distributions. It is also shown that the covariant derivative of a generalised scalar field commutes with this embedding at the level of association.	math.FA	math	NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS E. A. NIGSCH AND J. A. VICKERS Abstract. This paper lays the foundations for a nonlinear theory of differential geometry that is developed in a subsequent paper [1] which is based on Colombeau algebras of tensor distributions on manifolds. We adopt a new approach and construct a global theory of algebras of generalised functions on manifolds based on the concept of smoothing operators. This produces a generalisation of previous theories in a form which is suitable for applications to differential geometry. The generalised Lie derivative is introduced and shown to commute with the embedding of distributions. It is also shown that the covariant derivative of a generalised scalar field commutes with this embedding at the level of association. 1. Introduction The classical theory of distributions has proved a very powerful tool in the analysis of linear partial differential equations. However, the fact that in general one cannot multiply distributions makes them of limited use in theories such as general relativity whose underlying equations are inherently nonlinear. Geroch and Traschen [2] identified a class of regular metrics for which the components of the curvature tensor are well defined as distributions and showed that such regular metrics have curvature with singular support on a manifold of co-dimension at most one. Thus, one can describe shells of matter but not strings or particles with metrics in this class. However, by going outside con- ventional distribution theory Colombeau [3] showed that it is possible to construct associative, commutative differential algebras which con- tain the space of distributions as a linear subspace and the space of smooth functions as a subalgebra. Colombeau's theory of generalised functions has therefore increasingly had an important role to play in general relativity, enabling one to use distributions in situations where 2010 Mathematics Subject Classification. 46F30, 46T30. Key words and phrases. nonlinear generalised functions, Colombeau algebra, dif- feomorphism invariant. 1 2 E. A. NIGSCH AND J. A. VICKERS one has ill defined products according to the classical theory, but with- out having to resort to ad hoc regularisation procedures. Applications of Colombeau's theory to general relativity have included the calcula- tion of nonlinear distributional curvatures which correspond to metrics of low differentiability, such as those which occur in space-times with thin cosmic strings [4] and Kerr singularities [5], and the electromag- netic field tensor of the ultra-relativistic Reissner-Nordstrom solution [6]. For a review of applications of Colombeau algebras to general relativity see [7]. The basic idea is to represent generalised functions by families of smooth functions. In the special version of the theory the Colombeau algebra is denoted Gs and the basic space used for its construction con- sists of 1-parameter families (fε)ε∈(0,1] of smooth functions. However, this results in many different representations of what is essentially the same function so that one identifies families which differ by something negligible, i.e., by a family of functions whose derivatives vanish faster than any power of ε on any compact set. This identification is re- alised by factoring out by the set of such functions, but this is not an ideal unless one restricts the basic space to families of moderate func- tions whose derivatives are bounded on compact sets by some positive power of 1/ε. The (special) algebra of generalised functions is therefore defined to be moderate functions modulo negligible functions, see [8] for more details. Despite factoring out by negligible functions, the notion of gener- alised function within Colombeau algebras is finer than that within conventional distribution theory, and it is this feature that enables one to circumvent Schwartz's result on the impossibility of multiplying dis- tributions [9]. Although the pointwise product of smooth functions commutes with the embedding into the algebra, the pointwise prod- uct of continuous functions does not (and indeed this cannot be the case due to the Schwartz impossibility result). However, an important feature of Colombeau algebras is an equivalence relation known as as- sociation which coarse grains the algebra. At the level of association the pointwise product of continuous functions does indeed commute with the embedding. Furthermore, many (but not all) elements of the algebra are associated to conventional distributions. This feature has the advantage that in many cases one may use the mathematical power of the differential algebra to perform classically ill-defined calculations but then use the notion of association to give a physical interpretation to the answer. NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 3 Unfortunately, the special algebra suffers from the disadvantage that there is no canonical embedding of distributions into it. In some sit- uations this is not a problem because some mathematical or physical feature of the problem may be used to define a preferred embedding. However, in general there is no such preferred embedding into the spe- cial algebra, so in section 2 we will briefly describe the full Colombeau algebra G in which the generalised functions are parameterised by el- ements φ of a space of mollifiers Ak. This enables one to define an associative commutative differential algebra on Rn which contains the space of smooth functions as a subalgebra and has a canonical embed- ding of the space of distributions as a linear subspace. Furthermore, the embedding commutes with (distributional) partial derivatives. Within G one also has a notion of association which may be used to give a distributional interpretation to certain generalized functions. This al- gebra was used in [4] to show that the curvature of a cone is associated to a multiple of the delta distribution. Although the full Colombeau algebra on Rn permits a canonical em- bedding of the space of distributions as a linear subspace, this has been bought at the price of giving up manifest coordinate invariance. Indeed, the definition of the spaces Ak of mollifiers which are used to define the algebra is coordinate dependent. One approach to this problem is to regard the use of the Colombeau algebras as a purely intermediate part of the construction. For example, in the case of the cone one starts with the metric in a given coordinate system, calculates the regularised metric and uses this to calculate the curvature density in G. One can then show that the result is associated to a multiple of the delta distribution and that furthermore if one repeats the entire calculation in a different coordinate system the final result is just the transformed delta distribution (see [10] for details). However, there exist situations in which the generalised functions one obtains are not associated to any distribution and in which it is desir- able to have a coordinate invariant generalisation of the full algebra. Such an algebra was first proposed by Colombeau and Meril [11]. Their approach was to give a local description of the algebra together with a transformation law for the generalised functions which ensures that the embedding into the algebra commutes with coordinate transforma- tions. This work suffered from some technical problems but building on these ideas it was shown that one can construct a global Colombeau algebra of generalised functions on manifolds (see [12] for details) re- taining all the distinguishing features of the local theory in the global context. In section 3 we will present a new version of the algebra based on the idea of smoothing operators. This has a larger basic space than 4 E. A. NIGSCH AND J. A. VICKERS [12] which allows us to define a covariant derivative and can therefore be developed into a nonlinear theory of distributional differential ge- ometry [1]. In contrast to the theory on Rn the theory of generalised functions on manifolds involves a number of technical issues involving in particular the theory of differentiation in locally convex spaces. We will not go into the details here, but the approach will be to use the convenient setting of global analysis of [13]. For applications of the algebra to general relativity we are interested in Einstein's equations for metrics of low differentiability. These met- rics are tensorial rather than scalar objects. Because the embedding into the algebra does not commute with multiplication (except on the subalgebra of smooth functions) one cannot simply work with the coor- dinate components of a tensor and use the theory of generalised scalars. In a subsequent paper [1] we show how it is possible to define an alge- bra of generalised tensor fields on a manifold which contains the spaces of smooth tensor fields as a subalgebra and has a canonical coordinate independent embedding of the spaces of tensor distributions as linear subspaces. In order to make the presentation self-contained we begin in this paper by briefly reviewing the Colombeau theory of generalised func- tions on Rn emphasising the structural issues that will be important in generalising this to manifolds. 2. The full Colombeau algebra on Rn In this section we briefly describe the construction of the full Colombeau algebra in Rn (for further details and proofs see [3]). The starting point is the observation that one can smooth functions by taking the (anti)- convolution with a suitable mollifier. Let D(Rn) denote the space of smooth functions on Rn with compact support. We define A0(Rn) to be the set of those φ ∈ D(Rn) which satisfy the normalisation condition ZRn φ(x) dx = 1. Given ε > 0 we set φε(x) = 1 ε(cid:17) , εn φ(cid:16)x so that φε has support scaled by ε and its amplitude adjusted so that its integral is still one. Note that (φε)ε is an example of a net of smooth functions with the delta distribution as its limit in the sense that ε→0ZRn lim φε(x)Ψ(x) dx = Ψ(0) ∀Ψ ∈ D(Rn). NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 5 This is sometimes called a model delta net (see [14]). Provided f ∈ L1 loc (i.e., f is a locally integrable function), for each φ ∈ A0(Rn) we can define a 1-parameter family of smooth functions fε by (1) f (y)φε(y − x) dx fε(x) =ZRn which converges to f in D′(Rn). However, in what follows it will be important to regard φ as well as ε as a parameter so we write expression (1) as ef (φε, x). defined by It will also be convenient to introduce the translation operator τ (τxφ) (y) = φ(y − x) for x, y ∈ Rn and φ ∈ D(Rn). In order to match the notation of the theory on manifolds we will sometimes write φx,ε = τxφε, so that for fixed x, φx,ε is a 1-parameter family of smooth functions converging in D′(Rn) to δx, the delta distribution at x. A distribution T ∈ D′(Rn) is a linear functional on the space of smooth test functions D(Rn) and we may generalise equation (1) to distributions by defining a 1-parameter family of smooth functions Tε by (2) Tε = hT, τxφεi. Again, we will write this expression as T (φε, x). In order to construct the algebra of generalised functions we define a grading on the space of mollifiers in terms of moment conditions. Note that we will throughout use multi-index notation so that i = (i1, . . . in) and xi = xi1 1 . . . xin n . Definition 1. For q ∈ N we define Aq(Rn) to be the set of functions φ ∈ A0(Rn) such that ZRn xiφ(x) dx = 0 ∀i ∈ Nn 0 with i 6 q. We are now in a position to construct the full Colombeau algebra on Rn. Our basic space will be the following. Definition 2. E(Rn) is defined to be the set of all maps F : A0(Rn) × Rn → R (φ, x) 7→ F (φ, x) which for fixed φ are smooth as functions of x. 6 E. A. NIGSCH AND J. A. VICKERS The lack of any continuity requirement with respect to φ reflects their role as parameters rather than test functions. On E(Rn) we may define the product F G by (F G)(φ, x) = F (φ, x)G(φ, x) and the derivative operation (∂iF )(φ, x) = ∂ ∂xi (F (φ, x)) for i = 1 . . . n, which together give E(Rn) the structure of a differential algebra. However, as it stands, the space E(Rn) is much too large and, think- ing in terms of the limit ε → 0, contains many representations of what are essentially the same functions. For example, to represent a given smooth function f ∈ C ∞(Rn) we may define f ∈ E(Rn) by (3) f (φε, x) =ZRn f (y)φε(y − x) dy but since f is smooth we can also define another family f (φ, x) (which does not in fact depend on φ) by (4) f (φε, x) = f (x). Note that in the above equations (3) and (4) we have chosen to use the scaled mollifiers φε. Strictly speaking, however, when using these equations to define elements of E(Rn) one uses a general mollifier φ ∈ A0(Rn) (see below for details). We therefore want to introduce an equivalence relation such that f and f (and their derivatives) become equivalent. Expanding f (φε, x) in a Taylor series and using the moment conditions for φ ∈ Ak(Rn) we NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 7 see that f (φε, x) = f (y)φ(cid:18)y − x ε (cid:19) dy f (x + εy)φ(y) dy 1 f (x) + εnZRn =ZRn  =ZRn  + Xl=q+1 = f (x) + εq+1 Xl=q+1 q + 1 l! ∂lf (x) εlyl l! qXl=1 εq+1ylZ 1 (1 − t)q∂lf (x + tεy) dt  l! ZRnZ 1 yl(1 − t)q· 0 0 q + 1 φ(y)dy · ∂lf (x + tεy)φ(y) dt dy = f (φε, x) + O(εq+1) where ∂l is the derivative operator given by ∂l = ∂l1 n . Thus, by choosing φ to be in Ak(Rn) for suitably large k we can make f − f tend to zero like an arbitrary power of ε. Requiring a similar condition for the derivatives motivates the following definition. 1 . . . ∂ln Definition 3 (Negligible functions). N (Rn) is defined to be the set of functions F ∈ E(Rn) such that for all compact K ⊂ Rn, for all k ∈ Nn 0 and for all m ∈ N, there is some q ∈ N such that if φ ∈ Aq(Rn) then sup x∈K(cid:12)(cid:12)∂kF (φε, x)(cid:12)(cid:12) = O(εm) as ε → 0. Note that the derivative ∂k acts only on the x-variable here, contrary to the situation later on where we also have to consider derivatives with respect to φ. The key result that follows from this is that for a smooth function f we have that f − f is in N (Rn). However, in order to define an algebra we would like to factor out by N (Rn), and this requires it to be an ideal. Unfortunately, this is not the case because we can multiply elements of N (Rn) by elements of E(Rn) with rapid non-polynomial growth in 1/ε so that the conditions of Definition 3 are no longer satisfied. We therefore restrict E(Rn) to the subalgebra of functions of moderate growth in the following sense. Definition 4. EM (Rn) is defined to be the set of functions F ∈ E(Rn) such that for all compact K ⊂ Rn, for all k ∈ Nn 0 , there is some N ∈ N 8 E. A. NIGSCH AND J. A. VICKERS such that if φ ∈ AN (Rn) then sup x∈K(cid:12)(cid:12)∂kF (φε, x)(cid:12)(cid:12) = O(ε−N ) as ε → 0. Proposition 5. N (Rn) is an ideal in EM (Rn). We may therefore define the space of generalised functions G(Rn) as a factor algebra. Definition 6 (Generalised functions). G(Rn) = EM (Rn)/N (Rn). Although the definition of a negligible function F requires estimates for the derivatives(cid:12)(cid:12)∂kF (φε, x)(cid:12)(cid:12) these are in fact not needed as is shown by the following useful proposition. Proposition 7. Let F ∈ EM (Rn) be such that for all compact K ⊂ Rn, for all m ∈ N, there is some q ∈ N such that if φ ∈ Aq(Rn) then F (φε, x) = O(εm) as ε → 0. sup x∈K Then F ∈ N (Rn). Hence any moderate function which satisfies the negligibility con- is negligible. For the proof see [15, dition, without differentiating, Theorem 1.4.8]. One may now show that one has an embedding ι : D′(Rn) → G(Rn) T 7→ [ T ] where [ T ] denotes the equivalence class of T ∈ EM (Rn) and T (φ, x) := hT, τxφi. The only thing we need to establish is that T is moderate. As we may assume without limitation of generality that T = ∂lf for some continuous function f (y) in a neighborhood of a given compact set, differentiating the expression for T with respect to x we obtain ∂k T (φε, x) = h∂l yf, ∂k x(τxφε)i = hf, (−1)l∂l y∂k x(τxφε)i. Now τxφε(y) = 1 εn φ(cid:0) y−x ε (cid:1) so that (−1)l∂l y∂k x(τxφε(y)) = 1 ε (cid:19) . εn+l+k (−1)l+kφ(k+l)(cid:18) y − x NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 9 Thus, uniformly for x in a compact set we have ∂k T (φε, x) = = 1 1 εn+l+kZ f (y)(−1)l+kφ(k+l)(cid:18) y − x εl+k Z f (x + εy)(−1)l+kφ(k+l)(y) dy = O(ε−l−k) ε (cid:19) dy so that T is moderate. The main properties of G(Rn) are contained in the following propo- sition. For proofs and further details see [3] and [14]. Proposition 8. (a) G(Rn) is an associative commutative differential algebra. (b) The embedding ι defined above embeds D′(Rn) as a linear sub- space. (c) For smooth functions f ∈ C ∞(Rn) we have ι(f ) = [ f ] where f (φ, x) = f (x), so that G(Rn) contains the space of smooth functions as a subalgebra. (d) The embedding commutes with (distributional) partial differen- tiation so that ι(∂kT )(φ, x) = ∂k(ιT )(φ, x). As we remarked earlier an important concept is that of association. Definition 9 (Association). We say an element [F ] of G(Rn) is asso- ciated to 0 (denoted [F ] ≈ 0) if for each Ψ ∈ D(Rn) there exists some p > 0 with ε→0Zx∈Rn lim F (φε, x)Ψ(x)dx = 0 ∀φ ∈ Ap(Rn). We say two elements [F ], [G] are associated and write [F ] ≈ [G] if [F − G] ≈ 0. Definition 10 (Associated distribution). We say [F ] ∈ G(Rn) admits T ∈ D′(Rn) as an associated distribution if for each Ψ ∈ D(Rn) there exists some p > 0 with ε→0Zx∈Rn lim F (φε, x)Ψ(x)dx = hT, Ψi ∀φ ∈ Ap(Rn). These definitions do not depend on the choice of representative; moreover, note that not all generalised functions are associated to a distribution. At the level of association we regain the following compatibility re- sults for multiplication of distributions. 10 E. A. NIGSCH AND J. A. VICKERS Proposition 11. (a) If f ∈ C ∞(Rn) and T ∈ D′(Rn) then ι(f )ι(T ) ≈ ι(f T ). (b) If f, g ∈ C 0(Rn) then ι(f )ι(g) ≈ ι(f g). Although the partial derivative commutes with the embedding this is not true of the Lie derivative. Let X(x) ∈ X(Rn) be a smooth vector field on Rn and T ∈ D′(Rn), then (LX T )(φε, x) = X a(x)∂a( T (φε, x)) = X a(x)h∂aT, φx,εi. On the other hand, (]LXT )(φε, x) = hX a∂aT, φx,εi These two expressions are not the same in general since the first only involves the value of the vector field at x, while the second involves the values in a neighbourhood of x. In fact, if these expressions always were the same this would mean the embedding commutes with multiplica- tion by smooth functions, which contradicts the Schwartz impossibility result. However, by part (a) of Proposition 11 the two expressions are associated since X a is a smooth function for a = 1 . . . n. We also note that if f is a smooth function then LX f = gLXf since we may represent f by f and gLXf by LXf . We therefore have the following proposition. Proposition 12. (a) Let f ∈ C ∞(Rn) and X be a smooth vector field. Then, LX(ι(T )) = ι(LX T ). (b) Let T ∈ D′(Rn) and X be a smooth vector field. Then, LX(ι(T )) ≈ ι(LXT ). (5) (6) It is also possible to localise the entire construction to obtain G(Ω) for open sets Ω ⊂ Rn by restricting x to lie in Ω in the relevant definitions. The only technical complication relates to the embedding where one must first extend the distribution and then show that the result is independent of the extension (see [3] for details). NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 11 3. Smoothing distributions and the Colombeau algebra on manifolds A coordinate independent description of generalised functions on open sets Ω ⊂ Rn was proposed by Colombeau and Meril [11]. How- ever, this suffered from a number of minor defects; in particular, the definition of Ak(Ω) did not take into account the x-dependence of the mollifiers which meant that the definition of moderate functions was dependent on the coordinate system used. An explicit counterexample due to Jel´ınek [16] demonstrated that the construction was not in fact diffeomorphism invariant. In the same paper Jel´ınek gave an improved version of the theory which clarified a number of important issues but fell short of proving the existence of a coordinate invariant algebra. The existence of a (local) diffeomorphism invariant Colombeau alge- bra on open subsets Ω of Rn was finally established Grosser et al. [17]. In parallel with this, [18] proposed a definition of a global manifestly diffeomorphism invariant theory on manifolds. By making use of the characterisation results of [17] it was shown in [12] that one can con- struct a global Colombeau algebra G(M) of generalised functions on manifolds. There, it was demonstrated how to obtain a canonical lin- ear embedding of D′(M) into G(M) that renders C ∞(M) a faithful subalgebra of G(M). In addition, it was shown that this embedding commutes with the generalised Lie derivative, ensuring that the theory retains all the distinguishing features of the local theory in the global context. Although this theory has a well defined generalised Lie de- rivative it turns out that there is no natural definition of a generalised covariant derivative. In this section we describe a new approach to Colombeau algebras [19] based on the concept of smoothing operators that it is closer to the intuitive idea of a generalised function as a family of smooth functions. This results in a new basic space which allows us to define both a generalised Lie derivative and a covariant derivative. Replacing the spaces Ak(Rn) by suitable spaces of smoothing kernels we are able to use asymptotic versions of the moment conditions and hence do not need such a grading anymore, which results in a quantifier less in the definitions of moderateness, negligiblity and association. In contrast to [12], which made use of the local theory in a number of key places, in the current paper we give intrinsic definitions on the whole of the manifold M. As here we only outline the general theory, we refer for full proofs to [20]. On Rn the space of distributions D′(Rn) is dual to the space of smooth functions of compact support, whereas on an orientable man- ifold the space of distributions D′(M) is dual to Ωn c (M), the space of 12 E. A. NIGSCH AND J. A. VICKERS n-forms of compact support (note that on not necessarily orientable manifolds, one uses densities instead of n-forms; on an oriented mani- fold these are the same). In the Colombeau theory on Rn smoothness of the embedded functions is obtained by integrating against mollifier functions φ(y − x). The obvious generalisation on manifolds is to re- place the function by an n-form ω. However, on a manifold it does not make sense to look at ω(y − x) since y − x has no coordinate in- dependent meaning, so instead we will look at objects ωx(y) which are n-forms in y parameterised by x ∈ M. We therefore make the following definition. Definition 13. A smoothing kernel ω is a smooth map ω : M → Ωn x 7→ ωx c (M) and we denote the space of such objects SK(M). Thus SK(M) = C ∞(M, Ωn c (M)). The key new idea (see [19] for more details) is not to immediately try and generalise the Colombeau construction of Rn to a manifold M in some ad-hoc way, but to start with the notion of smoothing operator. Definition 14. A smoothing operator Φ on M is a linear continuous map We denote the space of such objects by L(D′(M), C ∞(M)). Φ : D′(M) → C ∞(M) Given a smoothing operator Φ we may associate to it a smoothing kernel ω in the following way: if for u ∈ D′(M) and x ∈ M we demand that Φ(u)(x) = hu, ωxi ∀u ∈ D′(M), ∀x ∈ M then this implies ωx(y) = hδy, ωxi = Φ(δy)(x). Thus given a smoothing operator Φ ∈ L(D′(M), C ∞(M)), this converts a distribution u ∈ D′(M) into a smooth function Φ(u) by the action of u on the smoothing kernel ω where ωx(y) := Φ(δy)(x). Conversely, given a smoothing kernel ω ∈ SK(M) we obtain a smoothing operator by Φ(u)(x) :=< u, ωx >. Indeed, it follows from a variant of the Schwartz kernel theorem that this correspondence is a topological isomorphism (7) Lb(D′(M), C ∞(M)) ∼= C ∞(M, Ωn c (M)) = SK(M) NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 13 where the left hand side has the topology of bounded convergence and the right hand side has the topology of uniform convergence on compact sets in all derivatives. We therefore take our basic space E(M) of generalised functions to consist of (smooth) maps from the space of smoothing kernels to the space of smooth functions, E(M) := C ∞(SK(M), C ∞(M)). Note that in this definition (and elsewhere in the paper where we con- sider smooth maps between infinite dimensional spaces) we will use the definition of smoothness based on the convenient setting of global anal- ysis of [13]. The basic idea of this approach is that a map f : E → F between locally convex spaces is smooth if it transports smooth curves in E to smooth curves in F (where the notion of smooth curves is straightforward via limits of difference quotients). Actually the basic space E(M) is somewhat larger than we would want since it allows F (ω) to depend on ω globally, which destroys the sheaf character of the algebra. We therefore restrict to a sub- algebra Eloc(M) consisting of local elements F ∈ E(M), defined by the property that if two smoothing kernels ω and ω agree on some open set U then F (ω) and F (ω) also agree on U. Note that all embedded elements satisfy this condition so that there is no real loss of generality in restricting to this space. Therefore, for the rest of the paper we will work exclusively with Eloc(M) but for ease of notation we will simply write it as E(M). For an in-depth exposition of this topic we refer to [21]. The basic space naturally contains both D′(M) and C ∞(M) via the linear embeddings ι and σ ι : D′(M) → E(M) σ : C ∞(M) → E(M) (ιu)(ω)(x) :=< u, ωx > (σf )(ω)(x) := f (x) and inherits the algebra structure from C ∞(M) through the product (F1 · F2)(ω) := F1(ω)F2(ω), F1, F2 ∈ E(M), ω ∈ SK(M). We may regard a smooth function as a regular distribution so that one may embed it either via σ to obtain (σf )(ω)(x) = f (x) or via ι to obtain (ιf )(ω)(x) =R f (y)ωx(y). In order to identify these expressions we would like to set ωx = δx. Strictly speaking this is not possible, but replacing ωx by a net (ωx,ε)ε of n-forms which tends to δx appropriately as ǫ → 0 and using suitable asymptotic estimates to define negligibility 14 E. A. NIGSCH AND J. A. VICKERS allows us to construct a quotient algebra in which the two embeddings of smooth functions agree. The next key concept required is therefore that of a delta net of smoothing kernels ωε which will play the role of the ε dependent mol- lifiers φx,ε used in the embedding of distributions on Rn. Since we are working on a manifold we do not have translation and scaling oper- ators available, so we need to consider carefully what properties are required. Again, rather than simply trying to copy the construction on Rn it is useful to look at what is required from the point of view of the corresponding family of smoothing operators. The key properties are that: (a) the family of smoothing operators should be localising, (b) in the limit the smoothing operator when applied to smooth functions should be the identity in C ∞(M), (c) the family of smoothing operators should satisfy some seminorm estimates which control the growth and (d) in the limit the smoothing of a distribution u should converge in D′(M) to u. Property (a) ensures that the support of the corresponding net of smoothing kernels shrinks, (b) ensures that (in the quotient algebra) the embeddings σ and ι coincide, (c) ensures that the embedding of distributions is moderate and property (d) shows that an embedded distribution is associated to the original distribution. More precisely, given a family of smoothing operators (Φε)ε∈(0,1] we require (a) on any compact K ⊂ M ∀r > 0 ∃ε0 > 0 ∀x ∈ K ∀ε 6 ε0 ∀u ∈ D′(M): (cid:0)uBr(x) = 0 ⇒ Φε(u)(x) = 0(cid:1) ; (b) for any continuous seminorm p on Lb(C ∞(M), C ∞(M)) and all m ∈ N we have p(ΦεC∞(M ) − id) = O(εm); (c) for any continuous seminorm p on Lb(D′(M), C ∞(M)) there is N ∈ N such that p(Φε) = O(ε−N ); (d) Φε → id in Lb(D′(M), D′(M)). Note that in the second condition we demand convergence like O(εm) for all m at once, contrary to Colombeau's original algebra presented above. We now use the topological isomorphism (7) to translate these con- ditions into conditions on a net (ωε)ε of smoothing kernels. The first NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 15 translates into the requirement that the support of the net shrinks, or more precisely that on any compact K ⊂ M ∀r > 0 ∃ε0 > 0 ∀x ∈ K ∀ε 6 ε0 : supp ωx,ε ⊆ Br(x). To do this we need to introduce a Riemannian metric h on M in order to measure the radius of the ball. However, it is not hard to see that the condition does not depend upon the particular choice of Riemannian metric. To formulate the next condition we need the Lie derivative of a smoothing kernel ω, which we will introduce in terms of the 1-parameter family of diffeomorphisms induced by a vector field. In principle we can consider two different diffeomorphisms µ and ν which act separately on the x and y variables of ω, i.e., the pullback action on the parameter x (for fixed y) given by (µ∗ω)x := ωµ(x) on the one hand and the pullback action on the form (for fixed x) given by ν∗(ωx) on the other hand. We will denote the combined pullback action on the smoothing kernel by (µ∗, ν∗)ω := ν∗(ωµ(x)). We can therefore also consider two different (complete) vector fields t acting on the x and y X and Y with corresponding flows FlX variables. This enables us to define the (double) Lie derivative t and FlY Varying the x and y variables separately we have two Lie derivatives L(X,Y )ω = t , (FlY )∗ t(cid:1) ω. (L(X,0)ω)x = ωFlX t (x) d d dt(cid:12)(cid:12)(cid:12)(cid:12)t=0(cid:0)(FlX)∗ dt(cid:12)(cid:12)(cid:12)(cid:12)t=0 dt(cid:12)(cid:12)(cid:12)(cid:12)t=0 d (L(0,Y )ω)x = (FlY )∗ t (ωx) L(X,Y )ω = L(X,0)ω + L(0,Y )ω. and and hence Since L(X,0)ω is given by the formula for the Lie derivative of a function we will denote this derivative by LC∞ X ω, and since L(0,Y )ω is given by the Lie derivative of an n-form we will denote this derivative by LΩn Y ω. Finally, we will often want to take the geometrically natural Lie derivative L(X,X)ω of a smoothing kernel which we denote LSK X ω. Note that L(X,0)ω is denoted L′ X ω and L(0,Y )ω is denoted LY ω in [12]. We may now define the convergence corresponding to the second condition above by demanding that for all compact subsets K ⊂ M, 16 E. A. NIGSCH AND J. A. VICKERS all m ∈ N0 and all smooth vector fields X1, . . . , Xm on M we have sup x∈K(cid:12)(cid:12)(cid:12)(cid:12) (cid:18)ZM as ε → 0. f LC∞ X1 . . . LC∞ Xm ωx,ε(cid:19) − LX1 . . . LXmf (x)(cid:12)(cid:12)(cid:12)(cid:12) = O(εm) It turns out that the third condition, which allows us to establish the fact that the embedding of a distribution is moderate, takes the following form. For any distribution u ∈ D′(M), on any compact subset K ⊂ M we require ∀k ∈ N0 ∀X1, . . . , Xk ∈ X(M) ∃N ∈ N such that LX1 . . . LXkhu, ωx,εi = O(ε−N ). sup x∈K Finally, the fourth condition which gives convergence in D′(M) is given by the condition that ∀u ∈ D′(M), ∀ω ∈ Ωn c (M) ε→0Zx∈M lim hu, ωε,xiω(x) = hu, ωi. We are now in a position to define a delta net of smoothing kernels (cf. [20] where the corresponding nets are called test objects). Definition 15 (Delta Nets of Smoothing kernels). (ωε)ε ∈ SK(M)(0,1] is called a delta net of smoothing kernels if on any compact subset K of M it satisfies the following conditions: (1) ∀r > 0 ∃ε0 ∀x ∈ K ∀ε 6 ε0: supp ωx,ε ⊆ Br(x); (2) ∀m ∈ N, as ε → 0: f LC∞ X1 . . . LC∞ Xk ωx,ε(cid:19) − LX1 . . . LXkf (x)(cid:12)(cid:12)(cid:12)(cid:12) = O(εm); (3) ∀u ∈ D′(M) ∀k ∈ N0 ∃N ∈ N ∀X1, . . . , Xk ∈ X(M): (cid:18)ZM sup x∈K(cid:12)(cid:12)(cid:12)(cid:12) LX1 . . . LXkhu, ωx,εi = O(ε−N ); sup x∈K (4) ∀u ∈ D′(M) ∀ω ∈ Ωn c (M): ε→0Zx∈Rn lim hu, ωε,xiω(x) = hu, ωi. The space of delta nets smoothing kernels on M is denoted A(M). Remark 16. We have seen in the previous section that the moment conditions on Rn allow one to show that for a smooth function f and for φ ∈ Aq(Rn) we have (in the case n = 1) (8) f (φε, x) = f (x) + εq+1 q! ZRZ 1 0 yq+1(1 − t)qf (q+1)(x + tεy)φ(y) dt dy NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 17 so that (9) f (φε, x) = f (x) + O(εq+1) with a similar argument giving the same estimate for the derivatives. By Proposition 7 this shows that f − f is negligible and hence that the two possible embeddings of a smooth function coincide in the algebra. On a manifold we have turned things round and instead used (9) to characterise the moment condition. As is the case in Rn we will use this condition to show that the two possible embeddings of smooth functions differ by a negligible function and hence coincide in the factor algebra. Although not necessary for the bare constrution of the theory, it is beneficial for practical calculations to add L1-conditions on the nets of smoothing kernels. For example, if we also require that (10) ωx,ε → 1 uniformly for x in compact subsets of M ZM so that (asymptotically) the L1-norm of the smoothing kernels is unity, one can then show that ε→0ZM lim f ωx,ε = f (x) ∀f ∈ C 0(M). Hence ι(f )(ωε) =< f, ωε > converges to f pointwise. However, this condition is different from condition (2) which involves convergence in C ∞(M) and requires that the derivatives (of arbitrary order) also converge to the derivatives of f . Another useful condition imitating the behaviour of scaled and translated mollifiers is (11) ∀K ⊆ M compact ∀α ∈ Nn x ωε(x) = O(ε−α). x∈KZ ∂α 0 : sup Before turning to the definition of moderate and negligible functions we consider the definition of the Lie derivative for elements of the basic space. There are two different ways of thinking about the Lie derivative of an element F ∈ E(M). The first comes from looking at the pullback action of the diffeomorphism group on the basic space (which we call the geometrical or generalised Lie derivative) while the second comes from thinking of F (ω) for fixed ω as a smooth function. The former has the advantage that it commutes with the embedding of distributions, but on the other hand it cannot be C ∞ linear in X (since having both properties would violate the Schwartz impossibility result). The latter is simply the ordinary Lie derivative of a smooth function and therefore agrees with the directional derivative or covariant derivative of a function. This will allow us to define the covariant derivative of a generalised tensor field as in [1]. Although the ordinary Lie derivative 18 E. A. NIGSCH AND J. A. VICKERS does not commute with the embedding of distributions, as is the case on Rn, it does so at the level of association. To consider the geometric Lie derivative we start by looking at the action of a diffeomorphism on a generalised function. Definition 17 (Pullback action). If ψ : M → N is a diffeomorphism then we define the pullback ψ∗ : E(N) → E(M) by (ψ∗F )(ω)(x) := F(cid:0)((ψ−1)∗, (ψ−1)∗)ω(cid:1)(ψ(x)). We are now in a position to define the Lie derivative. Definition 18 (Geometrical Lie derivative). Let FlX t be the flow gen- erated by the (complete) smooth vector field X. Then for F ∈ E(M) we set LXF = (FlX t )∗F. d dt(cid:12)(cid:12)(cid:12)(cid:12)t=0 Using the chain rule we may write this as ( LXF )(ω) = −dF (ω)(LSK X ω) + LX(F (ω)) and since this formula may also be applied to a non-complete vector field we take this as the definition in the general case. Definition 19 (Generalised Lie Derivative)). For any F ∈ E(M) and any X ∈ X(M) we set (12) ( LXF )(ω) := −dF (ω)(LSK X ω) + LX(F (ω)) Remark 20. In the terminology of [21], the basic space of [12] is given by the (ωx, x)-local elements of E(M). On these, the formula for the generalised Lie derivative is identical to that in [12] evaluated at ω = ωx. The other approach is to fix the smoothing kernel ω ∈ SK(M) so that x 7→ F (ω)(x) is a smooth function of x. We may then define another Lie derivative of F (which we denote LX F ) by fixing ω and taking the (ordinary) Lie derivative of F (ω), so that (13) ( LXF )(ω) := LX(F (ω)). Having defined suitable derivatives on E(M) and established that A(M) is non-void, we turn to the definition of moderate and negligi- ble functions on manifolds. We start with the definition of negligible functions. Consider a net Φε of smoothing operators converging to the identity. Then from this point of view the natural definition of a negligible function F is one that satisfies F (Φε) → 0 as k → ∞ NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 19 in C ∞(M) (i.e. in all derivatives). Writing this in terms of smooth- ing kernels we therefore require LX1 . . . LXk(F (ωε)) → 0 as ε → 0. Since ( LXF )(ω) = LX(F (ω)) this automatically gives stability of the subspace of negligible functions under the ordinary Lie derivative LX. However we also require stability of negligible functions under the gen- eralised Lie derivative LX . This suggests that we require ( L X1 . . . L Xk LX1 . . . LXℓF )(ωε) → 0 as ε → 0. However, by definition we have (( LX − LX)F )(ωε) = dF (ωε)(LSK X ωε) so that taking linear combinations of the two types of Lie derivative is equivalent to looking at dF and evaluating it on the tangent space to A(M). We therefore introduce the space A0(M) := {ω0 ∈ C ∞(M, Ωn and make the following definition: Definition 21 (Negligible functions). The function F ∈ E(M) is negli- gible if for any given compact K ⊂ M ∀k, j, m ∈ N0 ∀X1 . . . Xk ∈ X(M) ∀ω ∈ A(M) ∀ω1 . . . ωj ∈ A0(M): c (M))(0,1] : ω ∈ A(M) ⇒ ω0 + ω ∈ A(M)} as ε → 0. sup x∈K(cid:12)(cid:12)(cid:12)( LX1 . . . LXk(djF )(ωε)(ω1,ε . . . ωj,ε)(x)(cid:12)(cid:12)(cid:12) = O(εm) The set of negligible elements is denoted N (M). In order that the space of negligible functions is an ideal we also need to restrict to the space of moderate functions. Definition 22 (Moderate functions). The function F ∈ E(M) is mod- erate if for any given compact K ⊂ M ∀k, j ∈ N0 ∀ω ∈ A(M) ∀ω1 . . . ωj ∈ A0(M) ∃N ∈ N0 ∀X1 . . . Xk ∈ X(M): (14) sup x∈K(cid:12)(cid:12)(cid:12)( LX1 . . . LXk(djF )(ωε)(ω1,ε . . . ωj,ε)(x)(cid:12)(cid:12)(cid:12) = O(ε−N ) The set of moderate elements of E(M) is denoted EM (M). as ε → 0. Remark 23. Although the above definitions require one to consider derivatives djF of arbitrary order in practice one only needs to verify this condition is satisfied by objects that are embedded into the algebra via σ or ι. Since σ does not depend on ω and the embedding ι is linear in ω, this leaves the cases j = 0 and j = 1. Theorem 24. 20 E. A. NIGSCH AND J. A. VICKERS (a) EM (M) is a subalgebra of E(M). (b) N (M) is an ideal in EM (M). Proof. Because of the property of derivatives it is clear from the def- initions that that the product of two moderate functions is moderate and the product of a negligible function with a moderate function is negligible. (cid:3) Proposition 25. Let F ∈ EM (M) be such that for all compact K ⊂ M ∀m ∈ N0 ∀ω ∈ A(M) F (ωε)(x) = O(εm) as ε → 0. sup x∈K Then F ∈ N (M). Proof. This follows from looking at F (ωε + εkωε) where ω ∈ A(M), ω ∈ A0(M), applying the mean-value theorem and using the definition of moderateness of F with k suitably chosen. (cid:3) This result shows that one does not need derivatives to test negligi- bility of a moderate function. Theorem 26. Let X ∈ X(M). Then (a) LX EM (M) ⊆ EM (M) and LX EM (M) ⊆ EM (M). (b) LX N (M) ⊆ N (M) and LX N (M) ⊆ N (M). This follows immediately from the definitions. We are finally in a position to define generalised functions on mani- folds. Definition 27 (Generalised Functions). The space G(M) = EM (M) N (M) is called the Colombeau algebra of generalised functions on M. Theorem 28. The space of Colombeau generalised functions G(M) is a fine sheaf of associative commutative differential algebras on M. Proof. By construction the basic space E(M) is an associative commu- tative differential algebra, with derivative the generalised Lie derivative L given by equation (12). EM (M) is a subalgebra of E(M) and N (M) is an ideal in EM (M), hence G(M) is an algebra. Furthermore the spaces EM (M) and N (M) are stable under both the generalised and ordinary Lie derivatives so that G(M) is a differential algebra with respect to the both Lie derivatives. The sheaf properties of G(M) follow from the localisation results [21]. (cid:3) NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 21 We now want to show that we may embed the space of distributions D′(M) in the space of generalised functions G(M). Given a distribution T in D′(M) we define the function T ∈ E(M) by T (ω)(x) = hT, ωxi. We now need to show that T is moderate. For this we need to look at Lie derivatives of dj T (ω). For j = 0 we have T (ωε) = hT, ωεi and it then follows from property (3) of the smoothing kernels that LX1 . . . LXk( T (ωε))(x) = hT, LC∞ X1 . . . LC∞ Xk ωx,εi = O(ε−N ). Since the embedding is linear, we have for j = 1 that (d T )(ω)(Ψ) = hT, Ψi, so the above argument gives the desired bound on the growth, while for j > 2 we have dj T = 0. This shows that T is moderate and we have an embedding ι : D′(M) → G(M) T 7→ [ T ] where [ T ] is the equivalence class of T in G(M). By the definition of the generalised Lie derivative we have LX(ιT )(ω)(x) = −hT, (LSK = −hT, LΩn = hLXT, ωxi = ι(LXT )(ω)(x). X ω)xi + LXhT, ωxi X ωxi Hence, (15) LX(ιT ) = ι(LXT ) and thus the embedding ι commutes with the generalised Lie derivative. It is clear that if f is a smooth function on M then f defined by f (ω)(x) = f (x) is a moderate function. By passing to the equivalence class [ f ] we obtain the embedding σ : C ∞(M) → G(M) from above. Clearly σ gives an injective algebra homomorphism of the algebra of smooth functions on M into G(M), the algebra of generalised functions on M. Furthermore since σ(f ) has no dependence on ω we only have the second term in the formula for the definition of the generalised Lie derivative so σ also commutes with the Lie derivative. Finally, it easily follows from Definition 15 that for a smooth function the difference between f and f is negligible and hence on passing to the quotient ι coincides with σ on C ∞(M). Collecting these results together we have the following theorem: 22 E. A. NIGSCH AND J. A. VICKERS Theorem 29. ι : D′(M) → G(M) is a linear embedding that commutes with the generalised Lie derivative and coincides with σ : C ∞(M) → G(M) on C ∞(M). Thus ι renders D′(M) a linear subspace and C ∞(M) a faithful subalgebra of G(M). As explained in the introduction the concept of association is an important feature of the theory of Colombeau algebras on manifolds as in many cases it allows us to recover a description in terms of classical distributions by a method of 'coarse graining'. We now show how this notion may be extended to generalised functions on manifolds. Definition 30 (Association). We say an element [F ] of G(M) is as- sociated to 0 (denoted [F ] ≈ 0) if for each ω ∈ Ωn c (M) we have ε→0Zx∈M lim F (ωε)(x)ω(x) = 0 ∀ω ∈ A(M). We say two elements [F ], [G] are associated and write [F ] ≈ [G] if [F − G] ≈ 0. Definition 31 (Associated distribution). We say [F ] ∈ G(M) admits u ∈ D′(M) as an associated distribution if for each ω ∈ Ωn c (M) we have ε→0Zx∈M lim F (ωε)(x)ω(x) = hu, ωi ∀ω ∈ A(M). Again, these definitions do not depend on the representative of the class. As in Rn at the level of association we regain the usual results for multiplication of distributions, provided that suitable L1-conditions like (10) and (11) are used. Proposition 32. (a) If f ∈ C ∞(M) and T ∈ D′(M) then ι(f )ι(T ) ≈ ι(f T ). (b) If f, g ∈ C 0(M) then ι(f )ι(g) ≈ ι(f g). The above results establish almost everything we want at the scalar level. Before going on to look at the tensor theory and develop a theory of differential geometry there is one further ingredient we will require, which is the notion of directional (or covariant) derivative ∇XF of a generalised scalar field. Ideally this would be C ∞(M)-linear in X (so NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 23 that ∇f X F = f ∇XF ) and commute with the embedding. However, it is not hard to see that this is not possible since this would require that ι(f ∇X g) = ι(∇f X g) = ∇f X ι(g) = f ∇X (ιg) = ι(f )ι(∇X g) which cannot in general be true by the Schwartz impossibility result. However, in view of Proposition 32 a C ∞(M)-linear derivative that commutes with the embedding only at the level of association is not ruled out by the Schwartz result. By thinking of F (ω)(x) for fixed ω as a function of x we may make the following definition of generalised covariant derivative Definition 33 (Covariant derivative of a generalised scalar field). Let F ∈ G(M) be a generalised scalar field and X a smooth vector field. Then we define the covariant derivative ∇X F by ( ∇XF )(ω) = ∇X(F (ω)). We note that almost by definition this satisfies the requirements of a covariant derivative and for the case of a scalar field (which we are considering here) this is identical to the Lie derivative LXF given by (13) and hence is well defined. Although it is C ∞(M)-linear in X this derivative does not commute with the embedding into G(M). However as we now show this derivative does commute with the embedding at the level of association. Proposition 34. Let T ∈ D′(M) and X be a smooth vector field; then (16) ∇Xι(T ) = LX ι(T ) ≈ ι(LX T ) = ι(∇X T ) Proof. In the following calculation let ωε be a fixed delta net of smooth- ing kernels. Given µ a smooth n-form of compact support then ε→0ZM lim ( LXι(T ))(ωε)µ = lim (LX(ιT (ωε)))µ ε→0ZM ε→0ZM ε→0ZM = − lim = − lim (ιT )(ωε)(LXµ) hT, ωx,εi(LXµ)(x) = hT, −LX µi = hLXT, µi. (cid:3) 24 E. A. NIGSCH AND J. A. VICKERS 4. Conclusion In this paper we have reviewed the construction of the Colombeau algebra on Rn and adapted it to define the Colombeau algebra on a manifold M. The key idea has been to look at the construction on manifolds first of all in terms of smoothing operators and then translate this into the language of smoothing kernels. The result of this is to replace the mollifiers φ(y − x) by smoothing kernels ωx(y) and the scaled mollifiers φε(y−x) by delta nets of smoothing kernels ωx,ε(y). In this way, given a locally integrable function f we may approximate it by a 1-parameter family of smooth functions (depending on ω) according to fε(x) =Zy∈M f (y)ωx,ε(y). For fixed ω ∈ A(M) these may be treated just like smooth functions on manifolds so all the standard operations that may be carried out on smooth functions extend to the smoothed functions fε. The embedding extends to distributions T ∈ D′(M) by defining Tε(x) = hT, ωx,εi. The nets of smoothing kernels tend to δx as ε → 0 and by using the rate at which this happens to introduce a grading A(M) on the smoothing kernels we have a condition which corresponds to the vanishing mo- ment condition on Rn. We can therefore define the spaces of moderate and negligible functions which allows us to define G(M) as the quotient G(M) = EM (M)/ N (M). The algebra of generalised functions G(M) contains the space of smooth functions as a subalgebra and has the space of distributions as a canonically embedded linear subspace. We also introduced the generalised Lie derivative which commutes with the embedding and makes G(M) into a differential algebra. Finally we defined the covariant derivative of generalised scalar fields on the manifold M and showed that this commutes with the distributional (covariant) derivative at the level of association. In a subsequent pa- per [1] this theory will be extended to a nonlinear theory of tensor distributions on a manifold M where this is used to develop a theory of nonlinear distributional geometry. Acknowledgments. E. Nigsch was supported by grants P26859 and P30233 of the Austrian Science Fund (FWF). References [1] Nigsch EA, Vickers JA. 2019 A nonlinear theory of distributional geometry. Preprint. [2] Geroch R, Traschen J. 1987 Strings and other distributional sources in general relativity. Phys. Rev. D (3) 36, 1017 -- 1031. NONLINEAR GENERALISED FUNCTIONS ON MANIFOLDS 25 [3] Colombeau JF. 1984 New generalized functions and multiplication of distribu- tions. Number 84 in North-Holland Mathematics Studies. Amsterdam: North- Holland Publishing Co. [4] Clarke C, Vickers J, Wilson J. 1996 Generalized functions and distributional curvature of cosmic strings. Classical Quantum Gravity 13, 2485 -- 2498. [5] Balasin H. 1997 Distributional energy-momentum tensor of the extended Kerr geometry. Class. Quantum Grav. 14, 3353 -- 3362. [6] Steinbauer R. 1997 The ultrarelativistic Riessner-Nordstrøm field in the Colombeau algebra. J. Math. Phys. 38, 1614 -- 1622. [7] Steinbauer R, Vickers JA. 2006 The use of generalized functions and distribu- tions in general relativity. Classical Quantum Gravity 23, r91 -- r114. [8] Colombeau JF. 1992 Multiplication of distributions. Number 1532 in Lecture Notes in Mathematics. Berlin: Springer-Verlag. [9] Schwartz L. 1954 Sur l'impossibilit´e de la multiplication des distributions. Comptes Rendus de l'Acad´emie des Sciences 239, 847 -- 848. [10] Vickers J, Wilson J. 1999 Invariance of the distributional curvature of the cone under smooth diffeomorphisms. Classical Quantum Gravity 16, 579 -- 588. [11] Colombeau JF, Meril A. 1994 Generalized functions and multiplication of dis- tributions on C∞ manifolds. J. Math. Anal. Appl. 186, 357 -- 364. [12] Grosser M, Kunzinger M, Steinbauer R, Vickers JA. 2002 A Global Theory of Algebras of Generalized Functions. Adv. Math. 166, 50 -- 72. [13] Kriegl A, Michor PW. 1997 The convenient setting of global analysis. Num- ber 53 in Mathematical Surveys and Monographs. Providence, RI: American Mathematical Society. [14] Oberguggenberger M. 1992 Multiplication of Distributions and Applications to Partial Differential Equations. Number 259 in Pitman Research Notes in Mathematics. Harlow, U.K.: Longman. [15] Grosser M, Kunzinger M, Oberguggenberger M, Steinbauer R. 2001 Geometric theory of generalized functions with applications to general relativity. Number 537 in Mathematics and its Applications. Dordrecht: Kluwer Academic Pub- lishers. [16] Jel´ınek J. 1999 An intrinsic definition of the Colombeau generalized functions.. Commentat. Math. Univ. Carol. 40, 71 -- 95. [17] Grosser M, Farkas E, Kunzinger M, Steinbauer R. 2001 On the foundations of nonlinear generalized functions I, II. Mem. Am. Math. Soc. 729. [18] Vickers JA, Wilson JP. 1998 A nonlinear theory of tensor distributions. ESI- Preprint (available electronically at http://www.esi.ac.at/ESI-Preprints.html) 566. [19] Nigsch EA. 2015 The functional analytic foundation of Colombeau algebras. J. Math. Anal. Appl. 421, 415 -- 435. [20] Nigsch EA. 2019 Spacetimes with distributional semi-Riemannian metrics and their curvature. Submitted. [21] Grosser M, Nigsch EA. 2018 Full and special Colombeau Algebras. Proc. Edinb. Math. Soc 61, 961 -- 994. 26 E. A. NIGSCH AND J. A. VICKERS E. A. Nigsch, Institut fur Mathematik, Universitat Wien, Vienna, Austria E-mail address: [email protected] J. A. Vickers, School of Mathematics, University of Southampton, Southampton SO17 1BJ, UK E-mail address: [email protected]
1008.5238	1	1008	2010-08-31T07:43:02	Invariant and hyperinvariant subspaces for amenable operators	[ "math.FA" ]	There has been a long-standing conjecture in Banach algebra that every amenable operator is similar to a normal operator. In this paper, we study the structure of amenable operators on Hilbert spaces. At first, we show that the conjecture is equivalent to every non-scalar amenable operator has a non-trivial hyperinvariant subspace and equivalent to every amenable operator is similar to a reducible operator and has a non-trivial invariant subspace; and then, we give two decompositions for amenable operators, which supporting the conjecture.	math.FA	math	INVARIANT AND HYPERINVARIANT SUBSPACES FOR AMENABLE OPERATORS LUO YI SHI, YU JING WU, AND YOU QING JI Abstract. There has been a long-standing conjecture in Banach algebra that every amenable operator is similar to a normal operator. In this paper, we study the structure of amenable operators on Hilbert spaces. At first, we show that the conjecture is equivalent to every non-scalar amenable operator has a non-trivial hyperinvariant subspace and equivalent to every amenable operator is similar to a reducible operator and has a non-trivial invariant subspace; and then, we give two decompositions for amenable operators, which supporting the conjecture. 1. Introduction Throughout this paper, H denotes a complex separable infinite-dimension Hilbert space and B(H) denotes the bounded linear operators on H. For an algebra A in B(H), we write A′ for the commutant of A (i.e., A′ = {B ∈ B(H), BA = AB for allA ∈ A}) and A′′ for the double commutant of A (i.e., A′′ = (A′)′). We also write LatA for the collection of those subspaces which are invariant for every operator in A. We say A is completely reducible if for every subspace M in LatA there exists N in LatA such that H = M +N (i.e., M ∩ N = {0} and H is the algebraic direct sum of M and N); A is reducible if for every subspace M in LatA we have M ⊥ (the orthogonal complement of M) in LatA; A is transitive if LatA = {{0}, H}. If T ∈ B(H), we say that a subspace M of H is a hyperinvariant subspace for T if M is invariant under each operator in A′ T ; M is a reducible subspace for T if M, M ⊥ ∈ LatT . The concept of amenable Banach algebras was first introduced by B. E. Johnson in [10]. Suppose that A is a Banach algebra. A Banach A-bimodule is a Banach space X that is also an algebra A-bimodule for which there exists a constant K > 0 such that a· x ≤ Kax and x · a ≤ Kax for all a ∈ A and x ∈ X. We note that X ∗, the dual of X, is a Banach A-bimodule with respect to the dual actions [a · f ](x) = f (x · a), [f · a](x) = f (a · x), a ∈ A, x ∈ X, f ∈ X ∗. Such a Banach A-bimodule is called a dual A-bimodule. A derivation D : A → X is a continuous linear map such that D(ab) = a·D(b)+D(a)·b, for all a, b ∈ A. Given x ∈ X, the inner derivation δx : A → X, is defined by δx(a) = a · x − x · a. According to Johnsons original definition, a Banach algebra A is amenable if every deriva- tion from A into the dual A-bimodule X ∗ is inner for all Banach A-bimodules X. If T ∈ B(H), denote the norm-closure of span{T k : k ∈ {0} ∪ N} by AT , where N is the Date: 7.04. 2010. 2000 Mathematics Subject Classification. 47C05 (46H35 47A65 47A66 47B15). Key words and phrases. Amenable; Invariant subspaces; Hyperinvariant subspaces; Reduction property. Supported by NCET(040296), NNSF of China(10971079) and the Specialized Research Fund for the Doctoral Program of Higher Education(20050183002) . 1 set of natural numbers, T is said to be an amenable operator, if AT is an amenable Banach algebra. Ever since its introduction, the concept of amenability has played an important role in research in Banach algebras, operator algebras and harmonic analysis. There has been a long-standing conjecture in the Banach algebra community, stated as follows: Conjecture 1.1. A commutative Banach subalgebra of B(H) is amenable if and only if it is similar to a C ∗-algebra. One of the first result in this direction is due to Willis [16]. Willis showed that if T is an amenable compact operator, then T is similar to a normal operator. In [7] Gifford studied the reduction property for operator algebras consisting of compact operators and showed that if such an algebra is amenable then it is similar to a C ∗-algebras. In the recent papers [5], [6] Farenick, Forrest and Marcoux showed that if T is similar to a normal operator, then AT is amenable if and only if AT is similar to a C ∗-algebra and the spectrum of T has connected complement and empty interior; If T is a triangular operator with respect to an orthonormal basis of H, then AT is amenable if and only if T is similar to a normal operator whose spectrum has connected complement and empty interior. For further details see [5] and [6]. In this paper, we give the characterization of the structure of amenable operators. At first, we use the reduction theory of von Neumann to give two equivalent descriptions for Con- jecture 1.1; and then, we give two decompositions for amenable operators, which supporting the Conjecture 1.1. 2. An equivalent formulation of the conjecture 1.1 In this section we use the reduction theory of von Neumann to give two equivalent de- scriptions for Conjecture 1.1. We obtain that every amenable operator is similar to a normal operator if and only if every non-scalar amenable operator has a non-trivial hyperinvariant subspace if and only if every amenable operator is similar to a reducible operator and has a non-trivial invariant subspace. In order to proof the main theorem, we need to introduce von Neumann's reduction theory [15] and some lemmas. Let H1 ⊆ H2 ⊆ · · · ⊆ H∞ be a sequence of Hilbert spaces chosen once and for all, Hn having the dimension n. Let µ be a finite positive regular measure defined on the Borel sets of a separable metric space ∧, and let {En}∞ n=1 be a collection of disjoint Borel sets of ∧ with union ∧. Then the symbol Z ⊕ ∧ H(λ)µ(dλ) denotes the set of all functions f defined on ∧ such that (1)f (λ) ∈ Hn ⊆ H∞ if λ ∈ En; (2)f (λ) is a µ-measurable function with values in H∞; We put ∧ f (λ)2µ(dλ) < ∞. (3)R ⊕ (4)(f, g) =R ⊕ and dimension sets {En} and denoted by H =R ⊕ ∧ (f (λ), g(λ))µ(dλ). H(λ)µ(dλ). ∧ 2 The set of functions thus defined is called the direct integral Hilbert space with measure µ ∧ A(λ)µ(dλ) for the equivalence class corresponding to A(·). An operator on H is said to be decomposable if there exists a strongly µ-measurable operator-value function A(·) defined on ∧ such that A(λ) is a bounded operator on the space H(λ) = Hn when λ ∈ En, and for all f ∈ H, (Af )(λ) = A(λ)f (λ). We write If A(λ) is a scalar multiple of the identity on H(λ) for almost all λ, then A is called diagonal. The collection of all diagonal operator is called the diagonal algebra of ∧. In [15]I.3, Schwartz showed that H(λ)µ(dλ) is decomposable if and only if A belong A = R ⊕ an operator A on Hilbert space H =R ⊕ to the commutant of the diagonal algebra of ∧. And A = µ − ess.supλ∈∧A(λ). ∧ ∧ In [1], Azoff, Fong and Gilfeather used von Neumann's reduction theory to define the reduction theory for non-selfadjoint operator algebras: Fix a partitioned measure space ∧ and let D be the corresponding diagonal algebra. Given an algebra A of decomposable ∧ A(λ)µ(dλ). Chosse a countable generating set {An} for A. let A(λ) be the strongly closed algebra generated by the {An(λ)}. A(λ)µ(dλ) is called the decomposition of A respect to D. A decomposition A ∼ A(λ)µ(dλ) of an algebra is said to be maximal if the corresponding diagonal algebra is maximal among the abelian von Neumman subalgebras of A′. The following lemma is a basic result in [1] which will be used in this paper. operators. Each operator A ∈ A has a decomposition A =R ⊕ A ∼ R ⊕ R ⊕ Lemma 2.1. ([1], Theorem 4.1) Let A ∼R ⊕ A(λ)µ(dλ) be a decomposition of a reductive In particular, if the decomposition is algebra. Then almost all of {A(λ)} are reducible. maximal, then almost all of the algebras {A(λ)} are transitive. ∧ ∧ In [7] Gifford studied the reduction property for operator algebras and obtained the fol- lowing result: Lemma 2.2. ([7] Lemma 4.4, Lemma 4.12) If A is a commutative amenable operator algebra, then A′, A′′ are complete reducible and there exists M ≥ 1 so that for any idempotent p ∈ A′′ p ≤ M. Assume A is a operator algebra, let P (A) denote the idempotents in A and P(A) denote the strongly closed algebra generated by P (A). We get the following lemma: Lemma 2.3. If A is a commutative amenable operator algebra, then P(A′′) is similar to an abelian von Neumann algebra. Proof. By Lemma 2.2 and [2, Corollary 17.3], it follows that there exists X ∈ B(H) such that XpX −1 is selfadjoint for each p ∈ P (A′′). Hence P(A′′) is similar to a abelian von Neumann algebra. (cid:3) Lemma 2.4. ([5]) Let A and B be Banach algebras and suppose that ϕ : A −→ B is a continuous homomorphism with ϕ(A) dense in B. If A is amenable, then B is amenable. Notation 2.5. From Lemma 2.3,2.4 we always assume that P(A′′ algebra, and A′ T is a reducible operator algebra in this section. T ) is a abelian von Neumann Now we will proof the main result of this section: Theorem 2.6. The following are equivalent: (1) Every amenable operator is similar to a normal operator; (2) Every non-scalar amenable operator has a non-trivial hyperinvariant subspace; 3 (3) Every amenable Banach algebra which is generated by an operator is similar to a C ∗-algebra. Proof. (1) ⇔ (3) and (1) ⇒ (2) is clear by [5]. Therefore, in order to establish the theorem it suffices to show the implications (2) ⇒ (1). T ∼ R ⊕ ∧ Assume (2), by Lemma 2.3 choose a maximal decomposition for A′ A′ T (λ)µ(dλ) respect to the diagonal algebra P(A′′ T ). ∧ T (λ)µ(dλ) is the decomposition for T . Let {pn}∞ Assume T ∼R ⊕ polynomials. Then pn(T ) ∼R ⊕ n=1 denote the all rational ∧ pn(T )(λ)µ(dλ) is decomposable for all n and there exists a measurable E ⊆ ∧ such that µ(∧ − E) = 0 and for any λ ∈ E we have pn(T )(λ) = pn(T (λ)) and pn(T )(λ) ≤ pn(T ) by [15, Lemma I.3.1, I.3.2]. Define a mapping ϕλ : AT → AT (λ) by ϕλ(pn(T )) = pn(T (λ)) for each rational polynomial pn and λ ∈ E. Note that pn(T (λ)) ≤ pn(T ) for each rational polynomial pn and furthermore, {pn(T )} is dense in AT . Hence, ϕλ is well-defined and ϕλ a continuous homomorphism with ϕ(AT ) dense in AT (λ). By Lemma 2.4, T (λ) is amenable for almost all λ. Now for almost all λ, T (λ) is amenable and A′ T (λ) ⊆ A′ T (λ) and A′ Lemma 2.1. Thus almost all of T (λ) are scalar operators, i.e. T is a normal operator. T (λ) is transitive by (cid:3) Corollary 2.7. Every amenable operator is similar to a normal operator if and only if there exists a non-trivial idempotent in the double-commutant of every non-scalar amenable operator. Remark 2.8. In [5], Farenick, Forrest and Marcoux showed that if T ∈ B(H) is amenable and similar to a normal operator N, then the spectrum of N has connected complement and empty interior. According to [14, Theorem 1.23], N is a reducible operator. Hence, there exists an invertible operator X ∈ B(H) such that A′′ XT X −1 is a reducible algebra. The following theorem give the equivalent description for Conjecture 1.1 from the existence of invariant subspace for amenable operators. Theorem 2.9. The following are equivalent: (1) Every amenable operator is similar to a normal operator; (2) For every amenable operator T ∈ B(H), there exists an invertible operator X ∈ B(H) such that A′′ XT X −1 is a reducible algebra and T has a non-trivial invariant subspace. Proof. (1) ⇒ (2) is clear by Remark 2.8. (2) ⇒ (1) is trivial modifications adapt the proof of theorem 2.6. (cid:3) Remark 2.10. According to theorem 2.6, 2.9, we obtain that the Conjecture 1.1 for operator algebra which is generated by an operator is equivalent to the following statements: (1) Every amenable operator T has a non-trivial invariant subspace and renorm H with an equivalent Hilbert space norm so that under this norm LatAT becomes orthogonally complemented; (2) Every non-scalar amenable operator has a non-trivial hyperinvariant subspace. 3. Decomposition of Amenable operators In this section, we get two decompositions for amenable operators and prove that the two decompositions are the same which supporting Conjecture 1.1. 4 At first, we summarize some of the details of multiplicity theory for abelian von Neumann algebras. For the most part, we will follow [3]. If A is an operator on a Hilbert space K and n is a cardinal number, Let Kn denote the orthogonal direct sum of n copies of K, and A(n) be the operator on Kn which is the direct sum of n copies of A. Whenever A is an operator algebra on K, A(n) denotes the algebra {A(n), A ∈ A}. An abelian von Neumann algebra B is of uniform multiplicity n if it is (unitary equivalent to) A(n) for some maximal abelian von Neumann algebra A. By [3], for any abelian von Neumann algebra A, there exists a sequens of regular Borel measures {µn} on a sequens of separable metric space {Xn} such that A n=1 ⊕Bn ⊕ B∞, where Bn is a von Neumman algebra which has uniform multiplicity n for all 1 ≤ n ≤ ∞. For further details see [3, II.3]. is unitary equivalent toP∞ Proposition 3.1. Suppose that T is amenable operator and A′ T contains a subalgebra which is similar to an abelian von Neumman algebra with no direct summand of uniform multiplicity infinite, then T is similar to a normal operator. Proof. For the sake of simplicity, we assume A′ T contains a subalgebra B which is an abelian von Neumman subalgebra with no direct summand of uniform multiplicity infinite. Trivial modifications adapt the proof to the more general case. By [3, II.3], there exists a sequens of regular Borel measures {µn} on a sequens of separable n=1 ⊕Bn, where Bn is a von n=1 ⊕Tn, where n. It suffices to show that Tn is similar to a normal operator for all n, then by [4, metric space {Xn} such that B is unitarily equivalent to P∞ Neumman algebra which has uniform multiplicity n for all n. Hence, T =P∞ Tn ∈ B′ Corollary 26], it follows that T is similar to a normal operator. Since T ∈ B′ n, according to [14, Theorem 7.20], for any 1 ≤ n < ∞ there exists a unitary operator Un ∈ B′ n such that UnTn(Un)−1 = N11 N12 0 N22 0 ... 0 0 ... 0 · · · · · · . . . . . . · · · · · · N1n · · · N2n ... . . . ... . . . 0 Nnn   where Nij is a normal operator for all 1 ≤ i, j ≤ n. By [17, Proposition 3.1], it follows that Tn is similar to ⊕n (cid:3) i=1Nii, i.e. Tn is similar to a normal operator for all n. Corollary 3.2. Assume T is amenable operator, then there exists hyperinvariant subspaces M, N of T such that T has the form T = T1 +T2 respect to the space decomposition H = M +N, where T1, T2 are amenable operators, T1 is similar to a normal operator and P(A′′ T2) is similar to an abelian von Neumman algebra with uniform multiplicity infinite. The proof of the following lemma is straightforward and we omit it. Lemma 3.3. Suppose that A is a completely reductive operator algebra and p ∈ P (A′). Then pA is a completely reductive operator algebra on Ranp. We are in need of the following propositions before we can address the main theorem of this section. 5 respect to the space decomposition H = M ⊕ M ⊥. By [17, Lemma 2.8], there exists an M ⊥. Assume that S 1 . By [9, propsition Proposition 3.4. Assume that T is a amenable operator and there exists a space decompo- . Then T is similar to N T2 (cid:21) M Proof. Assume that T has the matrix form a normal operator if and only if T1 and T2 are similar to normal operators. sition H = M +N such that T has the matrix form T =(cid:20) T1 eT2 (cid:21) M T =(cid:20) T1 T12 M ⊥ such that S −1T S =(cid:20) T1 invertible operator S =(cid:20) I S12 I (cid:21) M S1 (cid:21) M (cid:20) I eT2 (cid:21) M , we obtain that T2 = S1eT2S −1 has the matrix form S = M M ⊥ M ⊥ N 6.5], we get that T is similar to a normal operator if and only if T1 and T2 are similar to normal operators. (cid:3) Proposition 3.5. Suppose that T is an amenable operator, M1 ∈ LatA′ Then M1 + M2 is closed. T and M2 ∈ LatA′′ T . Moreover, if T M1 and T M2 are similar to normal operators, then T M1+M2 is similar to a normal operator. Proof. Let N0 = M1 ∩ M2, according to Lemma 3.3, there exists N ∈ LatA′′ T such that M2 = N0 +N. Choose q ∈ P (A′ T . Hence qM1 ⊂ M1 ∩ N = {0}. Therefore M1 ⊂ (I − q)H. We see that M1 + M2 = M1 +N is closed. This establishes the first statement of the proposition. T ), such that Ranq = N. By the assumption, M1 ∈ LatA′ Since T M2 is similar to a normal operator, by proposition 3.4, we get that T N is similar to a normal operator. By the assumption T M1 is similar to a normal operator, using proposition 3.4 again, we obtain that T M1+M2 = T M1 +N is similar to a normal operator. (cid:3) Now we will obtain the main theorem of this section. Theorem 3.6. Assume T is an amenable operator, then there exists hyperinvariant sub- spaces M1, M2 of T such that T has the form T = T1 +T2 respect to the space decomposition H = M1 +M2 and satisfies that: (1) T1, T2 are amenable operators; (2) If M is a hyperinvariant subspace of T and T M is similar to a normal operator, then M ⊆ M1, i.e. M1 is the largest hyperinvariant subspace on which T is similar to a normal operator; (3) For any q ∈ P (A′′ (4) P(A′′ T2), T2Ranq is not similar to a normal operator; T2) is similar to an abelian von Neumman algebra with uniform multiplicity infi- nite; T = A′ (5) A′ (6) There exists no nonzero compact operator in A′ T = A′′ T2, A′′ T2; +A′′ +A′ T1 T1 T2. Proof. Case1. For any p ∈ P (A′′ the proof of Proposition 3.1, we obtain that P(A′′ algebra with uniform multiplicity infinite. Let M1 = 0. T ), T Ranp is not similar to normal operator. According to T ) is similar to an abelian von Neumman 6 Next we will prove that for any q ∈ P (A′′ Then according to Proposition 3.1 P(A′′ with uniform multiplicity infinite. Indeed, if there exists q ∈ P (A′′ has the form q =(cid:20) I 0 A11 Kerq 0 0 (cid:21) Ranq A = R = A22 A33 Ranp0 Ranq Kerq .  I I 0  Ranp0 Ranq Kerq . Let Case2. There exists p ∈ P (A′′ T ) such that T Ranp is similar to normal operator. Then, by Zorn's Lemma and the same method in the proof of [4, Corollary 26], we can show that there exists an element p0 ∈ P (A′′ T ) which is maximal with respect to the property that T Ranp0 is similar to a normal operator. Using Proposition 3.5, Ranp0 is the largest hyperinvariant subspace of T on which T is similar to a normal operator. Hence, T has the form T = T1 +T2 with respect to the space decomposition H = Ranp0 +Kerp0 where T1 is similar to a normal operator, T1, T2 are amenable operators. Let M1 = Ranp0, M2 = Kerp0. T2), T2Ranq is not similar to normal operator. T2) is similar to an abelian von Neumman algebra T2) such that T2Ranq is similar to a normal operator and q . Then for any A ∈ A′ T , A has the form Then R ∈ P (A′′ dicts to the maximal property of p0. T ). By the assumption T RanR is similar to a normal operator which contra- At last we will prove that there exists no nonzero compact operator in A′ Indeed, if there exists a nonzero compact operator k0 ∈ A′ T2, let L1 denote the subspace spanned by the ranges of all compact operators in A′ T2, and L2 the intersection of their T2 and L1 +L2 = Kerp0. Considering the kernel, by [13, Lemma 3.1], both L1, L2 lie in LatA′ restricting T2L1, assume T21 = T2L1, then T21 is an amenable operator and A′ T21 contain a sufficient set of compact operators. By Lemma 2.2 and [12, Theorem 9], T21 is similar to a normal operator which contradicts to the above discussion. (cid:3) T2. Trivial modifications adapt the proof of Theorem 3.6, we obtain the following theorem which decomposes amenable operators by the invariant subspaces of them. The proof is similar to Theorem 3.6 and we omit it. Theorem 3.7. Assume T is an amenable operator, then there exists invariant subspaces N1, N2 of T such that T has the form T = A1 +A2 respect to the space decomposition H = N1 +N2 and satisfies that: (1) A1, A2 are amenable operators; (2) If N is an invariant subspace of T such that N1 ⊆ N and T N is similar to a normal operator, then N = N1, i.e. N1 is the maximal invariant subspace on which T is similar to a normal operator; (3) For any q ∈ P (A′ (4) If P(A′ T2), T2Ranq is not similar to a normal operator; T2) contains a subalgebra which is similar to an abelian von Neumman algebra then the von Neumman algebra has the uniform multiplicity infinite. 7 Remark 3.8. If the answer to Conjecture 1.1 is positive, by Theorem 2.6, every amenable is similar to a normal operator. Then, for the above theorem M1 = N1 = H. That is to say, the two decompositions of theorem 3.6 and 3.7 are the same. The remainder of this section, we will prove that the two decompositions are the same which supporting Conjecture 1.1. Lemma 3.9. [4] If T ∈ B(H) is an amenable operator and there exist a one-to-one bounded linear map W : H → H2, a bounded linear map V : H1 → H with dense range and operators S1 ∈ B(H1), S2 ∈ B(H2) which are similar to normal operators such that T V = V S1 and W T = S2W , then T is similar to a normal operator. Corollary 3.10. Assume T = B1B2 is an amenable operator, where B1, B2 are positive operators, then T is similar to a normal operator. Proof. Assume B1, B2 have the forms B2 =(cid:20) 0 fB2 (cid:21) , B1 =(cid:20) B11 B12 12 B22 (cid:21) , B∗ space decomposition. Since T is an amenable operator, by [17, Lemma 2.8], T is similar to respect to the space decomposition H = KerB2 ⊕ (KerB2)⊥ where fB2 is one-to-one and B11, B22 are positive operators. Thus T has the form T = " 0 B12fB2 0 B22fB2 # respect to the (cid:20) 0 0 B22fB2 (cid:21). Thus without loss of generality, we may assume that B2 is one-to-one. Assume that B1, B2 has the form 0 B1 =(cid:20) fB1 0 (cid:21) , B2 =(cid:20) B11 B12 12 B22 (cid:21) , B∗ respect to the space decomposition. Since T is an amenable operator, by [17, Lemma 2.8], T respect to the space decomposition H = (KerB1)⊥ ⊕ KerB1 where fB1 is one-to-one and has dense range and B11, B22 are positive operators. Thus T has the form T =(cid:20) fB1B11 fB1B12 (cid:21) is similar to(cid:20) fB1B11 0 0 (cid:21) and there exists an operator S such that fB1B12 =fB1B11S. Note that fB1, B2 are one-to-one, hence B12 = B11S, and B11 is one-to-one. Thus without loss of generality, we may assume that B1 has dense range and B2 is one-to-one. 0 0 0 1 1 1 1 1 1 1 1 1 Note that B 2 1 = B 2 1 B 2 1 B2B 2 1 and B 1 1 B 2 2 B1B 2 2 B 1 2 2 2 1 , B 1 B2B 2 are positive operators and T B 2 , by Lemma 3.9, T is similar to a normal operator. 2 B1B 2 2 2 2 T = (cid:3) Theorem 3.11. The two decompositions for an amenable operator in Theorem 3.6, 3.7 are the same. Proof. According to Theorem 3.6, 3.7, and Proposition 3.5, it is suffices to proof that N1 ∈ LatA′ T . 8 N2 T2 (cid:21) N1 , T =   I 0 I 0 S =  I 0  0 (cid:21) N1 S = N2 Moreover, L =(cid:20) 0 X 0 In fact, if not. T has the form T =(cid:20) T1 where Y 6= 0. Note that S and T have the form and there exists S =(cid:20) 0 Y 0 (cid:21) N1 N2 0 ∈ A′ T 0 0 0 Y 0 0 0 0 0 N1 RanY N2 ⊖ RanY T1 0 0 T21 T22 T23 T24 N1 RanY , N2 ⊖ RanY   where Y has dense range. Note that T S = ST , we get that T23 = 0. Since T is amenable, by [17, Lemma 2.8] there exists an operator B : N2 ⊖ RanY → RanY such that 0 I B T1 I  I  0 I B I 0 0 0 T21 T22 T24 0 0 0 Y 0 0 0 0 0  . 0 I −B 0 I −B I  = I  = T1 0 T21 0 0 T24 0 0 0 Y 0 0 0 0 0  . Hence, we can assume that Y has dense range. Using T is amenable again, there exists ∈ A′ T , where X 6= 0, by [7, lemma 4.11]. Similar to the decomposition to S and T , we get that S, L and T have the form 0 0 0 Y1 0 0 Y2 0 0  , L = 0 0 0 0 0 0 X 0 0  , T = T1 0 0 T31 T32 T33 T34  , respect to the space decomposition H = N1 ⊕ KerX ⊕ (N2 ⊖ KerX), where X is one-to-one, and Y1, Y2 has dense range. Note that LT = T L, we get that T33 = 0. Using T is amenable again, there exists an operator C : N2 ⊖ KerX → KerX such that and 0 0 T31 T32 T34 0 0 0 0 0 0 X 0 0     I 0 0 I C I 0  I 0 0 I C T1 0 I C I  I  I  I 0 I 0 0 I −C 0 I −C I  = I  = I  = T1 0 T31 0 0 0 0 0 0 X 0 0 0 0 T34  0 0 0 Y1 + C Y2 0 0 0 0 Y2   . 0 0 0 Y1 0 0 Y2 0 0  I 0 0 I −C 9 Moreover, Y2T1 = T34 Y2, T1 X = XT34, and T1 is similar to a normal operator, by Lemma 3.9, T34 is similar to a normal operator, which contracts to Theorem 3.7. (cid:3) Corollary 3.12. Assume T is an amenable operator, then M is a maximal invariant sub- space such that T M is similar to a normal operator if and only if M is the largest invariant subspace such that T M is similar to a normal operator. Corollary 3.13. Assume T is an amenable operator and which is quasisimilar to a compact operator, then T is similar to a normal operator. Proof. Suppose, T V = V K, W T = KW with V, W injective operators having dense ranges and K is a compact operator. Then T V KW = V KW T . Let C = V KW , C ∈ A′ T , and C is a compact operator. According to Theorem 3.6, C has the form(cid:20) C1 0 (cid:21) respect to the space decomposition in the Theorem. If Cx = 0, V W T x = Cx = 0, thus T x = 0. It follows that there is no part of T2, i.e. T is similar to a normal operator. (cid:3) 4. (Essential) operator valued roots of abelian analytic functions In this section, we will study the structure of an operator which is an (essential) operator valued roots of abelian analytic functions and then we get that if such an operator is also amenable, then it is similar to a normal operator. In [8] Gilfeather introduce the concept of operator valued roots of abelian analytic functions as follows: Let A is an abelian von Neumann algebra and ψ(z), an A valued analytic function on a domain D in the complex plane. We may decompose A into a direct integral of factors such that for A ∈ A, there ∧ g(λ)I(λ)µ(dλ). If T ∈ A′ and σ(T ) ⊆ D, let exists a unique g ∈ L∞(∧, µ) such that A =R ⊕ ψ(T ) = (2πi)−1Z∧ (T − zI)−1ψ(z)dz. An operator T is called a (essential)roots of the abelian analytic function ψ, if ψ(T ) = 0(compact, respectively). The structure of roots of a locally nonzero abelian analytic function has been given in [8], in this section we main study the structure of essential roots of a locally nonzero abelian analytic function. Lemma 4.1. Assume T ∈ B(H), f is a locally nonzero analytic function on the neighborhood of σ(T ) and assume f (T ) is a compact operator, then T is a polynomial compact operator. Proof. Let bT denote the image of T in the Calkin algebra, then [f (T ) = 0. Since f is a locally nonzero analytic function on σ(T ), there exists a polynomial p such that [p(T ) = 0. i.e. T is a polynomial compact operator. (cid:3) Theorem 4.2. Let ψ be a locally nonzero abelian analytic function on D taking values in the von Neumann algebra A. If T is an essential roots of ψ and is amenable, then T is similar to a normal operator. n=1 ⊕Bn ⊕ B∞, where Bn is a von Neumman algebra which has uniform multiplicity n for all 1 ≤ n ≤ ∞. Note T ∈ A′ is an amenable operator, thus T = T1 ⊕ T2, where T1 is similar to ∞. Let σ1(σ2) denote the continuous (atom, respectively) parts of the spectrum of B∞, then B∞ = C∞ ⊕ D∞, where C∞ and Proof. Since A is an abelian von Neumann algebra, A is unitarily equivalent toP∞ a normal operator, and T1 ∈ (P∞ n=1 ⊕Bn)′, T2 ∈ B′ 10 D∞ are uniform multiplicity ∞ von Neumman algebra and σ(C∞) = σ1, σ(D∞) = σ2 and T2 = T3 ⊕ T4, where T3 ∈ σ(C∞)′, T4 ∈ σ(D∞)′. Assume ψ is a locally nonzero abelian analytic function on D and σ(T ) ⊆ D, then ψ(T ) = ψ(T1) ⊕ ψ(T3) ⊕ ψ(T4), note that ψ(T3) is a compact operator and σ(C∞) = σ1, so ψ(T3) = 0. Since D∞ are uniform multiplicity ∞ and σ(D∞) = σ2, by Lemma 4.1, it follows that T4 is direct sum of polynomial compact operators. According to [8, Theorem 2.1], there exists a sequence of mutually orthogonal projections {Pn, Qm} in A with I = P Pn +P Qm so that T Pn is finite type spectral operator and T Qm is polynomial compact operator. By [17, Theorem 3.5, 4.5], we get that T is similar to a normal operator. (cid:3) References [1] E. A. Azoff; C. K. Fong and F. Gilfeather, A reduction theory for non-self-adjoint operator algebras. Trans. Amer. Math. Soc. 224 (1976) 351 -- 366 (1977). [2] K. R. Davidson, Nest algebras, Longman group UK limited, Essex, 1988. [3] K. R. Davidson, (3-WTRL) C ∗-algebras by example. (English summary) Fields Institute Monographs, 6. American Mathematical Society, Providence, RI, 1996. [4] C. K. Fong, Operator algebras with complemented invariant subspace lattices. Indiana Univ. Math. J. 26 (1977) 1045 -- 1056. [5] D. R. Farenick, B.E. Forrest and L. W. Marcoux, Amenable operators on Hilbert spaces, J. reine angew. Math. 582(2005) 201-228. [6] D. R. Farenick, B.E. Forrest and L. W. Marcoux, Amenable operators on Hilbert spaces, J. reine angew. Math. 602(2007) 235. [7] J. A. Gifford, Operator algebras with a reduction proprety, J. Aust. Math. Soc 80 (2006) 279 -- 315. [8] F. Gilfeather, Operator valued roots of abelian analytic functions, Pacific J. Math. 55 (1974), 127 -- 148. [9] D. W. Hadwin, An asymptotic double commutant theorem for C ∗-algebras. Trans. Amer. Math. Soc. 244 (1978) 273 -- 297. [10] B. E. Johnson. Cohomology in Banach Algebras. Mem. Amer. Math. Soc. Vol. 127 (Amer. Math. Soc., 1972). [11] R. G. Douglas, On operators similar to normal operators, Rev. Roum. Math. Pures Appl. 14(1969) 193-197. [12] S. Rosenoer, Completely reducible operator algebras and spectral synthesis, Canad. J. Math. 34 (1982), no. 5, 1025 -- 1035. [13] S. Rosenoer, Completely reducible algebras containing compact operators, J. Operator Theory. 29 (1993), no. 2, 269 -- 285. [14] H. Radjavi and P. Rosenthal, Invariant subspaces. Second edition. Dover Publications, Inc., Mineola, NY, 2003. [15] J. T. Schwartz, W*-algebras, Gordon and Breach, New York, 1967. [16] G. A. Willis, When the algebra generated by an operator is amenable, J. Operator Theorey. 34 (1995) 239 -- 249. [17] Y. Q. Ji and L. Y. Shi, Amenable operators on Hilbert spaces, Houston Journal of Mathematics(to appear). Department of Mathematics, Tianjin Polytechnic University, Tianjin 300160, P.R. CHINA E-mail address: [email protected] Tianjin Vocational Institute, Tianjin 300160, P.R. CHINA E-mail address: [email protected] Department of Mathematics, Jilin University, Changchun 130012, P.R. CHINA E-mail address: [email protected] 11
1212.6761	2	1212	2013-04-25T19:08:20	On the Bishop-Phelps-Bollobas property for numerical radius in C(K)-spaces	[ "math.FA" ]	We study the Bishop-Phelps-Bollobas property for numerical radius within the framework of C(K) spaces. We present several sufficient conditions on a compact space K ensuring that C(K) has the Bishop-Phelps-Bollobas property for numerical radius. In particular, we show that C(K) has such property whenever K is metrizable.	math.FA	math	ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ Dedicated to Irene ABSTRACT. We study the Bishop-Phelps-Bollob´as property for numerical radius within the framework of C(K) spaces. We present several sufficient conditions on a compact space K ensuring that C(K) has the Bishop-Phelps-Bollob´as property for numerical ra- dius. In particular, we show that C(K) has such property whenever K is metrizable. 1. INTRODUCTION The Bishop-Phelps-Bollob´as property for numerical radius has been recently introduced in [14] as a quantitative way of studying the set of operators on a Banach space that attain their numerical radius (see below for precise definitions). Since Sims [19] raised the ques- tion of the norm denseness of the set of numerical radius attaining operators, several results have been obtained in this direction. Acosta initiated a systematic study of this problem in her Ph.D. Thesis [1], followed by [2] and joint works with Pay´a [4, 5]. Prior to them, Berg and Sims [6] gave a positive answer for uniformly convex spaces and Cardassi obtained positive answers for ℓ1, c0, C(K) (K compact metric space), L1(µ) and uniformly smooth spaces, see [9, 10, 11]. Note that Johnson and Wolfe [15] had already shown that the set of norm attaining operators T : C(K) → C(L) is norm dense in the space of operators L(C(K), C(L)), where K and L are arbitrary compact spaces. Acosta [1] pointed out that an operator T : C(K) → C(K) attains its norm if and only if it attains it numerical radius. This observation together with Johnson and Wolfe's result led her to conclude that the set of numerical radius attaining operators on C(K) is dense in L(C(K)). Using a renorming of c0, Pay´a [17] provided an example of a Banach space X such that the set of numerical radius attaining operators on X is not norm dense in L(X), answering in the negative Sims' question. Acosta, Aguirre and Pay´a [3] gave another counterexample: X = ℓ2 ⊕∞ G, where G is Gowers' space. Observe that these examples show that there exist Banach spaces failing the Bishop-Phelps-Bollob´as property for numerical radius. In [14] it is shown that ℓ1 and c0 have the Bishop-Phelps-Bollob´as property for numer- ical radius. In fact, the proof for c0 can be reduced to a duality argument from the proof for ℓ1. In this paper we focus on the Banach space C(K) and we discuss whether this space has the Bishop-Phelp-Bollob´as property for numerical radius. Trying to transfer the ideas in [14] to the C(K) case is clearly not enough. We now summarize briefly the contents of this paper. Date: November 10, 2018. 2010 Mathematics Subject Classification. 46B20, 47A12, 54E45. Key words and phrases. Bishop-Phelps-Bollob´as property, numerical radius, space of continuous functions, space of measures, compact space. Research supported by Ministerio de Econom´ıa y Competitividad and FEDER under project MTM2011- 25377. A. Avil´es was supported by Ram´on y Cajal contract (RYC-2008-02051). A. J. Guirao was supported by Generalitat Valenciana (GV/2010/036). 1 2 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ In Section 2 we introduce the concepts of compensation of a regular measure and of compact space admitting local compensation (Definition 2.1). These notions are essential tools for our proofs and are applied to obtain a parametric version of the classical Bishop- Phelps-Bollob´as theorem for functionals on C(K) (Lemma 2.9). Then we show that C(K) has the Bishop-Phelps-Bollob´as property for numerical radius whenever K admits local compensation (Theorem 2.2). In Section 3 we show that every compact metric space admits local compensation. In fact, a stronger result holds true, namely, that every compact metric space admits a com- pensation function (Definition 3.1). We rely on the constructive proof that the Cantor set admits a compensation function (Theorem 3.6) and the fact that compensation functions can be transferred to other compacta via regular averaging operators (Lemma 3.5). As a consequence of Theorem 2.2, it turns out that C(K) has the Bishop-Phelps-Bollob´as property for numerical radius whenever K is metrizable. In Section 4 we discuss the case of non-metrizable compacta. With the help of the aux- iliary concept of closeness function, we present two examples of compact spaces admitting local compensation but no compensation function (Theorems 4.8 and 4.11). We also show that there exist compact spaces that do not admit local compensation. We finish the paper with some open problems, see Subsection 4.3. Terminology. By countable we mean finite or countably infinite. The first uncountable ordinal is denoted by ω1. All our Banach spaces X are real. We write BX = {x ∈ X : kxk ≤ 1} and SX = {x ∈ X : kxk = 1}. The topological dual of X is denoted by X ∗ and the weak∗ topology on X ∗ is denoted by ω∗. The evaluation of x∗ ∈ X ∗ at x ∈ X is denoted by x∗(x) = hx∗, xi = hx, x∗i. We write Π(X) = {(x, x∗) ∈ SX × SX ∗ : x∗(x) = 1}. We write π2(x) = {x∗ ∈ BX ∗ : x∗(x) = 1} and π2(x, δ) = {x∗ ∈ BX ∗ : x∗(x) ≥ 1 − δ} for every x ∈ BX and δ > 0. By an operator on X we mean a linear continuous mapping T : X → X. Its numerical radius is defined by ν(T ) = sup{hx∗, T (x)i : (x, x∗) ∈ Π(X)}. The Banach space of all operators on X is denoted by L(X). It is well known that ν(·) is a continuous seminorm on L(X). In general, there exists a constant n(X) ≥ 0 (the numerical index of X) such that n(X) kT k ≤ ν(T ) ≤ kT k for all T ∈ L(X). For background in numerical radius (resp. Bishop-Phelps-Bollob´as property we are concerned about is defined as: index) we refer to [7, 8] (resp. [16]). The Definition 1.1. We say that a Banach space X has the Bishop-Phelps-Bollob´as (BPB) property for numerical radius if there is a function δ : (0, 1) → (0, 1) such that: for every 0 < ε < 1, T ∈ L(X) with ν(T ) = 1 and (x, x∗) ∈ Π(X) with hx∗, T (x)i ≥ 1 − δ(ε), there exist T0 ∈ L(X) with ν(T0) = 1 and (x0, x∗ 0, T0(x0)i = 1 such that ν(T − T0) ≤ ε, kx − x0k ≤ ε and kx∗ − x∗ 0) ∈ Π(X) with hx∗ 0k ≤ ε. Let K be a compact space (i.e. compact Hausdorff topological space). We denote by C(K) the Banach space of all continuous real-valued functions on K (equipped with the supremum norm). It is known that n(C(K)) = 1 and therefore ν(T ) = kT k for every T ∈ L(C(K)). Given any f ∈ C(K) and r ∈ R, we freely use notations like ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 3 {f ≤ r} = {t ∈ K : f (t) ≤ r}. The dual C(K)∗ is identified (via Riesz's theorem) with the Banach space M(K) of all regular Borel (signed) measures on K (equipped with the total variation norm). We write M+(K) = {µ ∈ M(K) : µ ≥ 0}. For every t ∈ K we denote by δt ∈ M(K) the Dirac measure at t. As usual, given any µ ∈ M(K), we write µ, µ+ and µ− to denote, respectively, the variation, positive part and negative part of µ. By a Hahn decomposition of µ we mean a partition (P, N ) of K into Borel sets such that µ(B) ≥ 0 (resp. µ(B) ≤ 0) for every Borel set B ⊆ P (resp. B ⊆ N ). The support of µ is denoted by supp(µ). Given µ1, µ2 ∈ M(K), we write µ1 ≪ µ2 (resp. µ1 ⊥ µ2) if µ1 is absolutely continuous with respect to µ2 (resp. µ1 and µ2 are mutually singular). 2. BPB PROPERTY FOR NUMERICAL RADIUS IN C(K) Throughout this section K is a fixed compact space. Our aim is to give a sufficient condition ensuring that C(K) has the BPB property for numerical radius, namely, that K admits local compensation (see the following definition). In Sections 3 and 4 we shall prove that K admits local compensation whenever it is metrizable, as well as in other cases. Definition 2.1. Let W (K) be the set of all ω∗-continuous functions F : K → BM(K). (i) We say that ν ∈ M(K) is a compensation of µ ∈ M(K) provided that: • 0 ≤ ν ≤ µ+ and ν(K) = µ(K) if µ(K) > 0; • ν = 0 if µ(K) ≤ 0. (ii) We say that G ∈ W (K) is a compensation of F ∈ W (K) if G(t) is a compensa- tion of F (t) for every t ∈ K. (iii) We say that K admits local compensation if every element of W (K) admits a compensation. Theorem 2.2. If K admits local compensation, then C(K) has the BPB property for numerical radius. In order to prove Theorem 2.2 we need several lemmas. Let us first point out that compensations of single measures always exist: Remark 2.3. If µ ∈ M(K) satisfies µ(K) > 0 and we set λ := µ(K) ν := λµ+ is a compensation of µ. µ+(K) ∈ (0, 1], then Lemma 2.4. If ν ∈ M(K) is a compensation of µ ∈ M(K), then kµ − νk ≤ 2kµ−k and kνk ≤ kµk. Proof. This is obvious if µ(K) ≤ 0. Suppose µ(K) > 0. Since (µ+ − ν) ⊥ µ−, we have kµ − νk =(cid:13)(cid:13)(µ+ − ν) − µ−(cid:13)(cid:13) =(cid:13)(cid:13)µ+ − ν(cid:13)(cid:13) +(cid:13)(cid:13)µ−(cid:13)(cid:13) = = (µ+ − ν)(K) + µ−(K) = µ+(K) − µ(K) + µ−(K) = 2µ−(K) = 2(cid:13)(cid:13)µ−(cid:13)(cid:13) . On the other hand, kνk = ν(K) = µ(K) ≤ kµk. (cid:3) Lemma 2.5. Let (f, µ) ∈ SC(K) × SM(K) and let (P, N ) be a Hahn decomposition of µ. Then µ(f ) = 1 if and only if Proof. Write A := ({f = 1} ∩ P ) ∪ ({f = −1} ∩ N ). Observe first that µ(cid:0)({f = 1} ∩ P ) ∪ ({f = −1} ∩ N )(cid:1) = 1. f dµ = Z{f =1}∩P f dµ + ZA (2.1) Z {f =−1}∩N f dµ = µ(A). 4 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ Therefore, if µ(A) = 1 then µ(f ) =RA f dµ = 1. Conversely, if µ(f ) = 1 then f dµ ≤ µ(cid:0){f 6= −1}∩N(cid:1) (2.1) and β := f dµ f dµ {f 6=−1}∩N = µ(A) +ZK\A Z and so µ(K \ A) =RK\A f dµ. Since we have α := Z{f 6=1}∩P α + β = ZK\A 1 = µ(f ) =ZA f dµ +ZK\A f dµ ≤ µ(cid:0){f 6= 1}∩P(cid:1), f dµ = µ(K \ A) = µ(cid:0){f 6= 1} ∩ P(cid:1) + µ(cid:0){f 6= −1} ∩ N(cid:1), µ(cid:0){f 6= 1} ∩ P(cid:1) = α = Z{f 6=1}∩P Z µ(cid:0){f 6= −1} ∩ N(cid:1) = β = − (2.2) and (2.3) it follows that {f 6=−1}∩N f dµ. f dµ Clearly, (2.2) yields µ(cid:0){f 6= 1} ∩ P(cid:1) = 0 and (2.3) yields µ(cid:0){f 6= −1} ∩ N(cid:1) = 0, so that µ(K \ A) = 0. Therefore µ(A) = 1. (cid:3) Definition 2.6. Let f ∈ C(K) and 0 < σ < ε. Since the sets {f ≥ 1−σ} and {f ≤ 1−ε} are closed and disjoint, Tietze extension theorem ensures the existence of a non-negative σ,ε ∈ BC(K) such that uf uf σ,ε{f ≥1−σ} ≡ 1 and uf σ,ε{f ≤1−ε} ≡ 0. In the same way, there is a non-negative vf σ,ε ∈ BC(K) such that vf σ,ε{f ≤−1+σ} ≡ 1 and vf σ,ε{f ≥−1+ε} ≡ 0. Given any µ ∈ M(K), we define µf,1 σ,ε, µf,2 σ,ε ∈ M(K) by µf,1 σ,ε(g) :=ZK σ,ε(g) :=ZK g · uf σ,ε dµ and µf,2 g · vf σ,ε dµ for all g ∈ C(K). Remark 2.7. (i) If ε < 1 then µf,1 σ,ε. σ,ε ⊥ µf,2 σ,ε and µ 7→ µf,2 σ,ε are ω∗-ω∗-continuous. Lemma 2.8. Let f ∈ BC(K), µ ∈ BM(K) and 0 < σ < ε < 1. Then: (ii) The mappings µ 7→ µf,1 (i) kµf,1 σ,εk ≤ 1 and kµf,2 σ,εk ≤ 1; σ,ε)+(cid:13)(cid:13) +(cid:13)(cid:13)(µf,2 σ,ε)−(cid:13)(cid:13) +(cid:13)(cid:13)(µf,2 (ii) (cid:13)(cid:13)(µf,1 (iii) (cid:13)(cid:13)(µf,1 (iv) (cid:13)(cid:13)µ − µf,1 σ,ε − µf,2 Proof. Write µ1 := µf,1 define σ,ε and µ2 := µf,2 σ,ε)−(cid:13)(cid:13) ≥ 1 − (1 − µ(f ))/σ; σ,ε)+(cid:13)(cid:13) ≤ (1 − µ(f ))/σ; σ,ε(cid:13)(cid:13) ≤ (1 − µ(f ))/σ. C :=(cid:0){f ≥ 1 − σ} ∩ P(cid:1) ∪(cid:0){f ≤ −1 + σ} ∩ N(cid:1). f dµ = µ(f ) −ZK\C f dµ −ZK\C f dµ =ZK (2.4) µ(C) ≥ZC We claim that µ(C) ≥ 1 + (1 − µ(f ))/σ. Indeed, we have f dµ. σ,ε. Let (P, N ) be a Hahn decomposition of µ and ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 5 Since ZK\C f dµ = = Z Z f dµ + f dµ + Z Z {f <1−σ}∩P {f >−1+σ}∩N {f <1−σ}∩P {f >−1+σ}∩N f dµ = (−f ) dµ ≤ (1 − σ)µ(K \ C), from (2.4) it follows that µ(C) ≥ µ(f ) − (1 − σ)(µ(K) − µ(C)) ≥ µ(f ) − (1 − σ)(1 − µ(C)), which implies that µ(C) ≥ 1 − (1 − µ(f ))/σ, as claimed. (ii). Observe that (P, N ) is also a Hahn decomposition of µ1 and µ2 (bear in mind that σ,ε ≥ 0) and that C ∩ P ⊆ {f ≥ 1 − σ} and C ∩ N ⊆ {f ≤ −1 + σ}. σ,ε ≥ 0 and vf uf Hence µ+ 1 (C) = µ1(C ∩ P ) =ZC∩P 2 (C) = −µ2(C ∩ N ) = −ZC∩N µ− uf σ,ε dµ = µ(C ∩ P ) = µ(C ∩ P ), vf σ,ε dµ = −µ(C ∩ N ) = µ(C ∩ N ), and therefore µ+ 1 (C) + µ− 2 k ≥ kµ+ 2 (C) = µ(C). We deduce that 1 + µ− 2 k ≥ (µ+ 1 + µ− kµ+ 1 k + kµ− 2 )(C) = µ(C) ≥ 1 − (1 − µ(f ))/σ. (i) and (iii). Since 0 ≤ uf σ,ε + vf σ,ε ≤ 1, we have kµ1 + µ2k ≤ kµk. On the other hand, the equality kµ1 + µ2k = kµ1k + kµ2k holds because µ1 ⊥ µ2. Hence 1 ≥ kµk ≥ kµ1k + kµ2k = = kµ+ 1 k + kµ− 1 k + kµ+ 2 k + kµ− 2 k (i) ≥ 1 − (1 − µ(f ))/σ + kµ− 1 k + kµ+ 2 k, which implies that kµ− (iv). Write h := 1 − uf 1 k + kµ+ σ,ε − vf 2 k ≤ (1 − µ(f ))/σ. σ,ε ∈ C(K), so that (µ − µ1 − µ2)(g) =RK gh dµ for all g ∈ C(K). Since 0 ≤ h ≤ 1 and h vanishes on C, we get kµ − µ1 − µ2k ≤ µ(K \ C) ≤ 1 − µ(C) ≤ (1 − µ(f ))/σ, which finishes the proof. (cid:3) Lemma 2.9. Suppose that K admits local compensation. Let f ∈ BC(K) \ {0} and take 1 − kf k < ε < 1. Then there exists f0 ∈ SC(K) such that for every F ∈ W (K) there is a ω∗-continuous function PF : F −1(π2(f, ε2/6)) → π2(f0) such that: (i) π2(f ) ⊆ π2(f0) and kf − f0k ≤ ε; (ii) kPF (t) − F (t)k ≤ ε for every t ∈ F −1(π2(f, ε2/6)). Proof. We divide the proof into several steps. Step 1. Fix ε < δ < 1. Note that K is the union of the following closed sets: A := {f ≥ 1 − ε}, B := {f ≤ −1 + ε}, C := {−1 + δ ≤ f ≤ 1 − δ}, D := {1 − δ ≤ f ≤ 1 − ε} ∪ {−1 + ε ≤ f ≤ −1 + δ}. By Tietze extension theorem, there is a continuous function g : D → [−ε, ε] such that g{f =1−ε} ≡ ε, g{f =1−δ} ≡ 0, g{f =−1+ε} ≡ −ε, g{f =−1+δ} ≡ 0. 6 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ Now, we can define f0 ∈ BC(K) by declaring f0(t) := 1 −1 f (t) f (t) + g(t) if t ∈ A, if t ∈ B, if t ∈ C, if t ∈ D.  It is straightforward that kf − f0k ≤ ε. Note also that A ∪ B 6= ∅ (because kf k > 1 − ε) and so kf0k = 1. To prove that π2(f ) ⊆ π2(f0), suppose that kf k = 1, fix any µ ∈ π2(f ) and take a Hahn decomposition (P, N ) of µ. By Lemma 2.5 we have µ(cid:0)({f = 1} ∩ P ) ∪ ({f = −1} ∩ N )(cid:1) = 1. Since {f = 1} ⊆ {f0 = 1} and {f = −1} ⊆ {f0 = −1}, another appeal to Lemma 2.5 yields µ ∈ π2(f0). Step 2. Fix F ∈ W (K). Set σ := 5ε/6 and consider F1, F2 ∈ W (K) defined by F1(t) := (F (t))f,1 σ,ε and F2(t) := (F (t))f,2 σ,ε. Define now a ω∗-continuous function Q : K → M(K) by the formula Q(t) := ξ1(t) − ξ2(t), where ξ1, ξ2 ∈ W (K) are compensations of F1 and −F2, respectively. For every t ∈ K we have supp(ξ1(t)) ⊆ supp(F1(t)) ⊆ A, supp(ξ2(t)) ⊆ supp(−F2(t)) ⊆ B, and A ∩ B = ∅, hence F1(t) ⊥ F2(t) and ξ1(t) ⊥ ξ2(t), and therefore 1 ≥ kF (t)k (∗) ≥ kF1(t) + F2(t)k = kF1(t)k + kF2(t)k ≥ (2.5) ≥ kξ1(t)k + kξ2(t)k = kQ(t)k = ξ1(t)(K) + ξ2(t)(K) ≥ ≥F1(t)(K) − F2(t)(K). (inequality (∗) was established in the proof of Lemma 2.8(iii)). It follows that hQ(t), f0i =ZA f0 dξ1(t) −ZB (2.6) = ξ1(t)(K) + ξ2(t)(K) (2.5) = kQ(t)k. f0 dξ2(t) = ξ1(t)(A) + ξ2(t)(B) = The ω∗-continuity of Q and (2.6) imply that the map t 7→ kQ(t)k is continuous. Step 3. Fix t ∈ K0 := F −1(π2(f, ε2/6)). By Lemmas 2.4 and 2.8(iii), we have (2.7) kQ(t) − (F1(t) + F2(t))k ≤ kξ1(t) − F1(t)k + kξ2(t) − (−F2(t))k ≤ ≤ 2(cid:0)(cid:13)(cid:13)(F1(t))−(cid:13)(cid:13) +(cid:13)(cid:13)(−F2(t))−(cid:13)(cid:13)(cid:1) = 2(cid:0)(cid:13)(cid:13)(F1(t))−(cid:13)(cid:13) +(cid:13)(cid:13)(F2(t))+(cid:13)(cid:13)(cid:1) ≤ (1 − hF (t), f i) ≤ 2 ≤ t∈K0 σ 2ε 5 . On the other hand, by (2.5) and Lemma 2.8(ii)-(iii), we get kQ(t)k ≥ F1(t)(K) − F2(t)(K) = =(cid:0)(cid:13)(cid:13)(F1(t))+(cid:13)(cid:13) +(cid:13)(cid:13)(F2(t))−(cid:13)(cid:13)(cid:1) −(cid:0)(cid:13)(cid:13)(F1(t))−(cid:13)(cid:13) +(cid:13)(cid:13)(F2(t))+(cid:13)(cid:13)(cid:1) ≥ (1 − hF (t), f i) ≥ 1 − 2 ≥ 1 − t∈K0 . 2ε 5 σ ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 7 Hence Q(t) 6= 0 and (2.8) Q(t) kQ(t)k (cid:13)(cid:13)(cid:13)(cid:13) − Q(t)(cid:13)(cid:13)(cid:13)(cid:13) = 1 − kQ(t)k ≤ 2ε 5 (bear in mind that kQ(t)k ≤ 1, as shown in (2.5)). But Lemma 2.8(iv) also yields (2.9) kF (t) − (F1(t) + F2(t))k ≤ Using (2.7), (2.8) and (2.9) we conclude that 1 − hF (t), f i σ t∈K0 ≤ ε 5 . Q(t) kQ(t)k (cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13) − F (t)(cid:13)(cid:13)(cid:13)(cid:13) ≤ − Q(t)(cid:13)(cid:13)(cid:13)(cid:13) + kQ(t) − (F1(t) + F2(t))k + k(F1(t) + F2(t)) − F (t)k ≤ ≤ + + = ε. 2ε 5 2ε 5 ε 5 Q(t) kQ(t)k Step 4. The previous step makes clear that the function PF : K0 → M(K), PF (t) := Q(t) kQ(t)k , is well-defined and satisfies kPF (t) − F (t)k ≤ ε for every t ∈ K0. Note that (2.6) says that PF (t) ∈ π2(f0) for every t ∈ K0. Since Q is ω∗-continuous and the map t 7→ kQ(t)k is continuous (Step 2), PF is ω∗-continuous as well. The proof is over. (cid:3) The following particular case of the classical Bishop-Phelps-Bollob´as theorem will be needed in the proof of Theorem 2.2. Corollary 2.10. Suppose that K admits local compensation. Let (f, µ) ∈ BC(K)×BM(K) such that µ(f ) ≥ 1 − ε2/6, where 0 < ε < 1. Then there is (f0, µ0) ∈ Π(C(K)) such that kf − f0k ≤ ε and kµ − µ0k ≤ ε. Proof. Apply Lemma 2.9 to f and the constant function F ∈ W (K) given by F (t) := µ for all t ∈ K, so that F −1(π2(f, ε2/6)) = K. Then we can take any µ0 ∈ PF (K). (cid:3) Remark 2.11. In the situation of Lemma 2.9, let t ∈ F −1(π2(f, ε2/6)). Then: (i) Every Hahn decomposition of F (t) is also a Hahn decomposition of PF (t). (ii) PF (t) ≪ F (t). Proof. (i) Let (P, N ) be a Hahn decomposition of F (t). As we pointed out in the proof of Lemma 2.8(ii), (P, N ) is a Hahn decomposition of both F1(t) and F2(t). We claim that for every Borel set B ⊆ P we have ξ2(t)(B) = 0. Indeed, this is obvious if F2(t)(K) ≥ 0, while if F2(t)(K) < 0 then 0 ≤ ξ2(t)(B) ≤ (−F2(t))+(B) = (F2(t))−(B) = F2(t)(B ∩ N ) = 0. Hence Q(t)(B) = ξ1(t)(B) ≥ 0 for every Borel set B ⊆ P . In the same way, we have Q(t)(B) = −ξ2(t)(B) ≤ 0 for every Borel set B ⊆ N . (ii) Obviously, F1(t) ≪ F (t) and F2(t) ≪ F (t). By the very definition of compensa- (cid:3) tion, we also have ξ1(t) ≪ F1(t) and ξ2(t) ≪ F2(t). Therefore Q(t) ≪ F (t). We are now ready to prove the main result of this section. 8 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ Proof of Theorem 2.2. We shall check that if K admits local compensation, then C(K) fulfills the requirements of Definition 1.1 with δ(ε) = (ε/6)4. Let T ∈ L(C(K)) with ν(T ) = 1 and (f, µ) ∈ Π(C(K)) such that hµ, T (f )i ≥ 1 − (ε/6)4, where 0 < ε < 1. Step 1. By Corollary 2.10 applied to (T (f ), µ) ∈ BC(K) × SM(K) and δ := ε2/7 (note that hµ, T (f )i ≥ 1 − δ2/6), there is (g, µ0) ∈ Π(C(K)) such that kT (f ) − gk ≤ δ and kµ − µ0k ≤ δ < ε. Let (P, N ) be a Hahn decomposition of µ, which in turn is also a Hahn decomposition of µ0 (see Remark 2.11(i)). Since µ(f ) = 1, an appeal to Lemma 2.5 yields The fact that µ0 ≪ µ (see Remark 2.11(ii)) implies µ(cid:0)K \ ({f = 1} ∩ P ) ∪ ({f = −1} ∩ N )(cid:1) = 0. µ0(cid:0)K \ ({f = 1} ∩ P ) ∪ ({f = −1} ∩ N )(cid:1) = 0 and so µ0(f ) = 1 (again, by Lemma 2.5). Writing D1 := {T (f ) ≥ 1 − δ} and D2 := {T (f ) ≤ −1 + δ}, the proof of Lemma 2.9 shows that supp(µ0) ⊆ D1 ∪ D2 and that µ0(B) ≥ 0 (resp. µ0(B) ≤ 0) for every Borel set B ⊆ D1 (resp. B ⊆ D2). Hence (2.10) µ0(D1) − µ0(D2) = µ0(D1 ∪ D2) = kµ0k = 1. Step 2. Let us consider the closed sets A1 := {T (f ) ≥ 1 − ε2/6} ⊇ D1, A2 := {T (f ) ≤ −1 + ε2/6} ⊇ D2, C := {−1 + ε2/6 ≤ T (f ) ≤ 1 − ε2/6}. Since D1 ∩ (C ∪ A2) = ∅ = D2 ∩ (C ∪ A1) = ∅, we can apply Tietze extension theorem to find two continuous functions g1 : K → [0, 1] and g2 : K → [−1, 0] such that g1D1 ≡ 1, g1C∪A2 ≡ 0, g2D2 ≡ −1, g2C∪A1 ≡ 0. Step 3. Let F, G ∈ W (K) be defined by F (t) := T ∗(δt) = δt ◦ T and G(t) := −F (t). It is clear that F (A1) ∪ G(A2) ⊆ π2(f, ε2/6). By Lemma 2.9 there is f0 ∈ SC(K) such that π2(f ) ⊆ π2(f0), kf − f0k ≤ ε and there exist two ω∗-continuous mappings PF : A1 → π2(f0) and PG : A2 → π2(f0) satisfying (2.11) kPF (t) − F (t)k ≤ ε and sup t∈A1 kPG(t) + F (t)k ≤ ε. sup t∈A2 Now, we can define a ω∗-continuous mapping eF : K → M(K) as follows: F (t) + g1(t)(cid:0)PF (t) − F (t)(cid:1) F (t) + g2(t)(cid:0)PG(t) + F (t)(cid:1) F (t) if t ∈ A1, if t ∈ A2, if t ∈ C. eF (t) := shall check that T0 satisfies the required properties. Define T0 ∈ L(C(K)) by T0(h)(t) := heF (t), hi for every h ∈ C(K) and t ∈ K. We Step 4. Note that eF (t) (resp. −eF (t)) is a convex combination of F (t) and PF (t) PF (A1) ∪ PG(A2) ⊆ π2(f0) ⊆ BM(K), we deduce eF (K) ⊆ BM(K), which implies that (resp. −F (t) and PG(t)) for every t ∈ A1 (resp. t ∈ A2). Since F (K) ⊆ BM(K) and kT0k = sup kT0(h)k = sup h∈BC(K) h∈BC(K) sup t∈K heF (t), hi ≤ 1. ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 9 On the other hand, kT0 − T k = sup h∈BC(K) sup t∈K heF (t) − F (t), hi ≤ sup t∈K(cid:13)(cid:13)eF (t) − F (t)(cid:13)(cid:13) (2.11) ≤ ε. Since (f, µ0) ∈ Π(C(K)) (as shown in Step 1) and π2(f ) ⊆ π2(f0), we deduce that (f0, µ0) ∈ Π(C(K)). Since g1D1 ≡ 1, g2D2 ≡ −1 and PF (A1) ∪ PG(A2) ⊆ π2(f0), we have T0(f0)(t) =(hPF (t), f0i = 1 if t ∈ D1, −hPG(t), f0i = −1 if t ∈ D2. Bearing in mind that supp(µ0) ⊆ D1 ∪ D2 (as pointed out in Step 1), it follows that hµ0, T0(f0)i =ZD1∪D2 =ZD1 T0(f0) dµ0 = T0(f0) dµ0 +ZD2 T0(f0) dµ0 = µ0(D1) − µ0(D2) (2.10) = 1. In particular, this implies that ν(T0) = 1. The proof is over. (cid:3) 3. EXISTENCE OF COMPENSATION FUNCTIONS FOR METRIC COMPACTA This section is devoted to proving that every compact metric space K admits local compensation. Actually, we shall show that a stronger property holds true, namely, that every F ∈ W (K) admits a compensation of the form ξ ◦ F , where ξ : M(K) → M(K) is a function (depending only on K) as in the following definition: Definition 3.1. Let K be a compact space and M ⊆ M(K). We say that ξ : M → M(K) is an M -compensation function if it is ω∗-ω∗-continuous and ξ(µ) is a compensation of µ for every µ ∈ M ; if in addition M = M(K), we say that ξ is a compensation function. Thus, in this section our goal is to prove the following: Theorem 3.2. Every compact metric space admits a compensation function. Corollary 3.3. If K is a compact metric space, then C(K) has the BPB property for numerical radius. Proof. Combine Theorems 2.2 and 3.2. (cid:3) Corollary 3.4. Let T be a topological space, K a compact space and F : T → M(K) a ω∗-continuous function. Suppose there is a compact metrizable set L ⊆ K such that supp(F (t)) ⊆ L for every t ∈ T . Then there is a w∗-continuous function G : T → M(K) such that G(t) is a compensation of F (t) for every t ∈ T . Proof. According to Theorem 3.2, L admits a compensation function ξ : M(L) → M(L). Let U : M(K) → M(L) and V : C(K) → C(L) be the restriction operators. Since supp(F (t)) ⊆ L for every t ∈ T , the composition U ◦ F is ω∗-continuous. It is now clear that G := V ∗ ◦ ξ ◦ U ◦ F satisfies the required properties. (cid:3) In order to prove Theorem 3.2 we need some previous work. Given a continuous onto mapping ϕ : K → L between compact spaces, let Cϕ : C(L) → C(K) be the operator defined by Cϕ(f ) := f ◦ ϕ for every f ∈ C(L). An operator u : C(K) → C(L) is called a regular averaging operator for ϕ provided that u is positive, u(1K) = 1L and u ◦ Cϕ = idC(L). 10 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ Lemma 3.5. Let K and L be compact spaces for which there is a continuous onto mapping ϕ : K → L with a regular averaging operator. If K admits a compensation function, then L admits a compensation function as well. Proof. Let ξ : M(K) → M(K) be a compensation function and u : C(K) → C(L) a regular averaging operator for ϕ. Define eξ : M(L) → M(L), eξ := C∗ ϕ ◦ ξ ◦ u∗. Clearly,eξ is ω∗-ω∗-continuous. Fix µ ∈ M(L). Since Cϕ is a positive operator, so is C∗ and thereforeeξ(µ) ≥ 0. Since u∗(µ)(K) = hu∗(µ), 1K i = hµ, u(1K)i = hµ, 1Li = µ(L) ϕ and non-negative f ∈ C(K) we have eξ(µ)(L) = heξ(µ), 1Li = hξ(u∗(µ)), Cϕ(1L)i = hξ(u∗(µ)), 1K i = ξ(u∗(µ))(K), we deduce that eξ(µ) = 0 if µ(L) ≤ 0 and eξ(µ)(L) = µ(L) if µ(L) > 0. For every (cid:10)eξ(µ), f(cid:11) =(cid:10)ξ(u∗(µ)), Cϕ(f )(cid:11) ≤(cid:10)(u∗(µ))+, Cϕ(f )(cid:11) ≤(cid:10)u∗(µ+), Cϕ(f )(cid:11) = =(cid:10)µ+, u(Cϕ(f ))(cid:11) =(cid:10)µ+, f(cid:11), because Cϕ and u∗ are positive operators and u ◦ Cϕ = idC(L). Hence eξ(µ) ≤ µ+. It follows thateξ is a compensation function. From now on we write C := 2N = {0, 1}N to denote the Cantor set. Pełczynski proved that a compact space L is metrizable if, and only if, there is a continuous onto mapping ϕ : C → L with a regular averaging operator, [18, Theorem 5.6]. This result and Lemma 3.5 show that Theorem 3.2 can be deduced from the following particular case: (cid:3) Theorem 3.6. The Cantor set C admits a compensation function. Such compensation function will be defined explicitly (Definition 3.13 and Proposi- tion 3.14). The rest of this section is devoted to proving Theorem 3.6. We divide the proof into three subsections for the convenience of the reader. We first need to introduce some notation. Definition 3.7. We define a continuous function d : R × R → R by d(s1, s2) :=(sign(s2) · min{s1, s2} 0 if s1 · s2 < 0, if s1 · s2 ≥ 0. Remark 3.8. The function d satisfies the following properties: (i) d(s1, s2) = −d(s2, s1). (ii) 0 ≤ 1 + d(s1, s2)/s1 ≤ 1 if s1 6= 0. (iii) 0 ≤ s1 + d(s1, s2) ≤ s1 if s1 ≥ 0. (iv) s1 ≤ s1 + d(s1, s2) ≤ 0 if s1 ≤ 0. (v) If s1 · s2 < 0 then either s1 + d(s1, s2) = 0 or s2 − d(s1, s2) = 0. As usual, we write 2<N to denote the set of all finite (maybe empty) sequences of 0s and 1s. Given σ = (σ(1), . . . , σ(n)) ∈ 2<N, we write length(σ) = n and σk := (σ(1), . . . , σ(k)) ∈ 2<N for every k ∈ {1, . . . , n}; ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 11 we use the convention σ0 = ∅. We denote σ a 0 := (σ(1), . . . , σ(n), 0) and σ a 1 := (σ(1), . . . , σ(n), 1). More generally, if τ = (τ (1), . . . , τ (m)) ∈ 2<N, we write σ a τ := (σ(1), . . . , σ(n), τ (1), . . . , τ (m)). Given any σ′ ∈ 2<N, the notation σ ⊆ σ′ means that length(σ′) ≥ n and σ′(k) = σ(k) for every k ∈ {1, . . . , n}. Analogously, given any t = (t(k))k∈N in the Cantor set C, the notation σ ⊆ t means that t(k) = σ(k) for every k ∈ {1, . . . , n}. Thus, the standard clopen basis for the topology of C consists of the sets Nσ := {t ∈ C : σ ⊆ t}, σ ∈ 2<N. For every n ∈ N ∪ {0} we write Cn := {σ ∈ 2<N : length(σ) = n}. 3.1. Construction. Fix µ ∈ M(C) and let mσ := µ(Nσ) for every σ ∈ 2<N. We next define a collection of real numbers { mσ : σ ∈ 2<N} satisfying some special properties which shall be discussed in Subsection 3.2. Fix n ∈ N ∪ {0}. In order to define the collection { mσ : σ ∈ Cn}, we construct certain real numbers { m(k) σ : σ ∈ Cn} for every k ∈ {0, . . . , n}. This is done inductively: • Case k = 0. Set m(0) • Case k = 1. For each τ ∈ Cn−1 we set σ := mσ for every σ ∈ Cn. m(1) τ a0 := mτ a0 + d(mτ a0, mτ a1), τ a1 := mτ a1 − d(mτ a0, mτ a1). m(1) • Assume that k ∈ {2, . . . , n} and that the collection { m(k−1) σ : σ ∈ Cn} is already constructed. Note that Cn is the disjoint union of the sets Cn,τ := {σ ∈ Cn : τ ⊆ σ}, τ ∈ Cn−k. Fix τ ∈ Cn−k. We define sn,τ,0 := Xσ∈Cn,τ,0 m(k−1) σ and sn,τ,1 := Xσ∈Cn,τ,1 m(k−1) σ , where Cn,τ,0 := {σ ∈ Cn,τ : σ(n − k + 1) = 0} and Cn,τ,1 := Cn,τ \ Cn,τ,0. We now distinguish two cases: -- If sn,τ,0 · sn,τ,1 = 0, then we set m(k) σ := m(k−1) σ for every σ ∈ Cn,τ . -- If sn,τ,0 · sn,τ,1 6= 0, then we set m(k) σ := m(k−1) σ m(k) σ := m(k−1) σ for every σ ∈ Cn,τ,0, for every σ ∈ Cn,τ,1. ·(cid:16)1 + ·(cid:16)1 − d(sn,τ,0, sn,τ,1) sn,τ,0 d(sn,τ,0, sn,τ,1) sn,τ,1 (cid:17) (cid:17) In this way, the collection { m(k) σ : σ ∈ Cn} is constructed. Finally, we define mσ := m(n) σ for every σ ∈ Cn and n ∈ N ∪ {0}. 12 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ 3.2. Properties. Fix µ ∈ M(C). We follow the notations introduced in Subsection 3.1. Lemma 3.9. mσ = mσa0 + mσa1 for every σ ∈ 2<N. Proof. Fix n ∈ N ∪ {0}. We shall prove that m(k) σ = m(k+1) σa0 + m(k+1) σa1 for every σ ∈ Cn and k ∈ {0, 1, . . . , n} by induction on k. Note that for k = 0 we have m(0) σ = mσ = µ(Nσ) = µ(Nσa0) + µ(Nσa1) = mσa0 + mσa1 = m(1) σa0 + m(1) σa1 for every σ ∈ Cn, by the very definition of m(1) and that the inductive hypothesis holds: σa0 and m(1) σa1. Suppose that k ∈ {1, . . . , n} (3.1) m(k−1) σ′ = m(k) σ′ a0 + m(k) σ′ a1 for every σ′ ∈ Cn. Fix σ ∈ Cn and let τ := σn−k, so that sn+1,τ,0 = Xσ′∈Cn+1,τ,0 m(k) m(k) σ′ σ′ = Xσ′∈Cn,τ,0 = Xσ′∈Cn,τ,0(cid:0) m(k) a0 + Xσ′∈Cn,τ,0 a1(cid:1) (3.1) a0 + m(k) σ′ σ′ m(k) σ′ a1 = = Xσ′∈Cn,τ,0 m(k−1) σ′ = sn,τ,0. In the same way, we have sn+1,τ,1 = sn,τ,1. If sn+1,τ,0 · sn+1,τ,1 = 0, then m(k+1) σa0 = m(k) σa0, m(k+1) σa1 = m(k) σa1 and m(k) σ = m(k−1) σ , hence σa0 + m(k+1) If sn+1,τ,0 · sn+1,τ,1 6= 0, then m(k+1) σa1 = m(k) σa0 + m(k) σa1 (3.1) = m(k−1) σ = m(k) σ . m(k+1) σa0 + m(k+1) σa1 = σa0 + m(k) =(cid:0) m(k) σa1(cid:1) ·(cid:16)1 + (−1)σ(n−k+1) d(sn+1,τ,0, sn+1,τ,1) sn+1,τ,σ(n−k+1) (cid:17) = ·(cid:16)1 + (−1)σ(n−k+1) d(sn,τ,0, sn,τ,1) sn,τ,σ(n−k+1) (cid:17) = m(k) (3.1) = m(k−1) σ σ , (cid:3) which finishes the proof. Lemma 3.10. µ(C) =Pσ∈Cn mσ for every n ∈ N ∪ {0}. Proof. By induction on n. The case n = 0 is obvious. Suppose n > 0 and the inductive hypothesis. By applying Lemma 3.9 we get Xτ ∈Cn mτ = Xσ∈Cn−1 mσa0 + Xσ∈Cn−1 as required. mσa1 = Xσ∈Cn−1(cid:0) mσa0 + mσa1(cid:1) = = Xσ∈Cn−1 mσ = µ(C), (cid:3) Lemma 3.11. Let n ∈ N, k ∈ {1, . . . , n} and τ ∈ Cn−k. Then the collection has constant sign. In particular, { mσ : σ ∈ Cn} has constant sign. { m(k) τ aτ ′ : τ ′ ∈ Ck} ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 13 Proof. We proceed by induction on k. The case k = 1 follows immediately from Re- mark 3.8(v). Suppose that k ∈ {2, . . . , n} and that the inductive hypothesis holds. Define τi := τ a i for i ∈ {0, 1}, so that τi ∈ Cn−k+1 and each of the collections has constant sign, which in turn coincides with the sign of { m(k−1) τiaτ ′ : τ ′ ∈ Ck−1} sn,τ,i = Xσ∈Cn,τ,i m(k−1) σ = Xτ ′∈Ck−1 m(k−1) τiaτ ′ . If sn,τ,0 · sn,τ,1 ≥ 0, then m(k) τ aτ ′ = m(k−1) τ aτ ′ for every τ ′ ∈ Ck and so the collection { m(k) τ aτ ′ : τ ′ ∈ Ck} = { m(k−1) τ0aτ ′ : τ ′ ∈ Ck−1} ∪ { m(k−1) τ1aτ ′ : τ ′ ∈ Ck−1} has constant sign, as required. If sn,τ,0 · sn,τ,1 < 0, then (3.2) m(k) τ0 aτ ′ := m(k−1) m(k) τ1 aτ ′ := m(k−1) τ0 aτ ′ ·(cid:16)1 + τ1 aτ ′ ·(cid:16)1 − d(sn,τ,0, sn,τ,1) sn,τ,0 d(sn,τ,0, sn,τ,1) sn,τ,1 (cid:17) (cid:17) for every τ ′ ∈ Ck−1, so each of the collections { m(k) the other hand, by Remark 3.8(v) and (3.2), we have either m(k) or m(k) τ1aτ ′ = 0 for every τ ′ ∈ Ck−1. It follows that the collection τ0aτ ′ : τ ′ ∈ Ck−1} ∪ { m(k) τ aτ ′ : τ ′ ∈ Ck} = { m(k) has constant sign and the proof is over. { m(k) τiaτ ′ : τ ′ ∈ Ck−1} has constant sign. On τ0 aτ ′ = 0 for every τ ′ ∈ Ck−1 τ1aτ ′ : τ ′ ∈ Ck−1} (cid:3) Lemma 3.12. Let σ ∈ 2<N. The following statements hold: (i) If mσ ≥ 0, then 0 ≤ mσ ≤ mσ. (ii) If µ(C) ≥ 0 and mσ ≤ 0, then mσ = 0. Proof. Write n := length(σ). (i) We shall prove that (3.3) 0 ≤ m(n) σ ≤ m(n−1) σ ≤ · · · ≤ m(1) σ ≤ mσ. σ ≤ mσ follow immediately from the very definition of m(1) The inequalities 0 ≤ m(1) Remark 3.8 (parts (i) and (iii)). Assume that 0 ≤ m(k−1) k ∈ {2, . . . , n}. Write τ := σn−k. If sn,τ,0 · sn,τ,1 ≥ 0, then m(k) we have σ ≤ · · · ≤ m(1) σ and σ ≤ mσ for some ; otherwise σ = m(k−1) σ m(k) σ = m(k−1) σ m(k−1) σ ·(cid:16)1 + d(sn,τ,0,sn,τ,1) ·(cid:16)1 − d(sn,τ,0,sn,τ,1) sn,τ,1 sn,τ,0 (cid:17) if σ(n − k + 1) = 0 (cid:17) if σ(n − k + 1) = 1 σ ≤ m(k−1) σ (by Remark 3.8 -- (i) and (ii) -- bearing in mind that and in either case 0 ≤ m(k) m(k−1) σ ≥ 0). This proves (3.3) and therefore 0 ≤ m(n) σ = mσ ≤ mσ. (ii) In the same way, the following chain of inequalities holds true: 0 ≥ mσ = m(n) σ ≥ m(n−1) σ ≥ · · · ≥ m(1) σ ≥ mσ. We now argue by contradiction. Suppose that mσ < 0. Then Lemma 3.11 ensures that mσ′ ≤ 0 for every σ′ ∈ Cn. Bearing in mind Lemma 3.10, we obtain 0 ≤ µ(C) = Xσ′∈Cn mσ′ ≤ mσ < 0, (cid:3) a contradiction. The proof is over. 14 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ 3.3. Compensation function. We follow the notations introduced in Subsection 3.1 with some obvious modifications to denote dependence with respect to µ ∈ M(C). Definition 3.13. Let µ ∈ M(C). We define ξ(µ) ∈ M(C) as follows: (i) If µ(C) < 0, then ξ(µ) := 0. (ii) If µ(C) ≥ 0, then ξ(µ) is the unique element of M(C) such that ξ(µ)(Nσ) = mσ(µ) for every σ ∈ 2<N. The existence of ξ(µ) is ensured by Lemma 3.9 via a standard argument. Proposition 3.14. ξ : M(C) → M(C) is a compensation function. Proof. Given any µ ∈ M(C) with µ(C) ≥ 0, we have 0 ≤ ξ(µ)(Nσ) ≤ µ+(Nσ) for every σ ∈ 2<N (thanks to Lemma 3.12) and, by the very definitions, ξ(µ)(C) = µ(C). Hence ξ(µ) is a compensation of µ for every µ ∈ M(C). To prove that ξ is a compensation function, it only remains to show that it is ω∗-ω∗- continuous. Of course, it suffices to check the continuity of ξ on which is equivalent to saying that, for every σ ∈ 2<N, the real-valued function H := {µ ∈ M(C) : µ(C) ≥ 0}, is ω∗-continuous on H. Fix n ∈ N ∪ {0}. We shall prove that µ 7→ ξ(µ)(Nσ) = mσ(µ) µ 7→ m(k) σ (µ) is ω∗-continuous on H for every σ ∈ Cn and k ∈ {0, 1, . . . , n} by induction on k. The cases k = 0 and k = 1 are obvious. Suppose k ∈ {2, . . . , n} and that the inductive hypothesis holds. Fix σ ∈ Cn and write τ := σn−k. Then the mappings sn,τ,0(·) = Xσ′∈Cn,τ,0 m(k−1) σ′ (·) and sn,τ,1(·) = Xσ′∈Cn,τ,1 m(k−1) σ′ (·) are ω∗-continuous on H. Suppose that σ(n − k + 1) = 0 (the other case is analogous). Then for every µ ∈ H we have (3.4) m(k) σ (µ) = m(k−1) σ (µ) ·(cid:16)1 + d(sn,τ,0(µ), sn,τ,1(µ)) sn,τ,0(µ) (cid:17) if sn,τ,0(µ) 6= 0, while m(k) once that m(k) σ (·) is ω∗-continuous at every µ ∈ H with sn,τ,0(µ) 6= 0. Take any µ0 ∈ H with sn,τ,0(µ0) = 0. Since { m(k−1) σ′ σ (µ) = m(k−1) σ (µ) if sn,τ,0(µ) = 0. From (3.4) it follows at stant sign (by Lemma 3.11 applied to τ a 0 and k − 1), we get m(k−1) m(k) (µ0) = 0. Bearing in mind (3.4) and Remark 3.8(ii), we obtain σ (µ0) = m(k−1) σ σ σ (µ) − m(k) (cid:12)(cid:12) m(k) σ (µ0)(cid:12)(cid:12) =(cid:12)(cid:12) m(k) σ (µ)(cid:12)(cid:12) ≤(cid:12)(cid:12) m(k−1) σ (µ)(cid:12)(cid:12) =(cid:12)(cid:12) m(k−1) σ for every µ ∈ H. This inequality and the inductive hypothesis imply that the mapping m(k) (cid:3) σ (·) is ω∗-continuous at µ0. The proof is finished. (µ0)(cid:12)(cid:12) (µ) − m(k−1) σ (µ0) : σ′ ∈ Cn,τ,0} has con- (µ0) = 0 and so 4. BEYOND THE METRIZABLE CASE In this section we discuss the existence of compensation functions in certain non- metrizable compacta. Specifically, we deal with one-point compactifications of discrete sets (Subsection 4.1) and ordinal intervals (Subsection 4.2). We shall provide examples of compact spaces K which admit local compensation but no BM(K)-compensation func- tion. Those examples and Proposition 4.1 below make clear that there exist compact spaces which do not admit local compensation. ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 15 Proposition 4.1. Let K be a compact space. If BM(K) (equipped with the ω∗-topology) admits local compensation, then K admits a BM(K)-compensation function. Proof. Write L := BM(K) and let φ : K → L be defined by φ(t) := δt, so that φ is a homeomorphism onto φ(K). Let F : L → BM(L) be the function defined by hF (µ), f i = hµ, f ◦ φi for every f ∈ C(L) and µ ∈ L. Observe that F (µ)(D) = µ(φ−1(D)) for every µ ∈ L and every Borel set D ⊆ L. Since F is ω∗-continuous and L admits local compensation, there is a ω∗-continuous func- tion G : L → BM(L) such that G(µ) is a compensation of F (µ) for every µ ∈ L. Let S : M(L) → M(φ(K)) be the restriction operator and U : C(K) → C(φ(K)) the iso- metric isomorphism given by U (g) := g ◦ φ−1. Define ξ : L → L by ξ := U ∗ ◦ S ◦ G. We shall check that ξ is a BM(K)-compensation function. Note first that S ◦ G is ω∗-ω∗-continuous, thanks to the ω∗-continuity of G and the fact that (4.1) supp(G(µ)) ⊆ supp(F (µ)) ⊆ φ(K) for every µ ∈ L. Hence ξ is ω∗-ω∗-continuous as well. On the other hand, take any µ ∈ L. Since G(µ) is a compensation of F (µ) and the inclusions (4.1) hold, it follows at once that S(G(µ)) is a compensation of S(F (µ)) Therefore, ξ(µ) is a compensation of U ∗(S(F (µ))) = µ. (cid:3) To go a bit further when studying the existence of compensation functions, we introduce the following definition. Definition 4.2. Let K be a compact space. A closeness function for K is a continuous function c : {(x, y, z) ∈ K 3 : y 6= z} → [−1, 1] such that: (i) c(x, y, z) = −c(x, z, y) whenever y 6= z; (ii) c(x, x, z) = 1 whenever x 6= z. Remark 4.3. If (K, ρ) is a compact metric space, then the formula c(x, y, z) := ρ(x, z) − ρ(x, y) max{ρ(x, y), ρ(x, z)} provides a closeness function for K. Next lemma gives a connection between closeness and compensation functions: Lemma 4.4. Let K be a compact space. If K admits a BM(K)-compensation function, then K admits a closeness function. Proof. Fix a BM(K)-compensation function ξ : BM(K) → BM(K). Define c : {(x, y, z) ∈ K 3 : y 6= z} → R, c(x, y, z) := 1 − 6 · ξ(f (x, y, z))({y}), where f (x, y, z) := 1 3 (δy + δz − δx). We will check that c is a closeness function for K. Fix (x, y, z) ∈ K 3 with y 6= z. Then f (x, y, z)(K) = 1 3 > 0, hence ξ(f (x, y, z))(K) = 1 3 and 0 ≤ ξ(f (x, y, z)) ≤ (f (x, y, z))+ = 1 3 (δy + δz). In particular, c(x, y, z) ∈ [−1, 1]. On one hand, if x = y then c(x, x, z) = 1 − 6 · ξ(f (x, x, z))({x}) = 1 − 2 · δz({x}) = 1 − 0 = 1. 16 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ On the other hand, since supp(f (x, y, z)) ⊆ {y, z}, we have 1 3 therefore = ξ(f (x, y, z))(K) = ξ(f (x, y, z))({y}) + ξ(f (x, y, z))({z}), c(x, y, z) = 6 · ξ(f (x, y, z))({z}) − 1 = −c(x, z, y). We finally check that c is continuous at (x, y, z). Since y 6= z, there exist disjoint open sets V, W ⊆ K with y ∈ V , z ∈ W , and a continuous function φ : K → [0, 1] such that φV ≡ 1 and φW ≡ 0. Then for every (x′, y′, z′) ∈ K × V × W we have (4.2) ξ(f (x′, y′, z′))({y′}) = hξ(f (x′, y′, z′)), φi, because supp(ξ(f (x′, y′, z′)) ⊆ {y′, z′}. Equality (4.2) and the ω∗-ω∗-continuity of ξ imply that c coincides with a continuous function on K × V × W , which is an open neighborhood of (x, y, z). This shows that c is a closeness function for K. (cid:3) Part (i) of the following proposition was pointed out to us by O. Kalenda and is included here with his kind permission. Proposition 4.5. Let K be a compact space admitting a closeness function. (i) K is first countable. (ii) If K is separable, then it is metrizable. Proof. Let c be a closeness function for K. We begin by proving the following: CLAIM. If x0 ∈ K belongs to the closure of a countable set D ⊆ K \ {x0}, then x0 is a Gδ-point. Indeed, for every z ∈ D we have c(x0, x0, z) = 1, hence we can take an open neigh- borhood Vz of x0 such that z 6∈ Vz and c(x0, x, z) > 0 for all x ∈ Vz. We claim thatTz∈D Vz = {x0}. By contradiction, suppose there is x ∈ Tz∈D Vz \ {x0}. Then c(x0, x, z) > 0 for all z ∈ D. Since x0 ∈ D and c is continuous, we get c(x0, x, x0) ≥ 0, which contradicts that c(x0, x, x0) = −1. This proves the claim. (i) Our proof is by contradiction. Suppose there is x ∈ K which is not a Gδ-point. Then we construct a sequence (xn) in K \{x} and a decreasing sequence (Hn) of closed Gδ-sets containing x as follows: Gδ-set containing x. Since x is not a Gδ-point, we can take xn+1 ∈ Hn \ {x}. j=1{y ∈ K : c(y, x, xj) = 1}. Then Hn is a closed • Pick an arbitrary x1 ∈ K \ {x}. • Given n ∈ N, set Hn :=Tn Now let x be a cluster point of (xn). By the CLAIM above, x 6= x. Since x ∈ Hn for every n ∈ N, we have c(x, x, xn) = 1 for every n ∈ N. From the continuity of c it follows that c(x, x, x) = 1, which contradicts that c(x, x, x) = −1. (ii) It is enough to find a countable subset of C(K) that separates the points of K. Let C be a countable dense subset of K. For every t, s ∈ C with t 6= s, let ft,s ∈ C(K) be defined by ft,s(x) := c(x, t, s). Let us check that the countable family {ft,s : t, s ∈ C, t 6= s} separates the points of K. Fix y 6= z in K. Since c(y, y, z) = 1, there are disjoint open sets V1, W1 ⊆ K such that y ∈ V1, z ∈ W1 and c(x′, y′, z′) > 0 for every (x′, y′, z′) ∈ V1 × V1 × W1. ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 17 On the other hand, since c(z, y, z) = −1, there are disjoint open sets V2, W2 ⊆ K such that y ∈ V2, z ∈ W2 and c(x′, y′, z′) < 0 for every (x′, y′, z′) ∈ W2 × V2 × W2. Pick t ∈ V1 ∩ V2 ∩ C and s ∈ W1 ∩ W2 ∩ C. Then ft,s(y) = c(y, t, s) > 0 while ft,s(z) = c(z, t, s) < 0. The proof is over. (cid:3) By combining Lemma 4.4, Proposition 4.5(ii) and Theorem 3.2 we get: Corollary 4.6. Let K be a compact space. The following statements are equivalent: (i) K is separable and admits a BM(K)-compensation function; (ii) K is metrizable. 4.1. One-point compactifications of discrete sets. Throughout this subsection Γ is a non-empty set and we denote by K := A(Γ) = Γ ∪ {∞} the one-point compactifica- tion of Γ equipped with the discrete topology. Since K is scattered, every element of M(K) is of the formPt∈K atδt for some (at)t∈K ∈ ℓ1(K), [13, Theorem 14.24]. It is well-known that a bounded net (µα) in M(K) is ω∗-convergent to µ ∈ M(K) if and only if µα(K) → µ(K) and µα({γ}) → µ({γ}) for all γ ∈ Γ. Note that if Γ is uncountable, then ∞ is not a Gδ-point of K and so Proposition 4.5(i) yields: Corollary 4.7. If Γ is uncountable set, then A(Γ) does not admit a closeness function. Hence, it neither admits a BM(A(Γ))-compensation function. However, we have the following: Theorem 4.8. A(Γ) admits local compensation and therefore C(A(Γ)) has the BPB prop- erty for numerical radius. Proof. The second statement will follow from Theorem 2.2 once we prove the first one. Let F ∈ W (K). If F (∞)(K) < 0, then there is a finite set Γ1 ⊆ Γ such that F (t)(K) < 0 for all t ∈ K \ Γ1 (because the function F (·)(K) : K → R is continuous). Fix an arbitrary compensation µt of F (t) for every t ∈ Γ1 (apply Remark 2.3). Define ξF ∈ W (K) by ξF (t) :=(µt 0 if t ∈ Γ1, if t ∈ K \ Γ1. Clearly, ξF is a compensation of F . Suppose now that F (∞)(K) ≥ 0. Since the function F (·)(K) is continuous, there is a countable set A ⊆ Γ such that (4.3) F (t)(K) = F (∞)(K) for every t ∈ K \ A. For every t ∈ K the set At := supp(F (t)) is countable. Write Γ0 := A∞ ∩ Γ = A∞ \ {∞}. For each s ∈ Γ0, the function F (·)({s}) : K → R is continuous (because {s} is a clopen subset of K) and so there is a countable set Bs ⊆ K such that (4.4) F (t)({s}) = F (∞)({s}) for every t ∈ K \ Bs. The set B := (Ss∈Γ0 Bs) ∪ A ∪ {∞} is countable, hence so isSt∈B At and therefore N := [t∈B At 18 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ is a compact metrizable (countable) subset of K. Observe that for every t ∈ B we have supp(F (t)) ⊆ N . An appeal to Corollary 3.4 ensures the existence of a ω∗-continuous function G : B → M(K) such that G(t) is a compensation of F (t) for every t ∈ B. Write ξ0 := G(∞). Let us define (4.5) C := {t ∈ K \ B : F (t)(Γ \ A∞) ≤ 0} and the mapping ξF : K → M(K) by (4.6) ξF (t)({s}) := G(t)({s}) ξ0({s}) ξ0({s}) F (t)+(K\Γ0) F (t)+({s}) ξ0({∞}) if t ∈ B and s ∈ K, if t ∈ C and s ∈ K, if t /∈ B ∪ C and s ∈ Γ0, if t /∈ B ∪ C and s ∈ K \ Γ0.  Observe that K \ Γ0 ⊇ Γ \ A∞, hence F (t)+(K \ Γ0) ≥ F (t)(Γ \ A∞) > 0 whenever t ∈ K \ (B ∪ C), so ξF is well-defined. Let us show that ξF is a compensation of F . STEP 1. ξF (t) is a compensation of F (t) for every t ∈ K. This is obvious for t ∈ B by the choice of G. In particular, ξ0(K) = F (∞)(K) ≥ 0 and 0 ≤ ξ0 ≤ F (∞)+. Let us analyze what happens for t ∈ K \ B. Case 1: Assume t ∈ C. Then t /∈ B ⊇ A, so ξF (t)(K) (4.6) = ξ0(K) = F (∞)(K) (4.3) = F (t)(K). On the other hand, take any s ∈ K. Then (4.6) = ξ0({s}) ≤ F (∞)+({s}). 0 ≤ ξF (t)({s}) We now distinguish several cases. • If s ∈ Γ0, then (4.4) implies that F (∞)+({s}) = F (t)+({s}). • If s /∈ A∞, then F (∞)+({s}) = 0 ≤ F (t)+({s}). • If ∞ ∈ A∞, then (4.7) F (∞)({∞}) = F (∞)(K) − Xr∈Γ0 = F (t)({∞}) + F (t)(Γ \ A∞) It follows that (4.3)&(4.4) = F (t)(K) − Xr∈Γ0 (4.5) ≤ F (t)({∞}). F (∞)({r}) F (t)({r}) = 0 ≤ ξF (t)({s}) ≤ F (∞)+({s}) ≤ F (t)+({s}) for all s ∈ K, hence 0 ≤ ξF (t) ≤ F (∞)+ ≤ F (t)+. Therefore, ξF (t) is a compensation of F (t). Case 2: Assume t ∈ K \ (B ∪ C). Then ξF (t)(K) =Xs∈K ξF (t)({s}) (4.6) = Xs∈Γ0 ξ0({s}) + ξ0({∞}) F (t)+(K \ Γ0) Xs∈K\Γ0 =ξ0(Γ0) + ξ0({∞}) = ξ0(Γ0 ∪ {∞}) = ξ0(K) = F (∞)(K) F (t)+({s}) (4.3) = F (t)(K). To prove that ξF (t) is a compensation of F (t) it remains to check that 0 ≤ ξF (t) ≤ F (t)+, which is equivalent to saying that 0 ≤ ξF (t)({s}) ≤ F (t)+({s}) for all s ∈ K. To this end, we distinguish two cases: • If s ∈ Γ0, then 0 ≤ ξF (t)({s}) (4.6) = ξ0({s}) ≤ F (∞)+({s}) = (4.4) = F (t)+({s}). ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 19 • If s ∈ K \ Γ0, then (4.8) 0 ≤ ξF (t)({s}) (4.6) = ξ0({∞}) F (t)+(K \ Γ0) F (t)+({s}) ≤ F (t)+({s}), because F (∞)({∞}) = F (∞)(K) − Xr∈Γ0 F (∞)({r}) (4.3)&(4.4) = F (t)(K \ Γ0) yields the inequalities ξ0({∞}) ≤ F (∞)+({∞}) ≤ F (t)+(K \ Γ0). STEP 2. ξF is ω∗-continuous. It suffices to check that ξF is ω∗-continuous when restricted to each of the closed sets B, C and K \ (B ∪ C). We already know that the restriction ξF B = G is ω∗-continuous. On the other hand, note that ξF (t) = ξF (∞) = ξ0 for every t ∈ C, hence ξF C is ω∗-continuous. Finally, let us show that ξF K\(B∪C) is also ω∗-continuous. To this end, it suffices to show that, if (tα) is a net in K \ (B ∪ C) converging to ∞, then ξF (tα) → ξF (∞) = ξ0 with respect to the ω∗-topology of M(K), which is equivalent to saying that ξF (tα)(K) → ξ0(K) and ξF (tα)({s}) → ξ0({s}) for all s ∈ Γ (note that (ξF (tα)) is bounded, because ξF (t) is a compensation of F (t) ∈ BM(K) for every t ∈ K). We know that ξF (t)(K) = F (t)(K) for all t ∈ K \ (B ∪ C) (see the proof of Step 1), hence ξF (tα)(K) = F (tα)(K) → F (∞)(K) = ξ0(K). Fix any s ∈ Γ. If s ∈ Γ0, then ξF (tα)({s}) = ξ0({s}) = ξF (∞)({s}) for all α. If s ∈ Γ \ Γ0, then by (4.8) we have 0 ≤ ξF (tα)({s}) ≤ F (tα)+({s}) for all α. Since F (tα)+({s}) → F (∞)+({s}) = 0 (because s /∈ A∞), it follows that This proves that ξF is ω∗-continuous and so ξF is a compensation of F . (cid:3) ξF (tα)({s}) → 0 = ξ0({s}). 4.2. Ordinal intervals. Throughout this subsection we work with the ordinal interval K := [0, ω1], which becomes a 0-dimensional scattered compact space when equipped with its order topology. Since K is scattered, every element of M(K) is of the form Pα∈K aαδα for some (aα)α∈K ∈ ℓ1(K), [13, Theorem 14.24]. We shall also need the following well-known fact, see e.g. [12, 3.1.27]. Lemma 4.9. If h : [0, ω1) → R is a continuous function, then there is α < ω1 such that h is constant on [α, ω1). Since ω1 is not a Gδ-point of K, an appeal to Proposition 4.5(i) yields: Corollary 4.10. [0, ω1] does not admit a closeness function. Hence, it neither admits a BM([0,ω1])-compensation function. On the other hand, we have: Theorem 4.11. [0, ω1] admits local compensation and therefore C([0, ω1]) has the BPB property for numerical radius. Proof. The second statement will follow from Theorem 2.2 once we prove the first one. Write Clop(K) to denote the algebra of all clopen subsets of K. Let F ∈ W (K). For every α ∈ K, define (4.9) s(α) := sup{γ < ω1 : F (α)({γ}) 6= 0} < ω1. 20 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ CLAIM 1. There exist b ∈ [−1, 1] and γ0 < ω1 such that (4.10) F (γ)({ω1}) = b whenever γ0 ≤ γ < ω1. Proof of Claim 1. By Lemma 4.9, we only have to check that the function F (·)({ω1}) is continuous on [0, ω1). To this end, it is enough to prove the continuity on [0, γ) for every γ < ω1. Let β := sup{s(α) : α < γ} < ω1. Notice that for every α < γ we have F (α)([β + 1, ω1]) = Xβ<γ≤ω1 F (α)({γ}) (4.9) = F (α)({ω1}). Since [β + 1, ω1] ∈ Clop(K) and F is ω∗-continuous, the previous equality ensures that the function F (·)({ω1}) is continuous on [0, γ). This finishes the proof of Claim 1. (cid:3) CLAIM 2. For every α < ω1 there is α < β(α) < ω1 such that F (γ)(A) = F (ω1)(A) for every β(α) ≤ γ < ω1 and every A ∈ Clop(K) such that A ⊆ [0, α] or A = [0, ω1]. Proof of Claim 2. Note that the set Aα := {A ∈ Clop(K) : A ⊆ [0, α]} ∪ {[0, ω1]} is countable (because [0, α] is a compact metric space and so it has countably many clopen subsets). For every A ∈ Clop(K) the function F (·)(A) : [0, ω1] → R is continuous, hence it is constant on [βA, ω1] for some α < βA < ω1 (apply Lemma 4.9). Now, the proof of Claim 2 finishes by taking β(α) := sup{βA : A ∈ Aα} < ω1. (cid:3) DEFINITION. We next define by transfinite induction a strictly increasing ω1-sequence of ordinals {λi : i < ω1} ⊆ [0, ω1). For convenience, we consider as the starting point of the induction. If i < ω1 is a limit ordinal, then we set λ0 := max{s(ω1), γ0, β(0)} < ω1 λi := sup{λj : j < i}. In the successor case, λi+1 is defined as λi+1 := sup{αn : n ∈ N} = sup{βn : n ∈ N} where λi =: α0 < β0 < α1 < β1 < . . . < λi+1 are defined as (given by Claim 2) (4.11) βn := β(αn) and (4.12) DEFINITION. We set αn := max(cid:8)βn−1 + 1, sup{s(γ) : γ ≤ βn−1}(cid:9). µ := F (ω1) − F (ω1)({ω1})δω1 ∈ M(K), (4.13) µα := F (α) − µ − bδω1 = F (α) − F (ω1) + aδω1 ∈ M(K), α ∈ [0, ω1], where we write a := F (ω1)({ω1}) − b. CLAIM 3. If i < ω1 and α ∈ (λi, λi+1), then µα is concentrated on [λi, λi+1] with µα(K) = a. Proof of Claim 3. By (4.10) (bear in mind that α > λ0 ≥ γ0) we have (4.14) F (α)({ω1}) = b. Let λi = α0 < β0 < α1 < β1 < . . . < λi+1 be the sequence of ordinals that defines λi+1 = sup{αn : n ∈ N} = sup{βn : n ∈ N}. Pick n ∈ N such that α ≤ βn−1. By (4.12) we have s(α) ≤ αn < λi+1, and so (4.15) for every λi+1 < γ < ω1. F (α)({γ}) = 0 ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 21 Note that we also have (4.16) F (ω1)({γ}) = 0 for every λi+1 < γ < ω1, because λi+1 > λ0 ≥ s(ω1). We next prove that (4.17) F (α)({γ}) = F (ω1)({γ}) for every γ < λi. 1 < β′ To this end, it suffices to check the equality for every γ < λj+1 and every ordinal j < i. 1 < . . . < λj+1 be the sequence of ordinals that defines Let λj = α′ n for some n ∈ N and λj+1 = sup{α′ n : n ∈ N}. Then γ < α′ n : n ∈ N} = sup{β′ 0 < α′ 0 < β′ Ak ⊆ [0, α′ so we can write {γ} = Tk∈N Ak for some decreasing sequence (Ak) in Clop(K) with {γ} =T{[γ1 + 1, γ2] : γ1 < γ ≤ γ2 < α′ n] for every k ∈ N. Indeed, this is obvious if γ = 0, while for γ 6= 0 we have n}. By the choice of β(α′ n) (Claim 2) and β(α′ n) (4.11) = β′ n < λj+1 ≤ λi < α, we have F (α)(Ak) = F (ω1)(Ak) for every k ∈ N and so F (α)({γ}) = lim k→∞ F (α)(Ak) = lim k→∞ F (ω1)(Ak) = F (ω1)({γ}). This proves (4.17). Since α > λ0 ≥ β(0), we have F (α)(K) = F (ω1)(K) (Claim 2) and therefore µα(K) = a (by (4.13)). Finally, from (4.14), (4.15) and (4.17) we get µα({γ}) = F (α)({γ}) − F (ω1)({γ}) + aδω1({γ}) = 0 for every γ ∈ K \ [λi, λi+1]. The proof of Claim 3 is over. CLAIM 4. For every 1 ≤ i < ω1 we have µλi = aδλi and so (4.18) F (λi) = µ + aδλi + bδω1. (cid:3) Proof of Claim 4. We proceed by transfinite induction on i. The limit ordinal case follows from the ω∗-continuity of F . Now, suppose (4.18) holds for some 1 ≤ i < ω1 and let us prove it for i + 1. Consider again the chain λi = α0 < β0 < α1 < β1 < . . . < λi+1 that defines λi+1 as its supremum. By the ω∗-continuity of F , the sequence (F (βn)) is w∗-convergent to F (λi+1), which (by (4.13)) is equivalent to saying that lim n→∞ µβn (A) = µλi+1 (A) for every A ∈ Clop(K). By Claim 3, each µβn is concentrated on [λi, λi+1] with µβn (K) = a. In particular, we get µλi+1 (K) = a. In order to prove that µλi+1 = aδλi+1 it only remains to check that µλi+1 is concentrated on {λi+1}. Fix any A ∈ Clop(K) with A ⊆ [0, λi+1). Since A is compact, we have A ⊆ [0, αn) for some n ∈ N. Then F (βm)(A) = F (ω1)(A) for every m ≥ n (by Claim 2 and (4.11)) and so F (λi+1)(A) = F (ω1)(A), hence µλi+1 (A) = 0 (by (4.13)). As A is an arbitrary clopen set contained in [0, λi+1), we conclude that µλi+1 is concentrated on [λi+1, ω1]. On the other hand, if we take any A ∈ Clop(K) with A ⊆ (λi+1, ω1], then µβn(A) = 0 for all n ∈ N and so µλi+1(A) = 0. It follows that µλi+1 is concentrated on {λi+1}, which finishes the proof of Claim 4. (cid:3) CLAIM 5. The restriction of F to [0, λ1] admits a compensation. Proof of Claim 5. Write L := [0, λ1] ∪ {ω1}. Notice first that (4.19) supp(F (α)) ⊆ L for every α ∈ [0, λ1]. Indeed, this is immediate for α = λ1 (by (4.18), bearing in mind that λ1 > λ0 ≥ s(ω1)). Let λ0 = α0 < β0 < α1 < β1 < . . . < λ1 be the chain that defines λ1 as its supremum. If we take any α < λ1, then α ≤ βn−1 for some n ∈ N and so s(α) ≤ αn < λ1 (by (4.12)), 22 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ hence supp(F (α)) ⊆ L. This proves (4.19). Since L is compact metrizable, the claim now follows from Corollary 3.4. (cid:3) CLAIM 6. There exist 0 ≤ a′ ≤ max{a, 0}, 0 ≤ b′ ≤ max{b, 0} and ν ∈ M(K) with 0 ≤ ν ≤ µ+ such that: (i) ν + a′δλi + b′δω1 is a compensation of F (λi) for every 1 ≤ i < ω1; (ii) ν + (a′ + b′)δω1 is a compensation of F (ω1). Proof of Claim 6. Write c := a + b = F (ω1)({ω1}). Observe that F (ω1) = µ + cδω1 and F (λi) (4.18) = µ + aδλi + bδω1 for every 1 ≤ i < ω1, hence F (λi)(K) = F (ω1)(K) (this equality also follows from Claim 2). Thus, the state- ment of Claim 6 holds trivially if F (ω1)(K) ≤ 0. We assume that F (ω1)(K) > 0 and distinguish two cases. Case 1. If c ≥ 0, choose a′′, b′′ ∈ R such that a′′δλ1 + b′′δω1 is a compensation of aδλ1 + bδω1. Then 0 ≤ a′′ ≤ max{a, 0}, 0 ≤ b′′ ≤ max{b, 0} and a′′ + b′′ = a + b = c. Set ν0 := µ + a′′δλ1 + b′′δω1 and note that ν0(K) = µ(K) + c = F (ω1)(K) > 0. Let ν1 be a compensation of ν0. Then ν1(K) = ν0(K) and 0 ≤ ν1 ≤ (µ + a′′δλ1 + b′′δω1)+ = µ+ + a′′δλ1 + b′′δω1 , so we can write ν1 = ν +a′δλ1 +b′δω1 for some 0 ≤ a′ ≤ a′′, 0 ≤ b′ ≤ b′′ and ν ∈ M(K) with 0 ≤ ν ≤ µ+. It is clear that ν + a′δλi + b′δω1 is a compensation of F (λi) for every 1 ≤ i < ω1 and that ν + (a′ + b′)δω1 is a compensation of F (ω1). Case 2. If c < 0, then let ν be a compensation of F (ω1), so that ν(K) = F (ω1)(K) and 0 ≤ ν ≤ (µ + cδω1)+ = µ+. In particular, ν is a compensation of F (λi) for every 1 ≤ i < ω1, so we can take a′ = b′ = 0 to conclude the proof of Claim 6. (cid:3) CLAIM 7. For every 1 ≤ i < ω1 there is a ω∗-continuous function ξi : [λi, λi+1] → M(K) such that ξi(α) is a compensation of µα for all α ∈ [λi, λi+1]. Proof of Claim 7. [λi, λi+1] is compact metrizable. Since supp(µα) ⊆ [λi, λi+1] for every α ∈ [λi, λi+1] (Claims 3 and 4) and the mapping α 7→ µα is ω∗-continuous (see (4.13)), the existence of ξi follows from Corollary 3.4. (cid:3) Let G : [0, λ1] → M(K) be a compensation of F [0,λ1] (Claim 5). We now define ξF : K → M(K) by ξF (α) := G(α) ν + a′δλi + b′δω1 ν + aξi(α) + b′δω1 ν + (a′ + b′)δω1 if α ∈ [0, λ1], if α = λi and 2 ≤ i < ω1, if α ∈ (λi, λi+1) and 1 ≤ i < ω1, if α = ω1,  where a := a′ a if a > 0 and a := 0 if a ≤ 0. We next check that ξF (α) is a compensation of F (α) for every α ∈ K. This is clear for α ∈ [0, λ1] (by the choice of G) and α ∈ {λi : 2 ≤ i < ω1} ∪ {∞} (by Claim 6). Take α ∈ (λi, λi+1) for some 1 ≤ i < ω1. Then µα is concentrated on [λi, λi+1] with µα(K) = a (Claim 3). Since ξi(α) is a compensation of µα (Claim 7), we have ξi(α) = 0 whenever a ≤ 0, while ξi(α)(K) = a and 0 ≤ ξi(α) ≤ µ+ α whenever a > 0. In any case, we have aξi(α)(K) = a′. Note also that F (α)(K) = F (ω1)(K) (by Claim 2, since α > λ0 ≥ β(0)). We now distinguish two cases. ON THE BISHOP-PHELPS-BOLLOB ´AS PROPERTY FOR NUMERICAL RADIUS IN C(K) SPACES 23 • If F (ω1)(K) > 0, then ξF (α)(K) = (ν + aξi(α) + b′δω1)(K) = ν(K) + a′ + b′ = F (ω1)(K) = F (α)(K) (bear in mind that ν + (a′ + b′)δω1 is a compensation of F (ω1), see Claim 6). Since ν ≤ µ+, aξi(α) ≤ ξi(α) ≤ µ+ α and b′ ≤ max{b, 0}, we conclude that 0 ≤ ξF (α) ≤ µ+ + µ+ α + max{b, 0}δω1 (4.13) = F (α)+. This shows that ξF (α) is a compensation of F (α). • If F (ω1)(K) ≤ 0, then ν = 0 and a′ = b′ = 0 (Claim 6), hence ξF (α) = 0 is the compensation of F (α). open set [0, λ1] ∪S1≤i<ω1 Finally, we check that ξF is ω∗-continuous. Observe that the continuity of ξF on the (λi, λi+1) follows at once from the ω∗-continuity of G and the ξis. On the other hand, for any 1 ≤ i < ω1, we have µλi+1 = aδλi+1 (Claim 4), hence ξi(λi+1) = max{a, 0}δλi+1 and so aξi(λi+1) = a′δλi+1. The last equality and the ω∗-continuity of ξi at λi+1 ensure that ξF is ω∗-continuous at λi+1. To finish the proof we show that ξF is continuous at ω1. Fix any A ∈ Clop(K). By Lemma 4.9 applied to the restriction of ξF (·)(A) to [0, ω1), there exists some αA < ω1 such that ξF (α)(A) = ξF (αA)(A) =: xA for all αA ≤ α < ω1. Choose 2 ≤ iA < ω1 such that λi ≥ αA for every iA ≤ i < ω1. Then xA = ξF (λi)(A) = (ν + a′δλi + b′δω1)(A) for every iA ≤ i < ω1, and so xA = (ν + a′δω1 + b′δω1)(A) = ξF (ω1)(A). The proof of the theorem is over. (cid:3) Remark 4.12. Using essentially the same arguments, one can prove by induction that the ordinal interval [0, ℵn] admits local compensation for each n ∈ N. It is not so clear to us what happens at ℵω. 4.3. Some open problems. Let K be an arbitrary compact space. (a) Does C(K) have the BPB property for numerical radius? (b) Does K admit local compensation if C(K) has the BPB property for numerical radius? (c) Is K metrizable if it admits a compensation function? (d) Does K admit a compensation function if it admits a BM(K)-compensation func- tion? (e) Does K admit a BM(K)-compensation function if it admits a closeness function? Acknowledgement. We wish to thank O. Kalenda for his valuable comments on Section 4 which led us to improve some results contained in a previous version of this manuscript. REFERENCES [1] M.D. Acosta, Operadores que alcanzan su radio num´erico, Ph.D. Thesis, Universidad de Granada, 1990. [2] M.D. Acosta, Every real Banach space can be renormed to satisfy the denseness of numerical radius attain- ing operators, Israel J. Math. 81 (1993), no. 3, 273 -- 280. MR 1231192 (94h:47003) [3] M.D. Acosta, F.J. Aguirre, and R. Pay´a, A space by W. Gowers and new results on norm and numerical ra- dius attaining operators, Acta Univ. Carolin. Math. Phys. 33 (1992), no. 2, 5 -- 14. MR 1287219 (95i:47009) [4] M.D. Acosta and R. Pay´a, Denseness of operators whose second adjoints attain their numerical radii, Proc. Amer. Math. Soc. 105 (1989), no. 1, 97 -- 101. MR 937841 (89f:47008) [5] M.D. Acosta and R. Pay´a, Numerical radius attaining operators and the Radon-Nikod´ym property, Bull. London Math. Soc. 25 (1993), no. 1, 67 -- 73. MR 1190367 (93j:47005) [6] I.D. Berg and B. Sims, Denseness of operators which attain their numerical radius, J. Austral. Math. Soc. Ser. A 36 (1984), no. 1, 130 -- 133. MR 720006 (84j:47004) 24 A. AVIL ´ES, A. J. GUIRAO, AND J. RODR´IGUEZ [7] F.F. Bonsall and J. Duncan, Numerical ranges of operators on normed spaces and of elements of normed algebras, London Mathematical Society Lecture Note Series, vol. 2, Cambridge University Press, London, 1971. MR 0288583 (44 #5779) [8] F.F. Bonsall and J. Duncan, Numerical ranges. II, Cambridge University Press, New York, 1973, London Mathematical Society Lecture Notes Series, No. 10. MR 0442682 (56 #1063) [9] C.S. Cardassi, Density of numerical radius attaining operators on some reflexive spaces, Bull. Austral. Math. Soc. 31 (1985), no. 1, 1 -- 3. MR 772627 (86e:47005) [10] C.S. Cardassi, Numerical radius attaining operators, Banach spaces (Columbia, Mo., 1984), Lecture Notes in Math., vol. 1166, Springer, Berlin, 1985, pp. 11 -- 14. MR 827753 (87c:47004) [11] C.S. Cardassi, Numerical radius-attaining operators on C(K), Proc. Amer. Math. Soc. 95 (1985), no. 4, 537 -- 543. MR 810159 (87a:47066) [12] R. Engelking, General topology, second ed., Sigma Series in Pure Mathematics, vol. 6, Heldermann Verlag, Berlin, 1989, Translated from the Polish by the author. MR 1039321 (91c:54001) [13] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, and V. Zizler, Banach space theory, CMS Books in Mathe- matics/Ouvrages de Math´ematiques de la SMC, Springer, New York, 2011, The basis for linear and nonlin- ear analysis. MR 2766381 [14] A.J. Guirao and O. Kozhushkina, The Bishop-Phelps-Bollob´as property for numerical radius in ℓ1(C), preprint, 2011. [15] J. Johnson and J. Wolfe, Norm attaining operators, Studia Math. 65 (1979), no. 1, 7 -- 19. MR 554537 (81a:47021) [16] V. Kadets, M. Mart´ın, and R. Pay´a, Recent progress and open questions on the numerical index of Banach spaces, RACSAM. Rev. R. Acad. Cienc. Exactas F´ıs. Nat. Ser. A Mat. 100 (2006), no. 1-2, 155 -- 182. MR 2267407 (2007h:46011) [17] R. Pay´a, A counterexample on numerical radius attaining operators, Israel J. Math. 79 (1992), no. 1, 83 -- 101. MR 1195254 (93j:47004) [18] A. Pełczy´nski, Linear extensions, linear averagings, and their applications to linear topological classifica- tion of spaces of continuous functions, Dissertationes Math. Rozprawy Mat. 58 (1968), 92. MR 0227751 (37 #3335) [19] B. Sims, On numerical range and its applications to Banach algebras, Ph.D. Thesis, University of Newcas- tle, 1972. DEPARTAMENTO DE MATEM ´ATICAS, FACULTAD DE MATEM ´ATICAS, UNIVERSIDAD DE MURCIA, 30100 ESPINARDO (MURCIA), SPAIN E-mail address: [email protected] INSTITUTO UNIVERSITARIO DE MATEM ´ATICA PURA Y APLICADA, UNIVERSIDAD POLIT ´ECNICA DE VA- LENCIA, CAMINO DE VERA S/N, 46022 VALENCIA, SPAIN E-mail address: [email protected] DEPARTAMENTO DE MATEM ´ATICA APLICADA, FACULTAD DE INFORM ´ATICA, UNIVERSIDAD DE MUR- CIA, 30100 ESPINARDO (MURCIA), SPAIN E-mail address: [email protected]
1806.03549	1	1806	2018-06-09T21:45:03	Constructions and properties of optimally spread subspace packings via symmetric and affine block designs and mutually unbiased bases	[ "math.FA", "math.MG" ]	We continue the study of optimal chordal packings, with emphasis on packing subspaces of dimension greater than one. Following a principle outlined in a previous work, where the authors use maximal affine block designs and maximal sets of mutually unbiased bases to construct Grassmannian $2$-designs, we show that their method extends to other types of block designs, leading to a plethora of optimal subspace packings characterized by the orthoplex bound. More generally, we show that any optimal chordal packing is necessarily a fusion frame and that its spatial complement is also optimal.	math.FA	math	CONSTRUCTIONS AND PROPERTIES OF OPTIMALLY SPREAD SUBSPACE PACKINGS VIA SYMMETRIC AND AFFINE BLOCK DESIGNS AND MUTUALLY UNBIASED BASES PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN ABSTRACT. We continue the study of optimal chordal packings, with emphasis on packing subspaces of dimension greater than one. Following a principle outlined in [12], where the authors use max- imal affine block designs and maximal sets of mutually unbiased bases to construct Grassmannian 2-designs, we show that their method extends to other types of block designs, leading to a plethora of optimal subspace packings characterized by the orthoplex bound. More generally, we show that any optimal chordal packing is necessarily a fusion frame and that its spatial complement is also optimal. 1. INTRODUCTION At a recent AMS meeting on Advances in Packings [54] at Ohio State University, several par- ticipants agreed that a seemingly inherent number-theoretic principle underpins the existence of most, if not all, Symmetric Informationally Complete Positive Operater Valued Measures, or simply SICs – a special type of chordal packing in complex projective space, CPm−1, consisting of m2 lines with constant pairwise chordal distance between its elements – objects which are of particular in- terest to quantum information theorists [68, 1, 58]. Accordingly, some speakers advocated that – by developing a tabulation of analytic and numerical constructions of SICs for different parameters (ie, for each dimension, m) – the community might eventually extrapolate an existential, or even constructive, proof of Zauner's conjecture, which states that SICs exist in every dimension. This work is motivated by a similar but generalized principal. By developing a parametric tabula- tion for all optimally spread packings with respect to the chordal distance in arbitrary Grassmannian manifolds, we hope to contribute to a fundamental theorem for the construction and structure of all optimally spread packings. With regard to developing a universal theory of said structure, in Section 3, we prove that all optimally spread packings are fusion frames. In addition, we show that their spatial complements are also optimally spread. Toward the development of the aforementioned tabulation of solutions, we demonstrate elemen- tary constructions of optimally spread subspace packings for parameter sets characterized by the simplex bound; less trivially, we construct optimal packings characterized by the orthoplex bound (see Section 4 for information about both bounds) by exploiting a technique presented in [12] which led to a special class of both real and complex solutions, so-called maximal orthoplectic fu- sion frames, which had previously been constructed for the real case in [59] by other means. In this vein, we re-present an abbreviated proof of the method of construction from [12], empha- sizing its dependence on the existence of certain families of mutually unbiased bases and certain block designs, along with a few other insights. Finally, we exploit the existence of numerous other classes of block designs which – omitted in [12] – which, to our surprise also satisfy the restrictive conditions of the construction, thereby by producing more optimally spread fusion frames charac- terized by the orthoplex bound. 1991 Mathematics Subject Classification. 42C15. Key words and phrases. optimal chordal packings, tight fusion frames, orthoplex bound, mutually unbiased bases, block designs. All authors of this paper were supported by NSF DMS 1609760, NSF DMS 1725455 and ARO W911NF-16-1-0008. 1 2 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN 2. PRELIMINARIES Throughout, we assume that l, m, and n are positive integers satisfying l ≤ m ≤ n and that F = R or F = C. Recall that the Grassmannian manifold, denoted G(l, Fm), is the space of all l-dimensional subspaces in the Hilbert space, Fm. We refer to a finite sequence of subspaces, {Wj}n j=1 ⊂ G(l, Fm), as an (n, l)-packing for Fm, or simply a packing when the context is clear. As a minor abuse of notation, we interchangeably refer to a given packing P either by its se- j=1, or its corresponding sequence of unique orthogonal projections, quence of subspaces, P = {Wj}n P = {Pj}n j=1: ie, im(Pj) = Wj, tr(Pj) = l, and Pj = P2 j = P∗ for every j ∈ {1, . . . , n}. To be precise, in Section 3, it is convenient to alternate between these equivalent interpretations, while in the remaining sections we primarily identify packings as sequences of orthogonal projec- tions. As considered in [25, 5], there are numerous notions of distance that one may define between two subspaces. This work concerns the study and construction of packings which are "optimally spread" with respect to the chordal distance. 2.1. Definition. Given two l-dimensional subspaces of Fm with corresponding orthogonal projec- tions P and P ′, the chordal distance between them is dc(P, P ′) := 1 √2kP − P ′k =(cid:0)l − tr(PP ′)(cid:1)1/2 . With respect to this objective function, our goal is to study and construct packings which maxi- mize the minimal pairwise distance over the space of all n-packings in G(l, Fm). To facilitate this, we define and denote the (chordal) coherence of an (n, l)-packing, P = {Pj}n j=1, for Fm as µ(P) := max 1≤j,j ′≤n tr(PjPj ′ ) and the packing constant as j6=j ′ µn,l,Fm := inf P⊂G(l,Fm) µ(P). Formally, we say an (n, l)-packing, P, for Fm is optimally spread if µ(P) = µn,l,Fm . An elementary topological argument ensures that optimally spread packings exist for all parameters satisfying l ≤ m ≤ n. 2.2. Proposition. An optimally spread (n, l)-packing for Fm exists. Proof. The Grassmannian manifold, G(l, Fm), is well known to be compact; whence, the space of all j=1G(l, Fm) – is also (n, l)-packings for Fm – which can be identified with the Cartesian product Πn compact. Because the coherence function is continuous, the claim follows by the extreme value theorem. (cid:3) These definitions and the preceding proposition imply that the following problem well-posed. Formally, the (constant-rank) chordal packing problem is stated as follows. 2.3. Problem. [The (constant-rank) chordal packing problem] For each quintuplet of parameters, (n, l, m, F), determine the corresponding packing constant, µn,l,Fm , and, if possible, construct and characterize an optimally spread n-packing, P ⊂ G(l, Fm). From an applications point of view, we are especially interested in (n, l)-packings endowed with additional spectral constraints, which engender numerous signal processing possibilities [22, 2, 51, 61, 16, 62, 37, 4, 11, 50, 3, 57, 23]; in particular, we are interested in fusion frames. OPTIMALLY SPREAD SUBSPACE PACKINGS 3 2.4. Definition. An (n, l)-packing for Fm, P = {Pj}n projections' sum, Pn Any (n, l) fusion frame is tight if this sum satisfies j=1, is an (n, l)-fusion frame for Fm if the j=1 Pj, is positive definite, and whenever l = 1, then P is also called a frame. nXj=1 Pj = nl m Im, where – henceforth – Im denotes the m × m identity matrix. An (n, l)-fusion frame for Fm is Grassmannian if it is optimally spread. 2.5. Remark. (1) In the case l = 1, where a packing is a frame, the projectors are usually - up to a choice of unimodular phasing - identified with unit vectors. (2) The term "Grassmannian fusion frame" is an acknowledgement of [62], where the authors were interested specifically in frames (ie, the case l = 1), or equivalently packings in pro- jective space where there is no ambiguity in the use of the term "Grassmannian"; however, because we are now studying packings in arbitrary Grassmannian manifolds, we prefer the term "optimally spread" over "Grassmannian" in order to avoid confusion when describing solutions to Problem 2.3. While it is false that all n-packing are fusion frames, fortunately for applications - as we show next - all optimally spread (n, l)-packings are fusion frames with the corresponding parameters. 3. SOME PROPERTIES OF OPTIMAL PACKINGS In this section, we prove two important facts about optimal packings. The first concerns their spanning properties. 3.1. All optimal packings are fusion frames. It is known [32] that if an (n, 1)-packing for Fm is an optimally spread packing, then it is a Grassmannian frame, but we are unaware of an analogous statement for the general case of optimal (n, l)-packings, 1 ≤ l ≤ m. In the following, we show that all optimal packings are fusion frames. To prove this, we begin with two lemmas. 3.1. Lemma. Suppose µn,l,Fm , δ and ǫ are positive real numbers satisfying µn,l,Fm > ǫ and (1) If {xj}l j=1 , {yj}l j=1 and {zj}l j=1 are sequences of unit vectors in Fm satisfying 2plδ(µn,l,Fm − ǫ) + lδ < ǫ 2 . (2) and (3) then lXj=1 lXk=1 hxj, yki2 < µn,l,Fm − ǫ lXj=1 kzj − yjk2 < δ, lXj=1 lXk=1 hxj, zki2 < µn,l,Fm − ǫ 2 . 4 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN Proof. Rewriting each zj as zj = yj + (zj − yj) yields lXj=1 lXk=1 lXj=1 lXk=1 hxj, zk − yki2. Next, the Cauchy-Schwarz inequality and (2) imply lXk=1 lXj=1 hxj, zki2 ≤vuut vuut hxj, zki2 ≤ √µn,l,Fm − ǫ +vuut lXk=1 hxj, yki2 +vuut lXk=1 lXj=1 kxjk2kzk − ykk2 < √µn,l,Fm − ǫ + √lδ, lXj=1 vuut (cid:16)√µn,l,Fm − ǫ + √lδ(cid:17)2 where the last inequality in the previous line follows from Equation (3) and the unit norm property of the vectors. The claim follows by squaring and applying Equation (1) to obtain = µn,l,Fm − ǫ + 2plδ(µn,l,Fm − ǫ) + lδ < µn,l,Fm − ǫ + ǫ 2 = µn,l,Fm − ǫ 2 . (cid:3) To facilitate the next lemma and the theorem that follows, we say an element Pk (or its image, Wk) of an optimally spread (n, l)-packing, P = {Pj}n j=1, for Fm achieves the packing constant if there exists k ′ ∈ {1, 2, ..., n} with k 6= k ′ such that tr(PkPk ′ ) = µn,l,Fm , and we call each element Pk ′ satisfying this condition a packing neighbor of Pk. 3.2. Lemma. Suppose P = {Pj}n j=1 is an optimally spread (n, l)-packing in Fm with corresponding subspaces {Wj}n j=1 and, furthermore, suppose that µn,l,Fm > 0. If Wk is an element of P that achieves the packing constant, then it contains a unit vector which is not orthogonal to any of its packing neighbors. Proof. Write I := {1 ≤ i ≤ n : tr(PiPk) = µn,l,Fm }, and for each i ∈ I, let Vi denote the maximal subspace of Wk which is orthogonal to Wi, ie Vi := Wk ∩ W⊥ i . The assumption µn,l,Fm > 0 implies that every Vi is a proper subspace of Wk, and since a linear space cannot be written as a finite union of proper subspaces, it follows that Wk\ ∪i∈I Vi is nonempty, so the claim follows. (cid:3) 3.3. Theorem. If P = {Pj}n {Wj}n j=1 is an optimally spread (n, l)-packing in Fm with corresponding subspaces j=1 and K := {1 ≤ k ≤ n : Wk achieves the packing constant} , then span {Wk : k ∈ K} = Fm, meaning {Wk}k∈K is an (n ′, l)-fusion frame for Fm, where n ′ := K. Consequently, P is an (n, l)-fusion frame for Fm and, in particular, a Grassmannian fusion frame. Proof. If µn,l,Fm = 0, then K = {1, 2, . . . , n}, so Wj is in the orthogonal complement of Wj ′ for every j 6= j ′, and since nl ≥ m, it follows that span {Wj}n j=1 = Fm. For the case µn,l,Fm > 0, we proceed by way of contradiction, iteratively replacing elements of P that achieve the packing constant in such a way that we eventually obtain a new (n, l)-packing in Fm with coherence strictly less than µn,l,Fm , which cannot exist. With the contradictory approach in mind, fix a unit vector z ∈ Fm so that z is in the orthogonal complement of span {Wj}j∈K. Next, we describe the replacement procedure. The first step is to choose some k ∈ K, write Ik := {1 ≤ i ≤ n : tr(PiPk) = µn,l,Fm } and I c k := {1 ≤ i ≤ n : tr(PiPk) < µn,l,Fm } , and for every i 6= k, fix an orthonormal basis {xi,j}l j=1 for Wi. By Lemma 3.2, there exists a unit vector xk,1 ∈ Wk which is nonorthogonal to all of Wk's packing neighbors, so apply the Gram- Schmidt algorithm to extend xk,1 to an orthonormal basis, {xk,j}l j=1, for Wk. As defined, there exists ǫ > 0 so that OPTIMALLY SPREAD SUBSPACE PACKINGS 5 lXj=1 lXj ′=1(cid:12)(cid:12)hxi,j, xk,j ′i(cid:12)(cid:12) 2 = tr(PiPk) < µn,l,Fm − ǫ, for all i ∈ I c k. Choose δ ∈ R so that 0 < δ < 1/2 and (1) from Lemma 3.1 is satisfied, define yk,1 :=p1 − δ2 xk,1 + δz and yk,j := xk,j for 2 ≤ j ≤ l, Noting that 1 − δ2 < √1 − δ2 implies 2(cid:16)1 − √1 − δ2(cid:17) < 2δ2, we estimate and define Vk := span {yk,j}l it follows by elementary computation that {yk,j}l j=1 is an orthonormal basis for Vk. lXj=1 kxk,j − yk,jk2 = kxk,1 − yk,1k2 = 2(cid:16)1 −p1 − δ2(cid:17) < 2δ2 < δ. j=1 with corresponding orthogonal projection, Qk. Because hxk,1, zi = 0, Therefore, Lemma 3.1 implies tr(PiQk) = lXj=1 Furthermore, the nonorthogonality of xk,1 with Wi for every i ∈ Ik implies 2 lXj ′=1(cid:12)(cid:12)hxi,j, yk,j ′i(cid:12)(cid:12) < µn,l,Fm − ǫ 2 for all i ∈ I c k. Thus, replacing Wk with Vk produces a new (n, l)-packing for Fm, where the replaced element no long achieves the packing constant. Now, we iterate this replacement procedure. After at most n ′ repetitions of this process, we obtain a final (n, l)-packing, with coherence strictly less than µn,l,Fm , the desired contradiction, so the claims follow. (cid:3) Next, we consider the spatial complements of Grassmannian frames. 3.2. Spatial complements of Grassmannian fusion frames. Given an (n, l)-packing, P = {Pj}n j=1, for Fm, its spatial complement is the (n, m − l)-packing, P ⊥ := {Im − Pj}n j=1, where Im denotes the m × m identity matrix. As one might expect, the spatial complement of a Grassmannian fusion frame is also optimally spread, which we show in the following theorem. 3.4. Theorem. If P = {Pj}n ment, P ⊥, is a Grassmannian (n, m − l)-fusion frame for Fm and µn,m−l,Fm = m − 2l + µn,l,Fm . Proof. Given any (n, l)-packing for Fm, say Q := {Qj}n complement, Q⊥, j=1 is a Grassmannian (n, l)-fusion frame for Fm, then its spatial comple- j=1, we compute the coherence of its spatial µ(cid:16)Q⊥(cid:17) = max j6=j ′ tr(cid:0)(Im − Qj)(Im − Qj ′ )(cid:1) = m − 2l + max j6=j ′ tr(cid:0)QjQj ′(cid:1) = m − 2l + µ (Q) , thereby showing that the coherence of Q⊥ depends only on the coherence of Q. Thus, µ(cid:0)P ⊥(cid:1) = m − 2l + µn,l,Fm is minimal over the space of all (n, m − l)-packings for Fm, so P ⊥ is optimally spread and therefore a Grassmannian (n, m − l)-fusion frame for Fm by Theorem 3.3. (cid:3) 0 < lXj=1 hyk,1, xi,ji2 = (1 − δ2) so it follows that tr(PiQk) = lXj=1 lXj ′=1(cid:12)(cid:12)hyk,j, xi,j ′i(cid:12)(cid:12) 2 < lXj=1 hxk,1, xi,ji2 < lXj=1 lXj ′=1(cid:12)(cid:12)hxk,j, xi,j ′i(cid:12)(cid:12) lXj=1 hxk,1, xi,ji2 for all i ∈ Ik, 2 = tr(PkPi) = µn,l,Fm for all i ∈ Ik. 6 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN 4. RECALLING A RECIPE With an equivalence between optimally spread n-packings and Grassmannian fusion frames of corresponding parameters established, this section is dedicated to the construction of infinite fam- ilies of tight, optimally spread fusion frames. Our constructions rely on three things: (I) the most successfully obtained lower chordal coherence bounds [67, 34, 15, 6] – the sim- plex and orthoplex bounds – often attributed to Rankin [55] and Conway, Hardin, and Sloane [25], and Welch [66], which motivate ours study, (II) two major ingredients, the existence of maximal sets of mutually unbiased bases in certain settings [67, 18] and the existence of certain 1-block designs [24], and (III) the technique, presented in [12], which combines the two ingredients to reveal numerous infinite families of optimally spread packings. 4.1. Two coherence bounds. The simplex and orthoplex bounds admit several manifestations, for example as seen in [55, 66, 25, 49, 8]. Of relevance to this work is their occurrence (i) in the coding problem, where one maximizes the minimal distance between a prescribed number of points on a real sphere of fixed dimension (see [55, 29, 8, 7]), (ii) its connection with a constrained version of the coding problem induced by the chordal packing problem via the so-called l-tracelesss map, henceforth denoted Tl, which seemingly laid the foundation for the mainstream chordal packing arena, due to Conway, Hardin, and Sloane in 1996 [25], and (iii) the manner with which the constrained coding problem is equivalent to the chordal packing problem. Given 1 ≤ l ≤ m, the l-traceless map is (4) Tl : G(l, Fm)→ RdFm : P 7→ V(cid:18)P − l m Im(cid:19) ∈ V(H), where H denotes the corresponding m×m matrix subspace of trace zero elements in Fm×m, V : H→ RdFm is a fixed vectorization isomorphism that maps the traceless symmetric/hermitian matrices of H to vectors in RdFm , and where the vanishing trace constraint implicates the two isomorphic spaces' real dimension, (5) dFm :=(cid:12) (m+2)(m−1) m2 − 1, 2 , F = R F = C . 4.1. Notation. Henceforth, as computed above, we denote by dFm the real dimension of the so- called "traceless' space", H, or equivalently, the dimension of real vectorized space, V(H) into which an element of G(l, Fm) embeds via the l-traceless map. It is elementary to verify that Tl is a (scaled) isometry [25, 14]; in particular, for elements and the vectorization, V, converts the trace inner product between points in the matrix subspace H into the standard inner product between points in RdFm . 4.2. Notation. As in the last line and Equation 6, if no other indexing scheme is established – which will occur occassionally in this work – projections embedded via Tl are sub-indexed by their under- lying projection, and - although unnessary - it is convenient to record the underlying projections' ranks in the superscripts. P, P ′ ∈ G(l, Fm), the l-traceless map enforces the traceless identity: P ′E , (6) where the embedded, rescaled-to-unit vectors are tr(cid:0)PP ′(cid:1) = l(m − l) l2 m + (l) (l) P , v m Dv kTl(P ′)k ∈ S dFm −1 ⊂ RdFm , Tl(P ′) (l) P := v Tl(P) kTl(P)k , v (l) P ′ := OPTIMALLY SPREAD SUBSPACE PACKINGS 7 Of significance here and in future work (where we are studying optimally spread mixed rank packings), placing the appropriate "min max" statements in front of the left and right hand sides of Equation 6 converts Problem 2.3, the chordal optimization problem, into a restricted coding problem. 4.3. Definition. Given any d, n ∈ N, an n-code, C, is a sequence of n unit vectors on the real unit sphere in Rd, ie C := {vi}n 4.4. Problem. [The restricted coding problem] Let Ω := Tl(G(l, Fm)) ⊂ S dFm −1 ⊂ RdFm , the normalized image of G(l, Fm) under the l-traceless map. Determine i=1 ⊂ S d−1 ⊂ Rd. σn,l,Fm := inf C:={vi}n i=1⊂Ω i6=j hvi, vji max 4.5. Remark. As with the chordal packing problem, we reserve the symbol σn,l,Fm to refer to the solution to this problem. Because the l-traceless map is continuous, an elementary topological proof similar to that of Proposition 2.2 verifies the existence of solutions for all parameters in the aforementioned problem. Perhaps the more obvious proof is that the chordal packing problem was already shown to be well- defined, so its equivalence to Problem 4.4 via the traceless relationship assures well-definedness. Indeed, any solution to Problem 4.4 resolves to a corresponding solution for Problem 2.3. 4.6. Solution (General solution). For every quadruple of parameters, (n, l, m, F), (7) µn,l,Fm = l2 m + l(m − l) m σn,l,Fm . Unfortunately, few – in fact, finitely many [39], except in R2 [10]) – solutions for this plethora of problems are known which are characterized neither by the simplex nor orthoplex bounds, which – incidentally – represent a complete set of tight bounds for the unrestricted coding problem, under suitable conditions. 4.7. Problem. [The unrestricted coding problem ] Let d, n ∈ N. Determine τn,d := C:={vi}n inf i=1⊂S d−1 i6=j hvi, vji max 4.8. Remark. As with preceding problems, the symbol τn,d is reserved for solutions to this problem. As early as 1955, Rankin [55] provided a perfect solution for the unrestricted coding problem for all dimensions d, n ∈ N satisfying 1 ≤ n ≤ 2d, providing sharp bounds, replete with examples of optimizers: namely the vertices of simplices and (partial) orthoplexes. 4.9. Remark. Orthoplex is increasingly used term (particularly within the packing community [60, 14]) to refer to the vectors corresponding to the vertices of an ℓ1-ball; in other words, an orthoplex is an orthonormal basis unioned with its antipodes. 4.10. Theorem ([55], [26]; see also [66]). Let d, n ∈ N. • Simplex Bound: If 1 ≤ n ≤ d+1, then τn,d = − 1 n−1 and any n-code, C ⊂ S d−1, that achieves this bound is necessarily a regular (n−1)-simplex, meaning all pairwise inner products among the code's elements equal τn,d and dim (span{C}) = n − 1. • Orthoplex Bound: If d + 1 < n ≤ 2d, then τn,d = 0 and any n-code, C ⊂ S d−1, that achieves this bound necessarily contains at least two orthogonal vectors, and all other pairwise inner products occurring within the code are non-positive. Moreover, if n = 2d, then the code is necessarily a complete orthoplex. • Beyond the orthoplex range If n > 2d, then τn,d > 0 and - except for a few sporadic instances [39, 32] – little is known other than computer-assisted punitively optimal examples, for example, those tabulated in Sloane's online tables of chordal packings [60]. 8 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN If solutions to the restricted and unrestricted coding problem equate, ie σn,l,Fm = τn,dFm , then Equation 7 yields the analytic solution to the chordal packing problem for the given parameters via the "traceless lifting". Whenever a given parameter quadruple (n, l, m, F), admits a simplex as a solution to both problems, the traceless identity implicates the existence of an equiangular tight fusion frame, where a packing is equiangular if the set of pairwise trace inner products between its elements is a singleton. Of particular note, for the case l = 1, such objects are more commonly referred to as equiangular tight frames (ETFs) and are probably the most famous and well-studied class of optimally spread packings [38, 9, 13, 17, 28, 31, 30, 33, 34, 35, 40, 43, 44, 45, 46, 51, 52, 53, 63, 64, 65]. Indeed, numerous infinite families are known to exist [34] and dozens [38, 9, 13, 17, 28, 31, 30, 33, 34, 35, 40, 43, 44, 45, 46, 51, 52, 53, 63, 64, 65] – if not hundreds – of mathematicians have contributed to this study, many of whom (see [34]) are engaged in ETF research concurrently with the preparation of this document. As for the orthoplex bound, again with l = 1 fixed, at least three infinite families [14, 67, 18] of tight frames and a few sporadic instances [41] of parameter quadruples are known to exist where solutions to Problem 4.4 and Problem 4.7 coincide at the orthoplex bound, meaning the conditions on the cardinality range, n > dFm + 1 and the coherence, σn,1,Fm = τn,dFm = 0, are satisfied. Included among these three families are maximal sets of mutually unbiased bases (MUBs) - to be discussed in further detail shortly, a key ingredient for the constructions of optimally spread packings we are building toward. Unfortunately, by the inherent restriction of Problem 4.4, solutions to the two coding problems do not always coincide, thereby preventing us from "pulling" Rankin's solutions back through the traceless identity; otherwise, ETFs would always exist for appropriate parameters. For example, it is known [10] that an ETF consisting of five elements in R3 cannot exist, implicating the inequality σ5,1,R3 > τ5,dR3 = − 1 4 . Similarly, the nonexistence of an ETF of eight elements in C3 was recently verified [65], enforcing the inequality σ8,1,C3 > τ8,dC3 = − 1 7 . For more information, we recommend the living table of known ETFs and their properties, located at [34]. More generally, for all parameter quadruples, the relationship between solutions to the restricted and unrestricted coding problem may be described succinctly by the inequality, Substituting this inequality into Equation 7 yields the corresponding lower bounds for the chordal coherence of packings for suitable parameter sets. σn,l,Fm ≥ τn,dFm . 4.11. Theorem. [[55], [26]; see also [66] and [12]] (1) Simplex bound: If P is an (n, l)-packing for Fm, then µ (P) ≥ nl2 − ml m(n − 1) , and equality is achieved if and only if the fusion frame is equiangular and tight. (2) Orthoplex bound: If P is an (n, l)-packing for Fm and n > dFm + 1, then µ (P) ≥ l2 m , and if equality is achieved then P is optimally spread and n ≤ 2dFm. We re-iterate that optimally spread fusion frames characterized by the simplex bound are called equiangular tight fusion frames (ETFFs). As the name suggests, they are indeed both equiangular and tight. Equiangularity follows immediately from the fact that the l-traceless map embeds a given ETFF into a regular simplex, which is itself equiangular. Moreover, regular simplexes are zero summing; if P is an equiangular, tight (n, l)-fusion frame for Fm and (cid:10)v (l) Pi(cid:11)n i=1 denotes the OPTIMALLY SPREAD SUBSPACE PACKINGS 9 v(l) l m lm − l2 nXi=1(cid:18)Pi − Im(cid:19) =r m Pi! =r m embedded simplex under Tl, then an application of the inverse of the vectorization isomorphism, V −1, applied to the overall sum of the embedded, rescaled vectors yields the de-vectorized identity, (8) 0m×m = V −1(0dFm ) = V −1 nXi=1 Im(cid:19) , whereq m More generally, if P ⊂ G(l, Fm) is any (n, l)-fusion frame, n > dFm + 1 and µ (P) = l2 lm−l2 is the rescaling factor, proving the tightness of all ETFFs according to Definition 2.4. m , then P is optimally spread according to the preceding theorem, in particular characterized by the orthoplex bound. We call such objects orthoplectic Grassmannian (ie, optimally spread) (n, l)-fusion frames or (n, l)-OGFFs for Fm, or simply n-OGFs when l = 1. In general, OGFFs are not tight; for a thorough examination of this phenomenon, see [21]. Nevertheless, the three infinite families of OGFs referred to in [14, 67] are tight and all of the OGFFs to be constructed in this paper are tight. Finally, we call an OGFF with n = 2dFm a maximal OGFF, because its tracelessly embedded vectors form a full orthoplex, which is zero-summing, thereby implying tightness according to the same argument for ETFFs; that is, Equation 8 applies for maximal OGFFs. nXi=1(cid:18)Pi − lm − l2 nl m With lower coherence bounds established, we commit the next two subsections to the develop- ment of a recipe, previously presented in [12], which generates tight OGFFs when certain criteria are satisfied. The recipe depends on two major ingredients: mutually unbiased bases and block designs. 4.2. Two ingredients. Because notation becomes somewhat cumbersome as we move forward, when given a natural number, m, we will occasionally denote the corresponding index set as fol- lows: (cid:2)[m](cid:3) := {1, 2, ..., m} . 4.2.1. Mutually Unbiased Bases. Motivated by consistency and, again, to mitigate notational and typographical issues that will arise in the next section, we make a slight modification to the usual definition [6, 18, 36, 42, 47, 56, 67] of mutually unbiased bases – they are usually defined as sets of orthonormal bases with a special property. Equivalently, we reformulate them as families of tight fusion frames comprised of rank one projectors. 4.12. Definition. If Π1 = (cid:10)π Fm, then they are mutually unbiased if for j, k ∈ (cid:2)[m](cid:3) and s, t ∈ (cid:2)[2](cid:3), the trace inner products are a pair of tight (m, 1)-fusion frames for j (cid:11)m j (cid:11)m satisfy j=1 j=1 (1) (2) A family of pairwise mutually unbiased bases are simply called mutually unbiased bases, or MUBs. 4.13. Remark. To be clear, by Pareseval's identity, the elements of any tight (m, 1)-fusion frame, Π = {πj}m j=1, for Fm (ie, noting m = n here) must arises from an orthonormal basis for Fm, because the tightness property in Definition 2.4 reduces to mXi=1 πi = Im. 4.14. Notation. In this vein, the symbol, Π, will always refer to a tight (m, 1)-fusion frame for Fm arising from an orthonormal basis. and Π2 = (cid:10)π k (cid:17) =  j π(t) 1 m , 0, 1, tr(cid:16)π(s) s 6= t s = t, j 6= k s = t, j = k . 10 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN Referring back to the orthoplex bound from Theorem 4.11, the coherence of a family MUBs equates with the orthoplex bound when l = 1, so a sufficiently large family of MUBs can form a Grassmannian frame by the theorem. However, there are many cases where the number of MUBs in Fm is known or conjectured [18, 42] to be too small to satisfy the cardinality requirement for the orthoplex bound in Theorem 4.11. Of seeming relevance is that the number of MUBs is bounded above in terms of the ambient dimension and underlying field. 4.15. Theorem. [Delsarte, Goethals and Seidel [27]] If {Πj}k j=1 is a family of MUBs for Fm, then [F : R] 2 k ≤ m + 1. Unfortunately, our recipe benefits most greatly from the existence of large families of MUBs, but we encounter frequent deficiencies. For example, this is especially evident in the real case, because most dimensions, m, admit no more than three MUBs in Rm [18]; fortunately, whenever m is a power of four [20], real families of MUBs exist that achieve the upper bound in Theorem 4.15. The complex case is a little less severe, as the MUBs' upper cardinality bound is achieved whenever the dimension, m, is a prime power [67], although – even for the complex case – evidence [42] suggests that 7 MUBs likely do not exist in C6, the first complex vector space of composite dimension. 4.16. Theorem. [[20, 67]] If m is a prime power, then a family of m + 1 pairwise mutually unbiased bases for Cm exists. If m is a power of 4, then a family of m/2 + 1 pairwise mutually unbiased bases exists for Rm. With Theorem 4.15 and Theorem 4.16 in mind, we abbreviate kRm := m/2 + 1 and kCm := m + 1, and say a family of k MUBs in Fm is maximal if k = kFm. The second ingredient of the recipe are block designs. 4.2.2. Block designs. As a further effort to avoid convoluted notation and also to illuminate the role of block designs in our construction of tight OGFFs, we diverge somewhat from the conventional symbology found in most combinatorial literature. Typically, the symbols v, b, r, k and λ are desig- nated as the parameters for specifying a given block t-design. To clarify, in the following definition, the symbol v is supplanted by m and l replaces k. 4.17. Definition. A t-(m, l, λ) block design, is a pair(cid:0)(cid:2)[m](cid:3), S(cid:1), where S, is a collection of subsets of (cid:2)[m](cid:3), called blocks, where each block J ∈ S has cardinality l, the cardinality of S, or number of blocks, is b = S, each element of(cid:2)[m](cid:3) occurs in exactly r blocks, and such that every subset of (cid:2)[m](cid:3) with cardinality t is contained in exactly λ blocks. When the parameters are not important or 4.18. Notation. From here on, the pair of symbols (cid:0)(cid:2)[m](cid:3), S(cid:1) refers to a t-block design, where the implied by the context, then S is also referred to as a t-block design. design's - often suppressed - parameters are prescribed as above. A few simple facts about block designs are collected below. 4.19. Proposition. Any such block design satisfies the following conditions: (1) mr = bl, and (2) r(l − 1) = λ(m − 1). Furthermore, it is immediate that for any t > 1, a t-block design is also a (t − 1)-block design. Although we will exploit t-block designs with many values for t, the forthcoming construction only depends on the existence of certain 1-block designs, so, henceforth, we will regard all t-block designs as 1-block designs. With the basic facts about MUBs and block designs surmised, we are ready to lay out the recipe for tight OGFFs. OPTIMALLY SPREAD SUBSPACE PACKINGS 11 4.3. A recipe for tight orthoplectic Grassmannian fusion frames. The key to the recipe is choos- ing a set of MUBs and a block design whose respective parameters align in such a way that each of the design's blocks inform us on how to choose l elements from a given MUB, which are then summed to form a rank l projector, constituting one of the n elements of the tight (n, l)-OGFF over Fm to be constructed; critical to this "alignment" of parameters is that the procedure produces sufficiently many projections in order for the orthoplex bound to apply (ie, we need n > dFm + 1) In pursuit of the idea of using a 1-design's blocks to select rank one projections for us, we define and that coherence ultimately equals the orthoplex bound (ie, we also need µ(P) = l2 m ). the J -block projection as follows. 4.20. Definition. Given a tight (m, 1)-fusion frame, Π = {πj}m block J ⊂ S, then the J -block projection with respect to Π is j=1, for Fm, a 1-(m, l, λ) design, and a PJ =Xj∈J πj. In this case, writing S = {J1,J2, . . . ,Jb}, then the n := b = S-packing, P :=(cid:8)PJj(cid:9)b j=1 ⊂ G(l, Fm), is called the block packing with respect to Π. 4.21. Proposition. Given a tight (m, 1)-fusion frame, Π = {πj}m j=1 for Fm and a 1-(m, l, λ) block design where S = {J1,J2, . . . ,Jb}, then the block packing with respect to Π forms a tight (n, l, m)- fusion frame for Fm, where n = b. Proof. By the preceding definition, every projection formed in this way has rank equal to l and, because S is a 1-design, every singleton {j} ⊂ (cid:2)[m](cid:3) occurs exactly λ times among the design's blocks. Thus, bXj=1 PJj = λ mXj=1 πj = λIm. (cid:3) The following fact is perhaps the most crucial and surprising aspect to the construction. Given a pair of mutually unbiased bases Fm and a suitable block design, one can select block projections from the respective MUBs to form a pair of tight (b, l, m)-fusion frames for Fm that achieve the orthoplex bound pair-wise. 4.22. Proposition. Given Πi = (cid:10)π 1-(m, l, λ) block design, then (i) j (cid:11)m j=1 l2 , m where PJ is the J -block projection with respect to Π1 and P ′ Proof. We compute J ′(cid:1) = , i ∈ (cid:2)[2](cid:3), a pair of MUBs for Fm, and J ,J ′ ⊂ S from a tr(cid:0)PJ P ′ J ′ is J ′-block projection for Π2. tr(PJ P ′ J ′) =Xj∈J Xj ′∈J ′ (1) j π tr(cid:16)π (2) j ′ (cid:17) = 1 mXj∈J Xj ′∈J ′ 1 = l2 m . (cid:3) With Propositions 4.21 and 4.22 in mind, we see that it is fairly simple to form families of tight fusion frames which achieve the orthoplex bound pairwise. However, in order to form optimally 12 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN spread packings, we must satisfy the sufficiency conditions for the orthoplex bound from Theo- rem 4.11. To reiterate, we need (a) n > dFm + 1 projections in our packing and we need (b) the internal coherence of each block packing arising from each MUB to be sufficiently low. To facilitate condition (b), we follow [12], recalling their notion of a block design's cohesion. 4.23. Definition. Let (X, S) be a 1-(m, l, λ) design. We say that (X, S) is c-cohesive if there exists c > 0 such that Finally, the desired recipe for tight OGFFs, originally presented in [12], is described in the fol- max J ,J ′∈S J 6=J ′ (cid:12)(cid:12)J ∩ J ′(cid:12)(cid:12) ≤ c. lowing theorem. (k) j (cid:11)m k∈(cid:2)[K](cid:3) for each k ∈ (cid:2)[K](cid:3). If P 4.24. Theorem. Let(cid:0)(cid:2)[m](cid:3), S(cid:1) be an l2/m-cohesive 1-(m, l, λ) design, where b := S and let {Πk} be a set of MUBs for Fm, where Kb > dFm + 1 and write Πk = (cid:10)π denotes the J -block projection with respect to Πk, then the set J (cid:11)J ∈S (cid:10)P P = [k∈(cid:2)[K](cid:3) Proof. By Proposition 4.21,(cid:10)P is a tight fusion frame for each k ∈(cid:2)[K](cid:3), so P is also a tight fusion frame. The cardinality requirement is satisfied since n > dFm + 1. Let k, k ′ ∈(cid:2)[K](cid:3). If k 6= k ′, forms a tight OGFF for Fm consisting of n rank l projectors, where n = Kb. J (cid:11)J ∈S then (k) J j=1 (k) (k) for every J ,J ′ ∈ S by Proposition 4.22. If k = k ′, then the fact that ((cid:2)[m](cid:3), S) is an l2/m-cohesive design yields tr(cid:16)P (k) J P (k ′) J ′ (cid:17) = l2 m max J ,J ′∈S J 6=J ′ tr(cid:16)P (k) J P (k) J ′(cid:17) = max J ,J ′∈S (k) j π (k) j ′ (cid:17) j ′∈J ′ J 6=J ′ Xj∈J tr(cid:16)π J 6=J ′ (cid:12)(cid:12)J ∩ J ′(cid:12)(cid:12) = max J ,J ′∈S l2 m , ≤ Altogether, with respect to the aforementioned (a) cardinality issue in mind, it is natural to begin which shows that P is a tight OGFF, as characterized by Theorem 4.11. with a maximal set of MUBs, and then seek a compatible l2/m-cohesive 1-block design ((cid:2)[m](cid:3), S) that produces a sufficient number of coordinate projections per MUB. This strategy was employed in the original work [12], and we similarly depend on the existence maximal MUBs in the following constructions, as well. Recall that the known cases of existence of maximal MUBs are described in Theorem 4.16. For each prime power q in the complex case or power of four, q = 4t, in the real case, we seek a l2 m -cohesive 1-(m, l, λ), where m = q and where the number of blocks, b, satisfies (cid:3) (9) b > dFm + 1 kFm . OPTIMALLY SPREAD SUBSPACE PACKINGS 13 5. SOLUTIONS FOR THE CHORDAL PACKING PROBLEM In [12], the authors laid out the preceding construction technique, and then went on to exploit the existence of maximal sets of MUBs in even prime power dimensions and a special class of affine block designs to construct an infinite family of maximal OGFFs, emphasizing that their construc- tion is highly rigid due to maximality properties of the MUBs and block designs they used, and highlighted the examples' cubature properties. In this section, we extend their work, noting that numerous other families of tight OGFFs arise from symmetric block designs via the final theorem of the preceding section; similarly, we observe examples arising from other types of affine block designs – all unmentioned in [12]. We begin with symmetric designs, which are surprising well-suited to this construction principle in a certain sense. 5.1. Symmetric block designs. 5.1. Definition. A t-(m, l, λ)-block design is symmetric if m = b or, equivalently, if l = r. As verified in [19], symmetric block designs have the useful property that the pairwise block intersections is constant. 5.2. Theorem. [[19]]For a symmetric t-(m, l, λ) block design,(cid:0)(cid:2)[m](cid:3), S(cid:1), every Jj,Jj ′ ∈ S with j 6= j ′ satisfies 5.1.1. Simple ETFFs from all t-block designs. An immediate – albeit trivial – corollary to this obser- vation is that every symmetric block designs yields an ETFF. 5.3. Corollary. Given any symmetric t-(m, l, λ) block design and a tight (m, 1)-fusion frame, Π = {πi}m i=1, for Fm, then the family of n = b block projections for Π forms an ETFF for Fm. Proof. By design, Proposition 4.21 ensures that the sequence of n = b rank l block projections, forms a tight fusion frame for Fm. By Theorem 4.11, these are chordally P = (cid:10)Pi :=Pj∈Ji equiangular, since, for i 6= i ′, we have πj(cid:11)m i=1 (cid:12)(cid:12)Jj ∩ Jj ′(cid:12)(cid:12) = λ. tr (PiPi ′ ) = Xj∈Ji Xj ′∈Ji ′ tr(cid:0)πjπj ′(cid:1) = λ, where one can confirm that λ is the packing constant for this special case. (cid:3) As implicated by the numerous examples of symmetric block designs in [24] and the references therein, these simple examples of ETFFs exist in abundance. 5.1.2. Tight OGFFs from certain block dseigns. Of more significance is the multitude of examples of tight OGFFs that arise from the construction outlined in theorem 4.24, depending heavily on the existence of maximal MUBs. The surprise here is that every nontrivial symmetric t-block-design satisfies the l2/m-cohesive property. 5.4. Proposition. Assume 1 < l < m − 1. Every symmetric t-(m, l, λ) block design, ((cid:2)[m](cid:3), S), is l2/m-cohesive. Proof. By Theorem 5.2, we only need to show that λ ≤ l2/m, which follows by applying the trivial necessary conditions from Proposition 4.19 to obtain the elementary estimation, λ = r(l − 1) m − 1 = l(l − 1) m − 1 ≤ l2 m . (cid:3) 14 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN Thus, following the strategy, wherein we employ maximal MUBs, as outlined after the statement of Theorem 4.24, we seek symmetric block designs for which m = b = q, where q is a prime power or a power of four, for the complex and real cases, respectively. 5.5. Corollary. Assume 1 < l < m − 1. • Every symmetric t-(m, l, λ) block design for which m is a prime power yields a tight (n, l)- OGFF for Cm, where n = m(m + 1), according to the construction described in Theorem 4.24. • Every symmetric t-(m, l, λ) block design for which m is a power of four yields a tight (n, l)- OGFF for Rm, where n = m(m/2 + 1), according to the construction described in Theo- rem 4.24. Examining the list of known families of symmetric designs in II.6.8 of [24], we identify numerous families of symmetric block designs for which Corollary 5.5 yields more tight OGFFs. 5.1.3. Point-hyperplane symmetric designs; see Family 1 of [24]. Given a prime power q and t ≥ 2, q−1 (cid:17) block design exists, known as a point-hyperplane design. a symmetric 1-(cid:16) qt+1−1 q−1(cid:17)-OGFF for Cm, comprised of n := (qt+1−1)(qt+1+q−2) • By Corollary 5.5, a tight (cid:16)m, qt−1 is a prime power, leading to numerous families of q−1 , qt−1−1 (q−1)2 q−1 , qt−1 projections, exists whenever m := qt+1−1 q−1 examples of tight OGFFs. • Unfortunately, the necessary conditions of these block designs seem too restrictive to admit real examples of OGFFs according to our approach. Contingent upon the open question regarding the existence of an infinitude of Fermat or Mersenne primes – a famous open problem in number theory [48] – complex examples arising from this family may be infinite by taking q = 2. 5.1.4. Hadamard symmetric designs; see Family 2 of [24]. Given t ≥ 1, a symmetric 1-(4t − 1, 2t − 1, t − 1) block design exists, known as a Hadamard design. • By Corollary 5.5, a tight (m, 2t − 1)-OGFF for Cm, comprised of n := 4t(4t − 1) projections, exists whenever m := 4t − 1 is a prime power, leading to an infinite family of complex examples of tight OGFFs. • Unfortunately, m := 4t − 1 is never a power of four, so these designs yield no real examples. 5.1.5. Menon symmetric designs; see Family 6 of [24]. The existence of a Hadamard matrix (see [24] for details) of order 4t2, t ∈ N, is equivalent to the existence of a symmetric t-(cid:0)4t2, 2t2 − t, t2 − t(cid:1) block design, called a Menon design. According to the well-known Hadamard conjecture [24], such block designs exist for all values of t. Besides the conjecture, a simple tensor construction [24], among other constructions, assures their existence when t = 2s for some s ∈ N. of n := 16t4 + 4t2 projections exists, leading to an infinite family of complex examples. • By Corollary 5.5, when t = 2s for some s ∈ N, a tight(cid:0)4t2, 2t2 − t(cid:1)-OGFF for Cm comprised • Conveniently, these designs are well-suited for the real case. Similarly, when t = 2s for some s ∈ N, where s is even, then a tight (cid:0)4t2, 2t2 − t(cid:1)-OGFF for Rm comprised of n := 5.1.6. Wallis symmetric designs; see Family 7 of [24]. Given a prime power q and t ∈ N, a symmetric t-(m, l, λ) block design exists, known as a Wallis design, where 4t2(2t2 + 1) projections exists, leading to an infinite family of tight OGFFs. • m := qt+1(cid:0)qt + qt−1 + ··· + q2 + q + 2(cid:1) , • l := qt(qt + qt−1 + ··· + q2 + q + 1), and • λ = qt(qt−1 + qt−2 + ··· + q2 + q + 1). OPTIMALLY SPREAD SUBSPACE PACKINGS 15 Taking q = 2 and computing (inductively or directly) that 2t+1(2t + 2t−1 + .... + 4 + 2 + 2) = 4t+1, the Wallis family yields infinite families of both real and complex tight OGFFs. 42t+2 + 4t+1 projections exists, leading to an infinite family of complex examples. • By Corollary 5.5, when q = 2 and t ∈ N, a tight (n, l)-OGFF for Cm comprised of n := • Once again, these designs are well-suited for the real case. When q = 2 and t ∈ N, a tight 2 + 1) projections exists, leading to another (n, l)-OGFF for Rm comprised of n := 4t+1( 4t+1 infinite family of real examples. 5.1.7. Wilson/Brouwer symmetric designs; see Family 1 of 11 [24]. Given an odd prime power q and t ∈ N, a symmetric 1-(m, l, λ) block design exists, which we refer to as a Wilson/Brouwer design, where • m := 2(cid:0)qt + qt−1 + ··· + q2 + q(cid:1) + 1, • l := qt, and • λ = qt−1 q−1 2 . Unfortunately, the addition of 1 to an even number makes it impossible for m to be a power of four here. Still, for various parameters – eg, q = 3, t = 1 yields m = 7 – this family produces various complex instances of tight OGFFs. A comprehensive list of all tight OGFFs obtained from symmetric designs is beyond the scope of this work – particularly due to outstanding open questions about their existence for various parameters [24]. Having demonstrated families arising in this fashion, we move to tight OGFFs arising from affine designs. 5.2. Affine Designs. In the work where this recipe is presented, the authors combined real and complex even powers of maximal MUBs of Hadamard 3-designs (see[12]), a special case of affine designs, to form maximal OGFFs according to Theorem 4.24. 5.6. Definition. A block design (cid:0)(cid:2)[m](cid:3), S(cid:1) is resolvable if S partitions into subsets, called parallel classes, such that • the blocks with in each class are disjoint, and • for each parallel class, every element of(cid:2)[m](cid:3) is contained in a block. Moreover, if the number of elements occurring in the intersection between blocks from different parallel classes is constant, then it is an affine design. A simple result in [24] assures the existence of analogous block designs for the odd prime power case. We restate these for our special case. 5.7. Proposition. [See II.7.10 [24]] If m = pt+1, l = pt and λ = pt−1 resolvable 1-(m, l, λ) block design exists and its remaining parameters are therefore r = Pt b =Pt+1 According to Bose's condition, such designs are l2/m-cohesive affine block designs. i=1 pi. p−1 for some prime, p, then a i=0 pi and 5.8. Theorem (Bose's condition; see Theorem II.7.28 of [24]). Given any resolvable 1-(m, l, λ) block design, the number of blocks is bounded by the other parameters according to and this lower bound is achieved if and only if the design is an l2/m-cohesive affine design. b ≥ m + r − 1, 16 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN One may verify that the designs produced by Proposition 5.7 achieve this lower bound. Noting that such designs satisfy kCmb = (pt+1 + 1) t+1Xi=1 pi > dCm + 1 = p2t+2, the conditions of Theorem 4.24 are satisfied. 5.9. Corollary. For every prime p and t ∈ N, a tight (n, l)-OGFF, comprised of n = (pt+1 + 1)Pt+1 exists for Cm, where m = pt+1 and l = pt. i=1 pi Although these tight OGFFs are not maximal OGFFs, they are maximal with respect to the con- struction in Theorem 4.24, as each instance is produced by a maximal set of MUBs and a maximal affine design, according to Bose's condition. REFERENCES [1] D. M. Appleby. SIC-POVMS and MUBS: geometrical relationships in prime dimension. In Foundations of probability and physics-5, volume 1101 of AIP Conf. Proc., pages 223–232. Amer. Inst. Phys., New York, 2009. [2] C. Bachoc, E. Bannai, and R. Coulangeon. Codes and designs in Grassmannian spaces. Discrete Math., 277(1-3):15– 28, 2004. [3] C. Bachoc and M. Ehler. Tight p-fusion frames. Appl. Comput. Harmon. Anal., 35(1):1–15, 2013. [4] C. Bachoc and M. Ehler. Signal reconstruction from the magnitude of subspace components. IEEE Trans. Inform. Theory, 61(7):4015–4027, 2015. [5] Waheed U. Bajwa, Robert Calderbank, and Dustin G. Mixon. Two are better than one: fundamental parameters of frame coherence. Appl. Comput. Harmon. Anal., 33(1):58–78, 2012. [6] S. Bandyopadhyay, P. O. Boykin, V. Roychowdhury, and F. Vatan. A new proof for the existence of mutually unbiased bases. Algorithmica, 34(4):512–528, 2002. [7] A. Barg and O. R. Musin. Codes in spherical caps. ArXiv Mathematics e-prints, June 2006. [8] A. Barg and D. Yu. Nogin. Bounds on packings of spheres in the Grassmann manifold. IEEE Trans. Inform. Theory, 48(9):2450–2454, 2002. [9] A. Barg and W.-H. Yu. New bounds for equiangular lines. In Discrete geometry and algebraic combinatorics, volume 625 of Contemp. Math., pages 111–121. Amer. Math. Soc., Providence, RI, 2014. [10] J. J. Benedetto and J. D. Kolesar. Geometric properties of Grassmannian frames for R2 and R3. EURASIP J. Appl. Signal Process., 2006:1–17, 2006. [11] John J. Benedetto and Andrew Kebo. The role of frame force in quantum detection. J. Fourier Anal. Appl., 14(3):443–474, 2008. [12] B. G. Bodmann and J. and Block Designs. Proc. Amer. Math. Soc., http://dx.doi.org/10.1090/proc/13956 (to appear in print)., 2017. I. Haas. Maximal Orthoplectic Fusion Frames from Mutually Unbiased Bases 2014, DOI: electronically published on February 26, [13] Bernhard G. Bodmann and Helen J. Elwood. Complex equiangular Parseval frames and Seidel matrices containing pth roots of unity. Proc. Amer. Math. Soc., 138(12):4387–4404, 2010. [14] Bernhard G. Bodmann and John Haas. Frame potentials and the geometry of frames. J. Fourier Anal. Appl., 21(6):1344–1383, 2015. [15] Bernhard G. Bodmann and John Haas. Achieving the orthoplex bound and constructing weighted complex projec- tive 2-designs with Singer sets. Linear Algebra Appl., 511:54–71, 2016. [16] B.G. Bodmann, P.G. Casazza, D. Edidin, and R. Balan. Frames for linear reconstruction without phase. In Informa- tion Sciences and Systems, 2008. CISS 2008. 42nd Annual Conference on, pages 721–726, March 2008. [17] Len Bos and Shayne Waldron. Some remarks on Heisenberg frames and sets of equiangular lines. New Zealand J. Math., 36:113–137, 2007. [18] P. O. Boykin, M. Sitharam, M. Tarifi, and P. Wocjan. Real mutually unbiased bases. arXiv:quant-ph/0502024v2. [19] P. J. Cameron and J. H. van Lint. Designs, Graphs, Codes and their Links. London Mathematical Society Student Texts. Cambridge University Press, 1991. [20] P. J. Cameron and J. J. Seidel. Quadratic forms over GF(2). Nederl. Akad. Wetensch. Proc. Ser. A 76=Indag. Math., 35:1–8, 1973. [21] P. G. Casazza and J. I. Haas. On the rigidity of geometric and spectral properties of Grassmannian frames. ArXiv e-prints, May 2016. [22] Peter G. Casazza and Gitta Kutyniok, editors. Finite frames. Applied and Numerical Harmonic Analysis. Birkhauser/Springer, New York, 2013. OPTIMALLY SPREAD SUBSPACE PACKINGS 17 [23] Amina Chebira and Jelena Kovacevic. Frames in bioimaging. In Information Sciences and Systems, 2008. CISS 2008. 42nd Annual Conference on, pages 727–732. IEEE, 2008. [24] C. J. Colbourn and J. H. Dinitz, editors. Handbook of combinatorial designs. Discrete Mathematics and its Applica- tions (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, second edition, 2007. [25] J. H. Conway, R. H. Hardin, and N. J. A. Sloane. Packing lines, planes, etc.: packings in Grassmannian spaces. Experiment. Math., 5(2):139–159, 1996. [26] Parker R. A. Sloane N. J. A. Conway, J. H. The covering radius of the leech lattice. Proc. Roy. Soc. London Ser. A, 380(1779):261–290, 1982. [27] P. Delsarte, J. M. Goethals, and J. J. Seidel. Bounds for systems of lines, and Jacobi polynomials. Philips Research Reports, 30:91, 1975. [28] Helen J. Elwood. Constructing complex equiangular Parsevel frames. ProQuest LLC, Ann Arbor, MI, 2011. Dissertation (Ph.D.)–University of Houston, Houston, TX. [29] T. Ericson and V. Zinoviev. Codes on Euclidean Spheres. North-Holland Mathematical Library. Elsevier Science, 2001. [30] B. Et-Taoui. Complex conference matrices, Complex Hadamard matrices and equiangular tight frames. ArXiv e- prints, September 2014. [31] Boumediene Et-Taoui. Complex conference matrices, complex hadamard matrices and equiangular tight frames. arXiv e-print, arXiv:1409.5720, 09 2014. [32] M. Fickus, J. Jasper, and D. G. Mixon. Packings in real projective spaces. ArXiv e-prints, July 2017. [33] M. Fickus, J. Jasper, D. G. Mixon, and J. Peterson. Tremain equiangular tight frames. ArXiv e-prints, February 2016. [34] Matthew Fickus and Dustin G. Mixon. Tables of the existence of equiangular tight frames. arXiv e-print, arXiv:1504.00253, 04 2015. [35] Matthew Fickus, Dustin G. Mixon, and Janet C. Tremain. Steiner equiangular tight frames. Linear Algebra Appl., 436(5):1014–1027, 2012. [36] C. Godsil and A. Roy. Equiangular lines, mutually unbiased bases, and spin models. European J. Combin., 30(1):246–262, 2009. [37] Vivek K. Goyal, Jelena Kovacevi´c, and Jonathan A. Kelner. Quantized frame expansions with erasures. Appl. Comput. Harmon. Anal., 10(3):203–233, 2001. [38] J. I. Haas, J. Cahill, J. Tremain, and P. G. Casazza. Constructions of biangular tight frames and their relationships with equiangular tight frames. ArXiv e-prints, March 2017. [39] John Haas, Nathaniel Hammen, and Dustin Mixon. The levenstein bound for packings in projective spaces. SPIE Conference Proceedings, 2017. [40] T. R. Hoffman and J. P. Solazzo. Complex equiangular tight frames and erasures. Linear Algebra Appl., 437(2):549– 558, 2012. [41] S. G. Hoggar. t-designs in projective spaces. European J. Combin., 3(3):233–254, 1982. [42] Philippe Jaming, M´at´e Matolcsi, and P´eter M´ora. The problem of mutually unbiased bases in dimension 6. Cryp- tography and Communications, 2(2):211–220, Sep 2010. [43] John Jasper, Dustin G. Mixon, and Matthew Fickus. Kirkman equiangular tight frames and codes. IEEE Trans. Inform. Theory, 60(1):170–181, 2014. [44] IV John Haas and Peter Casazza. On the structures of grassmannian frames. In 2017 International Conference on Sampling Theory and Applications (SampTA) (SampTA2017), Tallinn, Estonia, July 2017. [45] D. Kalra. Complex equiangular cyclic frames and erasures. Linear Algebra Appl., 419(2-3):373–399, 2006. [46] Mahdad Khatirinejad. On Weyl-Heisenberg orbits of equiangular lines. J. Algebraic Combin., 28(3):333–349, 2008. [47] A. Klappenecker and M. Rotteler. Mutually unbiased bases are complex projective 2-designs. In International Sym- posium on Information Theory, 2005. ISIT 2005., pages 1740–1744, Sept 2005. [48] H. V. Krishna. On Mersenne and Fermat numbers. Math. Student, 39:51–52 (1972), 1971. Volume dedicated to the memory of V. Ramaswami Aiyar. [49] V. I. Levenshtein. Designs as maximum codes in polynomial metric spaces. Acta Applicandae Mathematica, 29(1):1– 82, Nov 1992. [50] P. G. Massey, M. A. Ruiz, and D. Stojanoff. The structure of minimizers of the frame potential on fusion frames. J. Fourier Anal. Appl., 16(4):514–543, 2010. [51] D.G. Mixon, C. Quinn, N. Kiyavash, and M. Fickus. Equiangular tight frame fingerprinting codes. In Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on, pages 1856–1859, May 2011. [52] A. Neumaier. Graph representations, two-distance sets, and equiangular lines. Linear Algebra Appl., 114/115:141– 156, 1989. [53] Onur Oktay. Frame quantization theory and equiangular tight frames. ProQuest LLC, Ann Arbor, MI, 2007. Disserta- tion (Ph.D.)–University of Maryland, College Park, MD. [54] Joey Iverson Organizers: Dustin Mixon, John Jasper. Special ams session on recent advances in the packing prob- lem. Oh State, March 2018. [55] R. A. Rankin. The closest packing of spherical caps in n dimensions. Proc. Glasgow Math. Assoc., 2:139–144, 1955. 18 PETER G. CASAZZA, JOHN I. HAAS IV, JOSHUA STUECK, AND TIN T. TRAN [56] Asha Rao, Diane Donovan, and Joanne L. Hall. Mutually orthogonal latin squares and mutually unbiased bases in dimensions of odd prime power. Cryptography and Communications, 2(2):221–231, Sep 2010. [57] A. Roy and A. J. Scott. Weighted complex projective 2-designs from bases: optimal state determination by orthog- onal measurements. J. Math. Phys., 48(7):072110, 1–24, 2007. [58] A. J. Scott. SICs: Extending the list of solutions. ArXiv e-prints, March 2017. [59] P. W. Shor and N. J. A. Sloane. A family of optimal packings in Grassmannian manifolds. J. Algebraic Combin., 7(2):157–163, 1998. [60] Neil Sloane. Neil j. a. sloane: Home page. Online: http://neilsloane.com/. [61] Tobias Springer, Katja Ickstadt, and Joachim Stockler. Frame potential minimization for clustering short time series. Adv. Data Anal. Classif., 5(4):341–355, 2011. [62] T. Strohmer and R. W. Heath, Jr. Grassmannian frames with applications to coding and communication. Appl. Comput. Harmon. Anal., 14(3):257–275, 2003. [63] M. A. Sustik, J. A. Tropp, I. S. Dhillon, and R. W. Heath, Jr. On the existence of equiangular tight frames. Linear Algebra Appl., 426(2-3):619–635, 2007. [64] F. Szollosi. Complex Hadamard matrices and equiangular tight frames. Linear Algebra Appl., 438(4):1962–1967, 2013. [65] F. Szollosi. All complex equiangular tight frames in dimension 3. arXiv preprint arXiv:1402.6429, 2014. [66] L. R. Welch. Lower bounds on the maximum cross correlation of signals. IEEE Trans. on Information Theory, 20(3):397–9, May 1974. [67] W. K. Wootters and B. D. Fields. Optimal state-determination by mutually unbiased measurements. Ann. Physics, 191(2):363–381, 1989. [68] G. Zauner. Quantendesigns - Grundzuge einer nichtkommutativen Designtheorie. University Wien (Austria), 1999. Dissertation (Ph.D.), English translation in International Journal of Quantum Information (IJQI) 9 (1), 445–507, 2011. 219 MATHEMATICAL SCIENCES BUILDING, DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MISSOURI, COLUMBIA, MO 65211
1207.2662	1	1207	2012-07-11T14:56:08	A simple proof that the power $\frac{2m}{m+1}$ in the Bohnenblust--Hille inequalities is sharp	[ "math.FA" ]	The power $\frac{2m}{m+1}$ in the polynomial (and multilinear) Bohnenblust--Hille inequality is optimal. This result is well-known but its proof highly nontrivial. In this note we present a quite simple proof of this fact.	math.FA	math	A SIMPLE PROOF THAT THE POWER 2m m+1 IN THE BOHNENBLUST -- HILLE INEQUALITIES IS SHARP DANIEL NU NEZ-ALARC ´ON AND DANIEL PELLEGRINO Abstract. The power m+1 in the polynomial (and multilinear) Bohnenblust -- Hille inequality is optimal. This result is well-known but its proof highly nontrivial. In this note we present a quite simple proof of this fact. 2m 1. Introduction The polynomial and multilinear Bohnenblust -- Hille inequalities were proved by H.F. Bohnenblust and E. Hille in 1931 and play a crucial role in different fields as Fourier and Harmonic Analysis and Quantum Information Theory (see [4, 5, 7]). m+1 The polynomial Bohnenblust -- Hille inequality proves the existence of a positive function C : N → [1,∞) such that for every m-homogeneous polynomial P on CN , the ℓ 2m -norm of the set of coefficients of P is bounded above by Cm times the supre- mum norm of P on the unit polydisc. This result has important striking applications in different contexts (see [4]). The multilinear version of the Bohnenblust -- Hille in- equality asserts that for every positive integer m ≥ 1 there exists a sequence of positive scalars (Cm)∞ m=1 in [1,∞) such that N Xi1,...,im=1(cid:12)(cid:12)T (ei1 , . . . , eim)(cid:12)(cid:12) m+1 2m 2m m+1! ≤ Cm sup z1,...,zm∈DN T (z1, . . . , zm) for all m-linear forms T : CN × · · · × CN → C and every positive integer N, where (ei)N i=1 denotes the canonical basis of CN and DN represents the open unit polydisk in CN . The original proof ([3]) that the power 2m m+1 is optimal is quite puzzling. Ac- cording to Defant et al ([4, page 486]), Bohnenblust and Hille "showed, through 2m a highly nontrivial argument, that the exponent m+1 cannot be improved" or ac- cording to Defant and Schwarting [6, page 90], Bohnenblust and Hille showed "with a sophisticated argument that the exponent In [2] there is an alternative proof for the case of multilinear mappings, but the arguments are also nontrivial, involving p-Sidon sets and sub-Gaussian systems. The main goal of this note is to present a quite elementary proof (which solves simultaneously the cases of polynomials and multilinear mappings) of the optimality of 2m 2m m+1 is optimal". m+1. Key words and phrases. Bohnenblust -- Hille inequalities. 1 2 DANIEL NU NEZ-ALARC ´ON AND DANIEL PELLEGRINO 2. The new proof of the sharpness of 2m m+1 We will show that the optimality of the power 2m m+1 is a straightforward conse- quence of the following famous result known as Kahane-Salem-Zygmund inequal- ity (see [8, Theorem 4, Chapter 6] or [1, page 21]): Theorem 2.1 (Kahane-Salem-Zygmund inequality). Let m, n be positive integers. Then there are signs εα = ±1 so that the m-homogeneous polynomial given by satisfies Pm,n : ℓn ∞ → C Pm,n =Pα=mεαzα kPm,nk ≤ Cn(m+1)/2plog m where C is an universal constant (it does not depend on n or m). Theorem 2.2. The power 2m Proof. Let m ≥ 2 be a fixed positive integer. For each n, let Pm,n : ℓn ∞ → C be the m-homogeneous polynomial satisfying the Kahane-Salem-Zygmund inequality. For our goals it suffices to deal with the case n > m. m+1 in the Bohnenblust -- Hille inequalities is sharp. Let q < 2m m+1 . Then a simple combinatorial calculation shows that (cid:16)Pα=m εαq(cid:17)1/q =(cid:18)p(n) + 1 m! 1 q , (n − k)(cid:19) m−1 Qk=0 where p (n) > 0 is a polynomial of degree m−1. If the polynomial Bohnenblust -- Hille inequality was true with the power q, then there would exist a constant Cm,q > 0 so that Cm,qC ≥ 1 n(m+1)/2√log m(cid:18)p(n) + 1 m! for all n. If we raise both sides to the power of q and make n → ∞ we obtain m−1 (n − k)(cid:19)1/q Qk=0 nq(m+1)/2(cid:0)√log m(cid:1)q! , p(n) r(n) m!nq(m+1)/2(cid:0)√log m(cid:1)q + r(n) = m−1 Qk=0 (n − k). q(m + 1) 2 deg r = m > n→∞ (Cm,qC)q ≥ lim with Since we have n→∞ lim r(n) m!nq(m+1)/2(cid:0)√log m(cid:1)q + p(n) nq(m+1)/2(cid:0)√log m(cid:1)q! = ∞, a contradiction. Since the multilinear Bohnenblust -- Hille inequality (with a power q) implies the polynomial Bohnenblust -- Hille inequality with the same power, we conclude that 2m (cid:3) m+1 is also sharp in the multilinear case. A NEW PROOF THAT THE POWER 2m m+1 IS SHARP 3 References [1] F. Bayart, Maximum modulus of random polynomials, Quart. J. Math 63 (2012), 21 -- 39. [2] R. Blei, Analysis in integer and fractional dimensions, Cambridge Studies in Advances Mathe- matics, 2001. [3] H.F. Bohnenblust and E. Hille, On the absolute convergence of Dirichlet series, Ann. of Math. 32 (1931), 600-622. [4] A. Defant, L. Frerick, J. Ortega-Cerd´a, M. Ounaıes and K. Seip, The polynomial Bohnenblust -- Hille inequality is hypercontractive, Ann. of Math. (2) 174 (2011), 485 -- 497. [5] A. Defant. D. Popa and U. Schwarting, Coordinatewise multiple summing operators in Banach spaces, J. Funct. Anal. 259 (2010), 220 -- 242. [6] A. Defant, U. Schwarting, Bohr's radii and strips -- a microscopic and a macroscopic view, Note Mat. 31 (2011), 87 -- 101. [7] D. Diniz, G.A. Munoz-Fern´andez, D. Pellegrino and J.B. Seoane-Sep´ulveda, Lower bounds for the constants in the Bohnenblust -- Hille inequality: the case of real scalars, Proc. Amer. Math. Soc., in press. [8] J.-P. Kahane, Some Random Series of Functions, Cambridge Studies in Advanced Mathematics 5, Cambridge University Press, Cambridge, 1993. Departamento de Matem´atica, Universidade Federal da Para´ıba, 58.051-900 - Joao Pessoa, Brazil. E-mail address: [email protected] and [email protected]
1306.4761	2	1306	2015-06-30T14:27:08	On The Fu\v{c}ik Spectrum Of Non-Local Elliptic Operators	[ "math.FA", "math.AP" ]	In this article, we study the Fu\v{c}ik spectrum of fractional Laplace operator which is defined as the set of all $(\al,\ba)\in \mb R^2$ such that \begin{equation} \quad \left. \begin{array}{lr} \quad (-\De)^s u = \al u^{+} - \ba u^{-} \; \text{in}\; \Om \quad \quad \quad \quad u = 0 \; \mbox{in}\; \mb R^n \setminus\Om.\\ \end{array} \quad \right\} \end{equation} has a non-trivial solution $u$, where $\Om$ is a bounded domain in $\mb R^n$ with Lipschitz boundary, $n>2s$, $s\in(0,1)$. The existence of a first nontrivial curve $\mc C$ of this spectrum, some properties of this curve $\mc C$, e.g. Lipschitz continuous, strictly decreasing and asymptotic behavior are studied in this article. A variational characterization of second eigenvalue of the fractional eigenvalue problem is also obtained. At the end, we study a nonresonance problem with respect to Fu\v{c}ik spectrum.	math.FA	math	On The Fucik Spectrum Of Non-Local Elliptic Operators Sarika Goyal∗and K. Sreenadh† Department of Mathematics, Indian Institute of Technology Delhi Hauz Khaz, New Delhi-16, India In this article, we study the Fucik spectrum of fractional Laplace operator which is defined as the set of all (α, β) ∈ R2 such that Abstract (−∆)su = αu+ − βu− in Ω u = 0 in Rn \ Ω. ) has a non-trivial solution u, where Ω is a bounded domain in Rn with Lipschitz boundary, n > 2s, s ∈ (0, 1). The existence of a first nontrivial curve C of this spectrum, some properties of this curve C, e.g. Lipschitz continuous, strictly decreasing and asymptotic behavior are studied in this article. A variational characterization of second eigenvalue of the fractional eigenvalue problem is also obtained. At the end, we study a nonresonance problem with respect to Fucik spectrum. Key words: Non-local operator, fractional Laplacian, Fucik spectrum, Nonresonance. 2010 Mathematics Subject Classification: 35R11, 35R09, 35A15. ∗email: [email protected] †e-mail: [email protected] 1 Fucik Spectrum for non-local elliptic operators 2 1 Introduction The Fucik spectrum of fractional Laplace operator is defined as the set of all (α, β) ∈ R2 such that ( (−∆)su = αu+ − βu− in Ω u = 0 on Rn \ Ω. has a non-trivial solution u, where s ∈ (0, 1) and (−∆)s be fractional Laplacian operator defined as (−∆)su(x) = − 1 2ZRn u(x + y) + u(x − y) − 2u(x) yn+2s dy for all x ∈ Rn. In general, we study the Fucik spectrum of an equation driven by the non-local operator LK which is defined as LKu(x) = 1 2ZRn (u(x + y) + u(x − y) − 2u(x))K(y)dy for all x ∈ Rn, where K : Rn \ {0} → (0, ∞) satisfies the following: (i) mK ∈ L1(Rn), where m(x) = min{x2, 1}, (ii) There exist λ > 0 and s ∈ (0, 1) such that K(x) ≥ λx−(n+2s), (iii) K(x) = K(−x) for any x ∈ Rn \ {0}. In case K(x) = x−(n+2s), LK is the fractional Laplace operator −(−∆)s. The Fucik spectrum of the non-local operator LK is defined as the set PK of (α, β) ∈ R2 such that −LKu = αu+ − βu− in Ω u = 0 in Rn \ Ω. ) (1.1) (1.1) has a nontrivial solution u. Here u± = max(±u, 0) and Ω ⊂ Rn is a bounded domain with Lipshitz boundary. For α = β, Fucik spectrum of (1.1) becomes the usual spectrum of −LKu = λu in Ω u = 0 in Rn \ Ω. ) (1.2) Let 0 < λ1 < λ2 ≤ ... ≤ λk ≤ ... denote the distinct eigenvalues of (1.2). It is proved in [18] that the first eigenvalue λ1 of (1.2) is simple, isolated and can be characterized as follows λ1 = inf u∈X0(cid:26)ZQ (u(x) − u(y))2K(x − y)dxdy :ZΩ u2 = 1(cid:27) . The author also proved that the eigenfunction corresponding to λ1 are of constant sign. We is symmetric with respect to diagonal. In this paper we will prove that the two lines R × λ1 observe thatPK clearly contains (λk, λk) for each k ∈ N and two lines λ1 ×R and R×λ1. PK and λ1 × R are isolated in PK and give a variational characterization of second eigenvalue λ2 of −LK. Fucik Spectrum for non-local elliptic operators 3 When s = 1, the fractional Laplacian operator become the usual Laplace operator. Fucik spectrum is introduced by Fucik in 1976. The negative Laplacian in one dimension with peri- odic boundary condition is studied in [9]. Also study of Fucik spectrum in case of Laplacian, p-Laplacian equation with Dirichlet, Neumann and robin boundary condition has been stud- ied by many authors [1, 2, 3, 4, 6, 7, 13, 12, 15, 16]. A nonresonance problem with respect to Fucik spectrum is also discussed in many papers [3, 14, 17]. To best of our knowledge, no work has been done to find the Fucik spectrum for non-local operator. Recently a lot of attention is given to the study of fractional and non-local operator of elliptic type due to concrete real world applications in finance, thin obstacle problem, optimization, quasi-geostrophic flow etc [17, 19, 20, 21]. Here we use the similar approach to find Fucik spectrum that is used in [3]. In [17], Servadei and Valdinoci discussed Dirichlet boundary value problem in case of fractional Laplacian using the Variational techniques. We also used the similar variational technique to by Servadei. We introduce this space as follows: findPK. Due to non-localness of the fractional Laplacian, the space (X0, k.kX0 ) is introduced X =nu u : Rn → R is measurable, uΩ ∈ L2(Ω) and (u(x) − u(y))pK(x − y) ∈ L2(Q)o , where Q = R2n \ (CΩ × CΩ) and CΩ := Rn \ Ω. The space X is endowed with the norm defined as kukX = kukL2(Ω) +(cid:18)ZQ u(x) − u(y)2K(x − y)dxdy(cid:19) 1 2 . X0 = {u ∈ X : u = 0 a.e. in Rn \ Ω} 1 2 Then we define with the norm kukX0 =(cid:18)ZQ u(x) − u(y)2K(x − y)dxdy(cid:19) is an Hilbert space. Note that the norm k.kX0 involves the interaction between Ω and Rn \ Ω. Remark 1.1 (i) C 2 usual fractional Sobolev space endowed with the norm c (Ω) ⊆ X0, X ⊆ H s(Ω) and X0 ⊆ H s(Rn), where H s(Ω) denotes the kukH s(Ω) = kukL2 +(cid:18)ZΩ×Ω (u(x) − u(y))2 x − yn+2s dxdy(cid:19) 1 2 . (ii) The embedding X0 ֒→ L2∗(Rn) = L2∗(Ω) is continuous, where 2∗ = 2n detailed of these embeddings, one can refer [8, 17]. n−2s . To see the Definition 1.2 A function u ∈ X0 is a weak solution of (1.1), if for every v ∈ X0, u satisfies ZQ (u(x) − u(y))(v(x) − v(y))K(x − y)dxdy = αZΩ u+vdx − βZΩ u−vdx, Fucik Spectrum for non-local elliptic operators 4 Weak solutions of (1.1) are exactly the critical points of the functional J : X0 → R defined as J(u) = 1 2ZQ (u(x) − u(y))2K(x − y)dxdy − (u+)2dx − α 2 ZΩ (u−)2dx. β 2 ZΩ J is Fr´echet differentiable in X0 and hJ ′(u), φi =ZQ (u(x) − u(y))(φ(x) − φ(y))K(x − y)dxdy − αZΩ u+φdx − βZΩ u−φdx. described as (p + c(p), c(p)). In section 3 we prove that the lines λ1 × R and R × λ1 are The paper is organized as follows: In section 2 we construct a first nontrivial curve in PK , isolated in PK, the curve that we obtained in section 2 is the first nontrivial curve and give the variational characterization of second eigenvalue of −LK. In section 4 we prove some properties of the first curve. A non-resonance problem with respect to Fucik spectrum is also studied in section 5. We shall throughout use the function space X0 with the norm k.kX0 and we use the stan- dard Lp(Ω) space whose norms are denoted by kukLp. Also φ1 is the eigenfunction of −LK corresponding to λ1. 2 Fucik Spectrum PK for −LK In this section we study the existence of first nontrivial curve in the Fucik spectrum PK of −LK. We find that the points in PK are associated with the critical value of some restricted functional. For this we fix p ∈ R and for p ≥ 0, consider the functional Jp : X0 → R by Jp(u) =ZQ (u(x) − u(y))2K(x − y)dxdy − pZΩ (u+)2dx. (2.1) Then Jp ∈ C 1(X0, R) and for any φ ∈ X0 hJ ′ p(u), φi = 2ZQ (u(x) − u(y))(φ(x) − φ(y))K(x − y)dxdy − 2pZΩ u+(x)φ(x)dx. Also Jp := JpP is C 1(X0, R), where P is defined as P =(cid:26)u ∈ X0 : I(u) :=ZΩ u2dx = 1(cid:27) . We first note that u ∈ P is a critical point of Jp if and only if there exists t ∈ R such that ZQ (u(x) − u(y))(v(x) − v(y))K(x − y)dxdy − pZΩ u+vdx = tZΩ uvdx, (2.2) for all v ∈ X0. Hence u ∈ P is a nontrivial weak solution of the problem −LKu = (p + t)u+ − tu− in Ω; u = 0 on Rn \ Ω, Fucik Spectrum for non-local elliptic operators 5 the following result, which describe the relationship between the critical points of Jp and the which exactly means (p + t, t) ∈PK . Putting v = u in (2.2), we get t = Jp(u). Thus we have spectrum PK . Lemma 2.1 For p ≥ 0, (p + t, t) ∈ R2 belongs to the spectrum PK if and only if there exists a critical point u ∈ P of Jp such that t = Jp(u), a critical value. Now we look for the minimizers of Jp. Proposition 2.2 The first eigenfunction φ1 is a global minimum for Jp with Jp(φ1) = λ1 −p. The corresponding point in PK is (λ1, λ1 − p) which lies on the vertical line through (λ1, λ1). Proof. It is easy to see that Jp(φ1) = λ1 − p and Jp(u) =ZQ ≥λ1ZΩ (u(x) − u(y))2K(x − y)dxdy − pZΩ u2dx − pZΩ (u+)2dx ≥ λ1 − p. (u+)2dx Thus φ1 is a global minimum of Jp with Jp(φ1) = λ1 − p. (cid:3) Now we have a second critical point of Jp at −φ1 corresponding to strict local minimum. Proposition 2.3 The negative eigenfunction −φ1 is a strict local minimum for Jp with Jp(−φ1) = λ1. The corresponding point in PK is (λ1 + p, λ1), which lies on the horizontal line through (λ1, λ1). Proof. Let us suppose by contradiction that there exists a sequence uk ∈ P, uk 6= −φ1 with Jp(uk) ≤ λ1, uk → −φ1 in X0. Firstly, we show that uk changes sign for sufficiently large k. Since uk 6= −φ1, it must be ≤ 0 for some x ∈ X0. If uk ≤ 0 for a.e x ∈ Ω, then Jp(uk) =ZQ (uk(x) − uk(y))2K(x − y)dxdy > λ1, (2.3) since uk 6= ±φ1 and we get contradiction as Jp(uk) ≤ λ1. So uk changes sign for sufficiently large k. Define wk := and u+ k ku+ k kL2 rk :=ZQ (wk(x) − wk(y))2K(x − y)dxdy. Now we claim that rk → ∞. Let us suppose by contradiction that rk is bounded. Then there exists a subsequence of wk still denoted by wk and w ∈ X0 such that wk ⇀ w weakly in X0 and wk → w strongly in L2(Rn). Therefore RΩ w2dx = 1, w ≥ 0 a.e. and so for some ǫ > 0, δ = {x ∈ X0 : w(x) ≥ ǫ} > 0. As uk → −φ1 in X0 and hence in L2(Ω). Therefore Fucik Spectrum for non-local elliptic operators 6 {x ∈ Ω : uk(x) ≥ ǫ} → 0 as k → ∞ and so {x ∈ Ω : wk(x) ≥ ǫ} → 0 as k → ∞ which is a contradiction as δ > 0 . Hence the claim. Next, (uk(x)−uk(y))2 = ((u+ =(u+ =(u+ k (x) − u+ k (x) − u+ k (x) − u+ k (y))2 + (u− k (y))2 + (u− k (y)) − (u− k (x) − u− k (x) − u− k (x) − u− k (y))2 − 2(u+ k (y))2 + 2u+ k (y)))2 k (x) − u+ k (x)u− k (y))(u− k (y) + 2u− k (x) − u− k (x)u+ k (y), k (y)) where we have used u+ k (x)u− k (x) = 0. Using K(x) = K(−x) we have ZQ k (x)u− u+ k (y)K(x − y)dxdy =ZQ k (y)u− u+ k (x)K(x − y)dxdy. (2.4) Then from above estimates, we get (u+ k )2dx (u+ k )2dx (u+ k (x)u− u+ k (x) − u+ Jp(uk) =ZQ =ZQ + 4ZQ ≥(rk − p)ZΩ ≥(rk − p)ZΩ (uk(x) − uk(y))2K(x − y)dxdy − pZΩ k (y))2K(x − y)dxdy +ZQ k (y)K(x − y)dxdy − pZΩ k )2dx + 4ZQ k )2dx + λ1ZΩ k )2dx + λ1ZΩ Jp(uk) ≤ λ1 = λ1ZΩ k )2dx. (u− (u+ (u− (u+ (u+ k )2dx + λ1ZΩ (u− k )2dx. As uk ∈ P, we get (u− k (x) − u− k (y))2K(x − y)dxdy k (x)u− u+ k (y)K(x − y)dxdy Combining both the inequalities we have, (rk − p)ZΩ (u+ k )2dx + λ1ZΩ (u− k )2dx ≤ λ1ZΩ (u+ k )2dx + λ1ZΩ (u− k )2dx. This implies (rk − p − λ1)RΩ(u+ that rk → +∞, as required. k )2dx ≤ 0, and hence rk − p ≤ λ1, which contradicts the fact (cid:3) We will now find the third critical point based on mountain pass Theorem. A norm of derivative of the restriction Jp of Jp at u ∈ P is defined as k Jp(u)k∗ = min{k J ′ p(u) − tI ′(u)kX0 : t ∈ R}. Definition 2.4 We say that Jp satisfies the Palais-Smale (in short, (P.S)) condition on P if for any sequence uk ∈ P such that Jp(uk) is bounded and k J ′ p(uk)k∗ → 0, then there exists a subsequence that converges strongly in X0. Fucik Spectrum for non-local elliptic operators 7 Now we state here the version of mountain pass theorem, that will be used later. Proposition 2.5 Let E be a Banach space, g, f ∈ C 1(E, R), M = {u ∈ E g(u) = 1} and u0, u1 ∈ M . let ǫ > 0 such that ku1 − u0k > ǫ and inf{f (u) : u ∈ M and ku − u0kE = ǫ} > max{f (u0), f (u1)}. Assume that f satisfies the (P.S) condition on M and that Γ = {γ ∈ C([−1, 1], M ) : γ(−1) = u0 and γ(1) = u1} is non empty. Then c = inf γ∈Γ maxu∈γ[−1,1] f (u) is a critical value of f M . Lemma 2.6 Jp satisfies the (P.S) condition on P. Proof. Let {uk} be a (P.S) sequence. i.e., there exists K > 0 and tk such that Jp(uk) ≤ K, (2.5) ZQ (uk(x) − uk(y))(v(x) − v(y))K(x − y)dxdy − pZΩ u+ k v − tkZΩ ukv = ok(1)kvkX0 . (2.6) From (2.5), we get uk is bounded in X0. So we may assume that up to a subsequence uk ⇀ u0 weakly in X0, and uk → u0 strongly in L2(Ω). Putting v = uk in (2.6), we get tk is bounded and up to a subsequence tk converges to t. We now claim that uk → u0 strongly in X0. As uk ⇀ u0 weakly in X0, we have ZQ (uk(x) − uk(y))(v(x) − v(y))K(x − y)dxdy →ZQ (u0(x) − u0(y))(v(x) − v(y))K(x − y)dxdy (2.7) for all v ∈ X0. Also J ′ p(uk)(uk − u0) = ok(1). Therefore we get (cid:12)(cid:12)(cid:12)(cid:12) ZQ (uk(x) − uk(y))2 K(x − y)dxdy −ZQ ≤ O(ǫk) + pku+ (uk(x) − uk(y))(u0(x) − u0(y))K(x − y)dxdy(cid:12)(cid:12)(cid:12)(cid:12) k kL2kuk − u0kL2 + tkkukkL2 kuk − u0kL2 → 0 as k → ∞. Taking v = u0 in (2.7), we get ZQ (uk(x) − uk(y))(u0(x) − u0(y))K(x − y)dxdy →ZQ (u0(x) − u0(y))2K(x − y)dxdy. From above two equations, we have ZQ (uk(x) − uk(y))2K(x − y)dxdy →ZQ (u0(x) − u0(y))2K(x − y)dxdy. Fucik Spectrum for non-local elliptic operators Thus kukk2 X0 → ku0k2 X0. Now using this and v = u0 in (2.7), we get (uk(x) − uk(y))(u0(x) − u0(y))K(x − y)dxdy kuk − u0k2 X0 = kukk2 X0 + kukk2 X0 − 2ZQ −→ 0 as k → ∞. Hence uk → u0 strongly in X0. Lemma 2.7 Let ǫ0 > 0 be such that 8 (cid:3) Jp(u) > Jp(−φ1) for all u ∈ B(−φ1, ǫ0) ∩ P with u 6= −φ1, where the ball is taken in X0. Then for any 0 < ǫ < ǫ0, inf{ Jp(u) : u ∈ P and ku − (−φ1)kX0 = ǫ} > Jp(−φ1). (2.8) Proof. Assume by contradiction that infimum in (2.8) is equal to Jp(−φ1) = λ1 for some ǫ with 0 < ǫ < ǫ0. Then there exists a sequence uk ∈ P with kuk − (−φ1)kX0 = ǫ such that Jp(uk) ≤ λ1 + 1 2k2 . Consider the set C = {u ∈ P : ǫ − δ ≤ ku − (−φ1)kX0 ≤ ǫ + δ}, where δ is chosen such that ǫ − δ > 0 and ǫ + δ < ǫ0. From (2.8) and given hypotheses, it follows that inf{ Jp(u) : u ∈ C} = λ1. Now for each k, we apply Ekeland's variational principle to the functional Jp on C to get the existence of vk ∈ C such that Jp(vk) ≤ Jp(uk), kvk − ukkX0 ≤ 1 k , Jp(vk) ≤ Jp(u) + 1 k ku − vkkX0 ∀ u ∈ C We claim that vk is a (P.S) sequence for Jp on P i.e. Jp(vk) is bounded and k J ′ p(vk)k∗ → 0. Once this is proved we get by Lemma 2.6, up to a subsequence vk → v in X0. Clearly v ∈ P and satisfies kv − (−φ1)kX0 ≤ ǫ + δ < ǫ0 and Jp(v) = λ1 which contradicts the given hypotheses. Then proof of the claim can be proved similar as in Lemma 2.9 of [3] by replacing k.k1,p by k.kX0 . (cid:3) Proposition 2.8 Let ǫ > 0 such that kφ1 − (−φ1)kX0 > ǫ and inf{ Jp(u) : u ∈ P and ku − (−φ1)kX0 = ǫ} > max{ Jp(−φ1), Jp(φ1)}. Then Γ = {γ ∈ C([−1, 1], P) : γ(−1) = −φ1 and γ(1) = φ1} is non empty and is a critical value of Jp. Moreover c(p) > λ1. c(p) = inf γ∈Γ max u∈γ[−1,1] Jp(u) (2.9) Fucik Spectrum for non-local elliptic operators 9 Proof. Let φ ∈ X0 be such that φ 6∈ Rφ1 and consider the path γ(t) = tφ1+(1−t)φ , then γ(t) ∈ Γ . Moreover by Lemmas 2.6 and 2.7, Jp satisfies (P.S) condition and the geometric assumptions. Then by Proposition 2.5, c(p) is a critical value of Jp. Using the definition of c(p) we have c(p) > max{ Jp(−φ1), Jp(φ1)} = λ1. (cid:3) ktφ1+(1−t)φkL2 Thus we have proved the following: Theorem 2.9 For each p ≥ 0, the point (p + c(p), c(p)), where c(p) > λ1 is defined by the minimax formula (2.9), then the point (p + c(p), c(p)) belongs to PK. It is a trivial fact that PK is symmetric with respect to diagonal. The whole curve, that we obtain using Theorem 2.9 and symmetrizing, is denoted by C := {(p + c(p), c(p)), (c(p), p + c(p)) : p ≥ 0}. 3 First Nontrivial Curve Then we state some topological properties of the functional Jp and finally we prove that the We start this section by establishing that the lines R × {λ1} and {λ1} × R are isolated inPK . curve C constructed in the previous section is the first nontrivial curve in the spectrum PK . Proposition 3.1 The lines R × {λ1} and {λ1} × R are isolated in PK . In other words, there exists no sequence (αk, βk) ∈PK with αk > λ1 and βk > λ1 such that (αk, βk) → (α, β) with Suppose by contradiction that there exists a sequence (αk, βk) ∈ PK with αk, Proof. βk > λ1 and (αk, βk) → (α, β) with α or β = λ1. Let uk ∈ X0 be a solution of α = λ1 or β = λ1. −LKuk = αku+ k − βku− k in Ω, uk = 0 on Rn \ Ω (3.1) with kukkL2 = 1. Then we have ZQ (uk(x) − uk(y))2K(x − y)dxdy = αkZΩ (u+ k )2dx − βkZΩ (u− k )2dx ≤ αk, which shows that uk is bounded sequence in X0. Therefore up to a subsequence uk ⇀ u weakly in X0 and uk → u strongly in L2(Ω). Then the limit u satisfies ZQ (u(x) − u(y))2K(x − y)dxdy = λ1ZΩ (u+)2dx − βZΩ (u−)2dx, since uk ⇀ u weakly in X0 and h J ′ p(uk), uk − ui → 0 as k → ∞. i.e u is a weak solution of −LKu = αu+ − βu− in Ω, u = 0 on Rn \ Ω (3.2) Fucik Spectrum for non-local elliptic operators 10 where we have consider the case α = λ1. Multiplying by u+ in (3.2), integrate, using (u(x) − u(y))(u+(x) − u+(y)) = (u+(x) − u+(y))2 + u+(x)u−(y) + u+(y)u−(x) and (2.4), we get ZQ (u+(x) − u+(y))2K(x − y)dxdy + 2ZQ u+(x)u−(y)K(x − y)dxdy = λ1ZΩ (u+)2dx. Using this we have, λ1ZΩ Thus (u+)2dx ≤ZQ ZQ (u+(x) − u+(y))2K(x − y)dxdy ≤ λ1ZΩ (u+)2dx (u+(x) − u+(y))2K(x − y)dxdy = λ1ZΩ (u+)2dx, If u+ ≡ 0 then u ≤ 0 and (3.2) implies that u is an so that either u+ ≡ 0 or u = φ1. eigenfunction with u ≤ 0 so that u = −φ1. So in any case uk converges to either φ1 or −φ1 in Lp(Ω). Thus for every ǫ > 0 either {x ∈ Ω : uk(x) ≤ ǫ} → 0 or {x ∈ Ω : uk(x) ≥ ǫ} → 0. (3.3) On the other hand, taking u+ k as test function in (3.1), we get ZQ (u+ k (x) − u+ k (y))2K(x − y)dxdy + 2ZQ k (x)u− u+ k (y)K(x − y)dxdy = αkZΩ (u+ k )2dx. Using this, Holders inequality and Sobolev embeddings we get (u+ k (x)−u+ k (y))2K(x − y)dxdy ZQ ≤ZQ = αkZΩ (u+ k )2dx ≤ αkC{x ∈ Ω : uk(x) > 0}1− 2 X0 q ku+ k k2 (u+ k (x) − u+ k (y))2K(x − y)dxdy + 2ZQ k (x)u− u+ k (y)K(x − y)dxdy with a constant C > 0, 2 < q ≤ 2∗ = 2n n−2s . Then we have {x ∈ Ω : uk(x) > 0}1− 2 q ≥ α−1 k C −1. Similarly, one can show that {x ∈ Ω : uk(x) < 0}1− 2 q ≥ β−1 k C −1. have that uk changes sign. Hence, from the above inequalities, we get a contradiction with Since (αk, βk) does not belong to the trivial lines λ1 × R and R × λ1 of PK , by using (3.1) we (3.3). Hence the trivial lines λ1 × R and R × λ1 are isolated in PK. Fucik Spectrum for non-local elliptic operators 11 Lemma 3.2 [3] Let P = {u ∈ X0 :RΩ u2 = 1} then 1. P is locally arcwise connected. 2. Any open connected subset O of P is arcwise connected. 3. If O′ is any connected component of an open set O ⊂ P, then ∂O′ ∩ O = ∅. Lemma 3.3 Let O = {u ∈ P : Jp(u) < r}, then any connected component of O contains a critical point of Jp. Proof. Proof follows in the same lines as Lemma 3.6 of [3] by replacing k.k1,p by k.kX0 . (cid:3) Theorem 3.4 Let p ≥ 0 then the point (p + c(p), c(p)) is the first nontrivial point in the Proof. Assume by contradiction that there exists µ such that λ1 < µ < c(p) and (p+µ, µ) ∈ intersection between PK and the line (p, 0) + t(1, 1). PK. Using the fact that {λ1} × R and R × {λ1} are isolated in PK and PK is closed we In other words Jp has a critical value µ with can choose such a point with µ minimum. λ1 < µ < c(p), but there is no critical value in (λ1, µ). If we construct a path connecting from −φ1 to φ1 such that Jp ≤ µ, then we get a contradiction with the definition of c(p), which completes the proof. Let u ∈ P be a critical point of Jp at level µ. Then u satisfies, ZQ (u(x) − u(y))(v(x) − v(y))K(x − y)dxdy = (p + µ)ZΩ u+vdx − µZΩ u−vdx. for all v ∈ X0. Replacing v by u+ and u−, we have ZQ (u+(x) − u+(y))2K(x − y)dxdy + 2ZQ u+(x)u−(y)K(x − y)dxdy = (p + µ)ZΩ (u+)2dx, and ZQ (u−(x) − u−(y))2K(x − y)dxdy + 2ZQ u+(x)u−(y)K(x − y)dxdy = µZΩ (u−)2dx. Thus we obtain, Jp(u) = µ, Jp(cid:18) u+ ku+kL2(cid:19) = µ − 2RQ u+(x)u−(y)K(x − y)dxdy ku+k2 L2 , and Jp(cid:18) u− Jp(cid:18)− ku−kL2(cid:19) = µ − p − ku−kL2(cid:19) = µ − u− ku−k2 L2 2RQ u+(x)u−(y)K(x − y)dxdy 2RQ u+(x)u−(y)K(x − y)dxdy ku−k2 L2 , . Fucik Spectrum for non-local elliptic operators 12 Since u changes sign, the following paths are well-defined on P: u1(t) = (1 − t)u + tu+ k(1 − t)u + tu+kL2 , u2(t) = tu− + (1 − t)u+ ktu− + (1 − t)u+kL2 , u3(t) = −tu− + (1 − t)u k − tu− + (1 − t)ukL2 . Then by using above calculation one can easily get that for all t ∈ [0, 1], Jp(u1(t)) = RQ[(u+(x) − u+(y))2 + (1 − t)2(u−(x) − u−(y))2]K(x − y)dxdy ku+ − (1 − t)u−k2 L2 + 4(1 − t)RQ u−(x)u−(y)K(x − y)dxdy − pRΩ(u+)2dx 2t2RQ u+(x)u−(y)K(x − y)dxdy ku+ − (1 − t)u−k2 L2 ku+ − (1 − t)u−k2 L2 . = µ − Jp(u2(t)) = RQ[(1 − t)2(u+(x) − u+(y))2 + t2(u−(x) − u−(y))2]K(x − y)dxdy k(1 − t)u+ + tu−k2 L2 − 4t(1 − t)RQ u+(x)u−(y)K(x − y)dxdy + p(1 − t)2RΩ(u+)2 + pt2RΩ(u−)2 k(1 − t)u+ + tu−k2 L2 = µ − 2RQ u+(x)u−(y)K(x − y)dxdy k(1 − t)u+ + tu−k2 L2 − pt2RΩ(u−)2dx k(1 − t)u+ + tu−k2 L2 . Jp(u3(t)) = RQ[(1 − t)2(u+(x) − u+(y))2 + (u−(x) − u−(y))2]K(x − y)dxdy 4(1 − t)RQ u+(x)u−(y)K(x − y)dxdy − p(1 − t)2RΩ(u+)2dx 2t2RQ u+(x)u−(y)K(x − y)dxdy k(1 − t)u+ − u−k2 L2 k(1 − t)u+ − u−k2 L2 k(1 − t)u+ − u−k2 L2 = µ − + . Let O = {v ∈ P : Jp(v) < µ − p}. Then clearly φ1 ∈ O, while −φ1 ∈ O if µ − p > λ1. Moreover φ1 and −φ1 are the only possible critical points of Jp in O because of the choice of µ. u− ku−kL2(cid:17) ≤ µ−p, We note that Jp(cid:16) u− does not change sign and vanishes on a set of positive measure, it is not a critical point of Jp. Therefore there exists a C 1 path η : [−ǫ, ǫ] → P with η(0) = u− to a point v with Jp(v) < µ − p. Taking a connected component of O containing v and applying Lemma 3.3 we have that either φ1 or −φ1 is in this component. Let us assume that it is φ1. So we Jp(η(t))t=0 6= 0. Using this path we can move from u− and d dt ku−kL2 ku−kL2 ku−kL2 continue by a path u4(t) from (cid:16) u− −u4(t) connects (cid:16)− u− ku−kL2(cid:17) to −φ1. We observe that ku−kL2(cid:17) to φ1 which is at level less than µ. Then the path Jp(u) − Jp(−u) ≤ p. Fucik Spectrum for non-local elliptic operators 13 Then it follows that Jp(−u4(t)) ≤ Jp(u4(t)) + p ≤ µ − p + p = µ ∀ t. Connecting u1(t), u2(t) and u4(t), we get a path from u to φ1 and joining u3(t) and −u4(t) we get a path from u to −φ1. These yields a path γ(t) on P joining from −φ1 to φ1 such that Jp(γ(t)) ≤ µ for all t, which concludes the proof. (cid:3) Corollary 3.5 The second eigenvalue λ2 of (1.2) has the variational characterization given as λ2 = inf γ∈Γ (u(x) − u(y))2K(x − y)dxdy, max u∈γ[−1,1]ZQ where Γ is same as in Proposition 2.8. Proof. Take s = 0 in Theorem 3.4. Then we have c(0) = λ2 and (2.9) concludes the proof. (cid:3) 4 Properties of the curve In this section we prove that the curve C is Lipschitz continuous, has a certain asymptotic behavior and strictly decreasing. Proposition 4.1 The curve p → (p + c(p), c(p)), p ∈ R+ is Lipschitz continuous. Proof follows as in Proposition 4.1 of [3]. For completeness we give details. Let Proof. p1 < p2 then Jp1(u) > Jp2(u) for all u ∈ P. So we have c(p1) ≥ c(p2). Now for every ǫ > 0 there exists γ ∈ Γ such that max u∈γ[−1,1] Jp2(u) ≤ c(p2) + ǫ, and so 0 ≤ c(p1) − c(p2) ≤ max u∈γ[−1,1] Jp1(u) − max u∈γ[−1,1] Jp2(u) + ǫ. Let u0 ∈ γ[−1, 1] such that then max u∈γ[−1,1] Jp1(u) = Jp1(u0) 0 ≤ c(p1) − c(p2) ≤ Jp1(u0) − Jp2(u0) + ǫ ≤ p2 − p1 + ǫ, as ǫ > 0 is arbitrary so the curve C is Lipschitz continuous with constant ≤ 1. Lemma 4.2 Let A, B be two bounded open sets in Rn, with A ( B and B is connected then λ1(A) > λ1(B). Fucik Spectrum for non-local elliptic operators 14 Proof. From the Theorem 1 and 2 of [20], we see that φ1 is continuous and is a solution of (1.2) in viscosity sense. Then from Lemma 12 of [10], φ1 > 0. Now the variational characterization we see that for A ⊂ B, λ1(A) ≥ λ1(B). Since φ1(B) > 0 in B, we get the strict inequality as claimed. Lemma 4.3 Let (α, β) ∈ C, and let α(x), β(x) ∈ L∞(Ω) satisfying λ1 ≤ α(x) ≤ α, λ1 ≤ β(x) ≤ β. (4.1) Assume that λ1 < α(x) and λ1 < β(x) on subsets of positive measure. (4.2) Then any non-trivial solution u of − LKu = α(x)u+ − β(x)u− in Ω, u = 0 in Rn \ Ω. (4.3) changes sign in Ω and α(x) = α a.e. on {x ∈ Ω : u(x) > 0}, β(x) = β a.e. on {x ∈ Ω : u(x) < 0}. Proof. Let u be a nontrivial solution of (4.3). Replacing u by −u if necessary. we can assume that the point (α, β) in C is such that α ≥ β. We first claim that u changes sign in Ω. Suppose by contradiction that this is not true, first consider the case u ≥ 0,(case u ≤ 0 can be prove similarly). Then u solves −Lku = α(x)u in Ω u = 0 on Rn \ Ω. This implies that the first eigenvalue of −LK on X0 with respect to weight α(x) is equal to 1. i.e inf(RQ(v(x) − v(y))2K(x − y)dxdy : v ∈ X0, v 6= 0) = 1. (4.4) We deduce from (4.1), (4.2) and (4.4) that RΩ α(x)v2dx 1 = RQ(φ1(x) − φ1(y))2K(x − y)dxdy λ1 a contradiction and hence the claim. Again we assume by contradiction that either > RQ(φ1(x) − φ1(y))2K(x − y)dxdy 1dx RΩ α(x)φ2 or {x ∈ X0 : α(x) < α and u(x) > 0} > 0 {x ∈ X0 : β(x) < β and u(x) < 0} > 0. ≥ 1, (4.5) (4.6) Fucik Spectrum for non-local elliptic operators 15 Suppose that (4.5) holds (a similar argument will hold for (4.6)). Put α − β = p ≥ 0. Then β = c(p), where c(p) is given by (2.9). We show that there exists a path γ ∈ Γ such that max u∈γ[−1,1] Jp(u) < β, (4.7) which yields a contradiction with the definition of c(p). In order to construct γ we show that there exists of a function v ∈ X0 such that it changes sign and satisfies RQ(v+(x) − v+(y))2K(x − y)dxdy RΩ(v+)2dx < α and RQ(v−(x) − v−(y))2K(x − y)dxdy RΩ(v−)2dx < β. (4.8) For this let O1 be a component of {x ∈ Ω : u(x) > 0} satisfying x ∈ O1 : α(x) < α > 0, which is possible by (4.5). Let O2 be a component of {x ∈ Ω : u(x) < 0} satisfying which is possible by (4.6). Then we claim that x ∈ O1 : β(x) < β > 0, λ1(O1) < α and λ1(O2) ≤ β, (4.9) where λ1(Oi) denotes the first eigenvalue of −Lk on X0Oi = {u ∈ XOi : u = 0 on Rn \ Oi}. Clearly uOi ∈ X0Oi then we have (u(x) − u(y))2K(x − y)dxdy (u(x) − u(y))2K(x − y)dxdy RQO1 u2dx RO1 < αRQO1 RO1 α(x)u2dx = α which implies that λ1(O1) < α. The other inequality in (4.9) is proved similarly. Now with some modification on the sets O1 and O2, we construct the sets O1 and O2 such that O1 ∩ O2 = ∅ and λ1( O1) < α and λ1( O2) < β. For ν ≥ 0, O1(ν) = {x ∈ O1 : dist(x, Oc 1) > ν}. Clearly λ1(O1(ν)) ≥ λ1(O1)) and moreover λ1(O1(ν)) → λ1(O1)) as ν → 0. Then there exists ν0 > 0 such that λ1(O1(ν)) < α for all 0 ≤ ν ≤ ν0. (4.10) Let x0 ∈ ∂O2 ∩ Ω (is not empty as O1 ∩ O2 = ∅) and choose 0 < ν < min{ν0, dist(x0, Ωc)} and O1 = O1(ν) and O2 = O2 ∩ B(x0, ν 2 ). Then O1 ∩ O2 = ∅ and by (4.10), λ1( O1) < α. Since O2 is connected then by (4.9) and Lemma 4.2, we get λ1( O2) < β. Now we define v = v1 − v2, where vi are the the extension by zero outside Oi of the eigenfunctions associated to λi( Oi). Fucik Spectrum for non-local elliptic operators 16 Then v satisfies (4.8). Thus there exist v ∈ X0 which changes sign and satisfies condition (4.8), and moreover we have Jp(cid:18) v kvk2 L2 kvkL2(cid:19) = RQ(v+(x) − v+(y))2K(x − y)dxdy − 2RQ v+(x)v−(y)K(x − y)dxdy + βRΩ(v−)2dx < (α − p)RΩ(v+)2dx kvk2 L2 kvk2 L2 kvk2 L2 Jp(cid:18) v+ kv+kL2(cid:19) < α − p = β, kvk2 L2 kvk2 L2 + RQ(v−(x) − v−(y))2K(x − y)dxdy − pRΩ(v+)2dx − 2RQ v+(x)v−(y)K(x − y)dxdy Jp(cid:18) v− kv−kL2(cid:19) < β − p. kvk2 L2 < β. Using Lemma 3.3, we have that there exists a critical point in the connected component of the set O = {u ∈ P : Jp(u) < β − p}. As the point (α, β) ∈ C, the only possible critical point is φ1, then we can construct a path from −φ1 to φ1 exactly in the same manner as in Theorem 3.4 only by taking v in place of u. Thus we have construct a path satisfying (4.7), and hence the result follows. (cid:3) Corollary 4.4 Let (α, β) ∈ C and let α(x), β(x) ∈ L∞(Ω) satisfying λ1 ≤ α(x) ≤ α a.e, λ1 ≤ β(x) ≤ β a.e. Assume that λ1 < α(x) and λ1 < β(x) on subsets of positive measure. If either α(x) < α a.e in Ω or β(x) < β a.e. in Ω. Then (4.3) has only the trivial solution. Lemma 4.5 The curve p → (p + c(p), c(p)) is strictly decreasing, (in the sense that p1 < p2 implies p1 + c(p1) < p2 + c(p2) and c(p1) > c(p2)). Proof. Let p1 < p2 and suppose by contradiction that either (i) p1 +c(p1) ≥ p2 +c(p2) or (ii) c(p1) ≤ c(p2). In case (i) we deduce from p1+c(p1) ≥ p2+c(p2) > p1+c(p2) that c(p1) ≥ c(p2). If we take (α, β) = (p1 + c(p1), c(p1)) and (α(x), β(x)) = (p2 + c(p2), c(p2)), then by Corollary 4.4, only solution of (4.3) with (α(x), β(x)) is the trivial solution which contradicts the fact (α, β) = (p2 + c(p2), c(p2)) and (α(x), β(x)) = (p1 + c(p1), c(p1)), then only solution of (4.3) that (p2 + c(p2), c(p2)) ∈PK. If (ii) holds then p1 + c(p1) ≤ p1 + c(p2) < p2 + c(p2), if we take with (α(x), β(x)) is the trivial one which contradicts the fact that (p1 + c(p1), c(p1)) ∈ PK and hence the result follows. (cid:3) As c(p) is decreasing and positive so limit of c(p) exists as p → ∞. In the next Theorem we find the asymptotic behavior of the first nontrivial curve. Theorem 4.6 If n ≥ 2s then the limit of c(p) as p → ∞ is λ1. Proof. For n ≥ 2s, we can choose a function φ ∈ X0 such that there does not exist r ∈ R such that φ(x) ≤ rφ1(x) a.e. in Ω. For this it suffices to take φ ∈ X0 such that it is unbounded from above in a neighborhood of some point x ∈ X0. Then by contradiction argument, one can similarly show c(p) → λ1 as p → ∞ as in Proposition 4.4 of [3]. (cid:3) Fucik Spectrum for non-local elliptic operators 5 Non Resonance between (λ1, λ1) and C In this section we study the following problem ( −LKu = f (x, u) in Ω u = 0 on Rn \ Ω, 17 (5.1) where f (x, u)/u lies asymptotically between (λ1, λ1) and (α, β) ∈ C. Let f : Ω × R → R be a function satisfying L∞(Ω) Caratheodory conditions. Given a point (α, β) ∈ C, we assume following: γ±(x) ≤ lim inf s→±∞ f (x, s) s ≤ lim sup s→±∞ f (x, s) s ≤ Γ±(x) (5.2) hold uniformly with respect to x, where γ±(x) and Γ±(x) are L∞ functions which satisfy ( λ1 ≤ γ+(x) ≤ Γ+(x) ≤ α a.e. in Ω λ1 ≤ γ−(x) ≤ Γ−(x) ≤ β a.e. in Ω. Write F (x, s) =R s 0 f (x, t)dt, we also assume the following inequalities: δ±(x) ≤ lim inf s→±∞ 2F (x, s) s2 ≤ lim sup s→±∞ 2F (x, s) s2 ≤ ∆±(x) (5.3) (5.4) hold uniformly with respect to x, where δ±(x) and ∆±(x) are L∞ functions which satisfy λ1 ≤ δ+(x) ≤ ∆+(x) ≤ α a.e. in Ω λ1 ≤ δ−(x) ≤ ∆−(x) ≤ β a.e. in Ω δ+(x) > λ1 and δ−(x) > λ1 on subsets of positive measure, either ∆+(x) < α a.e. in Ω or ∆−(x) < β a.e. in Ω. (5.5)   Theorem 5.1 Let (5.2), (5.3), (5.4), (5.5) hold and (α, β) ∈ C. Then problem (5.1) admits at least one solution u in X0. Define the energy functional Ψ : X0 → R as Ψ(u) = 1 2ZQ (u(x) − u(y))2K(x − y)dxdy −ZΩ F (x, u)dx Then Ψ is a C 1 functional on X0 and ∀v ∈ X0 hΨ′(u), vi =ZQ (u(x) − u(y))(v(x) − v(y))K(x − y)dxdy −ZΩ f (x, u)vdx and critical points of Ψ are exactly the weak solutions of (5.1). Lemma 5.2 Ψ satisfies the (P.S) condition on X0. Fucik Spectrum for non-local elliptic operators Proof. Let uk be a (P.S) sequence in X0, i.e Ψ(uk) ≤ c, hΨ′(uk), vi ≤ ǫkkvkX0 , ∀ v ∈ X0, 18 (5.6) where c is a constant and ǫk → 0 as k → ∞. It suffices to show that uk is a bounded sequence in X0. Assume by contradiction that uk is not a bounded sequence. Then define vk = uk . kukkX0 Then vk is a bounded sequence. Therefore there exists a subsequence vk of vk and v0 ∈ X0 such that vk ⇀ v0 weakly in X0, vk → v0 strongly in L2(Ω) and vk(x) → v0(x) a.e. in Rn. Also by using (5.2) and (5.3), we have f (x, uk)/kukkX0 ⇀ f0(x) weakly in L2(Ω). Take v = vk − v0 in (5.6) and divide by kukkX0 we get vk → v0. In particular kv0kX0 = 1. One can easily seen from (5.6) that ZQ (v0(x) − v0(y))(v(x) − v(y))K(x − y)dxdy −ZΩ f0(x)vdx = 0 ∀ v ∈ X0. Now by standard argument based on assumption (5.2), f0(x) = α(x)v+ 0 for some L∞ functions α(x), β(x) satisfying (4.1). In the expression of f0(x), the value of α(x) (resp.β(x)) on {x : v0(x) ≤ 0} (resp. {x : v0(x) ≥ 0}) are irrelevant, and consequently we can assume that 0 −β(x)v− α(x) > λ1 on {x : v0(x) ≤ 0} and β(x) > λ1 on {x : v0(x) ≥ 0}. (5.7) So v0 is a nontrivial solution of equation (4.3). It then follows from Lemma 4.3 that either (i) α(x) = λ1 a.e in Ω or (ii) β(x) = λ1 a.e in Ω, or (iii) v0 is an eigenfunction associated to the point (α, β) ∈ C. We show that in each cases we get a contradiction. If (i) holds then 0, which by (5.7), v0 > 0 a.e. in Ω and (4.3) gives RQ(v0(x) − v0(y))2K(x − y)dxdy = λ1RΩ v2 implies that v0 is a multiple of φ1. Dividing (5.6) by kukk2 X0 and taking limit we get, λ1ZΩ v2 0 =ZQ (v0(x) − v0(y))2K(x − y)dxdy = lim k→∞ZΩ 2F (x, uk) kukk2 X0 ≥ZΩ δ+(x)v2 0dx. This contradicts assumption (5.5). The case (ii) is treated similarly. Now if (iii) holds, we deduce from (5.4) that ZΩ α(v+ 0 )2 + β(v− 0 )2 =ZQ ≤ZΩ (v0(x) − v0(y))2K(x − y)dxdy = lim k→∞ZΩ 2F (x, uk) kukk2 X0 ∆+(x)(v+ 0 )2 + ∆−(x)(v− 0 )2. This contradicts assumption (5.5), since v0 changes sign. Hence uk is bounded sequence in X0. (cid:3) Fucik Spectrum for non-local elliptic operators Now we study the geometry of Ψ. Lemma 5.3 There exists R > 0 such that max{Ψ(Rφ1), Ψ(−Rφ1)} < max u∈γ[−1,1] Ψ(u) for any γ ∈ Γ1 := {γ ∈ C([−1, 1], X0) : γ(±1) = ±Rφ1}. 19 (5.8) Proof. From (5.4), we have for any ǫ > 0 there exists aǫ(x) ∈ L2(Ω) such that for a.e x, ( (δ+(x) − ǫ) s2 (δ−(x) − ǫ) s2 2 − aǫ(x) ≤ F (x, s) ≤ (∆+(x) + ǫ) s2 2 − aǫ(x) ≤ F (x, s) ≤ (∆−(x) + ǫ) s2 2 + aǫ(x) ∀ s > 0 2 + aǫ(x) ∀ s < 0. (5.9) Now consider the following functional associated to the functions ∆±(x) as Φ(u) =ZQ (u(x) − u(y))2K(x − y)dxdy −ZΩ ∆+(x)(u+)2dx −ZΩ ∆−(x)(u−)2dx. Then we claim that d = inf γ∈Γ max u∈γ[−1,1] Ψ(u) > 0 (5.10) where Γ is the set of all continuous paths from −φ1 to φ1 in P. Write p = α − β ≥ 0. we can choose (α, β) ∈ C such that α ≥ β (as replacing u by −u if necessary), we have for any γ ∈ Γ max u∈γ[−1,1] Jp(u) ≥ c(p) = β. i.e. max u∈γ[−1,1](cid:18)ZQ (u(x) − u(y))2 K(x − y)dxdy −ZΩ α(u+)2dx −ZΩ β(u−)2dx(cid:19) ≥ 0 which implies by (5.5). So d ≥ 0. On the other hand, since δ±(x) ≤ ∆±(x), max u∈γ[−1,1] Φ(u) ≥ 0, Φ(±φ1) ≤ZΩ (λ1 − δ±(x))φ2 1dx < 0 by (5.5). Thus we have a mountain pass geometry for the restriction Φ of Φ to P, max{ Φ(φ1), Φ(−φ1)} < 0 ≤ max u∈γ[−1,1] Φ(u) for any path γ ∈ Γ and moreover one can verify exactly as in Lemma 2.6 that Φ satisfies the (P.S.) condition on X0. Then d is a critical value of Φ i.e there exists u ∈ P and µ ∈ R such that ( Φ(u) = d hΦ′(u), vi = µhI ′(u), vi ∀ v ∈ X0. Fucik Spectrum for non-local elliptic operators 20 Assume by contradiction that d = 0. Taking v = u in above, we get µ = 0 so u is a nontrivial solution of −LKu = ∆+(x)u+ − ∆−(x)u− in Ω u = 0 in Rn \ Ω Using (5.5), we get a contradiction with Lemma 4.3 . This completes the proof of claim. Next we show that (5.8) hold. From the left hand side of inequality (5.9), we have for R > 0 and η > 0, Ψ(±Rφ1) ≤ R2 2 ZΩ (λ1 − δ±(x))φ2 1 + + kaηkL1, ηR2 2 Then ,Ψ(±Rφ1) → −∞ as R → +∞, by (5.5) and letting η to be sufficiently small. Fix ǫ with 0 < ǫ < d. We can choose R = R(ǫ) so that Ψ(±Rφ1) < −kaǫkL1, (5.11) where aǫ is associated to ǫ using (5.9). Consider a path γ ∈ Γ1. Then if 0 ∈ γ[−1, 1], then by (5.11), Ψ(±Rφ1) < −kaǫkL1 ≤ 0 = Ψ(0) ≤ max u∈γ[−1,1] Ψ(u), so Lemma is proved in this case. If 0 6∈ γ[−1, 1], then we take the normalized path γ(t) = γ(t) kγ(t)kL2 belongs to Γ. Since by (5.9), Ψ(u) ≥ Φ(u) − ǫkuk2 L2 2 − kaǫkL1, we obtain max γ∈[−1,1] 2Ψ(u) + ǫkuk2 kuk2 L2 L2 + 2kaǫkL1 ≥ max γ∈[−1,1] Φ(v) ≥ d, and consequently, by choice of ǫ, This implies that max γ∈[−1,1] 2Ψ(u) + 2kaǫkL1 kuk2 L2 ≥ d − ǫ > 0, max u∈γ[−1,1] Ψ(u) > −kaǫkL1 > Ψ(±Rφ1), by (5.11) and hence the Lemma. Proof of Theorem 5.1: Lemmas 5.2 and 5.3 completes the proof. References [1] M. Alif, Fucik spectrum for the Neumann problem with indefinite weights, Partial differential equations, volume 229 of Lecture Notes in Pure and Appl. Math., Dekker, New York, (2002) pp. 45-62. [2] M. Arias and J. Campos, Radial Fucik spectrum of the Laplace operator, J. Math. Anal. Appl., 190 (1995) 654-666. Fucik Spectrum for non-local elliptic operators 21 [3] M. Cuesta, D. de Figueiredo and J.-P. Gossez, The Beginning of the Fucik Spectrum for the p-Laplacian, Journal of Differential Equations, 159 (1999) 212-238. [4] M. Cuesta and J.-P. Gossez, A variational approach to nonresonance with respect to the Fucik spectrum, Nonlinear Anal., 19 (1992) 487-500. [5] Norman Dancer and Kanishka Perera1, Some Remarks on the Fucik Spectrum of the p-Laplacian and Critical Groups, Journal of Mathematical Analysis and Applications, 254 (2001) 164177. [6] D. de Figueiredo and J.-P. Gossez, On the first curve of the Fucik spectrum of an elliptic operator, Differential Integral Equations, 7 (1994) 1285-1302. [7] Dumitru Motreanu and Patrick Winkert, On the Fucik spectrum of p-Laplacian with Robin boundary condition, Nonlinear Analysis, 74 (2011) 4671-4681. [8] E. Di Nezza, G. Palatucci and E. Valdinoci, Hitchhikers guide to the fractional Sobolev spaces, preprint, available at http://arxiv.org/abs/1104.4345 . [9] S. Fucik, Solvability of Nonlinear Equations and Boundary Value Problems, in: Mathematics and its Applications, vol. 4, D. Reidel Publishing Co., Dordrecht, 1980. [10] E. Lindgren, P. Lindqvist, Fractional Eigenvalues, to appear in Calculus of Variations 2013. [11] A. M. Micheletti and A. Pistoia, A note on the resonance set for a semilinear elliptic equation and an application to jumping nonlinearities, Topol. Methods Nonlinear Anal., 6 (1995) 67-80. [12] A. M. Micheletti and A. Pistoia, On the Fucik spectrum for the p-Laplacian, Differential Integral Equa- tions, 14 (2001) 867-882. [13] Sandra R. Martinez, Julio D. Rossi, On the Fucik spectrum and a resonance problem for the p-Laplacian with a nonlinear boundary condition, Nonlinear Analysis Theory Methods and Applications, 59, no. 6 (2004) 813-848. [14] K. Perera, Resonance problems with respect to the Fucik spectrum of the p-Laplacian. Electronic Journal of Differential Equations, pages No. 36, 10 pp. (electronic), 2002. [15] K. Perera., On the Fucik spectrum of the p-Laplacian, NoDEA Nonlinear Differential Equations Appl., 11 2 (2004) 259-270. [16] M. Schechter, The Fucik spectrum, Indiana Univ. Math. J., 43 (1994) 1139-1157. [17] R. Servadei and E. Valdinoci, Mountain pass solutions for non-local elliptic operators, J. Math. Anal. Appl. 389 (2012) 887-898. [18] R. Servadei and E. Valdinoci, Variational methods for non-local operators of elliptic type, Discrete Contin. Dyn. Syst., 33 no. 5 (2013) 2105-2137 [19] R. Servadei and E. Valdinoci, Lewy-Stampacchia type estimates for variational inequalities driven by non-local operators, to appear in Rev. Mat. Iberoam., 29 (2013). [20] R. Servadei and E. Valdinoci, Weak and Viscosity of the fractional Laplace equation, to appear. [21] R. Servadei, A Brezis-Nirenberg result for non-local critical equations in low dimension, Commu. Pure Appl. Math., 12 (2013) 2445-2464
1803.10321	1	1803	2018-03-27T20:58:38	Operator-valued local Hardy spaces	[ "math.FA" ]	This paper gives a systematic study of operator-valued local Hardy spaces. These spaces are localizations of the Hardy spaces defined by Tao Mei, and share many properties with Mei's Hardy spaces. We prove the ${\rm h}_1$-$\rm bmo$ duality, as well as the ${\rm h}_p$-${\rm h}_q$ duality for any conjugate pair $(p,q)$ when $1<p< \infty$. We show that ${\rm h}_1(\mathbb{R}^d, \mathcal M)$ and ${\rm bmo}(\mathbb{R}^d, \mathcal M)$ are also good endpoints of $L_p(L_\infty(\mathbb{R}^d) \overline{\otimes} \mathcal M)$ for interpolation. We obtain the local version of Calder\'on-Zygmund theory, and then deduce that the Poisson kernel in our definition of the local Hardy norms can be replaced by any reasonable test function. Finally, we establish the atomic decomposition of the local Hardy space ${\rm h}_1^c(\mathbb{R}^d,\mathcal M)$.	math.FA	math	OPERATOR-VALUED LOCAL HARDY SPACES RUNLIAN XIA AND XIAO XIONG Abstract. This paper gives a systematic study of operator-valued local Hardy spaces. These spaces are localizations of the Hardy spaces defined by Tao Mei, and share many properties with Mei's Hardy spaces. We prove the h1-bmo duality, as well as the hp-hq duality for any conjugate pair (p, q) when 1 < p < ∞. We show that h1(Rd, M) and bmo(Rd, M) are also good endpoints of Lp(L∞(Rd)⊗M) for interpolation. We obtain the local version of Calder´on-Zygmund theory, and then deduce that the Poisson kernel in our definition of the local Hardy norms can be replaced by any reasonable test function. Finally, we establish the atomic decomposition of the local Hardy space hc 1(Rd, M). Contents p for 1 ≤ p < 2 Introduction and Preliminaries 0. 0.1. Notation 0.2. Noncommutative Lp-spaces 0.3. Operator-valued Hardy spaces 1. Operator-valued local Hardy spaces 1.1. Operator-valued local Hardy spaces 1.2. Operator-valued bmo spaces 2. The dual space of hc 2.1. Definition of bmoc q 2.2. A bounded map 2.3. Duality 2.4. The equivalence hq = bmoq 3. 4. Calder´on-Zygmund theory 5. General characterizations 5.1. General characterizations 5.2. Discrete characterizations 5.3. The relation between hp(Rd,M) and Hp(Rd,M) 6. The atomic decomposition References Interpolation 1 2 3 3 6 6 7 10 10 11 16 20 24 26 29 29 34 37 39 41 0. Introduction and Preliminaries This paper is devoted to the study of operator-valued local Hardy spaces. It follows the current line of investigation of noncommutative harmonic analysis. This field arose from the noncommuta- tive integration theory developed by Murray and von Neumann, in order to provide a mathematical foundation for quantum mechanics. The objective was to construct and study a linear functional on an operator algebra which plays the role of the classical integral. In [37], Pisier and Xu developed a pioneering work on noncommutative martingale theory; since then, many classical results have been successfully transferred to the noncommutative setting, see for instance, [18, 19, 21, 22, 39, 34, 40]. Inspired by the above mentioned developments and the Littlewood-Paley-Stein theory of quan- [20, 25, 24]), Mei [30] studied operator-valued Hardy spaces, which tum Markov semigroups (cf. 2000 Mathematics Subject Classification: Primary: 46L52, 42B30. Secondary: 46L07, 47L65. Key words: Noncommutative Lp-spaces, operator-valued Hardy spaces, operator-valued bmo spaces, duality, interpolation, Calder´on-Zygmund theory, characterization, atomic decomposition. 1 2 R. Xia and X. Xiong are defined by the Littlewood-Paley g-function and Lusin area integral function associated to the Poisson kernel. These spaces are shown to be very useful for many aspects of noncommutative harmonic analysis. In [51], we obtain general characterizations of Mei's Hardy spaces, which state that the Poisson kernel can be replaced by any reasonable test function. This is done mainly by using the operator-valued Calder´on-Zygmund theory. In the classical setting, the theory of Hardy spaces is one of the most important topics in harmoic analysis. The local Hardy spaces hp(Rd) were first introduced by Goldberg [12]. These spaces are viewed as local or inhomogeneous counterparts of the classical real Hardy spaces Hp(Rd). Gold- berg's motivation of introducing these local spaces was the study of pseudo-differential operators. It is known that pseudo-differential operators are not necessarily bounded on the classical Hardy space H1(Rd), but bounded on h1(Rd) under some appropriate assumptions. Afterwards, many other inhomogeneous spaces have also been studied. Our references for the classical theory are [12, 46, 9]. However, they have not been investigated so far in the operator-valued case. Motivated by [52, 51, 30], we provide a localization of Mei's operator-valued Hardy spaces on Rd in this paper. The norms of these spaces are partly given by the truncated versions of the Littlewood-Paley g-function and Lusin area integral function. Some techniques that we use to deal with our local Hardy spaces are modelled after those of [51]; however, some highly non-trivial modifications are needed. Since with the truncation, we only know the Lp-norms of the Poisson integrals of functions on the strip Rd × (0, 1), and lose information when the time is large. This brings some substantial difficulties that the non-local case does not have, for example, the duality problem. Moreover, the noncommutative maximal function method is still unavailable in this setting, while in the classical case it is efficiently and frequently employed. However, based on tools developed recently, for instance, in [37, 18, 21, 39, 40, 20, 30, 31], we can overcome these difficulties. p-bmoc Let us present here the four main results of this paper. The first family of results concerns the p(Rd,M) and bmoc(Rd,M). The first major result of this operator-valued local Hardy spaces hc part is the hc q duality for 1 ≤ p < 2, where q denotes the conjugate index of p. In particular, when p = 1, we obtain the operator-valued local analogue of the classical Fefferman-Stein theorem. The pattern of the proof of this theorem is similar to that of Mei's non-local case. We also show that hc q(Rd,M) for 2 < q < ∞ like in the martingale and non-local settings. Thus the dual of hc The second major result shows that the local Hardy spaces behave well with both complex and q(Rd,M) when 1 < p ≤ 2. p(Rd,M) agrees with hc q(Rd,M) = bmoc real interpolations. In particular, we have (cid:0)bmoc(Rd,M), hc 1(Rd,M)(cid:1) 1 p = hc p(Rd,M) for 1 < p < ∞. We reduce this interpolation problem to the corresponding one on the non-local Hardy spaces in order to use Mei's interpolation result in [30]. This proof is quite simple. The third major result concerns the Calder´on-Zygmund theory. The usual M-valued Calder´on- Zygmund operators which satisfy the Hormander condition are in general not bounded on inho- mogeneous spaces. Thus, in order to guarantee the boundedness of a Calder´on-Zygmund operator on hc p(Rd,M), we need to impose an extra decay at infinity to the kernel. The Calder´on-Zygmund theory mentioned above will be applied to the general characterization of p(Rd,M) with the Poisson kernel replaced by any reasonable test function. This characterization hc will play an important role in our recent study of (inhomogeneous) Triebel-Lizorkin spaces on Rd, see [49]. 0.1. Notation. In the following,we collect some notation which will be frequently used in this paper. Throughout, we will use the notation A . B, which is an inequality up to a constant: A ≤ cB for some constant c > 0. The relevant constants in all such inequalities may depend on the dimension d, the test function Φ or p, etc, but never on the function f in consideration. The equivalence A ≈ B will mean A . B and B . A simultaneously. The Bessel and Riesz potentials are J α = (1− (2π)−2∆) α 2 and I α = (−(2π)−2∆) α If α = 1, we will abbreviate J 1 as J and I 1 as I. We denote also Jα(ξ) = (1 + ξ2) 2 , respectively. 2 on Rd and α Operator-valued local Hardy spaces 3 Iα(ξ) = ξα on Rd \ {0}. Then Jα and Iα are the symbols of the Fourier multipliers J α and I α, respectively. 2 (Rd) the potential Sobolev space, consisting of all tempered distributions f 2 (Rd) will serve as important convolution 2 , the elements in H σ We denote by H σ such that J σ(f ) ∈ L2(Rd). If σ > d kernels in the sequel. 0.2. Noncommutative Lp-spaces. We also recall some preliminaries on noncommutative Lp- spaces and operator-valued Hardy spaces. We start with a brief introduction of noncommutative Lp-spaces. Let M be a von Neumann algebra equipped with a normal semifinite faithful trace τ and S+ be the set of all positive elements x in M with τ (s(x)) < ∞, where s(x) denotes the M support of x, i.e., the smallest projection e such that exe = x. Let SM be the linear span of S+ . M Then every x ∈ SM has finite trace, and SM is a w*-dense ∗-subalgebra of M. define Let 1 ≤ p < ∞. For any x ∈ SM, the operator xp belongs to S+ M (recalling x = (x∗x) 2 ). We 1 kxkp =(cid:0)τ (xp)(cid:1) 1 p . One can prove that k · kp is a norm on SM. The completion of (SM,k · kp) is denoted by Lp(M), which is the usual noncommutative Lp-space associated to (M, τ ). In this paper, the norm of Lp(M) will be often denoted simply by k · kp if there is no confusion. But if different Lp-spaces appear in a same context, we will precise their norms in order to avoid possible ambiguity. We refer the reader to [54] and [38] for further information on noncommutative Lp-spaces. Now we introduce noncommutative Hilbert space-valued Lp-spaces Lp(M; H c) and Lp(M; H r), which are studied at length in [20]. Let H be a Hilbert space and v ∈ H with kvk = 1, and pv be the orthogonal projection onto the one-dimensional subspace generated by v. Then define the following row and column noncommutative Lp-spaces: Lp(M; H r) = (pv ⊗ 1M)Lp(B(H)⊗M) and Lp(M; H c) = Lp(B(H)⊗M)(pv ⊗ 1M), where the tensor product B(H)⊗M is equipped with the tensor trace while B(H) is equipped with the usual trace, and where 1M denotes the unit of M. For f ∈ Lp(M; H c), kfkLp(M;Hc) = k(f∗f ) 1 2 kp. A similar formula holds for the row space by passing to adjoint: f ∈ Lp(M; H r) if and only if f∗ ∈ Lp(M; H c), and kfkLp(M;Hr ) = kf∗kLp(M;Hc). It is clear that Lp(M; H c) and Lp(M; H r) are 1-complemented subspaces of Lp(B(H)⊗M) for any p. 0.3. Operator-valued Hardy spaces. Throughout the remainder of the paper, unless explicitly stated otherwise, (M, τ ) will be fixed as before and N = L∞(Rd)⊗M, equipped with the tensor trace. In this subsection, we introduce Mei's operator-valued Hardy spaces. Contrary to the custom, we will use letters s, t to denote variables of Rd since letters x, y are reserved for operators in noncommutative Lp-spaces. Accordingly, a generic element of the upper half-space Rd+1 + will be denoted by (s, ε) with ε > 0, where Rd+1 + = {(s, ε) : s ∈ Rd, ε > 0}. Let P be the Poisson kernel on Rd: with cd the usual normalizing constant and s the Euclidean norm of s. Let P(s) = cd 1 (s2 + 1) d+1 2 Pε(s) = ) = cd 1 εd P( s ε ε . d+1 2 (s2 + ε2) For any function f on Rd with values in L1(M) + M, its Poisson integral, whenever it exists, will be denoted by Pε(f ): Pε(f )(s) =ZRd Pε(s − t)f (t)dt, (s, ε) ∈ Rd+1 + . Note that the Poisson integral of f exists if f ∈ L1(cid:0)M; Lc 2(Rd, dt 1 + td+1 )(cid:1) + L∞(cid:0)M; Lc 2(Rd, dt 1 + td+1 )(cid:1). 4 R. Xia and X. Xiong This space is the right space in which all functions considered in this paper live as far as only column spaces are involved. As it will appear frequently later, to simplify notation, we will denote the Hilbert space L2(Rd, dt 1+td+1 ) by Rd: (0.1) (0.2) Thus The Lusin area square function of f is defined by Rd = L2(Rd, dt 1 + td+1 ). εd−1(cid:17) 1 Pε(f )(s + t)(cid:12)(cid:12)2 dt dε 2 Sc(f )(s) =(cid:16)ZΓ(cid:12)(cid:12) ∂ ∂ε , s ∈ Rd, where Γ is the cone {(t, ε) ∈ Rd+1 Hc p(Rd,M) to be + : t < ε}. For 1 ≤ p < ∞ define the column Hardy space p(Rd,M) =(cid:8)f : kfkHc Hc p = kSc(f )kp < ∞(cid:9). Note that [30] uses the gradient of Pε(f ) instead of the sole radial derivative in the definition of Sc above, but this does not affect Hc p(Rd,M) (up to equivalent norms). At the same time, it is proved in [30] that Hc (0.3) p(Rd,M) can be equally defined by the Littlewood-Paley g-function: , 2 Gc(f )(s) =(cid:16)Z ∞ 0 ε(cid:12)(cid:12) ∂ ∂ε Pε(f )(s)(cid:12)(cid:12)2 dε(cid:17) 1 p(Rd,M). f ∈ Hc s ∈ Rd. kfkHc p ≈ kGc(f )kp, The row Hardy space Hr p = kf∗kHc the norm kfkHr p(Rd,M) is the space of all f such that f∗ ∈ Hc . Finally, we define the mixture space Hp(Rd,M) as p p(Rd,M), equipped with equipped with the sum norm Hp(Rd,M) = Hc p(Rd,M) + Hr kfkHp = inf(cid:8)kf1kHc p + kf2kHr p(Rd,M) for 1 ≤ p ≤ 2 p : f = f1 + f2(cid:9), and Hp(Rd,M) = Hc equipped with the intersection norm Observe that Hc 2(Rd,M) = Hr It is proved in [30] that for 1 < p < ∞ p(Rd,M) for 2 < p < ∞ p(Rd,M) ∩ Hr kfkHp = max(cid:0)kfkHc 2(Rd,M) = L2(N ) with equivalent norms. ,kfkHr p(cid:1). p Hp(Rd,M) = Lp(N ) with equivalent norms. The operator-valued BMO spaces are also studied in [30]. Let Q be a cube in Rd (with sides parallel to the axes) and Q its volume. For a function f with values in M, fQ denotes its mean over Q: The column BMO norm of f is defined to be fQ = 1 f (t)dt. QZQ QZQ(cid:12)(cid:12)f (t) − fQ(cid:12)(cid:12)2 1 1 2 M . dt(cid:13)(cid:13)(cid:13) kfkBMOc = sup Q⊂Rd(cid:13)(cid:13)(cid:13) (0.4) Then BMOc(Rd,M) =(cid:8)f ∈ L∞(cid:0)M; Rc d(cid:1) : kfkBMOc < ∞(cid:9). Similarly, we define the row space BMOr(Rd,M) as the space of f such that f∗ lies in BMOc(Rd,M), and BMO(Rd,M) = BMOc(Rd,M) ∩ BMOr(Rd,M) with the intersection norm. 1(Rd,M) can be naturally identified with BMOc(Rd,M). In [30], it is showed that the dual of Hc This is the operator-valued analogue of the celebrated Fefferman-Stein H1-BMO duality theorem. On the other hand, one of the main results of [51] asserts that the Poisson kernel in the definition of Hardy spaces can be replaced by more general test functions. Operator-valued local Hardy spaces 5 Take any Schwartz function Φ with vanishing mean. We will assume that Φ is nondegenerate in the following sense: (0.5) Set Φε(s) = ε−dΦ( s defined by replacing the partial derivative of the Poisson kernel P in Sc(f ) and Gc(f ) by Φ : ε ) for ε > 0. The radial and conic square functions of f associated to Φ are Sc ∀ ξ ∈ Rd \ {0} ∃ ε > 0, s.t. bΦ(εξ) 6= 0. Φ(f )(s) =(cid:16)ZΓ Φε ∗ f (s + t)2 dtdε εd+1(cid:17) 1 ε (cid:17) 1 Φε ∗ f (s)2 dε Φ(f )(s) =(cid:16)Z ∞ Gc 0 , . 2 2 s ∈ Rd The following two lemmas are taken from [51]. The first one says that the two square functions above define equivalent norms in Hc Lemma 0.1. Let 1 ≤ p < ∞ and f ∈ L1(M; Rc if Gc p(Rd,M) if and only d). Then f ∈ Hc d) + L∞(M; Rc p(Rd,M): Φ(f ) ∈ Lp(N ) if and only if Sc Φ(f ) ∈ Lp(N ). If this is the case, then kGc Φ(f )kp ≈ kSc Φ(f )kp ≈ kfkHc p with the relevant constants depending only on p, d and Φ. The above square functions Gc Φ and Sc Φ can be discretized as follows: Gc,D Φ (f )(s) =(cid:16) ∞Xj=−∞ Φ (f )(s) =(cid:16) ∞Xj=−∞ Sc,D 2 Φ2−j ∗ f (s)2(cid:17) 1 2djZB(s,2−j ) Φ2−j ∗ f (t)2dt(cid:17) 1 2 . (0.6) and (0.7) (0.8) (0.9) (0.10) (0.12) Here B(s, r) denotes the ball of Rd with center s and radius r. To prove that these discrete square functions also describe our Hardy spaces, we need to impose the following condition on the previous Schwartz function Φ, which is stronger than (0.5): ∀ ξ ∈ Rd \ {0} ∃ 0 < 2a ≤ b < ∞ s.t. bΦ(εξ) 6= 0, ∀ ε ∈ (a, b]. The following is the discrete version of Lemma 0.1: Lemma 0.2. Let 1 ≤ p < ∞ and f ∈ L1(M; Rc if Gc,D Φ (f ) ∈ Lp(N ) if and only if Sc,D kGc,D d) + L∞(M; Rc Φ (f ) ∈ Lp(N ). Moreover, Φ (f )kp ≈ kSc,D Φ (f )kp ≈ kfkHc p with the relevant constants depending only on p, d and Φ. d). Then f ∈ Hc p(Rd,M) if and only Finally, let us give some easy facts on operator-valued functions. The first one is the following Cauchy-Schwarz type inequality for the operator-valued square function, where φ : Rd → C and f : Rd → L1(M) + M are functions such that all integrations of the above inequality make sense. We also require the operator-valued version of the Plancherel formula. For sufficiently nice functions f : Rd → L1(M) + M, for example, for f ∈ L2(Rd) ⊗ L2(M), we have (0.11) Given two nice functions f and g, the polarized version of the above equality is (cid:12)(cid:12)ZRd ≤ZRd φ(s)2dsZRd f (s)2ds, φ(s)f (s)ds(cid:12)(cid:12)2 ZRd f (s)2ds =ZRd bf (ξ)2dξ. f (s)g∗(s)ds =ZRd bf (ξ)bg(ξ)∗dξ. ZRd The paper is organized as follows. In the next section, we give the definitions of operator-valued local Hardy and bmo spaces. Section 2 is devoted to the proofs of duality results, including the 6 R. Xia and X. Xiong h1-bmo duality and the hp-hq duality for 1 < p < 2 and 1 q = 1. Section 3 gives the results of interpolation. In section 4, we develop Calder´on-Zymund theory that is suitable for our local p(Rd,M), and then version of Hardy spaces. In section 5, we prove general characterizations of hc connect the local Hardy spaces hc p(Rd,M). In the last section of this paper, we give the atomic decomposition of hc p(Rd,M) with Mei's non-local Hardy spaces Hc p + 1 1(Rd,M). 1. Operator-valued local Hardy spaces 1.1. Operator-valued local Hardy spaces. In this subsection, we give the definition of operator- valued local Hardy spaces as well as some basic facts of them. Let f ∈ L1(M; Rc d) + L∞(M; Rc d) (recalling that the Hilbert space Rd is defined by (0.1)). Then the Poisson integral of f is well- defined and takes values in L1(M) +M. Now we define the local analogue of the Lusin area square function of f by sc(f )(s) =(cid:16)ZeΓ(cid:12)(cid:12) ∂ where eΓ is the truncated cone {(t, ε) ∈ Rd+1 + : t < ε} and the strip S ⊂ Rd+1 {(t, ε) ∈ Rd+1 ∂ε εd−1(cid:17) 1 Pε(f )(s + t)(cid:12)(cid:12)2 dtdε : t < ε < 1}. + + defined by: 2 S = {(s, ε) : s ∈ Rd, 0 < ε < 1}. For 1 ≤ p < ∞ define the column local Hardy space hc hc p(Rd,M) = {f ∈ L1(M; Rc where the hc p(Rd,M)-norm of f is defined by p(Rd,M) to be d) : kfkhc d) + L∞(M; Rc p < ∞}, , s ∈ Rd, It is the intersection of the cone The row local Hardy space hr p = kf∗khc with the norm kfkhr p = ksc(f )kLp(N ) + kP ∗ fkLp(N ). kfkhc p(Rd,M) is the space of all f such that f∗ ∈ hc p. Moreover, define the mixture space hp(Rd,M) as follows: p(Rd,M), equipped hp(Rd,M) = hc p(Rd,M) + hr p(Rd,M) for 1 ≤ p ≤ 2 equipped with the sum norm kfkhp = inf{kgkhc and p + khkhr p : f = g + h, g ∈ hc p(Rd,M), h ∈ hr p(Rd,M)}, hp(Rd,M) = hc equipped with the intersection norm p(Rd,M) ∩ hr p(Rd,M) for 2 < p < ∞ The local analogue of the Littlewood-Paley g-function of f is defined by kfkhp = max{kfkhc p ,kfkhr p}. gc(f )(s) =(cid:16)Z 1 0 ∂ ∂ε Pε(f )(s)2εdε(cid:17) 1 2 , s ∈ Rd. ksc(f )kp + kP ∗ fkp ≈ kgc(f )kp + kP ∗ fkp We will see in section 5 that for all 1 ≤ p < ∞. In the following, we give some easy facts that will be frequently used later. Firstly, we have (1.1) Indeed, by (0.11), we have ksc(f )k2 2 + kP ∗ fk2 2 ≈ kfk2 2. Pε(f )(s)(cid:12)(cid:12)2 ∂ε ZRd(cid:12)(cid:12) ∂ Pε(f )(s)(cid:12)(cid:12)2 ds =ZRd(cid:12)(cid:12) =ZRd 4ZRd \∂ ∂ε bf (ξ)2dξ Pε(ξ)(cid:12)(cid:12)2 4π2ξ2bf (ξ)2e−4πεξdξ. (1 − e−4πξ − 4πξe−4πξ)bf (ξ)2dξ. 1 εdεds = ZRdZ 1 0 (cid:12)(cid:12) ∂ ∂ε Then Operator-valued local Hardy spaces 7 Therefore ksc(f )k2 = ds ds ∂ε ∂ε ∂ε cd 4 εdεds εd−1 εd−1 2 = τZRdZeΓ(cid:12)(cid:12) ∂ Pε(f )(s + t)(cid:12)(cid:12)2 dεdt = τZRdZ 1 0 ZB(s,ε)(cid:12)(cid:12) ∂ Pε(f )(t)(cid:12)(cid:12)2 dεdt = cd τZRdZ 1 Pε(f )(s)(cid:12)(cid:12)2 0 (cid:12)(cid:12) ∂ τZRd (1 − e−4πξ − 4πξe−4πξ)bf (ξ)2dξ, 2 = τZRd 2 = τZRd e−4πξbf (ξ)2dξ. (1 − 4πξe−4πξ)bf (ξ)2dξ 2 + kP ∗ fk2 where cd is the volume of the unit ball in Rd. Meanwhile, kP ∗ fk2 Then we deduce (1.1) from the equality 4 cd ksc(f )k2 and the fact that 0 ≤ 4πξe−4πξ ≤ 1 kfkhr 2(Rd,M) ≈ kf∗k2 = kfk2. Then we have e for every ξ ∈ Rd. Passing to adjoint, (1.1) also tells us that (1.2) with equivalent norms. hc 2(Rd,M) = hr 2(Rd,M) = L2(N ) Next, if we apply (0.12) instead of (0.11) in the above proof, we get the following polarized version of (1.1), P ∗ f (s)(I(P) ∗ g(s))∗ds. Pε(g)∗(s + t) dtdε εd−1 ds ZRd (1.3) Pε(f )(s) Pε(g)∗(s)ε dεds ∂ ∂ε ∂ ∂ε 0 f (s)g∗(s)ds = 4ZRdZ 1 +ZRd cdZRdZZeΓ +ZRd = 4 P ∗ f (s)(P ∗ g(s))∗ds + 4πZRd P ∗ f (s)(P ∗ g(s))∗ds + 4πZRd Pε(f )(s + t) ∂ ∂ε ∂ ∂ε P ∗ f (s)(I(P) ∗ g(s))∗ds d) (recalling that I is the Riesz potential of order 1). d) + L∞(M; Rc for nice f , g ∈ L1(M; Rc 1.2. Operator-valued bmo spaces. Now we introduce the noncommutative analogue of bmo spaces defined in [12]. For any cube Q ⊂ Rd, we denote its center by cQ, its side length by l(Q), and its volume by Q. Let f ∈ L∞(M; Rc d). The mean value of f over Q is denoted by fQ := 1 QRQ f (s)ds. We set (1.4) Then we define 1 kfkbmoc = maxn sup Q<1(cid:13)(cid:13)( QZQ f − fQ2dt) bmoc(Rd,M) = {f ∈ L∞(M; Rc 1 2(cid:13)(cid:13)M , sup Q=1(cid:13)(cid:13)(ZQ f2dt) 1 2(cid:13)(cid:13)Mo. d) : kfkbmoc < ∞}. Respectively, define bmor(Rd,M) to be the space of all f ∈ L∞(M; Rr d) such that kf∗kbmoc < ∞ with the norm kfkbmor = kf∗kbmoc . And bmo(Rd,M) is defined as the intersection of these two spaces equipped with the norm bmo(Rd,M) = bmoc(Rd,M) ∩ bmor(Rd,M) kfkbmo = max{kfkbmoc ,kfkbmor}. 8 R. Xia and X. Xiong Remark 1.1. Let Q be a cube with volume kd ≤ Q < (k + 1)d for some positive integer k. Then Q can be covered by at most (k + 1)d cubes with volume 1, say Qj's. Evidently, Hence, 1 (k+1)d QZQ f2dt ≤ k−dZQ f2dt ≤ k−d Q≥1(cid:13)(cid:13)( QZQ f2dt) Xj=1 ZQj f2dt. Q=1(cid:13)(cid:13)(ZQ f2dt) 2(cid:13)(cid:13)M 2(cid:13)(cid:13)M ≤ 2 2 sup sup 1 1 1 d . Thus, if we replace the second supremum in (1.4) over all cubes of volume one by that over all cubes of volume not less than one, we get an equivalent norm of bmoc(Rd,M). Proposition 1.2. Let f ∈ bmoc(Rd,M). Then kfkL∞(M;Rc d) . kfkbmoc . kfk2 L∞(M;Rc f (t)2 1 + td+1 Moreover, bmo(Rd,M), bmoc(Rd,M) and bmor(Rd,M) are Banach spaces. Proof. Let Q0 be the cube centered at the origin with side length 1 and Qm = Q0 + m for each m ∈ Zd. For f ∈ L∞(M; Rc d), d) =(cid:13)(cid:13)(cid:13)ZRd . Xm∈Zd(cid:13)(cid:13)(cid:13) dt(cid:13)(cid:13)(cid:13)M ≤ Xm∈Zd(cid:13)(cid:13)(cid:13)ZQm 1 + md+1ZQm f (t)2dt(cid:13)(cid:13)(cid:13)M bmoc Xm∈Zd 1 + md+1 . kfk2 It is then easy to check that bmoc(Rd,M) is a Banach space. Proposition 1.3. We have the inclusion bmoc(Rd,M) ⊂ BMOc(Rd,M). More precisely, there exists a constant C depending only on the dimension d, such that for any f ∈ bmoc(Rd,M), (1.5) f (t)2 1 + td+1 dt(cid:13)(cid:13)(cid:13)M . kfk2 bmoc . (cid:3) 1 1 kfkBMOc ≤ Ckfkbmoc . 1 for Q ≥ 1. By the triangle inequality and (0.10), we have Proof. By virtue of Remark 1.1, it suffices to compare the term(cid:13)(cid:13)(cid:13)( 1 QRQ f2dt) (cid:13)(cid:13)(cid:13)( 1 QRQ f − fQ2dt) QZQ f2dt) (cid:13)(cid:13)(cid:13)( 2(cid:13)(cid:13)(cid:13)M 2(cid:13)(cid:13)(cid:13)M ≤(cid:13)(cid:13)(cid:13)( QZQ f2dt) ≤ 2(cid:13)(cid:13)(cid:13)( 2(cid:13)(cid:13)(cid:13)M 2(cid:13)(cid:13)(cid:13)M QZQ f − fQ2dt) which leads immediately to (1.5). 2(cid:13)(cid:13)(cid:13)M + kfQkM 1 1 1 , 1 1 1 1 Classically, BMO functions are related to Carleson measures (see [11]). A similar relation still holds in the present noncommutative local setting. We say that an M-valued measure dλ on the strip S = Rd × (0, 1) is a Carleson measure if and the term (cid:3) N (λ) = sup Q<1{ 1 Q(cid:13)(cid:13)ZT (Q) dλ(cid:13)(cid:13)M where T (Q) = Q × (0, l(Q)]. Lemma 1.4. Let g ∈ bmoc(Rd,M). Then dλg = ∂ measure on the strip S and : Q ⊂ Rd cube } < ∞, ∂ε Pε(g)(s)2ε dsdε is an M-valued Carleson max{N (λg) 1 2 , kP ∗ gkL∞(N )} . kgkbmoc . Operator-valued local Hardy spaces 9 Proof. Given a cube Q with Q < 1, denote by 2Q the cube with the same center and twice the side length of Q. We decompose g = g1 + g2 + g3, where g1 = (g − g2Q)12Q and g2 = (g − g2Q)1Rd\2Q. ∂ε Pε(s)ds = 0 for any ε > 0, we have ∂ ∂ε Pε(g) = ∂ ∂ε Pε(g1) + ∂ ∂ε Pε(g2). By (0.10), SinceR ∂ We first deal with N (λg1 ). By (0.11) and (1.5), we have N (λg) ≤ 2(N (λg1 ) + N (λg2 )). ZT (Q)(cid:12)(cid:12) ∂ ∂ε Pε(g1)(s)(cid:12)(cid:12)2 2 0 0 ∂ε εdsdε \∂ ∂ε (cid:12)(cid:12) ∂ (cid:12)(cid:12)(cid:12) Pε(g1)(s)(cid:12)(cid:12)2 Pε(ξ)(cid:12)(cid:12)(cid:12) εdsdε ≤ZRdZ ∞ =ZRdZ ∞ .ZRd g1(s)2ds =Z2Q g − g2Q2ds . Qkgk2 ∂ε Pε(s)(cid:12)(cid:12) . Pε(g2)(s)(cid:12)(cid:12)2 (ε+s)d+1 , applying (0.10), we obtain 1 bg1(ξ)2εdεds g(t) − g2Q2 (ε + s − t)d+1 εZRd\2Q dt. . 1 bmoc . Since(cid:12)(cid:12) ∂ ∂ε (cid:12)(cid:12) ∂ bmoc . Thus, N (λg1 ) . kgk2 The integral on the right hand side of the above inequality can be treated by a standard argument as follows: for any (s, ε) ∈ T (Q), g(t) − g2Q2 (ε + s − t)d+1 g(t) − g2Q2 t − cQd+1 ZRd\2Q dt dt .ZRd\2Q .Xk≥1Z2k+1Q\2kQ l(Q)Xk≥1 1 l(Q)kgk2 bmoc , 2−k . . 1 g(t) − g2Q2 t − cQd+1 1 dt 2k+1QZ2k+1Q g(t) − g2Q2dt where cQ is the center of Q. Then, it follows that N (λg2 ) . kgk2 volume one centered at the origin, so Rd = ∪m∈Zd Qm. By (0.10), for any s ∈ Rd, we have Now we deal with the term kP ∗ g(s)kM. Let Qm = Q0 + m be the translate of the cube with bmoc. kP ∗ g(s)kM =(cid:13)(cid:13)Xm ZQm P(t)g(s − t)dt(cid:13)(cid:13)M ≤Xm (cid:0)ZQm P(t)2dt) m∈Zd k(ZQm g(s − t)2dt(cid:1) 1 QZQ g(t)2dt) Q=1(cid:13)(cid:13)(cid:13)( 2(cid:13)(cid:13)(cid:13)M 2 · sup . sup . kgkbmoc . 1 1 1 2 kM Thus, kP ∗ gkL∞(N ) = sups∈Rd kP ∗ g(s)kM . kgkbmoc , which completes the proof. (cid:3) Reexamining the above proof, we find that the facts used to prove N (λg) 1 2 . kgkbmoc are • RRd ε ∂ • supξ∈RdR ∞ • (cid:12)(cid:12)ε ∂ ∂ε Pε(s)(cid:12)(cid:12) . 0 ∂ε Pε(s)ds = 0 for ∀ ε < 0; ε < ∞; dε 2 ∂ε Pε(ξ)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)\ε ∂ (ε+s)d+1 . ε We can easily check that if we replace ε ∂ ∂ε Pε above by Ψε = 1 εd Ψ( ·ε ), where Ψ is a Schwartz function such that bΨ(0) = 0, the corresponding three conditions still hold. On the other hand, the only fact used for proving the inequality kP ∗ gkL∞(N ) . kgkbmoc is that (ZQm P(t)2dt) Xm 1 2 < ∞. 10 R. Xia and X. Xiong Recall that H σ J σ(f ) ∈ L2(Rd). It is equipped with the norm kfkHσ 2 (Rd) denotes the potential Sobolev space, consisting of distributions f such that 2 (Rd) = kJ σfkL2(Rd). If ψ is a function on Rd such that bψ ∈ H σ 2 (Rd) for some σ > d 2 , we have Xm (cid:0)ZQm ψ(s)2dt(cid:1) 1 2 .(cid:0)Xm 1 (1 + m2)σ(cid:1) 1 2(cid:0)ZRd (1 + s2)σψ(s)2ds(cid:1) 1 . 2 . kbψkHσ 2 Based on the above observation, we have the following generalization of Lemma 1.4: Lemma 1.5. Let ψ be the (inverse) Fourier transform of a function in H σ Schwartz function such that bΨ(0) = 0. If g ∈ bmoc(Rd,M), then dµg = Ψε ∗ g(s)2 dεds M-valued Carleson measure on the strip S and (1.6) ε 1 max{N (µg) 2 , kψ ∗ gkL∞(N )} . kgkbmoc . 2 (Rd), and Ψ be a is an In particular, (1.7) max{N (µg) 1 2 , kJ(P) ∗ gkL∞(N )} . kgkbmoc . 1 Proof. (1.6) follows from the above discussion; (1.7) is ensured by (1.6) and the fact that (1 + ξ2) (cid:3) Remark 1.6. We will see in the next section that the converse inequality of (1.7) also holds. 2 (Rd), which can be checked by a direct computation. 2 e−2πξ ∈ H σ 2. The dual space of hc p for 1 ≤ p < 2 In this section, we describe the dual of hc p(Rd,M) for 1 ≤ p < 2 as bmo type spaces. We call these spaces bmoc q(Rd,M) (with q the conjugate index of p). The argument used here is modelled on q(Rd,M). However, the one used in [12] when studying the duality between Hc due to the truncation of the square functions, some highly non-trivial modifications are needed. q(Rd,M) to be the space of all f ∈ q. Let 2 < q ≤ ∞. We define bmoc p(Rd,M) and BMOc 2.1. Definition of bmoc Lq(M; Rc d) such that kfkbmoc q =(cid:16)(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 1 QZQ f (t) − fQ2dt(cid:13)(cid:13)(cid:13) q 2 q 2 +(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q=1 1 QZQ f (t)2dt(cid:13)(cid:13)(cid:13) q q 2 2(cid:17) 1 q < ∞. q(Rd,M) coincides with the space bmoc(Rd,M) introduced in the previous section. If q = ∞, bmoc is just an intuitive notation since the pointwise supremum does not make any sense in the noncommutative setting. This is the norm of the Banach space L q 2 Note that the norm k sup+ (N ; ℓ∞); we refer to [36, 18, 22] for more information. If 1 ≤ p < ∞ and (ai)i∈Z is a sequence of positive elements in Lp(N ), it has been proved by i aik q 2 Junge (see [18], Remark 3.7) that (2.1) k sup i +aikp = sup(cid:8)Xi∈Z τ (aibi) : bi ∈ Lq(N ), bi ≥ 0, kXi∈Z bikq ≤ 1(cid:9). It is also known that a positive sequence (xi)i belongs to Lp(N ; ℓ∞) if and only if there is an a ∈ Lp(N ) such that xi ≤ a for all i, and moreover, k(xi)kLp(N ;ℓ∞) = inf{kakp : a ∈ Lp(N ), xi ≤ a,∀ i}. 2 1 and (2.2) (N ) s.t. ∃ a ∈ L q Then we get the following fact (which can be taken as an equivalent definition): f ∈ bmoc if and only if QZQ f (t) − fQ2dt ≤ a(s), ∀ s ∈ Q and ∀ Q ⊂ Rd with Q < 1 QZQ f (t)2dt ≤ b(s), ∀ s ∈ Q and ∀ Q ⊂ Rd with Q = 1. q : a, b as in (2.2) and (2.3) respectively(cid:9). 2(cid:1) 1 q = inf(cid:8)(cid:0)kak If this is the case, then ∃ b ∈ L q (N ) s.t. kfkbmoc + kbk (2.3) q 2 q 2 1 q 2 q 2 q(Rd,M) Operator-valued local Hardy spaces 11 In fact, the cubes considered in the definition of bmoc q(Rd,M) can be reduced to cubes with dyadic lengths. Let Qk 1 Qk Lemma 2.1. If q > 2, then f # k (s) = s denote the cube centered at s and with side length 2−k, k ∈ Z. Set sZQk sZQ0 and f #(s) = 1 Q0 dt. dt s(cid:12)(cid:12)f (t)(cid:12)(cid:12)2 s(cid:12)(cid:12)2 s(cid:12)(cid:12)f (t) − fQk (cid:16)k sup +f # k k q 2 q 2 + kf #k q q 2 q 2(cid:17) 1 gives an equivalent norm in bmoc k≥1 q(Rd,M). Proof. It is obvious from the definition that +f # k k k sup k≥1 1 2 q 2 ≤ kfkbmoc q and kf #k 1 2 q 2 ≤ kfkbmoc q . We notice that for any cube Q with Q < 1 and s ∈ Q, there exists k ≥ −1 such that Q ⊂ Qk Qk s ≤ 4dQ. Thus s and 1 4d(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 1 2 q 2 1 QZQ f (t) − fQ2dt(cid:13)(cid:13)(cid:13) 4d(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q=1 1 1 QZQ f (t)2dt(cid:13)(cid:13)(cid:13) Similarly, Thus the lemma is proved. . k sup k≥−1 +f # k k 1 2 q 2 . 2dk sup k≥1 +f # k k 1 2 q 2 . 1 2 2 ≤ 2dkf #k q 1 2 q 2 . (cid:3) From the proofs of Proposition 1.2 and Lemma 1.4, we can easily see that their q-analogues still hold in the present setting. We leave the proofs to the reader. Proposition 2.2. Let q > 2 and f ∈ bmoc q(Rd,M). Then d) . kfkbmoc q kfkLq(M;Rc . Lemma 2.3. Let f ∈ bmoc respectively. Then dλf is a q-Carleson measure in the following sense: q(Rd,M) and assume that the operators a and b satisfy (2.2) and (2.3) 1 QZT (Q) ∂ ∂ε Pε(f )(t)2εdtdε . a(s), ∀ s ∈ Q and ∀ Q ⊂ Rd with Q < 1. Moreover, ψ ∗ f (s)2 . b(s) for any s ∈ Rd, if ψ is the (inverse) Fourier transform of a function in H σ 2 (Rd). + εd+1 . For any 1 ≤ p < ∞, we will embed hc : s < p(Rd,M) into a larger space 2.2. A bounded map. In the sequel, we equip the truncated cone eΓ = {(s, ε) ∈ Rd+1 ε < 1} with the measure dtdε Lp(cid:0)N ; Lc Lp(cid:0)N ; Lc for f ∈ Lp(cid:0)N ; Lc 2(eΓ)(cid:1) ⊕p Lp(N ). Here Lp(cid:0)N ; Lc 2(eΓ)(cid:1) ⊕p Lp(N ) is the ℓp-direct sum of the Banach spaces 2(eΓ)(cid:1) and Lp(N ), equipped with the norm Lp(N )(cid:17) 1 k(f, g)k =(cid:16)kfkp Lp(cid:0)N ;Lc p(Rd,M) to Lp(cid:0)N ; Lc 2(eΓ)(cid:1) ⊕p Lp(N ) by Pε(f )(s + t), P ∗ f (s)(cid:1), 2(eΓ)(cid:1) and g ∈ Lp(N ), with the usual modification for p = ∞. E(f )(s, t, ε) =(cid:0)ε Definition 2.4. We define a map E from hc 2(eΓ)(cid:1) + kgkp ∂ ∂ε p and a map F for sufficiently nice h = (h′, h′′) ∈ Lp(cid:0)N ; Lc dtdε 2(eΓ)(cid:1) ⊕p Lp(N ) by εd + h′′(s)(P + 4πI(P))(s − u)ids . F(h)(u) =ZRdh 4 cdZZeΓ h′(s, t, ε) ∂ ∂ε Pε(s + t − u) 12 R. Xia and X. Xiong By definition, the map E embeds hc following results, Theorems 2.8 and 2.18 show that by identifying hc Lp(cid:0)N ; Lc 2(eΓ)(cid:1) ⊕p Lp(N ) via E, hc 1 < p < ∞ by virtue of the map F. Proposition 2.5. Let 1 ≤ p < ∞. Then for any nice f ∈ L1(M; Rc p(Rd,M) isometrically into Lp(cid:0)N ; Lc p(Rd,M) is complemented in Lp(cid:0)N ; Lc p(Rd,M) as a subspace of 2(eΓ)(cid:1) ⊕p Lp(N ). The 2(eΓ)(cid:1) ⊕p Lp(N ) for every d) + L∞(M; Rc d), we have F(E(f )) = f. Proof. Applying (1.3), we get, for any nice function g, ZRd ∂ ∂ε ∂ ∂ε dtdε εd−1 Pε(f )(s + t) Pε(s + t − u) cdZZeΓ cdZZeΓ F(E(f ))(u)g(u)du =ZRdh 4 + P ∗ f (s)Z (P(s − u) + 4πI(P)(s − u))g(u)duids =ZRdh 4 + P ∗ f (s)(P ∗ g + 4πI(P) ∗ g)(s)ids =ZRd Pε(f )(s + t) Pε(g)(s + t) f (u)g(u)du , dtdε εd−1 ∂ ∂ε ∂ ∂ε g(u)du which completes the proof. (cid:3) The following dyadic covering lemma is known. Tao Mei [30] proved this lemma for the d-torus and also for the real line. For the case Rd with d > 1, we refer the interested readers to [6, 16] for more details. In the following, we give a sketch of the way how we choose the dyadic covering. Lemma 2.6. There exist a constant C > 0, depending only on d, and d + 1 dyadic increasing filtrations Di = {Di j}j∈Z of σ-algebras on Rd for 0 ≤ i ≤ d, such that for any cube Q ⊂ Rd, there is a cube Di j satisfying Q ⊂ Di m,j ∈ Di Proof. Let {αi}d i=0 be a sequence in the interval (0, 1) such that m,j ≤ CQ. m,j and Di min i6=i′ αi − αi′ > 0. Then we define (2.4) αi j = αi, αi + 1 αi − 1 j ≥ 0, 3 (2−j − 1), 3 (2−j + 1), j < 0 and − j even, j < 0 and − j odd. j + (m1 + 1)2−j] × ··· × (αi j + md2−j, αi j + (md + 1)2−j], The σ-algebra Di j is generated by the cubes Di m,j = (αi j + m12−j, αi for all m = (m1,··· , md) ∈ Zd. cube Di m,j such that Q ⊂ Di For any cube Q ⊂ Rd, there exist a constant C, depending only on {αi}d i=0 and d, and a dyadic (cid:3) m,j and Di m,j ≤ CQ. To show the boundedness of the map F, we need the following assertion by Mei, see [30, Propo- sition 3.2]; we include a proof for this lemma, since the one in [30] is the one dimensional case. Let 1 ≤ p < ∞, and f ∈ Lp(N ) be a positive function. Let Q be a cube centered at the origin, and denote Qt = t + Q. Then we define f Q(t) = f (s)ds. 1 QZQt Lemma 2.7. Let 1 ≤ p < ∞ and let (fk)k∈Z be a positive sequence in Lp(N ) and (Qk)k∈Z be a sequence of cubes centered at the origin. Then (fk)Qk kXk∈Z kp . kXk∈Z fkkp. Operator-valued local Hardy spaces 13 m,jk−1, which has twice the side length of Di Proof. Similarly to the proof of [30, Proposition 3.2], we are going to apply [18, Theorem 0.1] for noncommutative martingales. By Lemma 2.6, we can cover every Qk by some Di , and thus by some Di m,jk−1 ≤ CQk. Obviously, t + Qk is still covered by t + Di jk−1. Let us adjust the translation vector t = (t1, ..., td) as follows. Write Qk = (−a, a] × ... × (−a, a] and m,jk−1 = (b1, b2] × ... × (b1, b2], then either b2 − a ≥ 2−jk or −a − b1 ≥ 2−jk . Without loss of Di generality, we can assume b2 − a ≥ 2−jk . Now setet = (et1, ...,etd) withetj the largest real number in the set 2−jk Z less than tj. Then we can check that t + Qk is covered byet + Di m,jk−1, but the later is not necessary a dyadic cube in Di m,jk−1 and that the later is a dyadic cube. Thus, . Moreover, Di m′,jk m′,jk (fk)Qk ≤ C X0≤i≤d E(fkDi jk ), where E(·Di from [18, Theorem 0.1]. j ) denotes the conditional expectation with respect to Di j. Then the lemma follows (cid:3) to bmoc p(Rd,M). Theorem 2.8. For 2 < p ≤ ∞, the map F extends to a bounded map from Lp(cid:0)N ; Lc Proof. We have to show that for any h = (h′, h′′) ∈ Lp(cid:0)N ; Lc Fix h = (h′, h′′) ∈ Lp(cid:0)N ; Lc p-norm of F(h). For v ∈ Rd and k ∈ N, denote by Qk 2(eΓ)(cid:1)⊕p Lp(N ) and set ϕ = F(h). We will apply Lemma 2.1 to estimate 2(eΓ)(cid:1) ⊕p Lp(N ), 2(eΓ)(cid:1)⊕p Lp(N ) the bmoc length 2−k, then we have Qk . khkLp(cid:0)N ;Lc 2(eΓ)(cid:1)⊕pLp(N ) v the cube centered at v and with side kF(h)kbmoc v = v + Qk 0. We set . p h′1(s, t, ε) = h′(s, t, ε)1Qk−1 v (s), h′2(s, t, ε) = h′(s, t, ε)1(Qk−1 v )c(s) and Let ϕ# k (v) = vZQk 1 Qk v(cid:12)(cid:12)2 v(cid:12)(cid:12)ϕ(u) − ϕQk du. with ( ∂ ∂ε Pε)Qk 0 (s, t, v) = 1 Qk 1 Qk 1 Qk + v ( ∂ v ds BQk ∂ ∂ε Pε)Qk dtdε εd 0 (s, t, v)h′2(s, t, ε) 0 (v) =ZRdZZeΓ vRQk ∂ε Pε(s + t − u)du. Then, we have vZQk ϕ(u) − BQk 0 (v)2du vZQk v(cid:12)(cid:12)(cid:12)Z(Qk−1 )cZZeΓ h′2(s, t, ε)(cid:2) ∂ vZQk ZZeΓ v(cid:12)(cid:12)(cid:12)ZQk−1 vZQk v(cid:12)(cid:12)(cid:12)ZRd h′′(s)[P(s − u) + 4πI(P)(s − u)]ds(cid:12)(cid:12)(cid:12) v and (t, ε) ∈ eΓ, we have s + t − u + ε ≈ s − v + ε with uniform Pε(s + t − u) − ( dtdε εd 0 (s, t, v)(cid:3) dtdε Pε(s + t − u) )c, u ∈ Qk ds(cid:12)(cid:12)(cid:12) ds(cid:12)(cid:12)(cid:12) h′1(s, t, ε) Pε)Qk ∂ ∂ε ∂ ∂ε du. du du ∂ε εd 2 2 2 v v ϕ# k (v) . . 1 Qk 1 Qk When s ∈ (Qk−1 constants. Then, + v ∂ ∂ε ZZeΓ(cid:12)(cid:12)(cid:12) .ZZeΓ(cid:16) = cdZ 1 0 Pε(s + t − u) − ( ∂ ∂ε Pε)Qk 2−k (s + t − u + ε)d+2(cid:17)2 dtdε 0 (s, t, v)(cid:12)(cid:12)(cid:12) εd−1 .Z 1 2 dtdε εd−1 0 ZB(0,ε) 2−2kε (s − v2 + ε2)d+2 dε . 2−2k . s − v2d+2 2−2k (s − v2 + ε2)d+2 dt dε εd−1 14 R. Xia and X. Xiong Let (ak)k∈N be a positive sequence such that kPk≥1 akk( p index of r. Let 2 )′ ≤ 1, where r′ denotes the conjugate A =Xk≥1 B =Xk≥1 C =Xk≥1 v v v )c )c 1 2−2k ∂ ∂ε h′1(s, t, ε) s − vd+1 1 Qk 1 Qk τZRdZ(Qk−1 ds ·Z(Qk−1 vZQk ZZeΓ v(cid:12)(cid:12)ZQk−1 τZRd τZRd v(cid:12)(cid:12)ZRd vZQk h′′(s)[P(s − u) + 4πI(P)(s − u)]ds(cid:12)(cid:12)2 τZ ϕ# Xk≥1 k (v)ak(v)dv . A + B + C. Pε(s + t − u) dtdε εd+1 Then, s − vd+1ZZeΓ h′2(s, t, ε)2 dtdε εd+1 ds · ak(v)dv du · ak(v)dv ds(cid:12)(cid:12)2 du · ak(v)dv. First, we estimate the term A. Applying the Fubini theorem and the Holder inequality, we arrive at s s ds ak(v)dv εd+1 . kh′k2 τZRd 2−kZ(Qk−1 2−kZ(Qk−1 )c v − s−d−1ZZeΓ h′2(s, t, ε)2 dtdε A .Xk≥1 ≤(cid:13)(cid:13)(cid:13)ZZeΓ h′2(·, t, ε)2 dtdε εd+1(cid:13)(cid:13)(cid:13) p 2 ·(cid:13)(cid:13)(cid:13)Xk≥1 2−kXj≤kZQj−2 2(eΓ)) ·(cid:13)(cid:13)(cid:13)Xk≥1 2(j−1)dZQj−2 Lp(N ;Lc \Qj−1 \Qj−1 2j−k−1ak(v)dv(cid:13)(cid:13)(cid:13)( p (cid:13)(cid:13)(cid:13)Xk≥1Xj≤k 2 )′ 2 )′ )c v − s−d−1ak(v)dv(cid:13)(cid:13)(cid:13)( p 2(j−1)(d+1)ak(v)dv(cid:13)(cid:13)(cid:13)( p .(cid:13)(cid:13)(cid:13)Xj∈ZXk≥j 2j−k−1ak(cid:13)(cid:13)(cid:13)( p .(cid:13)(cid:13)Xk≥1 ak(cid:13)(cid:13)( p 2 )′ ≤ 1. k≥1 . s s s s 2 )′ 2 )′ Then we move to the estimate of B: Here and in the context below, k·k( p Now we apply Lemma 2.7 to estimate the second factor of the last term: 2 )′ is the norm of L( p 2 )′ (N ) with respect to the variable s ∈ Rd. Since hc 2.7, we get 2(Rd,M) = L2(N ) with equivalent norms, by the Cauchy-Schwarz inequality and Lemma dtdε εd Pε(s + t − u) ∂ ∂ε Pε(f )∗(s + t) dudv 2 dv. 2 ds(cid:12)(cid:12)(cid:12) ds(cid:12)(cid:12)(cid:12) dtdε εd B ≤Xk≥1ZRd ≤Xk≥1ZRd v v 1 2 k (v) h′1(s, t, ε)a 1 2 k (v) ∂ ∂ε 2kd sup h′1(s, t, ε)a ZZeΓ 2kdτZRd(cid:12)(cid:12)(cid:12)ZQk−1 ZZeΓ kfk2=1(cid:12)(cid:12)(cid:12)τZQk−1 2kdτZQk−1 τZRdZZeΓ h′1(s, t, ε)2 dtdε 2kdZQk−1 2(eΓ)(cid:1)(cid:13)(cid:13)(cid:13)Xk≥1 Lp(cid:0)N ;Lc 2(eΓ)(cid:1)(cid:13)(cid:13)Xk≥1 ak(cid:13)(cid:13)( p Lp(cid:0)N ;Lc B ≤Xk≥1ZRd .Xk≥1 ≤ kh′k2 ≤ 2dkh′k2 εd+1 ZZeΓ h′1(s, t, ε)2 dtdε εd+1 2kdZQk−1 ak(v)dv(cid:13)(cid:13)(cid:13)( p 2 )′ ≤ 2dkh′k2 v s s ds ak(v)dv · kfkhc 2 ak(v)dvds 2 )′ Lp(cid:0)N ;Lc 2(eΓ)(cid:1). Operator-valued local Hardy spaces 15 dt, which is relatively easy. For The techniques used to estimate the term C are similar to that of B: v s 2 2 , ak(v)dvdu (cid:13)(cid:13)(cid:13)(cid:12)(cid:12)ZRd Take f ∈ Lp′ (N ) with norm one such that (cid:12)(cid:12)ZRd ak(v)dv(cid:13)(cid:13)(cid:13)( p h′′(s)[P(s − ·) + 4πI(P)(s − ·)]ds(cid:12)(cid:12)2(cid:13)(cid:13)(cid:13) p h′′(s)[P(s − u) + 4πI(P)(s − u)]ds(cid:12)(cid:12)2 2 )′(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)ZRd τZRd 2kdZQk−1 C =Xk≥1 2kdZQk−1 ≤(cid:13)(cid:13)(cid:13)Xk≥1 .(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)ZRd h′′(s)[P(s − ·) + 4πI(P)(s − ·)]ds(cid:12)(cid:12)2(cid:13)(cid:13)(cid:13) p =(cid:12)(cid:12)(cid:12)τZRd h′′(s)[P(s − u) + 4πI(P)(s − u)]ds(cid:12)(cid:12)2(cid:13)(cid:13)(cid:13) p (cid:12)(cid:12)(cid:12)τZRd h′′(s)[P ∗ f (s) + 4πI(P) ∗ f (s)]ds(cid:12)(cid:12)(cid:12) ≤ kh′′k2 pkP ∗ f + 4πI(P) ∗ fk2 . kh′′k2 pkfk2 Combining the estimates of A, B and C with (2.1), we obtain Lp(cid:0)N ;Lc 2(eΓ)(cid:1)⊕pLp(N ) s(cid:12)(cid:12)ϕ(t)(cid:12)(cid:12)2 sRQ0 any positive operator a such that kakL( p -norm of ϕ#(s) = 1 Q0 It remains to establish the L p 2 )′ (N ) ≤ 1, we have p′ = kh′′k2 p. . khk2 k sup k≥1 h′′(s)[P ∗ f (s) + 4πI(P) ∗ f (s)]ds(cid:12)(cid:12)(cid:12) k k p +ϕ# Then p′ 2 . 2 2 2 2 . The terms B′ and C′ are treated in the same way as B and C respectively. The results are ∂ ∂ε dtdε εd+1 h′(s, t, ε) def= B′ + C′. Pε(s + t − u) ds(cid:12)(cid:12)2 h′′(s)[P(s − u) + 4πI(P)(s − u)]ds(cid:12)(cid:12)2 τZ ϕ#(v)a(v)dv . τZRdZQ0 v(cid:12)(cid:12)ZRdZZeΓ + τZRdZQ0 v(cid:12)(cid:12)ZRd B′ ≤ τZRdZZeΓ h′(s, t, ε)2 dtdε εd+1ZQ0 2(eΓ)(cid:1)kak( p 2 )′(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)ZRd C′ ≤(cid:13)(cid:13)(cid:13)ZQ0 h′′(s)[P(s − ·) + 4πI(P)(s − ·)]ds(cid:12)(cid:12)2(cid:13)(cid:13)(cid:13) p a(v)dv(cid:13)(cid:13)(cid:13)( p a(v)dvds ≤ kh′k2 Lp(cid:0)N ;Lc 2 s s 2 )′ , . kh′′k2 p. du · a(v)dv du · a(v)dv So we obtain 2 . kϕ#k p kF(h)kbmoc which proves the theorem. Thus, Lemma 2.1 ensures that Lp(cid:0)N ;Lc . khk2 . khkLp(cid:0)N ;Lc Corollary 2.9. Let 1 ≤ p < 2. For any f ∈ Lp(cid:0)M; Lc . kfkLp(cid:0)M;Lc Proof. To simplify notation, we denote L2(cid:0)Rd, (1 +td+1)dt(cid:1) by Wd. Let q be the conjugate index of p. By duality, we can choose h = (h′, h′′) ∈ Lq(N ; Lc 2(eΓ)(cid:1)⊕pLp(N ) 2(eΓ)(cid:1)⊕pLp(N ) 2(Rd, (1 + td+1)dt)(cid:1), we have 2(Rd,(1+td+1)dt)(cid:1). 2) ⊕q Lq(N ) with norm one such that kfkhc (cid:3) , p p Pε(f )(s + t)h′∗(s, t, ε) ∂ ∂ε ksc(f )kp + kP ∗ fkp =(cid:12)(cid:12)(cid:12)τZRdZZeΓ =(cid:12)(cid:12)τZ f (u)eF(h)∗(u)du(cid:12)(cid:12), dtdε εd ds + τZRd P ∗ f (s)h′′∗(s)ds(cid:12)(cid:12)(cid:12) 16 where (2.5) R. Xia and X. Xiong h′(s, t, ε) ∂ ∂ε Pε(s + t − u) eF(h)(u) =ZRdhZZeΓ dtdε εd + h′′(s)P(s − u)ids . Following the proof of Theorem 2.8, we can easily check thateF is also bounded from Lq(cid:0)N ; Lc q(Rd,M). They applying Proposition 2.2 and Theorem 2.8, we have Lq(N ) to bmoc 2(eΓ)(cid:1)⊕q (cid:12)(cid:12)τZ f (s)eF(h)∗(s)ds(cid:12)(cid:12) . sup q (Rd ,M)≤1(cid:12)(cid:12)τZ f (s)ϕ∗(s)ds(cid:12)(cid:12) )≤1(cid:12)(cid:12)(cid:12)τZ (1 + sd+1)f (s)ϕ∗(s) d d) = kfkLp(M;Wc d). sup = k(1 + sd+1)fkLp(M;Rc kϕkbmoc . kϕkLq (M;Rc ds 1 + sd+1(cid:12)(cid:12)(cid:12) Thus we obtain the desired assertion. (cid:3) 2.3. Duality. Now we are going to present the hc subsection by a lemma which will be very useful in the sequel. p-bmoc q duality for 1 ≤ p < 2. We begin this Lemma 2.10. Let 1 ≤ p ≤ 2 and q be its conjugate index. For f ∈ hc g ∈ bmoc q(Rd,M), p(Rd,M) ∩ L2(N ) and (cid:12)(cid:12)τZRd f (s)g∗(s)ds(cid:12)(cid:12) . kfkhc pkgkbmoc q . Proof. It suffices to prove the lemma for compactly supported (relative to the variable of Rd) f ∈ hc p(Rd,M). We assume that f is sufficiently nice that all calculations below are legitimate. We need two auxiliary square functions. For s ∈ Rd and ε ∈ [0, 1], we define (2.6) (2.7) 2 ∂r sc(f )(s, ε) =(cid:16)Z 1 sc(f )(s, ε) =(cid:16)Z 1 ε ZB(s,r− ε rd−1(cid:17) 1 2 )(cid:12)(cid:12) ∂ Pr(f )(t)(cid:12)(cid:12)2 dtdr ε ZB(s, r rd−1(cid:17) 1 Pr(f )(t)(cid:12)(cid:12)2 dtdr 2 )(cid:12)(cid:12) ∂ ∂r . 2 , Both sc(f )(s, ε) and sc(f )(s, ε) are decreasing in ε and sc(f )(s, 0) = sc(f )(s). In addition, it is clear that sc(f )(s, ε) ≤ sc(f )(s, ε). Let (ei)i∈I be an increasing family of τ -finite projections of M such that ei converges to 1M in the strong operator topology. Then we can approximate sc(f )(s, ε) by sc(eif ei)(s, ε). Thus we can assume that τ is finite; under this finiteness assumption, for any small δ > 0 (which will tend to zero in the end of the proof), consider sc(f )(s, ε) + δ1M instead of sc(f )(s, ε), we can assume that sc(f )(s, ε) is invertible in M for every (s, ε) ∈ S. By (1.3) and the Fubini theorem, we have 0 (cid:12)(cid:12)τZ f (s)g∗(s)ds(cid:12)(cid:12) .(cid:12)(cid:12)(cid:12)τZRdZ 1 +(cid:12)(cid:12)(cid:12)τZRd τZRdZ 1 =(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)τZRd 2d cd ∂ ∂ε ∂ ∂ε Pε(f )(s) Pε(g)∗(s)ε dεds(cid:12)(cid:12)(cid:12) P ∗ f (s)(P ∗ g(s))∗ds + τZRd 0 ZB(s, ε P ∗ f (s)(P ∗ g(s))∗ds + τZRd Pε(f )(t) ∂ ∂ε ∂ ∂ε 2 ) Pε(g)∗(t) dεdt εd−1 P ∗ f (s)(I(P) ∗ g(s))∗ds(cid:12)(cid:12)(cid:12) P ∗ f (s)(I(P) ∗ g(s))∗ds(cid:12)(cid:12)(cid:12). ds(cid:12)(cid:12)(cid:12) Operator-valued local Hardy spaces 17 Then, 2d cd (cid:12)(cid:12)τZ f (s)g∗(s)ds(cid:12)(cid:12) τZRdZ 1 0 ZB(s, ε .(cid:12)(cid:12)(cid:12) P ∗ f (s)(P ∗ g(s))∗ds(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)τZRd +(cid:16)(cid:12)(cid:12)(cid:12)τZRd def= I + II. Pε(f )(t)sc(f )(s, ε) ∂ ∂ε 2 ) p−2 2 sc(f )(s, ε) 2−p 2 ∂ ∂ε Pε(g)∗(t) P ∗ f (s)(I(P) ∗ g(s))∗ds(cid:12)(cid:12)(cid:12)(cid:17) dεdt εd−1 ds(cid:12)(cid:12)(cid:12) The term II is easy to deal with. By the Holder inequality and (1.7), we get Then by [43, Proposition V.3 and Lemma V.3.2] we have II ≤ kP ∗ fkpkP ∗ gkq + kP ∗ fkpkI(P) ∗ gkq. Hence, by Lemma 2.3, kP ∗ gkq . kJ(P) ∗ gkq, and kI(P) ∗ gkq . kJ(P) ∗ gkq. Now we estimate the term I. By the Cauchy-Schwarz inequality II . kgkbmoc qkfkhc p . ∂ ∂ε c2 d 4d I2 ≤ τZRdZ 1 · τZRdZ 1 def= A · B. 0 (cid:16)ZB(s, ε 2 ) 0 (cid:16)ZB(s, ε 2 ) Pε(f )(t)2 dt Pε(g)(t)2 dt εd−1(cid:17)sc(f )(s, ε)p−2dεds εd−1(cid:17)sc(f )(s, ε)2−pdεds ∂ ∂ε Note here that sc(f )(s, ε) is the function of two variables defined by (2.6), which is differentiable in the w∗-sense. We first deal with A. Using sc(f )(s, ε) ≤ sc(f )(s, ε), we have ∂ ∂ε A ≤ τZRdZ 1 0 ZB(s, ε Pε(f )(t)2sc(f )(s, ε)p−2 dεdt 2 ) εd−1 = −τZRdZ 1 0 (cid:0) ∂ sc(f )(s, ε)2(cid:1)sc(f )(s, ε)p−2dεds = −2τZRdZ 1 ∂ε sc(f )(s, ε)p−1 ∂ ∂ε sc(f )(s, ε)dεds. 0 ds A . −τZRd . τZRd sc(f )(s, 0)p−1Z 1 sc(f )(s, 0)pds = kfkp 0 ∂ ∂ε hc p s(f )c(s, ε)dεds . Since 1 ≤ p < 2 and sc(f )(s, ε) is decreasing in ε, sc(f )(s, ε)p−1 ≤ sc(f )(s, 0)p−1. At the same time, ∂ ∂ε sc(f )(s, ε) ≤ 0. Therefore, The estimate of B is harder. For any j ∈ N, we need to create a square net partition in Rd as follows: 1 √d 1 √d md2−j] 1 √d 1 √d Qm,j = ( (m1 − 1)2−j, m12−j] × ··· × ( Sc(f )(s, j) =(cid:16)Z 1 if s ∈ Qm,j. For any s ∈ Rd and k ∈ N0 (N0 being the set of nonnegative integers), we define (md − 1)2−j, with m = (m1,··· , md) ∈ Zd. Let cm,j denote the center of Qm,j. Define rd−1(cid:17) 1 Pr(f )(t)2 dtdr (2.8) d(s, k) = Sc(f )(s, k)2−p − Sc(f )(s, k − 1)2−p. 2 ) ⊂ B(cm,j, r) whenever s ∈ Qm,j and ε ≥ 2−j, we have sc(f )(s, ε) ≤ Sc(f )(s, j), ∀ s ∈ Qm,j, ε ≥ 2−j. 2−jZB(cm,j ,r) Since B(s, r − ε ∂ ∂r 2 18 R. Xia and X. Xiong It is clear that Sc(f )(s, j) is increasing in j, so d(s, k) ≥ 0. At the same time, d(s, k) is constant on Qm,k andPk≤j d(s, k) = Sc(f )(s, j)2−p. Therefore, 2−j 2−j ∂ ∂ε B . τ Xm∈ZdXj≥1ZQm,jZ 2−j+1 = τZRdXj≥1 = τZRdXj≥1 X1≤k≤j = τZRdXk≥1 = τXm Xk≥1 (cid:16)ZB(s, ε 2 ) Sc(f )(s, j)2−pZ 2−j+1 (cid:16)ZB(s, ε 2 ) d(s, k)Z 2−j+1 (cid:16)ZB(s, ε 2 ) d(s, k)Xj≥kZ 2−j+1 (cid:16)ZB(s, ε 2 ) (cid:16)ZB(s, ε d(s, k)ZQm,kZ 2−k+1 2 ) ∂ ∂ε 2−j 2−j 0 Pε(g)(t)2 dt εd−1(cid:17)Sc(f )(s, j)2−pdεds Pε(g)(t)2 dt ∂ ∂ε εd−1(cid:17)dεds εd−1(cid:17)dεds εd−1(cid:17)dεds εd−1(cid:17)dεds . ∂ ∂ε Pε(g)(t)2 dt Pε(g)(t)2 dt ∂ ∂ε Pε(g)(t)2 dt 2 bmoc q . kgk2 Since g ∈ bmoc kak q QZT (Q) 1 q, Lemma 2.3 ensures the existence of a positive operator a ∈ L q and 2 (N ) such that ∂ ∂ε Pε(g)(t)2εdtdε ≤ a(s) and for s ∈ Q and for all cubes Q with Q < 1. Let eQm,k be the cube concentric with Qm,k and having side length 2−k+1. By the Fubini theorem and Lemma 1.4, we have ZQm,kZ 2−k+1 0 (cid:16)ZB(s, ε 2 ) ∂ ∂ε Pε(g)(t)2 dt 0 εd−1(cid:17)dεds ≤ 2dZ eQm,kZ 2−k+1 = 2dZT ( eQm,k) .ZQm,k a(s)ds. ∂ ∂ε ∂ ∂ε Pε(g)(s)2εdεds Pε(g)(s)2εdεds Then we deduce d(s, k)a(s)ds d(s, k)a(s)ds B . τXm Xk≥1ZQm,k = τZRdXk≥1 = τZRd = τZRd ≤ kfk2−p hc 2 Sc(f )(s, +∞)2−pa(s)ds Sc(f )(s)2−pa(s)ds ≤ kSc(f )k2−p p kak q . kfk2−p p kgk2 bmoc q hc p . kak q 2 Combining the estimates of A, B and II, we complete the proof. (cid:3) The following is the main theorem of this section. Theorem 2.11. Let 1 ≤ p < 2 and q be its conjugate index. We have hc with equivalent norms. More precisely, every g ∈ bmoc tional on hc q(Rd,M) q(Rd,M) defines a continuous linear func- p(Rd,M)∗ = bmoc p(Rd,M) by ℓg(f ) = τZ f (s)g∗(s)ds, ∀ f ∈ Lp(M; Wc d). Operator-valued local Hardy spaces 19 Conversely, every ℓ ∈ hc with p(Rd,M)∗ can be written as above and is associated to some g ∈ bmoc q(Rd,M) Proof. First, by Lemma 2.10, we get kℓk(hc p)∗ ≈ kgkbmoc q . (2.9) ℓg(f ) . kgkbmoc Now we prove the converse. Suppose that ℓ ∈ hc qkfkhc p(Rd,M)∗. By the Hahn-Banach theorem, ℓ p extends to a continuous functional on Lp(cid:0)N ; Lc exists h = (h′, h′′) ∈ Lq(cid:0)N ; Lc ℓ(f ) = τZRdZZeΓ ∂ ∂ε and 2(eΓ)(cid:1) ⊕p Lp(N ) with the same norm. Thus, there 2(eΓ)(cid:1) ⊕q Lq(N ) such that khkLq(cid:0)N ;Lc 2(eΓ)(cid:1)⊕q Lq(N ) = kℓk(hc p)∗ dtdε εd ds + τZRd P ∗ f (s)h′′∗(s)ds, Pε(f )(s + t)h′∗(s, t, ε) = τZRd f (u)eF(h)∗(u)du, where eF is the map defined in (2.5). Let g =eF(h). Following the proof of Theorem 2.8, we have ℓ(f ) = τZRd Thus, we have accomplished the proof of the theorem. d(cid:1). f (s)g∗(s)ds, ∀ f ∈ Lp(cid:0)M; Wc . kℓk(hc p)∗ kgkbmoc and q The following corollary gives an equivalent norm of the space bmoc q. Note that it is a strength- ening of the one-sided estimates in Lemmas 1.4 and 1.5. Corollary 2.12. Let 2 < q ≤ ∞. Then g ∈ bmoc is an M-valued Carleson q-measure on S and kJ(P) ∗ gkq < ∞. Furthermore, QZT (Q) q(Rd,M) if and only if dλg = ∂ + kJ(P) ∗ gkq. kgkbmoc ∂ ∂ε 1 q 2 1 2 Pε(g)(t)2εdtdε(cid:13)(cid:13)(cid:13) q ≈(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 Proof. From the proof of Lemma 2.10, we can see that if dλg = ∂ ∂ε Pε(g)(s)2εdsdε is an M- valued Carleson q-measure on S and J(P) ∗ g ∈ Lq(N ), then g defines a continuous functional on p(Rd,M): hc (cid:3) ∂ε Pε(g)(s)2εdsdε and and that According to Theorem 2.11, there exists a function g′ ∈ bmoc ℓ(f ) = τZRd QZT (Q) ∂ ∂ε 1 1 ∂ ∂ε QZT (Q) f (s)g∗(s)ds = τZRd QZT (Q) ∂ ∂ε 1 f (s)g∗(s)ds, 1 2 q 2 Pε(g)(t)2εdtdε(cid:13)(cid:13)(cid:13) Pε(g)(t)2εdtdε(cid:13)(cid:13)(cid:13) + kJ(P) ∗ gkq. q(Rd,M) such that + kJ(P) ∗ gkq 1 2 q 2 f (s)g′∗(s)ds, Pε(g)(t)2εdtdε(cid:13)(cid:13)(cid:13) 1 2 q 2 + kJ(P) ∗ gkq. kℓk(hc kg′kbmoc q s∈Q⊂Rd Q<1 s∈Q⊂Rd Q<1 p)∗ .(cid:13)(cid:13)(cid:13) sup + .(cid:13)(cid:13)(cid:13) sup + τZRd .(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 q kgkbmoc for any f ∈ hc p(Rd,M). Thus, g = g′ with 20 R. Xia and X. Xiong The inverse inequality is already contained in Lemmas 1.4 and 1.5. We obtain the desired assertion. (cid:3) 2.4. The equivalence hq = bmoq. We now show that hc q(Rd,M) for 2 < q < ∞. Thus according to the duality obtained in the last subsection, the dual of hc p(Rd,M) agrees with hc q(Rd,M) when 1 < p < 2. Let us begin with two lemmas concerning the comparison of sc(f ) and gc(f ). We require an auxiliary truncated square function. For s ∈ Rd and ε ∈ [0, 2 q(Rd,M) = bmoc 3 ], we define: (2.10) Lemma 2.13. We have 3 egc(f )(s, ε) =(cid:16)Z 2 ε ∂ ∂r Pr(f )(s)2rdr(cid:17) 1 2 . egc(f )(s, ε) . sc(f )(s, ε 2 ), where the relevant constant depends only on the dimension d. Proof. By translation, it suffices to prove this inequality for s = 0. Given ε ∈ [0, 2 that ε ≤ r ≤ 2 {(t, u) ∈ Rd+1 r√5 3 ], for any r such 3 , let us denote the ball centered at (0, r) and tangent to the boundary of the cone r u} by eBr. We notice that the radius of eBr is greater than or equal to . By the harmonicity of ∂ ∂r Pr(f ), we have + : t < r− ε 2 ∂ ∂r Pr(f )(0) = Pu(f )(t)dt. where cd+1 is the volume of the unit ball of Rd+1. Integrating the above inequality, we get 1 d+1 √5 ∂ ∂u eBrZ eBr cd+1rd+1Z eBr cd+1rdZ eBr √5 d+1 ε 3 ∂ ∂u Pu(f )(t)2dt, ∂ ∂u Pu(f )(t)2dtdudr. Then by (0.10), we arrive at ∂ ∂r Pr(f )(0)2 ≤ 3 ε ∂ ∂r (2.11) Pr(f )(0)2rdr ≤Z 2 Z 2 √5√5−1 √5√5+1 Since (t, u) ∈ eBr implies 2 Z eBr Pu(f )(t)2Z 2u cd+1 Z 1 u ≤ r ≤ majorized by ∂ ∂u u and ε √5 d+1 u 2 ε 2 ≤ u ≤ 1, the right hand side of (2.11) can be 1 rd drdtdu ≤ Csc(f )(0, ε 2 )2, 2 ). (cid:3) where C is a constant depending only on d. Therefore,egc(f )(0, ε) . sc(f )(0, ε Lemma 2.14. Let 1 ≤ p < ∞. Then for any f ∈ hc p(Rd,M), we have ksc(f )kp + kP ∗ fkp . kgc(f )kp + kP ∗ fkp. Proof. We first deal with the case when 1 ≤ p < 2. Let g be a function in bmoc conjugate index of p). Following a similar calculation as (1.3), we can easily check that τZRd Pε(g)∗(s)εdεds Pε(f )(s) ∂ ∂ε ∂ ∂ε 3 q(Rd,M) (q is the P ∗ f (s)(P 1 3 ∗ g(s))∗ds + P ∗ f (s)(I(P 1 3 8π 3 τZRd ) ∗ g(s))∗ds(cid:17) f (s)g∗(s)ds = 4τZRdZ 2 +(cid:16)τZRd 0 def = I + II. The term II can be treated in the same way as in the proof of Lemma 2.10: Applying Lemma 2.3, we have II . kP ∗ fkp · kJ(P 1 3 ) ∗ fkp. II . kP ∗ fkp · kgkbmoc q . Operator-valued local Hardy spaces 21 Concerning the term I, we have ∂ ∂ε Pε(f )(s)2egc(f )(s, ε)p−2εdεds · τZRdZ 2 0 3 ∂ ∂ε Pε(g)(s)2egc(f )(s, ε)2−pεdεds Following the argument for the estimate of A in the proof of Lemma 2.10, we deduce similarly that p. Now we deal with term B′. By Lemma 2.13, we have 3 0 I2 . τZRdZ 2 def= A′ · B′. A′ . kegc(f )kp B′ ≤ τZRdZ 2 0 3 ∂ ∂ε Pε(g)(s)2sc(f )(s, ε 2 )εdεds. Then we can apply almost the same argument as in the estimate of B. There is only one minor difference: when ε ≥ 2−j and s ∈ Qm,j, we have sc(f )(s, ε 2 ) ≤ Sc(f )(s, j + 1). We conclude that Combining the estimates of I, A′ and B′ with Theorem 2.11, we get B′ . kgk2 bmoc qksc(f )k2−p p . The case p = 2 is obvious. For p > 2, choose a positive g ∈ L( p ksc(f )kp + kP ∗ fkp . kegc(f )kp + kP ∗ fkp . kgc(f )kp + kP ∗ fkp. 2 )′ (N ) with norm one such that, Then the assertion for the case p > 2 is also proved. (cid:3) To proceed further, we introduce the definition of tent spaces. In the noncommutative setting, these spaces were first defined and studied by Mei [31]. Definition 2.15. For any function defined on Rd×(0, 1) = S with values in L1(M)+M, whenever it exists, we define For 1 ≤ p < ∞, we define equipped with the norm kfkT c T c ∞ norm of f as 2 , s ∈ Rd. Ac(f )(s) =(cid:16)ZeΓ f (t + s, ε)2 dtdε εd+1(cid:17) 1 p (Rd,M) = {f : Ac(f ) ∈ Lp(N )} T c QZT (Q) f (s, ε)2 dsdε 2(cid:13)(cid:13)(cid:13)M ε (cid:17) 1 Q≤1(cid:13)(cid:13)(cid:13)(cid:16) 1 ∞ = sup , kfkT c p (Rd,M) = kAc(f )kp. For p = ∞, define the operator-valued column By the noncommutative Hardy-Littlewood maximal inequality (the one dimension R case is given by [30, Theorem 3.3], the case Rd is a simple corollary of (2.1) and Lemma 2.7), there exists a positive a ∈ L( p 2 )′ (N ) such that kak( p 2 )′ ≤ 1 and Therefore, ksc(f )k2 ∂ ∂ε p =(cid:13)(cid:13)(cid:13)ZZeΓ = τZRdZZeΓ = τZRdZ 1 0 ∂ ∂ε Pε(f )(· + t)2 dtdε εd−1(cid:13)(cid:13)(cid:13) p Pε(f )(s + t)2 dtdε εd−1ZB(t,ε) Pε(f )(t)2 dtdε ∂ ∂ε 2 εd−1 g(s)ds g(s)ds. 1 ksc(f )k2 B(t, 2−k)ZB(t,2−k) p = τZRdZ 1 0 ≤ cdτZRdZ 1 0 ≤ cd(cid:13)(cid:13)Z 1 ∂ 0 ∂ε ≤ cdkgc(f )kp. ∂ ∂ε g(s)ds ≤ a(t), ∀ t ∈ Rd, ∀ ε > 0. εd−1ZB(t,ε) Pε(f )(t)2 dtdε Pε(f )(t)2εa(t)dtdε g(s)ds ∂ ∂ε Pε(f )(t)2εdtdε(cid:13)(cid:13) p 2 kak( p 2 )′ 22 R. Xia and X. Xiong and the corresponding space is T c ∞(Rd,M) = {f : kfkT c ∞ < ∞}. p + 1 ∂ε Pε(f )(s) and ∂ p (Rd,M)∗ = T c q (Rd,M) for 1 ≤ p < ∞ and 1 Remark 2.16. By the same arguments used in the proof of Theorem 2.11, we can prove the duality that T c q = 1. For the case p = 1, it suffices to replace ∂ ∂ε Pε(g)(s) in the proof of Lemma 2.10 by f (s, ε) and g(s, ε) 1 (Rd,M)∗. On respectively. A similar argument will give us the inclusion that T c ∞(Rd,M), we get the reverse inclusion. For 1 < p < ∞, the other hand, since L∞(N ; Lc the tent space T c p (Rd,M) we define above is a complemented subspace of the column tent space defined in [30]. So by Remark 4.6 in [51], we obtain the duality that T c Theorem 2.17. For 2 < q < ∞, hc Proof. First, we show the inclusion hc show that hc we have 2(eΓ)) ⊂ T c q(Rd,M) = bmoc q(Rd,M). By Theorem 2.11, it suffices to p(Rd,M), p (Rd,M)∗ = T c q(Rd,M) with equivalent norms. q(Rd,M) and g ∈ hc q(Rd,M) ⊂ bmoc ∞(Rd,M) ⊂ T c q(Rd,M) ⊂ hc q (Rd,M). τZRd g(s)f∗(s)ds = Pε(f )∗(s + t) 4 ∂ ∂ε ∂ ∂ε Pε(g)(s + t) p(Rd,M)∗. Applying (1.3), for any f ∈ hc cdZRdZZeΓ +ZRd g(s)f∗(s)ds(cid:12)(cid:12) ≤(cid:13)(cid:13)ε · P ∗ g(s)(P ∗ f (s))∗ds + 4πZRd Pε(g)(cid:13)(cid:13)Lp(cid:0)N ;L2(eΓ)(cid:1)kε · + k(P + I(P)) ∗ gkp · kP ∗ fkq ∂ ∂ε ∂ ∂ε (cid:12)(cid:12)τZRd dtdε εd−1 ds I(P) ∗ g(s)(P ∗ f (s))∗ds. Pε(f )kLq(cid:0)N ;L2(eΓ)(cid:1) Now, we show that for any 1 ≤ p < 2 and g ∈ hc Since 2 < q < ∞, we have 1 < q inequality, we get p(Rd,M), we have k(P + I(P)) ∗ gkp . kgkhc p. 2 < ∞. Applying the noncommutative Hardy-Littlewood maximal . q ≤(cid:16)(cid:13)(cid:13)sc(g)kp + k(P + I(P)) ∗ g(cid:13)(cid:13)p(cid:17)kfkhc .(cid:13)(cid:13)(cid:13) sup QZQ f (t)2dt(cid:13)(cid:13)(cid:13) . kf2k + 1 1 2 q 2 q 2 1 2 = kfkq. s∈Q⊂Rd q(Rd,M) for any 2 < q ≤ ∞. Then by Theorem 2.11, we get kfkbmoc q This implies that Lq(N ) ⊂ bmoc p(Rd,M) ⊂ Lp(N ). Therefore we deduce that hc (2.12) Then, by the Holder inequality, Thus, (2.13) Set k(P + I(P)) ∗ gkp . kgkp . kgkhc p . (cid:12)(cid:12)τZRd g(s)f∗(s)ds(cid:12)(cid:12) . kfkhc q(Rd,M) ⊂ hc qkgkhc p . We have proved hc q(Rd,M). Let us turn to the reverse inclusion bmoc q(Rd,M) ⊂ bmoc tent spaces in Definition 2.15. We claim that for q > 2, every f ∈ bmoc functional on T c p (Rd,M)⊕p Lp(N ). Indeed, for any h = (h′, h′′) ∈ T c q(Rd,M). We need to make use of the q(Rd,M) induces a linear p (Rd,M)⊕p Lp(N ), we define 0 h′(s, ε) ℓf (h) = τZRdZ 1 + τZRd Ac(h′)(s, ε) =Z 1 (h′)(s, ε) =Z 1 c A ∂ ∂ε Pε(f )∗(s)dεds h′′(s)[(P ∗ f )∗(s) + 4π(I(P) ∗ f )∗(s)]ds. ε ZB(s,r− ε ε ZB(s, r rd+1 2 ) h′(s, ε)2 dtdr 2 ) h′(s, ε)2 dtdr rd+1 . , Operator-valued local Hardy spaces 23 Then by the Cauchy-Schwarz inequality, we arrive at c ∂ ∂ε ℓf (h) .(cid:16)τZRdZ 1 ·(cid:16)τZRdZ 1 +(cid:12)(cid:12)τZRd 0 (cid:16)ZB(s, ε (h′)(s, ε)p−2dεds(cid:17) 1 εd−1(cid:17)A Pε(f )(t)2 dt 2 ) 0 (cid:16)ZB(s, ε εd−1(cid:17)A (h′)(s, ε)2−pdεds(cid:17) 1 Pε(f )(t)2 dt 2 ) h′′(s)(P ∗ f (s))∗ds(cid:12)(cid:12) +(cid:12)(cid:12)τZRd h′′(s)(I(P) ∗ f (s))∗ds(cid:12)(cid:12). ∂ ∂ε c 2 2 Following a similar argument as in the proof of Lemma 2.10, we obtain that p⊕pLp · kfkbmoc . khkT c q , ℓf (h) . (kh′kT c which implies that kℓfk ≤ cqkfkbmoc p + kh′′kLp)kfkbmoc q . So the claim is proved. q Next we show that kfkhc q ≤ Cqkℓfk. By definition, we can regard T c 2(eΓ)(cid:1) in the natural way. Then, ℓf extends to a linear functional on Lp(cid:0)N ; Lc Lp(cid:0)N ; Lc Lp(N ). Thus, there exists g = (g′, g′′) ∈ Lq(cid:0)N ; Lc kgkLq(cid:0)N ;Lc 2(eΓ)(cid:1) ⊕p Lp(N ), and for any h = (h′, h′′) ∈ Lp(cid:0)N ; Lc h′(s, ε)ZB(s,ε) ℓf (h) = τZRdZZeΓ = τZRdZ 1 2(eΓ)(cid:1) ⊕q Lq(N ) such that 2(eΓ)(cid:1)⊕q Lq(N ) ≤ kℓfk ds + τZRd εd+1 + τZRd h′(s, ε)g′∗(s, t, ε) h′′(s)g′′∗(s)ds. h′′(s)g′′∗(s)ds g′∗(s, t, ε)dt dtdε εd+1 dsdε 0 p as a closed subspace of 2(eΓ)(cid:1) ⊕p Comparing the equalities above with (2.13), we get ∂ ∂ε Pε(f )(s) = 1 εd+1ZB(0,ε) g′(s, t, ε)dt and By Lemma 2.14, we have kfkhc q ∂ ∂ε 0 .(cid:13)(cid:13)(cid:13)(cid:0)Z 1 ≤ cd(cid:13)(cid:13)(cid:13)(cid:0)Z 1 . kg′kLq(cid:0)N ;Lc 0 1 + kP ∗ fkq P ∗ f + 4πI(P) ∗ f = g′′. 2(cid:13)(cid:13)(cid:13)q Pε(f )2εdε(cid:1) 1 εd+1ZB(0,ε) g′(s, t, ε)2dtdε(cid:1) 1 2(cid:13)(cid:13)(cid:13)q 2(eΓ)(cid:1) + kP ∗ fkq. G(s) = 2πZ ∞ e−2πεPε(s)dε. 0 + kP ∗ fkq Now let us majorize the second term kP ∗ fkq by kg′′kq. Indeed, consider the function We can easily check that G ∈ L1(Rd), kGk1 ≤ 1 and bG(ξ) = (1 + ξ)−1. This means that the operator (1 + I)−1 is a contractive Fourier multiplier on Lq(N ). Therefore, kP ∗ fkq ≤ k(P + I(P)) ∗ fkq ≤ 4πkg′′kq. Finally, we conclude that kfkhc equivalent norms. q . kℓfk . kfkbmoc q and then hc q(Rd,M) = bmoc q(Rd,M) with (cid:3) Armed with the theorem above, we are able to extend the content of Theorem 2.8. Theorem 2.18. (1) The map F extends to a bounded map from L∞(cid:0)N ; Lc kF(h)kbmoc . khkL∞(cid:0)N ;Lc 2(eΓ)(cid:1) ⊕∞ L∞(N ) into bmoc(Rd,M) and 2(eΓ)(cid:1)⊕∞L∞(N ) . 24 R. Xia and X. Xiong 2(eΓ)(cid:1) ⊕p Lp(N ) into hc p(Rd,M) Proof. (1) is already contained in Theorem 2.8. When p > 2, (2) follows from Theorem 2.8 and Theorem 2.17. The case p = 2 is trivial. For the case 1 < p < 2, according to Theorem 2.11, we have p and kF(h)khc (2) For 1 < p < ∞, F extends to a bounded map from Lp(cid:0)N ; Lc 2(eΓ)(cid:1)⊕pLp(N ) F(h)(s)f∗(s)ds(cid:12)(cid:12). . khkLp(cid:0)N ;Lc q≤1(cid:12)(cid:12)τZRd Then, by Theorem 2.17 and (2.12), for h = (h′, h′′) ∈ Lp(cid:0)N ; Lc . sup kfkbmoc kF(h)khc . p h′(s, t, ε) sup kfkbmoc . sup kfkhc 2(eΓ)(cid:1) ⊕p Lp(N ), we have Pε(f )∗(s + t)dtdε + h′′(s)([P + 4πI(P)] ∗ f∗(s))(cid:3)ds(cid:12)(cid:12)(cid:12) q≤1(cid:12)(cid:12)τZRd F(h)(s)f∗(s)ds(cid:12)(cid:12) q ≤1(cid:12)(cid:12)(cid:12)τZRd(cid:2)ZZeΓ . khkLp(cid:0)N ;Lc 2(eΓ)(cid:1)⊕pLp(N ) 2(eΓ)(cid:1) ⊕p Lp(N ). Thus, we deduce the following duality theorem: (cid:3) p(Rd,M) is a complemented subspace of q(Rd,M) with equivalent norms for any 1 < p < ∞. The above theorem shows that, for any 1 < p < ∞, hc p(Rd,M)∗ = hc ∂ ∂ε . Lp(cid:0)N ; Lc The desired inequality is proved. Theorem 2.19. We have hc 3. Interpolation In this section we study the interpolation of local Hardy and bmo spaces by transferring the problem to that of the operator-valued Hardy and BMO spaces defined in [30]. We begin with an easy observation on the difference between bmoc Lemma 3.1. For 2 < q ≤ ∞, we have kgkbmoc q and BMOc q norms. BMOc q q . q ≈(cid:0)kgkq q(cid:1) 1 + kJ(P) ∗ gkq Proof. Repeating the proof of Proposition 1.3 with k · kM replaced by k · kL q q . By Lemma 2.3, it is also evident that kJ(P)∗ gkq . kgkbmoc kgkBMOc . kgkbmoc q 2 (N ;ℓ∞), we have q . Then we obtain On the other hand, by Corollary 2.12, we have BMOc q (cid:0)kgkq .(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 q q . kgkbmoc q(cid:1) 1 + kJ(P) ∗ gkq Pε(g)(t)2εdtdε(cid:13)(cid:13)(cid:13) ∂ ∂ε 1 QZT (Q) . kgkbmoc q 1 2 q 2 + kJ(P) ∗ gkq. Clearly, the first term on the right side can be estimated from above by kgkBMOq rem 3.4]). Therefore, c (see [51, Theo- kgkbmoc q . kgkBMOq c + kJ(P) ∗ gkq ≈(cid:0)kgkq BMOc q q(cid:1) 1 + kJ(P) ∗ gkq q . (cid:3) Thus, the lemma is proved. Define Fq(N ) to be the space of all f ∈ Lq(M; Rc q(Rd,M) and BMOc lemma, we see that bmoc interpolation between BMOc following lemma: Lemma 3.2. Let 2 < q < ∞ and 0 < θ < 1. Then d) such that kJ(P)∗ fkq < ∞. From the above q(Rd,M) ⊕q Fq(N ) have equivalent norms. By the q(Rd,M) and BMOc(Rd,M) (see [30] for more details), we deduce the (cid:0)bmoc q(Rd,M), bmoc(Rd,M)(cid:1)θ ⊂ bmoc (Rd,M) with = q 1 − θ . Operator-valued local Hardy spaces 25 Proof. By Lemma 3.1, we see that with equivalent norms. Define a map bmoc q(Rd,M) = BMOc q(Rd,M) ⊕q Fq(N ). Υq : Fq(N ) −→ Lq(N ) f 7−→ J(P) ∗ f. Thus, Υq defines an isometric embedding of Fq(N ) into Lq(N ). Then by the interpolation between BMOc q(Rd,M) ⊕q Fq(N ), BMOc(Rd,M) ⊕∞ F∞(N )(cid:1)θ q(Rd,M), BMOc(Rd,M)(cid:1)θ ⊕(cid:0)Fq(N ), F∞(N )(cid:1)θ (Rd,M) ⊕ F(N ) = bmoc (Rd,M), (cid:3) q(Rd,M) and BMOc(Rd,M), we get q(Rd,M), bmoc(Rd,M)(cid:1)θ =(cid:0)BMOc (cid:0)bmoc =(cid:0)BMOc ⊂ BMOc which completes the proof. Theorem 3.3. Let 1 < p < ∞. We have p′ = 1−θ (3.1) (3.3) bmoc , p′ (Rd,M). p(Rd,M), hc By Theorem 2.19, hc = hc p(Rd,M). 1(Rd,M)(cid:1) 1 p (3.1) is proper, we will get the proper inclusion which is in contradiction with Lemma 3.2. Thus, we have Proof. Let 1 < p < 2 and 1 hc p + θ. Since the map E in Definition 2.4 is an isometry from p is a reflexive Banach space. Then applying [2, Corollary 4.5.2], we know that p(Rd,M) to Lp(cid:0)N ; Lc the dual of(cid:0)hc (cid:0)bmoc(Rd,M), hc 2(eΓ)(cid:1) ⊕p Lp(N ), we have (cid:0)hc 1(Rd,M)(cid:1)θ ⊂ hc p(Rd,M), hc 1(Rd,M)(cid:1)θ is(cid:0)bmoc q(Rd,M), bmoc(Rd,M)(cid:1)θ. Therefore, if the inclusion (Rd,M) ((cid:0)bmoc q(Rd,M), bmoc(Rd,M)(cid:1)θ (cid:0)hc 1(Rd,M)(cid:1)θ = hc p(Rd,M), hc p′ (Rd,M). (cid:0)hc q(Rd,M), bmoc(Rd,M)(cid:1)θ = hc 2(eΓ)(cid:1) ⊕p Lp(N ) via the maps E and F in Definition 2.4. This implies mented subspace of Lp(cid:0)N ; Lc (cid:0)hc p1 (Rd,M), hc (3.2) By duality and [2, Corollary 4.5.2] again, the above equality implies that for q′ = q easier to handle. Indeed, by Theorem 2.18, we have, for any 1 < p < ∞, hc with 1 theorem (see [48]), we get the desired assertion. . Combining this equivalence with (3.2), (3.3), and applying Wolff's interpolation (cid:3) The following theorem is the mixed version of Theorem 3.3, which states that h1(Rd,M) and For the case where 1 < p1, p2 < ∞, the interpolation of hc q′ (Rd,M). p1 (Rd,M) and hc p2 (Rd,M) is much p(Rd,M) is a comple- p2 (Rd,M)(cid:1)θ = hc that, for any 1 < p1, p2 < ∞, bmo(Rd,M) are also good endpoints of Lp(N ). Theorem 3.4. Let 1 < p < ∞. We have (cid:0)X, Y(cid:1) 1 p X = bmo(Rd,M) or L∞(N ), and Y = h1(Rd,M) or L1(N ). Proof. By the same argument as in the proof of Theorem 3.3, we have the inclusion = Lp(N ) with equivalent norms, where p(Rd,M), p = 1−θ p1 + θ p2 1−θ , which ensures by duality that for 1 p′ = 1−θ p + θ. Then by Proposition 5.18, (cid:0)bmoq(Rd,M), bmo(Rd,M)(cid:1)θ ⊂ bmoq′ (Rd,M) (cid:0)hp(Rd,M), h1(Rd,M)(cid:1)θ ⊃ hp′ (Rd,M) = Lp′(N ) q′ = q θ , Lp′(N ) ⊂(cid:0)hp(Rd,M), h1(Rd,M)(cid:1)θ =(cid:0)Lp(N ), h1(Rd,M)(cid:1)θ . 26 R. Xia and X. Xiong Since h1(Rd,M) ⊂ L1(N ), then Combining the estimates above, we have (cid:0)hp(Rd,M), h1(Rd,M)(cid:1)θ ⊂(cid:0)Lp(N ), L1(N )(cid:1)θ = Lp′ (N ). (cid:0)hp(Rd,M), h1(Rd,M)(cid:1)θ = Lp′(N ). Again, using duality and Wolff's interpolation theorem, we conclude the proof by the same trick as in the proof of Theorem 3.3. (cid:3) We end this section by some real interpolation results. Corollary 3.5. Let 1 < p < ∞. Then we have (1) (cid:0)bmoc(Rd,M), hc (2) (cid:0)X, Y(cid:1) 1 h1(Rd,M) or L1(N ). 1(Rd,M)(cid:1) 1 p ,p = hc p(Rd,M) with equivalent norms. p ,p = Lp(N ) with equivalent norms, where X = bmo(Rd,M) or L∞(N ), and Y = Proof. Both (1) and (2) follow from [2, Theorem 4.7.2]; we only prove (1). Let 1 < p1 < p < p2 < ∞ with 1 . By [2, Theorem 4.7.2], we write p = 1−η p1 + η p2 p ,p (3.4) 1(Rd,M)(cid:1) 1 (cid:0)bmoc(Rd,M), hc =(cid:16)(cid:0)bmoc(Rd,M), hc p(Rd,M) is a complemented subspace of Lp(cid:0)N ; Lc 1(Rd,M)(cid:1) 1 ,(cid:0)bmoc(Rd,M), hc 4. Calder´on-Zygmund theory 1(Rd,M)(cid:1) 1 2(eΓ)(cid:1) ⊕p Lp(N ). Then the assertion (1) follows from Theorem 3.4 and the facts that(cid:0)Lp1(N ), Lp2 (N )(cid:1)η,p = Lp(N ) and that hc p2(cid:17)η,p (cid:3) p1 . We introduce the Calder´on-Zygmund theory on operator-valued local Hardy spaces in this sec- tion. It is closely related to the similar results of [14], [26], [33] and [53]. The results in the following will be used in the next section to investigate various square functions that characterize local Hardy spaces. Let K be an L1(M) + M)-valued tempered distribution which coincides on Rd \ {0} with a locally integrable L1(M) + M-valued function. We define the left singular integral operator K c associated to K by K(s − t)f (t)dt, and the right singular integral operator K r associated to K by K c(f )(s) =ZRd K r(f )(s) =ZRd f (t)K(s − t)dt. Let bmoc Both K c(f ) and K r(f ) are well-defined for sufficiently nice functions f with values in L1(M)∩M, for instance, for f ∈ S ⊗ (L1(M) ∩ M). 0(Rd,M) denote the subspace of bmoc(Rd,M) consisting of compactly supported func- tions. The following lemma is an analogue of Lemma 2.1 in [51] for inhomogeneous spaces. Notice that the usual Calder´on-Zygmund operators (the operators satisfying the condition (1) and (3) in 1(Rd,M). Thus, we need to impose an extra the following lemma) are not necessarily bounded on hc decay at infinity on the kernel K. Lemma 4.1. Assume that (1) the Fourier transform of K is bounded: supξ∈Rd kbK(ξ)kM < ∞; (2) K satisfies the size estimate at infinity: there exist C1 and ρ > 0 such that kK(s)kM ≤ , ∀s ≥ 1; C1 sd+ρ (3) K has the Lipschitz regularity: there exist C2 and γ > 0 such that kK(s − t) − K(s)kM ≤ C2 tγ s − td+γ , ∀s > 2t. Operator-valued local Hardy spaces 27 Then K c is bounded on hc A similar statement also holds for K r and the corresponding row spaces. p(Rd,M) for 1 ≤ p < ∞ and from bmoc 0(Rd,M) to bmoc(Rd,M). Proof. First suppose that K c maps constant functions to zero. This amounts to requiring that K c(1Rd ) = 0. Let Q ⊂ Rd be a cube with Q < 1. Since the assumption of Lemma 2.1 in [51] are included in the ones of this lemma, we get QZQ K c(f ) − K c(f )Q2dt(cid:1) 1 2(cid:13)(cid:13)(cid:13)M (cid:13)(cid:13)(cid:13)(cid:0) 1 . kfkBMOc . kfkbmoc . Now let us focus on the cubes with side length 1. Let Q be a cube with Q = 1 and eQ = 2Q be the cube concentric with Q and with side length 2. Decompose f as f = f1 + f2, where f1 = 1 eQf and f2 = 1Rd\ eQf . Then K c(f ) = K c(f1) + K c(f2). We have (cid:13)(cid:13)(cid:13) 1 QZQ K c(f )2ds(cid:13)(cid:13)(cid:13)M .(cid:13)(cid:13)(cid:13) 1 QZQ K c(f1)2ds(cid:13)(cid:13)(cid:13)M +(cid:13)(cid:13)(cid:13) 1 QZQ K c(f2)2ds(cid:13)(cid:13)(cid:13)M . The first term is easy to estimate. By assumption (1) and (0.10), (cid:13)(cid:13)(cid:13) 1 1 1 1 1 . sup QZRd bK(ξ)bf1(ξ)2dξ(cid:13)(cid:13)(cid:13)M QZQ K c(f1)2ds(cid:13)(cid:13)(cid:13)M ≤(cid:13)(cid:13)(cid:13) QZRd bf1(ξ)2dξ(cid:13)(cid:13)(cid:13)M .(cid:13)(cid:13)(cid:13) QZ eQ f (s)2ds(cid:13)(cid:13)(cid:13)M =(cid:13)(cid:13)(cid:13) QZQ f (s)2ds(cid:13)(cid:13)(cid:13)M Q=1(cid:13)(cid:13)(cid:13) =(cid:12)(cid:12)(cid:12)ZRd\ eQ K(s − t)f (t)dt(cid:12)(cid:12)(cid:12) ≤ZRd\ eQ kK(s − t)kMdt ·ZRd\ eQ kK(s − t)k−1 .ZRd\ eQ kK(s − t)kMf (t)2dt .ZRd\ eQ s − td+ρf (t)2dt. K(s − t)f2(t)dt(cid:12)(cid:12)(cid:12) 1 2 2 . K c(f2)(s)2 =(cid:12)(cid:12)(cid:12)ZRd To estimate the second term, using assumption (2) and (0.10) again, we have MK(s − t)f (t)2dt Set eQm = eQ + 2m for every m ∈ Zd. Then Rd \ eQ = ∪m6=0eQm. Continuing the estimate of K c(f2)(s)2, for any s ∈ Q, we have K c(f2)(s)2 ≤ Xm6=0Z eQm ≈ Xm6=0 1 s − td+ρf (t)2dt 1 md+ρZ eQm f (t)2dt . kfkbmoc . Combining the previous estimates, we deduce that K c is bounded from bmoc Now we illustrate that the additional requirement K c(1Rd ) = 0 is not needed. First, a similar argument as above ensures that for every compactly supported f ∈ L∞(N ), kK c(f )kbmoc . kfk∞. Then we follow the argument of [10, Proposition II.5.15] to extend K c on the whole L∞(N ), as 0(Rd,M) to bmoc(Rd,M). K c(f )(s) = lim j (cid:2)K c(f 1Bj )(s) −Z1<t≤j K(−t)f (t)dt(cid:3), ∀ s ∈ Rd, where Bj is the ball centered at the origin with radius j. Let us show that the sequence on the right hand side converges pointwise in the norm k ·kM and uniformly on any compact set Ω ⊂ Rd. 28 R. Xia and X. Xiong To this end, we denote by gj the j-th term of this sequence. Let l be the first natural number such that l ≥ 2 sups∈Ω s. Then for s ∈ Ω and j > l, we have gj(s) = gl(s) +Zl<t≤j(cid:0)K(s − t) − K(−t)(cid:1)f (t)dt. By assumption (3), the integral on the right hand side is bounded by a bounded multiple of kfk∞, uniformly on s ∈ Ω. This ensures the convergence of gj, so K c(f ) is a well-defined function. Now we have to estimate the bmoc-norm of K c(f ). Taking any cube Q ⊂ Rd, by the uniform convergence of gj on Q in M, we have Similarly, (cid:13)(cid:13)(cid:0)ZQ K c(f )(s) − (K c(f ))Q2ds(cid:1) 1 2(cid:13)(cid:13)M (cid:13)(cid:13)(cid:0)ZQ K c(f )(s)2ds(cid:1) 1 2(cid:13)(cid:13)M = lim . j (cid:13)(cid:13)(cid:0)ZQ gj(s) − (gj)Q2ds(cid:1) 1 2(cid:13)(cid:13)M j (cid:13)(cid:13)(cid:0)ZQ gj(s)2ds(cid:1) 1 2(cid:13)(cid:13)M . = lim Hence, by the fact that gj and K c(f 1Bj ) differ by a constant, we obtain kK c(f )kbmoc = lim j kgjkbmoc . lim sup j kK c(f 1Bj )kbmoc + kfk∞ . kfk∞. Therefore, K c defined above extends to a bounded operator from L∞(N ) to bmoc(Rd,M). In particular, K c(1Rd ) determines a function in bmoc(Rd,M). Then for f and Q as above, we have K c(f ) = K c(f1) + K c(f2) + K c(1Rd )f eQ, so kK c(f )kbmoc ≤ kK c(f1)kbmoc + kK c(f2)kbmoc + kK c(1Rd )kbmoc kf eQkM . kfkbmoc + kf eQkM . kfkbmoc . Thus we have proved the bmoc-boundedness of K c in the general case. By duality, the boundedness of K c on hc 1(Rd,M) is equivalent to that of its adjoint map (K c)∗ on bmoc 0(Rd,M). It is easy to see that (K c)∗ is also a singular integral operator: (K c)∗(g) =ZRd eK(s − t)g(t)dt, 0(Rd,M). Thus we get the boundedness of K c on hc 1(Rd,M) and bmoc(Rd,M) in Theorem 3.3, we get the boundedness of K c on hc where eK(s) = K∗(−s). Obviously, eK also satisfies the same assumption as K, so (K c)∗ is bounded on bmoc between hc for 1 < p < ∞. The assertion is proved. Remark 4.2. Under the assumption of the above lemma, K c(1Rd ) is a constant, so it is the zero element in BMOc(Rd,M). 1(Rd,M). Then, by the interpolation p(Rd,M) (cid:3) A special case of Lemma 4.1 concerns the Hilbert-valued kernel K. Let H be a Hilbert space and k : Rd → H be a H-valued kernel. We view the Hilbert space as the column matrices in B(H) with respect to a fixed orthonormal basis. Put K(s) = k(s) ⊗ 1M ∈ B(H)⊗M. For nice functions f : Rd → L1(M)+M, K c(f ) takes values in the column subspace of L1(B(H)⊗M)+ L∞(B(H)⊗M). Consequently, kK c(f )kLp(B(H)⊗N ) = kK c(f )kLp(N ;Hc). Since k(s) ⊗ 1M commutes with M, K c(f ) = K r(f ) for f ∈ L2(N ). Let us denote this common operator by kc. Here the superscript c refers to the previous convention that H is identified with the column matrices in B(H). Thus, Lemma 4.1 implies the following Corollary 4.3. Assume that (1) supξ∈Rd kbk(ξ)kH < ∞; (2) kk(s)kH . 1 (3) kk(s − t) − k(s)kH . tγ Then the operator kc is bounded (1) from bmoα (2) and from hc sd+ρ , ∀s ≥ 1, for some ρ > 0; s−td+γ , ∀s > 2t, for some γ > 0. 0 (Rd,M) to bmoα(Rd, B(H)⊗M), where α = c, α = r or we leave out α; p(Rd,M) to hc p(Rd, B(H)⊗M) for 1 ≤ p < ∞. Operator-valued local Hardy spaces 29 Proof. Since K c(f ) = K r(f ) on the subspace Lp(N ) ⊂ Lp(B(H)⊗N ), (1) follows immediately from Lemma 4.1. Denote the column subspace of bmoc(Rd, B(H)⊗M) (resp. hc p(Rd, B(H)⊗M)) by bmoc(Rd, H c⊗M) (resp. hc p(Rd, H c⊗M)). Consider the adjoint operator of kc which is de- noted by (kc)∗. It admits the convolution kernel eK(s) =ek(s)⊗ 1M, whereek(s) = k(−s)∗ (so it is a row matrix). Applying Lemma 4.1 to (kc)∗, we get that (kc)∗ is bounded from bmoc(Rd, H c⊗M) to bmoc(Rd,M). Then kc is bounded from hc 1(Rd, H c⊗M), and thus bounded Interpolating this with the boundedness of kc from from hc bmoc 0(Rd,M) to bmoc(Rd, B(H)⊗M), we deduce the desired assertion in (2). 1(Rd,M) into hc 1(Rd, B(H)⊗M). 1(Rd,M) to hc Remark 4.4. Let 1 ≤ p ≤ 2. Since L∞(N ) ⊆ bmoc(Rd,M), we get hc Theorem 3.3 and the fact that hc 4.3 ensures that 2(Rd,M) = L2(N ), we have hc kkc(f )kLp(N ;Hc) . kkc(f )khc p(Rd,B(H)⊗M) . kfkhc p(Rd,M) for any f ∈ hc p(Rd,M). 5. General characterizations (cid:3) 1(Rd,M) ⊆ L1(N ). By p(Rd,M) ⊆ Lp(N ). Then Corollary p(Rd,M) in [30]. We will use multi-index notation. For m = (m1,··· , md) ∈ Nd Applying the operator-valued Calder´on-Zygmund theory developed in the last section, we will p(Rd,M) can show that the Poisson kernel in the square functions which are used to define hc be replaced by any reasonable test function. As an application, we are able to compare the p(Rd,M) defined in this paper with the operator-valued Hardy operator-valued local Hardy spaces hc spaces Hc 0 and ··· ∂md s = (s1,··· , sd) ∈ Rd, we set sm = sm1 . 5.1. General characterizations. Let Φ be a complex-valued infinitely differentiable function defined on Rd\{0}. Recall that eΓ = {(t, ε) ∈ Rd+1 ε ). For any d), we define the local versions of the conic and radial square functions d . Let m1 = m1 +··· + md and Dm = ∂m1 + : t < ε < 1} and Φε(s) = ε−dΦ( s f ∈ L1(M; Rc of f associated to Φ by d) + L∞(M; Rc ··· smd md d m1 1 ∂s ∂s 1 sc gc Φ(f )(s) =(cid:16)ZZeΓ Φε ∗ f (s + t)2 dtdε εd+1(cid:17) 1 Φ(f )(s) =(cid:16)Z 1 ε (cid:17) 1 0 Φε ∗ f (s)2 dε , s ∈ Rd. 2 2 , s ∈ Rd, The function Φ that we use to characterize the operator-valued local Hardy spaces satisfies the following conditions: (1) Every DmΦ with 0 ≤ m1 ≤ d makes f 7→ sc singular integral operators in Corollary 4.3; (2) There exist functions Ψ, ψ and φ such that DmΦf and f 7→ gc DmΦf Calder´on-Zygmund (5.1) (3) The above Ψ and ψ make dµg = Ψε ∗ g(s)2 dεds ε bφ(ξ)bψ(ξ) +Z 1 QZT (Q) 1 = 1, dε ∀ ξ ∈ Rd; ε and φ ∗ g satisfy: , kψ ∗ gkqo . kgkbmoc 0 bΦ(εξ)bΨ(εξ) dµg(cid:13)(cid:13) 1 2 q 2 q maxn(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 for q > 2; (4) The above φ makes f 7→ φ∗ f a Calder´on-Zygmund singular integral operator in Corollary 4.3. Fix the four functions Φ, Ψ, φ, ψ as above. The following is one of our main results in this section, which states that the functions Φ, φ satisfying the above four conditions give a general characterization for hc Theorem 5.1. Let 1 ≤ p < ∞ and φ, Φ be as above. For any f ∈ L1(M; Rc f ∈ hc Φ(f ) ∈ Lp(N ) and φ ∗ f ∈ Lp(N ) if and only if gc φ ∗ f ∈ Lp(N ). If this is the case, then (5.2) d) + L∞(M; Rc d), Φ(f ) ∈ Lp(N ) and p(Rd,M) if and only if sc p(Rd,M). kfkhc p ≈ ksc Φ(f )kp + kφ ∗ fkp ≈ kgc Φ(f )kp + kφ ∗ fkp 30 R. Xia and X. Xiong with the relevant constants depending only on d, p, and the pairs (Φ, Ψ) and (φ, ψ). One implication of the above theorem is an easy consequence of conditions (1) and (4) that (5.3) (5.4) Φ(f )kp + kφ ∗ fkp . kfkhc ksc kgc Φ(f )kp + kφ ∗ fkp . kfkhc p p . In order to prove the converse inequalities, we need the following lemma, which can be seen as a generalization of Lemma 2.10. Lemma 5.2. Let 1 ≤ p < 2, q be its conjugate index and Φ, φ be the functions satisfying the above assumption. For f ∈ hc p(Rd,M) ∩ L2(N ) and g ∈ bmoc q(Rd,M), (cid:12)(cid:12)τZRd f (s)g∗(s)ds(cid:12)(cid:12) . (ksc Φ(f )kp + kφ ∗ fkp)kgkbmoc q . Proof. The proof of this Lemma is very similar to that of Lemma 2.10, we will just point out the necessary modifications to avoid duplication. We need two auxiliary square functions associated with Φ. For s ∈ Rd, ε ∈ [0, 1], we define (5.5) sc , 2 Φ(f )(s, ε) =(cid:16)Z 1 Φ(f )(s, ε) =(cid:16)Z 1 ε ZB(s,r− ε ε ZB(s, r rd+1(cid:17) 1 2 ) Φr ∗ f (t)2 dtdr rd+1(cid:17) 1 2 ) Φr ∗ f (t)2 dtdr sc . 2 (5.6) By assumption (2) of Φ, we have f (s)g∗(s)ds τZRd = τZRdZ 1 τZRdZ 1 + τZRd 2d cd = 0 def= I + II. ε Φε ∗ f (s)(Ψε ∗ g(s))∗ dsdε 0 ZB(s, ε φ ∗ f (s)(ψ ∗ g(s))∗ds Φε ∗ f (t)sc 2 ) + τZRd φ ∗ f (s)(ψ ∗ g(s))∗ds Φ(f )(s, ε) p−2 2 sc Φ(f )(s, ε) 2−p 2 (Ψε ∗ g(t))∗ dtdε εd+1 ds Then by the Cauchy-Schwarz inequality, I2 . τZRdZ 1 · τZRdZ 1 def= A · B. 0 (cid:0)ZB(s, ε 0 (cid:0)ZB(s, ε εd+1(cid:1)sc 2 ) Φε ∗ f (t)2 dt εd+1(cid:1)sc 2 ) Ψε ∗ g(t)2 dt Φ(f )(s, ε)p−2dεds Φ(f )(s, ε)2−pdεds ∂ε Pε(f ) and ε ∂ Replacing ε ∂ ∂ε Pε(g) in the proof of Lemma 2.10 by Φε ∗ f and Ψε ∗ g respectively and applying Lemma 5.7 and assumption (3) of Ψ and ψ, we get the estimates for the terms A and B that A . ksc Φ(f )kp p and B . kgk2 bmoc qksc Φ(f )k2−p p . The term II is easy to deal with. By the Holder inequality, Lemma 5.7 and assumption (3) again, we get (cid:12)(cid:12)(cid:12)τZRd φ ∗ f (s)(ψ ∗ g(s))∗ds(cid:12)(cid:12)(cid:12) ≤ kφ ∗ fkpkψ ∗ gkq . kφ ∗ fkpkgkbmoc q Combining the estimates for A, B and II, we finally get the desired inequality. . (cid:3) Operator-valued local Hardy spaces 31 We also need the radial version of Lemma 5.2. To this end, we need to majorize the radial square function by the conic one. When we consider the Poisson kernel, this result follows from the harmonicity of the Poisson integral (see Lemma 2.13). However, in the general case, the harmonicity is no longer available. To overcome this difficulty, a more sophisticated inequality has been developped in [51] to compare non-local radial and conic functions. Observe that the result given in [51, Lemma 4.3] is a pointwise one, which also works for the local version of square functions if we consider integration over the interval 0 < ε < 1. The following lemma is an obvious consequence of [51, Lemma 4.3]. Lemma 5.3. Let f ∈ L1(M; Rc d) + L∞(M; Rc d). Then DmΦ(f )(s)2, ∀ s ∈ Rd. sc Lemma 5.4. Let 1 ≤ p < 2. For f ∈ hc p(Rd,M) ∩ L2(N ) and g ∈ bmoc 2 kfk1− p Φ(f )kp + kφ ∗ fkp) hc p q(Rd,M), . 2 p kgkbmoc q Proof. This proof is similar to that of Lemma 5.2 and we keep the notation there. Let f ∈ hc p(Rd,M) with compact support (relative to the variable of Rd). We assume that f is sufficiently nice so that all calculations below are legitimate. Now we need the radial version of sc Φ(f )(s, ε), (cid:12)(cid:12)(cid:12)τZRd gc Φ(f )(s)2 . Xm1≤d f (s)g∗(s)ds(cid:12)(cid:12)(cid:12) . (kgc Φ(f )(s, ε) =(cid:0)Z 1 gc r (cid:1) 1 ε Φr ∗ f (s)2 dr 2 for s ∈ Rd and 0 ≤ ε ≤ 1. By approximation, we can assume that gc (s, ε) ∈ S. By (5.1), (0.12) and the Fubini theorem, we have Φ(f )(s, ε) is invertible for every 2 (cid:12)(cid:12)(cid:12)τZRd f (s)g∗(s)ds(cid:12)(cid:12)(cid:12) . τZRdZ 1 +(cid:12)(cid:12)(cid:12)τZRd φ ∗ f (s)(ψ ∗ g(s))∗ds(cid:12)(cid:12)(cid:12) 0 Φε ∗ f (s)2gc def= A′B′ + II′. 2 Φ(f )(s, ε)p−2 dεds ε · τZRdZ 1 0 Ψε ∗ g(s)2gc Φ(f )(s, ε)2−p dεds ε II′ is treated exactly in the same way as before, pkψ ∗ gk2 pkfk2−p A′ is also estimated similarly as in Lemma 5.2, we have A′ . kgc II′ . kφ ∗ fk2 q . kφ ∗ fkp hc To estimate B′, we notice that the proof of [51, Lemma 1.3] also gives . bmoc q p kgk2 Φ(f )kp p. gc Φ(f )(s, ε)2 . Xm1≤d DmΦ(f )(s, ε)2, sc where sc Lemma 5.7 and inequality (5.3) with DmΦ instead of Φ, we obtain DmΦ(f )(s, ε) is defined by (5.5) with DmΦ instead of Φ. Then by the above inequality, DmΦ(f )(s, ε)2−p dεds ε B′ . Xm1≤d . Xm1≤d . kgk2 τZRdZ 1 0 Ψε ∗ g(s)2sc DmΦ(f )k2−p kgk2 qksc qkfk2−p bmoc hc p p . bmoc Therefore, τZRd which completes the proof. f (s)g∗(s)ds2 . (kgc Φ(f )kp + kφ ∗ fkp)pkfk2−p p kgk2 hc bmoc q , (cid:3) 32 R. Xia and X. Xiong Proof of Theorem 5.1. From Lemmas 5.2, 5.4 and Theorem 2.11, we conclude that if 1 ≤ p ≤ 2, we have p kfkhc kfkhc p . ksc . kgc Φ(f )kp + kφ ∗ fkp, Φ(f )kp + kφ ∗ fkp. For the case 2 < p < ∞, by Theorem 2.19, we can choose g ∈ hc index of p) with norm one such that q(Rd,M) (with q the conjugate p ≈ τZRd kfkhc Then by the Holder inequality and (5.4) (applied to g, Ψ and q), Φε ∗ f (s) · (Ψε ∗ g(s))∗ dsdε f (s)g∗(s)ds = τZRdZ 1 ε 0 + τZRd φ ∗ f (s)(ψ ∗ g(s))∗ds. kfkhc p Similarly, we have . kgc . (kgc Ψ(g)kq + kφ ∗ fkpkψ ∗ gkq Φ(f )kpkgc Φ(f )kp + kφ ∗ fkp)kgkhc q = kgc Φ(f )kp + kφ ∗ fkp. kfkhc p . ksc Φ(f )kp + kφ ∗ fkp. Therefore, combined with (5.4) and (5.3), we have proved the assertion. (cid:3) The rest part of this subsection is devoted to explaining how Theorem 5.1 generalizes the char- acterization of hc p(Rd,M). Firstly and most naturally, we show how Theorem 5.1 covers the original definition of hc p(Rd,M). Let us take Φ = −2πI(P) and φ = P for example. A simple calculation shows that we can choose Ψ = −8πI(P) and ψ = 4πI(P) + P to fulfil (5.1). By the inverse Fourier transform formula, we have −2πf ∗ I(P)ε(t) = −2πZ e2πit·ξbf (ξ)εξe−2πεξdξ ∂εZ e2πit·ξbf (ξ)e−2πεξdξ = ε = ε ∂ ∂ ∂ε Pε(f )(t). So we return back to the original definition of hc p(Rd,M). Theorem 5.1 implies that Φ(f )kp + kφ ∗ fkp ≈ kgc In particular, we have the following equivalent norm of hc p ≈ ksc kfkhc Φ(f )kp + kφ ∗ fkp. p(Rd,M): Corollary 5.5. Let 1 ≤ p < ∞. Then for any f ∈ hc p(Rd,M), we have kfkhc p ≈ kgc(f )kp + kP ∗ fkp. Secondly, consider Φ to be a Schwartz function on Rd satisfying: (5.7) (Φ is of vanishing mean; Φ is nondegenerate in the sense of (0.5). Set Φε(s) = ε−dΦ( s ε ) for ε > 0. In the sequel, we will show that every Schwartz function satisfying (5.7) fulfils the four conditions in the beginning of this subsection. So they all can be used to characterize hp(Rd,M). of vanishing mean such that It is a well-known elementary fact (ef. e.g. [44, p. 186]) that there exists a Schwartz function Ψ (5.8) Lemma 5.6. R 1 at the origin as 0. Z ∞ 0 bΦ(ε·)bΨ(ε·) dε 0 dε ε = 1, ∀ ξ ∈ Rd \ {0} . bΦ(εξ)bΨ(εξ) ε is an infinitely differentiable function on Rd if we define its value with the remainder of integral form equal to Z 1 Since bΦ(0) = 0, the above Taylor series implies that (N + 1)εN +1 Rγ(εξ) = γ! εγ1ξγ Similarly, we have bΦ(εξ) = X1≤γ1≤N bΨ(εξ) = X1≤β1≤N DγbΦ(0) DβbΨ(0) Rγ(εξ) ξγ . R′β(εξ) ξβ, εβ1 ξβ 0 γ! β! (1 − θ)N DγbΦ(θεξ)dθ . + Xγ1=N +1 + Xβ1=N +1 0 bΦ(εξ)bΨ(εξ) dε 1 bΦ(ε·)bΨ(ε·) dε Operator-valued local Hardy spaces 33 Proof. To prove the assertion, it suffices to show thatR 1 the origin. Given ε ∈ (0, 1], we expand bΦ(ε·) in the Taylor series at the origin Rγ(εξ) ξγ , εγ1 ξγ 0 bΦ(ε·)bΨ(ε·) dε + Xγ1=N +1 bΦ(εξ) = Xγ1≤N DγbΦ(0) γ! ε is infinitely differentiable at ε (and the integrals of arbitrary 0 bΦ(εξ)bΨ(εξ) dε ε is infinitely differentiable at the origin ξ = 0. where R′β is the integral form remainder of bΨ. Thus, both bΦ(εξ) and bΨ(εξ) contain only powers of ε with order at least 1. Therefore, the integralR 1 order derivatives of bΦ(εξ) and bΨ(εξ)) converge uniformly for ξ ∈ Rd close to the origin. We then obtain thatR 1 It follows immediately from Lemma 5.6 thatR ∞ its value at the origin by 1. Then we can find two other functions φ, ψ such that bφ,bψ ∈ H σ bφ(0) > 0,bψ(0) > 0 and ∀ ξ ∈ Rd; Indeed, for β > 0 large enough, the function (1 + · 2)−β belongs to H σ if F ∈ S(Rd), the function (1 + · 2)β F is still in H σ the case 1 ≤ p ≤ 2. Let H = L2((0, 1), dε Φ·(s) : ε 7→ Φε(s). Then we can check that Now let show that conditions (1) and (4) hold for Φ, φ satisfying (5.7). First, we deal with ε ). Define the kernel k : Rd → H by k(s) = Φ·(s) with ε is a Schwartz function if we define 2 (Rd), bφ(ξ)bψ(ξ) +Z 1 0 bΦ(εξ)bΨ(εξ) 2 (Rd). Thus we obtain (5.1). 2 (Rd). On the other hand, dε ε (5.9) = 1, (cid:3) and that sup ξ∈Rd kbΦ(εξ)kH < ∞, kΦε(s)kH . 1 sd+1 , ∀ s ∈ Rd \ {0} k∇Φε(s)kH . , ∀ s ∈ Rd \ {0}. 1 sd+1 Thus, k satisfies the assumption of Corollary 4.3. By Remark 4.4, we have, for any 1 ≤ p ≤ 2, kΦε ∗ fkLp(N ;Hc) = kgc Φ(f )kp . kfkhc p . Φ is similar. In this case, we take the Hilbert space H = L2(eΓ, dtdε 2 (Rd) implies φ ∈ L1(Rd), then kφ ∗ fkLp(N ) . kfkLp(N ) . kfkhc εd+1 ). On the p. Thus, combining the above estimates, we obtain The treatment of sc other hand, bφ ∈ H σ kgc Φ(f )kp + kφ ∗ fkp . kfkhc ksc Φ(f )kp + kφ ∗ fkp . kfkhc . p p Then, a simple duality argument using (5.1) and Theorem 2.19 gives the above inequalities for the case p > 2. Moreover, it is obvious that if we replace Φ by DmΦ, the above two inequalities still hold for any 1 ≤ p < ∞. In the end, it remains to check the condition (3) for Ψ, ψ obtained in (5.8) and (5.9). This can be done by showing a Carleson measure characterization of bmoc q by general test functions. The proof of the following lemma has the same pattern with that of Lemma 1.5, so is left to the reader. 34 R. Xia and X. Xiong Lemma 5.7. Let 2 < q ≤ ∞, g ∈ bmoc ε . Then dµg is an M-valued q-Carleson measure on the strip Rd × (0, 1). Furthermore, let ψ be any function on Rd such that q(Rd,M) and dµg = Ψε ∗ g(s)2 dsdε (5.10) We have bψ ∈ H σ QZT (Q) 1 maxn(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 2 (Rd) with σ > d 2 . , kψ ∗ gkqo . kgkbmoc q . 2 q 2 dµg(cid:13)(cid:13) 1 Remark 5.8. It is worthwhile to note that, if Ψ and ψ are determined by (5.8) and (5.9), the opposite of the above lemma is also true. This can be deduced by a similar argument as that of Corollary 2.12; we omit the details. By the discussion above, we deduce the following corollary from Theorem 5.1. Corollary 5.9. Let Φ be the Schwartz function on Rd satisfying (5.7) and φ be the function given by (5.9). Then for any 1 ≤ p < ∞, we have (5.11) Φ(f )kp + kφ ∗ fkp ≈ kgc with the relevant constants depending only on d, p, Φ and φ. p ≈ ksc kfkhc Φ(f )kp + kφ ∗ fkp 5.2. Discrete characterizations. In this subsection, we give a discrete characterization for operator- valued local Hardy spaces. To this end, we need some modifications of the four conditions in the beginning of last subsection. The square functions sc Φ(f ) can be discretized as follows: Φ(f ) and gc gc,D Φ (f )(s) =(cid:16)Xj≥1 Φ (f )(s) =(cid:16)Xj≥1 sc,D 2 , Φj ∗ f (s)2(cid:17) 1 2djZB(s,2−j ) Φj ∗ f (t)2dt(cid:17) 1 2 . Here Φj is the inverse Fourier transform of Φ(2−j·). This time, to get a resolvent of the unit on Rd, we need to assume that Φ, Ψ, φ, ψ satisfy (5.12) ∞Xj=1bΦ(2−jξ)bΨ(2−jξ) +bφ(ξ)bψ(ξ) = 1, ∀ ξ ∈ Rd. In brief, the complex-valued infinitely differentiable function Φ considered in this subsection satis- fies: (1) Every DmΦ with 0 ≤ m1 ≤ d makes f 7→ sc,D (2) There exist functions Ψ, ψ and φ that fulfil (5.12); (3) The above Ψ and ψ make dµD singular integral operators in Corollary 4.3; DmΦf and f 7→ gc,D DmΦf Calder´on-Zygmund f =Pj≥1 Ψj ∗ f (s)2ds × dδ2−j (ε) (with δ2−j (ε) the unit Dirac QZT (Q) , kψ ∗ fkqo . kfkbmoc for q > 2; 2 q 2 q mass at the point 2−j) and φ ∗ f satisfy: dµD 1 maxn(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 f(cid:13)(cid:13) 1 (4) The above φ makes f 7→ φ∗ f a Calder´on-Zygmund singular integral operator in Corollary 4.3. Remark 5.10. Any Schwartz function that has vanishing mean and is nondegenerate in the sense of (0.5) satisfies all the four conditions above. The following discrete version of Theorem 5.1 will play a crucial role in the study of operator- valued Triebel-Lizorkin spaces on Rd in our forthcoming paper [49]. Now we fix the pairs (Φ, Ψ) and (φ, ψ) satisfying the above four conditions. Theorem 5.11. Let 1 ≤ p < ∞. Then for any f ∈ L1(M; Rc and only if sc,D Moreover, d) + L∞(M; Rc p(Rd,M) if Φ (f ) ∈ Lp(N ) and φ∗ f ∈ Lp(N ). Φ (f ) ∈ Lp(N ) and φ∗ f ∈ Lp(N ) if and only if gc,D d), f ∈ hc kfkhc p ≈ ksc,D Φ (f )kp + kφ ∗ fkp ≈ kgc,D Φ (f )kp + kφ ∗ fkp Operator-valued local Hardy spaces 35 with the relevant constants depending only on d, p, and the pairs (Φ, Ψ) and (φ, ψ). The following paragraphs are devoted to the proof of Theorem 5.11 which is similar to that of Theorem 5.1. We will just indicate the necessary modifications. We first prove the discrete counterparts of Lemmas 5.2 and 5.4. Lemma 5.12. Let 1 ≤ p < 2 and q be the conjugate index of p. For any f ∈ hc and g ∈ bmoc p(Rd,M) ∩ L2(N ) q(Rd,M), (cid:12)(cid:12)(cid:12)τZRd f (s)g∗(s)ds(cid:12)(cid:12)(cid:12) .(cid:0)ksc,D Φ (f )kp + kφ ∗ fkp(cid:1)kgkbmoc q . Proof. First, note that by (5.12), we have τZRd f (s)g∗(s)ds = τZRdXj≥1 Φj ∗ f (s)(cid:0)Ψj ∗ g(s)(cid:1)∗ ds + τZRd φ ∗ f (s)(cid:0)ψ ∗ g(s)(cid:1)∗ds. The second term on the right hand side of the above formula is exactly the same as the corre- sponding term II in the proof of Lemma 5.2. Now we need the discrete versions of sc Φ: For j ≥ 1, s ∈ Rd, let Φ and sc sc,D Φ (f )(s, j) =(cid:16) X1≤k≤j Φ (f )(s, j) =(cid:16) X1≤k≤j sc,D 2dkZB(s,2−k−2−j−1) Φj ∗ f (t)2dt(cid:17) 1 2dkZB(s,2−k−1) Φj ∗ f (t)2dt(cid:17) 1 . 2 2 Φ (f )(s, j) and sc,D Denote sc,D Φ (f )(s, j) simply by s(s, j) and s(s, j), respectively. By approximation, we may assume that s(s, j) and s(s, j) are invertible for every s ∈ Rd and j ≥ 1. By the Cauchy- Schwarz inequality, 2 2d cd (cid:12)(cid:12)(cid:12)τZRdXj≥1 Φj ∗ f (s)(cid:0)Ψj ∗ g(s)(cid:1)∗ ds(cid:12)(cid:12)(cid:12) 2djZB(s,2−j−1) τZRdXj Φj ∗ f (t)(cid:0)Ψj ∗ g(t)(cid:1)∗ dt ds(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) s(s, j)p−2(cid:16)2djZB(s,2−j−1)) Φj ∗ f (t)2 dt(cid:17)ds . τZRdXj s(s, j)2−p(cid:16)2djZB(s,2−j−1) Ψj ∗ g(t)2 dt(cid:17)ds · τZRdXj def= A · B. 2 The term A is less easy to estimate than the corresponding term A in the proof of Lemma 5.2. To deal with it we simply set sj = s(s, j) and s = s(s, +∞) ≤ sc,D(f )(s). Then sp−2 j j − s2 (s2 j−1)ds sp−2 j (s2 j − s2 j−1)ds A = τZRdXj≥1 ≤ τZRdXj = τZRdXj τZRdXj (sj − sj−1)ds + τZRdXj where s0 = 0. Since 1 ≤ p < 2, sp−1 j ≤ sp−1, we have (sj − sj−1)ds . τZRd sp−1 j sp−2 j sj−1(sj − sj−1)ds, spds ≤ ksc,D(f )kp p. 36 R. Xia and X. Xiong On the other hand, τZRdXj τZRdXj sp−2 j sp−2 j sj−1(sj − sj−1)ds = τZRdXj sj−1(sj − sj−1)ds ≤ τZRdXj 2 sp−2 1−p s since sj ≥ sj−1 for any j ≥ 1, we have s j p−1 s 1−p 2 sp−2 j 1−p 2 s sj−1s p−1 2 (sj − sj−1)s p−1 2 ds, 1−p sj−1s 2 ≤ 1. Thus, by the Holder inequality, Φ (f )kp p. spds ≤ ksc,D 2 ds = τZRd p−1 2 (sj − sj−1)s Combining the preceding inequalities, we get the desired estimate of A: A ≤ 2ksc,D Φ (f )kp p. The estimate of the term B is, however, almost identical to that of B in the proof of Lemma 5.2. There are only two minor differences. The first one concerns the square function Sc(f )(s, j) in (2.8): it is now replaced by Sc(f )(s, j) =(cid:16) X1≤k≤j 2dkZB(cm,j ,2−k) Φj ∗ f (t)2 dt(cid:17) 1 2 if s ∈ Qm,j. Then we have s(s, j) ≤ Sc(f )(s, j). The second difference is about the Carleson characterization of bmoc g . Apart from these two differences, the remainder of the argument is identical to that in the proof of Lemma 5.2. (cid:3) q; we now use its discrete analogue, namely, dµD Lemma 5.13. Let 1 ≤ p < 2 and f ∈ hc (cid:12)(cid:12)(cid:12)τZRd f (s)g∗(s)ds(cid:12)(cid:12)(cid:12) .(cid:16)kgc,D Proof. We use the truncated version of gc,D q(Rd,M). Then kgkbmoc q . 2 2 p(Rd,M) ∩ L2(N ), g ∈ bmoc Φ (f )kp + kφ ∗ fkp(cid:17) p kfk1− p Φk ∗ f (s)2(cid:17) 1 Φ (f ): hc p . 2 The proof of [51, Lemma 4.3] is easily adapted to the present setting to ensure Then where (cid:12)(cid:12)(cid:12)τZRd gc,D gc,D Φ (f )(s, j) =(cid:16)Xk≤j Φ (f )(s, j)2 . Xm1≤d f (s)g∗(s)ds(cid:12)(cid:12)(cid:12) I′ = τZRdXj II′ = τZRdXj 2 sc,D DmΦ(f )(s, j)2 . ≤ I′ · II′ +(cid:12)(cid:12)(cid:12)τZ φ ∗ f (s)(cid:0)ψ ∗ g(s)(cid:1)∗ds(cid:12)(cid:12)(cid:12), gc,D Φ (f )(s, j)p−2Φj ∗ f (s)2ds , gc,D Φ (f )(s, j)2−pΨj ∗ g(s)2ds . Both terms I′ and II′ are estimated exactly as before, so we have p kgk2 p and II′ . kfk2−p I′ ≤ 2kgc This gives the announced assertion. Φ(f )kp hc bmoc q . (cid:3) Armed with the previous two lemmas and the Calder´on-Zygmund theory in section 4, we can prove Theorem 5.11 in the same way as Theorem 5.1. Details are left to the reader. We also include a discrete Carleson measure characterization of bmoc q by general test functions. Much as the characterization in Lemma 5.7 and Remark 5.8, it is a byproduct of the proof of Theorem 5.11. Operator-valued local Hardy spaces 37 Corollary 5.14. Let 2 < q ≤ ∞, ψ and Ψ be given in (5.12). Assume further 2 (Rd) with σ > d 2 . Then for every g ∈ bmoc kgkbmoc q, we have bψ ∈ H σ q ≈ kψ ∗ gkq +(cid:13)(cid:13)(cid:13) sup + s∈Q⊂Rd Q<1 1 QZQ Xj≥− log2(l(Q)) Ψj ∗ g(s)2ds(cid:13)(cid:13)(cid:13) 1 2 q 2 . 5.3. The relation between hp(Rd,M) and Hp(Rd,M). Due to the noncommutativity, for any 1 < p < ∞ and p 6= 2, the column operator-valued local Hardy space hc p(Rd,M) and the column operator-valued Hardy space Hc p(Rd,M) are not equivalent. On the other hand, if we consider the mixture spaces hp(Rd,M) and Hp(Rd,M), then we will have the same situation as in the classical case. Since kP ∗ fkp . kfkp . kfkHc p for any 1 ≤ p ≤ 2, we deduce the inclusion (5.13) Hc p(Rd,M) ⊂ hc p(Rd,M) for 1 ≤ p ≤ 2. Then by the duality obtained in Theorem 2.19, we have (5.14) p(Rd,M) ⊂ Hc hc p(Rd,M) for 2 < p < ∞. However, we can see from the following proposition that we do not have the inverse inclusion of (5.13) nor (5.14). p(Rd,M), Let 2 < p < ∞. If for any f ∈ Hc (5.15) Proposition 5.15. Let φ be a function on Rd such that bφ(0) ≥ 0 and bφ ∈ H σ then we must have bφ(0) = 0. Proof. We prove the assertion by contradiction. Suppose that there exists φ such that bφ(0) > 0, bφ ∈ H σ 2 (Rd) and (5.15) holds for any f ∈ Hc p(Rd,M). Since both Hc p(Rd,M) and Lp(N ) are homogeneous spaces, we have, for any ε > 0, kφ ∗ fkp . kfkHc 2 (Rd) with σ > d 2 . , p kφ ∗ f (ε·)kp = k(φε ∗ f )(ε·)kp = ε− d pkφε ∗ fkp and kf (ε·)kHc p = ε− d pkfkHc p . This implies that (5.16) kφε ∗ fkp . kfkHc p , for any ε > 0 with the relevant constant independent of ε. Now we consider a function f ∈ Lp(N ) M and such that suppbf is compact, i.e. there exists a positive real number which takes values in S+ N such that suppbf ⊂ {ξ ∈ Rd : ξ ≤ N}. Since bφ(0) > 0, we can find ε0 > 0 and c > 0 such that bφ(ε0ξ) ≥ c whenever ξ ≤ N . Thus, in this case, kφε ∗ fkp ≥ ckfkp. Then by (5.16), we have which leads to a contradiction when p > 2. Therefore, bφ(0) = 0. p-norm and the duality in Theorem 2.19, we get the following result: By the definition of the hc kfkp . kfkHc (cid:3) , p Corollary 5.16. Let 1 ≤ p < ∞ and p 6= 2. hc p(Rd,M) and Hc p(Rd,M) are not equivalent. Although hc p(Rd,M) and Hc p(Rd,M) do not coincide when p 6= 2, for those functions whose Fourier transforms vanish at the origin, their hc p-norms and Hc p-norms are still equivalent. Theorem 5.17. Let φ ∈ S such thatRRd φ(s)ds = 1. (1) If 1 ≤ p ≤ 2 and f ∈ hc (2) If 2 < p < ∞ and f ∈ Hc p(Rd,M), then f − φ ∗ f ∈ Hc p(Rd,M), then f − φ ∗ f ∈ hc p(Rd,M) and kf − φ ∗ fkHc p(Rd,M) and kf − φ ∗ fkhc p p . kfkhc . . kfkHc p p . 38 R. Xia and X. Xiong p(Rd,M) and Φ be a nondegenerate Schwartz function with vanishing mean. Φ(f−φ∗f )kp. p(Rd,M) in Lemma 0.1, kf−φ∗fkHc p(Rd,M) ≈ kGc Proof. (1) Let f ∈ hc By the general characterization of Hc Let us split kGc kGc Φ(f − φ ∗ f )kp into two parts: Φ(f − φ ∗ f )kp p . . 1 1 .(cid:13)(cid:13)(cid:13)(cid:0)Z 1 =(cid:13)(cid:13)(cid:13)(cid:0)Z 1 +(cid:13)(cid:13)(cid:13)(cid:0)Z ∞ +(cid:13)(cid:13)(cid:13)(cid:0)Z ∞ 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 0 Φε ∗ (f − φ ∗ f )2 dε 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 0 Φε ∗ (f − φ ∗ f )2 dε p(Rd,M), thus we p(Rd,M). Then by Theorem 5.1, this term can be majorized from above by 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 Φε ∗ (f − φ ∗ f )2 dε 2(cid:13)(cid:13)(cid:13)p ε (cid:1) 1 (Φε − Φε ∗ φ) ∗ f2 dε In order to estimate the first term in the last equality, we notice that φ ∗ f ∈ hc have f − φ ∗ f ∈ hc kfkhc To deal with the second term, we express it as a Calder´on-Zygmund operator with Hilbert-valued kernel. Let H = L2((1, +∞), dε ε ) and define the kernel k : Rd → H by k(s) = Φ·(s) − Φ· ∗ φ(s) (Φ·(s) being the function ε 7→ Φε(s)). Now we prove that k satisfies the hypotheses of Corollary 4.3. The condition (1) of that corollary is easy to verify. So we only check the conditions (2) and (3) there. By the fact thatRRd φ(s)ds = 1 and the mean value theorem, we have (cid:1)(cid:12)(cid:12)φ(t) dt. ε (cid:17) 1 (cid:1)(cid:12)(cid:12)φ(t) dt(cid:1)2 dε ε (cid:17) 1 (cid:1)(cid:12)(cid:12)φ(t) dt(cid:1)2 dε [Φε(s) − Φε(s − t)] φ(t)dt(cid:12)(cid:12)(cid:12) 0<θ<1(cid:12)(cid:12)∇Φ(cid:0) s − θt 0<θ<1(cid:12)(cid:12)∇Φ(cid:0) s − θt 0<θ<1(cid:12)(cid:12)∇Φ(cid:0) s − θt (cid:12)(cid:12)(Φε − Φε ∗ φ)(s)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)ZRd ≤ZRd t (cid:0)Zt< s (cid:0)Zt> s (cid:13)(cid:13)(Φ· − Φ· ∗ φ)(s)(cid:13)(cid:13)H .(cid:16)Z ∞ +(cid:16)Z ∞ Then we split the last integral into two parts: 1 εd+1 sup 1 εd+1 sup εd+1 t t sup 1 ε ε ε 1 2 2 2 2 1 def= I + II. 2 2 ε ) . εd+ 1 sd+ 1 If t < s2 , we have s − θt ≥ s2 , thus ∇Φ( s−θt dε(cid:1) 1 sd+ 1 2 tφ(t) dt . 1 sd+ 1 ε (cid:1) 1 1 I .(cid:0)Z ∞ When t > s2 , since φ ∈ S, we haveRt> s II .(cid:0)Z ∞ sd+ 1 sd+ 1 ε2d+2 1 ε2 2 · dε . . 1 1 1 1 2 2 2 2 2 The estimates of I and II imply . Hence 1 . sd+ 1 2 for any 0 ≤ θ ≤ 1. Then 1 . k(Φε − Φε ∗ φ)(s)kH . . 1 sd+ 1 2 In a similar way, we obtain k∇(Φε − Φε ∗ φ)(s)kH . Thus, it follows from Corollary 4.3 that (cid:13)(cid:13)(cid:13)(R ∞ above by kfkhc 1 . p 1 sd+1 . 1 2(cid:13)(cid:13)(cid:13)p (Φε − Φε ∗ φ) ∗ f2 dε ε ) is also majorized from (2) The case p > 2 can be deduced from the duality between hc p and hc between Hc p and Hc q (q being the conjugate index of p). There exists g ∈ hc q (Theorem 2.19) and that q(Rd,M) with norm Operator-valued local Hardy spaces one such that kf − φ ∗ fkhc p =(cid:12)(cid:12)(cid:12)τZRd =(cid:12)(cid:12)(cid:12)τZRd ≤ kfkHc (f − φ ∗ f )(s)g∗(s)ds(cid:12)(cid:12)(cid:12) f (s)(g∗ − φ ∗ g∗)(s)ds(cid:12)(cid:12)(cid:12) pkg − φ ∗ gkHc . kfkHc q pkgkhc q = kfkHc p which completes the proof. 39 (cid:3) , From the interpolation result of mixture local hardy spaces in Proposition 3.4, we can deduce the equivalence between mixture local Hardy spaces and Lp-spaces. Proposition 5.18. For any 1 < p < ∞, hp(Rd,M) = Hp(Rd,M) = Lp(N ) with equivalent norms. Proof. It is known that Hp(Rd,M) = Lp(N ) with equivalent norms. One can see [30, Corollary 5.4] for more details. One the other hand, since L∞(N ) ⊂ bmoc(Rd,M), by duality, we get 1(Rd,M) ⊂ L1(N ). Combining (1.2) and the interpolation result in Theorem 3.3, we deduce that hc hc p(Rd,M) ⊂ Lp(N ) for any 1 < p ≤ 2 and Lp(N ) ⊂ hc p(Rd,M) for any 2 < p < ∞. Similarly, we also have hr p(Rd,M) for any 2 < p < ∞. Combined with (5.13) and (5.14), we get p(Rd,M) ⊂ Lp(N ) for any 1 < p ≤ 2 and Lp(N ) ⊂ hr Hp(Rd,M) ⊂ hp(Rd,M) ⊂ Lp(N ) for 1 < p ≤ 2, Lp(N ) ⊂ hp(Rd,M) ⊂ Hp(Rd,M) Then (5.17), (5.18) and [30, Corollary 5.4] imply that which completes the proof. hp(Rd,M) = Hp(Rd,M) = Lp(N ) for 2 < p < ∞. for 1 < p < ∞, (cid:3) 6. The atomic decomposition 1(Rd,M). The atomic decomposition of In this section, we give the atomic decomposition of hc 1(Rd,M) studied in [30] and the characterizations obtained in the last section will be the main Hc tools for us. Definition 6.1. Let Q be a cube in Rd with Q ≤ 1. If Q = 1, an hc a function a ∈ L1(M; Lc 1-atom associated with Q is 2(Rd)) such that (5.17) and (5.18) • supp a ⊂ Q; 2 ≤ Q− 1 2 . If Q < 1, we assume additionally: • τ(cid:0)RQ a(s)2ds(cid:1) 1 • RQ a(s)ds = 0. Let hc 1,at(Rd,M) be the space of all f admitting a representation of the form f = λj aj, ∞Xj=1 where the aj's are hc the sense of distribution. We equip hc 1-atoms and λj ∈ C such thatP∞j=1 λj < ∞. The above series converges in 1,at(Rd,M) with the following norm: kfkhc 1,at = inf{ λj : f = λjaj; aj's are hc 1 -atoms, λj ∈ C}. ∞Xj=1 ∞Xj=1 Similarly, we define the row version hr 1,at(Rd,M). Then we set Theorem 6.2. We have hc h1,at(Rd,M) = hc 1,at(Rd,M) = hc 1,at(Rd,M) + hr 1(Rd,M) with equivalent norms. 1,at(Rd,M). 40 R. Xia and X. Xiong Proof. First, we show the inclusion hc that for any atom a in Definition 6.1, we have 1,at(Rd,M) ⊂ hc 1(Rd,M). To this end, it suffices to prove (6.1) Recall that the atomic decomposition of Hc function b ∈ L1(M; Lc 2(Rd)) such that, for some cube Q, 1 . 1. kakhc 1(Rd,M) has been considered in [30]. An Hc 1-atom is a • supp b ⊂ Q; • RQ b(s)ds = 0; • τ(cid:0)RQ b(s)2ds(cid:1) 1 2 ≤ Q− 1 2 . If a is supported in Q with Q < 1, then a is also an Hc . 1. Now assume that the supporting cube Q of a is of side length one. We use the discrete characterization obtained in Theorem 5.11, i.e. 1-atom, so kakhc . kakHc 1 1 Apart from the assumption on Φ and φ in Theorem 5.1, we may take Φ and φ satisfying kakhc ∞Xj=1 1 ≈(cid:13)(cid:13)( Φj ∗ a2) 1 2(cid:13)(cid:13)1 + kφ ∗ ak1. Then supp Φ, supp φ ⊂ B1 = {s ∈ Rd : s ≤ 1}. By the Cauchy-Schwarz inequality we have supp φ ∗ a ⊂ 3Q and supp Φε ∗ a ⊂ 3Q for any 0 < ε < 1. Similarly, 2 dt . 1. 1 ( Φj ∗ a2) 2 · τ(cid:0)Z a(s)2ds(cid:1) 1 kφ ∗ ak1 ≤Z3Q(cid:0)ZQ φ(t − s)2ds(cid:1) 1 2(cid:13)(cid:13)1 = τZ3Q ∞Xj=1 ∞Xj=1 (cid:13)(cid:13)( . τ(cid:16)Z3Q ∞Xj=1 = τ(cid:16)ZRd ∞Xj=1 ≤ τ(cid:0)Z a(s)2ds(cid:1) 1 Φj ∗ a(s)2) Φj ∗ a(s)2ds(cid:17) 1 bΦ(2−jξ)ba(ξ)2dξ(cid:17) 1 2 ≤ 1. 1 2 ds 2 2 Therefore, hc 1,at(Rd,M) ⊂ hc 1(Rd,M). the atomic decomposition of Hc every continuous functional ℓ on hc Now we turn to proving the reverse inclusion. Observe that Hc over, since for any cube Q with side length one, L1(cid:0)M; Lc continuous functional on L1(cid:0)M; Lc 1(Rd,M) and the duality between Hc 1-atoms. Then by 1(Rd,M) and BMOc(Rd,M), 1,at(Rd,M) corresponds to a function g ∈ BMOc(Rd,M). More- 1,at(Rd,M), ℓ induces a 2(Q)(cid:1) with norm less than or equal to kℓk(hc 1,at)∗ . Thus, the 2(Q)(cid:1) ⊂ hc function g satisfies the condition that 1-atoms are also hc (6.2) g ∈ BMOc(Rd,M) and kgQkL∞(M;Lc 2(Q)) ≤ kℓk(hc 1,at)∗. sup Q⊂Rd Q=1 Consequently, g ∈ bmoc(Rd,M) and ℓ(f ) = τZRd 1,at(Rd,M)∗. Thus, hc 1(Rd,M) densely, we deduce that hc Thus, hc bmoc(Rd,M) ⊂ hc hc 1,at(Rd,M) ⊂ hc norms. 1,at(Rd,M)∗ ⊂ bmoc(Rd,M). On the other hand, by the previous result, we have 1,at(Rd,M)∗ = bmoc(Rd,M) with equivalent norms. Since 1(Rd,M) with equivalent (cid:3) f (s)g∗(s)ds, ∀ f ∈ hc 1,at(Rd,M) = hc 1,at(Rd,M). Operator-valued local Hardy spaces 41 Acknowledgements. The authors are greatly indebted to Professor Quanhua Xu for having suggested to them the subject of this paper, for many helpful discussions and very careful reading of this paper. The authors are partially supported by the the National Natural Science Foundation of China (grant no. 11301401). References [1] T. Bekjan, Z. Chen, M. Perrin and Z. Yin. Atomic decomposition and interpolation for Hardy spaces of non- commutative martingales. J. Funct. Anal. 258 (2010), no. 7, 2483-2505. [2] J. Bergh and J. Lofstrom. Interpolation Spaces: An Introduction. Springer, Berlin, 1976. [3] Z. Chen, Q. Xu and Z. Yin. Harmonic analysis on quantum tori. Comm. Math. Phys. 322 (2013), no. 3, 755-805. [4] R. Coifman, Y. Meyer and E. M. Stein. Some new function spaces and their applications to Harmonic analysis. J. Funct. Anal. 62 (1985), no. 2, 304-335. [5] R. Coifman and G. Weiss. Extensions of Hardy spaces and their use in analysis. Bull. Amer. Math. Soc. 83 (1977), no. 4, 569-645. [6] J. Conde. A note on dyadic coverings and nondoubling Caldern-Zygmund theory. J. Math. Anal. Appl. 397 (2013), 785-790. [7] R. E. Edwards and G. I. Gaudry. Littlewood-Paley and multiplier theory. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 90. Springer-Verlag, Berlin-New York, 1977. [8] C. Fefferman and E. M. Stein. Hp spaces of several variables. Acta Math. 129 (1972), no. 3-4, 137-193. [9] M. Frazier, R. Torres and G. Weiss. The boundedness of Caldern-Zygmund operators on the spaces F α,q p . Rev. Mat. Iberoam. Vol. 4 (1988), 41-72. [10] J. Garca-Cuerva and J.L. Rubio de Francia. Weighted norm inequalities and related topics. North-Holland Mathematics Studies, 116. Notas de Matemtica, 104. North-Holland Publishing Co., Amsterdam, 1985. x+604 pp. [11] J. B. Garnett. Bounded analytic functions. Pure and Applied Mathematics, 96. Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York-London, 1981. xvi+467 pp. [12] D. Goldberg. A local version of real Hardy spaces. Duke Math. J. 46 (1979), no. 1, 27-42. [13] L. Grafakos. Classical Fourier analysis. Second Edition. Springer-Verlag, New York, 2008. [14] G. Hong, L. D. L´opez-Snchez, J. M. Martell, and J. Parcet. Caldern-Zygmund operator associated to matrix- valued kernels. Int. Math. Res. Not. 2014, no. 5, 1221-1252. [15] G. Hong and T. Mei. John-Nirenberg inequality and atomic decomposition for noncommutative martingales. J. Funct. Anal. 263 (2012), no. 4, 1064-1097. [16] T. Hytonen. The real-variable Hardy space and BMO. Lecture notes of a course at the University of Helsinki, Winter 2010. [17] Y. Jiao, F. Sukochev, D. Zanin and D. Zhou. Johnson-Schechtman inequalities for noncommutative martingales. J. Funct. Anal. 272 (2017), no. 3, 976-1016. [18] M. Junge. Doob's inequality for non-commutative martingales. J. Reine Angew. Math. 549 (2002), 149-190. [19] M. Junge and M. Musat. A noncommutative version of the John-Nirenberg theorem. Trans. Amer. Math. Soc. 359 (2007), no. 1, 115-142. [20] M. Junge, C. Le Merdy and Q. Xu. H∞-functional calculus and square functions on noncommutative Lp-spaces. Astrisque No. 305 (2006), vi+138 pp. [21] M. Junge and Q. Xu. Non-commutative Burkholder/Rosenthal Inequalities. Ann. Prob. 31 (2003), no. 2, 948- 995. [22] M. Junge and Q. Xu. Noncommutative Burkholder/Rosenthal inequalities. II. Applications. Israel J. Math. 167 (2008), 227-282. [23] M. Junge and Q. Xu. Noncommutative maximal ergodic theorems. J. Amer. Math. Soc. 20 (2007), 385-439. [24] M. Junge and T. Mei. Noncommutative Riesz transforms - a probabilistic approach. Amer. J. Math. 132 (2010), 611-681. [25] M. Junge and T. Mei. BMO spaces associated with semigroups of operators. Math. Ann. 352 (2012), 691-743. [26] M. Junge and T. Mei and J. Parcet. Smooth Fourier multipliers on group von Neumann algebras. GAFA. 24 (2014), 1913-1980. [27] C. Lance. Hilbert C ∗-modules. A toolkit for operator algebraists. London Mathematical Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995. x+130 pp. [28] F. Lust-Piquard and G. Pisier. Noncommutative Khintchine and Paley Inequalities. Ark. Mat. 29 (1991), 241- 260. [29] T. Mei. BMO is the intersection of two translates of dyadic BMO. C. R. Math. Acad. Sci. Paris 336 (2003), no. 12, 1003-1006. [30] T. Mei. Operator-valued Hardy Spaces. Mem. Amer. Math. Soc. 188 (2007), no. 881, vi+64 pp. [31] T. Mei. Tent spaces associated with semigroups of operators. J. Funct. Anal. 255 (2008), no. 12, 3356-3406. [32] M. Musat. Interpolation between non-commutative BMO and non-commutative Lp-spaces. J. Funct. Anal. 202 (2003), no. 1, 195-225. [33] J. Parcet. Pseudo-localization of singular integrals and noncommutative Caldern-Zygmund theory. J. Funct. Anal. 256 (2009), no. 2, 509-593. [34] J. Parcet and N. Randrianantoanina. Gundy's decomposition for non-commutative martingales and applications. Proc. London Math. Soc. 93 (2006), 227-252. 42 R. Xia and X. Xiong [35] G. Pisier. Martingales in Banach spaces. Cambridge Studies in Advanced Mathematics, 155. Cambridge Uni- versity Press, Cambridge, 2016. xxviii+561 pp. [36] G. Pisier. Noncommutative vector valued Lp spaces and completely p-summing maps. Ast´erisque. 247 (1998), vi+131 pp. [37] G. Pisier and Q. Xu. Non-commutative Martingale Inequalities. Comm. Math. Phys. 189 (1997), 667-698. [38] G. Pisier and Q. Xu. Noncommutative Lp-spaces. Handbook of the geometry of Banach spaces, Vol. 2, 1459- 1517, North-Holland, Amsterdam, 2003. [39] N. Randrianantoanina. Noncommutative martingale transforms. J. Funct. Anal. 194 (2002), 181-212. [40] N. Randrianantoanina. Conditional square functions for noncommutative martingales. Ann. Prob. 35 (2007), 1039-1070. [41] N. Randrianantoanina and L. Wu. Noncommutative Burkholder/Rosenthal inequalities associated with convex functions. Ann. Inst. Henri Poincar´e Probab. Stat. 53 (2017), no. 4, 1575-1605. [42] N. Randrianantoanina, L. Wu and Q. Xu. Noncommutative Davis type decompositions and applications. arXiv:1712.01374. [43] E. M. Stein. Singular integrals and differentiability properties of functions. Princeton Mathematical Series, No. 30 Princeton University Press, Princeton, N.J. 1970. [44] E. M. Stein. Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals. Princeton Math- ematical Series, 43. Monographs in Harmonic Analysis, III. Princeton University Press, Princeton, NJ, 1993. [45] E. M. Stein and G. Weiss. Introduction to Fourier analysis on Euclidean spaces. Princeton Mathematical Series, No. 32. Princeton University Press, Princeton, N.J., 1971. [46] R. H. Torres. Boundedness results for operators with singular kernels on distribution spaces. Mem. Amer. Math. Soc. 90 (1991), no. 442, viii+172 pp. [47] H. Triebel. Theory of function spaces. II. Monographs in Mathematics, 84. Birkhauser Verlag, Basel, 1992. [48] T. Wolff. A note on interpolation spaces. Harmonic analysis (Minneapolis, Minn., 1981), pp. 199-204, Lecture Notes in Math., 908, Springer, Berlin-New York, 1982. [49] R. Xia and X. Xiong. Operator-valued inhomogeneous Triebel-Lizorkin spaces. Preprint. [50] R. Xia and X. Xiong. Mapping properties of operator-valued pseudo-differential operators. Preprint. [51] R. Xia, X. Xiong and Q. Xu. Characterizations of operator-valued Hardy spaces and applications to harmonic analysis on quantum tori. Adv. Math. 291 (2016), 183-227. [52] X. Xiong, Q. Xu and Z. Yin. Function spaces on quantum tori. C. R. Math. Acad. Sci. Paris 353 (2015), no. 8, 729-734. [53] X. Xiong, Q. Xu and Z. Yin. Sobolev, Besov and Triebel-Lizorkin spaces on quantum tori. Mem. Amer. Math. Soc. 252 (2018), no. 1203, vi+118 pp. [54] Q. Xu. Noncommutative Lp-spaces and martingale inequalities. Book manuscript, 2007. Laboratoire de Math´ematiques, Universit´e de Franche-Comt´e, 25030 Besanc¸on Cedex, France, and Instituto de Ciencias Matem´aticas, 28049 Madrid, Spain E-mail address: [email protected] Department of Mathematics and Statistics, University of Saskatchewan, Saskatoon, Saskatchewan, S7N 5E6, Canada E-mail address: [email protected]
0901.1403	2	0901	2010-04-30T12:20:30	The Logarithmic Sobolev Inequality in Infinite dimensions for Unbounded Spin Systems on the Lattice with non Quadratic Interactions	[ "math.FA", "math.PR" ]	We are interested in the Logarithmic Sobolev Inequality for the infinite volume Gibbs measure with no quadratic interactions. We consider unbounded spin systems on the one dimensional Lattice with interactions that go beyond the usual strict convexity and without uniform bound on the second derivative. We assume that the one dimensional single-site measure with boundaries satisfies the Log-Sobolev inequality uniformly on the boundary conditions and we determine conditions under which the Log-Sobolev Inequality can be extended to the infinite volume Gibbs measure.	math.FA	math	The Logarithmic Sobolev Inequality in Infinite dimensions for Unbounded Spin Systems on the Lattice with non Quadratic Interactions. Ioannis Papageorgiou ∗ Abstract We are interested in the Logarithmic Sobolev Inequality for the infi- nite volume Gibbs measure with no quadratic interactions. We consider unbounded spin systems on the one dimensional Lattice with interactions that go beyond the usual strict convexity and without uniform bound on the second derivative. We assume that the one dimensional single-site mea- sure with boundaries satisfies the Log-Sobolev inequality uniformly on the boundary conditions and we determine conditions under which the Log- Sobolev Inequality can be extended to the infinite volume Gibbs measure. Keywords: Logarithmic Sobolev inequality, Gibbs measure, Infinite dimensions, Spin systems. Mathematics Subject Classification (2000): 60E15, 26D10 1 Introduction We are interested in the q Logarithmic Sobolev Inequality (LSq) for measures related to systems of unbounded spins on the one dimensional Lattice with near- est neighbour interactions that are not strictly convex. Suppose that the Log- Sobolev Inequality is true for the single site measure with a constant uniformly bound on the boundary conditions. The aim of this paper is to present a cri- terion under which the inequality can be extended to the infinite volume Gibbs measure. More specifically, we extend the already know results for interactions V that satisfy k∇i∇jV (xi, xj)k∞ < ∞ to the more general case of interactions with k∇i∇jV (xi, xj)k∞ = ∞. ∗Department of Mathematics, Imperial College London, 180 Queen's Gate, London, SW7 2AZ. Email: [email protected] 1 Regarding the Log-Sobolev Inequality for the local specification {EΛ,ω}Λ⊂⊂Zd,ω∈Ω on a d-dimensional Lattice, criterions and examples of measures EΛ,ω that satisfy the Log-Sobolev -with a constant uniformly on the set Λ and the boundary con- ditions ω− are investigated in [Z2], [B-E], [B-L], [Y] and [B-H]. For k∇i∇jV (xi, xj)k∞ < ∞ the Log-Sobolev is proved when the phase φ is strictly convex and convex at infinity. Furthermore, in [G-R] the Spectral Gap Inequality is proved to be true for phases beyond the convexity at infinity, while in [M-M] and [B-J-S] the Decay of Correlation is studied. For the measure E{i},ω on the real line, necessary and sufficient conditions are presented in [B-G], [B-Z] and [R-Z], so that the Log-Sobolev Inequality is satisfied uniformly on the boundary conditions ω. The problem of the Log-Sobolev inequality for the Infinite dimensional Gibbs measure on the Lattice is examined in [G-Z], [Z1] and [Z2]. The first two study the LS for measures on a d-dimensional Lattice for bounded spin systems, while the third one looks at continuous spins systems on the one dimensional Lattice. In [M] and [O-R], criterions are presented in order to pass from the Log-Sobolev Inequality for the single-site measure E{i},ω to the LS2 for the Gibbs measure νN on a finite N-dimensional product space. Furthermore, using these criterions one can conclude the Log-Sobolev Inequality for the family {νN , N ∈ N} with a constant uniformly on N. Concerning the same problem for the LSq (q ∈ (1, 2]) inequality in the case of Heisenberg groups with quadratic interactions in [I-P] a similar criterion is presented for the Gibbs measure based on the methods developed in [Z1] and [Z2]. All the pre mentioned developments refer to measures with interactions V that satisfy k∇i∇jV (xi, xj)k∞ < ∞. The question that arises is whether similar assertions can be verified for the infinite dimensional Gibbs measure in the case where k∇i∇jV (xi, xj)k∞ = ∞ and in this paper we present a strategy to solve this problem. Consider the one dimensional measure E{i},ω(dxi) = e−φ(xi)−Pj∼i Jij V (xi,ωj)dXi Z {i},ω with k∂x∂yV (x, y)k∞ = ∞ Assume that E{i},ω satisfies the (LS) inequality with a constant uniformly on ω. Our aim is to set conditions, so that the infinite volume Gibbs measure ν for the local specification {EΛ,ω}Λ⊂⊂Z,ω∈Ω satisfies the LS inequality. We will focus on measures on the the one dimensional Lattice, but our result can also be easily extended on trees. Our general setting is as follows: The Lattice. When we refer to the Lattice we mean the 1-dimensional Lattice Z. 2 The Configuration space. We consider continuous unbounded random variables in R, representing spins. Our configuration space is Ω = RZ. For any ω ∈ Ω and Λ ⊂ Z we denote ω = (ωi)i∈Z, ωΛ = (ωi)i∈Λ, ωΛc = (ωi)i∈Λc and ω = ωΛ ◦ ωΛc where ωi ∈ R. When Λ = {i} we will write ωi = ω{i}. Furthermore, we will write i ∼ j when the nodes i and j are nearest neighbours, that means, they are connected with a vertex, while we will denote the set of the neighbours of k as {∼ k} = {r : r ∼ k}. The functions of the configuration. We consider integrable functions f that depend on a finite set of variables {xi}, i ∈ Σf for a finite subset Σf ⊂⊂ Z. The symbol ⊂⊂ is used to denote a finite subset. The Measure on Z. For any subset Λ ⊂⊂ Z we define the probability measure EΛ,ω(dxΛ) = e−H Λ,ω dxΛ Z Λ,ω where • xΛ = (xi)i∈Λ and dxΛ =Qi∈Λ dxi • Z Λ,ω =R e−H Λ,ω dxΛ • H Λ,ω =Pi∈Λ φ(xi) +Pi∈Λ,j∼i JijV (xi, zj) • zj = xΛ ◦ ωΛc =(xj , i ∈ Λ , i /∈ Λ and ωj We call φ the phase and V the potential of the interaction. For convenience we will frequently omit the boundary symbol from the measure and will write EΛ ≡ EΛ,ω. The Infinite Volume Gibbs Measure. The Gibbs measure ν for the local spec- ification {EΛ,ω}Λ⊂Z,ω∈Ω is defined as the probability measure which solves the Dobrushin-Lanford-Ruelle (DLR) equation νEΛ,⋆ = ν for finite sets Λ ⊂ Z (see [P]). For conditions on the existence and uniqueness of the Gibbs measure see e.g. [B-HK] and [D]. In this paper we consider local specifications for which the Gibbs measure exists and it is unique. It should be noted that {EΛ,ω}Λ⊂⊂Z,ω∈Ω always satisfies the DLR equation, in the sense that EΛ,ω EM,∗ = EΛ,ω 3 for every M ⊂ Λ. [P]. The gradient ∇ for continuous spins systems. For any subset Λ ⊂ Z we define the gradient ∇Λf q =Xi∈Λ ∇if q , where ∇i = ∂ ∂xi When Λ = Z we will simply write ∇ = ∇Z. We denote EΛ,ωf =Z f dEΛ,ω(xΛ) We can define the following inequalities The q Log-Sobolev Inequality (LSq). We say that the measure EΛ,ω satisfies the q Log-Sobolev Inequality for q ∈ (1, 2], if there exists a constant CLS such that for any function f , the following holds EΛ,ω f q log f q EΛ,ω f q ≤ CLS EΛ,ω ∇Λf q with a constant CLS ∈ (0, ∞) uniformly on the set Λ and the boundary conditions ω. The q Spectral Gap Inequality. We say that the measure EΛ,ω satisfies the q Spectral Gap Inequality for q ∈ (1, 2], if there exists a constant CSG such that for any function f , the following holds q ≤ CSGEΛ,ω ∇Λf q EΛ,ω(cid:12)(cid:12)f − EΛ,ωf(cid:12)(cid:12) with a constant CSG ∈ (0, ∞) uniformly on the set Λ and the boundary conditions ω. Remark 1.1. We will frequently use the following two well known properties about the Log-Sobolev and the Spectral Gap Inequality. If the probability measure µ sat- isfies the Log-Sobolev Inequality with constant c then it also satisfies the Spectral Gap Inequality with a constant c = 4c log 2. More detailed, in the case where q = 2 the optimal constant is less or equal to c 2 < c, while in the case 1 < q < 2 it is less or equal to 4c log 2. The constant c does not depend on the value of the parameter q ∈ (1, 2]. Furthermore, if for a family I of sets Λi ⊂ Z, dist(Λi, Λj) > 1 , i 6= j the measures EΛi,ω, i ∈ I satisfy the Log-Sobolev Inequality with constants ci, i ∈ I, then the probability measure E{∪i∈I Λi},ω also satisfies the (LS) Inequality with constant c = maxi∈I ci. The last result is also true for the Spectral Gap Inequality. The proofs of these two properties can be found in [G] and [G-Z] for q = 2 and in [B-Z] for 1 < q < 2. 4 2 The Main Result We want to extend the Log-Sobolev Inequality from the single-site measure E{i},ω to the Gibbs measure for the local specification {EΛ,ω}Λ⊂⊂Z,ω∈Ω on the entire one dimensional Lattice. Hypothesis We consider four main hypothesis: (H0): The one dimensional measures Ei,ω satisfies the Log-Sobolev-q Inequality with a constant c uniformly with respect to the boundary conditions ω. (H1): The restriction νΛ(k) of the Gibbs measure ν to the σ−algebra ΣΛ(k), Λ(k) = {k − 2, k − 1, k, k + 1, k + 2} satisfies the Log-Sobolev-q Inequality with a constant C ∈ (0, ∞). (H2): For some ǫ > 0 and K > 0 νΛ(i)e2q+2ǫV (xr ,xs) ≤ eK and νΛ(i)e2q+2ǫ∇r V (xr ,xs)q ≤ eK for r, s ∈ {i − 2, i − 1, i, i + 1, i + 2} (H3): The coefficients Ji,j are such that Ji,j ∈ [0, J] for some J < 1 sufficiently small. Remark 2.1. From Hypothesis (H2) and Jensen's inequality it follows that νeǫ(F (r)+ES(r),ωF (r))q ≤ eK, for r = i − 2, i − 1, i, i + 1, i + 2 where the functions F (r) are defined by F (r) =(∇rV (xi−1, xi) + ∇rV (xi+1, xi) ∇rV (xs, xr)Is∼r:s∈{i−3,i+3} for r = i − 1, i, i + 1 for r = i − 2, i + 2 and the sets S(r) by {∼ i} {i + 3, i + 4, ...} {..., i − 4, i − 3} for r = i − 1, i, i + 1 for r = i + 2 and s = i + 3 for r = i − 2 and s = i − 3 S(r) =  These bounds will be frequently used through out the paper. 5 Remark 2.2. Throughout this paper we will consider differentiable functions that satisfy ν f q < ∞ and ν ∇f q < ∞ The main theorem follows. Theorem 2.3. If hypothesis (H0)-(H3) are satisfied, then the infinite dimen- sional Gibbs measure ν for the local specification {EΛ,ω}Λ⊂⊂Z,ω∈Ω satisfies the q Log-Sobolev inequality ν f q log for some positive constant C. f q ν f q ≤ C ν ∇f q Proof. For the proof of the theorem it is sufficient to consider f ≥ 0. This is an assumption that we will make through all the proofs presented in this paper. We want to extend the Log-Sobolev Inequality from the single-site measure E{i},ω to the Gibbs measure for the local specification {EΛ,ω}Λ⊂⊂Z,ω∈Ω on the entire one dimensional lattice. To do so, we will follow the iterative method developed by Zegarlinski in [Z1] and [Z2]. Define the following sets Γ0 = even integers, Γ1 = Z r Γ0 One can notice that {dist(i, j) > 1, ∀i, j ∈ Γk, k = 0, 1}, Γ0 ∩ Γ1 = ∅ and Z = Γ0 ∪ Γ1. For convenience we will write EΓi = EΓi,ω for i = 0, 2.We will denote P = EΓ1 EΓ0 In order to prove the Log-Sobolev Inequality for the measure ν, we will express the entropy with respect to the measure ν as the sum of the entropies of the measures EΓ0 and EΓ1 which are easier to handle. We can write ν(f qlog f q νf q ) =νEΓ0(f qlog f q EΓ0f q ) + νEΓ1(EΓ0f qlog EΓ0f q EΓ1 EΓ0f q ) + ν(EΓ1 EΓ0f qlogEΓ1 EΓ0f q) − ν(f qlogνf q) (2.1) According to hypothesis (H0), the Log-Sobolev Inequality is satisfied for the single- state measures E{j} and the sets Γ0 and Γ1 are unions of one dimensional sets of distance greater than the length of the interaction one. Thus, as we mentioned in Remark 1.1 in the introduction, the (LS) holds for the product measures EΓ0 and EΓ1 with the same constant c. If we use the LS for EΓi, i = 0, 1 we get (2.1) ≤cν(EΓ0 ∇Γ0f q) + cνEΓ1(cid:12)(cid:12)(cid:12) ∇Γ1(EΓ0f q) q 1 q(cid:12)(cid:12)(cid:12) + ν(EΓ1 EΓ0f qlogEΓ1 EΓ0f q) − ν(f qlogνf q) (2.2) 6 For the third term of (2.2) we can write ν(Pf qlogPf q) =νEΓ0(Pf qlog Pf q EΓ0Pf q ) + νEΓ1(EΓ0Pf qlog EΓ0Pf q EΓ1 EΓ0Pf q ) + ν(EΓ1 EΓ0Pf qlogEΓ1 EΓ0Pf q) If we use again the Log-Sobolev Inequality for the measures EΓi, i = 0, 1 we get ∇Γ0(Pf q) ∇Γ1(EΓ0Pf q) q + ν(P 2f qlogP 2f q) (2.3) ν(Pf qlogPf q) ≤ cν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) + cν(cid:12)(cid:12)(cid:12) If we work similarly for the last term ν(P 2f qlogP 2f q) of (2.3) and inductively for any term ν(P kf qlogP kf q), then after n steps (2.2) and (2.3) will give 1 q(cid:12)(cid:12)(cid:12) ν(f qlog f q νf q ) ≤ν(P nf q log P nf q) − ν(f qlogνf q) + cν ∇Γ0f q + c ∇Γ0(P kf q) + c ∇Γ1(EΓ0P kf q) (2.4) n−1 Xk=1 ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) n−1 Xk=0 ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) In order to calculate the fourth and fifth term on the right-hand side of (2.4) we will use the following proposition Proposition 2.4. Suppose that hypothesis (H0)-(H3) are satisfied. Then the fol- lowing bound holds for {i, j} = {0, 1} and constants C1 ∈ (0, ∞) and 0 < C2 < 1. ∇Γi(EΓj f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) ≤ C1ν ∇Γif q + C2ν(cid:12)(cid:12)∇Γj f(cid:12)(cid:12) q The proof of Proposition 2.4 will be the subject of Section 4. If we apply inductively relationship (2.5) k times to the fourth and the fifth term of (2.4) we obtain ∇Γ0(P kf q) ≤ C 2k−1 2 C1ν ∇Γ1f q + C 2k 2 ν ∇Γ0f q and ∇Γ1(EΓ0P kf q) ≤ C 2k 2 C1ν ∇Γ1f q + C 2k+1 2 ν ∇Γ0f q (2.5) (2.6) (2.7) ν(cid:12)(cid:12)(cid:12) ν(cid:12)(cid:12)(cid:12) q q 1 1 q(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) If we plug (2.6) and (2.7) in (2.4) we get ν(f qlog f q νf q ) ≤ν(P nf qlogP nf q) − ν(f qlogνf q) C 2k−1 2 )C1ν ∇Γ1f q + c( C 2k 2 )ν ∇Γ0f q + c( + c( n−1 n−1 Xk=0 Xk=0 n−1 n−1 Xk=0 Xk=0 C 2k 2 )C1ν ∇Γ1f q + c( 7 C 2k+1 2 )ν ∇Γ0f q (2.8) If we take the limit of n to infinity in (2.8) the first two terms on the right hand side cancel with each other, as explained on the proposition bellow. Proposition 2.5. Under hypothesis (H0)-(H3), P nf converges ν-almost every- where to νf . The proof of this proposition will be presented in Section 3. So, taking the limit of n to infinity in (2.8) leads to ν(f qlog f q νf q ) ≤ cA(cid:18)C1 C2 + C2 + C1(cid:19) ν ∇Γ1f q + cAν ∇Γ0f q where A = limn→∞Pn−1 constant C = max{cA(cid:16) C1 C2 + C2 + C1(cid:17) , cA} k=0 C 2k 2 < ∞ for C2 < 1, and the theorem follows for a 3 Proof of Proposition 2.5. Before proving Proposition 2.5 we will present three useful lemmata. These lem- mata will also be used in the next section 4 where Proposition 2.4 is proved. In the case of quadratic interactions V (x, y) = (x − y)2 one can calculate Ei,ω(cid:0)f 2(∇jV (xi − xj) − Ei,ω∇jV (xi − xj))2(cid:1) (see [B-H] and [H]) with the use of the Deuschel-Stroock relative entropy inequality (see [D-S]) and the Herbst argument (see [L] and [H]). Herbst's arguement states that if a probability measure µ satisfies the LS2 inequality and a function F is Lipschitz continues with kF kLips ≤ 1 and such that µ(F ) = 0, then for some small ǫ we have < ∞ For µ = Ei,ω and F = ∇j V (xi−xj)−Ei,ω ∇j V (xi−xj) µeǫF 2 2 we then obtain Ei,ωe ǫ 4 (∇j V (xi−xj )−Ei,ω∇j V (xi−xj))2 < ∞ uniformly on the boundary conditions ω, because of hypothesis (H0). In the more general case however of non quadratic interactions that we examine in this work, the Herbst argument cannot be applied. In this and next sections we show how one can bound exponential quantities like the last one with the use of the projection of the infinite dimensional Gibbs measure and hypothesis (H1) and (H2). For every probability measure µ, we define the correlation function µ(f ; g) ≡ µ(f g) − µ(f )µ(g) 8 If for the set M(k) = Z r Λ(k) and hk := f − E{∼k}f we define Q(u, k) ≡ νΛ(u)(cid:12)(cid:12)(cid:12) ∇Λ(u)(cid:0)EM (u)hkq(cid:1) q 1 q(cid:12)(cid:12)(cid:12) then the following lemma presents an estimate for the correlation function, in terms of Q(k, k). Lemma 3.1. For any functions u localised in Λ(k) for which νΛ(k)e2qǫuq < ∞ the following inequalities are satisfied (a) under hypothesis (H1) (b) under hypothesis (H0) and (H1) q Ek−1Ek+1(f ; u)(cid:12)(cid:12) ν(cid:12)(cid:12) Ek−1Ek+1(f ; u)(cid:12)(cid:12) ν(cid:12)(cid:12) where c = 4c log 2 . ≤ C ǫ Q(k, k) + 1 q ≤ C ǫ Q(k, k) + q ǫ (cid:16)logνΛ(k)eǫu−Ek−1Ek+1uq(cid:17) ν(cid:12)(cid:12)f − Ek−1Ek+1f(cid:12)(cid:12) ǫ(cid:16)logνΛ(k)eǫu−Ek−1Ek+1uq(cid:17) Xi=k−1,k+1 ν ∇if q c Proof. From the definition of the correlation function we can write ν(cid:12)(cid:12) q Ek−1Ek+1(f ; u)(cid:12)(cid:12) Ek−1Ek+1((f − Ek−1Ek+1f )(u − Ek−1Ek+1u))(cid:12)(cid:12) =ν(cid:12)(cid:12) ≤νEk−1Ek+1(cid:0)f − Ek−1Ek+1f qu − Ek−1Ek+1uq(cid:1) =ν(cid:0)f − Ek−1Ek+1f qu − Ek−1Ek+1uq(cid:1) q (3.1) where above we first used the Jensen's Inequality and then the fact that the Gibbs measure ν satisfies the DLR equation. Because the function u is localised in Λ(k) and the measure E{k−1,k+1},ω = Ek−1Ek+1 has boundary in {k − 2, k, k + 2} ⊂ Λ(k), we have that u − Ek−1Ek+1u is also localised in Λ(k) and so for M(k) being the complementary of Λ(k) we can write ν(f − Ek−1Ek+1f qu − Ek−1Ek+1uq) = νΛ(k)(cid:0)(cid:0)EM (k)f − Ek−1Ek+1f q(cid:1) u − Ek−1Ek+1uq(cid:1) (3.2) On the right hand side of (3.2) we can use the following entropic inequality (see [D-S]) ∀t > 0, µ(uy) ≤ µ(y log y) (3.3) 1 t 1 t log(cid:0)µ(etu)(cid:1) + 9 for any probability measure µ and y ≥ 0, µy = 1. Then from (3.1) and (3.2) we will obtain q ν(cid:12)(cid:12) 1 ǫ ≤ νΛ(k)EM (k)f − Ek−1Ek+1f q log Ek−1Ek+1(f ; u)(cid:12)(cid:12) ǫ (cid:16)logνΛ(k)eǫu−Ek−1Ek+1uq(cid:17) νΛ(k)EM (k)f − Ek−1Ek+1f q EM (k)f − Ek−1Ek+1f q νΛ(k)EM (k)f − Ek−1Ek+1f q 1 + (3.4) The first term on the right hand side of (3.4) can be bounded from hypothesis (H1) by the Log-Sobolev inequality for νΛ(k) νΛ(k)EM (k)f − Ek−1Ek+1f q log EM (k)f − Ek−1Ek+1f q νΛ(k)EM (k)f − Ek−1Ek+1f q ∇Λ(k)(EM (k)f − Ek−1Ek+1f q) ≤CνΛ(k)(cid:12)(cid:12)(cid:12) Using (3.4) and (3.5) we get = CQ(k, k) (3.5) q 1 q(cid:12)(cid:12)(cid:12) ν(cid:12)(cid:12) Ek−1Ek+1(f ; u)(cid:12)(cid:12) q ≤ C ǫ Q(k, k) + 1 ǫ (cid:16)logνeǫu−Ek−1Ek+1uq(cid:17) ν(cid:12)(cid:12)f − Ek−1Ek+1f(cid:12)(cid:12) q (3.6) which proves (a). If we assume hypothesis (H0), then we can bound the second term on the right hand side of (3.6) from the SGq for the measures Ek−1, Ek+1 from hypothesis (H0) and the product property for the SGq (Remark 1.1), to obtain νf − Ek−1Ek+1f q =νEk−1Ek+1f − Ek−1Ek+1f q ≤ c Xi=k−1,k+1 where c = 4c log 2. Using (3.6) and (3.7) we finally get (b) ν ∇if q (3.7) q Ek−1Ek+1(f ; u)(cid:12)(cid:12) ν(cid:12)(cid:12) ≤ C ǫ Q(k, k) + c ǫ(cid:16)logνeǫu−Ek−1Ek+1uq(cid:17) Xi=k−1,k+1 ν ∇if q The following lemma gives an explicit bound for the quantity Q(k, k). Lemma 3.2. Suppose that hypothesis (H0)-(H3) are satisfied. Then k+2 Q(k, k) ≤D ν ∇rf q Xr=k−2 Xn=0 ∞ + D J (n+1)(q−1) for some positive constant D. (ν ∇k+3+4n+rf q + ν ∇k−3−4n−rf q) 3 Xr=0 10 The proof of this lemma will be the subject of Section 5. Lemma 3.3. Suppose that hypothesis (H0)-(H3) are satisfied. Then for {i, j} = {0, 1} holds for constants D1 ∈ (0, ∞) and 0 < D2 < 1. q ν(cid:12)(cid:12)∇Γi(EΓj f )(cid:12)(cid:12) ≤ D1ν ∇Γif q + D2ν(cid:12)(cid:12)∇Γj f(cid:12)(cid:12) q Proof. Assume i = 1, j = 0. We have q ν(cid:12)(cid:12)∇Γ1(EΓ0f )(cid:12)(cid:12) = Xi∈Γ1 q ν(cid:12)(cid:12)∇i(EΓ0f )(cid:12)(cid:12) ≤ Xi∈Γ1 q ν(cid:12)(cid:12)∇i(Ei−1Ei+1f )(cid:12)(cid:12) e−H(xi−1)e−H(xi+1) R e−H(xi−1)dxi R e−H(xi+1)dxi the density of the measure Ei−1Ei+1 we (3.8) If we denote ρi = can then write q ≤ q ∇i(Z Z ρif dxi−1dxi+1)(cid:12)(cid:12)(cid:12)(cid:12) = ν(cid:12)(cid:12)(cid:12)(cid:12) ν(cid:12)(cid:12)∇i(Ei−1Ei+1f )(cid:12)(cid:12) + 2q−1ν(cid:12)(cid:12)(cid:12)(cid:12) Z Z (∇if )ρidxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) 2q−1ν(cid:12)(cid:12)(cid:12)(cid:12) Ei−1Ei+1(∇if )(cid:12)(cid:12) Ei−1Ei+1(f ; ∇iV (xi−1, xi) + ∇iV (xi+1, xi))(cid:12)(cid:12) + c1J qν(cid:12)(cid:12) c1ν(cid:12)(cid:12) Z Z f (∇iρi)dxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) ≤ q q q q (3.9) (3.10) where in (3.10) we used hypothesis (H3) to bound the coefficients Ji,j and we have denoted c1 = 24q. If we apply the Holder Inequality to the first term of (3.10) and Lemma 3.1 (b) to the second term, we obtain ν(cid:12)(cid:12)∇i(Ei−1Ei+1f )(cid:12)(cid:12) q ≤ c1ν ∇if q + J qc1C ǫ Q(i, i)+ J qcc1K ǫ Xk=i−1,i+1 ν ∇kf q (3.11) where the constant K as in hypothesis (H2). From (3.8) and (3.11) we have ≤ c1ν ∇Γ1f q + J qc1C ǫ Xi∈Γ1 Q(i, i) + J qcc1K ǫ Xi∈Γ1 Xk=i−1,i+1 ν ∇kf q If we use Lemma 3.2 to replace Q(k, k) in the above expression we get q ν(cid:12)(cid:12)∇Γ1(EΓ0f )(cid:12)(cid:12) q ν(cid:12)(cid:12)∇Γ1(EΓ0f )(cid:12)(cid:12) ≤ c1ν ∇Γ1f q + J qc2c1K ǫ ν ∇Γ0f q + J qc1DC ǫ Xi∈Γ1 ∞ Xn=0 J (n+1)(q−1) 3 Xr=0 11 J qc1DC ǫ Xi∈Γ1 i+2 Xr=i−2 ν ∇rf q + (ν ∇i+3+4n+rf q + ν ∇i−3−4n−rf q) for constant D > 0 as in Lemma 3.2. For coefficients Ji,j sufficiently small such that J < 1 in (H3) we finally obtain ν(cid:12)(cid:12)∇Γ1(EΓ0f )(cid:12)(cid:12) q ≤ J q(cid:16) 2cc1K ǫ + 2c1CD ǫ + 2D c1C ǫ +(cid:16)c1q + J qc1C ǫ 3D + D 2J qc1C ǫ J (q−1) 1−J (q−1)(cid:17) ν ∇Γ0f q J (q−1) 1−J (q−1)(cid:17) ν ∇Γ1f q and the lemma follows for J sufficiently small such that D2 = J q(cid:18) cc1K ǫ 2 + 2c1CD ǫ + D 2c1C ǫ J (q−1) 1 − J (q−1)(cid:19) < 1 Now we can prove Proposition 2.5. Proof of Proposition 2.5. Following [G-Z] we will show that in L1(ν) we have limn→∞P n = ν. For i 6= j we have that νEΓj f − EΓi EΓj f q = νEΓiEΓj f − EΓi EΓj f q q ≤ cν(cid:12)(cid:12)∇Γi(EΓj f )(cid:12)(cid:12) (3.12) The last inequality due to the fact that both the measures EΓ0 and EΓ1 satisfy the Log-Sobolev Inequality and the Spectral Gap inequality with constants indepen- dently of the boundary conditions. If we use Lemma 3.3 we get νEΓj f − EΓi EΓj f q ≤ cD1ν∇Γif q + cD2ν∇Γj f q From the last inequality we obtain that for any n ∈ N, νP nf − EΓ0P nf q ≤ cD1ν∇Γ0(EΓ0P n−1f )q + cD2ν∇Γ1(EΓ0P n−1f )q = cD2ν∇Γ1(EΓ0P n−1f )q If we use Lemma 3.3 to bound the last expression we have the following νP nf − EΓ0P nf q ≤ cDn 2 (D1ν ∇Γ1f q + D2ν ∇Γ0f q) (3.13) Similarly we obtain νEΓ0P nf − P n+1f q ≤ cDn 2 (D1ν ∇Γ1f q + D2ν ∇Γ0f q) (3.14) 12 Consider the sequence {Qn}n∈N defined as Qnf =(P n 2 f EΓ0P n−1 2 f if n even if n odd for every n ∈ N. Hence, if we define the sets An = {Qnf − Qn+1f ≥ ( 1 2 )n} we obtain ν(An) = ν(cid:18){Qnf − Qn+1f ≥ ( 1 2 )n}(cid:19) ≤ 2qnνQnf − Qn+1f q by Chebyshev inequality. If we use (3.13) and (3.14) to bound the last we have ν(An) ≤ (2qD 1 2 2 )nc (D1ν ∇Γ1f q + D2ν ∇Γ0f q) We can choose J sufficiently small such that 2qD 1 2 2 < 1 2 in which case we get that ∞ Xn=0 ν(An) ≤ ∞ Xn=0 ( 1 2 )n! c (D1ν ∇Γ1f q + D2ν ∇Γ0f q) < ∞ From the Borel-Cantelli lemma, only finite number of the sets An can occur, which implies that the sequence {Qnf }n∈N is a Cauchy sequence and that it converges ν−almost surely. Say Qnf → θ(f ) ν − a.e. We will first show that θ(f ) is a constant, i.e. it does not depend on variables on Γ0 or Γ1. To show that, first notice that Qn(f ) is a function on Γ1 and Γ0 when n is odd and even respectively, which implies that the limits θo(f ) = lim n odd,n→∞ Qnf and θe(f ) = lim n even,n→∞ Qnf do not depend on variables on Γ0 and Γ1 respectively. Since both the subsequences {Qnf }n even and {Qnf }n odd converge to θ(f ) ν−a.e. we have that which implies that θ(f ) is a constant. From that we obtain that θo(f ) = θ(f ) = θe(f ) ν (θ(f )) = θ(f ) (3.15) 13 Since the sequence {Qnf }n∈N converges ν−almost, the same holds for the sequence {Qnf − νQnf }n∈N. We have (Qnf − νQnf ) = θ(f ) − ν (θ(f )) = θ(f ) − θ(f ) = 0 lim n→∞ where above we used (3.15). On the other side, we also have lim n→∞ (Qnf − νQnf ) = lim n→∞ (Qnf − νf ) = θ(f ) − ν(f ) (3.16) From (3.15) and (3.16) we get that θ(f ) = ν(f ) We finally get lim n→∞ P nf = lim n even,n→∞ Qnf = νf, ν a.e. 4 Proof of Proposition 2.4 Before we prove Proposition 2.4 we present some useful lemmata. First we define Wk = ∇kV (xk, xk−1) + ∇kV (xk, xk+1) and Uk = Wkq + E{∼k} Wkq (4.1) where {∼ k} ≡ {j : j ∼ k} = {k − 1, k + 1}. Lemma 4.1. The following inequality holds E{∼k}(f q; Wk) ≤ c0(cid:0)E{∼k}f q(cid:1) p (cid:0)E{∼k}(f − E{∼k}f qUk)(cid:1) for some constant c0 uniformly on the boundary conditions and 1 q + 1 p = 1. 1 1 q Proof. We can write E{∼k}(f q; Wk) = 1 2 where E{∼k} is an isomorphic copy of E{∼k}. If we define the function F to be F (s) = sf + (1 − s) f then E{∼k} ⊗ E{∼k}(cid:16)(f q − f q)(Wk − Wk)(cid:17) (4.2) (4.2) = = = 1 2 1 2 1 2 0 E{∼k} ⊗ E{∼k}(cid:18)(cid:18)Z 1 E{∼k} ⊗ E{∼k}(cid:18)(cid:18)Z 1 E{∼k} ⊗ E{∼k}(cid:18)(cid:18)qZ 1 0 0 ds d ds F (s)q(cid:19) (Wk − Wk)(cid:19) F (s)(cid:19) (Wk − Wk)(cid:19) dsqF (s)q−1 d ds dsF (s)q−1(f − f )(cid:19) (Wk − Wk)(cid:19) 14 If we use the Holder inequality for the conjugate numbers p and q, then the last quantity can be bounded by q 2nE{∼k} ⊗ E{∼k}(cid:16)R 1 1 p × 0 dsF (s)q−1(cid:17)po nE{∼k} ⊗ E{∼k}(cid:12)(cid:12)(cid:12) For the first term in the above product, by Jensen's Inequality and 1 obtain q + 1 p = 1, we qo (f − f )(Wk − Wk)(cid:12)(cid:12)(cid:12) 1 q (4.3) (cid:26)E{∼k} ⊗ E{∼k}(cid:18)Z 1 0 1 p dsF (s)q−1(cid:19)p(cid:27) dsF (s)q(cid:27) 0 ≤(cid:26)E{∼k} ⊗ E{∼k}Z 1 ≤(cid:18)2qZ 1 0 dsE{∼k} ⊗ E{∼k}(cid:16)sf q + (1 − s) f q(cid:17)(cid:19) 1 p 1 p =(cid:18)Z 1 0 dsE{∼k} ⊗ E{∼k}F (s)q(cid:19) 1 p = 2 q p (E{∼k}f q) 1 p (4.4) If we plug (4.4) into (4.3) we finally get E{∼k}(f q; Wk) ≤ (E{∼k}f q) q 2 p q 2 ≤262 q p q(E{∼k}f q) 1 1 p nE{∼k} ⊗ E{∼k}(cid:16)f − f Wk − Wk(cid:17)qo p (cid:8)E{∼k}(cid:0)f − E{∼k}f q(Wkq + E{∼k} Wkq)(cid:1)(cid:9) 1 q 1 q The lemma follows for constant c0 = 262 q p q. Define now the quantity The next lemma presents an estimate of A(k) involving Q(k, k). A(k) = ν(cid:0)E{∼k}f q(cid:1)− q E{∼k}(f q; Wk)(cid:12)(cid:12) p (cid:12)(cid:12) q Lemma 4.2. Suppose that that hypothesis (H0)-(H2) are satisfied. Then A(k) ≤ cq 0C ǫ Q(k, k) + cq 0cK ǫ Xi=k−1,k+1 ν ∇if q where the constants ǫ and K are as in hypothesis (H2). Proof. We can initially bound A(k) with the use of Lemma 4.1 A(k) =ν(cid:0)E{∼k}f q(cid:1)− q E{∼k}(f q; Wk)(cid:12)(cid:12) p (cid:12)(cid:12) 0νΛ(k)(cid:0)(EM (k)f − E{∼k}f q)Uk(cid:1) =cq q 15 ≤ cq 0νEk−1Ek+1(f − Ek−1Ek+1f qUk) (4.5) because Uk is localized in Λ(k). If we use the entropy inequality (3.3) and hypoth- esis (H1) for νΛ(k) as well as (H2), as we did in Lemma 3.1, then for K as in (H2), we can bound (4.5) by (4.5) ≤ ≤ cq 0C ǫ cq 0C ǫ Q(k, k) + Q(k, k) + cq 0K ǫ cq 0cK νΛ(k)E{∼k}f − E{∼k}f q ǫ Xi=k−1,k+1 ν ∇if q where above we used that E{∼k} = Ek−1Ek+1 satisfies the SGq with constant c uniformly on the boundary conditions, by hypothesis (H0) and Remark 1.1. Lemma 4.3. The following inequality holds ∇i(Ei−1Ei+1f q) ν(cid:12)(cid:12)(cid:12) Proof. We have q 1 q(cid:12)(cid:12)(cid:12) ≤ c1ν ∇if q + J qc1 qq A(i) ∇i(Ei−1Ei+1f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) =ν(cid:12)(cid:12)(cid:12)(cid:12) = 1 1 q (Ei−1Ei+1f q) qq ν(Ei−1Ei+1f q)− q q q 1 q −1∇i(Ei−1Ei+1f q)(cid:12)(cid:12)(cid:12)(cid:12) p (cid:12)(cid:12)∇i(Ei−1Ei+1f q)(cid:12)(cid:12) (4.6) But from relationship (3.9) of Lemma 3.3, for ρi being the density of E{∼i} we have q q q q ≤ For the second term in (4.7) we have Z Z f q(∇iρi)dxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) ∇i(Ei−1Ei+1f q)q =(cid:12)(cid:12)∇i(R R ρif qdxi−1dxi+1)(cid:12)(cid:12) + 22q−2(cid:12)(cid:12)(cid:12)(cid:12) Z Z ∇i(f q)ρidxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) 22q−2(cid:12)(cid:12)(cid:12)(cid:12) Z Z f q(∇iρi)dxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) Ei−1Ei+1(f q; ∇iV (xi−1, xi) + ∇iV (xi+1, xi))(cid:12)(cid:12) ≤ J q(cid:12)(cid:12) Z Z ∇i(f q)ρidxi−1dxi+1(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) While for the first term of (4.7) the following bound holds Ei−1Ei+1(f q−1(∇if ))(cid:12)(cid:12) = qq(cid:12)(cid:12) ≤ qq(cid:0)Ei−1Ei+1f (q−1)p(cid:1) = qq(cid:0)Ei−1Ei+1f q(cid:1) p (cid:0)Ei−1Ei+1 ∇if q(cid:1) p (cid:0)Ei−1Ei+1 ∇if q(cid:1) q q q q (4.7) q (4.8) (4.9) 16 where above we used the Holder inequality and that p is the conjugate of q. If we plug (4.8) and (4.9) in (4.7) we get (cid:12)(cid:12)∇i(Ei−1Ei+1f q)(cid:12)(cid:12) q ≤22q−2qq(cid:0)Ei−1Ei+1f q(cid:1) + 22q−2J q(cid:12)(cid:12) q p (cid:0)Ei−1Ei+1 ∇if q(cid:1) Ei−1Ei+1(f q; ∇iV (xi−1, xi) + ∇iV (xi+1, xi))(cid:12)(cid:12) q From the last relationship and (4.6) the lemma follows. Now we can prove Proposition 2.4. Proof of Proposition 2.4. We have ∇Γ1(EΓ0f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) =Xi∈Γ1 ≤Xi∈Γ1 ν(cid:12)(cid:12)(cid:12) ∇i(EΓ0f q) c1ν ∇if q + ∇i(E{∼i}f q) ν(cid:12)(cid:12)(cid:12) A(i) q 1 q(cid:12)(cid:12)(cid:12) ≤ Xi∈Γ1 qq Xi∈Γ1 J qc1 q 1 q(cid:12)(cid:12)(cid:12) where the last inequality is due to Lemma 4.3. If we use Lemma 4.2 to bound A(i) we get ∇Γ1(EΓ0f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) ≤Xi∈Γ1 c1ν ∇if q + cq 0cK ǫ J qc1 qq Xi∈Γ1 Xr=i−1,i+1 ν ∇rf q + J qc1 qq cq 0C ǫ Xi∈Γ1 Q(i, i) Furthermore, if we use Lemma 3.2 to bound Q(i, i) we obtain ∇Γ1(EΓ0f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) ≤Xi∈Γ1 c1ν ∇if q + cq 0cK ǫ k+2 J qc1 qq Xi∈Γ1 Xr=i−1,i+1 ν ∇rf q + + + J qc1 qq J qc1 qq J qc1 qq cq 0CD cq 0CD ǫ Xi∈Γ1 ǫ Xi∈Γ1 ǫ Xi∈Γ1 ∞ Xr=k−2 Xn=0 Xn=0 ∞ cq 0CD ν ∇rf q J (n+1)(q−1) J (n+1)(q−1) ν ∇i+3+4n+rf q ν ∇i−3−4n−rf q (4.10) 3 3 Xr=0 Xr=0 If we set R = c1 + c1 qq ( cq 0CD ǫ + cq 0cK ǫ ) and we choose J < 1, relationship (4.10) gives ∇Γ1(EΓ0f q) ν(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) ≤ (R + RJ q4 + R8J q 1 − J q−1 )ν ∇Γ1f q + RJ q(4 + 8 1 − J q−1 )ν ∇Γ0f q 17 For J sufficiently small (H3) such that RJ q(4 + 8 constants 1−J q−1 ) < 1 the lemma follows for C1 = R + RJ q4 + R8J q 1−J q−1 and C2 = RJ q(4 + 8 1−J q−1 ) < 1. 5 Proof of Lemma 3.2 This section is dedicated in the proof of Lemma 3.2 under the assumptions (H0)- (H3). We begin by showing the weaker result of Lemma 5.1 under the weaker assumptions (H1)-(H3). Lemma 5.1. Suppose that hypothesis (H1)-(H3) are satisfied. Then Q(k, k) ≤J qSν(cid:12)(cid:12)f − Ek−1Ek+1f(cid:12)(cid:12) Xr=0 J (n+1)(q−1) Xn=0 + S q + S k+2 Xr=k−2 ν ∇rf q (ν ∇k+3+4n+rf q + ν ∇k−3−4n−rf q) ∞ 3 for some positive constant S. Lemma 3.2 follows for some constant D > 0 directly from the last lemma and the Spectral Gap inequality implied from (H0). The remaining of this section is dedicated to the proof of Lemma 5.1. At first we prove some lemmata. To start, for any k ∈ Z, we define the sets Ms(k) for s = k − 3, k + 3 as Ms(k) =({j ∈ Z : j ≥ k + 3} = {k + 3, k + 4, ...} {j ∈ Z : j ≤ k − 3} = {..., k − 4, k − 3} if s = k + 3 if s = k − 3 (5.1) Remark 5.2. Since Λ(k) = {k − 2, k − 1, k, k + 1, k + 2} and M(k) = Z r Λ(k), with the use of the definition (5.1) we can write M(k) = {j ∈ Z : j ≤ k − 3} ∪ {j ∈ Z : j ≥ k + 3} = Mk−3(k) ∪ Mk+3(k) Since the sets Mk−3(k) and Mk+3(k) are disjoint we obtain that EM (k) is a product measure, and for every function f we can write EM (k)f = EMk−3(k) ⊗ EMk+3(k)f (5.2) Accordingly, for functions, say fk−3 and fk+3, that depend on variables xi with i /∈ Mk+3(k) and i /∈ Mk−3(k) respectively, we obtain EM (k)fk−3 = EMk−3(k)fk−3 18 and EM (k)fk+3 = EMk+3(k)fk+3 For instance, for r = k − 2, k + 2 and s ∈ {k − 3, k + 3} : s ∼ r, that is for the couples (r, s) = (k − 2, k − 3) and (r, s) = (k + 2, k + 3), we have EM (k)∇rV (xs, xr) = EMs(k)∇rV (xs, xr) (5.3) Remark 5.3. Consider couples (r, s) that take the values (k − 2, k − 3) and (k + 2, k + 3). We then have that ∇rV (xs, xr) is localised in Λ(k − 4) when (r, s) = (k − 2, k − 3) and in Λ(k + 4) when (r, s) = (k + 2, k + 3). Furthermore, from Remark 5.2, for (r, s) = (k − 2, k − 3) we get that EMk−3(k)∇k−2V (xk−3, xk−2) = E{...,k−4,k−3}∇k−2V (xk−3, xk−2) is localised in Λ(k − 4), while for (r, s) = (k + 2, k + 3) we get that EMk+3(k)∇k+2V (xk+3, xk+2) = E{k+3,k+4,...}∇k+2V (xk+3, xk+2) is localised in Λ(k + 4). So, if we set Ys(xs, xr) = ∇rV (xs, xr) − EMs(k)∇rV (xs, xr) we then have that Yk+3(xk+2, xk+3) and Yk−3(xk−2, xk−3) are localised in Λ(k + 4) and Λ(k − 4) respectively. Thus, we have If we combine the last two together we can write ν(f qY q ν(f qY q ν(f qY q k+3(xk+3, xk+2)) = νΛ(k+4)(cid:0)(EM (k+4)f q)Y q k−3(xk−3, xk−2)) = νΛ(k−4)(cid:0)(EM (k−4)f q)Y q s (xs, xr)) = νΛ(t)(cid:0)(EM (t)f q)Y q k+3(xk+3, xk+2)(cid:1) k−3(xk−3, xk−2)(cid:1) s (xs, xr)It∈{k−4,k+4}∩Ms(k)(cid:1) for (r, s) ∈ {(k + 2, k + 3), (k − 2, k − 3)}. Lemma 5.4. Suppose conditions (H1) and (H2) are satisfied. Then for r = k − 2, k + 2 and s ∈ {k − 3, k + 3} : s ∼ r the following inequality is true νΛ(k)(cid:0)EM (k)f q(cid:1)− q p (cid:12)(cid:12) q q 1 C ǫ EM (k)(f q; ∇rV (xs, xr))(cid:12)(cid:12) νΛ(t)(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) ∇Λ(t)(EM (t)f q) ≤ It∈{k−4,k+4}∩Ms(k) + K ǫ νf q where IA denotes the characteristic function of a set A and the set Ms(k) as in (5.1). 19 Proof. For any two function f and g the covariance with respect to a measure µ can be computed as bellow µ(f ; g) = µ ((f − µf )(g − µg)) = µ (f (g − µg)) − µ (µf (g − µg)) = µ (f (g − µg)) − (µf )µ (g − µg) = µ (f (g − µg)) Using this expression we can write EM (k)(f q; ∇rV (xs, xr)) = EM (k)(f q(∇rV (xs, xr) − EM (k)∇rV (xs, xr))) (5.4) If we use (5.3) from Remark 5.2, (5.4) becomes EM (k)(f q; ∇rV (xs, xr)) = EM (k)(f q(∇rV (xs, xr) − EMs(k)∇rV (xs, xr))) (5.5) If we set Ys(xs, xr) = ∇rV (xs, xr) − EMs(k)∇rV (xs, xr) then for (5.5) we can write EM (k)(f q; ∇rV (xs, xr))(cid:12)(cid:12) ≤ EM (k)(f q−1+1Ys(xs, xr)) (cid:12)(cid:12) p (cid:0)EM (k)(f qY q ≤(cid:0)EM (k)f (q−1)p(cid:1) =(cid:0)EM (k)f q(cid:1) p (cid:0)EM (k)(f qY q 1 1 s (xs, xr))(cid:1) s (xs, xr))(cid:1) 1 q 1 q (5.6) p + 1 q = 1. So, for s = k+3, k−3 where above we used the Holder inequality and that 1 from relationship (5.6) we obtain νΛ(k)(cid:0)EM (k)f q(cid:1)− q EM (k)(f q; ∇rV (xs, xr))(cid:12)(cid:12) p (cid:12)(cid:12) q ≤ νΛ(k)EM (k)(f qY q = ν(f qY q s (xs, xr)) s (xs, xr)) If we combine the last inequality together with Remark 5.3 we finally obtain νΛ(k)(cid:0)EM (k)f q(cid:1)− q EM (k)(f q; ∇rV (xs, xr))(cid:12)(cid:12) p (cid:12)(cid:12) νΛ(t)(cid:0)(EM (t)f q)Y q q ≤ s (xs, xr)It∈{k−4,k+4}∩Ms(k)(cid:1) If in (5.7) we use the Entropy Inequality and the LSq for νΛ(s) from hypothesis (H1) and (H2), we get (5.7) νΛ(k)(cid:0)EM (k)f q(cid:1)− q ≤ C ǫ p (cid:0)EM (k)(f q; ∇rV (xs, xr))(cid:1)q νΛ(t)(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) ∇Λ(t)(EM (t)f q) q 1 It∈{k−4,k+4}∩Ms(k) + K ǫ νf q and s ∈ {k − 3, k + 3} : s ∼ r and K and ǫ as in hypothesis (H2). 20 Lemma 5.5. Suppose P and G are positive functions with domain on N such that for constants J, K ′ > 0 and for n = 4k for k ∈ N ∩ [2, ∞) P (4) ≤ G(4) + J qK ′P (8) P (n) ≤ G(n) + J qK ′P (n − 4) + J qK ′P (n + 4) (5.8) (5.9) Then for J sufficiently small such that J ≤ 1 and JK ′ + J qK ′J q−1 6 1 (5.10) the following inequality holds P (4n) ≤ 1 n−2 Xm=0 1 − J qK ′J q−1 + J q−1P (4n + 4) J mq−mG(4n − 4m) + J (n−1)q−(n−1)G(4) (5.11) for any n ∈ N, n ≥ 2 . Proof. In order to show (5.11) we will work inductively. Step 1: The base case of the induction (n=2). We prove (5.11) for n = 2. For k = 8 in (5.9) we have P (8) ≤ G(8) + J qK ′P (12) + J qK ′P (4) If we bound P (4) in the above inequality by (5.8) we obtain P (8) ≤ G(8) + J qK ′P (12) + J qK ′G(4) + (J qK ′)2P (8) ⇒ P (8) ≤ 1 1 − (J qK ′)2 G(8) + J qK ′ 1 − (J qK ′)2 G(4) + J qK ′ 1 − (J qK ′)2 P (12) (5.12) For J satisfying properties (5.10), we have JK ′ + J qK ′J q−1 ≤ 1 and JK ′ < 1 which implies JK ′ + (J qK ′)2 ≤ 1 ⇒ J qK ′ 1 − (J qK ′)2 ≤ J q−1 (5.13) From (5.12) and (5.13) we have P (8) ≤ ≤ 1 1 − (J qK ′)2 G(8) + J q−1G(4) + J q−1P (12) 1 − J qK ′J q−1 G(8) + J q−1G(4) + J q−1P (12) 1 21 because of (5.10). This proves (5.11) for n = 2. Step 2: The induction step. Suppose the inequality (5.11) is true for n = k. Then we will show it is also true for n = k + 1. If we use (5.9) for n = 4k + 4 we have P (4k + 4) ≤ G(4k + 4) + J qK ′P (4k) + J qK ′P (4k + 8) (5.14) If we use (5.11) for n = k to bound P (4k) in (5.14) we get P (4k + 4) ≤G(4k + 4) + J qK ′ 1 − J qK ′J q−1 J mq−mG(4k − 4m) k−2 Xm=0 + J qK ′J (k−1)q−(k−1)G(4) + J qK ′J q−1P (4k + 4) + J qK ′P (4k + 8) This implies P (4k + 4) ≤ 1 1 − J qK ′J q−1 G(4k + 4) + J qK ′ 1 − J qK ′J q−1 + J qK ′J (k−1)q−(k−1) 1 − J qK ′J q−1 G(4) + J qK ′ 1 − J qK ′J q−1 P (4k + 8) J mq−m 1 − J qK ′J q−1 G(4k − 4m) k−2 Xm=0 (5.15) If we use condition (5.10) for J, (5.15) becomes P (4k + 4) ≤ m=0 J mq−mG(4k + 4 − 4m) + J kq−kG(4) 1 1−J q K ′J q−1 Pk−1 +J q−1P (4k + 8) which proves (5.11) for n = k + 1. This finishes the proof of (5.11). Lemma 5.6. Suppose P and G are positive functions with domain on N such that for constants J, K ′ > 0 one has P (n) < ∞ sup n∈N (5.16) as well as (5.8) and (5.9) for n = 4k for k ∈ N ∩ [2, ∞). Then for J sufficiently small such that (5.10) is true, the following inequality holds P (4) ≤ 1 1 − J 2q−2 +∞ Xn=0 22 J nq−nG(4n + 4) Proof. We can use relationship (5.11) from Lemma 5.5 to prove the lemma. We first replace the bound of P (8) from (5.11) in (5.8), to obtain P (4) ≤ G(4) + J qK ′ 1 1 − J qK ′J q−1 G(8) + J qK ′J q−1G(4) + J qK ′J q−1P (12) ≤ (1 + J qK ′J q−1)G(4) + J 2q−2G(8) + JK ′J 2q−2P (12) where at the last inequality we used (5.10). If we now bound in the above expression P (12) from (5.11), then P (16) from (5.11) and so on, we will finally obtain +∞ P (4) ≤(1 + J qK ′ Xn=0 J (2n+1)q−(2n+1))G(4) +∞ + J qK ′ 1 − J qK ′J q−1 Xn=1 J 2q−1K ′ 1 − J 2q−2 )G(4) + =(1 + J (n−1)q−(n−1)( J 2sq−2s)G(4n + 4) +∞ Xs=0 J qK ′ 1 1 − J qK ′J q−1 1 − J 2q−2 J (n−1)q−(n−1)G(4n + 4) +∞ Xn=1 where above we used that J < 1, as well as that J nq−nP (8 + 4n) = 0 lim n→∞ since (5.16) is true. Furthermore, if we use again (5.10) we then get P (4) ≤ 1 1 − J 2q−2 +∞ Xn=0 J nq−nG(4n + 4) The next lemma presents a bound for Q(u, k) = νΛ(u)(cid:12)(cid:12)(cid:12) ∇Λ(u)(cid:0)EM (u)hkq(cid:1) q 1 q(cid:12)(cid:12)(cid:12) in terms of Q(t, k)Idist(u,t)=4. Lemma 5.7. Under hypothesis (H1) and (H2) the following bound for Q(u, k) holds Q(u, k) ≤νΛ(u) ∇uhkq + Xr=u−1,u+1 + c1 Xr=u−2,u+2 νΛ(u) ∇rhkq + where hk = f − E{∼k}f . 23 νΛ(u) ∇rhkq + J qc12cK ǫ ν hkq J qc1C ǫ Xdist(u,t)=4 Q(t, k) Proof. We have Q(u, k) =νΛ(u)(cid:12)(cid:12)(cid:12) ∇Λ(u)(cid:0)EM (u)hkq(cid:1) + Xr=u−1,u+1 νΛ(u)(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) 1 q = νΛ(u)(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) ∇u(EM (u)hkq) + Xr=u−2,u+2 q 1 q(cid:12)(cid:12)(cid:12) νΛ(u)(cid:12)(cid:12)(cid:12) ∇r(EM (u)hkq) ∇r(EM (u)hkq) q 1 q(cid:12)(cid:12)(cid:12) (5.17) For r = u − 1, u, u + 1 For r = u − 2, u + 2 ∇r(EM (u)hkq) J qc1 νΛ(u)(cid:12)(cid:12)(cid:12) νΛ(u)(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) q 1 ∇r(EM (u)hkq) ≤ νΛ(u) ∇rhkq (5.18) q 1 q(cid:12)(cid:12)(cid:12) ≤ c1νΛ(u) ∇rhkq + qq νΛ(u)(cid:0)EM (u)hkq(cid:1)− q p (cid:0)EM (u)(hkq; ∇rV (xr, xs))(cid:1)q For s ∈ {u − 3, u + 3} : s ∼ r, if we use Lemma 5.4 we obtain Is∈{u−3,u+3}:s∼r (5.19) νΛ(u)(cid:0)EM (u)hkq(cid:1)− q EM (u)(hkq; V ′(xr, xs))(cid:12)(cid:12) p (cid:12)(cid:12) ∇Λ(t)(EM (t)hkq) q 1 q C ǫ Is∈{u−1,u+1}:s∼r ≤ K ǫ I(s,t)=(u+1,u+4)∪(u−1,u−4):s∼r + νhkq (5.20) From (5.19) and (5.20) we get νΛ(t)(cid:12)(cid:12)(cid:12) νΛ(u)(cid:12)(cid:12)(cid:12) + + J qc1C ǫ J qc1K ǫ q(cid:12)(cid:12)(cid:12) q(cid:12)(cid:12)(cid:12) q 1 ∇r(EM (u)hkq) ≤ c1νΛ(u) ∇rhkq Q(t, k)I(s,t)=(u+1,u+4)∪(u−1,u−4):s∼rIs∈{u−1,u+1}:s∼r ν hkq Is∈{u−1,u+1}:s∼r (5.21) To summarise, if we plug (5.18) and (5.21) in (5.17) we finally obtain Q(u, k) ≤ J qc1C ǫ Xdist(u,t)=4 + Xr=u−1,u+1 Q(t, k) + J qc12K ǫ νΛ(u) ∇rhkq + c1 Xr=u−2,u+2 q + ν ∇uhkq ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) νΛ(u) ∇rhkq 24 Lemma 5.8. Suppose conditions (H1) is satisfied. Then for r ∈ Λ(k), the follow- ing statements are true (a) When r = {k − 2, k, k + 2} ν ∇rhkq ≤ c1ν ∇rf q + J qCc1 ǫ Q(k, k) + J qc1K ǫ (b) When r ∈ {k − 1, k + 1} q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) where hk = f − E{∼k}f . ν ∇rhkq = ν ∇rf q Proof. We will show (a). For general r ∈ Λ(k) r {k − 1, k + 1} we have ν ∇rhkq ≤ 2q−1ν ∇rf q + 2q−1ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) q q for the separate cases of r ∈ {k − 2, k + 2} (5.22) We will now compute ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) and r = k. Consider r = {k − 2, k + 2}. In this case ≤2q−1ν ∇rf q q ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) + J q2q−1ν(cid:12)(cid:12) E{∼k}(f ; ∇rV (xs, xr))(cid:12)(cid:12) q Is∈{k−1,k+1}:s∼r (5.23) If we use Lemma 3.1 (a) to bound the second term on the right hand side of (5.23) we obtain q ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) ≤2q−1ν ∇rf q + J q2q−1C ǫ Q(k, k) + J q2q−1K ǫ q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) Combining (5.22) and (5.24) together we derive ν ∇rhkq ≤ c1ν ∇rf q + J qCc1 ǫ Q(k, k) + J qc1K ǫ for K as in (H2). Consider r = k. In this case q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) (5.24) q ν(cid:12)(cid:12)∇kE{∼k}f(cid:12)(cid:12) ≤2q−1ν ∇kf q + J q2q−1ν(cid:0)E{∼k}(f ; Wk)(cid:1)q J qC2q−1 ≤2q−1ν ∇kf q + Q(k, k) + J q2q−1K ǫ q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) (5.25) ǫ 25 where in the last inequality the Lemma 3.1 (a) was used for K as in (H2). From (5.22) and (5.25) ν ∇khkq ≤ c1ν ∇kf q + J qCc1 ǫ Q(k, k) + J qc1K ǫ q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) We can now prove Lemma 5.1. Proof of Lemma 5.1. If we combine the bound for Q(k, k) from Lemma 5.7, together with the bounds for ν ∇rhkq , r = k − 2, k − 1, k, k + 1, k + 2 from Lemma 5.8, we obtain Q(k, k) ≤ Xr=k−1,k+1 ν ∇rf q + c1ν ∇kf q + J qCc1 ǫ Q(k, k) + J qc1K q ǫ ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) q(cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) ǫ Xdist(k,t)=4 J qc1C Q(t, k) (5.26) + c1 Xr=k−2,k+2(cid:18)c1ν ∇rf q + J qc1C + ǫ Xdist(k,t)=4 Q(t, k) + J qc12K ν ∇rf q + c1ν ∇kf q + J qCc1 ǫ Q(k, k) + J qc1K ǫ ǫ q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) 12)K J q(c13 + c2 ǫ ν ∇rf q + J q2C(c1 + c2 1) ǫ Q(k, k) + = Xr=k−1,k+1 + c2 1 Xr=k−2,k+2 In order to bound Pdist(k,t)=4 Q(t, k) in the above quantity the lemma bellow will be used. Lemma 5.9. Under conditions (H1)-(H3) the following inequality Xt:dist(t,k)=4 Q(t, k) ≤J qT Q(k, k) + J qT ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) ∞ 3 q + T Xr=k−2,k+2 ν ∇rf q + T J n(q−1) (ν ∇k+3+4n+rf q + ν ∇k−3−4n−rf q) Xn=0 Xr=0 is satisfied for some positive constant T independent of k. 26 The proof of Lemma 5.9 will be presented later in the section. If we use the bound of Lemma 5.9 in (5.26), we obtain + + 1 ǫ ǫ Q(k, k) ≤J q(cid:18) T J qc1C + J q(cid:18) 2Cc2 + Xr=k−1,k+1 Xn=0 J qc1C + T ∞ ǫ J n(q−1) 3 Xr=0 c13cK c2 12cK + ǫ Cc1 ǫ (cid:19) Q(k, k) + ǫ q ǫ (cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) J qc1C J qT Q(k, k) ν ∇rf q + c1ν ∇kf q + ( J qc1C ǫ T + c2 1) Xr=k−2,k+2 ν ∇rf q (ν ∇k+3+4n+rf q + ν ∇k−3−4n−rf q) (5.27) If we choose J sufficiently small such that 1 − J q(cid:18) 2Cc2 ǫ 1 + J qc1CT ǫ + Cc1 ǫ (cid:19) > 1 2 then from (5.27) we have + ǫ Q(k, k) ≤2J q(cid:18) T J qc1C + 2 Xr=k−1,k+1 Xn=0 2J qc1C + T ∞ ǫ J n(q−1) 3 Xr=0 c13cK ǫ + ν ∇rf q + 2( c2 12cK J qc1C ǫ (cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) 1) Xr=k−2,k+2 T + c2 ǫ q + 2c1ν ∇kf q ν ∇rf q (ν ∇k+3+4n+rf q + ν ∇k−3−4n−rf q) and the lemma follows for an appropriate positive constant D. It remains to show Lemma 5.9. For this we will need the following lemmata. Lemma 5.10. Under conditions (H1)-(H3) the following two bounds for Q(u, k) hold. (a) For u such that dist(u, k) ≥ 8 νΛ(u) ∇rf q + c2 νΛ(u) ∇rf q Q(u, k) ≤c1νΛ(u) ∇uf q + c1 Xr=u−1,u+1 J qc1C + ǫ Xdist(u,t)=4 Q(t, k) + J qc12K ǫ 1 Xr=u−2,u+2 ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) q 27 (b) For u such that dist(u, k) = 4 Q(u, k) ≤c1ν ∇uf q + c2 1 Xr=u−2,u+2 + c1 Xr=u−1,u+1 ν ∇rf q + ν ∇rf q + J q(cid:18)c12K ǫ + J qCc2 1 ǫ Q(k, k) + J qc1C ǫ q c2 1K ǫ (cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) Xdist(u,t)=4,t6=k Q(t, k) Proof. The lemma follows from the bound of Q(u, k) in Lemma 5.7. In the case where dist(u, k) ≥ 8, for r = u − 2, u − 1, u, u + 1, u + 2 we have that ν ∇rhkq ≤ 2q−12ν ∇rf q (5.28) Substituting (5.28) in the expression from Lemma 5.7 we immediately obtain (a). Consider the case where dist(u, k) = 4. Then for r = u − 1, u, u + 1 ν ∇rhkq ≤ 2q−12ν ∇rf q (5.29) While for r = {u−2, u+2} we can bound ν ∇rhkq from Lemma 5.8 (a). If we plug the bounds from (5.29) and Lemma 5.8 (a) into the expression from Lemma 5.7, we obtain Q(u, k) ≤J q(cid:18) c12K ǫ + + c1 Xr=u−1,u+1 c2 1K q ǫ (cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) 1 Xr=u−2,u+2 ν ∇rf q + c2 + J qCc2 1 ǫ Q(k, k) + c1ν ∇uf q ν ∇rf q + J qc1C ǫ Xdist(u,t)=4 Q(t, k) Before proving Lemma 5.9, we will also need to show that for any k ∈ N sup n∈N Xdist(u,k)=n Q(u, k) < Cf < ∞ for Cf a constant which depends on the function f but not on n, u and k. To show this we first need the following lemma. Lemma 5.11. For any r, k ∈ Z we have ν ∇rhkq ≤ Cf < ∞ where Cf depends on the function f but not on r and k. 28 Proof. For general r ∈ {k − 2, k, k + 2} ν ∇rhkq ≤ 2q−1ν ∇rf q + 2q−1ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) q since hk = f − E{∼k}f . For the second term on the right hand side of (5.30) we have (5.30) q ν(cid:12)(cid:12)∇rE{∼k}f(cid:12)(cid:12) where ≤2q−1ν ∇rf q + J q2q−1ν(cid:12)(cid:12) E{∼k}(f ; Zk)(cid:12)(cid:12) q (5.31) Zk = ∇k−2V (xk−2, xk−1)Ir=k−1 + ∇k+2V (xk+2, xk+1)Ir=k+1 + WkIr=k where Wk as in (4.1). We will now compute the last term on the right hand side of (5.31) q q ≤ νf qZk − E{∼k}Zkq If we use the entropic inequality (3.3) we obtain q E{∼k}(f ; Zk)(cid:12)(cid:12) ν(cid:12)(cid:12) E{∼k}(f − E{∼k}f )(Zk − E{∼k}Zk)(cid:12)(cid:12) =ν(cid:12)(cid:12) E{∼k}(cid:0)f (Zk − E{∼k}Zk)(cid:1)(cid:12)(cid:12) =ν(cid:12)(cid:12) E{∼k}(f ; Zk)(cid:12)(cid:12) ν(cid:12)(cid:12) νf q log ≤ q ≤ νf q log f q νf q + f q νf q + 1 ǫ K ǫ 1 ǫ 1 ǫ νf q νf q log νeǫZk−E{∼k}Zkq (5.32) where K as in (H2). r ∈ {k − 2, k, k + 2} If we combine (5.30), (5.31) and (5.32) we get that for ν ∇rhkq ≤ 2qν ∇rf q + J q22q−2 ǫ νf q log f q νf q + J q22q−2 ǫ νf q For r /∈ {k − 2, k, k + 2} we have ν ∇rhkq ≤ 2qνf q (5.33) (5.34) From (5.33) and (5.34) the lemma follows since functions f are as in Remark 2.2. Lemma 5.12. If (H2) is satisfied, then for any k ∈ N sup n∈N Xdist(u,k)=n Q(u, k) < Cf < ∞ where Cf is a constant which depends on the function f but not on u and k. 29 Proof. Since we work on the one dimensional lattice, it is sufficient to show that Q(u, k) < C ′ f < ∞ sup n∈N f depends only on the functions f . To compute Q(u, k) we can use (5.17) for C ′ and (5.18) to obtain Q(u, k) ≤ Xr=u−1,u,u+1 ν ∇rhkq + Xr=u−2,u+2 Furthermore, from (5.19) for r = u − 2, u + 2 we have ∇r(EM (u)hkq) νΛ(u)(cid:12)(cid:12)(cid:12) ∇r(EM (u)hkq) νΛ(u)(cid:12)(cid:12)(cid:12) where q 1 q(cid:12)(cid:12)(cid:12) ≤ c1ν ∇rhkq + J qc1 qq I0 q 1 q(cid:12)(cid:12)(cid:12) (5.35) (5.36) I0 := νΛ(u)(cid:0)EM (u)hkq(cid:1)− q p (cid:0)EM (u)(hkq; ∇rV (xr, xs))(cid:1)q In order to bound the second term on the right hand side of (5.36) we compute Is∈{u−3,u+3}:s∼r EM (u)(hkq; ∇rV (xr, xs)) = EM (u)(cid:0)hk(q−1)+1(cid:0)∇rV (xr, xs) − EM (u)∇rV (xr, xs)(cid:1)(cid:1) 1 q 1 From the last bound, since p and q are conjugate, we get q(cid:17)(cid:17) p (cid:16)EM (u)(cid:16)hkq(cid:12)(cid:12)∇rV (xr, xs) − EM (u)∇rV (xr, xs)(cid:12)(cid:12) q(cid:17) Is∈{u−3,u+3}:s∼r ≤(cid:0)EM (u)hkpq−p(cid:1) I0 ≤ νΛ(u)EM (u)(cid:16)hkq(cid:12)(cid:12)∇rV (xr, xs) − EM (u)∇rV (xr, xs)(cid:12)(cid:12) where above we denoted Nr =(cid:12)(cid:12)∇rV (xr, xs) − EM (u)∇rV (xr, xs)(cid:12)(cid:12) = ν(hkqNr) ≤ 2q−1ν(f qNr) + 2q−1ν((E{∼k}f q)Nr) If we use again the entropic inequality (3.3) we obtain q Is∈{u−3,u+3}:s∼r. I0 ≤ νf q log νf q log νeǫNr + νE{∼k}f q log 2q−1 ǫ E{∼k}f q νE{∼k}f q 2q−1 ǫ + f q νf q + 2q−1 ǫ 2q−1 ǫ log νeǫNr νE{∼k}f q ≤ 2q−1 ǫ νf q log f q νf q + 2qK ǫ νf q + 2q−1 ǫ νE{∼k}f q log E{∼k}f q νE{∼k}f q (5.37) where K as in (H2). For the last term on the right hand side of (5.37) we can write νE{∼k}f q log E{∼k}f q νE{∼k}f q = νf q log E{∼k}f q νE{∼k}f q ≤ νf q log f q νf q (5.38) 30 Combining together (5.37) and (5.38) we obtain I0 ≤ 2q ǫ νf q log f q νf q + 2qK ǫ νf q (5.39) From (5.36), and (5.39) we then get that for r = u − 2, u + 2 J q2qc1 qqǫ νf q log f q νf q + J qc12qK qqǫ νf q (5.40) ≤ c1ν ∇rhkq + q 1 q(cid:12)(cid:12)(cid:12) ∇r(EM (u)hkq) νΛ(u)(cid:12)(cid:12)(cid:12) any function f there is a bound of νΛ(u)(cid:12)(cid:12)(cid:12) q 1 q(cid:12)(cid:12)(cid:12) If we combine (5.35) and (5.40) together with Lemma 5.11 we conclude that for ∇r(EM (u)hkq) uniformly with respect to the set M(u) depending only on νf q, maxi∈Z ν ∇if q and νf q log f q νf q . We can now prove Lemma 5.9. Proof of Lemma 5.9. For every u s.t. dist(u, k) ≥ 8 define G(u, k) :=c1νΛ(u) ∇uf q + c1 Xr=u−1,u+1 ν ∇rf q + c2 1 Xr=u−2,u+2 νΛ(u) ∇rf q + and for every u s.t. dist(u, k) = 4 define J qc12K ǫ q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) G(u, k) :=c1ν ∇uf q + c1 Xr=u−1,u+1 ν ∇uf q + J qCc2 1 ǫ Q(k, k) + c2 1 Xi=u−2,u+2 ν ∇rf q + J q(cid:18)c12K ǫ + c2 1K q ǫ (cid:19) ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) If we set K ′ = c1C ǫ , then from Lemma 5.10 (a) and (b) respectively we can write Q(u, k) ≤ G(u, k) + J qK ′ Xdist(u,t)=4 Q(t, k), for dist(u, k) ≥ 8 (5.41) and Q(u, k) ≤ G(u, k) + J qK ′Q(t, k)Idist(t,u)=4,t6=k, for dist(u, k) = 4 (5.42) 31 From equation (5.41) we obtain Xdist(u,k)=n Q(u, k) ≤ Xdist(u,k)=n G(u, k) + J qK ′ Xdist(u,k)=n Xdist(t,u)=4 Q(t, k) or equivalently Xdist(u,k)=n Q(u, k) ≤ Xdist(u,k)=n G(u, k) + J qK ′ Xdist(t,k)=n+4 Q(t, k) + J qK ′ Xdist(t,k)=n−4 Q(t, k) which implies where we denote Q(n) ≤ G(n) + J qK ′ Q(n − 4) + J qK ′ Q(n + 4) (5.43) Q(n) = Xdist(u,k)=n Q(u, k) and G(n) = Xdist(u,k)=n G(u, k) While from equation (5.42), we have Xdist(u,k)=4 Q(u, k) ≤ Xdist(u,k)=4 G(u, k) + J qK ′ Xdist(u,k)=4 Q(t, k)Idist(t,u)=4,t6=k This implies Xdist(u,k)=4 Q(u, k) ≤ Xdist(u,k)=4 G(u, k) + J qK ′ Xdist(t,k)=8 Q(t, k) which is equivalent to Q(4) ≤ G(4) + J qK ′ Q(8) (5.44) Choose J in (H3) sufficiently small such that hypothesis (5.10) of Lemma 5.6 is satisfied. Then, since relationships (5.43), (5.44) and Lemma 5.12 are true, the conditions of Lemma 5.6 are satisfied for P = Q and G = G and so we obtain Q(4) ≤ J +∞ Xn=0 J nq−n G(4n + 4) 1 1−J 2q−2 . This is equivalent to where J = G(u, k) Xt:dist(t,k)=4 Q(t, k) ≤ J Xdist(u,k)=4 Xn=1 + J +∞ J nq−n Xdist(u,k)=4n+4 G(u, k) (5.45) 32 Substituting G(u, k) leads to Q(t, k) ≤ J qCc2 1 Xt:dist(t,k)=4 νΛ(u) ∇uf q ǫ Q(k, k) J Xdist(u,k)=4 J nq−n Xdist(u,k)=4n+4 J nq−n Xdist(u,k)=4n+4 Xr=u−1,u+1 J nq−n Xdist(u,k)=4n+4 Xr=u−2,u+2 +∞ + J c1 + J c1 + J c2 1 +∞ +∞ Xn=0 Xn=0 Xn=0 +∞ + J q J c1K ǫ (c1 + 2) Xn=0 J nq−n Xdist(u,k)=4n+4 νΛ(u) ∇rf q νΛ(u) ∇rf q q ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) n=0 J nq−n < ∞. (5.46) (5.46) then implies But for J in (H3) we have J q−1 < 1 which implies J = P+∞ Pt:dist(t,k)=4 Q(t, k) ≤ J q Cc2 ǫ 2 JQ(k, k) 1 + Jc1 JPdist(u,k)=4n+4Pr=u−1,u,u+1 ν ∇rf q JPdist(u,k)=4n+4Pr=u−2,u+2 ν ∇rf q + Jc2 1 q +J q c1K ǫ (c1 + 2) J 2 1−J q−1 ν(cid:12)(cid:12)f − E{∼k}f(cid:12)(cid:12) and the lemma follows for appropriate constant T > 0. 6 Conclusion In the present work, we have determined conditions for the infinite volume Gibbs measure to satisfy the Log-Sobolev Inequality. As explained in the introduction, the criterion presented in Theorem 2.3 can in particular be applied in the case of local specifications {EΛ,ω}Λ⊂⊂Z,ω∈Ω with no quadratic interactions for which k∇i∇jV (xi, xj)k∞ = ∞ Thus, we have shown that our results can go beyond the usual uniform boundness of the second derivative of the interactions considered in [Z1], [Z2], [M] and [O-R]. 33 Concerning the additional conditions (H1) and (H2) placed here to handle the exotic interactions, they refer to finite dimensional measures with no boundary conditions which are easier to handle than the {EΛ,ω}Λ⊂⊂Z,ω∈Ω measures or the infinite dimensional Gibbs measure ν. In fact, the following results concerning the conditions can be proven. This is a work in progress that will consist the material of a forthcoming paper. Proposition 6.1. The hypothesis (H0), (H3) and (H2) imply hypothesis (H1). Consequently, the main result of Theorem 2.3 is then reduced to the following Theorem 6.2. If hypothesis (H0), (H3) and (H2) are satisfied, then the infinite dimensional Gibbs measure ν for the local specification {EΛ,ω}Λ⊂⊂Z,ω∈Ω satisfies the q Log-Sobolev inequality ν f q log f q ν f q ≤ C ν ∇f q for some positive constant C independent of f . Concerning examples of measures that satisfy the above conditions, one can consider measures with phase φ(x) = xt with t ≥ q q−1 and interaction V (x, y) = x − yr, with max{r, (r − 1)q} < t. The main idea of the proof of the Proposition 6.1 follows in main lines the method followed in the current paper. Although some of the details are more involved because of the lack of hypothesis (H1), the fact that in Proposition 6.1 the Gibbs measure is localised and thus the approximation procedure starts from a finite set compensates for the loss of the LSq for νΛ(i). In this paper we have been concerned with the q Logarithmic Sobolev inequality for measures on the 1 dimensional Lattice Z. It is interesting to try to extend the current result to a higher dimensional lattice on Zd, d ≥ 2, although this does not appear to be immediate. In a different direction, we can consider the following class of modified Logarithmic Sobolev inequalities presented in [G-G-M]: ν f 2 log f 2 ν f 2 ≤ C Z Ha,c(cid:18)∇f f (cid:19) f 2dν (6.1) for some positive constant C, where x2 2 a2−β xβ +∞ β + a2 β−2 2β if if if x ≤ a x ≥ a and c 6= 1 x ≥ a and c = 1 Ha,c(x) =  34 for c ∈ [1, 2], a > 0 and β satisfying 1 β = 1 (β ≥ 2). This new class of inequal- ities is an interpolation between Log-Sobolev (LS2) and Spectral Gap inequalities c + 1 (SG2), which retains the basic properties of the Log-Sobolev inequalities mentioned in Remark 1.1. Some preliminary results suggest that on Zd, d ≥ 2, the infinite dimensional Gibbs measure satisfies a [G-G-M] type inequality with β = 2q, under hypothesis (H0) for LSq (1 < q < 2) and some hypothesis stronger than (H2). This is work in early stages, but hopefully a modified LS inequality comparable to the [G-G-M] inequalities can be obtained in the case of the higher dimensional lattice. In addition, it is interesting to investigate whether the result presented in this paper can be extended to the family of weaker inequalities presented in [G-G-M], assuming (H0) and (H1) for the (6.1) inequality instead of the LSq. However, this does not seem to be immediate especially in showing the sweeping out relationships and so more work needs to be done towards this direction. Furthermore, concerning the hypothesis on the single-site measure, the main hypothesis (H0) for E{i},ω can be reduced to the same assumption for the boundary free single-site measure, that is (H0′): The single-site measure e−φ(x)dx R e−φ(x)dx satisfies the LSq Inequality. Measures as in (H0′) do not involve boundary conditions and for this reason it is easier to show that they satisfy the Log-Sobolev inequality. For instance, when in R one can think of phases that are convex and increase sufficiently fast, like φ(x) = xp for p > 2 (see [B-Z]). In the case of the Heisenberg group H one can consider φ(x) = βd(x)p with p conjugate of q (see [H-Z]). However, that does not mean that condition (H0′) is in general weaker than condition (H0) as there are examples of single-site boundary free measures e−φ(x)dx R e−φ(x)dx that do not satisfy the LSq inequality, which when perturbed with interactions, give new measures E{i},ω that satisfy the Log-Sobolev-q inequality uniformly on the boundary conditions, that is condition (H0) is satisfied. In addition, in the case of hypothesis (H0′), it seems that the analogues of Proposition 6.1 and Theorem 6.2 will be more to difficult to be shown. Acknowledgements: The author would like to thank Prof. Boguslaw Zegarlinski for his valuable comments and suggestions. References [B-J-S] V. Bach, T. Jecko and J. Sjostrand, Correlation Asymptotics of Classical Lattice spin Systems with Nonconvex Hamilton function at Low Temperature, Ann. Henri Poincare, 1, 59-100 (2000) 35 [B-E] [B-HK] [B-G] [B-L] [B-Z] [B-H] [D-S] [D] D. Bakry and M. Emery, Diffusions hypercontractives ,S´eminaire de Probabilit´es XIX, Springer Lecture Notes in Math., 1123, 177-206 (1985) J. Bellisard and R. Hoegn-Krohn, Compactness and the maximal Gibbs state for random fields on the Lattice, Commun. Math. Phys., 84, 297-327 (1982). S.G. Bobkov and F. Gotze, Exponential integrability and transporta- tion cost related to logarithmic sobolev inequalities , J of Funct Anal., 163, 1-28 (1999) S.G. Bobkov and M. Ledoux, From Brunn-Minkowski to Brascamp- Lieb and to Logarithmic Sobolev Inequalities, Geom. funct. anal., 10, 1028-1052 (2000). S.G. Bobkov and B. Zegarlinski, Entropy Bounds and Isoperime- try. Memoirs of the American Mathematical Society, 176, 1 - 69 (2005) T. Bodineau and B. Helffer, Log-Sobolev inequality for unbounded spin systems , J of Funct Anal., 166, 168-178 (1999) J.D. Deuschel and D. Stroock, Large Deviations , Academic Press, San Diego, (1989). R. L. Dobrushin, The problem of uniqueness of a Gibbs random field and the problem of phase transition, Funct. Anal. Apll., 2, 302-312 (1968). [G-G-M] I. Gentil, A. Guillin and L. Miclo, Modified logarithmic Sobolev inequalities and transportation inequalities, Probab.Theor. Relat. Fields 133, 409-436 (2005). [G-R] [G] [G-Z] I. Gentil and C. Roberto, Spectral Gaps for Spin Systems: Some Non-convex Phase Examples, J. Func. Anal., 180, 66-84 (2001). L. Gross, Logarithmic Sobolev inequalities, Am. J. Math., 97, 1061- 1083 (1976) A.Guionnet and B.Zegarlinski, Lectures on Logarithmic Sobolev Inequalities, IHP Course 98, S´eminaire de Probabilit´es XXVI, Lecture Notes in Mathematics 1801, Springer, 1-134 (2003). 36 [H] B. Helffer, Semiclassical Analysis, Witten Laplacians and Statisti- cal Mechanics. Partial Differential Equations and Applications. World Scientific, Singapore (2002) [H-Z] W.Hebisch and B.Zegarlinski, Coercive inequalities on metric mea- sure spaces. J of Funct Anal. (to appear) [I-P] [L] [M] [M-M] [O-R] J. Inglis and I. Papageorgiou, Logarithmic Sobolev Inequalities for Infinite Dimensional Hormander Type Generators on the Heisenberg Group. Potential Anal., 31, 79-102 (2009) M. Ledoux, Concentration of measure and logarithmic Sobolev inequal- ities, S´eminaire de Probabilit´es, XXXIII, Lecture Notes in Math. 1709, Springer-Verlag, 120-216. (1999) K.Marton, Logarithmic Sobolev Inequality for Weakly Dependent Ran- dom Variables. (preprint) O. Matte and J. S. Moller, On the spectrum of semi-classical Witten-Laplacians and Schrdinger operators in large dimension, J of Funct Anal, Volume 220 (2), 243-264 (2005) F. Otto and M. Reznikoff, A new criterion for the Logarithmic Sobolev Inequality and two Applications , J. Func. Anal., 243, 121-157 (2007). [P] C.J.Preston,Random Fields, LNM 534, Springer (1976) [R-Z] [Y] [Z1] [Z2] C. Roberto and B. Zegarlinski, Orlicz-Sobolev inequalities for sub- Gaussian measures and ergodicity of Markov semi-groups , J. Func. Anal., 243 (1), 28 -- 66 (2007). N.Yoshida, The log-Sobolev inequality for weakly coupled lattice field, Probab.Theor. Relat. Fields 115 , 1-40 (1999) B. Zegarlinski, On log-Sobolev Inequalities for Infinite Lattice Sys- tems, Lett. Math. Phys. 20, 173-182 (1990) B. Zegarlinski, The strong decay to equilibrium for the stochastic dynamics of unbounded spin systems on a lattice, Comm. Math. Phys. 175, 401-432 (1996) 37
1512.04709	1	1512	2015-12-15T10:25:01	On the stability of the linear functional equation in a single variable on complete metric groups	[ "math.FA" ]	In this paper we obtain a result on Hyers-Ulam stability of the linear functional equation in a single variable $f(\varphi(x)) = g(x) \cdot f(x)$ on a complete metric group.	math.FA	math	On the stability of the linear functional equation in a single variable on complete metric groups Soon-Mo Jung, Dorian Popa, Michael Th. Rassias Mathematics Section, College of Science and Technology, Hongik University, 339–701 Sejong, Republic of Korea E-mail: [email protected] Department of Mathematics, Technical University of Cluj-Napoca, 28 Memorandumului, 400114 Cluj-Napoca, Romania E-mail: [email protected] Department of Mathematics, ETH–Zurich, Ramistrasse 101, 8092 Zurich, Switzerland E-mail: [email protected] Abstract. In this paper we obtain a result on Hyers-Ulam stability of the linear functional equation in a single variable f (ϕ(x)) = g(x) · f (x) on a complete metric group. 1 Introduction Hyers-Ulam stability is one of the main topics in the theory of functional equations. Generally a functional equation is said to be stable provided, for any function f satisfying the perturbed functional equation, there exists an exact solution f0 of that equation which is not far from the 0Mathematics Subject Classification (2010): 65J15. 0Key words and phrases: Nonlinear operator, stability, functional equation, com- plete metric group, inequalities, Banach spaces, operator mapping, Euler-Mascheroni constant. 1 given f . Based on this concept, the study of the stability of functional equations can be regarded as a branch of optimization theory. (We can find some applications of the Hyers-Ulam stability to optimization theory and economics in [11].) It seems that the first result on the stability of functional equations appeared in the famous book by Gy. P´olya and G. Szego [18] and con- cerns the Cauchy functional equation on the set of positive integers. But the starting point of the stability theory of functional equations is due to S.M. Ulam who formulated a question concerning the perturbation of homomorphisms on metric groups. The first result for Ulam’s problem was obtained by D.H. Hyers for the Cauchy functional equation on Ba- nach spaces. Due to the question of Ulam and the answer of Hyers the stability of functional equations is called after their names. For more details on Hyers-Ulam stability of functional equations and optimization theory we refer the reader to [2, 4, 12, 13, 15, 17, 19]. The functional equation f (ϕ(x)) = g(x)f (x) + h(x), (1.1) where f is the unknown function and g, h, ϕ are given functions, is called the linear functional equation in a single variable. For particular cases of g and h in (1.1) we obtain some classical functional equations. We mention here some of them as • Abel’s equation • Schroder’s equation f (ϕ(x)) = f (x) + c f (ϕ(x)) = cf (x) • Gamma functional equation f (x + 1) = xf (x) • Digamma functional equation f (x + 1) = f (x) + 1 x . 2 (1.2) (1.3) (1.4) (1.5) Recall that Digamma function ψ0 : R∗ + → R is defined by ψ0(x) = d dx ln Γ(x) = Γ′(x) Γ(x) , ∀ x ∈ R∗ +, (1.6) where Γ(x) =Z ∞ 0 tx−1e−tdt, ∀ x ∈ R∗ +, (1.7) R+ stands for the set of all nonnegative numbers, i.e., R+ = [0, ∞) and R∗ + = (0, ∞). For more details on the functional equation (1.1) and its particular cases we refer to [16] and the references therein. It seems that the first result on stability for the equation (1.1) was obtained in 1970 by J. Brydak [3]. A generalized Hyers-Ulam stability of the gamma func- tional equation was obtained by S.-M. Jung in [14]. A nice result on generalized Hyers-Ulam stability of the equation (1.1) was obtained by T. Trif [20] for functions f acting from an arbitrary nonempty set S into a Banach space X. Some recent results on the stability and nonstability of the equation (1.1) and the linear functional equation of higher order in a single variable were obtained by J. Brzdek, D. Popa, B. Xu (see [6, 7, 8, 9, 10]). The goal of this paper is to study the Hyers-Ulam stability of the homogeneous linear functional equation (1.1) for functions defined from an arbitrary nonempty set S into a complete metric group (G, ·, d), i.e., (G, ·) is a group, (G, d) is a complete metric space, the group’s binary operation and the inverse operation are continuous with respect to the product topology on G × G and the topology generated by the metric d on G, respectively. 2 Stability of linear functional equation Let S be a nonempty set, (G, ·, d) a complete metric group with the metric d invariant to left translations, i.e., d(x · y, x · z) = d(y, z), ∀ x, y, z ∈ G, (2.1) and let ϕ : S → S, g : S → G be given functions. An example of metric invariant to left translations is the metric induced by a norm. 3 We deal with the Hyers-Ulam stability of the linear functional equa- tion f (ϕ(x)) = g(x) · f (x), (2.2) where f : S → G is the unknown function. Let RS + be the class of all functions ε : S → R+. We study the generalized Hyers-Ulam stability of the equation (2.2) in the sense defined in [4]. Definition 2.1 Let C ⊆ RS C into RS respectively) provided for every ε ∈ C and f : S → G with + be nonempty and T be an operator mapping +. We say that the equation (2.2) is T -stable (with uniqueness, d(f (ϕ(x)), g(x) · f (x)) ≤ ε(x), ∀ x ∈ S there exists a (unique, respectively) solution f0 : S → G of the equation (2.2) such that d(f (x), f0(x)) ≤ T ε(x), ∀ x ∈ S. If ε is a constant function in the previous definition then the equation (2.2) is said to be stable in Hyers-Ulam sense. By ϕk, k ∈ N0 = N ∪ {0} we denote the k-th iterate of the function ϕ, ϕ0 = 1S, ϕk = ϕ ◦ ϕk−1, k ∈ N. The main result is contained in the next theorem. Theorem 2.2 Let ε : S → R+ be a function with the property ε(ϕn(x)) = Φ(x), ∀ x ∈ S, (2.3) ∞ Xn=0 where Φ : S → R+. Then for every function f : S → G satisfying the inequality d(f (ϕ(x)), g(x) · f (x)) ≤ ε(x), ∀ x ∈ S, (2.4) there exists a unique solution f0 : S → G of the functional equation (2.2) such that d(f (x), f0(x)) ≤ Φ(x), ∀ x ∈ S. (2.5) 4 Proof. Existence. Let f : S → G be a function satisfying (2.4). Then the following relation holds: d f(cid:0)ϕn(x)(cid:1), n Yk=1 g(cid:0)ϕk−1(x)(cid:1) · f (x)! ≤ n Xk=1 ε(cid:0)ϕk−1(x)(cid:1) (2.6) for all x ∈ S and n ∈ N. We prove (2.6) by induction on n. Since the group (G, ·) is not generally commutative, we let ak := an · an−1 · . . . · ap, n Yk=p where ak ∈ G for p ≤ k ≤ n. For n = 1 the relation (2.6) holds in view of (2.4). We suppose that (2.6) holds for some n ∈ N and for all x ∈ S, and we prove that x ∈ S. Hence (2.6) holds for all x ∈ S and n ∈ N. Now let (εn)n≥1 be the sequence of functions defined by εn(x) := n Yk=1 g(cid:0)ϕk−1(x)(cid:1)!−1 · f (ϕn(x)), n ∈ N, x ∈ S. (2.7) 5 Indeed, it follows from (2.1), (2.4) and (2.6) that d f(cid:0)ϕn+1(x)(cid:1), n+1 n+1 n+1 Yk=1 ε(cid:0)ϕk−1(x)(cid:1), g(cid:0)ϕk−1(x)(cid:1) · f (x)! ≤ Xk=1 Yk=1 d f(cid:0)ϕn+1(x)(cid:1), g(cid:0)ϕk−1(x)(cid:1) · f (x)! ≤ d(cid:0)f(cid:0)ϕn+1(x)(cid:1), g(ϕn(x)) · f (ϕn(x))(cid:1) + d g(ϕn(x)) · f (ϕn(x)), g(cid:0)ϕk−1(x)(cid:1) · f (x)! Yk=1 ≤ ε(ϕn(x)) + d f (ϕn(x)), g(cid:0)ϕk−1(x)(cid:1) · f (x)! Yk=1 Xk=1 ε(cid:0)ϕk−1(x)(cid:1), x ∈ S. n+1 ≤ n+1 n We prove that (εn(x))n≥1 is a Cauchy sequence in (G, ·, d) for all x ∈ S, where a−1 means the inverse of the element a in the group G. Using (2.1) and (2.6), we have g(cid:0)ϕk−1(x)(cid:1)!−1 · f (ϕn(x))  (2.8) d(cid:0)εn+p(x), εn(x)(cid:1) = d g(cid:0)ϕk−1(x)(cid:1)!−1  n+p · f(cid:0)ϕn+p(x)(cid:1), n Yk=1 Yk=1 = d · f(cid:0)ϕn+p(x)(cid:1), f (ϕn(x)) g(cid:0)ϕk−1(x)(cid:1)!−1 n+p Yk=n+1   Xk=0 Xk=1 ε(cid:0)ϕn+k(x)(cid:1) ε(cid:0)ϕk−1(ϕn(x))(cid:1) ≤ Now rn(x) := P∞ for x ∈ S and n, p ∈ N. ∞ p ≤ k=0 ε(ϕn+k(x)), n ∈ N, is the remainder of order n of the convergent series (2.3), so limn→∞ rn(x) = 0 for all x ∈ S. We conclude that (εn(x))n≥1 is a Cauchy sequence, therefore it is convergent since G is a complete metric group. Define the function f0 by f0(x) = lim n→∞ εn(x), x ∈ S. The relation (2.8), for p = 1, leads to d(εn+1(x), εn(x)) ≤ ∞ Xk=0 ε(cid:0)ϕn+k(x)(cid:1), n ∈ N, x ∈ S. (2.9) Taking account of εn+1(x) = g(x)−1 · εn(ϕ(x)) and letting n → ∞ in (2.9) it follows that d(cid:0)g(x)−1 · f0(ϕ(x)), f0(x)(cid:1) = 0 which is equivalent to f0(ϕ(x)) = g(x) · f0(x), x ∈ S, i.e., f0 is a solution of the equation (2.2). On the other hand, the relations (2.1) and (2.6) lead to d(εn(x), f (x)) ≤ n Xk=1 ε(cid:0)ϕk−1(x)(cid:1) 6 (2.10) for all x ∈ S and n ∈ N, therefore letting n → ∞ in (2.10) we get d(f0(x), f (x)) ≤ Φ(x), which completes the proof of the existence. Uniqueness. Assume that for a function f satisfying (2.4) there exist two solutions f1, f2 of the equation (2.2) satisfying d(f (x), fi(x)) ≤ Φ(x), ∀ x ∈ S, i ∈ {1, 2} and f1 6= f2. Taking into account that f1, f2 satisfy (2.2), it follows easily that fi(ϕn(x)) = n Yk=1 g(cid:0)ϕk−1(x)(cid:1) · fi(x), n ∈ N, x ∈ S, i ∈ {1, 2}, and hence d(f1(x), f2(x)) · f1(ϕn(x)), n Yk=1 = d g(cid:0)ϕk−1(x)(cid:1)!−1  n Yk=1 = d(cid:0)f1(ϕn(x)), f2(ϕn(x))(cid:1) ≤ d(cid:0)f1(ϕn(x)), f (ϕn(x))(cid:1) + d(cid:0)f (ϕn(x)), f2(ϕn(x))(cid:1) g(cid:0)ϕk−1(x)(cid:1)!−1 ≤ 2Φ(ϕn(x)), x ∈ S, n ∈ N. · f2(ϕn(x))  Since limn→∞ Φ(ϕn(x)) = limn→∞ rn(x) = 0, x ∈ S, it follows that f1(x) = f2(x), which completes the proof. (cid:3) The Digamma function ψ0 : R∗ + → R is defined by (1.6). The Digamma function is frequently called the psi function and it satisfies the Digamma functional equation (1.5) for all x ∈ R∗ +. Indeed, we know that ψ0 is the unique solution of the functional equation (1.5) which is monotone on R∗ + and satisfies ψ0(1) = −γ, where γ = 0.577215 . . . is the Euler-Mascheroni constant (see [1, §6.3] and [21, §6.11.5]). The gamma function defined by (1.7) satisfies the functional equation Γ(x + 1) = xΓ(x), ∀ x ∈ R∗ +. 7 If we take the logarithmic values from both sides of the last equation, then we have ln Γ(x + 1) = ln Γ(x) + ln x, ∀ x ∈ R∗ +. We differentiate each side of the above equality with respect to x to get d dx ln Γ(x + 1) = d dx ln Γ(x) + 1 x , ∀ x ∈ R∗ +. In view of (1.6), we know that the Digamma function ψ0 is a solution of the Digamma functional equation (1.5). The generalized Hyers-Ulam stability of the Digamma functional equation (1.5) follows from Theorem 2.2. Corollary 2.3 Let ε : R∗ + → R+ be a function with the property ε(x + n) = Φ(x), ∀ x ∈ R∗ +. ∞ Xn=0 Then for every function f : R∗ + → R satisfying f (x + 1) − f (x) − ≤ ε(x), x ∈ R∗ + (cid:12)(cid:12)(cid:12)(cid:12) 1 x(cid:12)(cid:12)(cid:12)(cid:12) there exists a unique solution f0 : R∗ that + → R of the equation (1.5) such f (x) − f0(x) ≤ Φ(x), ∀ x ∈ R∗ +. Proof. Take S = R∗ d the Euclidian metric on R and g(x) = 1/x, x ∈ R∗ follows in view of Theorem 2.2. +, ϕ(x) = x + 1, G = R with the usual addition and +. Then the result (cid:3) References [1] M. Abramowitz, I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, Dover Publ., New York, 1965. 8 [2] R.P. Agarwal, B. Xu, W. Zhang, Stability of functional equations in single variable, J. Math. Anal. Appl., 288 (2003), 852–869. [3] J. Brydak, On the stability of the functional equation ϕ[f (x)] = g(x)ϕ(x) + F (x), Proc. Amer. Math. Soc., 26 (1970), 455–460. [4] J. Brzdek, N. Brillouet-Bellout, K. Ciepli´nski, On some recent develop- ments in Ulam’s type stability, Abstr. Appl. Anal. 2012 (2012), Art. ID 716936. [5] J. Brzdek, J. Chudziak, Zs. P´ales, A fixed point approach to stability of functional equations, Nonlinear Anal., 74 (2011), 6728–6732. [6] J. Brzdek, D. Popa, B. Xu, The Hyers-Ulam stability of linear equations of higher orders, Acta Math. Hungar., 120 (2008), 1–8. [7] J. Brzdek, S.-M. Jung, A note on stability of an operator equation of the second order, Abstr. Appl. Anal., 2011 (2011), Art. ID 602713, 15 pp. [8] J. Brzdek, D. Popa, B. Xu, On approximate solutions of the linear func- tional equation of higher order, J. Math. Anal. Appl., 373 (2011), 680–689. [9] J. Brzdek, D. Popa, B. Xu, Note on nonstability of the linear functional equation of higher order, Comput. Math. Appl., 62 (2011), 2648–2657. [10] J. Brzdek, D. Popa, B. Xu, Selections of set-valued maps satisfying a linear inclusion in a single variable, Nonlinear Anal., 74 (2011), 324–330. [11] E. Castillo and M. R. Ruiz-Cobo, Functional Equations and Modelling in Science and Engineering, Marcel Dekker, New York/Basel/Hong Kong, 1992. [12] L. Cesaro, Optimization Theory and Applications. Problems with Or- dinary Differential Equations, Springer, New York/Heidelberg/Berlin, 1983. [13] S. Czerwik, Functional Equations and Inequalities in Several Variables, World Scientific, River Edge, NJ, 2002. [14] S.-M. Jung, On the modified Hyers-Ulam-Rassias stability of the func- tional equation for gamma function, Mathematica (Cluj), 39(62) (1997), 235–239. 9 [15] S.-M. Jung, Hyers-Ulam-Rassias Stability of Functional Equations in Nonlinear Analysis, Springer Optimization and its Applications, Vol. 48, Springer, New York, 2011. [16] M. Kuczma, Functional Equations in a Single Variable, Pa´nstwowe Wydawnictwo Naukowe, Warszawa, 1968. [17] P. M. Pardalos and T. F. Coleman (eds.), Lectures on Global Optimiza- tion, Fields Institute Communications, Amer. Math. Soc., 2009. [18] G. Polya, G. Szego, Aufgaben und Lehrsatze aus der Analysis I, Julius Springer, Berlin, Germany, 1925. [19] T. T. Rockafellar, Convex Analysis, Princeton Univ. Press, Princeton, 1972. [20] T. Trif, On the stability of a general gamma-type functional equation, Publ. Math. Debrecen, 60 (2002), 47–61. [21] D. Zwillinger, Standard Mathematical Tables and Formulae (31th Ed.), Chapman & Hall/CRC, 2003. 10
1809.00651	1	1809	2018-09-03T16:43:46	A note on (asymptotically) Weyl-almost periodic properties of convolution products	[ "math.FA" ]	The main aim of this paper is to investigate Weyl-$p$-almost periodic properties and asymptotically Weyl-$p$-almost periodic properties of convolution products. In such a way, we continue several recent research studies of ours which do concern a similar problematic.	math.FA	math	A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC PROPERTIES OF CONVOLUTION PRODUCTS VLADIMIR E. FEDOROV AND MARKO KOSTI ´C Abstract. The main aim of this paper is to investigate Weyl-p-almost peri- odic properties and asymptotically Weyl-p-almost periodic properties of con- volution products. In such a way, we continue several recent research studies of ours which do concern a similar problematic. 1. Introduction and preliminaries In a series of recent research papers, the second named author has considered the invariance of (asymptotical) Weyl-p-almost periodicity under the action of (finite) infinite convolution product, where 1 ≤ p < ∞. It has been perceived that the case p > 1 is much more delicate for the analysis and, before proceeding any further, we would like to stress that Proposition 2.1 in [10] is not correctly proved in the case that p > 1. In a recent erratum and addendum to the paper [10], we have introduced the class of quasi asymptotically almost periodic functions and considered quasi asymptotical almost periodicity of infinite convolution product (1.1) G(t) ≡ t 7→Z t −∞ R(t − s)g(s) ds, t ∈ R, where p ≥ 1 and g(·) is (equi-)Weyl-p-almost periodic. In this paper, we consider the invariance of (asymptotical) Weyl-p-almost period- icity under the action of (finite) infinite convolution product by assuming that the corresponding resolvent operator family (R(t))t>0 ⊆ L(X, Y ) has a certain growth order at zero and infinity (by (X,k · k), (Y,k · kY ) and L(X, Y ) we denote two non-trivial complex Banach spaces and the space consisting of all linear continu- ous operators from X into Y, respectively; in the sequel, we will use the standard terminology from the monograph [9]). We specifically consider the following two types of growth rates: (1.2) kR(t)kL(X,Y ) ≤ M e−cttβ−1, t > 0 for some finite constants c > 0, β ∈ (0, 1], M > 0, or a substantially weaker one (1.3) tβ−1 1 + tγ , t > 0 for some finite constants γ > 1, β ∈ (0, 1], M > 0. kR(t)kL(X,Y ) ≤ M The estimate (1.2) appears in the theoretical studies of abstract degenerate differ- ential equations of first order with multivalued linear operators A satisfying the 2010 Mathematics Subject Classification. Primary 43A60; Secondary 47D06. Key words and phrases. Weyl-p-almost periodic functions, asymptotically Weyl-p-almost pe- riodic functions, convolution products. 1 2 VLADIMIR E. FEDOROV AND MARKO KOSTI ´C condition (P) clarifed below, while the estimate (1.3) appears in the theoretical studies of abstract degenerate fractional relaxation differential equations with mul- tivalued linear operators A satisfying the same condition. The genesis of paper is stimulated by reading some recent results of F. Bedouhene, N. Challali, O. Mellah, P. Raynaud de Fitte and M. Smaali, which are still unpub- lished and where the cases that X = Y and (R(t))t≥0 is an exponentially decaying strongly continuous non-degenerate C0-semigroup have been analyzed. The main novelty of this paper is the consideration of growth rate (1.3) for solution opera- tor families (R(t))t>0 not necessarily strongly continuous at zero, with regards to the existence and uniqueness of (asymptotically) Weyl-p-almost periodic solutions of abstract fractional differential equations; see e.g. [4]-[7], [12] and [17] for some recent results treating the abstract degenerate fractional differential equations. There is no need to say that fractional calculus and fractional differential equa- tions are rapidly growing fields of research, due to their invaluable importance in modeling real world phenomena appearing in many fields such as astrophysics, electronics, diffusion, chemistry, biology, electricity and thermodynamics. It is almost impossible to summarize here all relevant contributions made recently in these fields; see e.g. [8]-[9], [16] and references cited therein for the basic and not- updated information on the subject. Throughout the paper, we use two different types of fractional derivatives. The Weyl-Liouville fractional derivative Dγ t,+u(t) of order γ ∈ (0, 1) is defined for those continuous functions u : R → X satisfying that t 7→R t −∞ g1−γ(t − s)u(s) ds, t ∈ R is a well-defined continuously differentiable mapping, by g1−γ(t − s)u(s) ds, t ∈ R. −∞ Dγ t,+u(t) := d dtZ t Set D1 t,+u(t) := −(d/dt)u(t). For further information about Weyl-Liouville frac- tional derivatives, we refer the reader to the paper [15] by J. Mu, Y. Zhoa and L. Peng. If α > 0 and m = ⌈α⌉, then the Caputo fractional derivative Dα those functions u ∈ Cm−1([0,∞) : X) satisfying that gm−α ∗ (u −Pm−1 Cm([0,∞) : X), by t u(t) is defined for k=0 ukgk+1) ∈ Dα t u(t) = dm dtm"gm−α ∗ u − ukgk+1!#. m−1 Xk=0 For more details about the abstract fractional differential equations with Caputo derivatives, the reader may consult the monograph [9] and references cited therein. The organization of this paper can be simply described as follows. In Subsection 1.1, we recall the basic definitions and results about generalized (asymptotically) almost periodic functions which will be necessary for our further work. Our main contributions are Theorem 2.1 and Proposition 2.3; besides them, Section 2 contains a great deal of other remarks and observations about problems considered. It is clear that our results are applicable in the analysis of a wide class of abstract inho- mogenous integro-differential equations, which can be degenerate or non-degenerate in time-variable and which may or may not contain fractional derivatives. The main aim of Section 3 is to present certain applications of Theorem 2.1 and Proposition 2.3 in the analysis of existence and uniqueness of Weyl-p-almost periodic solutions A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC... 3 of the abstract fractional differential inclusion (3.3) and asymptotically Weyl-p- almost periodic solutions of the abstract fractional differential inclusion (DFP)f,γ clarifed below. 1.1. Generalized almost periodic functions and asymptotically general- ized almost periodic functions. Unless stated otherwise, we will always assume henceforth that 1 ≤ p < ∞. Let I = [0,∞) or I = R. A function f ∈ Lp loc(I : X) is said to be Stepanov p-bounded iff kfkSp := sup t∈I Z t+1 t kf (s)kp ds!1/p < ∞. The space Lp equipped with the above norm. S(I : X) consisted of all Sp-bounded functions becomes a Banach space Let 1 ≤ p < ∞, let l > 0, and let f, g ∈ Lp 'metric' by loc(I : X). We define the Stepanov Dp x∈I" 1 Sl(cid:2)f (·), g(·)(cid:3) := sup l Z x+l x Then it is well-known that there exists dt#1/p . p (cid:13)(cid:13)f (t) − g(t)(cid:13)(cid:13) Sl(cid:2)f (·), g(·)(cid:3) Dp Dp l→∞ W(cid:2)f (·), g(·)(cid:3) := lim W [f (·), g(·)] appearing above is called the Weyl distance in [0,∞]. The distance Dp of f (·) and g(·). The Stepanov and Weyl 'norm' of f (·) are defined by W(cid:2)f (·), 0(cid:3), Sl(cid:2)f (·), 0(cid:3) and (cid:13)(cid:13)f(cid:13)(cid:13)W p := Dp (cid:13)(cid:13)f(cid:13)(cid:13)Sp respectively. The notions of Stepanov p-boundedness and Weyl p-boundedness are mutually equivalent, i.e., for any function f ∈ Lp loc(I : X), we have := Dp l < ∞ iff l (cid:13)(cid:13)f(cid:13)(cid:13)Sp (cid:13)(cid:13)f(cid:13)(cid:13)W p < ∞. The notion of an (equi-)Weyl-p-almost periodic function is given below (see e.g. [10] and references cited therein). Definition 1.1. Let f ∈ Lp loc(I : X). (i) We say that the function f (·) is equi-Weyl-p-almost periodic, f ∈ e−W p ap(I : X) for short, iff for each ǫ > 0 we can find two real numbers l > 0 and L > 0 such that any interval I ′ ⊆ I of length L contains a point τ ∈ I ′ such that sup x∈I" 1 l Z x+l x (cid:13)(cid:13)f (t + τ ) − f (t)(cid:13)(cid:13) i.e., Dp p ≤ ǫ, dt#1/p Sl(cid:2)f (· + τ ), f (·)(cid:3) ≤ ǫ. (ii) We say that the function f (·) is Weyl-p-almost periodic, f ∈ W p ap(I : X) for short, iff for each ǫ > 0 we can find a real number L > 0 such that any 4 VLADIMIR E. FEDOROV AND MARKO KOSTI ´C interval I ′ ⊆ I of length L contains a point τ ∈ I ′ such that lim l→∞ sup x∈I" 1 l Z x+l x p (cid:13)(cid:13)f (t + τ ) − f (t)(cid:13)(cid:13) i.e., lim l→∞ ≤ ǫ, dt#1/p Sl(cid:2)f (· + τ ), f (·)(cid:3) ≤ ǫ. Dp Denote by AP Sp(I : X) the space consisting of all Stepanov p-almost periodic functions from the interval I into X ([11]). Then it is well known that AP Sp(I : X) ⊆ e − W p ap(I : X) in the set theoretical sense and that any of these two inclusions can be strict. ap(I : X) ⊆ W p Denote by C0([0,∞) : X) the vector space consisting of all bounded continuous functions from [0,∞) into X which vanish at infinity. We say that an Sp-bounded function q : [0,∞) → X is Stepanov p-vanishing iff the function t 7→ q(t + ·), t ≥ 0 belongs to the class C0([0,∞) : Lp([0, 1] : X)). We denote by Sp 0 ([0,∞) : X) the vector space consisting of all Stepanov p-vanishing functions. If q ∈ Lp loc([0,∞) : X), then we define the function q(·,·) : [0,∞) × [0,∞) → X by q(t, s) := q(t + s), t, s ≥ 0. The class of (equi-)Weyl-p-vanishing functions has been recently introduced as fol- lows (see e.g. [10]): Definition 1.2. Let q ∈ Lp loc([0,∞) : X). lim (i) It is said that q(·) is Weyl-p-vanishing iff x≥0" 1 t→∞(cid:13)(cid:13)q(t,·)(cid:13)(cid:13)W p = 0, i.e., (ii) It is said that q ∈ Lp lim t→∞ sup x lim l→∞ l Z x+l ds#1/p loc([0,∞) : X) is equi-Weyl-p-vanishing iff x≥0" 1 (cid:13)(cid:13)q(t + s)(cid:13)(cid:13) ds#1/p = 0. p sup lim t→∞ lim l→∞ l Z x+l (cid:13)(cid:13)q(t + s)(cid:13)(cid:13) We know that C0([0,∞) : X) ⊆ Sp 0 ([0,∞) : X) ⊆ e − W p 0 ([0,∞) : X) and that any of these three inclusions can be strict. 2. Formulation and proof of main results x p W p = 0. 0 ([0,∞) : X) ⊆ Albeit given with a relatively non-complicated proof, the following theorem can be viewed as the main result of this paper: Theorem 2.1. Let 1/p + 1/q = 1 and let (R(t))t>0 ⊆ L(X, Y ) satisfy (1.3). Let a function g : R → X be (equi-)Weyl-p-almost periodic and Weyl p-bounded, and let q(β − 1) > −1 provided that p > 1, resp. β = 1, provided that p = 1. Then the function G : R → Y, defined through (1.1), is bounded continuous and (equi-)Weyl- p-almost periodic. Proof. We will consider the case that g(·) is Weyl-p-almost periodic with p > 1 and explain the main differences in the case that p = 1. Without loss of generality, we may assume that X = Y. Since we have assumed that g(·) is Weyl p-bounded (equivalently, Stepanov p-bounded) and q(β − 1) > −1, we can repeat literally the arguments given in the proof of [10, Proposition 2.1] to deduce that G(·) is A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC... 5 bounded and continuous on the real line (a similar argumentation works in the case that p = 1). Therefore, it remains to be proved that G(·) is Weyl-p-almost periodic. In order to do that, fix a number ǫ > 0. By definition, we can find a real number L > 0 such that any interval I ′ ⊆ I of length L contains a point τ ∈ I ′ such that there exists a number l(ǫ, τ ) > 0 so that (2.1) Dp Sl(cid:2)g(· + τ ), g(·)(cid:3) ≤ ǫ, l ≥ l(ǫ, τ ). On the other hand, it is clear that ku(t + τ ) − u(t)k ≤Z 0 ≤ MZ 0 −∞ kR(−s)kkg(s + t + τ ) − g(s + t)k ds −∞ sβ−1kg(s + t + τ ) − g(s + t)k/(1 + sγ) ds, t ∈ R. (2.2) Since γ > 1 and β ∈ (0, 1], we have the existence of a positive real number ζ > 0 satisfying 1 p < ζ < 1 p + γ − β (in the case that p = 1, we can take any number ζ ∈ (1, γ) and repeat the same procedure). This implies that the function s 7→ sβ−1(1+sζ ) , s < 0 belongs to the space Lq((−∞, 0)) as well as that the function s 7→ 1 1+sζ , s < 0 belongs to the space Lp((−∞, 0)). The integral 1+sγ Z 0 −∞ kg(s + t + τ ) − g(s + t)kp/(1 + sζ)p ds converges for any t ∈ R, which follows from the following computation ∞ Z 0 −∞ kg(s + t + τ ) − g(s + t)kp/(1 + sζ)p ds Xk=0Z −k ≤2p−1kgkp −(k+1) kg(s + t + τ ) − g(s + t)kp/(1 + sζ)p ds 1 + kγ . ≤ Sp 1 ∞ Xk=0 Further on, applying (2.2) and the Holder inequality we get that ku(t + τ ) − u(t)k ≤ M(cid:13)(cid:13)(cid:13)(cid:16) · β−1(1 + · ζ)/(1 + · γ)(cid:17)(cid:13)(cid:13)(cid:13)Lq((−∞,0)) −∞ kg(s + t + τ ) − g(s + t)kp/(1 + sζ )p ds#1/p , t ∈ R. ×"Z 0 6 VLADIMIR E. FEDOROV AND MARKO KOSTI ´C Making use of the Fubini theorem and (2.1), we get that x sup ku(t + τ ) − u(t)kp dt#1/p x∈R" 1 l Z x+l ≤M(cid:13)(cid:13)(cid:13)(cid:16) · β−1(1 + · ζ)/(1 + · γ)(cid:17)(cid:13)(cid:13)(cid:13)Lq((−∞,0)) ×"Z 0 ≤M ǫ(cid:13)(cid:13)(cid:13)(cid:16) · β−1(1 + · ζ)/(1 + · γ)(cid:17)(cid:13)(cid:13)(cid:13)Lq((−∞,0))Z 0 (1 + sζ)p sup l Z x+l x∈R −∞ 1 1 x −∞ kg(s + t + τ ) − g(s + t)kpdt! ds#1/p ds (1 + sζ)p , for any l(ǫ, τ ) > 0. This completes the proof of theorem in a routine manner. Remark 2.2. Let 1 ≤ p < ∞. Then any equi-Weyl-p-almost periodic function is automatically Weyl-p-bounded, which seems to be still unknown for Weyl-p-almost periodic functions. (cid:3) The analysis of asymptotically Stepanov-p-almost periodic and asymptotically (equi-)Weyl-p-almost periodic properties of finite convolution product is not trivial in general case and we need some extra conditions on the ergodic part of function under consideration (denoted henceforth by q(·)) in order to obtain any relevant result in this direction; the situation is, unfortunately, similar if the resolvent family (R(t))t>0 ⊆ L(X, Y ) satisfies the estimate (1.2) or (1.3). In this paper, we will prove the following general proposition with regards to this question: Proposition 2.3. Let q ∈ Lp loc([0,∞) : X), 1/p + 1/q = 1 and let (R(t))t>0 ⊆ L(X, Y ) satisfy (1.3). Let a function g : R → X be (equi-)Weyl-p-almost periodic and Weyl p-bounded, and let q(β − 1) > −1 provided that p > 1, resp. β = 1, provided that p = 1. Suppose that the function belongs to the space FY , which equals to C0([0,∞) : Y ), Sp W p 0 ([0,∞) : Y ) or W p 0 ([0,∞) : Y ), e − R(t − s)q(s) ds, t ≥ 0 0 t 7→ Q(t) ≡Z t 0 ([0,∞) : Y ). Then the function H(t) ≡Z t 0 R(t − s)[g(s) + q(s)] ds, t ≥ 0 ap ap (2.3) is continuous and belongs to the class (e−)W [0,∞),p (Y ) stands for the space of all restictions of Y -valued (equi-)Weyl-p-almost periodic functions from the real line to the interval [0,∞). Proof. Without loss of generality, we may assume that X = Y. Define R(t − s)q(s) ds −Z ∞ (Y )+FY , where (e−)W [0,∞),p R(s)g(t − s) ds, t ≥ 0. F (t) :=Z t The local integrability of convolution products in (2.3) follows from the arguments given in the proofs of [2, Proposition 1.3.4, Proposition 1.3.5]. By Theorem 2.1, the function G : R → X, defined by (1.1), is bounded continuous and (equi-)Weyl-p- almost periodic. Due to the facts that H(t) = G(t) + F (t), t ≥ 0, FX + C0([0,∞) : X) = FX and our assumption that the function Q(·) belongs to the space FX , it 0 t A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC... 7 Z ∞ t t+1 t t t+1 sβ−1 t sβ−1 k=0R t+k+2 R(s)g(t − s) ds, t ≥ 0. sβ−1 1 + sγ kg(tn − s)k ds 0 R(s)q(t− s) ds, t ≥ 0, inclusion q ∈ Lp suffices to show that the mapping Q(t) is continuous for t ≥ 0 as well as that the mapping t 7→R ∞ t R(s)g(t− s) ds, t ≥ 0 is in class C0([0,∞) : X). The continuity of mapping Q(t) for t ≥ 0 can be proved as in the final part of proof of Theorem 2.1, by using the equality Q(t) =R t loc([0,∞) : X) and Holder inequality. To prove that the mapping t 7→R ∞ t R(s)g(t − s) ds, t ≥ 0 is in class C0([0,∞) : X), observe first that R(s)g(t − s) ds =Z ∞ R(s)g(t − s) ds +Z t+1 The continuity of mapping t 7→ R ∞ t+1 R(s)g(t − s) ds = P∞ t+k+1 R(s)g(t − s) ds := P∞ k=0 Fk(t), t ≥ 0 can be shown following the lines of the proof of [13, Proposition 5] since the mapping Fk(·) is continuous by the dominated conver- gence theorem and the series P∞ k=0 Fk(t) converges uniformly in t ≥ 0 due to the Weierstrass criterion. To prove the continuity of mapping R ·+1 · R(s)g(· − s) ds, fix a number t ≥ 0 and a sequence (tn)n∈N in [t, t + 1] converging to t as n → +∞. Then an elementary argumentation involving the Holder inequality shows that: R(s)g(t − s) ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) R(s)g(tn − s) ds −Z t+1 Z tn+1 ≤M"Z tn 1 + sγ kg(t − s)k ds +Z tn+1 1 + sγ(cid:13)(cid:13)g(tn − s) − g(t − s)(cid:13)(cid:13) ds# +Z t+1 kgkLp[0,tn−t] +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[t,tn] ≤M"(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[0,t+2] kg(tn − ·) − g(t − ·)kLp[0,t+2]# kgkLp[0,tn−t] +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[t,t+1] ≤M"(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[0,t+2] kg(tn − ·) − g(t − ·)kLp[0,t+2]#. The right continuity of mappingR ·+1 · R(s)g(·− s) ds at point t follows from the evi- dent equalities limn→+∞ kgkLp[0,tn−t] = 0 and limn→+∞ kg(tn−·)−g(t−·)kLp[0,t+2] = 0, while the left continuity can be proved analogously. The vanishing of function t 7→R ∞ R(s)g(t − s) ds(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=0 t R(s)g(t − s) ds, t ≥ 0 at plus infinity follows from the estimates ·β−1 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[t+1,tn+1] 1 + ·γ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lq[t+1,t+2] ·β−1 kR(·)kLq[t+k,t+k+1], t > 0 kR(·)kLq[t+k,t+k+1] = 0; (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Z ∞ t ∞ lim t→+∞ kgkLp[0,tn−t] ≤ kgkSp ∞ Xk=0 ·β−1 ·β−1 ·β−1 ·β−1 tn tn and kgkLp[0,tn−t] 8 VLADIMIR E. FEDOROV AND MARKO KOSTI ´C see [11, Remark 2.14(ii)] and the proof of [10, Proposition 2.5]. (cid:3) ap ap (Y ). (Y ) is contained in (e−)W p ap([0,∞) : Y ) ⊆ (e−)W [0,∞) Remark 2.4. The space (e−)W [0,∞) ap([0,∞) : Y ). It is not clear whether an (equi-)Weyl-p-almost periodic function defined on [0,∞) can be extended to an (equi-)Weyl-p-almost periodic function defined on R. Therefore, it is not clear whether (e−)W p Remark 2.5. Suppose that q ∈ C0([0,∞) : X) and (R(t))t>0 ⊆ L(X, Y ) satisfies (1.3). The argumentation contained in the proof of [1, Lemma 2.13] combined with the fact that R ∞ 0 sβ−1/(1 + sγ) ds < ∞ shows that limt→+∞ Q(t) = 0, so that Q ∈ C0([0,∞) : Y ). Some conditions on the function q(·) which ensure that the function Q(·) is Stepanov p-vanishing have been analyzed in [11, Remark 2.14]. Remark 2.6. In [14], the second named author has recently introduced various classes of generalized (asymptotically) C(n)-almost periodic functions and inves- tigated the invariance of generalized (asymptotical) C(n)-almost periodicity under the action of convolution products; in particular, the class of (asymptotically) C(n)- Weyl-p-almost periodic functions has been introduced and analyzed. It is worth noting that Theorem 2.1 and Proposition 2.3 can be reconsidered for this class of functions, which can be left to the interested readers as an easy exercise (cf., especially, [14, Proposition 4.2, Proposition 4.5, Proposition 4.7]). Remark 2.7. Suppose that q ∈ Lp loc([0,∞) : X), 1/p + 1/q = 1, the mapping a : (0,∞) → (0,∞) satisfies 0 < a(t) < t, t > 0 as well as that (R(t))t>0 ⊆ L(X, Y ) satisfies (1.3). Let p > 1, let Bp(0) := 0 and Bp(t) := a(t)1/q(cid:0)t − a(t)(cid:1)β−1−γ Z a(t) q Z t +(cid:0)t − a(t)(cid:1)β−1−γ+ 1 0 kq(s)kp ds!1/p a(t) kq(s)kp ds!1/p , t > 0. Since Q(t) =R t a(t) R(s)q(t − s) ds for all t ≥ 0, applying the Holder inequality and (1.3) we may conclude that kQ(t)kY ≤ Bp(t) for all t > 0. In the case that p = 1, set B1(0) := 0 and 0 R(t − s)q(s) ds =R a(t) B1(t) :=(cid:0)t − a(t)(cid:1)β−1−γZ a(t) 0 R(s)q(t − s) ds +R t kq(s)k ds +Z t a(t) kq(s)k ds, t > 0. Then we can similarly prove that kQ(t)kY ≤ B1(t) for all t > 0. This information can be useful to describe the long time behaviour of function Q(·). 0 For example, the function q(t) := χ[n2,n2+1](t), t ≥ 0 ∞ Xn=0 is not Stepanov p-vanishing but it is equi-Weyl-p-vanishing for any finite number p ≥ 1 (using a mollification, we can simply addapt this example to construct an example of an equi-Weyl-p-vanishing function that belongs to the space C∞([0,∞)) 0 kq(s)kp ds ≤ 2 + √t, t ≥ 0 and that is not Stepanov p-vanishing). Furthermore, R t and we can use this estimate as well as the estimate obtained in the first part of A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC... 9 this remark (with a(t) = t/2, t > 0) to see that Q ∈ C0([0,∞) : Y ), provided that the inequality 1 The interested reader may try to construct some examples in which we have that p + β − γ < 0 holds true. the function Q(·) belongs to the class e − W p 0 ([0,∞) : Y ) or W p 0 ([0,∞) : Y ). 3. An application For the beginning, we need to remind ourselves of the notion of a multivalued linear operator (cf. the monographs [3] by R. Cross and [4] by A. Favini, A. Yagi for more details on the subject). A multivalued map (multimap) A : X → P (Y ) is said to be a multivalued linear operator (MLO) iff the following holds: (i) D(A) := {x ∈ X : Ax 6= ∅} is a linear subspace of X; (ii) Ax + Ay ⊆ A(x + y), x, y ∈ D(A) and λAx ⊆ A(λx), λ ∈ C, x ∈ D(A). If X = Y, then we say that A is an MLO in X. If x, y ∈ D(A) and λ, η ∈ C with λ + η 6= 0, then λAx + ηAy = A(λx + ηy). Furthermore, if A is an MLO, then A0 is a linear submanifold of Y and Ax = f +A0 for any x ∈ D(A) and f ∈ Ax. Put R(A) := {Ax : x ∈ D(A)}. Then the set A−10 = {x ∈ D(A) : 0 ∈ Ax} is called the kernel of A and it is denoted by N (A). The inverse A−1 of an MLO is defined by D(A−1) := R(A) and A−1y := {x ∈ D(A) : y ∈ Ax}. Let A be an MLO in X. Then the resolvent set of A, ρ(A) for short, is defined as the union of those complex numbers λ ∈ C for which (i) R(λ − A) = X; (ii) (λ − A)−1 is a single-valued linear continuous operator on X. The operator λ 7→ (λ − A)−1 is called the resolvent of A (λ ∈ ρ(A)); R(λ : A) ≡ (λ − A)−1 (λ ∈ ρ(A)). The basic properties of resolvents of single-valued linear operators continue to hold in the multivalued linear setting ([4]). In the remaining part of paper, it will be supposed that A is a multivalued linear operator on a Banach space X satisfying the following condition (see e.g. [4, p. 47]): (P) There exist finite constants c, M > 0 and β ∈ (0, 1] such that Ψ := Ψc :=nλ ∈ C : ℜλ ≥ −c(cid:0)ℑλ + 1(cid:1)o ⊆ ρ(A) and kR(λ : A)k ≤ M(cid:0)1 + λ(cid:1)−β , λ ∈ Ψ. Suppose that γ ∈ (0, 1) and β > θ. Then the degenerate strongly continuous semigroup (T (t))t>0 ⊆ L(X) generated by A satisfies the estimate (3.1) kT (t)k ≤ M0e−cttβ−1, t > 0 for some finite constant M0 > 0 ([11]). Define Tγ,ν(t)x := tγνZ ∞ 0 sνΦγ(s)T(cid:0)stγ(cid:1)x ds, t > 0, x ∈ X, Sγ(t) := Tγ,0(t), t > 0 ; Sγ(0) := I, Pγ(t) := γTγ,1(t)/tγ, t > 0 ; Rγ(t) := tγ−1Pγ(t), t > 0. Then there exist two finite constants M1 > 0 and M2 > 0 such that (3.2) (cid:13)(cid:13)Rγ(t)(cid:13)(cid:13) ≤ M1tγβ−1, t ∈ (0, 1] and (cid:13)(cid:13)Rγ(t)(cid:13)(cid:13) ≤ M2t−1−γ, t ≥ 1. 10 VLADIMIR E. FEDOROV AND MARKO KOSTI ´C We will use the following notions: A continuous function u : R → X is a mild solution of the abstract first order inclusion u(t) ∈ Au(t) + f (t), t ∈ R Let γ ∈ (0, 1). A continuous function u : R → X is a mild solution of the abstract fractional relaxation inclusion iff (3.3) iff u(t) =Z t −∞ T (t − s)f (s) ds, t ∈ R. Dγ t,+u(t) ∈ −Au(t) + f (t), t ∈ R u(t) =Z t Rγ(t − s)f (s) ds, t ∈ R. −∞ Let Dγ t denote the Caputo fractional derivative of order γ ∈ (0, 1), let x0 ∈ X and let f : [0,∞) → X be asymptotically (equi-)Weyl-p-almost periodic. Suppose, further, that x0 is a point of continuity of (Sγ(t))t>0, i.e., limt→0+ Sγ(t)x0 = x0. By a mild solution of the abstract fractional relaxation inclusion (DFP)f,γ :(cid:26) Dγ t u(t) ∈ Au(t) + f (t), t > 0, u(0) = x0, we mean any function u ∈ C([0,∞) : X) satisfying that Rγ(t − s)f (s) ds, u(t) = Sγ(t)x0 +Z t 0 t ≥ 0. Keeping in mind the estimates (3.1)-(3.2) and the fact that limt→+∞ kSγ(t)k = 0, it is clear how we can apply the main results of Section 3 in the study of ex- istence and uniqueness of (equi-)Weyl-p-almost periodic solutions of the abstract fractional inclusion (3.3) and asymptotically (equi-)Weyl-p-almost periodic solu- tions of the abstract fractional inclusion (DFP)f,γ (solutions of problem (DFP)f,1, t u(t) ≡ u′(t) and the meaning clear, can be also examined). Applications with D1 can be simply incorporated in the study of qualitative properties of solutions of the following fractional Poisson heat equations with the Dirichlet Laplacian ∆: and (cid:26) Dγ   t,+[m(x)v(t, x)] = (∆ − b)v(t, x) + f (t, x), t ∈ R, x ∈ Ω; v(t, x) = 0, (t, x) ∈ [0,∞) × ∂Ω, t [m(x)v(t, x)] = (∆ − b)v(t, x) + f (t, x), t ≥ 0, x ∈ Ω; Dγ v(t, x) = 0, m(x)v(0, x) = u0(x), (t, x) ∈ [0,∞) × ∂Ω, x ∈ Ω, in the space X := Lp(Ω), where Ω is a bounded domain in Rn with smooth bound- ary, b > 0, m(x) ≥ 0 a.e. x ∈ Ω, m ∈ L∞(Ω), γ ∈ (0, 1) and 1 < p < ∞. Further on, let A(x; D) be a second order linear differential operator on Ω with coefficients continuous on Ω; see [4, Example 6.1] for more details. Based on the examina- tion carried out in [13, Example 4], we can apply our main results in the study of existence and uniqueness of asymptotically (equi-)Weyl-p-almost periodic solu- tions of the following fractional damped Poisson-wave type equation in the space A NOTE ON (ASYMPTOTICALLY) WEYL-ALMOST PERIODIC... 11 X := H −1(Ω) or X := Lp(Ω) :   t u(cid:1) +(cid:0)2ωm(x) − ∆(cid:1)Dγ t(cid:0)m(x)Dγ Dγ t ≥ 0, x ∈ Ω ; u = Dγ u(0, x) = u0(x), m(x)(cid:2)Dγ t = 0, t u +(cid:0)A(x; D) − ω∆ + ω2m(x)(cid:1)u(x, t) = f (x, t), (x, t) ∈ ∂Ω × [0,∞), t u(x, 0) + ωu0(cid:3) = m(x)u1(x), x ∈ Ω. References 1. R. Agarwal, B. de Andrade, C. Cuevas, On type of periodicity and ergodicity to a class of fractional order differential equations, Adv. Difference Equ., Vol. 2010, Article ID 179750, 25 pp., DOI:10.1155/2010/179750. 2. W. Arendt, C. J. K. Batty, M. Hieber, F. Neubrander, Vector-valued Laplace Transforms and Cauchy Problems, Birkhauser/Springer Basel AG, Basel, 2001. 3. R. Cross, Multivalued Linear Operators, Marcel Dekker Inc., New York, 1998. 4. A. Favini, A. Yagi, Degenerate Differential Equations in Banach Spaces, Chapman and Hall/CRC Pure and Applied Mathematics, New York, 1998. 5. V. E. Fedorov, L. V. Borel, Solvability of loaded linear evolution equations with a degen- erate operator at the derivative, St. Petersburg Math. J. 26 (2015), 487 -- 497. 6. V. E. Fedorov, D. M. Gordievskikh, M. K. Plekhanova, Equations in Banach spaces with a degenerate operator under a fractional derivative, Differ. Uravn. 51 (2015), 1360 -- 1368. 7. V. E. Fedorov, E. A. Romanova, A. Debbouche, Analytic in a sector resolving families of operators for degenerate evolution fractional equations, J. Math. Sci. 288 (2018), 380 -- 394. 8. A. A. Kilbas, H. M. Srivastava, J. J. Trujillo, Theory and Applications of Fractional Differential Equations, Elsevier Science B.V., Amsterdam, 2006. 9. M. Kosti´c, Abstract Volterra Integro-Differential Equations, Taylor and Francis Group/CRC Press, Boca Raton, Fl., 2015. 10. M. Kosti´c, Weyl-almost periodic and asymptotically Weyl-almost periodic properties of so- lutions to linear and semilinear abstract Volterra integro-differential equations, Math. Notes NEFU 25(2) (2018), 65 -- 84. 11. M. Kosti´c, Existence of generalized almost periodic and asymptotic almost periodic solutions to abstract Volterra integro-differential equations, Electron. J. Differential Equations, vol. 2017, no. 239 (2017), 1 -- 30. 12. M. Kosti´c, Abstract degenerate fractional differential inclusions, Appl. Anal. Discrete Math. 11 (2017), 39 -- 61. 13. M. Kosti´c, Generalized almost automorphic and generalized asymptotically almost auto- morphic solutions of abstract Volterra integro-differential inclusions, Fractional Diff. Calc. 8 (2018), 255 -- 284. 14. M. Kosti´c, On generalized C(n)-almost periodic solutions of abstract Volterra integro- differential equations, Novi Sad J. Math. 48 (2018), 73 -- 91. 15. J. Mu, Y. Zhoa, L. Peng, Periodic solutions and S-asymptotically periodic solutions to fractional evolution equations, Discrete Dyn. Nat. Soc., Volume 2017, Article ID 1364532, 12 pages, https://doi.org/10.1155/2017/1364532. 16. S. G. Samko, A. A. Kilbas, O. I. Marichev, Fractional Derivatives and Integrals: Theory and Applications, Gordon and Breach, New York, 1993. 17. G. A. Sviridyuk, V. E. Fedorov, Linear Sobolev Type Equations and Degenerate Semigroups of Operators, Inverse and Ill-Posed Problems (Book 42). VSP, Utrecht, Boston, 2003. Chelyabinsk State State University, Chelyabinsk, 454 080, Russia E-mail address: [email protected] Faculty of Technical Sciences, University of Novi Sad, Trg D. Obradovi´ca 6, 21125 Novi Sad, Serbia E-mail address: [email protected]
1104.2418	1	1104	2011-04-13T08:47:44	Operator approach to Vlasov scaling for some models of spatial ecology	[ "math.FA", "math-ph", "math.DS", "math-ph" ]	We consider Vlasov-type scaling for Markov evolution of birth-and-death type in continuum, which is based on a proper scaling of corresponding Markov generators and has an algorithmic realization in terms of related hierarchical chains of correlation functions equations. The existence of rescaled and limiting evolutions of correlation functions as well as convergence to the limiting evolution are shown. The obtained results enable to derive a non-linear Vlasov-type equation for the density of the limiting system.	math.FA	math	Operator approach to Vlasov scaling for some models of spatial ecology Dmitri Finkelshtein∗ Yuri Kondratiev† Oleksandr Kutoviy‡ Key words. Continuous systems, Spatial birth-and-death processes, Individ- ual based models, Vlasov scaling, Vlasov equation, Correlation functions MSC (2010): 47D06, 60J25, 60J35, 60J80, 60K35 Abstract We consider Vlasov-type scaling for Markov evolution of birth- and-death type in continuum, which is based on a proper scaling of corresponding Markov generators and has an algorithmic realization in terms of related hierarchical chains of correlation functions equations. The existence of rescaled and limiting evolutions of correlation func- tions as well as convergence to the limiting evolution are shown. The obtained results enable to derive a non-linear Vlasov-type equation for the density of the limiting system. ∗Institute of Mathematics, National Academy of Sciences of Ukraine, Kyiv, Ukraine ([email protected]). †Fakultat fur Mathematik, Universitat Bielefeld, 33615 Bielefeld, Germany ([email protected]) ‡Fakultat fur Mathematik, Universitat Bielefeld, 33615 Bielefeld, Germany ([email protected]). 1 Operator approach to Vlasov scaling for models of spatial ecology 2 1 Introduction The Vlasov equation is a famous example of a kinetic equation which de- scribes the dynamical behavior of a many-body system. In physics, it char- acterizes the Hamiltonian motion of an infinite particle system influenced by weak long-range forces in the mean field scaling limit. The detailed exposition of the Vlasov scaling for the Hamiltonian dynamics was given by W.Braun and K.Hepp [3] and later by R.L.Dobrushin [5] for more general deterministic dynamical systems. The limiting Vlasov-type equations for particle densities in both papers are considered in classes of integrable functions (or finite mea- sures in the weak form). This corresponds, actually, to the situation of finite volume systems or systems with zero mean density in an infinite volume. The Vlasov equation for the integrable functions was investigated in details by V.V.Kozlov [16]. An excellent review about kinetic equations which describe dynamical multi-body systems was given by H.Spohn [22], [23]. Note that in the framework of interacting diffusions a similar problem is known as the McKean -- Vlasov limit. Motivated by the study of Vlasov scaling for some classes of stochastic evolutions in continuum for which the use of the mentioned above approaches breaks down (even in the finite volumes) we developed the general approach to study the Vlasov-type dynamics (see [8]). It is based on a proper scaling of the hierarchical equations for the evolution of correlation functions and can be interpreted in the terms of the rescaled Markov generators. Up to our knowledge, at the present time only this technique may give a possibility to control the convergence in the Vlasov limit in the case of non-integrable densities which is generic for infinite volume infinite particle systems. Say- ing about the evolutions, the kinetic equations of which can not be studied by the classical techniques described in [3] and [5], we have in mind, first of all, spatial birth-and-death Markov processes (e.g., continuous Glauber dynamics, spatial ecological models) and hopping particles Markov evolu- tions (e.g., Kawasaki dynamics in continuum). The main difficulty to carry out the approach proposed by W.Braun, K.Hepp [3] and R.L.Dobrushin [5] for such models is absence of the proper descriptions in terms of stochastic evolutional equations. Another problem concerns the possible variation of particles number in the evolution. The important point to note also is that an application of the technique proposed in [8] leads to a limiting hierarchy which posses a chaos preservation property. Operator approach to Vlasov scaling for models of spatial ecology 3 The aim of this paper is to study the Vlasov scaling for the individual based model (IBM) in spatial ecology introduced by B.Bolker and S.Pacala [1, 2], U.Dieckmann and R.Law [4] (BDLP model) using the scheme developed in [8]. A population in this model is represented by a configuration of motionless organisms (plants) located in an infinite habitat (an Euclidean space in our considerations). The unbounded habitat is taken to avoid boundary effects in the population evolution. The evolution equation for the correlation functions of the BDLP model was studied in details in [9]. In [1, 2], [4] this system was called the system of spatial moment equations for plant competition and, actually, this sys- tem itself was taking as a definition of the dynamics in the BDLP model. The mathematical structure of the correlation functions evolution equation is close to other well-known hierarchical systems in mathematical physics, e.g., BBGKY hierarchy for the Hamiltonian dynamics (see, e.g. [6]). As in all hierarchical chains of equations, we can not expect the explicit form of the solution, and even more, the existence problem for these equations is a highly delicate question. According to the general scheme (see [8]), we state conditions on struc- tural coefficients of the BDLP Markov generator, which give a weak con- vergence of the rescaled generator to the limiting generator of the related Vlasov hierarchy. Next, we may compute limiting Vlasov type equation for the BDLP model leaving the question about the strong convergence of the hierarchy solutions for a separate analysis. A control of the strong conver- gence of the rescaled hierarchy is, in general, a difficult technical problem. In particular, this problem remains be open for BBGKY hierarchy for the case of Hamiltonian dynamics as well as for Bogoliubov -- Streltsova hierarchy corresponding to the gradient diffusion model. In the present paper we show the existence of the rescaled and limiting evolutions of correlation functions related to the Vlasov scaling of the BDLP model and the convergence to the limiting evolution. With this evolution for special class of initial conditions is related a non-linear equation for the density, which is called Vlasov equation for the considered stochastic dynamics. Let us mention that a version of the BDLP model for the case of finite populations was studied in the paper [11]. In this work the authors developed a probabilistic representation for the finite BDLP process and applied this technique to analyze a mean-field limit in the spirit of classical Dobrushin or McKean -- Vlasov schemes. They obtained an integro-differential equation for the limiting deterministic process corresponding to an integrable initial Operator approach to Vlasov scaling for models of spatial ecology 4 condition. The latter equation coincides with the Vlasov equation for the BDLP model derived below in our approach. The present paper is organized in the following way. Section 2 is devoted to the general settings required for the description of the model which we study. In Subsection 3.1 we discuss the general Vlasov scaling approach for spatial continuos models. Subsection 3.2 is devoted to the abstract conver- gence result for semigroups in Banach spaces which will be crucial to prove the main statements of the paper presented in Subsection 3.3. The corre- sponding proofs are given in Subsection 3.4. 2 Basic fact and description of model 2.1 General facts and notations Let B(Rd) be the family of all Borel sets in Rd and Bb(Rd) denotes the system of all bounded sets in B(Rd). The space of n-point configuration is Γ(n) 0 = Γ(n) 0,Rd :=(cid:8)η ⊂ Rd(cid:12)(cid:12) η = n(cid:9) , n ∈ N0 := N ∪ {0}, where A denotes the cardinality of the set A. The space Γ(n) Λ for Λ ∈ Bb(Rd) is defined analogously to the space Γ(n) equivalent to the symmetrization of 0 . As a set, Γ(n) 0 := Γ(n) 0,Λ is ^(Rd)n =(cid:8)(x1, . . . , xn) ∈ (Rd)n(cid:12)(cid:12) xk 6= xl if k 6= l(cid:9) , i.e. ^(Rd)n/Sn, where Sn is the permutation group over {1, . . . , n}. Hence one can introduce the corresponding topology and Borel σ-algebra, which we denote by O(Γ(n) 0 ), respectively. Also one can define a mea- sure m(n) as an image of the product of Lebesgue measures dm(x) = dx on 0 ) and B(Γ(n) (cid:0)Rd, B(Rd)(cid:1). The space of finite configurations Γ(n) 0 Γ0 := Gn∈N0 is equipped with the topology which has structure of disjoint union. There- fore, one can define the corresponding Borel σ-algebra B(Γ0). Operator approach to Vlasov scaling for models of spatial ecology 5 A set B ∈ B(Γ0) is called bounded if there exists Λ ∈ Bb(Rd) and N ∈ N Λ . The Lebesgue -- Poisson measure λz on Γ0 is defined n=0 Γ(n) such that B ⊂FN as λz := zn n! m(n). ∞ Xn=0 Here z > 0 is the so called activity parameter. The restriction of λz to ΓΛ will be also denoted by λz. The configuration space Γ :=(cid:8)γ ⊂ Rd (cid:12)(cid:12) γ ∩ Λ < ∞, for all Λ ∈ Bb(Rd)(cid:9) is equipped with the vague topology. It is a Polish space (see e.g. [14]). The corresponding Borel σ-algebra B(Γ) is defined as the smallest σ-algebra for which all mappings NΛ : Γ → N0, NΛ(γ) := γ ∩ Λ are measurable, i.e., One can also show that Γ is the projective limit of the spaces {ΓΛ}Λ∈Bb(Rd) w.r.t. the projections pΛ : Γ → ΓΛ, pΛ(γ) := γΛ, Λ ∈ Bb(Rd). B(Γ) = σ(cid:0)NΛ(cid:12)(cid:12)Λ ∈ Bb(Rd)(cid:1) . the family of measures {πΛ by πΛ The Poisson measure πz on (Γ, B(Γ)) is given as the projective limit of z is the measure on ΓΛ defined z }Λ∈Bb(Rd), where πΛ z := e−zm(Λ)λz. We will use the following classes of functions: L0 ls(Γ0) is the set of all measurable functions on Γ0 which have a local support, i.e. G ∈ L0 ls(Γ0) if there exists Λ ∈ Bb(Rd) such that G ↾Γ0\ΓΛ= 0; Bbs(Γ0) is the set of bounded measurable functions with bounded support, i.e. G ↾Γ0\B= 0 for some bounded B ∈ B(Γ0). On Γ we consider the set of cylinder functions Fcyl(Γ), i.e. the set of all measurable functions G on (cid:0)Γ, B(Γ))(cid:1) which are measurable w.r.t. BΛ(Γ) for some Λ ∈ Bb(Rd). These functions are characterized by the following relation: F (γ) = F ↾ΓΛ (γΛ). The following mapping between functions on Γ0, e.g. L0 ls(Γ0), and func- tions on Γ, e.g. Fcyl(Γ), plays the key role in our further considerations: KG(γ) :=Xη⋐γ G(η), γ ∈ Γ, (2.1) where G ∈ L0 [13, 17, 18]. The summation in the latter expression is taken over all finite subconfigurations of γ, which is denoted ls(Γ0), see e.g. Operator approach to Vlasov scaling for models of spatial ecology 6 by the symbol η ⋐ γ. The mapping K is linear, positivity preserving, and invertible, with K −1F (η) :=Xξ⊂η (−1)η\ξF (ξ), η ∈ Γ0. (2.2) Let M1 fm(Γ) be the set of all probability measures µ on (cid:0)Γ, B(Γ)(cid:1) which have finite local moments of all orders, i.e. RΓ γΛnµ(dγ) < +∞ for all Λ ∈ Bb(Rd) and n ∈ N0. A measure ρ on (cid:0)Γ0, B(Γ0)(cid:1) is called locally finite iff ρ(A) < ∞ for all bounded sets A from B(Γ0). The set of such measures is denoted by Mlf(Γ0). One can define a transform K ∗ : M1 fm(Γ) → Mlf(Γ0), which is dual to the K-transform, i.e., for every µ ∈ M1 fm(Γ), G ∈ Bbs(Γ0) we have ZΓ KG(γ)µ(dγ) =ZΓ0 G(η) (K ∗µ)(dη). The measure ρµ := K ∗µ is called the correlation measure of µ. As shown in [13] for µ ∈ M1 fm(Γ) and any G ∈ L1(Γ0, ρµ) the series (2.1) is µ-a.s. absolutely convergent. Furthermore, KG ∈ L1(Γ, µ) and ZΓ0 G(η) ρµ(dη) =ZΓ (KG)(γ) µ(dγ). (2.3) A measure µ ∈ M1 µΛ := µ ◦ p−1 In this case ρµ := K ∗µ is absolutely continuous w.r.t λz. We denote Λ is absolutely continuous with respect to πΛ fm(Γ) is called locally absolutely continuous w.r.t. πz iff z for all Λ ∈ BΛ(Rd). kµ(η) := dρµ dλz (η), η ∈ Γ0. The functions k(n) µ : (Rd)n −→ R+ (2.4) k(n) µ (x1, . . . , xn) :=( kµ({x1, . . . , xn}), 0, if (x1, . . . , xn) ∈ ^(Rd)n otherwise are the correlation functions well known in statistical physics, see e.g. [21]. [20], Operator approach to Vlasov scaling for models of spatial ecology 7 2.2 Description of model We consider the evolving in time system of interacting individuals (particles) in the space Rd. The state of the system at the fixed moment of time t > 0 is described by the random configuration γt from Γ. Heuristically, the mechanism of the evolution is given by a Markov generator, which has the following form L := L− + L+, where  (L−F )(γ) := (L−(m, κ−, a−)F )(γ) :=Xx∈γ m + κ− Xy∈γ\x (L+F )(γ) := (L+(κ+, a−)F )(γ) := κ+ZRdXy∈γ a−(x − y)  D− a+(x − y)D+ x F (γ)dx. x F (γ), (2.5) Here 0 ≤ a−, a+ ∈ L1(Rd) are arbitrary, even functions such that ZRd a−(x)dx =ZRd a+(x)dx = 1 (in other words, a−, a+ are probability densities) and m, κ−, κ+ > 0 are some positive constants. The pre-generator L describes the Bolker -- Dieckmann -- Law -- Pacala BDLP model, which was introduced in [1, 2, 4]. During the corresponding stochastic evolution the birth of individuals occurs independently and the death is ruled not only by the global regulation (mortality) but also by the local regulation with the kernel κ−a−. This regulation may be described as a competition (e.g., for resources) between individuals in the population. The evolution of the one dimensional distribution for such systems can be expressed in terms of their characteristics, e.g. the correlation functions (see (2.4)). The dynamics of correlation functions for the BDLP model was studied in [9]. The main result of this paper informally says the following: If the mortality m and the competition kernel κ−a− are large enough, then the dynamics of correlation functions, associated with the pre-generator (2.5), exists and preserves (sub-)Poissonian bound. For the readers convenience we repeat below the relevant material from [9] without proofs. Operator approach to Vlasov scaling for models of spatial ecology 8 Let L± := K −1L±K be the K-image of L±, which can be initially defined on functions from Bbs(Γ0). For arbitrary and fixed C > 0 we consider the operator L := L+ + L− in the functional space LC = L1(cid:0)Γ0, C ηdλ (η)(cid:1) . Below, symbol k·kC stands for the norm of this space. For any ω > 0 we define H(ω) to be the set of all densely defined closed operators T on LC, the resolvent set ρ(T ) of which contains the sector Sect(cid:16)π 2 and for any ε > 0 + ω(cid:17) :=nζ ∈ C (cid:12)(cid:12)(cid:12) arg ζ < π 2 + ωo , (T − ζ11)−1 ≤ Mε ζ , arg ζ ≤ π 2 + ω − ε, where Mε does not depend on ζ. The first non-trivial result, which is based on the perturbation theory, says that the operator L with the domain D( L) :=nG ∈ LC (cid:12)(cid:12)(cid:12) · G(·) ∈ LC, Ea− (·)G(·) ∈ LCo is a generator of a holomorphic C0-semigroup Ut on LC. To construct the corresponding evolution of correlation functions we note that the dual space (LC)′ = (cid:0)L1(Γ0, dλC)(cid:1)′ = L∞(Γ0, dλC), where dλC := C ·dλ. The space (LC)′ is isometrically isomorphic to the Banach space with the norm KC :=nk : Γ0 → R (cid:12)(cid:12)(cid:12) k(·)C −· ∈ L∞(Γ0, λ)o kkkKC := kC −·k(·)kL∞(Γ0,λ), where the isomorphism is provided by the isometry RC (LC)′ ∋ k 7−→ RCk := k(·)C · ∈ KC. (2.6) In fact, we have duality between Banach spaces LC and KC given by the following expression hhG, kii :=ZΓ0 G · k dλ, G ∈ LC, k ∈ KC (2.7) Operator approach to Vlasov scaling for models of spatial ecology 9 with hhG, kii ≤ kGkC · kkkKC . It is clear that for any k ∈ KC k(η) ≤ kkkKC C η for λ-a.a. η ∈ Γ0. (2.8) (2.9) Let L′ be the adjoint operator to L in (LC)′ with domain D( L′). Its image in KC under the isometry RC we denote by L∗ = RC L′RC−1. It is evident that the domain of L∗ will be D( L∗) = RCD( L′), correspondingly. Then, for any G ∈ LC, k ∈ D( L∗) ZΓ0 G · L∗kdλ =ZΓ0 =ZΓ0 G · RC L′RC−1kdλ =ZΓ0 LG · RC−1kdλC =ZΓ0 LG · kdλ, G · L′RC−1kdλC therefore, L∗ is the dual operator to L w.r.t. the duality (2.7). By [10], we have the precise form of L∗: ( L∗k)(η) = −(cid:16)mη + κ−Ea− (η)(cid:17) k(η) a+(x − y)k(η \ x) (2.10) + κ+Xx∈η Xy∈η\x + κ+ZRdXy∈η + κ−ZRdXy∈η a+(x − y)k((η \ y) ∪ x)dx a−(x − y)k(η ∪ x)dx. Now we consider the adjoint semigroup T ′(t) on (LC)′ and its image T ∗(t) in KC. The latter one describes the evolution of correlation functions. Trans- ferring the general results about adjoint semigroups (see, e.g., [7]) onto semigroup T ∗(t) we deduce that it will be weak-continuous and weak- differentiable at 0. Moreover, L∗ will be the weak-generator of T ∗(t). Here and subsequently we mean "weak-properties" w.r.t. the duality (2.7). Operator approach to Vlasov scaling for models of spatial ecology 10 3 Vlasov scaling 3.1 Description of Vlasov scaling We begin with a general idea of the Vlasov-type scaling. It is of interest to construct some scaling Lε, ε > 0 of the generator L, such that the following scheme works. Suppose that we know the proper scaling of L and we are able to prove the existence of the semigroup Tε(t) with the generator Lε := K −1LεK in the space LC for some C > 0. Let us consider the Cauchy problem correspond- ing to the adjoint operator L∗ and take an initial function with the strong singularity in ε. Namely, k(ε) 0 (η) ∼ ε−ηr0(η), ε → 0, η ∈ Γ0, where the function r0 is independent of ε. The solution to this problem is described by the dual semigroup T ∗ ε (t). The scaling L 7→ Lε has to be chosen in such a way that T ∗ ε (t) preserves the order of the singularity: ( T ∗ ε (t)k(ε) 0 )(η) ∼ ε−ηrt(η), ε → 0, η ∈ Γ0. Another very important requirement for the proper scaling concerns the dy- namics r0 7→ rt. It should preserve Lebesgue -- Poisson exponents: if r0(η) = eλ(ρ0, η) then rt(η) = eλ(ρt, η) and there exists explicit (nonlinear, in general) differential equation for ρt ∂ ∂t ρt(x) = υ(ρt(x)), (3.1) which will be called the Vlasov-type equation. Now let us explain the main technical steps to realize Vlasov-type scaling. Let us consider for any ε > 0 the following mapping (cf. (2.6)) on functions on Γ0 (Rεr)(η) := εηr(η). (3.2) This mapping is "self-dual" w.r.t. the duality (2.7), moreover, R−1 Then we have k(ε) 0 ∼ Rε T ∗ Therefore, we have to show that for any t ≥ 0 the operator family Rε T ∗ ε > 0 has limiting (in a proper sense) operator U(t) and 0 ∼ Rε−1r0, and we need rt ∼ Rε T ∗ ε (t)k(ε) ε = Rε−1. ε (t)Rε−1r0. ε (t)Rε−1, U(t)eλ(ρ0) = eλ(ρt). (3.3) Operator approach to Vlasov scaling for models of spatial ecology 11 But, informally, T ∗ us consider the "renormalized" operator ε (t) = exp {t L∗ ε} and Rε T ∗ ε (t)Rε−1 = exp {tRε L∗ εRε−1}. Let L∗ ε, ren := Rε L∗ εRε−1. (3.4) In fact, we need that there exists an operator V ∗ (called Vlasov operator) such εRε−1} → exp {t V ∗} =: U(t) for which (3.3) holds. Hence, that exp {tRε L∗ heuristic way to produce the scaling L 7→ Lε is to demand that lim ε→0(cid:18) ∂ ∂t eλ(ρt, η) − L∗ ε, reneλ(ρt, η)(cid:19) = 0, η ∈ Γ0, if ρt satisfies (3.1). The point-wise limit of L∗ ε, ren will be natural candidate for V ∗. Having chosen the proper scaling we proceed to the following technical steps which give the rigorous meaning to the idea introduced above. Note that definition (3.4) implies Lε, ren = Rε−1 LεRε. We prove that "renormal- ized" operator Lε, ren is a generator of a contraction semigroup Tε, ren(t) on LC. Next we show that this semigroup converges strongly to some semigroup TV (t) with the generator V . This limiting semigroup leads us directly to the solution for the Vlasov-type equation. Below we show how to realize this scheme in details. 3.2 Approximation in Banach space In this subsection we study general question about the strong convergence of semigroups in Banach spaces. The obtained results will be crucial in the realization of the Vlasov-type scaling for the BDLP model. Let {U ε t , t ≥ 0} , ε ≥ 0 be a family of semigroups on a Banach space E. We set (Lε, D(Lε)) to be the generator of {U ε t , t ≥ 0} for each ε ≥ 0. Our purpose now is to describe the strong convergence of semigroups {U ε t , t ≥ 0} , ε ≥ 0 in terms of the corresponding generators as ε tends to 0. According to the classical result (see e.g. [12]), it is enough to show that there exists β > 0 and λ : Re λ > β such that (Lε − λ11)−1 s−→ (L0 − λ11)−1 , ε → 0, (3.5) where 11 is the identical operator. In this subsection we show how to prove (3.5) under the following assumptions on the family (Lε, D(Lε)), ε ≥ 0: Operator approach to Vlasov scaling for models of spatial ecology 12 Assumptions (A): 1. For any ε ≥ 0, the operator (Lε, D(Lε)) admits representation Lε = A1(ε) + A2(ε), where A1(ε) is a closed operator and D(A1(ε)) = D(A2(ε)) := D(Lε). 2. There exists β > 0 and λ: Re λ > β such that (a) λ belongs to the resolvent set of A1(ε) for any ε ≥ 0 and (A1(ε) − λ11)−1 s−→ (A1(0) − λ11)−1 , ε → 0, (b) sup ε>0(cid:13)(cid:13)(A1(ε) − λ11)−1(cid:13)(cid:13) ≤(cid:13)(cid:13)(A1(0) − λ11)−1(cid:13)(cid:13) , (c) for any ε ≥ 0 (cid:13)(cid:13)A2(ε) (A1(ε) − λ11)−1(cid:13)(cid:13) < 1, converges strongly to the operator as ε → 0. (d) (cid:0)A2(ε) (A1(ε) − λ11)−1 + 11(cid:1)−1 (cid:0)A2(0) (A1(0) − λ11)−1 + 11(cid:1)−1 The strong convergence result for the family {U ε t , t ≥ 0} , ε ≥ 0 is estab- lished by our next theorem Theorem 3.1. Let (Lε, D(Lε)), ε ≥ 0 be the family of generators corre- sponding to the family of C0-semigroups {U ε t con- verges strongly to U 0 t as ε → 0 uniformly on each finite interval of time, provided assumptions (A) are satisfied. t , t ≥ 0} , ε ≥ 0. Then, U ε Proof. The proof is completed by showing (3.5). For any ε ≥ 0 and λ from the resolvent set of A1(ε) we have Ran(cid:0)(A1(ε) − λ11)−1(cid:1) = D (A1(ε)) = D (A2(ε)) = D(Lε). Hence, Lε − λ11 = A1(ε) + A2(ε) − λ11 =(cid:0)A2(ε) (A1(ε) − λ11)−1 + 11(cid:1) (A1(ε) − λ11) . (3.6) Operator approach to Vlasov scaling for models of spatial ecology 13 Combining (3.6) with the assumption 2(c) of (A) we get the following rep- resentations for the resolvent (Lε − λ11)−1 = (A1(ε) + A2(ε) − λ11)−1 = (A1(ε) − λ11)−1(cid:0)A2(ε) (A1(ε) − λ11)−1 + 11(cid:1)−1 . (3.7) From this formula, triangle inequality and assumptions 2(a), 2(b) and 2(d) of (A) we conclude the assertion of the theorem. 3.3 Main results We check at once that the proper scaling for the BDLP pre-generator is the following one (LεF )(γ) :=Xx∈γ  m + εκ− Xy∈γ\x + κ+ZRdXy∈γ a−(x − y)  D− x F (γ) a+(x − y)D+ x F (γ)dx, (3.8) ε > 0. Next we consider the formal K-image of Lε and the corresponding renor- malized operator on Bbs(Γ0): LεG := K −1LεKG; Lε, renG := Rε−1 LεRεG. In the proposition below we calculate the precise form of the operator Lε, ren for the BDLP model. Proposition 3.2. For any ε > 0 and any G ∈ Bbs (Γ0) Lε,renG =A1G + A2G + ε (B1G + B2G) , Operator approach to Vlasov scaling for models of spatial ecology 14 where (A1G) (η) = − m η G (η) , (A2G) (η) = − κ−Xx∈η Xy∈η\x + κ+Xy∈η ZRd (B2G) (η) = κ+Xy∈ηZRd (B1G) (η) = − κ−Ea− a− (x − y) G (η \ x) a+ (x − y) G (η \ y ∪ x) dx, (η) G (η) , a+ (x − y) G (η ∪ x) dx. Proof. According to the definition, we have Lε,ren = Rε−1 LεRε, where Lε = L−(cid:0)m, εκ−a−(cid:1) + ε−1 L+(cid:0)εκ+a+(cid:1) . As a result, ( LεG) (η) = (A1G) (η) + ε(B1G) (η) + (B2G) (η) a− (x − y) G (η \ x) − εκ−Xx∈η Xy∈η\x + κ+Xy∈η ZRd a+ (x − y) G (η \ y ∪ x) dx. and hence ( Lε,renG) (η) = (A1G) (η) + (A2G) (η) + ε((B1 + B2) G) (η) , which completes the proof. Remark 3.3. It is easily seen that the operator V := A1 + A2 will be the point-wise limit of Lε, ren as ε tends to 0. Therefore, the adjoint operator to V w.r.t. to the duality (2.7) (if it exists) can be considered as a candidate for the Vlasov operator in our model. Below we give the rigorous meaning to the operator Lε,ren. Let us define the set D1 :=nG ∈ LC Ea− (·) G (·) ∈ LC, · G (·) ∈ LCo Operator approach to Vlasov scaling for models of spatial ecology 15 Proposition 3.4. For any ε, m, κ−, C > 0 the operator with the domain D1 is a generator of a contraction C0-semigroup on LC. A1(ε) := A1 + εB1 (3.9) Moreover, A1(ε) ∈ H (ω) for all ω ∈(cid:0)0; π 2(cid:1). Proof. See the proof of Proposition 4.2 in [9]. Remark 3.5. It is a simple matter to check that Proposition 3.4 holds also in the case ε = 0, provided the domain of the operator A1(0) := A1 is changed to D0 := {G ∈ LC · G ∈ LC} ⊃ D1. The next task is to show that for any ε > 0 the operator A2(ε) := Lε,ren − A1(ε) = A2 + εB2 (3.10) with the domain D1 as well as the operator A2(0) := A2 with the domain D0 are relatively bounded w.r.t. the operator (A1(ε), D1) and (A1, D0), correspondingly. This is demonstrated in Propositions 3.6 and 3.7, which can be proved similarly to Lemmas 4.4 and 4.5 in [9]. Proposition 3.6. For any δ > 0 and any κ−, κ+, m, C > 0 such that κ−C m + κ+ m ≤ δ the following estimate holds kA2GkC ≤ δ kA1GkC , G ∈ D0. Moreover, for all ε > 0 kA2GkC ≤ δ kA1(ε)GkC , G ∈ D1. Now, the operator (A2, D0) is well-defined on LC. Proposition 3.7. For any ε, δ > 0 and any κ−, κ+, m, C > 0 such that εκ+Ea+ the following estimate holds (η) < δC(cid:16)εκ−Ea− (η) + m η(cid:17) , η 6= ∅ kεB2GkC ≤ a kA1(ε)GkC , G ∈ D1 with a < δ. Operator approach to Vlasov scaling for models of spatial ecology 16 Remark 3.8. Proposition 3.7 enables us to take D(B2) = D1. As a result, Remark 3.5 shows that the domain of the operator A2(ε) will be D0 ∩ D1 = D1. We are now in a position to show that the operator ( Lε,ren, D1) generates semigroup on LC. To this end we use the classical result about the pertur- bation of holomorphic semigroups (see, e.g. [12]). For the convenience of the reader we formulate below the main statement without proof: For any T ∈ H(ω), ω ∈ (0; π 2 ) and for any ǫ > 0 there exist positive constants α, δ such that if the operator A satisfies Au ≤ aT u + bu, u ∈ D(T ) ⊂ D(A), with a < δ, b < δ, then T + A is a generator of a holomorphic semigroup. In particular, if b = 0, then T + A ∈ H(ω − ǫ). Theorem 3.9. Let the functions a−, a+ and the constants m, κ−, κ+, C > 0 satisfy Cκ−a− (x) ≥ 4κ+a+ (x) , x ∈ Rd. m > 4(cid:0)κ−C + κ+(cid:1) , (3.11) (3.12) Then, for any ε > 0 the operator ( Lε,ren, D1) is a generator of a holomorphic semigroup Ut,ε, t ≥ 0 on LC. Proof. Let ε > 0 be arbitrary and fixed. By definition, Lε,ren = A1(ε) + A2(ε). The direct application of the theorem about perturbation of holomorphic semigroups (see the formulation above the assertion of Theorem 3.9) to T = A1(ε) and A = A2(ε) gives now the desired claim. It is important to note that Proposition 3.4 enables us to consider δ equal to 1 2 in the formulation of the classical theorem introduced above. The appearance of the multiplicand 4 on the left-hand side of the both assumptions in assertion of Theorem 3.9 is motivated exactly by the latter fact. Theorem 3.10. Assume that the constants m, κ−, κ+, C > 0 satisfy Then, the operator V = A1 + A2 with the domain D0 is a generator of a holomorphic semigroup U V t , t ≥ 0 on LC. m > 2(cid:0)κ−C + κ+(cid:1) . Operator approach to Vlasov scaling for models of spatial ecology 17 Proof. We use the same classical result as for Theorem 3.9 in the case: A1 is a generator of holomorphic semigroup, A2 is relatively bounded w.r.t. A1 with the boundary less then 1 2. Now we may repeat the same considerations as at the end of Section 2. Namely, transferring the general results about adjoint semigroups (see, e.g., [7]) onto semigroup ( U V t )∗ in KC we deduce that it will be weak-continuous and weak-differentiable at 0. Moreover, V ∗ will be the weak-generator of T ∗(t). This means, in particular, that for any G ∈ D( V ) ⊂ LC, k ∈ D( V ∗) ⊂ KC d dtDDG, ( U V t )∗kEE =DDG, V ∗( U V t )∗kEE . (3.13) The explicit form of V ∗ follows from (2.10), namely, for any k ∈ D( V ∗) V ∗k(η) = −mηk(η) − κ−ZRdXx∈η + κ+Xx∈ηZRd a−(x − y)k(η ∪ y)dy a+(x − y)k(η \ x ∪ y)dy. (3.14) As a result, we have that for any k0 ∈ D( V ∗) the function kt = ( U V provides a weak solution of the following Cauchy problem t )∗k0 ∂ ∂t kt = V ∗kt (3.15)   kt(cid:12)(cid:12)t=0 = k0. In the next theorem we show that the limiting Vlasov dynamics has chaos preservation property, i.e. preserves the Lebesgue -- Poisson exponents. Theorem 3.11. Let conditions of Theorem 3.9 be satisfied and, additionally, 16e−1. Let ρ0 ≥ 0 be a measurable nonnegative function on Rd such that C ≥ 4 ess supx∈Rd ρ0(x) ≤ C. Then the Cauchy problem (3.15) with k0 = eλ(ρ0) has a weak* solution kt = eλ(ρt) ∈ KC, where ρt is a unique nonnegative solution to the Cauchy problem ρt(x) = κ+(a+ ∗ ρt)(x) − κ−ρt(x)(a− ∗ ρt)(x) − mρt(x), ∂ ∂t  ρt(cid:12)(cid:12)t=0(x) = ρ0(x),  and ess supx∈Rd ρt(x) ≤ C, t ≥ 0. (3.16) Operator approach to Vlasov scaling for models of spatial ecology 18 Proof. First of all, if (3.16) has a solution ρt(x) ≥ 0 then ∂ ∂t ρt(x) ≤ κ+(a+ ∗ ρt)(x) − mρt(x) and, therefore, ρt(x) ≤ rt(x) where rt(x) is a solution of the Cauchy problem rt(x) = κ+(a+ ∗ rt)(x) − mrt(x), ∂ ∂t  rt(cid:12)(cid:12)t=0(x) = ρ0(x) ≥ 0,  for a.a. x ∈ Rd. Hence, rt(x) = e−(m−κ+)teκ+tLa+ ρ0(x), where (La+f )(x) :=ZRd a+(x − y)[f (y) − f (x)]dy. Since for f ∈ L∞(Rd) we have (cid:12)(cid:12)(La+f )(x) ≤ 2kf kL∞(Rd) then, by (3.11), rt(x) ≤ Ce−(m−κ+)te2κ+t ≤ C, that yields 0 ≤ ρt(x) ≤ C. To prove the existence and uniqueness of the solution of (3.16) let us fix some T > 0 and define the Banach space XT = C([0; T ], L∞(Rd)) of all continuous functions on [0; T ] with values in L∞(Rd); the norm on XT is given by kukT := max T the cone of all nonnegative t∈[0;T ] kutkL∞(Rd). We denote by X + functions from XT . Let Φ be a mapping which assign to any v ∈ XT the solution ut of the linear Cauchy problem ut(x) = κ+(a+ ∗ vt)(x) − κ−ut(x)(a− ∗ vt)(x) − mut(x), ∂ ∂t  ut(cid:12)(cid:12)t=0(x) = ρ0(x),  (Φv)t(x) = exp(cid:26)−Z t for a.a. x ∈ Rd. Therefore, (3.17) (3.18) 0 (cid:0)m + κ−(a− ∗ vs)(x)(cid:1)ds(cid:27) ρ0(x) exp(cid:26)−Z t s (cid:0)m + κ−(a− ∗ vτ )(x)(cid:1)dτ(cid:27) κ+(a+ ∗ vs)(x)ds. +Z t 0 Operator approach to Vlasov scaling for models of spatial ecology 19 We have that v ∈ X + T implies Φv ≥ 0 as well as the estimate (Φv)t(x) ≤ ρ0(x) + κ+kvkT Z t 0 e−(t−s)mds ≤ C + κ+ m kvkT , where we use the trivial inequality kf ∗ gkL∞(Rd) ≤ kf kL1(Rd)kgkL∞(Rd), f ∈ L1(Rd), g ∈ L∞(Rd). (3.19) Therefore, Φv ∈ X + T . For simplicity of notations we denote for v ∈ X + T (Bv)(t, x) = m + κ−(a− ∗ vt)(x) ≥ m > 0. Then, for any v, w ∈ X + T (Bv)(s, x)ds(cid:27) − exp(cid:26)−Z t 0 (Bw)(s, x)ds(cid:27)(cid:12)(cid:12)(cid:12)(cid:12) ρ0(x) ds. We have 0 (cid:12)(cid:12)(Φv)t(x) − (Φw)t(x)(cid:12)(cid:12) exp(cid:26)−Z t ≤(cid:12)(cid:12)(cid:12)(cid:12) +Z t exp(cid:26)−Z t 0 (cid:12)(cid:12)(cid:12)(cid:12) − exp(cid:26)−Z t s s (Bv)(τ, x)dτ(cid:27) κ+(a+ ∗ vs)(x) (Bw)(τ, x)dτ(cid:27) κ+(a+ ∗ ws)(x)(cid:12)(cid:12)(cid:12)(cid:12) (Bw)(s, x)ds(cid:27)(cid:12)(cid:12)(cid:12)(cid:12) κ−(a− ∗ vs)(x)ds(cid:27) − exp(cid:26)−Z t κ−(a− ∗ ws)(x)ds(cid:12)(cid:12)(cid:12)(cid:12) kv − wkT , κ− (Bv)(s, x)ds(cid:27) − exp(cid:26)−Z t exp(cid:26)−Z t (cid:12)(cid:12)(cid:12)(cid:12) exp(cid:26)−Z t ≤e−mt(cid:12)(cid:12)(cid:12)(cid:12) Z t κ−(a− ∗ vs)(x)ds −Z t ≤e−mt(cid:12)(cid:12)(cid:12)(cid:12) ≤e−mtκ−kv − wkT · t ≤ 0 0 0 0 0 em 0 κ−(a− ∗ ws)(x)ds(cid:27)(cid:12)(cid:12)(cid:12)(cid:12) where we used (3.19) and obvious inequalities e−a −e−b ≤ a−b for a, b ≥ 0; e−xx ≤ e−1 for x ≥ 0. Next, using another simple estimates for any a, b, p, q ≥ 0 pe−a − qe−b ≤ e−ap − q + qe−be−(a−b) − 1 ≤ e−ap − q + qe−ba − b, Operator approach to Vlasov scaling for models of spatial ecology 20 we obtain s (Bv)(τ, x)dτ(cid:27) κ+(a+ ∗ vs)(x) (Bw)(τ, x)dτ(cid:27) κ+(a+ ∗ ws)(x)(cid:12)(cid:12)(cid:12)(cid:12) (Bv)(τ, x)dτ(cid:27)(cid:12)(cid:12)a+ ∗ (vs − ws)(cid:12)(cid:12)(x)ds (Bw)(τ, x)dτ(cid:27) (κ+a+ ∗ ws)(x) ds s s ds (Bw)(τ, x)dτ(cid:12)(cid:12)(cid:12)(cid:12) s 0 0 Z t exp(cid:26)−Z t 0 (cid:12)(cid:12)(cid:12)(cid:12) − exp(cid:26)−Z t ≤κ+Z t exp(cid:26)−Z t +Z t exp(cid:26)−Z t Z t (Bv)(τ, x)dτ −Z t ×(cid:12)(cid:12)(cid:12)(cid:12) ≤κ+kv − wkT Z t +Z t exp(cid:26)−Z t × e−m(t−s)Z t e−m(t−s)ds 0 0 s s s s κ−(a− ∗ wτ )(x)dτ(cid:27) (κ+a+ ∗ ws)(x) κ−(a− ∗ vτ − wτ )(x)dτ ds and, using (3.12) and the inequalities above, one can continue ≤ κ+ m kv − wkT + C 4 κ− em kv − wkT ×Z t 0 exp(cid:26)−Z t κ−(a− ∗ wτ )(x)dτ(cid:27) κ−(a− ∗ ws)(x)ds = κ+ m kv − wkT + kv − wkT s κ− em C 4 0 ×Z t ≤(cid:18) κ+ m + κ−(a− ∗ wτ )(x)dτ(cid:27) ds ∂ ∂s C 4 exp(cid:26)−Z t em(cid:19) kv − wkT . κ− s Therefore, for v, w ∈ X + T kΦv − ΦwkT ≤(cid:18) κ+ m +(cid:16)1 + C 4(cid:17) κ− em(cid:19) kv − wkT ≤ 4(κ+ + Cκ−) m kv − wkT , Operator approach to Vlasov scaling for models of spatial ecology 21 if, e.g., 1 + C 4 ≤ 4Ce, that means C ≥ 4 16e−1 . As a result, by (3.11), Φ is a contraction mapping on the cone X + T . Taking, as usual, v(n) = Φnv(0), n ≥ 1 for v(0) ∈ X + T we obtain that {v(n)} ⊂ X + T is a fundamental sequence in XT which has, therefore, a unique limit point v ∈ XT . Since X + T . Then, identically to the classical Banach fixed point theorem, v will be a fixed point of Φ on XT and a unique fixed point on X + T . Then, this v is the nonnegative solution of (3.16) on the interval [0; T ]. By the note above, vt(x) ≤ C. Changing initial T is a closed cone we have that v ∈ X + value in (3.16) onto ρt(cid:12)(cid:12)t=T (x) = vT (x) we may extend all our considerations on the time-interval [T ; 2T ] with the same estimate vt(x) ≤ C; and so on. As a a result, (3.16) has a unique global bounded non-negative solution ρt(x) on R+. Consider now then ∂ ∂t kt(η) = eλ(ρt, η) ∈ KC, eλ(ρt, η) =Xx∈η ∂ρt ∂t (x)eλ(ρt, η \ x). Using (3.16) and (3.14), we immediately conclude that kt(η) = eλ(ρt, η) is a solution to (3.15). The main result of the paper is formulated in the next theorem. Its proof will be given in Subsection 3.4. Theorem 3.12. Under conditions of Theorem 3.9 the semigroup Ut,ε con- verges strongly to the semigroup U V t as ε → 0 uniformly on any finite inter- vals of time. 3.4 Proofs According to Theorem 3.1, the statement of Theorem 3.12 will be proved once we verify Assumptions (A) for the operators (A1(ε), D1), (A2(ε), D1), ε > 0, defined in the previous subsection. Note, that A1(0) = A1 and A2(0) = A2 are defined on the domain D0. In the following proposition we verify Assumption 2(a) of (A). Proposition 3.13. Let λ > 0 then (A1(ε) − λ11)−1 s−→ (A1 − λ11)−1 , ε → 0. Operator approach to Vlasov scaling for models of spatial ecology 22 Proof. For any G ∈ LC G (η)(cid:18) (cid:13)(cid:13)(A1(ε) − λ11)−1 G − (A1 − λ11)−1 G(cid:13)(cid:13)C =ZΓ0(cid:12)(cid:12)(cid:12)(cid:12) =ZΓ0 −m η − εκ−Ea− (η) − λ G (η) Fε (η) C ηdλ (η) , 1 − C ηdλ (η) 1 −m η − λ(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) where Fε (η) := εκ−Ea− (η) (m η + εκ−Ea− (η) + λ) (m η + λ) , η ∈ Γ0. Since 0 ≤ Fε (η) < 1/λ and Fε (η) → 0 as ε → 0 for any η ∈ Γ0, we get the desired statement. Next we check Assumption 2(b) of (A). Proposition 3.14. Let λ > 0 be arbitrary and fixed. Then sup ε≥0(cid:13)(cid:13)(A1(ε) − λ11)−1(cid:13)(cid:13) ≤(cid:13)(cid:13)(A1 − λ11)−1(cid:13)(cid:13) . Proof. For any G ∈ LC and any ε > 0 1 G (η) (cid:13)(cid:13)(A1(ε) − λ11)−1 G(cid:13)(cid:13)C =ZΓ0 ≤ZΓ0 ≤(cid:13)(cid:13)(A1 − λ11)−1(cid:13)(cid:13) · kGkC . m η + λ G (η) 1 This finishes the proof. m η + εκ−Ea− (η) + λ C ηdλ (η) C ηdλ (η) =(cid:13)(cid:13)(A1 − λ11)−1 G(cid:13)(cid:13)C Assumption 2(c) of (A) is proved in the next Proposition. Proposition 3.15. Let conditions of Theorem 3.9 be satisfied. Then, for any λ > 0 sup ε≥0(cid:13)(cid:13)A2(ε) (A1(ε) − λ11)−1(cid:13)(cid:13) < 1 2 (3.20) Operator approach to Vlasov scaling for models of spatial ecology 23 Proof. First we prove assertion for ε = 0. Since D (A1) = D (A2) = D0 and Ran(cid:0)(A1 − λ11)−1(cid:1) = D (A1), the operator A2 (A1 − λ11)−1 is well defined. Next, inequality (3.11) and Proposition 3.6 yields (3.21) Indeed, (cid:13)(cid:13)A2 (A1 − λ11)−1(cid:13)(cid:13) < 1 4 . kA2GkC ≤ a kA1GkC < a k(A1 − λ11) GkC with a < 1 4 . Therefore, (cid:13)(cid:13)A2 (A1 − λ11)−1 G(cid:13)(cid:13)C < 1 4 kGkC , and (3.21) is proved. Now, let ε > 0 be arbitrary and fixed. The main arguments we use to show are the following: (cid:13)(cid:13)A2(ε) (A1(ε) − λ11)−1(cid:13)(cid:13) < 1 2 1) D (A1(ε)) = D1 ⊂ D0 = D (A2). Hence, A2 (A1(ε) − λ11)−1 is well- defined on LC. Moreover, Proposition 3.6 implies 2) D (B2) = D (A1(ε)) = D1 and for any ε > 0 1 4 ε > 0. , (cid:13)(cid:13)A2 (A1(ε) − λ11)−1(cid:13)(cid:13) < (cid:13)(cid:13)εB2 (A1(ε) − λ11)−1(cid:13)(cid:13) < 1 4 , which follows from Proposition 3.7. 3) Since A2(ε) := A2 + εB2, we have The latter concludes the proof. (cid:13)(cid:13)A2(ε) (A1(ε) − λ11)−1(cid:13)(cid:13) < 1 2 . (3.22) We set Qε =(cid:0)A2(ε) (A1(ε) − λ11)−1 + 1(cid:1)−1 , Q =(cid:0)A2 (A1 − λ11)−1 + 11(cid:1)−1 . Operator approach to Vlasov scaling for models of spatial ecology 24 In order to verify Assumption 2(d) of (A) we have to show that Qε as ε → 0. Suppose that we can show that A2 (A1(ε) − λ11)−1 εB2 (A1(ε) − λ11)−1 s−→ A2 (A1 − λ11)−1 , ε → 0. s−→ 0, ε → 0. s−→ Q (3.23) Then, Cε :=A2(ε) (A1(ε) − λ11)−1 =A2 (A1 + εB1 − λ11)−1 + εB2 (A1 + εB1 − λ11)−1 s−→ A2 (A1 − λ11)−1 To check we proceed as follows: (Cε + 11)−1 − Q Qε = (Cε + 11)−1 s−→ Q (3.24) = (Cε + 11)−1 −(cid:0)A2 (A1 − λ11)−1 + 11(cid:1)−1 = (Cε + 11)−1(cid:0)A2 (A1 − λ11)−1 + 11 − Cε − 11(cid:1)(cid:0)A2 (A1 − λ11)−1 + 11(cid:1)−1 = (Cε + 11)−1(cid:0)A2 (A1 − λ11)−1 − Cε(cid:1)(cid:0)A2 (A1 − λ11)−1 + 11(cid:1)−1 Assuming (3.23) it is obvious now that convergence (3.24) is equivalent to . which is clear from sup ε>0(cid:13)(cid:13)(Cε + 11)−1(cid:13)(cid:13) < ∞, (cid:13)(cid:13)(Cε + 11)−1(cid:13)(cid:13) ≤ 1 1 − kCεk and kCεk < 1 2 . The last bound we conclude from (3.22). As a result we shall have established Theorem 3.12 if we show (3.23). Lemma 3.16. A2 (A1(ε) − λ11)−1 s−→ A2 (A1 − λ11)−1 , as ε → 0. Proof of Lemma 3.16. Proposition 3.6 and D (A1(ε)) = D1 ⊂ D (A1) = D (A2) = D0 Operator approach to Vlasov scaling for models of spatial ecology 25 leads to the following formula A2 (A1(ε) − λ11)−1 = A2 (A1 − λ11)−1 (A1 − λ11) (A1(ε) − λ11)−1 . Now, we are left with the task to show that (A1 − λ11) (A1(ε) − λ11)−1 s−→ 1, as ε → 0. But, for any G ∈ LC m η + λ (cid:13)(cid:13)(cid:0)(A1 − λ11) (A1(ε) − λ11)−1 − 11(cid:1) G(cid:13)(cid:13)C =ZΓ0(cid:12)(cid:12)(cid:12)(cid:12) − 1(cid:12)(cid:12)(cid:12)(cid:12) =ZΓ0 m η + εκ−Ea− (η) + λ m η + εκ−Ea− (η) + λ εκ−Ea− (η) G (η) C ηdλ (η) G (η) C ηdλ (η) → 0, as ε → 0 due to the Lebesgue's dominated convergence theorem. Lemma 3.17. εB2 (A1(ε) − λ11)−1 s−→ 0, as ε → 0. Proof of Lemma 3.17. Since kB2GkC ≤ 1 4 kB1GkC, we have to show that But, (cid:13)(cid:13)εB1 (A1(ε) − λ11)−1 G(cid:13)(cid:13)C → 0, as ε → 0. (cid:13)(cid:13)εB1 (A1(ε) − λ11)−1 G(cid:13)(cid:13)C =ZΓ0 εκ−Ea− (η) m η + εκ−Ea− (η) + λ G (η) C ηdλ (η) → 0, ε → 0. The last two lemmas conclude the proof of the main Theorem. Remark 3.18. Under assumptions of Proposition 3.15 we get the following representation for the resolvents of V and Lε,ren (cid:16) V − λ11(cid:17)−1 (cid:16) Lε,ren − λ11(cid:17)−1 = (A1 + A2 − λ11)−1 = (A1 − λ11)−1(cid:0)A2 (A1 − λ11)−1 + 11(cid:1)−1 , = (A1(ε) + A2(ε) − λ11)−1 = (A1(ε) − λ11)−1(cid:0)A2(ε) (A1(ε) − λ11)−1 + 11(cid:1)−1 , (3.25) λ > 0. Operator approach to Vlasov scaling for models of spatial ecology 26 Acknowledgements The financial support of DFG through the SFB 701 (Bielefeld University) and German-Ukrainian Projects 436 UKR 113/97 is gratefully acknowledged. This work was partially supported by the Marie Curie "Transfer of Knowl- edge" programme, project TODEQ MTKD-CT-2005-030042 (Warsaw, IM- PAN). O.K. is very thankful to Prof. J. Zemanek for fruitful and stimulating discussions. References [1] B. Bolker and S. W. Pacala. Using moment equations to understand stochastically driven spatial pattern formation in ecological systems. Theor. Popul. Biol., 52(3):179 -- 197, 1997. [2] B. Bolker and S. W. Pacala. Spatial moment equations for plant com- petitions: Understanding spatial strategies and the advantages of short dispersal. American Naturalist, 153:575 -- 602, 1999. [3] W. Braun and K. Hepp. The Vlasov dynamics and its fluctuations in the 1/N limit of interacting classical particles. Comm. Math. Phys., 56(2):101 -- 113, 1977. [4] U. Dieckmann and R. Law. Relaxation projections and the method of moments. In The Geometry of Ecological Interactions, pages 412 -- 455. Cambridge University Press, Cambridge, UK, 2000. [5] R. L. Dobrushin. Vlasov equations. Functional Anal. Appl., 13(2):115 -- 123, 1979. [6] R. L. Dobrushin, Y. G. Sinai, and Y. M. Sukhov. Dynamical systems of statistical mechanics. In Y. G. Sinai, editor, Ergodic Theory with Applications to Dynamical Systems and Statistical Mechanics, volume II of Encyclopaedia Math. Sci., Berlin, Heidelberg, 1989. Springer. [7] K.-J. Engel and R. Nagel. One-parameter semigroups for linear evolution equations, volume 194 of Graduate Texts in Mathematics. Springer- Verlag, New York, 2000. Operator approach to Vlasov scaling for models of spatial ecology 27 [8] D. Finkelshtein, Y. Kondratiev, and O. Kutoviy. Vlasov scaling for the Glauber dynamics in continuum. CRC-701 Preprint 10055, University of Bielefeld; arXiv:1002.4762v3, 2011. [9] D. Finkelshtein, Y. Kondratiev, and O. Kutoviy. Individual based model with competition in spatial ecology. SIAM J. Math. Anal., 41(1):297 -- 317, 2009. [10] D. Finkelshtein, Y. Kondratiev, and M. J. Oliveira. Markov evolutions and hierarchical equations in the continuum. I. One-component systems. J. Evol. Equ., 9(2):197 -- 233, 2009. [11] N. Fournier and S. Meleard. A microscopic probabilistic description of a locally regulated population and macroscopic approximations. The Annals of Applied Probability, 14(4):1880 -- 1919, 2004. [12] T. Kato. Perturbation theory for linear operators. Springer-Verlag, Berlin, second edition, 1976. Grundlehren der Mathematischen Wis- senschaften, Band 132. [13] Y. Kondratiev and T. Kuna. Harmonic analysis on configuration space. I. General theory. Infin. Dimens. Anal. Quantum Probab. Relat. Top., 5(2):201 -- 233, 2002. [14] Y. Kondratiev and O. Kutoviy. On the metrical properties of the con- figuration space. Math. Nachr., 279(7):774 -- 783, 2006. [15] Y. Kondratiev, R. Minlos, and E. Zhizhina. One-particle subspace of the Glauber dynamics generator for continuous particle systems. Rev. Math. Phys., 16(9):1073 -- 1114, 2004. [16] V. V. Kozlov. The generalized Vlasov kinetic equation. Russian Math. Surveys, 63(4):691 -- 726, 2008. [17] A. Lenard. States of classical statistical mechanical systems of infinitely many particles. I. Arch. Rational Mech. Anal., 59(3):219 -- 239, 1975. [18] A. Lenard. States of classical statistical mechanical systems of infinitely many particles. II. Characterization of correlation measures. Arch. Ra- tional Mech. Anal., 59(3):241 -- 256, 1975. Operator approach to Vlasov scaling for models of spatial ecology 28 [19] M. Reed and B. Simon. Methods of Modern Mathematical Physics, vol- ume Vol. 1: Functional Analysis. Academic Press, 1981. [20] D. Ruelle. Statistical mechanics: Rigorous results. W. A. Benjamin, Inc., New York-Amsterdam, 1969. [21] D. Ruelle. Superstable interactions in classical statistical mechanics. Comm. Math. Phys., 18:127 -- 159, 1970. [22] H. Spohn. Kinetic equations from Hamiltonian dynamics: Markovian limits. Rev. Modern Phys., 52(3):569 -- 615, 1980. [23] H. Spohn. Large Scale Dynamics of Interacting Particles. Texts and Monographs in Physics. Springer, November 1991.
1209.3639	2	1209	2013-05-03T14:39:43	An algebraic construction of quantum flows with unbounded generators	[ "math.FA", "math-ph", "math-ph", "math.OA" ]	It is shown how to construct -homomorphic quantum stochastic Feller cocycles for certain unbounded generators, and so obtain dilations of strongly continuous quantum dynamical semigroups on C algebras; this generalises the construction of a classical Feller process and semigroup from a given generator. The construction is possible provided the generator satisfies an invariance property for some dense subalgebra A_0 of the C* algebra A and obeys the necessary structure relations; the iterates of the generator, when applied to a generating set for A_0, must satisfy a growth condition. Furthermore, it is assumed that either the subalgebra A_0 is generated by isometries and A is universal, or A_0 contains its square roots. These conditions are verified in four cases: classical random walks on discrete groups, Rebolledo's symmetric quantum exclusion processes and flows on the non-commutative torus and the universal rotation algebra.	math.FA	math	An algebraic construction of quantum flows with unbounded generators Alexander C. R. Belton Department of Mathematics and Statistics Lancaster University, United Kingdom Stephen J. Wills School of Mathematical Sciences University College Cork, Ireland [email protected] [email protected] 6th November 2018 Abstract It is shown how to construct ∗-homomorphic quantum stochastic Feller cocycles for certain unbounded generators, and so obtain dilations of strongly continuous quantum dynamical semigroups on C ∗ algebras; this generalises the construction of a classical Feller process and semigroup from a given generator. The construction is possible provided the generator satisfies an invariance property for some dense subalgebra A0 of the C ∗ algebra A and obeys the necessary structure relations; the iterates of the generator, when applied to a generating set for A0, must satisfy a growth condition. Furthermore, it is assumed that either the subalgebra A0 is generated by isometries and A is universal, or A0 contains its square roots. These conditions are verified in four cases: classical random walks on discrete groups, Rebolledo's symmetric quantum exclusion processes and flows on the non-commutative torus and the universal rotation algebra. Key words: quantum dynamical semigroup; quantum Markov semigroup; CPC semigroup; strongly continuous semigroup; semigroup dilation; Feller cocycle; higher-order Ito product formula; random walks on discrete groups; quantum exclusion process; non-commutative torus MSC 2000: 81S25 (primary); 46L53, 46N50, 47D06, 60J27 (secondary). 1 Introduction The connexion between time-homogeneous Markov processes and one-parameter contraction semigroups is an excellent example of the interplay between probability theory and functional analysis. Given a measurable space (E,E), a Markov semigroup T with state space E is a family (Tt)t>0 of positive contraction operators on L∞(E) such that Ts+t = Ts ◦ Tt for all s, t > 0 and T0f = f for all f ∈ L∞(E); 3 1 0 2 y a M 3 ] . A F h t a m [ 2 v 9 3 6 3 . 9 0 2 1 : v i X r a the semigroup is conservative if Tt1 = 1 for all t > 0. Typically, such a semigroup is defined by setting (Ttf )(x) =ZE f (y)pt(x, dy) 1 for a family of transition kernels pt : E ×E → [0, 1]. Given a time-homogeneous Markov process (Xt)t>0 with values in E, the associated Markov semigroup is obtained from the prescription (Ttf )(x) = E[f (Xt)X0 = x], (1.1) so that pt(x, A) = P(Xt ∈ AX0 = x) is the probability of moving from x into A in time t. When the state space E is a locally compact Hausdorff space we may specialise further: a Feller semigroup is a Markov semigroup T such that Tt(cid:0)C0(E)(cid:1) ⊆ C0(E) for all t > 0 and kTtf − fk∞ → 0 as t → 0 for all f ∈ C0(E). Any sufficiently nice Markov process, such as a L´evy process, gives rise to a Feller semigroup; conversely, if E is separable then any Feller semigroup gives rise to a Markov process with c`adl`ag paths. A celebrated theorem of Gelfand and Naimark states that every commutative C ∗ algebra is of the form C0(E) for some locally compact Hausdorff space E. Thus the first step in generalising Feller semigroups, and so Markov processes, to a non-commutative setting is to replace C0(E) with a general C ∗ algebra A. Moreover, a strengthening of positivity, called complete positivity, is required for a satisfactory theory: a map φ : A → B between C ∗ algebras is completely positive if the ampliation φ(n) : Mn(A) → Mn(B); (xij) 7→(cid:0)φ(xij)(cid:1) is positive for all n > 1. This property is justified on physical grounds and is equivalent to the usual form of positivity when either algebra A or B is commutative. The resulting object, a semigroup of completely positive contractions on a C ∗ algebra A, is known as a quantum dynamical semigroup or, when conservative, a quantum Markov semigroup; such semigroups are used to describe the evolution of quantum-mechanical systems which interact irreversibly with their environment. Any strongly continuous quantum dynamical semigroup T is characterised by its infinitesimal generator τ , the closed linear operator such that dom τ =nf ∈ A : lim t→0 Ttf − f t existso and τ f = lim t→0 Ttf − f t . For a Feller semigroup, the form of the generator τ may reveal properties of the corresponding process; for instance, a classical L´evy process may be specified, via the L´evy -- Khintchine formula, by the characteristics of its generator, viz. a drift vector, a diffusion matrix describing the Brownian-motion component and a L´evy measure characterising its jumps. If we start with a putative generator τ then operator-theoretic methods may be used to construct the semigroup, although there are often considerable analytical challenges to be met. Verifying that τ satisfies the hypotheses of the Hille -- Yosida theorem, the key analytical tool for this construction, is often difficult. In this paper we provide, for a suitable class of generators, another method of constructing quantum dynamical semigroups and the corresponding non-commutative Markov processes. 2 To understand how the relationship between semigroups and Markov processes generalises to the non-commutative framework, recall first that any locally compact Hausdorff space E may be made compact by adjoining a point at infinity, which corresponds to adding an identity to the algebra C0(E) or adding a coffin state for an E-valued Markov process; it is sufficient, therefore, to restrict our attention to compact Hausdorff spaces or, equivalently, unital C ∗ algebras. The correct analogue of an E-valued random variable X is then a unital ∗-homomorphism j from A to some unital C ∗ algebra B; classically, j is the map f 7→ f ◦ X, where f ∈ A = C0(E) and B is L∞(P) for some probability measure P. A family of unital ∗-homomorphisms (jt : A → B)t>0, i.e., a non-commutative stochastic process, is said to dilate the quantum dynamical semigroup T if A is a subalgebra of B and E ◦ jt = Tt for all t > 0, where E is a conditional expectation from B to A; the relationship to (1.1) is clear. Thus finding a dilation for a given semigroup is analogous to constructing a Markov process from a family of transition kernels. The tool used here for constructing semigroups and their dilations is a stochastic calculus: the quantum stochastic calculus introduced by Hudson and Parthasarathy in their 1984 paper [15]. In its simplest form, this is a non-commutative theory of stochastic integration with respect to three operator martingales which correspond to the creation, annihilation and gauge processes of quantum field theory. It generalises simultaneously the Ito -- Doob L2 integral with respect to either Brownian motion or the compensated Poisson process; as emphasised by Meyer [25] and Attal [3], the L2 theory of any normal martingale having the chaotic-representation property, such as Brownian motion, the compensated Poisson process or Az´ema's martingale, gives a classical probabilistic interpretation of Boson Fock space, the ambient space of quantum stochastic calculus. We develop below new techniques for obtaining ∗-homomorphic solutions to the Evans -- Hudson quantum stochastic differential equation (QSDE) djt = (jt ⊗ ιB(bk)) ◦ φ dΛt, (1.2) where the solution jt acts on a unital C ∗ algebra A. In this way, we obtain the process j and the quantum dynamical semigroup T simultaneously. The components of the flow generator φ include τ , the restriction of a semigroup generator, and δ, a bimodule derivation, which are related to one another through the Bakry -- ´Emery carr´e du champ operator: see Remark 2.4. If A is commutative then, by Theorem 3.19, the process j is classical, in the sense that the algebra generated by {jt(a) : t > 0, a ∈ A} is also commutative. The recent expository paper [4] on quantum stochastic methods, written for an audience of probabilists, includes Parthasarathy and Sinha's method [26] for constructing continuous-time Markov chains with finite state spaces by solving quantum stochastic differential equations. To quote Biane, "It may seem strange to the classical probabilist to use noncommutative objects in order to describe a perfectly commutative situation, however, this seems to be necessary if one wants to deal with processes with jumps . . . The right mathem- atical notion . . . , which generalizes to the noncommutative situation, is that of a derivation into a bimodule . . . Using this formalism, we can use the Fock space as 3 a uniform source of noise, and construct general Markov processes (both continuous and discontinuous) using stochastic differential equations.". The results herein give a further illustration of this philosophy. The use of quantum stochastic calculus to produce dilations has now been studied for nearly thirty years. Most results, by Hudson and Parthasarathy, Fagnola, Mohari, Sinha et cetera, are obtained in the case that A = B(h) by first solving an operator-valued QSDE, the Hudson -- Parthasarathy equation, to obtain a unitary process U , and defining j through conjugation by U ; see [10] and references therein. The corresponding theory for the Heisenberg rather than the Schrodinger viewpoint, solving the Evans -- Hudson equation (1.2), has mainly been developed under the standing assumption that the generator φ is completely bounded, which is necessary if the corresponding semigroup T is norm continuous [21]. When one deviates from this assumption, which is analytically convenient but very restrictive, there are few results. The earliest general method is due to Fagnola and Sinha [11], with later results by Goswami, Sahu and Sinha for a particular model [14] and a more general method developed by Goswami and Sinha in [29]. Another approach based on semigroup methods has yet to yield existence results for the Evans -- Hudson equation: see [1] and [24]. Our method here has an attractive simplicity, imposing minimal conditions on the generator φ. It must be a ∗-linear map φ : A0 → A0 ⊗ B(C ⊕ k), where A0 is a dense ∗-subalgebra of the unital C ∗ algebra A ⊆ B(h) which contains 1 = 1h and k is a Hilbert space, called the multiplicity space, the dimension of which measures the amount of noise available in the system. This incorporates an assumption that, if φ is viewed as a matrix of maps, its components leave A0 invariant, a hypothesis also used in [11]. Furthermore, φ must be such that φ(1) = 0 and the first-order Ito product formula holds: φ(xy) = φ(x)(y ⊗ 1bk) + (x ⊗ 1bk)φ(y) + φ(x)∆φ(y) for all x, y ∈ A0, (1.3) where bk := C ⊕ k and ∆ ∈ A0 ⊗ B(bk) is the orthogonal projection onto h ¯⊗ k. Both these conditions are known to be necessary if φ is to generate a family of unital ∗-homomorphisms. Finally, a growth bound must be established for the iterates of φ applied to elements taken from a suitable subset of A0. Our approach is an elementary one for those adept in quantum stochastic calculus, relying on familiar techniques such as representing the solution to the Evans -- Hudson QSDE as a sum of quantum Wiener integrals. An essential tool is the higher-order Ito product formula, presented in Section 2. This formula was first stated, for finite-dimensional noise, in [8], was proved for that case in [16] and reached its definitive form in [23]. In that last paper it was shown that (1.3) is but the first of a sequence of identities that must be satisfied in order to show that the solution j of the QSDE is weakly multiplicative. However, there are many situations in which the validity of (1.3) implies that the other identities hold [23, Corollary 4.2], and this is the case for φ as above. Moreover, one of our main observations, Corollary 2.12, is that, by exploiting the algebraic structure imposed by this sequence of identities, it is sufficient to 4 establish pointwise growth bounds on a ∗-generating set of A0; this is a major simplification when compared with [11]. Also, by using the coordinate-free approach to quantum stochastic analysis given in [19], we can take k to be any Hilbert space, removing the restriction in [11] that k be finite dimensional. The growth bounds obtained Section 2 are employed in Section 3 to produce a family of weakly multiplicative ∗-linear maps from the algebra A0 into the space of linear operators in h ¯⊗ F, where F is the Boson Fock space over L2(R+; k). It is shown that these maps extend to unital ∗- homomorphisms in two distinct situations. Theorem 3.9, which includes the case of AF algebras, exploits a square-root trick that is well known in the literature; Theorem 3.12, which applies to universal C ∗ algebras such as the non-commutative torus or the Cuntz algebras, is believed to be novel. Uniqueness of the solution is proved, and it is also shown that j is a cocycle, i.e., it satisfies the evolution equation js+t = (js ¯⊗ ιB(F[s,∞))) ◦ σs ◦ jt for all s, t > 0, (1.4) where (σt)t>0 is the shift semigroup on the algebra of all bounded operators on F. At this point we see another novel feature of our work in contrast to previous results, all of which start with a particular quantum dynamical semigroup T . In these other papers the generator τ of T is then augmented to produce φ, and the QSDE solved to give a dilation of T . For example, in [11] it is assumed that T is an analytic semigroup and that the composition of τ with the other components of φ is well behaved in a certain sense; in [29] it is assumed that T is covariant with respect to some group action on A. For us, the starting point is the map φ, which yields the cocycle j, and hence, by compression, a quantum dynamical semigroup T generated by the closure of τ , which has core A0; this semigroup, a fortiori, is dilated by j. Thus we do not have to check that τ is a semigroup generator with good properties at the outset, thereby rendering our method easier to apply. Our first application of Theorem 3.9, in Section 4, is to construct the Markov semigroups which correspond to certain random walks on discrete groups. Theorem 3.9 is also employed in Section 5 to produce a dilation of the symmetric quantum exclusion semigroup. This object, a model for systems of interacting quantum particles, was introduced by Rebolledo [27] as a non-commutative generalisation of the classical exclusion process [18] and has generated much interest: see [13] and [12]. The multiplicity space k is required to be infinite dimensional for this process, as in previous work on processes arising from quantum interacting particle systems, e.g., [14]. In Section 6 we use Theorem 3.12 to obtain flows on some universal C ∗ algebras, namely the non- commutative torus and the universal rotation algebra [2]; the former is a particularly important example in non-commutative geometry. Quantum flows on these algebras have previously been considered by Goswami, Sahu and Sinha [14] and by Hudson and Robinson [17], respectively. 1.1 Conventions and notation The quantity := is to be read as 'is defined to be' or similarly. The quantity 1P equals 1 if the proposition P is true and 0 if P is false, where 1 and 0 are the appropriate multiplicative 5 and additive identities. The set of natural numbers is denoted by N := {1, 2, 3, . . .}; the set of non-negative integers is denoted by Z+ := {0, 1, 2, . . .}. The linear span of the set S in the vector space V is denoted by lin S; all vector spaces have complex scalar field and inner products are linear on the right. The algebraic tensor product is denoted by ⊗; the Hilbert- space tensor product is denoted by ¯⊗, as is the ultraweak tensor product. The domain of the linear operator T is denoted by dom T . The identity transformation on the vector space V is denoted by 1V . If P is an orthogonal projection on the inner-product space V then the complement P ⊥ := 1V − P , the projection onto the orthogonal complement of the range of P . The Banach space of bounded operators from the Banach space X1 to the Banach space X2 is denoted by B(X1; X2), or by B(X1) if X1 and X2 are equal. The identity automorphism on the algebra A is denoted by ιA. If a and b are elements in an algebra A then [a, b] := ab − ba and {a, b} := ab + ba denote their commutator and anti-commutator, respectively. If A0 is a ∗-algebra, H1 and H2 are Hilbert spaces and α : A0 → B(H1; H2) is a linear map then the adjoint map α† : A0 → B(H2; H1) is such that α†(a) := α(a∗)∗ for all a ∈ A0. 2 A higher-order product formula Notation 2.1. The Dirac bra-ket notation will be useful: for any Hilbert space H and vectors ξ, χ ∈ H, let Hi := B(C; H), hH := B(H; C), ξi : C → H; λ 7→ λξ hχ : H → C; η 7→ hχ, ηi (ket) (bra). and In particular, we have the linear map ξihχ ∈ B(H) such that ξihχη = hχ, ηiξ for all η ∈ H. Let A ⊆ B(h) be a unital C ∗ algebra with identity 1 = 1h, whose elements act as bounded operators on the initial space h, a Hilbert space. Let A0 ⊆ A be a norm-dense ∗-subalgebra of A which contains 1. Let the extended multiplicity spacebk := C⊕ k, where the multiplicity space k is a Hilbert space, and distinguish the unit vector ω := (1, 0). For brevity, let B := B(bk). Let ∆ := 1 ⊗ Pk ∈ A0 ⊗ B, where Pk := ωihω⊥ ∈ B is the orthogonal projection onto k ⊂bk. Lemma 2.2. The map φ : A0 → A0 ⊗ B is ∗-linear, such that φ(1) = 0 and such that φ(xy) = φ(x)(y ⊗ 1k) + (x ⊗ 1k)φ(y) + φ(x)∆φ(y) for all x, y ∈ A0 if and only if (2.1) (2.2) where π : A0 → A0 ⊗ B(k) is a unital ∗-homomorphism, δ : A0 → A0 ⊗ ki is a π-derivation, i.e., a linear map such that φ(x) ="τ (x) δ(x) π(x) − x ⊗ 1k# δ†(x) for all x ∈ A0, δ(xy) = δ(x)y + π(x)δ(y) for all x, y ∈ A0, 6 and τ : A0 → A0 is a ∗-linear map such that τ (xy) − τ (x)y − xτ (y) = δ†(x)δ(y) for all x, y ∈ A0. (2.3) Proof. This is a straightforward exercise in elementary algebra. Definition 2.3. A ∗-linear map φ : A0 → A0 ⊗ B such that φ(1) = 0 and such that (2.1) holds is a flow generator. Remark 2.4. Condition (2.3) may be expressed in terms of the Bakry -- ´Emery carr´e du champ operator Γ : A0 × A0 → A0; (x, y) 7→ 1 2(cid:0)τ (xy) − τ (x)y − xτ (y)(cid:1); for (2.3) to be satisfied, it is necessary and sufficient that 2Γ(x, y) = δ†(x)δ(y) for all x, y ∈ A0. The π-derivation δ becomes a bimodule derivation if A0 ⊗ ki is made into an A0-A0 bimodule by setting x · z · y := π(x)zy for all x, y ∈ A0 and z ∈ A0 ⊗ ki. Lemma 2.5. Let A0 = A, let π : A → A ⊗ B(k) be a unital ∗-homomorphism, let z ∈ A ⊗ ki and let h ∈ A be self adjoint. Define and δ : A → A ⊗ ki; x 7→ zx − π(x)z τ : A → A; x 7→ i[h, x] − 1 2{z∗z, x} + z∗π(x)z. Then the map φ : A → A ⊗ B defined in terms of π, δ and τ through (2.2) is a flow generator. Proof. This is another straightforward exercise. Remark 2.6. Modulo important considerations regarding tensor products and the ranges of δ and τ , the above form for φ is, essentially, the only one possible [20, Lemma 6.4]. The quantum exclusion process in Section 5 has a generator of the same form but with unbounded z and h. Definition 2.7. Given a flow generator φ : A0 → A0 ⊗B, the quantum random walk (cid:0)φn(cid:1)n∈Z+ is a family of ∗-linear maps defined by setting φ0 := ιA0 and φn : A0 → A0 ⊗ B⊗n φn+1 :=(cid:0)φn ⊗ ιB(cid:1) ◦ φ for all n ∈ Z+. The following identity is useful: if ξ1, χ1, . . . , ξn, χn ∈bk and x ∈ A0 then (cid:0)1h ⊗ hξ1 ⊗ ··· ⊗ hξn(cid:1)φn(x)(cid:0)1h ⊗ χ1i ⊗ ··· ⊗ χni(cid:1) = φξ1 where φξ χ : A0 → A0; x 7→ (1h ⊗ hξ)φ(x)(1h ⊗ χi) is a linear map for each choice of ξ, χ ∈bk. 7 χ1 ◦ ··· ◦ φξn χn(x), (2.4) Remark 2.8. The paper [23], results from which will be employed below, uses a different convention to that adopted in Definition 2.7: the components of the product B⊗n appear in the reverse order to how they do above. Notation 2.9. Let α ⊆ {1, . . . , n}, with elements arranged in increasing order, and denote its cardinality by α. The unital ∗-homomorphism is defined by linear extension of the map A0 ⊗ B⊗α → A0 ⊗ B⊗n; T 7→ T (n, α) where A ⊗ B1 ⊗ ··· ⊗ Bα 7→ A ⊗ C1 ⊗ ··· ⊗ Cn, Ci :=( Bj 1bk if i is the jth element of α, if i is not an element of α. For example, if α = {1, 3, 4} and n = 5 then (A ⊗ B1 ⊗ B2 ⊗ B3)(5, α) = A ⊗ B1 ⊗ 1bk ⊗ B2 ⊗ B3 ⊗ 1bk. Given a flow generator φ : A0 → A0 ⊗ B, for all n ∈ Z+ and α ⊆ {1, . . . , n}, let and let φα(x; n, α) :=(cid:0)φα(x)(cid:1)(n, α) ∆(n, α) := (1h ⊗ P ⊗α k for all x ∈ A0 )(n, α), the others. so that, in the latter, Pk acts on the components ofbk⊗n which have indices in α and 1bk acts on Theorem 2.10. Let(cid:0)φn(cid:1)n∈Z+ be the quantum random walk given by the flow generator φ. For all n ∈ Z+ and x, y ∈ A0, φn(xy) = Xα∪β={1,...,n} φα(x; n, α)∆(n, α ∩ β)φβ(y; n, β), (2.5) where the summation is taken over all sets α and β whose union is {1, . . . n}. Proof. This may be established inductively: see [23, Proof of Theorem 4.1]. Definition 2.11. The set S ⊆ A0 is ∗-generating for A0 if A0 is the smallest unital ∗-algebra which contains S. Corollary 2.12. For a flow generator φ : A0 → A0 ⊗ B, let Aφ := {x ∈ A0 : there exist Cx, Mx > 0 such that kφn(x)k 6 CxM n x for all n ∈ Z+}. (2.6) Then Aφ is a unital ∗-subalgebra of A0, which is equal to A0 if Aφ contains a ∗-generating set for A0. 8 Proof. It suffices to demonstrate that Aφ is closed under products. To see this, let x, y ∈ Aφ and suppose Cx, Mx and Cy, My are as in (2.6). Then (2.5) implies that kφα(x)kkφβ(y)k kφn(xy)k 6 Xα∪β={1,...,n} nXk=0(cid:18)n nXk=0(cid:18)n 6 CxCy = CxCy x k(cid:19)M k k(cid:19)M k kXl=0(cid:18)k l(cid:19)M n−k+l y (k = α, l = α ∩ β) x M n−k y (1 + My)k = CxCy(Mx + MxMy + My)n for all n ∈ Z+, as required. Lemma 2.13. If the flow generator φ is as defined in Lemma 2.5 then Aφ = A0. Proof. This follows immediately, since φ is completely bounded and kφnk 6 kφnkcb 6 kφkn all n ∈ Z+. The following result shows that, given a flow generator φ and vectors χ, ξ ∈bk, the elements of Aφ are entire vectors for φξ χ. Lemma 2.14. Let φ : A0 → A0⊗B be a flow generator. For all ξ, χ ∈bk we have φξ χ(Aφ) ⊆ Aφ, and the series cb for χ)n exp(zφξ χ) := zn(φξ n! ∞Xn=0 is strongly absolutely convergent on Aφ for all z ∈ C. Proof. Suppose kφn(x)k 6 CxM n (2.7) (2.8) (2.9) so x for all n ∈ Z+. It follows from (2.4) that χ(x)(cid:1), kχk = (kξkkχkCxMx)M n (cid:0)1h ¯⊗bk ¯⊗n ⊗ hξ(cid:1)φn+1(x)(cid:0)1h ¯⊗bk ¯⊗n ⊗ χi(cid:1) = φn(cid:0)φξ χ(x)(cid:1)(cid:13)(cid:13) 6 kξkCxM n+1 (cid:13)(cid:13)φn(cid:0)φξ (cid:13)(cid:13)(cid:0)φξ1 χn(cid:1)(x)(cid:13)(cid:13) 6 kξ1k··· kξnkkχ1k··· kχnkCxM n χ1 ◦ ··· ◦ φξn x x x , and φξ χ(x) ∈ Aφ. Moreover (2.4) also gives that hence the series (2.7) converges as claimed. 9 3 Quantum flows Notation 3.1. Let F denote Boson Fock space over L2(R+; k), the Hilbert space of k-valued, square-integrable functions on the half line, and let E := lin{ε(f ) : f ∈ L2(R+; k)} denote the linear span of the total set of exponential vectors in F. As is customary, elementary tensors in h ⊗ F are written without a tensor-product sign: in other words, uε(f ) := u ⊗ ε(f ) for all u ∈ h and f ∈ L2(R+; k), et cetera. If f ∈ L2(R+; k) and t > 0 then bf (t) := df (t), where bξ := ω + ξ ∈bk for all ξ ∈ k. Given f ∈ L2(R+; k) and an interval I ⊆ R+, let fI ∈ L2(R+; k) be defined to equal f on I and 0 elsewhere, with ft) := f[0,t) and f[t := f[t,∞) for all t > 0. Definition 3.2. A family of linear operators (Tt)t>0 in h ¯⊗ F with domains including h ⊗ E is adapted if huε(f ), Ttvε(g)i = huε(ft)), Ttvε(gt))ihε(f[t), ε(g[t)i for all u, v ∈ h, f , g ∈ L2(R+; k) and t > 0. linear operators in h ¯⊗ F , with domains including h ⊗ E, that is adapted and such that Theorem 3.3. For all n ∈ N and T ∈ B(h ¯⊗bk ¯⊗n) there exists a family Λn(T ) =(cid:0)Λn t (T )vε(g)i =ZDn(t)hu ⊗ bf ⊗n(t), T v ⊗bg⊗n(t)i dthε(f ), ε(g)i for all u, v ∈ h, f , g ∈ L2(R+; k) and t > 0. Here the simplex huε(f ), Λn t (T )(cid:1)t>0 of Dn(t) := {t := (t1, . . . , tn) ∈ [0, t]n : t1 < ··· < tn} bf ⊗n(t) := bf (t1) ⊗ ··· ⊗ bf (tn), et cetera. t (T ) := T ⊗ 1F for all t > 0. (3.1) (3.2) and We extend this definition to include n = 0 by setting Λ0 If f ∈ L2(R+; k) then kΛn K n f,t√n! kTkkuε(f )k where Kf,t :=p(2 + 4kfk2)(t + kfk2), and the map R+ → B(h; h ¯⊗ F); t 7→ Λn t (T )uε(f )k 6 is norm continuous. 10 for all t > 0 and u ∈ h, t (T )(cid:0)1h ⊗ ε(f )i(cid:1) Proof. This is an extension of Proposition 3.18 of [19], from which we borrow the notation; as for Remark 2.8, the ordering of the components in tensor products is different but this is no more than a convention. For each f ∈ L2(R+; k) define Cf > 0 so that and note that, by inequality (3.21) of [19], C 2 f =(cid:0)kfk +p1 + kfk2(cid:1)2 6 2 + 4kfk2, kΛn t (T )uε(f )k2 6(cid:0)Cft)(cid:1)2nZDn(t) kT u ⊗ bf ⊗n(t)k2 dtkε(f )k2 6 K 2n f,t n! kTk2kuε(f )k2. To show continuity, let eT denote T considered as an operator on (h ¯⊗bk) ¯⊗bk ¯⊗(n−1), where the right-most copy ofbk in the n-fold tensor product has moved next to the initial space h. Then Λn t (T ) − Λn and so, using Theorem 3.13 of [19], k(cid:0)Λn t (T ) − Λn s (T )(cid:1)uε(f )k2 6 2(t + C 2 s (T ) = Λt(cid:0)1(s,t](·)Λn−1 · (eT )(cid:1), (eT )(cid:0)u ⊗ bf (r)(cid:1)ε(f )k2 dr r f )Z t s kΛn−1 f )(cid:16)Z t s kbf (r)k2 dr(cid:17) K 2n−2 f,t (n − 1)! kTk2kuε(f )k2. 6 2(t + C 2 The family Λn(T ) is the n-fold quantum Wiener integral of T . Remark 3.4. It may be shown [23, Proof of Theorem 2.2] that dom Λl t(S)∗ ⊇ Λm t (T )(h ⊗ E) Theorem 3.5. Let φ : A0 → A0 ⊗ B be a flow generator. If x ∈ Aφ then the series for all l, m ∈ Z+, S ∈ B(h ¯⊗bk ¯⊗l), T ∈ B(h ¯⊗bk ¯⊗m) and t > 0. t(cid:0)φn(x)(cid:1) ∞Xn=0 jt(x) := Λn is strongly absolutely convergent on h ⊗ E for all t > 0, uniformly so on compact subsets of R+. The map R+ → B(h; h ¯⊗ F); t 7→ jt(x)(cid:0)1h ⊗ ε(f )i(cid:1) is norm continuous for all f ∈ L2(R+; k), the family(cid:0)jt(x)(cid:1)t>0 0 (cid:10)uε(f ), js(cid:0)φ for all u, v ∈ h, f , g ∈ L2(R+; k), x ∈ Aφ and t > 0. Furthermore, (cid:0)1h ⊗ hε(f )(cid:1)jt(x)(cid:0)1h ⊗ ε(g)i(cid:1) ∈ A huε(f ), jt(x)vε(g)i = huε(f ), (xv)ε(g)i +Z t for all x ∈ Aφ, f , g ∈ L2(R+; k) and t > 0. is adapted and bf (s) bg(s) (x)(cid:1)vε(g)(cid:11) ds (3.3) (3.4) (3.5) 11 Proof. The first two claims are a consequence of the estimate (3.2), the definition of Aφ and the continuity result from Theorem 3.3; adaptedness is inherited from the adaptedness of the quantum Wiener integrals. Lemma 2.14 implies that the integrand on the right-hand side of (3.4) is well defined and, by (2.8), (cid:10)uε(f ), Λn s(cid:0)φn(cid:0)φ bf (s) bg(s) (x)(cid:1)(cid:1)vε(g)(cid:11) =ZDn(s)hu ⊗ bf ⊗n(t), φn(cid:0)φ =ZDn(s)hu ⊗ bf ⊗n(t) ⊗ bf (s), φn+1(x)v ⊗bg⊗n(t) ⊗bg(s)i dthε(f ), ε(g)i; bg(s) (x)(cid:1)v ⊗bg⊗n(t)i dthε(f ), ε(g)i bf (s) integrating with respect to s then taking the sum of these terms gives (3.4). For the final claim, note that for any f , g ∈ L2(R+; k), the A0-valued map (3.6) Dn(t) ∋ t 7→ φ is Bochner integrable, hence bf (tn) bf (t1) bg(t1) ◦ ··· ◦ φ bg(tn) (x) =(cid:0)1h ⊗ hbf ⊗n(t)(cid:1)φn(x)(cid:0)1h ⊗ bg⊗n(t)i(cid:1) bg(tn)(cid:1)(x) dt ∈ A. bf (t1) bg(t1) ◦ ··· ◦ φ t(cid:0)φn(x)(cid:1)(cid:0)1h ⊗ ε(g)i(cid:1) = ehf,giZDn(t)(cid:0)φ bf (tn) (cid:0)1h ⊗ hε(f )(cid:1)Λn By (2.9), we may sum (3.6) over all n ∈ Z+, with the resulting series being norm convergent, and so the final claim follows. Remark 3.6. For all t > 0, let jt be as in Theorem 3.5. Since Aφ is a subspace of A0 containing 1, and each φn is linear with φn(1) = 0, it follows from (3.3) and Theorem 3.3 that each jt is linear and unital, as a map into the space of operators with domain h ⊗ E. Moreover, the maps jt are weakly ∗-homomorphic in the following sense. Lemma 3.7. Let φ : A0 → A0 ⊗ B be a flow generator and let jt be as in Theorem 3.5 for all t > 0. If x, y ∈ Aφ then x∗y ∈ Aφ, with hjt(x)uε(f ), jt(y)vε(g)i = huε(f ), jt(x∗y)vε(g)i (3.7) for all u, v ∈ h and f , g ∈ L2(R+; k). In particular, if x ∈ Aφ then jt(x)∗ ⊇ jt(x∗). Proof. As Aφ is a ∗-algebra, so x∗y ∈ Aφ. Let N ∈ Z+ and note that, by [23, Theorem 2.2], where Λl t(cid:0)φl(x)(cid:1)∗Λm NXl,m=0 t (cid:0)φm(y)(cid:1) = φn,N ](x∗y) := Xα∪β={1,...,n} α, β6N Λn t(cid:0)φn,N ](x∗y)(cid:1) 2NXn=0 φα(x∗; n, α)∆(n, α ∩ β)φβ(y; n, β). on h ⊗ E, (3.8) 12 Working as in the proof of Corollary 2.12 yields the inequality kφn,N ](x∗y)k 6 Cx∗Cy(Mx∗ + Mx∗My + My)n, and so, by (3.2), huε(f ), Λn t(cid:0)φn,N ](x∗y)(cid:1)vε(g)i 6 As φn,N ] = φn if n ∈ {0, 1, . . . , N}, it follows that K n g,t(Mx∗ + Mx∗My + My)n √n! Cx∗Cy kuε(f )kkvε(g)k. (3.9) hjt(x)uε(f ), jt(y)vε(g)i = lim N→∞ t (cid:0)φm(y)(cid:1)vε(g)i = lim N→∞ huε(f ), Λl NXl,m=0 t(cid:0)φl(x)(cid:1)∗Λm NXn=0 t(cid:0)φn(x∗y)(cid:1)vε(g)i huε(f ), Λn 2NXn=N +1 huε(f ), Λn = huε(f ), jt(x∗y)vε(g)i, + lim N→∞ t(cid:0)φn,N ](x∗y)(cid:1)vε(g)i since the final limit is zero by (3.9). Lemma 3.8. If Aφ is dense in A then there is at most one family of ∗-homomorphisms (¯t)t>0 from A to B(h ¯⊗ F) that satisfies (3.4). Proof. Suppose that j(1) and j(2) are two families of ∗-homomorphisms from A to B(h ⊗ F) that satisfy (3.4). Set kt := j(1) and note we have that t − j(2) t huε(f ), kt(x)vε(g)i =Z t 0 huε(f ), ks(cid:0)φ bf (s) bg(s) (x)(cid:1)vε(g)i ds for all u, v ∈ h, f , g ∈ L2(R+; k) and x ∈ Aφ. Iterating the above, and using the fact that kktk 6 2 for all t > 0, we obtain the inequality bf (t1) bg(t1) ◦ ··· ◦ φ huε(f ), kt(x)vε(g)i 6 2ZDn(t) kφ bf (tn) bg(tn) (x)k dtkuε(f )kkvε(g)k. However (2.9) now gives that huε(f ), kt(x)vε(g)i 6 2Cx(cid:0)Mxkbft)kkbgt)k(cid:1)n n! and the result follows by letting n → ∞. kuε(f )kkvε(g)k 13 Theorem 3.9. Let φ : A0 → A0 ⊗ B be a flow generator and suppose A0 contains its square If Aφ = A0 then, for roots: all t > 0, there exists a unital ∗-homomorphism for all non-negative x ∈ A0, the square root x1/2 lies in A0. ¯t : A → B(h ¯⊗ F) such that ¯t(x) = jt(x) on h ⊗ E for all x ∈ A0, where jt(x) is as defined in Theorem 3.5. Proof. Let x ∈ A0 and suppose first that x > 0. If y := (kxk1 − x)1/2, which lies in A0 by assumption, then Lemma 3.7 and Remark 3.6 imply that 0 6 kjt(y)θk2 = hθ, jt(y2)θi = kxkkθk2 − hθ, jt(x)θi for all θ ∈ h ⊗ E. If x is now an arbitrary element of A0, it follows that kjt(x)θk2 = hθ, jt(x∗x)θi 6 kx∗xkkθk2 = kxk2kθk2. Thus jt(x) extends to ¯t(x) ∈ B(h ¯⊗ F), which has norm at most kxk, and the map is a ∗-linear contraction, which itself extends to a ∗-linear contraction A0 → B(h ¯⊗ F); x 7→ ¯t(x) ¯t : A → B(h ¯⊗ F). Furthermore, if x, y ∈ A0 and θ, ζ ∈ h ⊗ E then, by Lemma 3.7, hθ, ¯t(x)¯t(y)ζi = h¯t(x∗)θ, ¯(y)ζi = hjt(x∗)θ, jt(y)ζi = hθ, jt(xy)ζi = hθ, ¯t(xy)ζi, so ¯t is multiplicative on A0. An approximation argument now gives that ¯t is multiplicative on the whole of A. Remark 3.10. If A is an AF algebra, i.e., the norm closure of an increasing sequence of finite- dimensional ∗-subalgebras, then its local algebra A0, the union of these subalgebras, contains its square roots, since every finite-dimensional C ∗ algebra is closed in A. Definition 3.11. The unital C ∗ algebra A has generators {ai : i ∈ I} if A is the smallest unital C ∗ algebra which contains {ai : i ∈ I}. These generators satisfy the relations {pk : k ∈ K} if each pk is a complex polynomial in the non-commuting indeterminate hXi, X ∗ i : i ∈ Ii and, for i : i ∈ I), obtained from pk by replacing Xi by ai and X ∗ all k ∈ K, the algebra element pk(ai, a∗ by a∗ Suppose A has generators {ai : i ∈ I} which satisfy the relations {pk : k ∈ K}. Then A is generated by isometries if {X ∗ i Xi − 1 : i ∈ I} ⊆ {pk : k ∈ K} and is generated by unitaries if {X ∗ i − 1 : i ∈ I} ⊆ {pk : k ∈ K}. The algebra A is universal if, given any unital C ∗ algebra B containing a set of elements {bi : i ∈ I} which satisfies the relations {pk : k ∈ K}, i.e., pk(bi, b∗ i : i ∈ I) = 0 for all k ∈ K, there exists a unique ∗-homomorphism π : A → B such that π(ai) = bi for all i ∈ I. i for all i ∈ I, is equal to 0. i i Xi − 1, XiX ∗ 14 Theorem 3.12. Let A be the universal C ∗ algebra generated by isometries {si : i ∈ I} which satisfy the relations {pk : k ∈ K}, and let A0 be the ∗-algebra generated by {si : i ∈ I}. If φ : A0 → A0 ⊗ B is a flow generator such that Aφ = A0 then, for all t > 0, there exists a unital ∗-homomorphism ¯t : A → B(h ¯⊗ F) such that ¯t(x) = jt(x) on h ⊗ E for all x ∈ A0, where jt(x) is as defined in Theorem 3.5. Proof. Remark 3.6 and Lemma 3.7 imply that jt(si) is isometric and that jt(s∗ i ) is contractive for all i ∈ I. Repeated application of (3.7) then shows that jt(x) is bounded for each x ∈ A0, and that jt extends to a unital ∗-homomorphism from A0 to B(h ¯⊗ F). Furthermore, the set {jt(si) : i ∈ I} satisfies the relations {pk : k ∈ K} so, by the universal nature of A, there exists a ∗-homomorphism π from A into B(h ¯⊗ F) such that π(si) = jt(si) for all i ∈ I and ¯t := π is as required. Corollary 3.13. The family (cid:0)¯t : A → B(h ¯⊗ F)(cid:1)t>0 a strong solution of the QSDE (1.2). constructed in Theorems 3.9 and 3.12 is Proof. Fix x ∈ Aφ and let Lt := Σ(cid:0)(¯t ⊗ ιB)(φ(x))(cid:1) for all t > 0, where Σ : B(h ¯⊗ F ¯⊗bk) → B(h ¯⊗bk ¯⊗ F) is the isomorphism that swaps the last two components of simple tensors. If f ∈ L2(R+; k) then (3.10) kLtu ⊗ bf (t) ⊗ ε(f )k 6 kφ(x)kkbf (t)kkuε(f )k, so if t 7→ Ltu⊗bf (t)⊗ε(f ) is strongly measurable then t 7→ Lt is quantum stochastically integrable [19, p.232] and ¯ satisfies the QSDE in the strong sense, since we already have from (3.4) that it is a weak solution. Now, Theorem 3.5 implies that for each x ∈ Aφ = A0 and θ ∈ h ⊗ E the map t 7→ ¯t(x)θ is continuous, hence so is t 7→ (¯t ⊗ ιB)(y ⊗ T )(θ ⊗ ξ) = ¯t(y)θ ⊗ T ξ for all y ∈ A0, T ∈ B(bk) and ξ ∈bk. As kLtk = kφ(x)k for all t > 0, it follows that t 7→ Lt and t are strongly continuous on h ¯⊗bk ¯⊗F. Hence t 7→ Lt(u⊗bf (t)⊗ ε(f )) is separably valued t 7→ L∗ and weakly measurable, so Pettis's theorem gives the result. Remark 3.14. Property (3.5) implies that the homomorphism ¯t given by Theorems 3.9 and 3.12 takes values in the matrix space A ⊗M B(F) [19]. Notation 3.15. For all t > 0, f , g ∈ L2(R+; k) and a ∈ A, let ¯t[f, g](a) :=(cid:0)1h ⊗ hε(ft))(cid:1)¯t(a)(cid:0)1h ⊗ ε(gt))i(cid:1). 15 Theorem 3.16. The family of ∗-homomorphisms (¯t)t>0 given by Theorems 3.9 and 3.12 forms a Feller cocycle [22, Section 2.4] for the shift semigroup on B(F): for all s, t > 0, f , g ∈ L2(R+; k) and a ∈ A, (i) ¯0[0, 0](a) = a, (ii) ¯t[f, g](a) ∈ A, (iii) (iv) ¯s+t[f, g] = ¯s[f, g] ◦ ¯t[f (· + s), g(· + s)]. t 7→ ¯t[f, g](a) is norm continuous and Consequently, setting Tt(a) := ¯t[0, 0](a) =(cid:0)1h ⊗ hε(0)(cid:1)¯t(a)(cid:0)1h ⊗ ε(0)i(cid:1) for all a ∈ A gives a strongly continuous semigroup T = (Tt)t>0 of completely positive contractions on A such that Tt(x) = exp(tφω ω)(x) for all x ∈ A0 and t > 0. In particular, Tt(1) = 1 for all t > 0 and A0 is a core for the generator of T . Proof. Properties (i) and (ii) are immediate consequences of (3.4) and (3.5) respectively. For (iii), note that if x ∈ A0 and f , g ∈ L2(R+; k) then Theorem 3.5 implies that t 7→ ¯t[f, g](x) =(cid:0)1h ⊗ hε(f )(cid:1)jt(x)(cid:0)1h ⊗ ε(g)i(cid:1) exp(cid:16)−Z ∞ hf (s), g(s)i ds(cid:17) is norm continuous; the general case follows by approximation. In order to establish (iv), fix s > 0 and continuous functions f , g ∈ L2(R+; k), and let t Jt := ¯s[f, g] ◦ ¯t[f (· + s), g(· + s)] for all t > 0. We will show that Jt = ¯s+t[f, g]. First note that for any x ∈ A0 and t > 0, the map F : [0, t] → A; r 7→ ¯r[f (· + s), g(· + s)](cid:0)φ is continuous, hence Bochner integrable, and so bf (r+s) bg(r+s) (x)(cid:1)hε(f[s+r,s+t)), ε(g[s+r,s+t))i xhε(f[s,s+t)), ε(g[s,s+t))i +Z t 0 F (r) dr ∈ A. By the adaptedness of ¯t(x) and (3.4), hu,(cid:16)xhε(f[s,s+t)), ε(g[s,s+t))i +Z t +Z t 0 huε(f (· + s)r)), jr(cid:0)φ 0 F (r) dr(cid:17)vi bg(r+s) (x)(cid:1)vε(g(· + s)r))hε(f (· + s)[r,t)), ε(g(· + s)[r,t))i dr bf (r+s) = hu, xvihε(f (· + s)t)), ε(g(· + s)t))i = huε(f (· + s)t)), jt(x)vε(g(· + s)t))i = hu, ¯t[f (· + s), g(· + s)](x)vi. 16 Consequently, hu, Jt(x)vi = hu, ¯s[f, g](x)vihε(f[s,s+t)), ε(g[s,s+t))i = hu, ¯s[f, g](x)vihε(f[s,s+t)), ε(g[s,s+t))i +Z t 0 hu, ¯s[f, g] ◦ ¯r[f (· + s), g(· + s)](cid:0)φ +Z t 0 hu, Jr(cid:0)φ bf (r+s) bf (r+s) bg(r+s) (x)(cid:1)vihε(f[s+r,s+t)), ε(g[s+r,s+t))i dr bg(r+s) (x)(cid:1)vihε(f[s+r,s+t)), ε(g[s+r,s+t))i dr. 0 huε(fs+t)), jr(cid:0)φ bg(r) (x)(cid:1)vε(gs+t))i dr 0 huε(fs+t)), jq+s(cid:0)φ huε(fs+t)), jr(cid:0)φ bf (r) bf (r) bg(r) (x)(cid:1)vε(gs+t))i dr bg(q+s) (x)(cid:1)vε(gs+t))i dq bf (q+s) = hu, ¯s[f, g](x)vihε(f[s,s+t)), ε(g[s,s+t))i On the other hand, by (3.4), s hu, ¯s+t[f, g](x)vi = hu, xvihε(fs+t)), ε(gs+t))i +Z s +Z s+t = huε(fs+t)), js(x)vε(gs+t))i +Z t +Z t 0 hu, ¯q+s[f, g](cid:0)φ hu, Kt(x)vi =Z t kKtk 6 2 exp(cid:16) 1 Now set Kt := Jt − ¯s+t[f, g], so that where G : r 7→ hε(f[s+r,s+t)), ε(g[s+r,s+t))i is continuous. As bf (r+s) bf (q+s) bg(q+s) (x)(cid:1)vihε(f[s+q,s+t)), ε(g[s+q,s+t)) dq. 0 hu, Kr(cid:0)φ 2(cid:0)kfk2 + kgk2(cid:1)(cid:17) bg(r+s) (x)(cid:1)vi G(r) dr, for all t > 0, iterating the above and estimating as in the proof of Lemma 3.8 shows that K ≡ 0. The density of A0 in A and of continuous functions in L2(R+; k) now gives (iv). That T is a semigroup follows from this cocycle property (iv): note that Ts+t = ¯s+t[0, 0] = ¯s[0, 0] ◦ ¯t[0, 0] = Ts ◦ Tt for all s, t > 0. Contractivity, complete positivity and strong continuity of T are immediate; the exponential identity holds because hu, Tt(x)vi = hu, xvi +Z t 0 hu, Ts(cid:0)φω ω(x)(cid:1)vi ds (3.11) for all u, v ∈ h, t > 0 and x ∈ A0, by (3.4). That A0 is a core for the generator of T follows from Lemma 2.14 and [5, Corollary 3.1.20]. 17 Remark 3.17. A ∗-homomorphic Feller cocycle as in Theorem 3.16 is called a quantum flow ; a strongly continuous semigroup (Tt)t>0 of completely positive contractions is known as a quantum dynamical semigroup, and the condition Tt(1) = 1 for all t > 0 means that the semigroup is conservative; conservative quantum dynamical semigroups are also known as quantum Markov semigroups. Hence Theorem 3.16 gives the existence of a quantum flow which dilates a quantum Markov semigroup on the C ∗ algebra A. Remark 3.18. By Theorem 3.16, the component φω ω = τ of the flow generator φ is closable, with τ being the generator of the quantum Markov semigroup T . However, closability of the bimodule map δ seems to be a much more delicate issue and remains an open question. Theorem 3.19. Consider the family of ∗-homomorphisms (¯t)t>0 constructed in Theorems 3.9 and 3.12. If Ac is a commutative ∗-subalgebra of A such that (i) φ(Ac ∩ A0) ⊆ Ac ⊗ B (ii) Ac ∩ A0 is dense in Ac and then the family {¯t(a) : t > 0, a ∈ Ac} is commutative, i.e., the commutator [¯s(a), ¯t(b)] = 0 for all s, t > 0 and a, b ∈ Ac. Proof. The result is immediate when s = t, so assume without loss of generality that s < t and let b ∈ Ac ∩ A0; if Kt(b) := huε(f ), [¯s(a), ¯t(b)]vε(g)i = 0, where u, v ∈ h, f , g ∈ L2(R+; k) and a ∈ Ac are arbitrary, then the result follows by (ii) and the continuity of ¯t. Write ¯t(b) = ¯s(b) +R t AZ t s Lr dΛr, where L = (Lr)r>0 is the process defined in (3.10) with x changed to b. It is straightforward, using adaptedness, to show that Lr dΛrB =Z t Σ(A ⊗ 1bk) Lr Σ(B ⊗ 1bk) dΛr for any A, B ∈ B(h ¯⊗ F[0,s)) ¯⊗ 1F[s,∞), where Σ is the swap isomorphism defined after (3.10). Since ¯ is a strong solution of the QSDE (1.2), by Corollary 3.13, it follows that s s Assumption (i) allows us to iterate this identity; noting also that Kt(b) =Z t s Kr(cid:0)φ bf (r) bg(r) (b)(cid:1) dr. Kr(c) 6 2kukkvkkε(f )kkε(g)kkakkck for all c ∈ Ac ∩ A0, one readily obtains the estimate Kt(b) 6 2kukkvkkε(f )kkε(g)kkakCb M n b 1 n!(cid:16)Z t s kbf (r)kkbg(r)k dr(cid:17)n , where Cb and Mb are constants associated to b through its membership of Aφ. Letting n → ∞ gives the result. 18 Remark 3.20. If A is commutative then conditions (i) and (ii) of Theorem 3.19 are satisfied automatically when Ac = A, so Theorems 3.9 and 3.12 produce classical Markov semigroups in this case. However, Theorem 3.19 also allows for the possibility of dealing with different commutative subalgebras that do not commute with one another, a necessary feature of quantum dynamics. 4 Random walks on groups Definition 4.1. Let A = C0(G) ⊕ C1 ⊆ B(cid:0)ℓ2(G)(cid:1), where G is a discrete group and x ∈ C0(G) acts on ℓ2(G) by multiplication, and let A0 = lin{1, eg : g ∈ G}, where eg(h) := 1g=h for all h ∈ G. That is, A is the unitisation of the C ∗ algebra of functions on G which vanish at infinity and A0 is the dense unital subalgebra generated by the functions with finite support; as positivity in the C ∗-algebraic sense corresponds here to the pointwise positivity of functions, A0 contains its square roots. Let H be a non-empty finite subset of G \ {e} and let the Hilbert space k have orthonormal basis {fh : h ∈ H}; the maps correspond to the permitted moves in the random walk constructed on G. λh : G → G; g 7→ hg (h ∈ H) Lemma 4.2. Given a transition function the map t : H × G → C; (h, g) 7→ th(g), φ : A0 → A0 ⊗ B; x 7→" Ph∈H th2(x ◦ λh − x) Ph∈H th(x ◦ λh − x) ⊗ fhi Ph∈H(x ◦ λh − x) ⊗ fhihfh# Ph∈H th(x ◦ λh − x) ⊗ hfh is a flow generator such that φ(eg) = eg ⊗ me(g) +Xh∈H eh−1g ⊗ mh(h−1g) for all g ∈ G, where me(g) :=" −Ph∈H th(g)2 −Ph∈H th(g)hfh −1k −Ph∈H th(g)fhi Hence # and mh(g) :=" th(g)2 th(g)fhi th(g)hfh fhihfh# . φn(eg) = Xh1∈H∪{e} ··· Xhn∈H∪{e} for all n ∈ N and g ∈ G. eh−1 n ···h−1 1 g ⊗ mhn(h−1 n ··· h−1 1 g) ⊗ ··· ⊗ mh1(h−1 1 g) 19 Proof. The first claim is readily verified with the aid of Lemma 2.2; the second is immediate. Theorem 4.3. Let A be as in Definition 4.1 and φ as in Lemma 4.2. If the transition function t is chosen such that Aφ = A0 then there exists an adapted family of unital ∗-homomorphisms which forms a Feller cocycle in the sense of Theorem 3.16 and satisfies the quantum stochastic differential equation (1.2) in the strong sense on A0 for all t > 0. Setting (cid:0)¯t : A → B(h ¯⊗ F)(cid:1)t>0 gives a classical Markov semigroup T on A whose generator is the closure of Tt(a) :=(cid:0)1h ⊗ hε(0)(cid:1)¯t(a)(cid:0)1h ⊗ ε(0)i(cid:1) τ : A0 → A0; x 7→ Xh∈H th2(x ◦ λh − x). for all a ∈ A and t > 0 Proof. This follows from Theorems 3.9 and 3.19 together with Lemma 4.2. Remark 4.4. Given g ∈ G, let A :=(cid:2)B 1k(cid:3) ∈ B(C ⊕ k; k), where B :=Ph∈H th(g)fhi. Then me(g) = −A∗A and kme(g)k = kAA∗k = kBB∗ + 1kk = kB∗Bk + 1 = 1 +Xh∈H th(g)2. It may be shown similarly that kmh(g)k = 1 + th(g)2 for all g ∈ G and h ∈ H, so if 1 g) : h1, . . . , hn ∈ H ∪ {e}, h ∈ H(cid:9) < ∞ n ··· h−1 Mg := lim n→∞ sup(cid:8)th(h−1 g )n kφn(eg)k 6 (1 + H + 2HM 2 for all n ∈ Z+, then (4.1) where H denotes the cardinality of H. Hence Aφ = A0 if (4.1) holds for all g ∈ G. Remark 4.5. If t is bounded then clearly (4.1) holds for all g ∈ G. In this case, there exist bounded operators L ∈ B(h; h ¯⊗ k), S ∈ B(h ¯⊗ k) and F ∈ B(h ¯⊗bk) such that Sh ⊗ fhihfh and F ="− 1 2 L∗L SL L = Xh∈H th ⊗ fhi, S = Xh∈H −L∗ S − 1h⊗k# , where th acts by multiplication and Sh is the unitary operator on ℓ2(G) such that eg 7→ ehg. It follows from [20, Theorems 7.1 and 7.5] that the Hudson -- Parthasarathy QSDE U0 = Ih⊗F , dUt = (F ⊗ 1F )Σ(Ut ⊗ Ibk) dΛt, where Σ is the swap isomorphism defined after (3.10), has a unique solution which is a unitary cocycle. Furthermore, by [20, Theorem 7.4], setting kt(a) := U ∗ t (a ⊗ 1F )Ut for all a ∈ B(h) and t > 0 20 defines a quantum flow k with generator ϕ : B(h) → B(h ¯⊗bk); a 7→ (a ⊗ 1bk)F + F ∗(a ⊗ 1bk) + F ∗∆(a ⊗ 1bk)F. A short calculation shows that ϕ is of the form covered by Lemma 2.5, with π(a) = S∗(a ⊗ 1k)S, δ(a) = −La + π(a)L and τ (a) = − 1 2{L∗L, a} + L∗π(a)L for all a ∈ B(h). It follows that ϕA0 = φ, where φ is the flow generator of Lemma 4.2, and so the cocycle ¯ given by Theorem 4.3 is the restriction of k to A. However, this construction by conjugation does not give the Feller property, that A is preserved by k. Example 4.6. If G = (Z, +), H = {±1} and the transition function t is bounded, with t+1(g) = 0 for all g < 0 and t−1(g) = 0 for all g 6 0, then the Markov semigroup T given by Theorem 4.3 corresponds to the classical birth-death process with birth and dates rates t+12 and t−12, respectively. The cocycle constructed here is Feller, as it acts on A = C0(Z) ⊕ C1, in contrast to [26, Example 3.3], where the cocycle acts on the whole of ℓ∞(Z). Remark 4.7. If G = (Z, +), H = {+1} and t+1 : g 7→ 2g then Mg = 2g and the condition (4.1) holds for all g ∈ G. Thus Theorem 4.3 applies to examples where the transition function t is unbounded. 5 The symmetric quantum exclusion process This section was inspired by Rebolledo's treatment of the quantum exclusion process: see [27, Examples 2.4.3 and 4.1.3]. Definition 5.1. Let I be a non-empty set. The CAR algebra is the unital C ∗ algebra A with generators {bi : i ∈ I}, subject to the anti-commutation relations and {bi, b∗ (5.1) for all i, j ∈ I. It follows from (5.1) that the bi are nonzero partial isometries for all i ∈ I. {bi, bj} = 0 j} = 1i=j As is well known [6, Proposition 5.2.2], A is represented faithfully and irreducibly on F−(cid:0)ℓ2(I)(cid:1), the Fermionic Fock space over ℓ2(I); in other words, we may (and do) suppose that A ⊆ B(h), where h := F−(cid:0)ℓ2(I)(cid:1), and the algebra identity 1 = 1h. Remark 5.2. The elements of I may be taken to correspond to sites at which Fermionic particles may exist, with the operators bi and b∗ i representing the annihilation and creation, respectively, of a particle at site i. Notation 5.3. Let A0 be the unital algebra generated by {bi, b∗ a norm-dense unital ∗-subalgebra of A. i : i ∈ I}; by definition, this is 21 Lemma 5.4. For each x ∈ A0 there exists a finite subset J ⊆ I such that x lies in the finite- dimensional ∗-subalgebra j1 ··· b∗ jq bi1 ··· bip : 0 6 p, q 6 J, {i1, . . . , ip} ∈ J (p), {j1, . . . , jq} ∈ J (q)} ⊆ A0, AJ := lin(cid:8)b∗ where J (p) denote the set of subsets of J with cardinality p et cetera. Consequently, A is an AF algebra and A0 contains its square roots. Proof. By employing the anti-commutation relations (5.1), any finite product of terms from the generating set {bi, b∗ i : i ∈ I} may be reduced to a linear combination of words of the form b∗ j1 ··· b∗ jq bi1 ··· bip, (5.2) where i1, . . . , ip are distinct elements of I, as are j1, . . . , jq, and p, q ∈ Z+, with an empty product equal to 1. As every element of A0 is a finite linear combination of such terms, the first claim follows. The second claim holds by Remark 3.10. Definition 5.5. Let {αi,j : i, j ∈ I} ⊆ C be a fixed collection of amplitudes. We may view (I,{αi,j : i, j ∈ I}) as a weighted directed graph, where I is the set of vertices, an edge exists from i to j if αi,j 6= 0 and αi,j is a complex weight on the edge from vertex i to vertex j, which may differ from the weight αj,i from j to i. For all i ∈ I, let supp(i) := {j ∈ I : αi,j 6= 0} and supp+(i) := supp(i) ∪ {i}. Thus supp(i) is the set of sites with which site i interacts and supp(i) is the valency of the vertex i. We require that the valencies are finite: supp(i) < ∞ for all i ∈ I. (5.3) The transport of a particle from site i to site j with amplitude αi,j is described by the operator ti,j := αi,j b∗ j bi. Definition 5.6. Let {ηi : i ∈ I} ⊆ R be fixed. The total energy in the system is given by h :=Xi∈I ηi b∗ i bi, If the set {i ∈ I : ηi 6= 0} is infinite then where ηi gives the energy of a particle at site i. the proper interpretation of h involves issues of convergence; below it will only appear in a commutator with elements of A0, which is sufficient to give a well-defined quantity. Lemma 5.7. Let τi,j(x) := t∗ i,j[ti,j, x] + [x, t∗ i bj[b∗ j bi, x] + [x, b∗ i bj]b∗ i,j]ti,j = αi,j2(cid:0)b∗ 22 j bi(cid:1) defines a ∗-linear map τ : A0 → A0. Proof. Let x ∈ A0 and note that x ∈ AJ for some finite set J ⊆ I, by Lemma 5.4. Furthermore, [b∗ j bi, x] = b∗ j{bi, x} − {b∗ j , x}bi = 0 whenever i 6∈ J and j 6∈ J, so where [h, x] =Xi∈J ηi[b∗ i bi, x] ∈ AJ and τ (x) = i[h, x] − 1 2 Xi,j∈J + τi,j(x) ∈ AJ +, J + := [k∈J supp+(k). (5.6) Hence τ (AJ ) ⊆ AJ + and, as (5.3) implies that J + is finite, it follows that A0 is invariant under τ . The ∗-linearity of τ is immediately verified. Lemma 5.8. Let for all i, j ∈ I and x ∈ A, and let k be a Hilbert space with orthonormal basis {fi,j : i, j ∈ I}. Setting δi,j(x) := [ti,j, x] = αi,j(b∗ j bix − xb∗ j bi) δ(x) := Xi,j∈I δi,j(x) ⊗ fi,ji (5.4) (5.5) (5.7) (5.8) (5.9) for all i, j ∈ I and x ∈ A, and let for all x ∈ A0. Setting [h, x] :=Xi∈I τ (x) := i[h, x] − 1 ηi[b∗ i bi, x] 2 Xi,j∈I τi,j(x) for all x ∈ A0 defines a linear map δ : A0 → A0 ⊗ ki such that δ(xy) = δ(x)y + (x ⊗ 1k)δ(y) δ†(x)δ(y) = τ (xy) − τ (x)y − xτ (y) and for all x, y ∈ A0, where τ is as defined in Lemma 5.7. Proof. The series in (5.7) contains only finitely many terms, since if x ∈ AJ then δi,j(x) = 0 when {i, j} 6⊆ J +. Hence δ is well defined, and (5.8) holds because each δi,j is a derivation. A short calculation shows that τi,j(xy) − τi,j(x)y − xτi,j(y) = −2δ† i,j(x)δi,j(y) (5.10) 23 for all x, y ∈ A. Since x 7→ [b∗ i bi, x] is a derivation for all i ∈ I, it follows from (5.10) that τ (xy) − τ (x)y − xτ (y) = Xi,j∈I δ† i,j(x)δi,j(y) = δ†(x)δ(y) for all x, y ∈ A0. Lemma 5.9. The map φ : A0 → A0 ⊗ B; x 7→"τ (x) δ(x) δ†(x) 0 # , where τ , δ and δ† are as defined in Lemmas 5.7 and 5.8, is a flow generator. If the amplitudes satisfy the symmetry condition αi,j = αj,i for all i, j ∈ I (5.11) (5.12) then, for all n ∈ N and i0 ∈ I, φn(bi0) = Xi1∈supp+(i0) ··· Xin∈supp+(in−1) bin ⊗ Bin−1,in ⊗ ··· ⊗ Bi0,i1, (5.13) where and for all i, j ∈ I. Bi,j := 1j=iλiωihω + ωihαi,jfi,j − αj,ifj,iihω λi := −iηi − 1 2 Xj∈supp(i) αj,i2 Proof. The first claim is an immediate consequence of Lemmas 5.7, 5.8 and 2.2. If i, j, k ∈ I then a short calculation shows that αi,i2bi αj,i2b∗ j bjbi αi,k2bkb∗ kbi 0 (j = i, k = i), (j 6= i, k = i), (j = i, k 6= i), (j 6= i, k 6= i). τj,k(bi) =  Since [h, bi] =Xj∈I ηj[b∗ j bj, bi] = ηi[b∗ i bi, bi] = −ηibi, the symmetry condition (5.12) implies that τ (bi) = λibi for all i ∈ I. 24 Furthermore, if i, j, k ∈ I then δj,k(bi) = αj,k(b∗ kbjbi − bib∗ kbj) = −αj,k{b∗ k, bi}bj = −1k=iαj,i bj and thus and Hence δ† j,k(bi) = αj,k(bib∗ j bkbi) = αj,k{bi, b∗ j bk − b∗ δj,k(bi) ⊗ fj,ki = − Xj∈supp(i) j,k(bi) ⊗ hfj,k = Xk∈supp(i) δ† δ(bi) = Xj,k∈I δ†(bi) = Xj,k∈I j}bk = 1j=iαi,k bk; αj,i bj ⊗ fj,ii αi,k bk ⊗ hfi,k. φ(bi) = λibi ⊗ ωihω − Xj∈supp(i) αj,ibj ⊗ fj,iihω + Xk∈supp(i) = Xj∈supp+(i) bj ⊗(cid:0)1j=iλiωihω + ωihαi,jfi,j − αj,ifj,iihω(cid:1) αi,kbk ⊗ ωihfi,k and the identity (5.13) follows. Theorem 5.10. Let A be the CAR algebra and let φ be defined as in Lemma 5.9. If the amplitudes {αi,j} and energies {ηi} are chosen so that Aφ = A0 then there exists an adapted which forms a Feller cocycle in the sense of Theorem 3.16 and satisfies the quantum stochastic differential equation (1.2) in the strong sense on A0 for all t > 0. Setting family of unital ∗-homomorphisms (cid:0)jt : A → B(h ¯⊗ F)(cid:1)t>0 Tt(a) :=(cid:0)1h ⊗ hε(0)(cid:1)jt(a)(cid:0)1h ⊗ ε(0)i(cid:1) 2 Xi,j∈I τ : A0 → A0; x 7→ iXi∈I i bi, x] − 1 gives a quantum Markov semigroup T on A whose generator is the closure of j bi, x] + [x, b∗ i bj[b∗ ηi[b∗ αi,j2(cid:0)b∗ for all a ∈ A and t > 0 i bj]b∗ j bi(cid:1). Proof. This is an immediate consequence of Theorem 3.9, Theorem 3.16 and Lemma 5.9. Example 5.11. Suppose that the amplitudes satisfy the symmetry condition (5.12), and further that there are uniform bounds on the amplitudes, valencies and energies: M := sup i,j∈I αi,j < ∞, V := sup i∈I supp(i) < ∞ and H := sup i∈I ηi < ∞. (5.14) It follows that λi 6 ηi + 1 2 V M 2 and kBi,jk 6 λi + 2M 6 H + 1 2 V M 2 + 2M 25 for all i, j ∈ I. Hence, for all n ∈ Z+, kφn(bi)k 6 (V + 1)n(cid:0)H + 1 2 V M 2 + 2M(cid:1)n and so Aφ = A0, by Corollary 2.12. Hence there is a flow on A for this generator. Example 5.12. We can lift the boundedness assumptions in Example 5.11 by taking I to be a disjoint union of subsets, I = Gk∈K Ik, such that there is no transport between any of these subsets, i.e., αi,j 6= 0 only if there is some k ∈ K such that i, j ∈ Ik. Assume the symmetry condition (5.12) once again. Suppose that in each Ik the conditions of (5.14) are satisfied, but with respect to constants Mk, Vk and Hk that depend on k. Then, if i ∈ Ik, we get the estimate kφn(bi)k 6 (Vk + 1)n(cid:0)Hk + 1 2 VkM 2 k + 2Mk(cid:1)n and so Aφ = A0 once more, but now it is possible that M = ∞ et cetera. Example 5.13. To create an example where the graph associated to I has only one component, but where we do not assume M < ∞ as in Example 5.11, assume once again that I is decomposed into a disjoint union: I = Gk∈Z+ Ik with Ik < ∞ for all k ∈ Z+. This time assume, as well as the symmetry condition (5.12), that αi,j = 0 unless there is some k ∈ Z+ such that i ∈ Ik and j ∈ Ik+1, or j ∈ Ik and i ∈ Ik+1, so that there is transport only between neighbouring levels in I. Set ak = sup{αi,j : i ∈ Ik, j ∈ Ik+1} for all k ∈ Z+, and furthermore assume that the energies are bounded, i.e., H < ∞. Now if k ∈ N and i ∈ Ik then Xj∈supp+(i) kBi,jk 6 kBi,ik + Xj∈Ik−1 kBi,jk + Xj∈Ik+1 kBi,jk 6 λi + 2Ik−1ak−1 + 2Ik+1ak, with a similar estimate holding if i ∈ I0. Furthermore, k−1 + 1 λi 6 H + 1 2Ik−1a2 2Ik+1a2 k. 26 As in Example 5.11, if it can be shown that Xj∈supp+(i) kBi,jk 6 C for some constant C that does not depend on i, it follows that kφn(bi)k 6 C n for each n ∈ Z+ and i ∈ I, and so Aφ = A0 once more. Here, the previous working shows this will hold if there are constants a > 0, b > 0 and p > 1 such that ak 6 a (k + 2)p and Ik 6 b(k + 1)p for all k ∈ Z+. It is clear that this can yield an example where M = ∞, i.e., there is no upper bound on the valencies. 6 Flows on universal C ∗ algebras 6.1 The non-commutative torus Definition 6.1. Let λ ∈ T, the set of complex numbers with unit modulus. The non- commutative torus is the universal C ∗ algebra A generated by unitaries U and V which satisfy the relation U V = λV U. Let A0 denote the dense ∗-subalgebra of A generated by U and V . There is a faithful trace tr on A such that τ (U mV n) = 1m=n=0 for all m, n ∈ Z; the proof of this in [9, pp.166 -- 168] is valid for all λ. Consequently {U mV n : m, n ∈ Z} is a basis for A0. Lemma 6.2. Let h := ℓ2(Z2), let (Ucu)m,n = um+1,n and (Vcu)m,n = λmum,n+1 for all u ∈ h and m, n ∈ Z, and let Ac ⊆ B(h) be the C ∗ algebra generated by Uc and Vc. There is a C ∗ isomorphism from A to Ac such that U 7→ Uc and V 7→ Vc. Moreover, under this map the trace tr corresponds to the vector state given by e ∈ h such that em,n = 1m=n=0 for all m, n ∈ Z. Proof. Unitarity of Uc and Vc is immediately verified, as is the identity UcVc = λVcUc, so the universality of A gives a surjective ∗-homomorphism from A to Ac. Injectivity is a consequence of the final observation, that tr corresponds to the vector state given by e. From now on we will identify A and Ac. Definition 6.3. For each (µ, ν) ∈ T2, let πµ,ν be the automorphism of A such that πµ,ν(U mV n) = µmνnU mV n for all m, n ∈ Z; the existence of πµ,ν is an immediate consequence of universality. 27 The proofs of the next two lemmas are a matter of routine algebraic computation. Lemma 6.4. For all a, b ∈ Z, define maps aδ : A0 → A0 and δb : A0 → A0 by linear extension of the identities aδ(U mV n) = mU a+mV n and δb(U mV n) = nλ−bmU mV b+n for all m, n ∈ Z. Then aδ is a π1,λa -derivation and δb is a πλ−b,1-derivation; moreover, their adjoints are such that aδ†(U mV n) = −mλanU −a+mV n and δ† b(U mV n) = −nU mV −b+n for all m, n ∈ Z. Remark 6.5. The sufficient condition in Lemma 6.4 is also necessary. It is easy to show that if aδ is a πµ,ν-derivation then µ = 1 and ν = λa; similarly, if δb is a πµ,ν-derivation then µ = λ−b and ν = 1. Lemma 6.6. With A0 as in Definition 6.1, and aδ and δb as in Lemma 6.4, fix c1, c2 ∈ C and let φ : A0 → A0 ⊗ B(C3); x 7→ c1 aδ†(x) τ (x) c1 aδ(x) π1,λa(x) − x c2 δb(x) 0 c2 δ† b (x) 0 πλ−b,1(x) − x  , where the map τ : A0 → A0; U mV n 7→ − 1 2(cid:0)c12m2 + c22n2(cid:1)U mV n. Then τ is ∗-linear and φ is a flow generator. Lemma 6.7. Let φ be as in Lemma 6.6. If a = b = 0 then Aφ = A0; conversely, if a 6= 0 and c1 6= 0 then U /∈ Aφ, and if b 6= 0 and c2 6= 0 then V /∈ Aφ. Proof. When a = b = 0, note that φ(U ) = U ⊗ mU and φ(V ) = V ⊗ mV , where − 1 2c22 0 −c2 0 0 − 1 2c12 −c1 0 0 c1 and 0 0 0 0 0 c2 0 0 mU :=  mV :=  . Hence φn(U ) = U ⊗ m⊗n Corollary 2.12. U and φn(V ) = V ⊗ m⊗n V , so U , V ∈ Aφ, as claimed, and Aφ = A0, by If a > 0 then, by induction, one gets that aδn(U ) = n−1Yi=0(cid:0)ia + 1(cid:1)U an+1 for all n ∈ N. Let e = [1 0 0]T and f = [0 1 0]T be unit vectors in C3, and note that (cid:0)1h ⊗ hf ⊗ ··· ⊗ hf(cid:1)φn(x)(cid:0)1h ⊗ ei ⊗ ··· ⊗ ei(cid:1) = cn 1 aδn(x) 28 for all x ∈ A0, so kφn(U )k > c1n n−1Yi=0(cid:0)ia + 1(cid:1) > c1nn!. If a < 0 then, by considering aδ† instead, we see that kφn(U )k > k(cid:0)1h ⊗ he ⊗ ··· ⊗ he(cid:1)φn(U )(cid:0)1h ⊗ fi ⊗ ··· ⊗ fi(cid:1)k > c1nn!. A similar proof shows that V /∈ Aφ when b 6= 0. Remark 6.8. The lower bounds obtained in Lemma 6.7 when a 6= 0 or b 6= 0 show that our techniques do not apply in these cases. The same problem arises if one attempts to use the results of [11] instead. The following theorem gives the existence of a quantum flow used by Goswami, Sahu and Sinha [14, Theorem 2.1(i)]. Theorem 6.9. Let A be as in Definition 6.1 and φ as in Lemma 6.6 for a = b = 0. There exists an adapted family j of unital ∗-homomorphisms from A to B(h ¯⊗ F) such that huε(f ), jt(x)vε(g)i = huε(f ), (xv)ε(g)i +Z t for all u, v ∈ h, f , g ∈ L2(R+; k), x ∈ A0 and t > 0. Proof. This follows from Theorem 3.12, Lemma 6.6 and Lemma 6.7. 0 huε(f ), js(cid:0)φ bf (s) bg(s) (x)(cid:1)vε(g)i ds Remark 6.10. The cocycle constructed in Theorem 6.9 is essentially a classical object: as noted in [7, Theorem 2.1], when c1 = c2 = i one may take jt(x) := β(cid:0)exp(2πiB1 t ), exp(2πiB2 t )(cid:1)(x) for all x ∈ A and t > 0, where β : T2 → Aut(A) is the natural action of the 2-torus T2 on A, so that β(z, w)(U mV n) = zmwnU mV n for all (z, w) ∈ T2, and the Fock space F is identified in the usual manner with the L2 space of the two-dimensional classical Brownian motion (B1, B2). The existence of flows where the generator has non-zero gauge part may also be established. Lemma 6.11. Fix (µ, ν) ∈ T2 with µ 6= 1. Let A0 be as in Definition 6.1 and πµ,ν as in Definition 6.3. There exists a flow generator φ : A0 → A0 ⊗ B(C2); x 7→"τ (x) δ(x) πµ,ν(x) − x# , −µδ(x) 29 φ(U ) = U ⊗ 1 − µ −µ µ 1 µ − 1 and φ(V ) = V ⊗ 1 − ν 1 − µ 1 − µ −µ µ 1 µ − 1 , where the πµ,ν-derivation δ : A0 → A0; U mV n 7→ 1 − µmνn 1 − µ U mV n (6.1) is such that δ† = −µδ, and the map τ := µ 1 − µ δ : A0 → A0; U mV n 7→ µ(1 − µmνn) (1 − µ)2 U mV n. Furthermore, U , V ∈ Aφ and so Aφ = A0. Proof. Using the basis {U mV n : m, n ∈ Z}, one can readily verify that δ is a πµ,ν-derivation such that δ† = −µδ, and hence φ is a flow generator. Since the fact that {U, V } ⊆ Aφ follows as in the proof of Lemma 6.7. Remark 6.12. It is curious that for φ as in Lemma 6.11 we have τ = µ(1 − µ)−1δ, and so τ is first rather than second order. Whether or not φ or, equivalently, δ is bounded is an open question; our existence result obviates the need to determine this. Theorem 6.13. Let A be as in Definition 6.1 and φ as in Lemma 6.11. There exists an adapted family j of unital ∗-homomorphisms from A to B(h ¯⊗ F) such that huε(f ), jt(x)vε(g)i = huε(f ), (xv)ε(g)i +Z t for all u, v ∈ h, f , g ∈ L2(R+; k), x ∈ A0 and t > 0. As noted by Hudson and Robinson [17], the following result makes clear why in Theorem 6.9 it is necessary to use two dimensions of noise to obtain a process whose flow generator includes both of the derivations c1 0δ and c2 δ0: the linear combination δ = c1 0δ + c2 δ0 can appear on the right-hand side of (2.3) only when the coefficients c1 and c2 satisfy a particular algebraic relation. bg(s) (x)(cid:1)vε(g)i ds 0 huε(f ), js(cid:0)φ bf (s) Proposition 6.14. Let 0δ and δ0 be as in Lemma 6.4, and let δ = c1 0δ + c2 δ0 for complex numbers c1 and c2. A necessary and sufficient condition for the existence of a linear map τ : A0 → A such that τ (xy) − τ (x)y − xτ (y) = δ†(x)δ(y) for all x, y ∈ A0 is the equality c1c2 = c1c2. Proof. This may be established by adapting slightly the proof of [28, Theorem 2.2]. 30 6.2 The universal rotation algebra To avoid the issue of Proposition 6.14, Hudson and Robinson work with the universal rotation algebra. Definition 6.15. Let A be the universal rotation algebra [2]: this is the universal C ∗ algebra with unitary generators U , V and Z satisfying the relations U V = ZV U, U Z = ZU and V Z = ZV. It may be viewed as the group C ∗ algebra corresponding to the discrete Heisenberg group Γ := hu, v, z uv = zvu, uz = zu, vz = zvi; from this perspective, its universal nature is immediately apparent. Letting A0 denote the ∗-subalgebra generated by U , V and Z, there are skew-adjoint derivations δ1 : A0 → A0; U mV nZ p 7→ mU mV nZ p and δ2 : A0 → A0; U mV nZ p 7→ nU mV nZ p for all m, n, p ∈ Z. Remark 6.16. For a concrete version of the universal rotation algebra, let h := ℓ2(Z3) and define operators Uc, Vc and Zc by setting (Ucu)m,n,p = um+1,n,p, (Vcu)m,n,p = um,n+1,m+p and (Zcu)m,n,p = um,n,p+1 for all u ∈ h and m, n, p ∈ Z. It is readily verified that Uc, Vc and Zc are unitary and satisfy the commutation relations as claimed; let Ac be the C ∗ algebra generated by these operators. Universality gives a surjective ∗-homomorphism from A to Ac such that U 7→ Uc, V 7→ Vc and Z 7→ Zc, and injectivity may be established in the same manner as for the non-commutative torus: there is a faithful state τ on A such that τ (U mV nZ p) = 1m=n=p=0 and this corresponds to the vector state given by e ∈ h such that em,n,p = 1m=n=p=0. Lemma 6.17. With A0, δ1 and δ2 as in Definition 6.15, fix c1, c2 ∈ C, let δ = c1δ1 + c2δ2 and define the Bellissard map τ : A0 → A0; U mV nZ p 7→ −(cid:0) 1 Then τ is ∗-linear and such that 2c12m2 + 1 2c22n2 + c1c2mn + (c1c2 − c1c2)p(cid:1)U mV nZ p, so the map τ (xy) − τ (x)y − xτ (y) = δ†(x)δ(y) for all x, y ∈ A0, φ : A0 → A0 ⊗ B(C2); x 7→"τ (x) δ(x) δ†(x) 0 # is a flow generator. Furthermore, U , V , Z ∈ Aφ and Aφ = A0. 31 Proof. The algebraic statements are readily verified, and a short calculation shows that φ(U ) = U ⊗ mU , φ(V ) = V ⊗ mV and φ(Z) = Z ⊗ mZ, where mU ="− 1 0 # , 2c12 −c1 c1 mV ="− 1 0 # 2c22 −c2 c2 and 0# . mZ ="c1c2 − c1c2 0 0 Hence φn(U ) = U ⊗ m⊗n U , φn(V ) = V ⊗ m⊗n V and φn(Z) = Z ⊗ m⊗n Z for all n ∈ Z+, so U , V , Z ∈ Aφ and Aφ = A0, by Corollary 2.12. The following theorem is an algebraic version of the result presented by Hudson and Robinson in [17, Section 4]. Theorem 6.18. Let A be as in Definition 6.15 and φ as in Lemma 6.17. There exists an adapted family j of unital ∗-homomorphisms from A to B(h ¯⊗ F) such that huε(f ), jt(x)vε(g)i = huε(f ), (xv)ε(g)i +Z t for all u, v ∈ h, f , g ∈ L2(R+; k), x ∈ A0 and t > 0. 0 huε(f ), js(cid:0)φ bf (s) bg(s) (x)(cid:1)vε(g)i ds (cid:3) Acknowledgements ACRB thanks Professors Kalyan Sinha and Tirthankar Bhattacharyya for hospitality at the Indian Institute of Science, Bangalore, and in Munnar, Kerala; part of this work was completed during a visit to India supported by the UKIERI research network Quantum Probability, Non- commutative Geometry and Quantum Information. Thanks are also due to Professor Martin Lindsay for helpful discussions. Funding from Lancaster University's Research Support Office and Faculty of Science and Technology is gratefully acknowledged. SJW thanks Professor Rolando Rebolledo for a very pleasant visit to Santiago in 2006 where thoughts about the quantum exclusion process were first encouraged. Both authors are indebted to the two anonymous referees and the associate editor for their constructive comments on an earlier draft of this paper. 7 References [1] L. Accardi and S. V. Kozyrev, On the structure of Markov flows, Chaos Solitons Fractals 12 (2001), no. 14 -- 15, 2639 -- 2655. 32 [2] J. Anderson and W. Paschke, The rotation algebra, Houston J. Math. 15 (1989), no. 1, 1 -- 26. [3] S. Attal, Classical and quantum stochastic calculus, in: Quantum Probability Commu- nications X, R. L. Hudson and J. M. Lindsay (eds.), World Scientific, Singapore, 1998, 1 -- 52. [4] P. Biane, Ito's stochastic calculus and Heisenberg commutation relations, Stochastic Pro- cess. Appl. 120 (2010), no. 5, 698 -- 720. [5] O. Bratteli and D. W. Robinson, Operator algebras and quantum statistical mechan- ics 1. C ∗- and W ∗-algebras. Symmetry groups. Decomposition of states, second printing of the second edition, Springer, Berlin, 2002. [6] O. Bratteli and D. W. Robinson, Operator algebras and quantum statistical mechan- ics 2. Equilibrium states. Models in quantum statistical mechanics, second printing of the second edition, Springer, Berlin, 2002. [7] P. S. Chakraborty, D. Goswami and K. B. Sinha, Probability and geometry on some noncommutative manifolds, J. Operator Theory 49 (2003), no. 1, 185 -- 201. [8] P. Beazley Cohen, T. M. W. Eyre and R. L. Hudson, Higher order Ito product formula and generators of evolutions and flows, Internat. J. Theoret. Phys. 34 (1995), no. 8, 1481 -- 1486. [9] K. R. Davidson, C ∗-algebras by example, Fields Institute Monographs 6, American Math- ematical Society, Providence, Rhode Island, 1996. [10] F. Fagnola, Quantum Markov semigroups and quantum flows, Proyecciones 18 (1999), no. 3, 144pp. [11] F. Fagnola and K. B. Sinha, Quantum flows with unbounded structure maps and finite degrees of freedom, J. London Math. Soc. (2) 48 (1993), no. 3, 537 -- 551. [12] J. C. Garc´ıa, R. Quezada and L. Pantale´on-Mart´ınez, Sufficient condition for the existence of invariant states for the asymmetric exclusion QMS, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 14 (2011), no. 2, 337-343. [13] L. Pantale´on-Mart´ınez and R. Quezada, The asymmetric exclusion quantum Markov semigroup, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 12 (2009), no. 3, 367-385. [14] D. Goswami, L. Sahu and K. B. Sinha, Dilation of a class of quantum dynamical semigroups with unbounded generators on UHF algebras, Ann. Inst. H. Poincar´e Probab. Statist. 41 (2005), no. 3, 505 -- 522. [15] R. L. Hudson and K. R. Parthasarathy, Quantum Ito's formula and stochastic evol- utions, Comm. Math. Phys. 93 (1984), no. 3, 301 -- 323. 33 [16] R. L. Hudson and S. Pulmannov´a, Chaotic expansion of elements of the universal enveloping algebra of a Lie algebra associated with a quantum stochastic calculus, Proc. London Math. Soc. (3) 77 (1998), no. 2, 462 -- 480. [17] R. L. Hudson and P. Robinson, Quantum diffusions and the noncommutative torus, Lett. Math. Phys. 15 (1988), no. 1, 47 -- 53. [18] T. M. Liggett, Stochastic interacting systems: contact, voter and exclusion processes, Springer, Berlin, 1999. [19] J. M. Lindsay, Quantum stochastic analysis -- an introduction, in: Quantum Independ- ent Increment Processes I, M. Schurmann & U. Franz (eds.), Lecture Notes in Mathemat- ics 1865, Springer, Berlin, 2005, 181 -- 271. [20] J. M. Lindsay and S. J. Wills, Existence, positivity and contractivity for quantum stochastic flows with infinite dimensional noise, Probab. Theory Related Fields 116 (2000), no. 4, 505 -- 543. [21] J. M. Lindsay and S. J. Wills, Markovian cocycles on operator algebras adapted to a Fock filtration, J. Funct. Anal. 178 (2000), no. 2, 269-305. [22] J. M. Lindsay and S. J. Wills, Existence of Feller cocycles on a C ∗-algebra, Bull. London Math. Soc. 33 (2001), no. 5, 613 -- 621. [23] J. M. Lindsay and S. J. Wills, Homomorphic Feller cocycles on a C ∗-algebra, J. London Math. Soc. (2) 68 (2003), no. 1, 255-272. [24] J. M. Lindsay and S. J. Wills, Quantum stochastic cocycles and completely bounded semigroups on operator spaces, preprint, 2012. [25] P.-A. Meyer, Quantum probability for probabilists, second edition, Lecture Notes in Math- ematics 1538, Springer, Berlin, 1995. [26] K. R. Parthasarathy and K. B. Sinha, Markov chains as Evan -- Hudson diffusions in Fock space, in: S´eminaire de Probabilit´es XXIV, J. Az´ema, P.-A. Meyer and M. Yor (eds.), Lecture Notes in Mathematics 1426, Springer, Berlin, 1990, 362 -- 369. [27] R. Rebolledo, Decoherence of quantum Markov semigroups, Ann. Inst. H. Poincar´e Probab. Statist. 41 (2005), no. 3, 349 -- 373. [28] P. Robinson, Quantum diffusions on the rotation algebras and the quantum Hall effect, in: Quantum Probability and Applications V, L. Accardi & W. von Waldenfels (eds.), Lecture Notes in Mathematics 1442, Springer, Berlin, 1990, 326 -- 333. [29] K. B. Sinha and D. Goswami, Quantum stochastic processes and noncommutative geo- metry, Cambridge University Press, Cambridge, 2007. 34
1610.03929	1	1610	2016-10-13T03:54:07	Quantum information inequalities via tracial positive linear maps	[ "math.FA", "cs.IT", "math-ph", "cs.IT", "math-ph", "math.OA" ]	We present some generalizations of quantum information inequalities involving tracial positive linear maps between $C^$-algebras. Among several results, we establish a noncommutative Heisenberg uncertainty relation. More precisely, we show that if $\Phi: \mathcal{A} \to \mathcal{B}$ is a tracial positive linear map between $C^$-algebras , $\rho \in \mathcal{A}$ is a $\Phi$-density element and $A,B$ are self-adjoint operators of $\mathcal{A}$ such that $ {\rm sp}(\mbox{-i}\rho^\frac{1}{2}[A,B]\rho^\frac{1}{2}) \subseteq [m,M] $ for some scalers $0<m<M$, then under some conditions \begin{eqnarray}\label{inemain1} V_{\rho,\Phi}(A)\sharp V_{\rho,\Phi}(B)\geq \frac{1}{2\sqrt{K_{m,M}(\rho[A,B])}} \left\|\Phi(\rho [A,B])\right\|, \end{eqnarray} where $K_{m,M}(\rho[A,B])$ is the Kantorovich constant of the operator $\mbox{-i}\rho^\frac{1}{2}[A,B]\rho^\frac{1}{2}$ and $V_{\rho,\Phi}(X)$ is the generalized variance of $X$.\\ In addition, we use some arguments differing from the scalar theory to present some inequalities related to the generalized correlation and the generalized Wigner--Yanase--Dyson skew information.	math.FA	math	QUANTUM INFORMATION INEQUALITIES VIA TRACIAL POSITIVE LINEAR MAPS A. DADKHAH AND M. S. MOSLEHIAN Abstract. We present some generalizations of quantum information inequal- ities involving tracial positive linear maps between C ∗-algebras. Among several results, we establish a noncommutative Heisenberg uncertainty relation. More precisely, we show that if Φ : A → B is a tracial positive linear map between C ∗-algebras , ρ ∈ A is a Φ-density element and A, B are self-adjoint operators 1 2 ) ⊆ [m, M ] for some scalers 0 < m < M , then of A such that sp(-iρ under some conditions 1 2 [A, B]ρ Vρ,Φ(A)♯Vρ,Φ(B) ≥ 1 2pKm,M (ρ[A, B]) Φ(ρ[A, B]) , (0.1) where Km,M (ρ[A, B]) is the Kantorovich constant of the operator -iρ and Vρ,Φ(X) is the generalized variance of X. In addition, we use some arguments differing from the scalar theory to present 2 [A, B]ρ 2 1 1 some inequalities related to the generalized correlation and the generalized Wigner -- Yanase -- Dyson skew information. 1. Introduction and preliminaries In quantum measurement theory, the classical expectation value of an observ- able (self-adjoint operator) A in a quantum state (density operator) ρ is expressed by Tr(ρA). Also, the classical variance for a quantum state ρ and an observable operator A is defined by Vρ(A) := Tr(ρA2) − (Tr(ρA))2. The Heisenberg uncer- tainty relation asserts that Vρ(A)Vρ(B) ≥ 1 4 Tr(ρ[A, B])2 (1.1) for a quantum state ρ and two observables A and B; see [6]. It gives a fundamental limit for the measurements of incompatible observables. A further strong result was given by Schrodinger [14] as Vρ(A)Vρ(B) − Re(Covρ(A, B))2 ≥ 1 4 Tr(ρ[A, B])2, (1.2) 2010 Mathematics Subject Classification. 46L05, 47A63, 81P15. Key words and phrases. Tracial positive linear map; trace; generalized covariance, gen- eralized variance; generalized correlation; generalized Wigner -- Yanase skew information; C ∗- algebra. 1 2 A. DADKHAH AND M.S. MOSLEHIAN where [A, B] := AB − BA is the commutator of A, B and the classical covariance is defined by Covρ(A) := Tr(ρAB) − Tr(ρA)Tr(ρB). Yanagi et al. [18] defined the one-parameter correlation and the one-parameter Wigner -- Yanase skew information (is known as the Wigner -- Yanase -- Dyson skew information; cf. [17]) for operators A, B, respectively, as follows Corrα ρ (A, B) := Tr(ρA∗B) − Tr(ρ1−αA∗ραB) and ρ (A) := Corrα I α ρ (A, A), where α ∈ [0, 1]. They showed a trace inequality representing the relation between these two quantities as Re(Corrα ≤ I α ρ (A)I α ρ (B). (1.3) (cid:12)(cid:12)(cid:12) 2 ρ (A, B))(cid:12)(cid:12)(cid:12) In the case that α = 1 2, we get the classical notions of the correlation Corrρ(A, B) and the Wigner -- Yanase skew information Iρ(A). The classical Wigner -- Yanase skew information represents a for non-commutativity between a quantum state ρ and an observable A. Luo [10] introduced the quantity Uρ(A) as a measure of uncertainty by Uρ(A) =qVρ(A)2 − (Vρ(A) − Iρ(A))2. He then showed a Heisenberg-type uncertainty relation on Uρ(A) as Uρ(A)Uρ(B) ≥ 1 4 Tr(ρ[A, B])2. (1.4) These inequalities was studied and extended by a number of mathematicians. For further information we refer interested readers to [3, 5, 7, 16]. Let B(H) denote the C ∗-algebra of all bounded linear operators on a complex Hilbert space (H, h·, ·i) with the unit I. If H = Cn, we identify B(Cn) with the matrix algebra of n × n complex matrices Mn(C). We consider the usual Lowner order ≤ on the real space of self-adjoint operators. Throughout the paper, a capital letter means an operator in B(H). An operator A is said to be strictly positive (denoted by A > 0) if it is a positive invertible operator. According to the Gelfand -- Naimark -- Segal theorem, every C ∗-algebra can be regarded as a C ∗-subalgebra of B(H) for some Hilbert space H. We use A, B, · · · to denote C ∗-algebras. We denote by Re(A) and Im(A) the real and imaginary parts of A, respectively, so we may consider elements of A as Hilbert space operators. The geometric mean is defined by A♯B = A 2 for operators A > 0 and B ≥ 0. A W ∗-algebra is a ∗-algebra of bounded operators on a Hilbert space that is closed in the weak operator topology and contains the identity operator. The C ∗-algebra of complex valued continuous functions on the compact Hausdorff A 2 BA− 1 1 2 (cid:16)A− 1 1 2 1 2(cid:17) QUANTUM INFORMATION INEQUALITIES 3 space Ω is denoted by C(Ω). A linear map Φ : A −→ B between C ∗-algebras is said to be ∗-linear if Φ(A∗) = Φ(A)∗. It is positive if Φ(A) ≥ 0 whenever A ≥ 0. It is called strictly positive if A > 0, then Φ(A) > 0. We say that Φ is unital if A, B are unital and Φ preserves the unit. A linear map Φ is called n-positive if the map Φn : Mn(A) −→ Mn(B) defined by Φn([aij]) = [Φ(aij)] is positive, where Mn(A) stands for the C ∗-algebra of n × n matrices with entries in A. A map Φ is said to be completely positive if it is n-positive for every n ∈ N. According to [15, Theorem 1.2.4] if the range of the positive linear map Φ is commutative, then Φ is completely positive. It is known (see, e.g., [4]) that if Φ is a unital positive linear map, then Φ(A♯B) ≤ Φ(A)♯Φ(B). (1.5) A map Φ is called tracial if Φ(AB) = Φ(BA). The usual trace on the trace class operators acting on a Hilbert space is a tracial positive linear functional. For a given closed two sided ideal I of a C ∗-algebra A, the existence of a tracial positive linear map Φ : A → A satisfying Φ(Φ(A)) = Φ(A) and Φ(A) − A ∈ I is equivalent to the commutativity of the quotient A/I; see [2] for more examples and implications of the definition. For a tracial positive linear map Φ, a positive operator ρ ∈ A is said to be Φ-density if Φ(ρ) = I. A unital C ∗-algebra B is said to be injective whenever for every unital C ∗-algebra A and for every self-adjoint subspace S of A, each unital completely positive linear map from S into B, can be extended to a completely positive linear map from A into B. Our investigation is based on the following definition. Definition 1.1. Let Φ : A −→ B be a tracial positive linear map and ρ be a Φ-density operator. Then Covρ,Φ(A, B) := Φ(ρA∗B) − Φ(ρA∗)Φ(ρB) and Vρ,Φ(A) := Covρ,Φ(A, A), are called the generalized covariance and the generalized variance A, B, respec- tively. Further, the generalized correlation and the generalized Wigner -- Yanase -- Dyson skew information of two operators A, B are defined by Corrα ρ,Φ(A, B) := Φ(ρA∗B) − Φ(ρ1−αA∗ραB) and I α ρ,Φ(A) := Corrα ρ,Φ(A, A) , respectively. It is known that for every tracial positive linear map, the matrix " Vρ,Φ(A) Covρ,Φ(A, B) Covρ,Φ(B, A) Vρ,Φ(B) # 4 A. DADKHAH AND M.S. MOSLEHIAN is positive, which is equivalent to Vρ,Φ(A) ≥ Covρ,Φ(B, A)(Vρ,Φ(B))−1Covρ,Φ(A, B), (1.6) which is called the variance-covariance inequality; see [13, 12] for technical dis- cussions. If A is a C ∗-algebra and B is a C ∗-subalgebra of A, then a conditional expec- tation E : A −→ B is a positive contractive linear map such that E(BAC) = BE(A)C for every A ∈ A and all B, C ∈ B. If (X , h·, ·i) is a semi-inner product module over a C ∗-algebra A, then the Cauchy -- Schwarz inequality for x, y ∈ X asserts that (see [9, 1]) hx, yihy, xi ≤ khy, yikhx, xi. If hy, yi ∈ Z(A), where Z(A) is the center of the C ∗-algebra A, then the latter inequality turns into (see [8]) hx, yihy, xi ≤ hy, yihx, xi. (1.7) In Section 1, we use some techniques in the non-commutative setting to give some generalizations of inequalities (1.2) and (1.1) for tracial positive linear maps between C ∗-algebras. More precisely, for a tracial positive linear map Φ between C ∗-algebras under ceratin conditions we show that Vρ,Φ(A)♯Vρ,Φ(B) ≥ Φ(ρ[A, B]) 1 2pKm,M (ρ[A, B]) for every self adjoint operators A, B. Section 2 deals with generalizations of inequalities (1.3) and (1.4) for conditional expectation maps. Among other things, we prove that 0 ≤ I α ρ,Φ(A) ≤ Iρ,Φ(A) ≤ Vρ,Φ(A). for every self-adjoint operator A. In addition, we generalize some significant in- equalities for trace in the quantum mechanical systems to inequalities for tracial positive linear maps between C ∗ -algebras. We indeed use some arguments differ- ing from the classical theory to present some inequalities related to the generalized correlation and the generalized Wigner -- Yanase -- Dyson skew information. 2. Inequalities for generalized covariance and variance We start this section by giving a generalization of inequality (1.2). In fact we prove inequality (1.2) for a tracial positive linear map between C ∗-algebras under some mild conditions. We need the following notions slightly differing from the notions defined in Definition 1.1. QUANTUM INFORMATION INEQUALITIES 5 Definition 2.1. For a tracial positive linear map Φ from a C ∗-algebra A into a C ∗-algebra B and positive operator ρ ∈ A and for operators A, B ∈ A we set Cov′ ρ,Φ(A, B) = Φ(ρA∗B) − Φ(ρA∗)Φ(ρ)−1Φ(ρB) and V ′ ρ,Φ(A) = Cov′ ρ,Φ(A, A). To achieve our result we need the following lemma. Lemma 2.2. [2, Lemma 2.1] Let A > 0, B ≥ 0 be two operators in A. Then the block matrix " A X X ∗ B# is positive if and only if B ≥ X ∗A−1X. We are ready to prove our first result. Theorem 2.3. Let Φ : A −→ B be a tracial positive linear map between C ∗- algebras and ρ ∈ A be a positive operator such that Φ(ρ) > 0. If Φ(A) is a commutative subspace of B, then V ′ ρ,Φ(A)V ′ ρ,Φ(B) − Re(Cov′ ρ,Φ(A, B))2 ≥ 1 4 Φ(ρ[A, B])2 for all self-adjoint operators A, B. In particular, V ′ ρ,Φ(A)V ′ ρ,Φ(B) ≥ 1 4 Φ(ρ[A, B])2. Proof. A simple calculation shows that Cov′ ρ,Φ(A, B) − Cov′ ρ,Φ(B, A) = Φ(ρAB) − Φ(ρA)Φ(ρ)−1Φ(ρB) −Φ(ρBA) + Φ(ρB)Φ(ρ)−1Φ(ρA) = Φ(ρ[A, B]) (since, Φ(A) is commutative) and Cov′ ρ,Ψ(A, B) + Cov′ ρ,Φ(B, A) = Cov′ = 2Re(Cov′ ρ,Ψ(A, B) +(cid:0)Cov′ ρ,Φ(A, B)). ρ,Φ(A, B)(cid:1)∗ Summing both sides of the above inequalities, we get 2Cov′ ρ,Φ(A, B) = Φ(ρ[A, B]) + 2Re(Cov′ ρ,Φ(A, B)). Since Φ(ρ[A, B])∗ = −Φ(ρ[A, B]) and 2Re(Cov′ from the commutativity of Φ(A) that ρΦ(A, B)) is self-adjoint, it follows Cov′ ρ,Φ(A, B)2 = Re(Cov′ ρ,Φ(A, B))2 + 1 4 Φ(ρ[A, B])2. (2.1) 6 A. DADKHAH AND M.S. MOSLEHIAN Moreover, ρ ρ 1 1 2 A∗Aρ 2 B ∗Aρ ρ 2 Aρ 1 2 1 1 2 ρ 2 ρ 1 1 2 ρ 2 ρ 1 1 1 2 A∗Bρ 2 B ∗Bρ ρ 2 Bρ 1 2 1 1 2 A∗ρ 2 B ∗ρ 1 1 2 1 2 ρ     =  ρ ρ 1 1 2 A∗ 0 0 2 B ∗ 0 0 ρ 0 0 1 2 Aρ 0 0 1 2 Bρ 0 0 1 2 1 2 ρ 0 0     ≥ 0.   It follows from the complete positivity (and indeed the 3-positivity) and the tracial property of Φ that Φ(ρA∗A) Φ(ρA∗B) Φ(ρA∗) Φ(ρB ∗A) Φ(ρB ∗B) Φ(ρB ∗)   "Φ(ρA∗A) Φ(ρA∗B) Φ(ρB ∗A) Φ(ρB ∗B)# ≥"Φ(ρA)∗ 0   Φ(ρB)∗ 0#"Φ(ρ)−1 0 Hence, by applying Lemma 2.2, we have Φ(ρB) Φ(ρA) Φ(ρ) 0 ≥ 0. 0#"Φ(ρA) Φ(ρB) 0 # , 0 whence Φ(ρB ∗A) Φ(ρB ∗B)# ≥"Φ(ρA)∗Φ(ρ)−1Φ(ρA) Φ(ρA)∗Φ(ρ)−1Φ(ρB) "Φ(ρA∗A) Φ(ρA∗B) Φ(ρB)∗Φ(ρ)−1Φ(ρA) Φ(ρB)∗Φ(ρ)−1Φ(ρB)# , or equivalently , " V ′ ρ,Φ(A) Cov′ ρ,Φ(A, B) Cov′ ρ,Φ(B, A) V ′ ρ,Φ(B) # ≥ 0. It follows from Lemma 2.2 that ρ,Φ(A) ≥ Cov′ V ′ ρ,Φ(A, B)∗(cid:0)V ′ ρ,Φ(B)(cid:1)−1 Applying the commutativity Φ(A), we get Cov′ ρ,Φ(A, B). V ′ ρ,Φ(A)V ′ ρ,Φ(B) ≥ Cov′ ρ,Φ(A, B)2. Consequently, if A, B are self-adjoint operators, then (cid:0)Φ(ρA2) − Φ(ρA)2Φ(ρ)−1(cid:1)(cid:0)Φ(ρB2) − Φ(ρB)2Φ(ρ)−1(cid:1) ρ,Φ(A)V ′ ρ,Φ(B) = V ′ ≥ Re(Cov′ ρ,Ψ(A, B))2 + 1 4 Φ(ρ[A, B])2 (by equality (2.1)). (cid:3) QUANTUM INFORMATION INEQUALITIES 7 Corollary 2.4. Let Φ : A −→ B be a tracial positive linear map between C ∗- algebras and ρ ∈ A be a Φ-density operator. If Φ(A) is a commutative subspace of B, then Vρ,Φ(A)Vρ,Φ(B) − Re(Covρ,Φ(A, B))2 ≥ 1 4 Φ(ρ[A, B])2 for all self-adjoint operators A, B. Proof. Obviously, if ρ is a Φ-density operator, then Φ(ρ)−1 = I. Now Theorem 2.3 yields the required inequality. (cid:3) Let A be a C ∗-algebra and B be a C ∗-subalgebra of A. If E : A −→ B is a tracial conditional expectation, then BE(A) = E(BA) = E(AB) = E(A)B (2.2) for every A ∈ A and B ∈ B. Using this fact we give the following corollary. Corollary 2.5. Let A be a C ∗-algebra and B be a C ∗-subalgebra of A. If E : A −→ B is a tracial conditional expectation, then Vρ,E(A)Vρ,E(B) − Re(Covρ,E(A, B))2 ≥ 1 4 E(ρ[A, B])2 for all self-adjoint operators A, B ∈ A and each E -density operator ρ ∈ A. Now we give a version of Heisenberg's uncertainty relation, in the case that B is not a commutative C ∗-algebra. To get this result we need some lemmas. Lemma 2.6 (Choi -- Tsui). [2, pp. 59-60] Let A, B be C ∗-algebras such that either one of them is W ∗-algebra or B is an injective C ∗-algebra. Let Φ : A −→ B be a tracial positive linear map. Then there exist a commutative C ∗-algebra C(X) and tracial positive linear maps φ1 : A −→ C(X) and φ2 : C(X) −→ B such that Φ = φ2 ◦ φ1. Moreover, in the case that Φ is unital, then φ1 and φ2 can be chosen to be unital. In particular, Φ is completely positive. Lemma 2.7. (Kadison's inequality) [4, Chapter 1] If Φ : A −→ B is a unital 2-positive linear map between unital C ∗-algebras, then Φ(A2) ≥ Φ(A)2 for every A ∈ A. In the case that A is a positive operator of A satisfying 0 < mI ≤ A ≤ MI for some scalers m < M, by [4, Theorem 1.32], the reverse inequality Φ(A2) ≤ (M + m)2 4Mm Φ(A)2 (2.3) 8 holds. A. DADKHAH AND M.S. MOSLEHIAN Lemma 2.8. Let Φ : A −→ B be a unital 2-positive linear map between unital C ∗-algebras. If A is an operator of A satisfying sp(A) ⊆ [m, M] ∪ [−M, −m] for some scalers 0 < m < M , then Φ(A) ≤r (M + m)2 4Mm Φ(A). (2.4) Proof. Using Lemma 2.7 and inequality (2.3), we get Φ(A)2 ≤ Φ(A2) ≤ (M + m)2 4Mm (Φ(A))2. Taking square roots of both sides of the latter inequality we obtain inequality (2.4). (cid:3) (M + m)2 4Mm In this paper, we denote for an operator m ≤ A ≤ M by KM,m(A), which is called the Kantorovich constant of A. The next theorem gives a Heisenberg's type uncertainty relation for tracial posi- tive linear maps between C ∗-algebras. Theorem 2.9. Let A, B be unital C ∗-algebras and Ω be a compact Hausdorff space. Let φ1 : A −→ C(Ω) be a unital tracial positive linear map and φ2 : C(Ω) −→ B be a unital positive linear map and Φ := φ2 ◦ φ1. If ρ ∈ A is a Φ- density operator and A, B are self-adjoint operators of A such that sp(-iρ 2 [A, B]ρ [m, M] for some scalers 0 < m < M , then 1 1 2 ) ⊆ Vρ,Φ(A)♯Vρ,Φ(B) ≥ Φ(ρ[A, B]) , (2.5) 1 2pKm,M (ρ[A, B]) where Km,M (ρ[A, B]) is the Kantorovich constant of the operator -iρ 1 2 [A, B]ρ 1 2 . 1 Proof. By a continuity argument we can assume that ρ > 0. Due to 0 < mI ≤ 2 ≤ MI and φ1 is unital and tracial positive linear, we infer that -iρ mI ≤ -iφ1(ρ[A, B]) ≤ MI. It follows that 2 [A, B]ρ 1 mI ≤ φ1(ρ[A, B]) ≤ MI. (2.6) Using the fact that φ1 is a unital positive linear map (and so strictly positive) and applying Theorem 2.3 for φ1 we get Applying the commutativity of range of φ1 we get (cid:0)φ1(ρA2) − φ1(ρA)2φ1(ρ)−1(cid:1)(cid:0)φ1(ρB2) − φ1(ρB)2φ1(ρ)−1(cid:1) ≥ (cid:0)φ1(ρA2) − φ1(ρA)2φ1(ρ)−1(cid:1) ♯(cid:0)φ1(ρB2) − φ1(ρB)2φ1(ρ)−1(cid:1) ≥ 1 4 φ1(ρ[A, B])2. 1 2 φ1(ρ[A, B]).(2.7) QUANTUM INFORMATION INEQUALITIES 9 Using the fact that (φ2 ◦ φ1)(ρ) = I, we can write (cid:0)Φ(ρA2) − Φ(ρA)2(cid:1)♯(cid:0)Φ(ρB2) − Φ(ρB)2(cid:1) =(cid:0)(φ2 ◦ φ1)(ρA2) − (φ2 ◦ φ1)(ρA)2(cid:1) ♯(cid:0)(φ2 ◦ φ1)(ρB2) − (φ2 ◦ φ1)(ρB)2(cid:1) =(cid:0)(φ2 ◦ φ1)(ρA2) − (φ2 ◦ φ1)(ρA)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρA)(cid:1) ♯(cid:0)(φ2 ◦ φ1)(ρB2) − (φ2 ◦ φ1)(ρB)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρB)(cid:1) . We claim that (cid:0)(φ2 ◦ φ1)(ρA2) − (φ2 ◦ φ1)(ρA)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρA)(cid:1) ♯(cid:0)(φ2 ◦ φ1)(ρB2) − (φ2 ◦ φ1)(ρB)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρB)(cid:1) ≥ φ2(cid:0)φ1(ρA2) − φ1(ρA)2φ1(ρ)−1(cid:1)♯φ2(cid:0)φ1(ρB2) − φ1(ρB)2φ1(ρ)−1(cid:1). Since the range of φ1 is commutative, to prove our claim, it is enough to show that φ2(cid:0)φ1(ρX)φ1(ρ)−1φ1(ρX)(cid:1) ≥ (φ2 ◦ φ1)(ρX)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρX) (2.8) for every self-adjoint operator X ∈ A and (φ2 ◦ φ1)-density operator ρ ∈ A. Clearly, matrix φ1(ρX)φ1(ρ)−1φ1(ρX) φ1(ρX) φ1(ρ) ! is positive. Since Φ2 is com- φ1(ρX) pletely positive (and so 2-positive), we get φ2(φ1(ρX)φ1(ρ)−1φ1(ρX)) (φ2 ◦ φ1)(ρX) (φ2 ◦ φ1)(ρX) (φ2 ◦ φ1)(ρ) ! ≥ 0, 10 A. DADKHAH AND M.S. MOSLEHIAN which ensures the validity of inequality (2.8). Therefore, Vρ,Φ(A)♯Vρ,Φ(B) = (cid:0)Φ(ρA2) − Φ(ρA)2(cid:1) ♯(cid:0)Φ(ρB2) − Φ(ρB)2(cid:1) = (cid:0)(φ2 ◦ φ1)(ρA2) − (φ2 ◦ φ1)(ρA)2(cid:1) ♯(cid:0)(φ2 ◦ φ1)(ρB2) − (φ2 ◦ φ1)(ρB)2(cid:1) = (cid:0)(φ2 ◦ φ1)(ρA2) − (φ2 ◦ φ1)(ρA)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρA)(cid:1) ♯(cid:0)(φ2 ◦ φ1)(ρB2) − (φ2 ◦ φ1)(ρB)(φ2 ◦ φ1)(ρ)(φ2 ◦ φ1)(ρB)(cid:1) ≥ φ2(cid:0)φ1(ρA2) − φ1(ρA)2φ1(ρ)−1(cid:1) ♯φ2(cid:0)φ1(ρB2) − φ1(ρB)2φ1(ρ)−1(cid:1)(by inequality (2.8)) ≥ φ2(cid:16)(cid:0)φ1(ρA2) − φ1(ρA)2φ1(ρ)−1(cid:1) ♯(cid:0)φ1(ρB2) − φ1(ρB)2φ1(ρ)−1(cid:1)(cid:17)(by inequality(1.5)) ≥ ≥ = 1 2 φ2(φ1(ρ[A, B])) 1 2pKm,M (ρ[A, B]) 2pKm,M (ρ[A, B]) 1 (by inequality (2.7) φ2 ◦ φ1(ρ[A, B]) (by inequality (2.4)) Φ(ρ[A, B]). (2.9) (cid:3) Applying Lemma 2.6, we immediately get the following corollary. Corollary 2.10. Let A, B be C ∗-algebras such that either one of them is W ∗- algebra or B is an injective C ∗-algebra. Let Φ : A −→ B be a tracial positive linear map and ρ ∈ A be a Φ-density operator and A, B are self-adjoint operators in A such that sp(-iρ 2 ) ⊆ [m, M] for some scalars 0 < m < M , then 1 1 2 [A, B]ρ Vρ,Φ(A)♯Vρ,Φ(B) ≥ Φ(ρ[A, B]) , 1 2pKm,M (ρ[A, B]) where K[m,M ](ρ[A, B]) is the Kantorovich constant of the operator -iρ 1 2 [A, B]ρ 1 2 . If φ1 := Tr is the usual trace and φ2 is the identity map on C, then it immedi- ately follows from inequality (2.9) that Vρ(A) 1 2 Vρ(B) 1 2 ≥ 1 2 Tr(ρ[A, B]) . QUANTUM INFORMATION INEQUALITIES 11 So we get the following result. Corollary 2.11. For every self-adjoint operators A, B and each density operator ρ it holds that Vρ(A)Vρ(B) ≥ 1 4 Tr(ρ[A, B])2 . Remark 2.12. Let A and B be unital C ∗-algebras. By readout the proof of The- orem 2.3, Theorem 2.9 and Corollary 2.4, if we put ρ := I, then we can replace the condition "tracial positive linear map" with "unital positive linear map". 3. Inequalities for generalized correlation and Wigner -- Yanase skew information We aim to give generalizations of inequality (1.3) and inequality (1.4). In ad- dition we generalize some inequalities related to classical Wigner -- Yanase -- Dyson skew information to tracial positive linear map. As mentioned in the introduc- tion, in the case that α = 1 ρ,Φ by Iρ,Φ. In some cases we prove our result by assuming that ρ is only a positive operator. To get these generalizations we need some lemmas. The technique of the first lemma is classic. 2 we denote I α Lemma 3.1. Let A and B be C ∗-algebras. If Φ : A −→ B is a tracial positive linear map, then ρ,Φ(A) = Φ(ρA2) − Φ(ρ1−αAραA) ≥ 0 I α (α ∈ [0, 1]) (3.1) for every self-adjoint operator A ∈ A and each positive operator ρ ∈ A. Proof. Put ∆ = {α ∈ [0, 1] : Φ(ρA2) ≥ Φ(ραAρ1−αA)}. Clearly {0, 1} ⊆ ∆ and the set ∆ is closed, since the map α −→ Φ(ραAρ1−αA) is norm continuous. Therefore, to prove [0, 1] ⊆ ∆ it is enough to show that α, β ∈ ∆ implies α+β 2 ∈ ∆. It follows from the tracial positivity of Φ that 1−α 2 Aρ α 2 − ρ 1−β 2 Aρ β 2 )∗(ρ 1−α 2 Aρ α 2 − ρ 1−β 2 Aρ α 2 Aρ 1−α 2 − ρ β 2 Aρ 1−β 2 )(ρ 1−α 2 Aρ α 2 − ρ 1−β 2 Aρ β β 2 )(cid:17) 2 )(cid:17) 0 ≤ Φ(cid:16)(ρ = Φ(cid:16)(ρ = Φ(cid:16)ρ α 2 Aρ1−αAρ α 2 − ρ α 2 Aρ1− α+β 2 Aρ β 2 − ρ β 2 Aρ1− α+β 2 Aρ α 2 + ρ β 2 Aρ1−βAρ = Φ(ραAρ1−αA) − Φ(ρ α+β 2 Aρ1− α+β 2 A) − Φ(ρ α+β 2 Aρ1− α+β 2 A) + Φ(ρβAρ1−βA). β 2(cid:17) Hence, Φ(ραAρ1−αA) + Φ(ρβAρ1−βA) ≥ 2Φ(ρ α+β 2 Aρ1− α+β 2 A), (3.2) 12 A. DADKHAH AND M.S. MOSLEHIAN since α, β ∈ ∆, we infer Φ(ρA2) ≥ Φ(ρ α+β 2 Aρ1− α+β 2 A), which implies that α+β 2 ∈ ∆. (cid:3) Remark 3.2. Inequality (3.2) shows that the map α → Φ(ραAρ1−αA) is convex. In addition, we have I α ρ,Φ(A) ≤ Iρ,Φ(A) (α ∈ R) (3.3) for every self-adjoint operator A and each positive operator ρ. Indeed, using the positivity and the tracial property of Φ, we get α 2 Aρ 1−α 2 − ρ 1−α 2 Aρ α 2 )(ρ 1−α 2 Aρ α 2 − ρ α 2 Aρ 0 ≤ Φ(cid:16)(ρ = 2Φ(ραAρ1−αA) − 2Φ(ρ 1 2 Aρ 1 2 A). α 2 )(cid:17) Therefore Φ(ραAρ1−αA) ≥ Φ(ρ 1 2 Aρ 1 2 A) and so ρ,Φ(A) = Φ(ρA2) − Φ(ραAρ1−αA) ≤ Φ(ρA2) − Φ(ρ I α 1 2 Aρ 1 2 A) = Iρ,Φ(A). Let Φ be a tracial positive linear map between C ∗ algebras. It follows from Lemma 3.1 that I α ρ,Φ(A) ≥ 0 for every self-adjoint operator A, but it is not true in general when A is an arbitrary operator. Hence even if Φ is a tracial positive functional, then Corrα ρ,Φ(·, ·) cannot induce a complex valued semi-inner product and we cannot use the Cauchy -- Schwarz inequality; see [18, Remark IV.2]. The next lemma helps us to give a positive definite version of the generalized correlation. Lemma 3.3. Let Φ : A −→ B be a tracial positive linear map and ρ ∈ A be a positive operator. Then I α ρ,Φ(A) + I α ρ,Φ(A∗) ≥ 0 for every A ∈ A. Proof. We define Φ : M2(A) −→ B by Φ "A B that Φ is a tracial positive linear map. Take A = " 0 A∗ C D#! = 1 2 Φ(A + D). It is obvious A 0 # and ρ = "ρ 0 0 ρ#. Clearly A is a self-adjoint operator in M2(A) and ρ is a Φ-density operator. QUANTUM INFORMATION INEQUALITIES 13 Using Lemma 3.1 for Φ, we get 1 2 (Φ(ρA∗A) + Φ(ρAA∗)) = Φ "ρ 0 0 ρ#" 0 A∗ A 0 #! A 0 #" 0 A∗ = Φ(ρ A2) ≥ Φ(ρ1−α Aρα A) (by Lemma 3.1) = Φ "ρ1−α 0 0 ρ1−α#" 0 A∗ A 0 #"ρα 0 0 A 0 #! ρα#" 0 A∗ = 1 2 (Φ(ρ1−αA∗ραA) + Φ(ρ1−αAραA∗)). Hence, I α ρ,Φ(A) + I α ρ,Φ(A∗) = Φ(ρA∗A) + Φ(ρAA∗) −Φ(ρ1−αA∗ραA) − Φ(ρ1−αAραA∗) ≥ 0. Definition 3.4. Let Φ : A −→ B be a tracial positive linear map and ρ ∈ A be a Φ-density operator. Then for every operators A, B ∈ A, we set (cid:3) Corr′α ρ,Φ(A, B) := ρ,Φ(A, B) + Corrα 1 2(cid:0)Corrα ρ,Φ(B ∗, A∗)(cid:1) and I ′α ρ,Φ(A) := Corr′α ρ,Φ(A, A). It is easy to check that Corr′α ρ,Φ(A, B) has the following properties: (i) Corr′α (ii) Corr′α ρ,Φ(A, A) ≥ 0, for every A ∈ A, (by Lemma 3.3), ρ,Φ(A, B + λC) = Corr′α ρ,Φ(A, B) + λCorr′α ρ,Φ(A, C), for all A, B ∈ A and every λ ∈ C, (iii) Corr′α ρ,Φ(A, B)∗ = Corr′α ρ,Φ(B, A). Next we give a generalization of inequality (1.3). Theorem 3.5. Let A be a C ∗-algebra and B be C ∗-subalgebra of A. If E : A −→ B is a tracial conditional expectation, then (cid:12)(cid:12)Re(Corrα ρ,E(A, B))(cid:12)(cid:12) 2 ≤ I α ρ,E(A)I α ρ,E(B) for all self-adjoint operators A, B ∈ A and each E -density operator ρ ∈ A. 14 A. DADKHAH AND M.S. MOSLEHIAN Proof. Define the map h·, ·i : A×A −→ B by hA, Bi = Corr′α and C ∈ B, then ρ,E(A, B). If A, B ∈ A hA, BCi = Corr′α ρ,E(A, BC) = = = = 1 1 1 ρ,E(A, BC) + Corrα ρ,E(C ∗B ∗, A∗)(cid:1) 2(cid:0)Corrα 2(cid:0)E(ρA∗BC) − E(ρ1−αA∗ραBC) + E(ρBCA∗) − E(ρ1−αBCραA∗)(cid:1) 2(cid:0)E(ρA∗BC) − E(ρ1−αA∗ραBC) + E(CA∗ρB) − E(CραA∗ρ1−αB)(cid:1) 2(cid:0)E(ρA∗B)C − E(ρ1−αA∗ραB)C + E(ρBA∗)C − E(ρ1−αBραA∗)C(cid:1) (since E is a conditional expectation and by equality (2.2)) (since E is tracial) 1 = Corr′α ρ,E(A, B)C = hA, BiC. Using this fact and Definition 3.4 we see that (A, h·, ·i) is a semi-inner product B-module. Moreover, equality (2.2) shows that ran(Z) ⊆ Z(B). If A and B are self-adjoint operators in A, then we get (cid:12)(cid:12)Re(Corrα ρ,E(A, B))(cid:12)(cid:12) 2 ρ,E(B, A)(cid:1)(cid:12)(cid:12)(cid:12)(cid:12) ρ,E(A, B) + Corrα 1 2 = (cid:12)(cid:12)(cid:12)(cid:12) 2(cid:0)Corrα = (cid:12)(cid:12)Corr′α = hA, Bi2 ρ,E(A, B)(cid:12)(cid:12) 2 ≤ hA, AihB, Bi (by inequality (1.7)) = I α ρ,E(A)I α ρ,E(B) (since A, B are self − adjoint). (cid:3) Let A be a self-adjoint operator and ρ be a density operator. According to [11, Section III] we have Iρ(A) ≤ Vρ(A). We give a generalization of this inequality for a tracial 2-positive linear map Φ and a Φ-density operator ρ. It follows from Lemma 2.2 that the matrix "ρ 1 4 Aρ 1 1 4 Aρ 2 ρ ρ 4 ρ 4 Aρ 1 1 1 4 1 4 ρ 1 4 ρ 1 2 1 4 Aρ ρ # is positive. Since Φ is 2-positive, we have "Φ(ρ 1 1 1 4 Aρ 2 ρ 2 ρ 4 Aρ 2 Aρ 4 ) 1 1 1 Φ(ρ 1 4 ) Φ(ρ 1 4 ρ 1 4 Aρ Φ(ρ) 1 2 ) # ≥ 0. QUANTUM INFORMATION INEQUALITIES 15 Therefore, by using Lemma 2.2 and applying the tracial property of Φ we get Φ(ρ 1 2 Aρ 1 2 A) ≥ Φ(ρA)Φ(ρ)−1Φ(ρA) = Φ(ρA)2, (3.4) which implies that Iρ,Φ(A) ≤ Vρ,Φ(A). Consequently, by using inequality (3.3), we reach I α For a tracial positive linear map Φ and a self-adjoint operator A, we set ρ,Φ(A) ≤ Vρ,Φ(A). Jρ,Φ(A) := 2Vρ,Φ(A) − Iρ,Φ(A) and Uρ,Φ(A) := Iρ,Φ(A)♯Jρ,Φ(A). Since Uρ,Φ(A) ≤ Vρ,Φ(A) (by the arithmetic-geometric mean inequality), the next theorem is a refinement of Theorem 2.9 in the case that Φ is a conditional ex- pectation. To establish it, we model the classical techniques (see [10]) to the non commutative framework. Theorem 3.6. Let A be a C ∗-algebra and B be a C ∗-subalgebra of A. If E : A −→ B is a tracial conditional expectation, then Uρ,E(A)Uρ,E (B) ≥ 1 4 E(ρ[A, B])2 (3.5) for all self-adjoint operators A, B ∈ A and each E -density operator ρ ∈ A. Proof. Consider self-adjoint operators A0 = A − E(ρA) and B0 = B − E(ρB). A simple calculation shows that Iρ,E(A) = 1 2 E((i[ρ 1 2 , A0])2) and Jρ,E(B) = E({ρ 1 2 , B0}2), (3.6) 1 2 where {ρ 1 2 , B} = ρ 1 2 B0 + B0ρ 1 2 . Indeed, E((i[ρ 1 2 , A0])2) = −E(ρ 1 2 A0ρ 1 2 A0 − ρ 1 2 A2 0ρ 1 2 − A0ρ 1 2 ρ 1 2 A0 + A0ρ 1 2 A0ρ 1 2 ) = 2E(ρA2 0) − 2E(ρ 1 2 A0ρ 1 2 A0) (by the tracial property of E) = 2E(cid:16)ρ(A − E(ρA))2(cid:17) − 2E(cid:16)ρ = 2E(ρA2) − 2(E(ρA))2 − 2E(ρ 1 1 2 Aρ 1 2 (A − E(ρA))ρ 1 2 (A − E(ρA))(cid:17) 2 A) + 2(E(ρA))2 (since E is a conditional expectation) = 2E(ρA2) − 2E(ρ 1 2 Aρ 1 2 A). 16 A. DADKHAH AND M.S. MOSLEHIAN Similarly, we can establish the other inequality in (3.6). Let Z ∈ B. Then 1 2 , A0]{ρ 1 2 , B0} + i{ρ 1 2 , B0}[ρ 1 2 , A0]Z(cid:17) 1 2 A0ρ 2 B0 + ρ 1 2 A0B0ρ 1 2 − A0ρ 1 2 ρ 1 2 B0 1 = ZE(cid:16)i(cid:0)ρ −A0ρ 1 2 B0ρ 1 2 + ρ 1 2 B0ρ 1 2 A0 −ρ 1 2 B0A0ρ 1 2 + B0ρ 1 2 ρ 1 2 A0 − B0ρ 1 2 A0ρ (by equality (2.2)) 1 2(cid:1)(cid:17) = 2iZE(ρ[A0, B0]) = 2iZE(cid:0)ρ [A − E(ρA), B − E(ρB)](cid:1) = 2iZE(cid:16)ρ(cid:0)(A − E(ρA))(B − E(ρB)) −(B − E(ρB))(A − E(ρA))(cid:1)(cid:17) = 2iZE(cid:16)ρ(cid:0)AB − AE(ρB) − E(ρA)B + E(ρA)E(ρB) −BA + BE(ρA) + E(ρB)A − E(ρB)E(ρA)(cid:1)(cid:17) = 2iZ(cid:0)E(ρAB) − E(ρA)E(ρB) − E(ρA)E(ρB) +E(ρ)E(ρA)E(ρB) − E(ρBA) + E(ρB)E(ρA) +E(ρB)E(ρA) − E(ρ)E(ρB)E(ρA)(cid:1) = 2iZE(ρ[A, B]) (by equality 2.2). E(cid:16)iZ[ρ Hence, 1 2 , A0]{ρ 1 2 , B0} + i{ρ 1 2 , B0}[ρ E(cid:16)iZ[ρ 1 2 , A0]Z(cid:17) = 2iZE(ρ[A, B]). 1 1 (3.7) Let Z ∈ B be a self-adjoint operator and X = i[ρ 2 , A0]Z + {ρ 2 , B0}. Then 1 0 ≤ E(X ∗X) = E(cid:16)(cid:0)iZ[ρ = E(cid:16) − Z[ρ 1 2 , A0] + {ρ 1 2 , A0]Z + {ρ 1 2 , B0}(cid:1)(cid:0)i[ρ 1 2 , B0}(cid:1)(cid:17) (because i[ρ 2 , A0] is self − adjoint) 1 1 2 , A0][ρ 1 2 , A0]Z + iZ[ρ 1 2 , A0]{ρ 1 2 , B0} +i{ρ 2 , B0}[ρ 1 2 , A0]Z + {ρ 1 2 , B0}2(cid:17) = 2Iρ,E(A)Z 2 + 2iE(ρ[A, B])Z + 2Jρ,E(B) (by equality (3.6) and equality (3.7)). QUANTUM INFORMATION INEQUALITIES 17 Without loss of the generality we can assume that Iρ,E(A) > 0. If we put Z := − i 2 Iρ,E(A)−1E(ρ[A, B]), then we get 1 2 Iρ,E(A)−1E(ρ[A, B])2 + 1 4 − Iρ,E(A)−1E(ρ[A, B])2 + Jρ,E(B) ≥ 0. or equivalently, Iρ,E(A)Jρ,E(B) ≥ − 1 4 E(ρ[A, B])2 = 1 4 E(ρ[A, B])2, (3.8) since E(ρ[A, B])∗ = −E(ρ[A, B]). It follows from the fact that for every X ∈ A, E(X) ⊆ Z(B) (equality (2.2)), we have Uρ,E(A)Uρ,E (B) = (cid:0)Iρ,E(A)♯Jρ,E(A)(cid:1)(cid:0)Iρ,E(B)♯Jρ,E(B)(cid:1) 2(cid:0)Iρ,E(B)Jρ,E(A)(cid:1) = (cid:0)Iρ,E(A)Jρ,E(B)(cid:1) 1 1 2 (by the commutivity property in equality (2.2)) ≥ 1 4 E(ρ[A, B])2 (by inequlity (3.8)). As a consequence we get the following result of Luo [10]. Corollary 3.7. [10, p. 2] If A, B are two self-adjoint operators, then (cid:3) Uρ(A)Uρ(B) ≥ 1 4 Tr(ρ[A, B])2. References [1] J.M. Aldaz, S. Barza, M. Fujii and M. S. Moslehian, Advances in operator Cauchy-Schwarz inequalities and their reverses, Ann. Funct. Anal. 6 (2015), no. 3, 275 -- 295. [2] M. D. Choi and S. K. Tsui, Tracial positive linear maps of C ∗-algebaras, Proc. Amer. Math. Soc. 87 (1983), no. 1, 57 -- 61. [3] S. Friedland, V. Gheorghiu and G. Gour, Universal uncertainty relations, Phys. Rev. Lett. 111 (2013), 230401. [4] T. Furuta, J. Mi´ci´c Hot, J. Pecari´c and Y. Seo, Mond-Pecari´c Method in Operator Inequal- ities. Inequalities for Bounded Self-adjoint Operators on a Hilbert Space, Element, Zagreb, Croatia, 2005. [5] P. Gibilisco and T. Isola, How to distinguish quantum covariances using uncertainty rela- tions, J. Math. Anal. Appl. 384 (2011), 670 -- 676. [6] W. Heisenberg, Uber den anschaulichen Inhalt der quantum mechanischen Kinematik und Mechanik, Zeitschrift fur Physik, 43 (1927), 172 -- 198. [7] C. K. Ko and H. J. Yoo, Uncertainty relation associated with a monotone pair skew infor- mation, J. Math. Anal. Appl. 383 (2011) 208 -- 214. 18 A. DADKHAH AND M.S. MOSLEHIAN [8] D. Ilisevi´c and S. Varosanec, On the Cauchy -- Schwarz inequality and its reverse in semi- inner product C ∗-modules, Banach J. Math. Anal. 1 (2007), 78 -- 84. [9] E. C. Lance, Hilbert C ∗-Modules, London Math. Soc. Lecture Note Series, Vol.210, Cam- bridge Univ. Press, 1995. [10] S. Luo, Heisenberg uncertainty relation for mixed states, Phys. Rev. A 72 (2005), 042110, pp.3. [11] S. Luo and Q. Zhang, On skew information, IEEE Trans. Inf. Theo. 50 (2004), 1778 -- 1782, and Correction to "On skew information", IEEE Trans. Info. Theo. 51 (2005), p.4432. [12] J. S. Matharu and M. S. Moslehian, Gruss inequality for some types of positive linear maps, J. Operator Theory 73 (2015), no. 1, 265 -- 278. [13] M. S. Moslehian and R. Raji´c, A Gruss inequality for n-positive linear maps, Linear Algebra Appl. 433 (2010), 1555 -- 1560. [14] E. Schrodinger, About Heisenberg uncertainty relation, Proc. Russian Acad. Sci. Phys. Math. Section XIX (1930), p.293. [15] E. Størmer, Positive Linear Maps of Operator Algebras, Springer-Verlag Berlin Heidelberg, 2013. [16] G. T´oth and D. Petz, Extremal Properties of the Variance and the Quantum Fisher Infor- mation, Phys. Rev. A 87, (2013) 0323 -- 24. [17] E. H. Lieb, Convex trace functions and the Wigner -- Yanas -- Dyson conjecture, Adv. Math. 11 (1973), 267 -- 288. [18] K. Yanagi, S. Furuichi and K. Kuriyama, A generalized skew information and uncertainty relation, IEEE Trans. Info. Theo. 51 (2005), 4401 -- 4404. Department of Pure Mathematics, Center Of Excellence in Analysis on Al- gebraic Structures (CEAAS), Ferdowsi University of Mashhad, P. O. Box 1159, Mashhad 91775, Iran E-mail address: [email protected] E-mail address: [email protected]
1009.0853	1	1009	2010-09-04T17:42:16	Decompositions of Measures on Pseudo Effect Algebras	[ "math.FA" ]	Recently in \cite{Dvu3} it was shown that if a pseudo effect algebra satisfies a kind of the Riesz Decomposition Property ((RDP) for short), then its state space is either empty or a nonempty simplex. This will allow us to prove a Yosida-Hewitt type and a Lebesgue type decomposition for measures on pseudo effect algebra with (RDP). The simplex structure of the state space will entail not only the existence of such a decomposition but also its uniqueness.	math.FA	math	DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS ANATOLIJ DVURE CENSKIJ 1 Mathematical Institute, Slovak Academy of Sciences Stef´anikova 49, SK-814 73 Bratislava, Slovakia E-mail: [email protected], Abstract. Recently in [Dvu3] it was shown that if a pseudo effect algebra satisfies a kind of the Riesz Decomposition Property ((RDP) for short), then its state space is either empty or a nonempty simplex. This will allow us to prove a Yosida -- Hewitt type and a Lebesgue type decomposition for measures on pseudo effect algebra with (RDP). The simplex structure of the state space will entail not only the existence of such a decomposition but also its uniqueness. 1. Introduction The classical decomposition theorems of or Yosida -- Hewitt [YoHe] initiated in the last two decades interest of authors studying finitely additive measures on quantum structures like orthomodular lattices or posets to study an interesting problem of decomposition measures. There appeared a whole series of papers studying Lebesgue and Yosida-Hewitt type decompositions, see e.g. [DDP, DeMo, Rut1, Rut2]. They exhibited at least the existence of such a decomposition. To prove even the uniqueness of decompositions, some sufficient conditions are presented in [Rut2]. Quantum structures were inspired by the research of the mathematical founda- tions of quantum structures. An analogue of a probability measure is a state. One of the most important examples of orthomodular lattices or of the Hilbert space quantum mechanics is the system L(H) of all closed subspaces of a Hilbert space (real, complex or quaternionic) H. The Gleason Theorem, see e.g. [Dvu0], says that every σ-additive state on L(H) is uniquely expressible via a Hermitian trace operator on H if 3 ≤ dim H ≤ ℵ0. The Aarnes Theorem, [Dvu0, Thm 3.2.28] says that every (finitely additive) state on L(H) is a unique convex combination of two states, s1 and s2, where s1 is a completely additive state and s2 is a finitely additive state that vanishes on each finite-dimensional subspace of H. In the Nineties, Foulis and Bennett [FoBe] introduced effect algebras that are par- tial structures with a partially defined operation + that models the join of mutually excluding events. They generalize orthomodular lattices and posets, orthoalgebras, 1Keywords: Pseudo effect algebra; effect algebra; Riesz Decomposition Properties; signed measure, state; Jordan signed measure, unital po-group; simplex; Yosida -- Hewitt decomposition; Lebesgue decomposition AMS classification: 81P15, 03G12, 03B50 The author thanks for the support by Center of Excellence SAS - Quantum Technologies -, ERDF OP R&D Projects CE QUTE ITMS 26240120009 and meta-QUTE ITMS 26240120022, the grant VEGA No. 2/0032/09 SAV. 1 2 ANATOLIJ DVURE CENSKIJ and the basic example important for so-called POV-measures of quantum mechan- ics is the system, E(H), of all Hermitian operators of a Hilbert space H that are between the zero operator and the identity one. These commutative structures were extended in [DvVe1, DvVe2] to so-called pseudo effect algebras where the partial addition, +, is not more assumed to be commutative. In many important examples they are intervals in po-groups (= partially ordered groups). E.g. E(H) is an interval in the po-group B(H) of all Hermitian operators on H. A sufficient condition for a pseudo effect algebra to be an interval of a unital po- group is a variant of the Riesz Decomposition Property ((RDP) in abbreviation), see [Dvu2, DvVe2]. It is a weaker form of the distributivity that allows to do a joint refinement of two decompositions of the unit element 1. For example, (RDP) on an orthomodular poset entails that it has to be a Boolean algebra, and therefore, (RDP) fails to hold on L(H) or on E(H). We recall that every effect algebra with (RDP) has at least one state, however the state space of a pseudo effect algebra with a stronger type of (RDP) can be empty, [Dvu1]. Recently in [Dvu3, Thm 5.1], it was shown that the state space of every pseudo effect algebra with (RDP) is a simplex, more precisely a Choquet simplex. The simplex is a special type of a convex set that generalizes the classical one in Rn. For a comprehensive source on simplices see [Alf]. We note that the state space of E(H) is not a simplex, however the state space of any commutative C∗-algebra is a simplex, see e.g. [AlSc, Thm 4.4, p. 7]. The simplex structure of the state spaces allows also to represent uniquely states as an integral [Dvu4] through a regular Borel probability measure. The simplex structure of the state space of a pseudo effect algebra E satisfying (RDP) entails that the space of all Jordan measures on E is an Abelian Dedekind complete ℓ-group, [Dvu3, Thm 3.5, Thm 3.6]. This new fact is our basic tool to present a Yosida -- Hewitt type and a Lebesgue type of decompositions of finitely additive measures on E. The property (RDP) as we show is a sufficient condition to prove not only the existence but also the uniqueness of such a decomposition that is a main goal of the present paper. The paper is organized as follows. Section 2 is an introduction to the theory of pseudo effect algebras gathering the necessary latest results. Section 3 describes the faces of the state space of pseudo effect algebras and it gives a general result on a decomposition and it will be applied in Section 4 to present the main body of the paper - the Yosida -- Hewitt type of decomposition and the Lebesque type decomposition. 2. Elements of Pseudo Effect Algebras Pseudo effect algebras were introduced in [DvVe1, DvVe2]. We say that a pseudo effect algebra is a partial algebra (E; +, 0, 1), where + is a partial binary operation and 0 and 1 are constants, such that for all a, b, c ∈ E, the following holds (i) a + b and (a + b) + c exist if and only if b + c and a + (b + c) exist, and in this case (a + b) + c = a + (b + c); (ii) there is exactly one d ∈ E and exactly one e ∈ E such that a+d = e+a = 1; (iii) if a + b exists, there are elements d, e ∈ E such that a + b = d + a = b + e; (iv) if 1 + a or a + 1 exists, then a = 0. DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 3 If we define a ≤ b if and only if there exists an element c ∈ E such that a + c = b, then ≤ is a partial ordering on E such that 0 ≤ a ≤ 1 for any a ∈ E. It is possible to show that a ≤ b if and only if b = a + c = d + a for some c, d ∈ E. We write c = a / b and d = b \ a. Then (b \ a) + a = a + (a / b) = b, and we write a− = 1 \ a and a∼ = a / 1 for any a ∈ E. For basic properties of pseudo effect algebras see [DvVe1, DvVe2]. We recall that if + is commutative, E is said to be an effect algebra; for a guide overview on effect algebras we recommend e.g. [DvPu]. For example, let (G, u) be a unital po-group (= partially ordered group) with strong unit u that is not necessarily Abelian. We recall that a po-group (= partially ordered group) is a group with a partial ordering ≤ such that if a ≤ b, then x + a + y ≤ x + b + y for all x, y ∈ G; and an element u ∈ G+ := {g ∈ G : g ≥ 0} is said to be a strong unit if given G =Sn[−nu, nu]. Then for (G, u) we define Γ(G, u) = [0, u] and we endow it with + that is the restriction of the group addition to the set of all those (x, y) ∈ Γ(G, u) × Γ(G, u) that x ≤ u − y. Then (Γ(G, u); +, 0, u) is a pseudo effect algebra with possible two negations: a− = u − a and a∼ = −a + u. Any pseudo effect algebra of the form Γ(G, u) is said to be an interval pseudo effect algebra. In [DvVe1, DvVe2], we have some sufficient conditions posed to a pseudo effect algebra to be an interval. They are analogues of the Riesz Decomposition Properties. Roughly speaking they are a weaker form of distributivity that allows a joint refinement of two partitions of 1. This is a reason why they fail to hold for L(H) and E(H). Now we introduce according to [DvVe1] the following types of the Riesz Decom- position properties for pseudo effect algebras that in the case of effect algebras may coincide, but not for pseudo effect algebras, in general. (a) For a, b ∈ E, we write a com b to mean that for all a1 ≤ a and b1 ≤ b, a1 and b1 commute. (b) We say that E fulfils the Riesz Interpolation Property, (RIP) for short, if for any a1, a2, b1, b2 ∈ E such that a1, a2 ≤ b1, b2 there is a c ∈ E such that a1, a2 ≤ c ≤ b1, b2. (c) We say that E fulfils the weak Riesz Decomposition Property, (RDP0) for short, if for any a, b1, b2 ∈ E such that a ≤ b1 + b2 there are d1, d2 ∈ E such that d1 ≤ b1, d2 ≤ b2 and a = d1 + d2. (d) We say that E fulfils the Riesz Decomposition Property, (RDP) for short, if for any a1, a2, b1, b2 ∈ E such that a1 + a2 = b1 + b2 there are d1, d2, d3, d4 ∈ E such that d1 + d2 = a1, d3 + d4 = a2, d1 + d3 = b1, d2 + d4 = b2. (e) We say that E fulfils the commutational Riesz Decomposition Property, (RDP1) for short, if for any a1, a2, b1, b2 ∈ E such that a1 + a2 = b1 + b2 there are d1, d2, d3, d4 ∈ E such that (i) d1 + d2 = a1, d3 + d4 = a2, d1 + d3 = b1, d2 + d4 = b2, and (ii) d2 com d3. (f) We say that E fulfils the strong Riesz Decomposition Property, (RDP2) for short, if for any a1, a2, b1, b2 ∈ E such that a1 + a2 = b1 + b2 there are d1, d2, d3, d4 ∈ E such that (i) d1 + d2 = a1, d3 + d4 = a2, d1 + d3 = b1, d2 + d4 = b2, and (ii) d2 ∧ d3 = 0. We have the implications (RDP2) ⇒ (RDP1) ⇒ (RDP) ⇒ (RDP0) ⇒ (RIP). 4 ANATOLIJ DVURE CENSKIJ The converse of any of these implications does not hold. For commutative effect algebras we have (RDP2) ⇒ (RDP1) ⇔ (RDP) ⇔ (RDP0) ⇒ (RIP). If in the above definitions of (RDP)'s we change E to G+, we have po-groups with the corresponding forms of the Riesz Decomposition Properties. According to [DvVe2, Thm 5.7], every pseudo effect algebra with (RDP)1 is an interval in some unital po-group with (RDP)1, so is any effect algebra with (RDP) in an Abelian po-group with (RDP), see see [Rav] or [DvPu, Thm 1.7.17]. Any effect algebra with (RDP) that is a lattice, or equivalently, with (RDP)2 is an MV-algebra, for a definition see e.g. [DvPu]. Any pseudo effect algebra with (RDP)2 is a so-called pseudo MV-algebra, see [GeIo], and the group G is an ℓ-group (= lattice ordered group). A signed measure on a pseudo effect algebra E is any mapping m : E → R such that m(a + b) = m(a) + m(b) provided a + b is defined in E. We have s(0) = 0 and s(a−) = s(a∼). If a signed measure m is positive, i.e., m(a) ≥ 0 for each a ∈ E, we call it a measure, and any normalized measure, i.e. a measure s such that s(1) = 1, is said to be a state. For any measure m, we have m(a) ≤ m(b) whenever a ≤ b. If m1 and m2 are two measures on E, then the signed measure m = m1 − m2 is said to be a Jordan signed measure. We denote by M(E), M+(E), S(E), and J (E) the sets of all signed measures, or measures, or states or Jordan signed measures on E. We recall that it can happen that M(E) = {0} = J (E). The set S(E) is convex, i.e., any convex combination s = λs1+(1−λ)s2, λ ∈ [0, 1], of two states s1 and s2 and λ ∈ [0, 1] is a state. If s cannot be expressed by a convex combination of two distinct states, it is called an extremal state. Let ∂eS(E) denote the set of all extremal states. On M(E) we introduce a weak topology: We say that a net of measures {mα} converges weakly to a measure m if limα mα(a) = m(a). Then S(E) is a convex compact Hausdorff space, and due to the Krein -- Mil'man Theorem, see [Goo, Thm 5.17], every state on E is a weak limit of a net of convex combinations of extremal states. If E is a pseudo effect algebra with (RDP), then S(E) is either empty or a non- void simplex, [Dvu3, Thm 5.1], for definition of a simplex and its basic properties, see [Goo, Chap 10]. For two signed measures m1 and m2, we write m1 ≤+ m2 if m1(a) ≤ m2(a) for each a ∈ E. The following important statement was proved in [Dvu3, Thm 3.5, Thm 3.6]: Theorem 2.1. Let E be a pseudo effect algebra with (RDP). Then J (E) is an Abelian Dedekind complete ℓ-group such that if {mi}i∈I is a nonempty system of mi! (x) =_{d(x1) + · · · + d(xn) : x = x1 + · · · + xn, x1, . . . , xn ∈ E} J (E) that is bounded above, and if d(x) =Wi mi(x) for all x ∈ E, then _i And if e(x) =Vi fi(x) for all x ∈ E, then ^i mi! (x) =^{e(x1) + · · · + e(xn) : x = x1 + · · · + xn, x1, . . . , xn ∈ E} for all x ∈ E. for all x ∈ E. Given m1, . . . , mn ∈ J (E), DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 5 n mi! (x) = sup{m1(x1) + · · · + mn(xn) : x = x1 + · · · + xn, x1, . . . , xn ∈ E}, _i=1 mi! (x) = inf{m1(x1) + · · · + mn(xn) : x = x1 + · · · + xn, x1, . . . , xn ∈ E}, n ^i=1 for all x ∈ E. 3. Faces of the State Space The present section describes the faces of the state space of a pseudo effect algebra with (RDP). Since our state space is a simplex, we know that if F is a closed face, then every simplex is a direct convex sum of F and its complementary face, [Goo, Thm 11.28]. However, not every face is closed, we present a general decomposition, Theorem 3.4, where a weaker form of the closedness, ∨-closedness, allows to obtain a unique decomposition of measures. This result will apply in the next section to obtain Yosida -- Hewitt and Lebesgue types of decomposition. A face of a convex set K is a convex subset F of K such that if x = λx1 + (1 − λ)x2 ∈ F for λ ∈ (0, 1), then also x1, x2 ∈ F. We note that if x ∈ K, then {x} is a face iff x ∈ ∂eK. For any X ⊆ K, there is the face generated by X. Due to [Goo, Prop 5.7], the face F generated by X is the set of those points x ∈ K for which there exists a positive convex combination λx + (1 − λ)y = z with y ∈ K and z belongs to the convex hull of X. If S(E) 6= ∅, then a state s ∈ S(E) belongs to the face generated by X if and only if s ≤+ αt for some positive constant α and some state t in the convex hull of X. In particular, the face of S(E) generated by a state s is the set of states s′ ∈ S(E) such that s′ ≤+ αs for some real number α > 0. Let s be a state on a pseudo effect algebra E. The kernel of s is the set Ker(s) := {x ∈ E : s(x) = 0}. Then Ker(s) is a normal ideal of E. We note that a subset I of a pseudo effect algebra is an ideal if (i) 0 ∈ I, (ii) a, b ∈ I and a + b ∈ E imply a + b ∈ I, and (iii) a ∈ E, b ∈ I and a ≤ b entail a ∈ I. An ideal I is normal, if a + I := {a + b ∈ E : b ∈ I} = I + a := {b + a ∈ E : b ∈ I} for any a ∈ E. Proposition 3.1. Let E be a pseudo effect algebra and let X be a subset of E. Then the set is a closed face of S(E). F = {s ∈ S(E) : X ⊆ Ker(s)} Proof. If F = ∅, then F is trivially a closed face. Assume F 6= ∅. If s = λs1 + (1 − λ)s2 for λ ∈ [0, 1] and for two states s1 and s2, then Ker(s) ⊆ Ker(s1) ∩ Ker(s2) which proves that F is a convex set. If {sα} is a net of states from F that converges weakly to a state s on E, then for each x ∈ X, 0 = limα sα(x) = s(x) so that F is closed. If now s = λs1 + (1 − λ)s2 ∈ F for λ ∈ (0, 1) and s1, s2 ∈ F, then for each x ∈ X we have 0 = s(x) = λs1(x) + (1 − λ)s2(x) so that s1(x) = s2(x) = 0 and therefore, s1, s2 ∈ F. (cid:3) Fi!′ \i =[i F ′ i (3.0). 6 ANATOLIJ DVURE CENSKIJ Let F be a face of a simplex K, and let F ′ be the union of those faces of K that are disjoint from F. According to [Goo, Prop 10.12], F ′ is a face of K and it is the largest face of K that is disjoint from F. If a point x ∈ K can be expressed as convex combinations x = α1x1 + α2x2 = β1y1 + β2y2 with x1, y1 ∈ F and x2, y2 ∈ F ′, then αi = βi for i = 1, 2 and xi = yi for those i such that αi > 0. The face F ′ is said to be a complementary face of F in K. It is possible to show that if F1 ⊆ F2 are two faces of S(E), then F ′ 2 ⊆ F ′ 1 and for a set {Fi} of faces, we have F =Ti Fi is a face, and Proposition 3.2. Let F be a face of the state space of a pseudo effect algebra E with (RDP) and let S(E) 6= ∅. Then its complementary face F ′ consists of all states s′ ∈ S(E) such that s′ ∧ s = 0 for every state s ∈ F. Equivalently, F ′ consists of all states s′ ∈ S(E) such that if s ∈ F is such that αs ≤+ s′ for some constant α ≥ 0, then α = 0. Proof. We know that S(E) is a simplex. According to [Goo, Prop 10.12], F ′ is the union of all faces of S(E) that are disjoint from F. If s′ ∈ F ′ and s ∈ F, and if s′ and s belong to mutually disjoint faces, then s′ ∧ s = 0, see [Dvu3, Prop 4.2]. Conversely, let s′ be a state on E such that s′ ∧ s = 0 for each s ∈ F. Then by [Dvu3, Prop 4.2], s′ and s belong to mutually disjoint faces. Let F (s′) and F (s) be the faces generated by s′ and s. Therefore, F (s′) ∩ F (s) = ∅. But F =Ss∈F F (s), so that F (s′) ∩ F =Ss∈F F (s′) ∩ F (s) = ∅ that gives s′ ∈ F ′ Now let s′ ∈ F ′ and let s ∈ F be such that αs ≤ s′ for some α ≥ 0. If α > 0, then s ≤ s′/α that s belongs to the face generated by s′ that is impossible, whence α = 0. Assume that F0 is the set of all those states s′ on E such that if αs ≤+ s′ for s ∈ F and α ≥ 0, then α = 0. We have seen that F ′ ⊆ F0. Let s′ ∈ F0 and s ∈ F If s′ ∧ s >+ 0, then s belongs to the face generated by s′, hence, s ≤+ ts′ for some t > 0. This gives s/t ≤+ s so that 1/t = 0 that is absurd. Hence, F0 ⊆ F ′. (cid:3) Let K1, . . . , Kn be convex subsets of K. We say that K is the direct convex sum i=1 Ki and every element x ∈ K can be uniquely expressed as a convex combination of some elements xi ∈ Ki, i = 1, . . . , n. That is, if of K1, . . . , Kn if (i) K equals the convex hull of Sn α1x1 + · · · + αnxn = β1y1 + · · · + βnyn are convex combinations of xi, yi ∈ Ki, for i = 1, . . . , n, then αi = βi for all i's and xi = yj for those i such that αi > 0. Theorem 3.3. Let E be a pseudo effect algebra with (RDP) and let S(E) 6= ∅. Let I be a normal ideal of E, and set F = {s ∈ S(E) : I ⊆ Ker(s)}. Then F is a closed face of S(E), and S(E) is the direct convex sum of F and its complementary face. DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 7 Proof. By Proposition 3.1, F is a closed face of S(E). If F is empty, then F ′ = S(E). Now assume F is non-void. Let J (E) be the set of all Jordan signed measures on E. By Theorem 2.1, J (E) is an Abelian Dedekind complete ℓ-group, and S(E) is a base for the positive cone J (E)+ = {αs : α ≥ 0, s ∈ S(E)}. We show that the set C = {αs : α ≥ 0, s ∈ S(E), s ∈ F } contains the supremum of any subset of C which is bounded above in J (E). Given a nonempty system of positive measures {mi}i from C that is bounded in J (E), let m =Wi mi. According to Theorem 2.1, m(x) = sup{d(x1) + · · · + d(xn) : x = x1 + · · · + xn}, x ∈ E, where d(x) = supi mi(x), x ∈ E. If x ∈ I, then x = x1 + · · · + xn entails x1, . . . , xn ∈ I so that d(xj ) = supi mi(xj) = 0 for each j = 1, . . . , n. Hence, m(x) = 0 and m ∈ C. According to [Goo, Prop 10.14], S(E) is the direct convex sum of F and its (cid:3) complementary face. Let F be a face of S(E), where E is a pseudo effect algebra with (RDP). We set V (F ) := {αs : α ≥ 0, s ∈ F }. It is a cone that is a subcone of M+(S), i.e., if (i) m1, m2 ∈ V (F ), then m1 + m2 ∈ V (F ), (ii) R+(F ) ⊆ V (F ) (ii) −V (F ) ∩ V (F ) = {0}. We say that V (F ) is ∨-closed if, for any chain {mi} from V (F ) bounded in M+(E), Wi mi ∈ V (F ). If F ′ is the complementary face of F, according to Proposition 3.2, m ∈ V (F ′) iff for any t ∈ V (F ) with t ≤+ m, we have t = 0. For an arbitrary cone V of S(E), we denote by V ♯ the set of all those measures t ∈ M+(E) that m ≤+ t for m ∈ V entails m = 0. The elements of V ♯ are said to be V -singular measures. Hence, V (F )♯ = V (F ′) for any face F of S(E) of a pseudo effect algebra E with (RDP). Theorem 3.4. Let F be a face of the state space S(E) of a pseudo effect algebra E with (RDP). If V (F ) is ∨-closed, then S(E) equals the direct convex sum of F and its complementary face F ′. In addition, every measure m on E can be uniquely decomposed as a sum m = m1 + m1 (3.1) of two measures such that m1 ∈ V (F ) and m2 ∈ V (F ′). Proof. Existence: Let m be a measure on E, and let Γ(m) = {m1 ∈ V (F ) : m1 ≤+ m}. Since the zero measure belongs to Γ(m), Γ(m) is non-void. Let {mi} be a chain of elements from Γ(m). The measure m is an upper bound for {mi}. By Theorem 2.1, m0 := Wi mi is a measure on E and by the hypotheses, m0 belongs to V (F ). It follows from the Zorn's Lemma that Γ(m) contains a maximal element m1 such that m1 ≤+ m. If we set m2 = m−m1, we show that m2 ∈ V (F ′). Let t ∈ V (F ) and let t ≤+ m2. Then m1 + t ≤+ m1 + m2 = m. Since m1 + t ∈ V (F ), the maximality of m1 in Γ(m) implies t = 0, and therefore, m2 ∈ V (F ). 8 ANATOLIJ DVURE CENSKIJ Uniqueness: Now let s be an arbitrary state on E. According to the first part of the present proof, we see that s can be decomposed in the convex form s = λ1s1 + λ2s2, (3.2) where s1 ∈ F and s2 ∈ F ′. This implies that S(E) is a direct convex sum of F and F ′. In particular, according to [Goo, Prop 10.12], this yields that the decomposition 1 ∈ F and (3.2) is unique, i.e. if s = α1s′ 2 ∈ F ′, then αi = λi for i = 1, 2, and if αi > 0 implies si = s′ s′ i. In particular, this implies that the decomposition in (3.1) is unique. (cid:3) 2 is another convex combination of s′ 1 + α2s′ 4. Decomposition of States on Pseudo Effect Algebras This section presents the main results of the paper: Yosida -- Hewitt type and Lebesgue type of decomposition of finitely additive measures and states. Using closed faces F , we can decompose any state s on a pseudo effect algebra with (RDP) in the unique form: s = λs1 + (1 − λ)s2, where s1 ∈ F and s2 ∈ F ′, where F ′ is the complementary face of F, see [Goo, Thm 11.28]. But not every face of S(E) is closed. E.g. the set of all σ-additive states on E is a face that is not necessarily closed. However, also for some such situations we show that S(E) can be the direct convex sum of the face F and its complementary face. A non-empty set X of a poset E is directed downwards (directed upwards), and we write D ↓ (D ↑), if for any x, y ∈ X there exists z ∈ D such that z ≤ x, z ≤ y (z ≥ x, z ≥ y). Two downwards directed sets {xt : t ∈ T } and {yt : t ∈ T } indexed by the same index set T are called downwards equidirected if, for any s, t ∈ T, there exists v ∈ T such that xv ≤ xs and xv ≤ xt as well as yv ≤ ys and yv ≤ yt. A similar definition holds for upwards directed sets. Let {an} be a sequence of elements of a pseudo effect E such that bn = a1+· · ·+an Let x ∈ E and D ⊆ E. We say that D ↑ x if D ↑ and x =W D. Dually we define D ↓ x, i.e. D ↓ and x =V D. exists in E for each n ≥ 1 and if a =Wn bn exists in E, we write a =Pn an. i.e. an ≤ an+1 for each n ≥ 1 and Wn an = a, then m(a) = limn m(an). A signed A signed measure m on a pseudo effect algebra E is σ-additive if, {an} ր a, measure m is σ-additive iff {an} ց 0 entails limn m(an) = 0. Now let E be an effect algebra (not a pseudo effect algebra). We say that a system {at}t∈T of elements of E is summable if, for each finite subset F ⊆ T, the element A signed measure m on an effect algebra is said to be completely additive if aF =Pt∈F at is defined in E. If there exists the element a = W{aF : F ⊆ T }, we called it the sum of the summable system {at}t∈T , and write a =Pt∈T at. m(a) =Pt∈T m(at) whenever a =Pt∈T at. We denote by J (E)σ, J (E)ca the sets of all σ-additive Jordan signed measures and completely additive Jordan signed measures on E, respectively. In the same way we define M+(E)σ, M+(E)ca and similarly for the states: S(E)σ and S(E)ca denote the systems of all σ-additive and completely additive states on E. Then S(E)ca ⊆ S(E)σ ⊆ S(E). Each of these sets can be empty. Moreover, S(E)σ and S(E)ca are also faces of S(E), and V (S(E)σ) = M+(E)σ and V (S(E)ca) = M+(E)ca. (4.1) Proposition 4.1. Let m1, . . . , mn be completely additive measures on an effect algebra E that satisfies (RDP). Then m = m1 ∨ · · · ∨ mn is also a completely additive measure on E. DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 9 The same is true if m1, . . . , mn are σ-additive measures on a pseudo effect algebra E with (RDP). Proof. Without loss of generality, we can assume that n = 2. Let a =Pt∈T at and let F be any finite subset of T and let aF =Pt∈F at. Given ǫ > 0, there is F0 such that, for each finite F ⊇ F0, m1(a − aF ) < ǫ/2 and m2(a − aF ) < ǫ/2. By Theorem 2.1, m(a − aF ) = sup{m1(x) + m2(y) : x + y = a − aF }. Since x ≤ a − aF and y ≤ a − aF , we have m1(x) ≤ m1(a − aF ) < ǫ/2 and m2(y) ≤ m2(a − aF ) < ǫ/2. Then m1(x) + m2(y) < ǫ, consequently m(a − aF ) ≤ ǫ. Hence, m(a) − m(aF ) ≤ ǫ and m is completely additive. (cid:3) Lemma 4.2. If {ft}t∈T ↑ f and {gt}t∈T ↑ g, where ft, gt, f, g ∈ R+ for all t ∈ T. Then If, in addition, {ft}t∈T and {gt}t∈T are upwards equidirected, then (fs + gt) = f + g, _s,t {ft + gt}t∈T ↑ f + g. (4.2) (4.3) Proof. We have fs + gt ≤ f + g for all s, t ∈ T. If fs + gt ≤ x for some x ∈ R, then fs ≤ x − gt so that f ≤ x − gt and hence, gt ≤ x − f so that g ≤ x − f and f + g ≤ x that gives, (4.2). Now assume {ft} and {gt} are upwards equidirected. It is clear that ft + gt ≤ f + g. The equidirectness entails that, for all indices s0, t0 ∈ T , there exists an index t such that fs0 ≤ ft and ft0 ≤ gt. Therefore, for all indices s0 and t0, fs0 + gt0 ≤ ft + gt ≤ f + g which by (4.2) gives (4.3). (cid:3) Lemma 4.3. Let E be a pseudo effect algebra with (RDP). If {mi} is a chain of measures in M+(E) that is bounded above, then for m0 =Wi mi we have m0(a) = sup mi(a), (4.4) for each a ∈ E. i Proof. We assert that if d(x) = Wi mi(x), x ∈ E, then d is additive, i.e., d(x + y) = d(x) + d(y) whenever x + y is defined in E. This follows from the fact that {mi(x)} ↑ d(x) and {mi(y)} ↑ d(y), are upwards equidirected because {mi} is a chain, and {mi(x) + mi(y)} = {mi(x + y)} ↑ (d(x) + d(y)) by (4.3). Since, d is a measure such that mi ≤+ d ≤+ m0, we conclude d = m0. (cid:3) Now we present a Yosida-Hewitt type of decomposition for measures on E. Theorem 4.4. Let E be an effect algebra with (RDP). Then every measure m on E can be uniquely expressed in the form (4.5) where m1 ∈ M+(E)ca and m2 is a finitely additive measure on E such that if t ∈ M+(E)ca, such t ≤+ m2, then t = 0. m = m1 + m2, In particular, every state s on E can be uniquely expressed as a convex combi- nation (4.6) where s1 is a completely additive state and s2 is a finitely additive state such that if αs′ ≤+ s2 for some completely additive state s′ on E and for some constant α ≥ 0, then α = 0. s = λ1s1 + λ2s2, 10 ANATOLIJ DVURE CENSKIJ Proof. First we show that M+(E)ca is a ∨-closed cone in M+(E). Thus let {mi} be a chain of completely additive measures on E that is bounded above by a finitely and m0 ≤+ m′. additive measure m′. By Theorem 2.1, there is m0 =Wi mi that is finitely additive We assert that m0 is completely additive. Let a = Pt∈T at exist in L. Then from the monotonicity of m0 we have m0(aF ) ≤ m0(a), where aF := Pt∈F at for Assume that m0(aF ) ≤ x for some x ∈ R+ and for any finite F. Then mi(aF ) ≤ x for any i and any F. The complete additivity of mi entails that mi(a) ≤ x for any i, so that by (4.4), m0(a) = supi mi(a) ≤ x. This gives any F finite subset of the index set T . m0(a) = sup m0(at), F Xt∈F consequently m0 ∈ M+(E)ca, and M+(E)ca is a ∨-closed cone of J (E). Now let m be an arbitrary finitely additive measure on E. Since all the conditions of Theorem 3.4 are satisfied, we obtain the unique decomposition m = m1 + m2 in question. Similarly we have (4.6). (cid:3) Theorem 4.5. Let E be a pseudo effect algebra with (RDP). Then every measure m on E can be uniquely expressed in the form (4.7) where m1 ∈ M+(E)σ and m2 is a finitely additive measure on E such that if t ∈ M+(E)σ, t ≥ 0, such t ≤+ m2, then t = 0. m = m1 + m2, In particular, every state s on E can be uniquely expressed as a convex combi- nation (4.8) where s1 is a σ-additive state and s2 is a finitely additive state such that if αs′ ≤+ s2 for some completely additive state s′ on E and for some constant α ≥ 0, then α = 0. s = λ1s1 + λ2s2, Proof. It follows the same steps as the proof of Theorem 4.4. (cid:3) Remark 4.6. Theorems 4.5 and 4.5 have been proved in [11, 12, 9, 17]. They are analogues of the classical Yosida -- Hewitt decomposition from [YoHe]. In [DeMo], the component m2 from Theorem 4.4 is said to be a weakly purely additive measure and that from Theorem 4.5 a purely additive measure. Now we present another Yosida -- Hewitt type decomposition for an analogue of complete additivity of measures for pseudo effect algebra. We say that a measure m on E is upwards continuous if {at} ↑ a entails {m(at)} ↑ m(a). A measure m is upwards continuous iff {at} ↓ 0 implies {m(at)} ↓ 0. For example, if E is an effect algebra, then m is completely additive whenever m is upwards continuous. Indeed, let m be an upwards continuous measure and let a = Pt∈T at. Given any finite subset F of indices we define aF = Pt∈F at. Then {aF }F is upwards directed and {aF } ↑ a, so that m(a) =Pt m(at). Theorem 4.7. Let E be a pseudo effect algebra with (RDP). Then every measure m on E can be uniquely expressed in the form m = m1 + m2, where m1 is an upwards continuous measure and m2 is a finitely additive measure on E such that if t is an upwards continuous measure with t ≤+ m2, then t = 0. DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 11 In particular, every state s on E can be uniquely expressed as a convex combi- nation s = λ1s1 + λ2s2, where s1 is an upwards continuous state and s2 is a finitely additive state such that if αs′ ≤+ s2 for some upwards continuous state s′ on E and for some constant α ≥ 0, then α = 0. Proof. Let M+(E)uc and S(E)uc be the sets of upwards continuous measures and states, respectively, on E. Then S(E)uc is a face and V (S(E)uc) = M+(E). We show that M+(E)uc is a ∨-closed cone. Indeed, let {mi} be a chain in M+(E)uc. Then supt m(at) ≤ m(a). By (4.4), for m =Wi mi, we have m(a) = supi mi(a) for each a ∈ E. Assume {at} ↑ a. Given ǫ > 0, there is i such that mi(a) > m(a) − ǫ/2. Since mi is upwards continuous, there is at such that mi(at) > mi(a) − ǫ/2. Then mi(at) ≥ m(a) − ǫ and by (4.4), m(at) > m(a) − ǫ. Therefore, m(a) = supt m(at). (cid:3) For the final desired result, we apply Theorem 3.4. In what follows, we present two types of the Lebesgue decomposition. Let m1 and m2 be measures on a pseudo effect algebra E. We say that (i) m1 is absolutely continuous with respect to m2, and we write m1 ≪ m2 if m2(a) = 0 implies m1(a) = 0 for a ∈ E. (ii) m1 is m2-continuous, and we write m1 ≪ǫ m2 provided given ǫ > 0, there is a δ > 0 such that m2(a) < δ yields m1(a) < ǫ. (iii) m1 ⊥ m2 if there is an element a ∈ E such that m2(a) = 0 = m1(a−). It is clear that m1 ≪ǫ m2 entails m1 ≪ m2. Theorem 4.8. Let E be a pseudo effect algebra with (RDP). Let t be a fixed measure on E. Then every measure m on E can be uniquely expressed in the form m = m1 + m2, (4.9) where m1 and m2 are finitely additive measures on E such that m1 ≪ǫ t and if m′ is any measure such that m′ ≪ǫ t and m1 ≤+ m2, then m′ = 0. Moreover, m2 ∧ t = 0. In particular, every state s on E can be uniquely expressed as a convex combi- nation of two states s1 and s1 on E, s = λ1s1 + λ2s2, (4.10) where s1 ≪ǫ t and if s′ is any state such that s′ ≪ǫ t and αs′ ≤+ s2, then α = 0. Proof. If t = 0, the statement is trivial. Suppose that t(1) > 0 and let F (t) = {s ∈ S(E) : s ≪ǫ t}. Then t0 = t/t(1) ∈ F (t) and F (t) is a non-empty face of S(E). Let C(t) := {s ∈ M+(E) : s ≪ǫ t}. Then C(t) is a nonempty cone that is a subcone of M+(E), and C(t) = V (F (t)). We claim that C(t) is ∨-closed. Let {mi} be a chain from C(t) that is bounded above by m′ ∈ M+(E). If we define d(x) = Wi mi(x), according to Lemma 4.3, we have d = m0 := Wi mi and {mi(1)} ↑ d(1). Since J (E) is an Abelian Dedekind complete ℓ-group, we have {m0 − mi} ↓ 0. Therefore, {(m0 − mi)(1)} = {m0(1) − mi(1)} ↓ 0. Thus given ǫ > 0, there is an index i0 such that, for each mi with mi ≥+ mi0, we have m0(1) − mi(1) < ǫ/2. Fix the index i0. Given ǫ > 0, there is δ > 0 such that t(a) < δ implies mi0 (a) < ǫ/2. Calculate, 12 ANATOLIJ DVURE CENSKIJ m0(a) = mi0 (a)+(m0 −mi0 )(a) ≤ mi0 (a)+(m0 −mi0 )(1) ≤ ǫ/2+(m0 −mi0 )(1) < ǫ that gives m0 ≪ǫ t and m0 ∈ C(t). Applying the general result from Theorem 3.4, we have the existence and unique- ness of (4.9). Now we show that m2 ∧ t = 0. Let k be a measure on E such that k ≤+ m2 and (cid:3) k ≤+ t. Then k ≪ǫ t, so that k = 0 and whence m2 ∧ t = 0. Theorem 4.9. Let E be a pseudo effect algebra with (RDP). Let t be a fixed measure on E. Then every measure m on E can be uniquely expressed in the form (4.11) where m1 and m2 are finitely additive measures on E such m1 ≪ t and if m′ is any measure on E such that m′ ≪ t and m′ ≤+ m2, then m′ = 0. m = m1 + m2, Proof. Let us define V (t) = {m ∈ M+(E) : m ≪ t} and C(t) := {s ∈ S(E) : s ≪ t}. Then F (t) is a face of S(E) and it generates V (t) = V (F (t)). Assume that {mi} is a chain from V (t) that is bounded in J (E). If we set d(x) =Wi mi(x) for each x ∈ E, then according to (4.4), d = m0 :=Wi mi. Therefore, if t(a) = 0, then mi(a) = 0 for each i so that m(a) = d(a) = 0 and m0 ∈ V (t). The existence and uniqueness of (4.11) follows from Theorem 3.4. (cid:3) Finally, we show a relation among two types of continuity of measures. Proposition 4.10. Let s1 and ss be two σ-additive measures on a σ-complete MV-algebra E. Then s1 ≪ s2 if and only if s1 ≪ǫ s2. Proof. It is clear that s1 ≪ǫ s2 entails s1 ≪ s2. Now let us suppose s1 ≪ s2 and let (ad absurdum) s1 6≪ǫ s2. Then there is an ǫ > 0 such that for each n ≥ 1, there is n=k ak. Then an an ∈ E such that s2(an) < 1/2n and s1(an) ≥ ǫ. Set a =V∞ n=1W∞ s2(a) ≤ s2(an ∨ an+1 ∨ · · · ) ≤ s2(ak) < 1/2n−1, ∞ Xk=n so that s2(a) = 0. On the other hand, s1(a) = limn s1(an ∨ an+1 ∨ · · · ) ≥ lim supn s1(an) ≥ ǫ that (cid:3) contradicts s1 ≪ s2. We say that a measure t on E is Jauch-Piron if t(a) = t(b) = 0 entails there is an element c ∈ E such that a, b ≤ c and m(c). Every measure on an MV-algebra is Jauch-Piron. Let E be a pseudo effect with (RDP). According to [Dvu, Thm 3.2], we say that an element a ∈ E is said to be central or Boolean, if a∧a− = 0. Then also a∧a∼ = 0 and a− = a∼. Moreover, for any x ∈ E, x ∧ a is defined in E, and x = (x ∧ a) + (x ∧ a−). Let C(E) be the set of all central elements of E. Then it is a Boolean algebra that is a subalgebra of E. If, in addition, E is monotone σ-complete, then C(E) is a Boolean σ-algebra, [Dvu, Thm 5.11]. We recall that a pseudo effect algebra E is said to be monotone σ-complete provided, for any sequence {an} from E such that an ≤ an+1 for each n ≥ 1, a =Wn an is defined in E. DECOMPOSITIONS OF MEASURES ON PSEUDO EFFECT ALGEBRAS 13 If a, b are central elements and t is a measure, then t(a)+ t(b) = t(a∧b)+ t(a∨b), so that if, in addition, t(a) = t(b) = 0, then t(a ∨ b) = 0. Now we present another Lebesgue type of decomposition for measures. For two measures m and t on E we write m ≪C t provided t(a) = 0 for a ∈ C(E) implies m(a) = 0. It is a weaker form of m ≪ t. Theorem 4.11. Let E be a monotone σ-complete pseudo effect algebra with (RDP) and let t be a σ-additive measure on E. Every σ-additive measure m on E can be uniquely decomposed in the form such that m1, m2 are σ-additive measures, m1 ≪C t, and m2 ⊥ t. In particular, every σ-additive measure s can be uniquely decompose in the form m = m1 + m2 (4.12) s = λs1 + (1 − λ)s2, (4.13) where s1 and s2 are σ-additive measures on E such that s1 ≪ t and s2 ⊥ t. Proof. Existence: Let Ker(t)C = {a ∈ C(E) : t(a) = 0}. The zero element 0 is central, so that 0 ∈ Ker(t)C . We order the elements of Ker(t)C by a (cid:22) b iff a, b ∈ Ker(t)C and m(a) ≤ m(b). Then (cid:22) is a partial order and now let {ai} be a chain of elements from Ker(t)C with respect to (cid:22), and let δ = supi m(ai). Then either there is an upper bound a of {ai} itself or there is a sequence {an} in {ai} such that m(an) < m(an+1) ր δ. Set a = Wn an; then a ∈ Ker(t)C and m(a) = limn m(an) = δ, so that a is an upper bound in Ker(t)C for the chain {ai}. Applying the Zorn Lemma, Ker(t)C contains a maximal element, say a0. Set m1(x) = m(x ∧ a− 0 ) and m2(x) = m(x ∧ a0) for each x ∈ E. Since C(E) is a Boolean σ-algebra, and x = (x ∧ a− 0 ) + (x ∧ a0) for each x ∈ E, we see that m1 and m2 are σ-additive measures on E such that m = m1 + m2. Since t(a0) = 0, then m2(a− 0 ∧ a0) = 0 so that m2 ⊥ t. We assert that m1 ≪C t. If not, there is an element a ∈ Ker(t)C such that m1(a) = m(a ∧ a− 0 ) > 0. Since a0 ≤ a0 ∨ (a ∧ a− 0 ) ∈ Ker(t)C which contradicts the maximality of a0. This proves that m1 ≪C t. 0 ) = m(a− Uniqueness: Let F (t)σ be the set of all σ-additive states s on E such that s ≪C t. Then F (t)σ is a face and due to Theorems 4.4 -- 4.5, if {mi} is a bounded chain from V (F (t)σ), then m0 =Wi mi is a σ-additive measure, and in view of (4.4), m0 ≪C t. This gives that V (F (t)σ) is a ∨-closed cone. Now let m′ be an arbitrary σ-additive measure from V (F (t)σ) such that m′ ≤+ m2. Then m′(a0) = 0 while t(a0) and m′(a− σ) so that by Theorem 3.4 we have that the decomposition (4.12) is unique. (cid:3) 0 ) = 0 so that m′ = 0. Therefore, m1 ∈ V (F (t)′ 0 ) ≤ m1(a− Corollary 4.12. Let X be any subset of a pseudo effect algebra with (RDP), and let F = {s ∈ S(E) : X ⊆ Ker(s)}. Every measure m on E can be uniquely decomposed in the form m = m1 + m2, where m1 ∈ V (F ) and m2 is V (F )-singular. In particular, every state s on E can be uniquely expressed as a convex combi- nation where s1 ∈ F and s2 ∈ F ′. m = λs1 + (1 − λ)s2, 14 ANATOLIJ DVURE CENSKIJ Proof. Since V (F ) = {m ∈ M+(E) : X ⊆ Ker(m)} is by (4.4) ∨-closed, the statements follow from Theorem 3.4. (cid:3) References [Alf] [AlSc] E.M. Alfsen, "Compact Convex Sets and Boundary Integrals", Springer-Verlag, Berlin, 1971. E.M. Alfsen, F.W. Schultz, "State Spaces of Operator Algebras", Birkhauser, Boston- Basel-Berlin, 2001. [Dvu0] A. Dvurecenskij, "Gleason's Theorem and Its Applications", Kluwer Academic Publisher, [Dvu] Dordrecht/Boston/London, 1993, 325+xv pp. A. Dvurecenskij, Central elements and Cantor -- Bernstein's theorem for pseudo-effect algebras, J. Austral. Math. Soc. 74 (2003), 121 -- 143. [Dvu1] A. Dvurecenskij, States on pseudo MV-algebras, Studia Logica 68 (2001), 301 -- 327. [Dvu2] A. Dvurecenskij, Pseudo MV-algebras are intervals in ℓ-groups, J. Austral. Math. Soc. 72 (2002), 427 -- 445. [Dvu3] A. Dvurecenskij, The lattice and simplex structure of states on pseudo effect algebras, submitted. http://arxiv.org/submit/103087 [Dvu4] A. Dvurecenskij, Every state on interval effect algebra is integral, J. Math. Phys. 51 (2010), 083508 -- 12. DOI: 10.1063/1.3467463 [DvPu] A. Dvurecenskij, S. Pulmannov´a, "New Trends in Quantum Structures", Kluwer Acad. Publ., Dordrecht, Ister Science, Bratislava, 2000. [DvVe1] A. Dvurecenskij, T. Vetterlein, Pseudoeffect algebras. I. Basic properties, Inter. J. Theor. Phys. 40 (2001), 685 -- 701. [DvVe2] A. Dvurecenskij, T. Vetterlein, Pseudoeffect algebras. II. Group representation, Inter. J. Theor. Phys. 40 (2001), 703 -- 726. [DDP] A. Dvurecenskij, P. de Lucia, E. Pap, On a decomposition theorem and its applications, Math. Japonica 44 (1996), 145 -- 164. [DeMo] P. De Lucia, P. Morales, Decomposition theorems in Riesz spaces, Proc. Amer. Math. Soc. 120 (1994), 193 -- 202. [FoBe] D.J. Foulis, M.K. Bennett, Effect algebras and unsharp quantum logics, Found. Phys. 24 (1994), 1325 -- 1346. [GeIo] G. Georgescu, A. Iorgulescu, Pseudo-MV algebras, Multi. Val. Logic 6 (2001), 95 -- 135. [Goo] K.R. Goodearl, "Partially Ordered Abelian Groups with Interpolation", Math. Surveys and Monographs No. 20, Amer. Math. Soc., Providence, Rhode Island, 1986. P.R. Halmos, "Measure Theory", Springer-Verlag, New York, Heidelberg, Berlin, 1988. K. Ravindran, On a structure theory of effect algebras, PhD thesis, Kansas State Univ., Manhattan, Kansas, 1996. [Hal] [Rav] [Rut1] G.T. Ruttimann, Decomposition of cone of measures, Atti Sem. Mat. Fis. Univ. Modena 38 (1990), 109 -- 121. [Rut2] G.T. Ruttimann, Weakly purely additive measures, Canad. J. Math. 46 (1994), 872 -- 885. [YoHe] K. Yosida, E. Hewitt,Finitely additive measures, Trans. Amer. Math. Soc. 72 (1952), 44 -- 66.
1611.02584	1	1611	2016-11-08T16:10:05	Local affine selections of convex multifunctions	[ "math.FA" ]	It is well known that not every convex multifunction admits an affine selection. One could ask whether there exists at least local affine selection. The answer is positive in the finite-dimensional case. The main part of this note consists of two examples of non-existence of local affine selections of convex multifunctions defined on certain infinite-dimensional Banach spaces.	math.FA	math	Manuscript LOCAL AFFINE SELECTIONS OF CONVEX MULTIFUNCTIONS SZYMON W ASOWICZ ABSTRACT. It is well known that not every convex multifunction admits an affine selection. One could ask whether there exists at least local affine selection. The answer is positive in the finite-dimensional case. The main part of this note consists of two examples of non-existence of local affine selections of convex multifunctions defined on certain infinite-dimensional Banach spaces. 1. INTRODUCTION Given two non-void sets X and Y , a map F : X → 2Y is called a multifunc- tion or a set-valued function. A (single-valued) function f : X → Y is a selection for F , if f (x) ∈ F (x) for all x ∈ X. There is a plethora of results concern- ing selections of various kinds, with the Micheal Selection Principle concerning lower semi-continuous selections and the Kuratowski -- Ryll-Nardzewski Selection Principle concerning measurable selections as probably the most prominent ones. More recent results connected with Michael Selection Principle were established by Zippin [12]. When X and Y carry a vector-space structure, it is natural to study affine selec- tions or, at least local affine selections for multifunctions F : X → 2Y , which are the objective of this note. This topic was investigated (among others) by A. Lazar [6], A. Smajdor and W. Smajdor [10], E. Behrends and K. Nikodem [3], M. Balaj and K. Nikodem [2] and the present author (cf. [11]). We denote by n(X) the family of all non-empty subsets of a set X. Now, if X, Y are (real) vector spaces and D ⊂ X is a convex set, then the multifunction F : D → n(Y ) is said to be convex, if (1) for any x, y ∈ D and t ∈ [0, 1]. When the reversed inclusion is stipulated, F is then called concave. Of course, the notation A + B and tA is meant in the Minkowski sense, i.e., A + B = {a + b : a ∈ A, b ∈ B} and tA = {ta : a ∈ A} for any t ∈ R. Observe that a single-valued function f : D → Y is convex (as a multifunction, i.e., f (x) is identified with a singleton {f (x)}) if and only if f is affine, which means that tF (x) + (1 − t)F (y) ⊂ F(cid:0)tx + (1 − t)y(cid:1) 6 1 0 2 v o N 8 ] . A F h t a m [ 1 v 4 8 5 2 0 . 1 1 6 1 : v i X r a tf (x) + (1 − t)f (y) = f(cid:0)tx + (1 − t)y(cid:1) It is easy to see that a multifunction F is convex if and only if its graph (cid:0)x, y ∈ D, t ∈ [0, 1](cid:1). Gr F = {(x, y) : x ∈ D, y ∈ F (x)} Date: November 08, 2016. 2010 Mathematics Subject Classification. 54C60, 54C65. Key words and phrases. multifunction, selection, convexity, extension of a function, Cech -- Stone compactification. 1 2 S. W ASOWICZ is a convex subset of X × Y . Moreover, if F is convex, then F (x) is a convex subset of Y for any x ∈ D. Indeed, if y1, y2 ∈ F (x) and t ∈ [0, 1], then by (1) we get ty1 + (1 − t)y2 ∈ tF (x) + (1 − t)F (x) ⊂ F (x). The condition (2) (cid:16)tF (x) + (1 − t)F (y)(cid:17) ∩ F(cid:0)tx + (1 − t)y(cid:1) 6= ∅ seems to be the weakest one to guarantee the existence of an affine selection for the multifunction F . Indeed, if F (x) = {f (x)}, where f : X → Y is affine, the above intersection is a singleton(cid:8)f(cid:0)tx + (1 − t)y(cid:1)(cid:9). It is proved in [11, Theorem 1] that the multifunction F mapping a real interval I into the family of all compact intervals in R, admits an affine selection if and only if the condition (2) is satisfied. In particular, if either F is convex or concave, then F admits an affine selection. One could ask whether a convex multifunction defined on more general domain admits an affine selection. There is a number of results going in this direction. One of the versions of the classical Hahn -- Banach Separation Theorem guarantees the existence of the linear functional separating two convex subsets of a topological vector space. It could be easily utilised to prove the existence of a linear (and hence afine) selection of the certain convex multifunction. Since the problem of extending functions is strongly related to the problem of the existence of selections of multifunctions, we notice that Pełczy´nski in his PhD dissertation [9] dealt with linear versions of the classical Tietze -- Urysohn theorem (and extended further the classical Borsuk -- Dugundji theorem). It is worth mentioning that Edwards [5] proved in 1965 the following separation theorem: Theorem 1. Let X be a Choquet simplex, f : X → [−∞, ∞) a convex upper semicontinuous function and let g : X → (−∞, ∞] be a concave lower semicon- tinuous function such that f 6 g on X. Then there exists a continuous affine function a : X → R such that f 6 a 6 g on X. This result, read in the context of multifunctions, states that the lower semicon- tinuous convex set-valued function defined on a Choquet simplex, whose values are compact intervals, admits a continuous affine selection. This multivalued ver- sion of Theorem 1 was extended in 1968 by Lazar (cf. [6, Theorem 3.1]) to more general codomains. Theorem 2. Let ϕ : X → 2E be a lower semicontinuous affine mapping from a Choquet simplex X to a Fréchet space E that takes non-empty closed values. Then there exists a continuous affine mapping h : X → E such that h(x) ∈ ϕ(x) for each x ∈ X. In fact, Edwards proved his result in a form of the necessary and sufficient con- dition for X to be a Choquet simplex. This means that if a convex set X is not a simplex, one could find two functions f, g (as considered in Theorem 1), which cannot be separated by the continuous affine function. Hence, in general, a convex multifunction defined on a convex subset of a vector space need not to admit the LOCAL AFFINE SELECTIONS OF CONVEX MULTIFUNCTIONS 3 affine selection. Let us have a look at the well known example due to Olsen [8] (see also Nikodem [7, Remark 1]). Consider the square D = {(x, y) ∈ R2 : x + y 6 1} and the simplex S ⊂ R3 with vertices (−1, 0, 0), (1, 0, 0), (0, −1, 1), (0, 1, 1). Observe that S is a graph of a convex multifunction F : D → n(R) (whose values are compact intervals) with no affine selection. Nevertheless, locally it is possible to put a piece of a plane into S. It means that F admits a local affine selection at every x0 ∈ Int D. We develop this observation in the next section. A. Smajdor and W. Smajdor proved in [10, Theorem 6] that if F is defined on a cone with the cone-basis in a (real) vector space and F takes the non-empty, closed (and necessarily convex) values in a (real) locally convex space, then F admits an affine selection. 2. CONVEX MULTIFUNCTIONS WITH LOCAL SELECTIONS Let X be a topological vector space and let D be a non-empty, convex subset of X with non-empty interior. Moreover, let Y be a real vector space. A multifunction F : D → n(Y ) admits a local affine selection at a point x0 ∈ Int D, if there exist an open neighourhood U ⊂ D of x0 and an affine function f : X → Y such that f (x) ∈ F (x) for every x ∈ U. The following finite-dimensional version of Edwards' theorem is quite easy to obtain. Lemma 3. Let S be an n-dimensional simplex in Rn and Y be a (real) vector space. Any convex multifunction F : S → n(Y ) admits the affine selection. Proof. Let a0, . . . , an ∈ Rn be the vertices of S and let us choose yi ∈ F (ai) (i = 0, . . . , n). Then there exists (the unique) affine function f : Rn → Y such that f (ai) = yi (i = 0, . . . , n). Take x ∈ S. Expressing x as a convex combination of a0, . . . , an (with coefficients λ0, . . . , λn > 0, λ0 + . . . + λn = 1) we get Since (ai, yi) ∈ Gr F , by convexity of this graph we arrive at λiai! = n n λif (ai) = λiyi. Xi=0 f (x) = f n Xi=0 (cid:0)x, f (x)(cid:1) = n Xi=0 λiai, n Xi=0 Xi=0 λiyi! = n Xi=0 λi(ai, yi) ∈ Gr F, whence f (x) ∈ F (x), x ∈ X and the proof is complete. (cid:3) The above lemma allows us to prove the existence of local affine selections in the finite-dimensional case. Theorem 4. Let D ⊂ Rn be a convex set with a non-empty interior and Y be a (real) vector space. Any convex multifunction F : D → n(Y ) admits a local affine selection at every interior point of D. Proof. Let x0 ∈ Int D. Then x0 is the interior point of some n-dimensional sim- plex S ⊂ D. By Lemma 3 there exists the affine function f : Rn → Y such that f (x) ∈ F (x) for any x ∈ S. In particular, f (x) ∈ F (x) for any x ∈ U = Int S and f is a desired local affine selection of F . (cid:3) 4 S. W ASOWICZ 3. CONVEX MULTIFUNCTIONS WITHOUT LOCAL SELECTIONS In the infinite-dimensional case the problem of the existence of local affine se- lection looks completely different. Namely, there are convex multifunctions with no local affine selection. The following observations are due to Tomasz Kania (Warwick) who has kindly permitted us to include them here. Let X be a closed linear subspace of a Banach space Y . In the light of the Hahn -- Banach theorem, the multifunction F : X ∗ → 2Y ∗ given by (3) F (f ) = {g ∈ Y ∗ : gX = f and kgk = kf k} (f ∈ X ∗) assumes always non-void values. Certainly, F (f ) is convex and weak*-compact for each f ∈ X ∗. It is easy to prove that F is a convex multifunction. Proposition 5. Suppose that X is a closed linear subspace of a Banach space Y such that X ∗ does not embed isometrically into Y ∗. Then F , as defined by (3), admits no local affine selection. Proof. Assume contrapositively that there exists an open neighbourhood U of the origin such that F admits an affine selection ϕ say, when restricted to U. In partic- ular, kϕ(g)k = kgk for all g ∈ U; thus ϕ is isometric. Denote by T the affine map E∗ → F ∗ that extends ϕ. As T 0 = 0, T is a linear, isometric embedding of X ∗ into Y ∗. (cid:3) Remark 6. The hypotheses of Proposition 5 are easily met when Y = C[0, 1]. Indeed, by the Banach -- Mazur theorem, C[0, 1] contains isometric copies of all separable Banach spaces. The dual space of C[0, 1] is isometric to L1(µ) for some measure µ and this, in turn, prevents many Banach spaces to embed into it (cf. [1, proof of Proposition 4.3.8]). For instance take X = ℓ1. Then X ∗ ∼= ℓ∞ which contains isometrically all In this case it is plain that any isometric copy of ℓ1 separable Banach spaces. inside of Y = C[0, 1] meets the hypotheses of Proposition 5. Remark 7. When the hypotheses of Proposition 5 are met, the mutlifunction is not lower semicontinuous. Indeed, otherwise by Lazar's theorem (Theorem 2 in this note) it would have admitted an affine selection. Let βN denote the Cech -- Stone compactification of the discrete space of natural numbers. Set (4) F (f ) = {g ∈ C(βN) : gβN\N = f and kgk = kf k} (f ∈ C(βN \ N)). Then by the Tietze -- Urysohn theorem, F (f ) is non-empty for every f ∈ C(βN\N). (It is also closed and convex.) Certainly the multifunction F is convex. Proposition 8. The multifunction F given by (4) does not admit a local affine selection. Proof. Arguing as in the proof of Proposition 5, we would get a linear embedding of C(βN \ N) into C(βN), however it is known that no such operator exists. (For example, the former space does not have a strictly convex renorming but the latter does; the possibility of finding a strictly convex renorming passes to subspaces, cf. [4]; here we use the fact that C(βN \ N) is isometric to ℓ∞/c0.) (cid:3) LOCAL AFFINE SELECTIONS OF CONVEX MULTIFUNCTIONS 5 REFERENCES [1] Fernando Albiac and Nigel J. Kalton. Topics in Banach space theory, volume 233 of Graduate Texts in Mathematics. Springer, New York, 2006. [2] Mircea Balaj and Kazimierz Nikodem. Remarks on Bárány's theorem and affine selections. Discrete Math., 224(1-3):259 -- 263, 2000. [3] Ehrhard Behrends and Kazimierz Nikodem. A selection theorem of Helly type and its applica- tions. Studia Math., 116(1):43 -- 48, 1995. [4] Jean Bourgain. l∞/c0 has no equivalent strictly convex norm. Proc. Amer. Math. Soc., 78(2):225 -- 226, 1980. [5] David Albert Edwards. Séparation des fonctions réelles définies sur un simplexe de Choquet. C. R. Acad. Sci. Paris, 261:2798 -- 2800, 1965. [6] Aldo J. Lazar. Spaces of affine continuous functions on simplexes. Trans. Amer. Math. Soc., 134:503 -- 525, 1968. [7] Kazimierz Nikodem. A characterization of midconvex set -- valued functions. Acta Univ. Car- olin. Math. Phys., 30(2):125 -- 129, 1989. 17th Winter School on Abstract Analysis (Srní, 1989). [8] Gunnar Hans Olsen. On simplices and the Poulsen simplex. In Functional analysis: surveys and recent results, II (Proc. Second Conf. Functional Anal., Univ. Paderborn, Paderborn, 1979), volume 68 of Notas Mat., pages 31 -- 52. North-Holland, Amsterdam-New York, 1980. [9] Aleksander Pełczy´nski. Linear extensions, linear averagings, and their applications to linear topological classification of spaces of continuous functions. Dissertationes Math. Rozprawy Mat., 58:92, 1968. [10] Andrzej Smajdor and Wilhelmina Smajdor. Affine selections of convex set-valued functions. Aequationes Math., 51(1-2):12 -- 20, 1996. [11] Szymon W asowicz. On affine selections of set -- valued functions. J. Appl. Anal., 1(2):173 -- 179, 1995. [12] Mori Zippin. The embedding of Banach spaces into spaces with structure. Illinois J. Math., 34(3):586 -- 606, 1990. DEPARTMENT OF MATHEMATICS, UNIVERSITY OF BIELSKO-BIAŁA, WILLOWA 2, 43-309 BIELSKO-BIAŁA, POLAND E-mail address: [email protected]
1711.10702	1	1711	2017-11-29T06:46:28	A new variation on statistical ward continuity	[ "math.FA" ]	A real valued function defined on a subset $E$ of $\mathbb{R}$, the set of real numbers, is $\rho$-statistically downward continuous if it preserves $\rho$-statistical downward quasi-Cauchy sequences of points in $E$, where a sequence $(\alpha_{k})$ of real numbers is called ${\rho}$-statistically downward quasi-Cauchy if $\lim_{n\rightarrow\infty}\frac{1}{\rho_{n} }\|\{k\leq n: \Delta \alpha_{k} \geq \varepsilon\}\|=0 $ for every $\varepsilon>0$, in which $(\rho_{n})$ is a non-decreasing sequence of positive real numbers tending to $\infty$ such that $\limsup _{n} \frac{\rho_{n}}{n}<\infty $, $\Delta \rho_{n}=O(1)$, and $\Delta \alpha _{k} =\alpha _{k+1} - \alpha _{k}$ for each positive integer $k$. It turns out that a function is uniformly continuous if it is $\rho$-statistical downward continuous on an above bounded set.	math.FA	math	A NEW VARIATION ON STATISTICAL WARD CONTINUITY HUSEYIN CAKALLI MALTEPE UNIVERSITY, GRADUATE SCHOOL OF SCIENCE AND ENGINEERING, DEPARTMENT OF MATHEMATICS, MALTEPE, ISTANBUL-TURKEY Abstract. A real valued function defined on a subset E of R, the set of real numbers, is ρ-statistically downward continuous if it preserves ρ-statistical downward quasi-Cauchy sequences of points in E, where a sequence (αk) of real numbers is called ρ-statistically downward quasi-Cauchy if limn→∞ 1 ρn {k ≤ n : ∆αk ≥ ε} = 0 for every ε > 0, in which (ρn) is a non-decreasing sequence of positive real numbers tending to ∞ such that lim supn ρn n < ∞, ∆ρn = O(1), and ∆αk = αk+1 − αk for each positive integer k. It turns out that a function is uniformly continuous if it is ρ-statistical downward continuous on an above bounded set. 1. Introduction The concept of continuity and any concept involving continuity play a very im- portant role not only in pure mathematics but also in other branches of sciences involving mathematics especially in computer science, information theory, econom- ics, and biological science. The notion of strongly lacunary convergence or Nθ convergence was introduced, and studied by Freedman, Sember, and M. Raphael in [42] in the sense that a sequence (αk) of points in R is strongly lacunary convergent or Nθ convergent to an L ∈ R, which is denoted by Nθ − lim αk = L, if limr→∞ Pk∈Ir αk − L = 0, where Ir = (kr−1, kr], and k0 6= 0, hr := kr − kr−1 → ∞ as r → ∞ and θ = (kr) 1 hr is an increasing sequence of positive integers (see also [45], and [46]). In [43] Fridy and Orhan introduced the concept of lacunary statistically convergence in the sense that a sequence (αk) of points in R is called lacunary statistically convergent, or Date: September 27, 2018. 2010 Mathematics Subject Classification. Primary: 40A05; Secondaries: 40D25; 40A35; 40A30; 26A15. Key words and phrases. Sequences, series, summability. 1 2 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY Sθ-convergent, to an element L of R if limr→∞ 1 hr {k ∈ Ir : αk − L ≥ ε} = 0 for every positive real number ε where Ir = (kr−1, kr] and k0 = 0, hr : kr − kr−1 → ∞ as r → ∞ and θ = (kr) is an increasing sequence of positive integers (see also [44], [40], [41], [27], [39], and [5]). Modifying the definitions of a forward Cauchy sequence introduced in [58] and [10], Palladino ([54]) gave the concept of downward half Cauchyness in the following way: a real sequence (αk) is called downward half Cauchy if for every ε > 0 there exists an n0 ∈ N so that αk − αm < ε for m ≥ n ≥ n0 (see also [58]). A sequence (αk) of points in R, the set of real numbers, is called statistically downward quasi- Cauchy if limn→∞ 1 n {k ≤ n : αk+1 − αk ≥ ε} = 0 for every ε > 0 ([19]). Recently, many kinds of continuities were introduced and investigated, not all but some of them we recall in the following: slowly oscillating continuity ([13], [60]), quasi-slowly oscillating continuity ([37]), ward continuity ([23], [4]), δ-ward conti- nuity ([14]), p-ward continuity ([20]), statistical ward continuity ([16]), ρ-statistical ward continuity ([6], lacunary statistical δ2-ward continuity ([65]), arithmetic con- tinuity ([24], [25]) Abel continuity ([28]), which enabled some authors to obtain conditions on the domain of a function for some characterizations of uniform conti- nuity in terms of sequences in the sense that a function preserves a certain kind of sequences (see [60, Theorem 6], [4, Theorem 1 and Theorem 2], [37, Theorem 2.3]). The aim of this paper is to introduce, and investigate the concepts of ρ-statistical downward continuity and ρ-statistical downward compactness. 2. ρ-statistical downward compactness The definition of a Cauchy sequence is often misunderstood by the students who first encounter it in an introductory real analysis course. In particular, some fail to understand that it involves far more than that the distance between successive terms is tending to zero. Nevertheless, sequences which satisfy this weaker property are interesting in their own right. In [4] the authors call them "quasi-Cauchy", while they were called "forward convergent to 0" sequences in [23] (see also [66]), where a sequence (αk) is called quasi-Cauchy if for given any ε > 0, there exists an integer K > 0 such that k ≥ K implies that αk+1 − αk < ε. A subset E of R is compact if and only if any sequence in E has a convergent subsequence whose limit is in E. Boundedness of a subset E of R coincides with that any sequence of points in E has A NEW VARIATION ON STATISTICAL WARD CONTINUITY 3 either a Cauchy subsequence, or a quasi-Cauchy subsequence. What is the case for above boundedness? ρ-statistical downward quasi Cauchy sequences provide with the answer. Weakening the condition on the definition of a ρ-statistical quasi-Cauchy se- quence, omitting the absolute value symbol, i.e. replacing αk+1 − αk < ε with αk+1 − αk < ε in the definition of a ρ-statistical quasi-Cauchy sequence given in [6], we introduce the following definition. Definition 1. A sequence (αk) of points in R is called ρ-statistically downward quasi-Cauchy if for every ε > 0. lim n→∞ 1 ρn {k ≤ n : αk+1 − αk ≥ ε} = 0 ∆ρ− will denote the set of all ρ-statistically downward quasi-Cauchy sequences of points in R. As an example, the sequence (ρn) is a ρ-statistically downward quasi- Cauchy sequence. Any ρ-statistically convergent sequence is ρ-statistically down- ward quasi-Cauchy. Any ρ-statistically quasi-Cauchy sequence is ρ-statistically downward quasi-Cauchy, so any slowly oscillating sequence is ρ-statistically down- ward quasi-Cauchy, so any Cauchy sequence is, so any convergent sequence is. Any downward Cauchy sequence is ρ-statistically downward quasi-Cauchy. Now we give some interesting examples that show importance of the interest. Example 1. Let n be a positive integer. In a group of n people, each person selects at random and simultaneously another person of the group. All of the selected persons are then removed from the group, leaving a random number n1 < n of people which form a new group. The new group then repeats independently the selection and removal thus described, leaving n2 < n1 persons, and so forth until either one person remains, or no persons remain. Denote by αn the probability that, at the end of this iteration initiated with a group of n persons, one person remains. Then the sequence (α1, α2, , αn, ...) is a ρ-statistically downward quasi-Cauchy sequence (see also [61]). Example 2. In a group of k people, k is a positive integer, each person selects independently and at random one of three subgroups to which to belong, resulting in three groups with random numbers k1, k2, k3 of members; k1 + k2 + k3 = k. 4 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY Each of the subgroups is then partitioned independently in the same manner to form three sub subgroups, and so forth. Subgroups having no members or having only one member are removed from the process. Denote by αk the expected value of the number of iterations up to complete removal, starting initially with a group of k people. Then the sequence (α1, α2 n , ...) is a bounded non-convergent 2 , α3 3 , ..., αn ρ-statistically downward quasi-Cauchy sequence ([48]). It is well known that a subset of R is compact if and only if any sequence of points in E has a convergence subsequence, whose limit is in E. By using this idea in the definition of sequential compactness, now we introduce a definition of ρ-statistically downward compactness of a subset of R. Definition 2. A subset E of R is called ρ-statistically downward compact if any sequence of points in E has a ρ-statistically downward quasi-Cauchy subsequence. First, we note that any finite subset of R is ρ-statistically downward compact, the union of finite number of ρ-statistically downward compact subsets of R is ρ- statistically downward compact, and the intersection of any family of ρ-statistically downward compact subsets of R is ρ-statistically downward compact. Further- more any subset of a ρ-statistically downward compact set is ρ-statistically down- ward compact, any compact subset of R is ρ-statistically downward compact, any bounded subset of R is ρ-statistically downward compact, and any slowly oscillating compact subset of R is ρ-statistically downward compact (see [13] for the defini- tion of slowly oscillating compactness). The sum of finite number of ρ-statistically downward compact subsets of R is ρ-statistically downward compact. Any bounded above subset of R is ρ-statistically downward compact. These observations suggest to us giving the following result. Theorem 1. A subset of R is ρ-statistically downward compact if and only if it is bounded above. Proof 1. Let E be a bounded above subset of R. If E is also bounded below, then it follows from [16, Lemma 2] and [17, Theorem 3] that any sequence of points in E has a quasi Cauchy subsequence which is also ρ-statistically downward quasi-Cauchy. If E is unbounded below, and (αn) is an unbounded below sequence of points in E, A NEW VARIATION ON STATISTICAL WARD CONTINUITY 5 then for k = 1 we can find an αn1 less than 0. For k=2 we can pick an αn2 such that αn2 < −ρ2 + αn1 . We can successively find for each k ∈ N an αnk such that αnk+1 < −ρk+1 + αnk . Then αnk+1 − αnk < −ρk+1 for each k ∈ N. Therefore we see that lim n→∞ 1 ρn {k ≤ n : αnk+1 − αnk ≥ ε} = 0 for every ε > 0. Conversely, suppose that E is not bounded above. Pick an element α1 of E. Then we can choose an element α2 of E such that α2 > ρ2 + α1. Similarly we can choose an element α3 of E such that α3 > ρ3 + α2. We can inductively choose αk satisfying αk+1 > ρk + αk for each positive integer k. Then the sequence (αk) does not have any ρ-statistically downward quasi-Cauchy subsequence. Thus E is not ρ-statistically downward compact. This contradiction completes the proof. It follows from the above theorem that if a closed subset E of R is ρ-statistically downward compact, and −A is ρ-statistically downward compact, then any se- quence of points in E has a (Pn, s)-absolutely almost convergent subsequence (see [36], [52], [62], [3], [64], [63], and [65]). Corollary 1. A subset of R is ρ-statistically downward compact if and only if it is statistically downward compact. Proof 2. The proof of this theorem follows from Theorem 1 and [19, Theorem 3.3], so it is omitted. Corollary 2. A subset of R is ρ-statistically downward compact if and only if it is lacunary statistically downward compact. Proof 3. The proof of this theorem follows from Theorem 1 and [30, Theorem 1.9], so is omitted. Corollary 3. A subset of R is ρ-statistically downward compact if and only if it is downward compact. Proof 4. The proof of this theorem follows from Theorem 1 and [21, Theorem 2.8 ], so is omitted. Corollary 4. A subset A of R is bounded if and only if the sets A and −A are ρ-statistically downward compact. 6 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY Proof 5. The proof of this theorem follows from the fact that a subset A of R is bounded if and only if the sets A and −A are bounded above, so is omitted. 3. ρ-statistical downward continuity A real valued function f defined on a subset of R is ρ-statistically continuous, or Sρ-continuous if for each point ℓ in the domain, Sρ−limn→∞ f (αk) = f (ℓ) whenever Sρ − limn→∞ αk = ℓ ([6]). This is equivalent to the statement that (f (αk)) is a convergent sequence whenever (αk) is. This is also equivalent to the statement that (f (αk)) is a Cauchy sequence whenever (αk) is Cauchy provided that the domain of the function is complete. These known results for ρ-statistically-continuity and continuity for real functions in terms of sequences might suggest to us introducing a new type of continuity, namely, ρ-statistically-downward continuity, weakening the condition on the definition of a ρ-statistically ward continuity, omitting the absolute value symbol, i.e. replacing "αk+1 − αk" with αk+1 − αk" in the definition of ρ- statistically ward continuity given in [6]. Definition 3. A real valued function f is called ρ-statistically downward contin- uous, or S− ρ -continuous on a subset E of R if it preserves ρ-statistically down- ward quasi-Cauchy sequences, i.e. the sequence (f (αk)) is ρ-statistically-downward quasi-Cauchy whenever (αk) is a ρ-statistically-downward quasi-Cauchy sequence of points in E, writing in symbols, limn→∞ 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} = 0, whenever (αk) is a sequence of points in E such that limn→∞ 1 ρn {k ≤ n : αk+1 − αk ≥ ε} = 0 for every positive real number ε. It should be noted that ρ-statistically-downward continuity cannot be given by any A-continuity in the manner of [11]. We see that the composition of two ρ- statistically-downward continuous functions is ρ-statistically-downward continuous, and for every positive real number c, cf is ρ-statistically-downward continuous, if f is ρ-statistically-downward continuous. We see in the following that the sum of two ρ-statistically-downward continuous functions is ρ-statistically-downward continuous A NEW VARIATION ON STATISTICAL WARD CONTINUITY 7 Proposition 1. If f and g are ρ-statistically-downward continuous functions, then f + g is ρ-statistically-downward continuous. Proof 6. Let f , g be ρ-statistically-downward continuous functions on a subset E of R. To prove that f + g is ρ-statistically-downward continuous on E, take any ρ-statistically-downward quasi-Cauchy sequence (αk) of points in E. Then (f (αk)) and (g(αk)) are ρ-statistically-downward quasi-Cauchy sequences. Let ε > 0 be given. Since (f (αk)) and (g(αk)) are ρ-statistically-downward quasi-Cauchy, we have and Hence lim n→∞ 1 ρn lim n→∞ 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ {k ≤ n : g(αk+1) − g(αk) ≥ ε 2 } = 0 ε 2 } = 0. lim n→∞ 1 ρn {k ≤ n : [f (αk+1) − f (αk)] + [g(αk+1 − g(αk)]) ≥ ε} = 0 which follows from the inclusion {k ≤ n : (f +g)(αk+1)−(f +g)(αk ≥ ε} ⊆ {k ≤ n : f (αk+1) − f (αk) ≥ ε 2 } ∪ {k ≤ n : g(αk+1) − g(αk) ≥ ε 2 }. This completes the proof. Proposition 2. The composition of two ρ-statistically-downward continuous func- tions is ρ-statistically-downward continuous, i.e. if f and g are ρ-statistically- downward continuous functions on R, then the composition gof of f and g is ρ-statistically-downward continuous. Proof 7. Let f and g be ρ-statistically-downward continuous functions on R, and (αn) be a ρ-statistically-downward quasi Cauchy sequence of points in R. As f is ρ-statistically-downward continuous, the transformed sequence (f (αn)) is a ρ- statistically downward quasi Cauchy sequence. Since g is ρ-statistically-downward continuous, the transformed sequence g(f (αn)) of the sequence (f (αn)) is a ρ- statistically downward quasi Cauchy sequence. This completes the proof of the theorem. 8 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY In connection with ρ-statistically-downward quasi-Cauchy sequences, ρ-statistically- quasi-Cauchy sequences, ρ-statistically-statistical convergent sequences, and con- vergent sequences the problem arises to investigate the following types of continuity of functions on R: (δS− ρ ): (αk) ∈ ∆S− ρ ⇒ (f (αk)) ∈ ∆S− ρ (δS− ρ c): (αk) ∈ ∆S− ρ ⇒ (f (αk)) ∈ c (c): (αk) ∈ c ⇒ (f (αk)) ∈ c (cδS− ρ ): (αk) ∈ c ⇒ (f (αk)) ∈ ∆S− ρ (Sρ): (αk) ∈ Sρ ⇒ (f (αk)) ∈ Sρ (δSρ): (αk) ∈ ∆Sρ ⇒ (f (αk)) ∈ ∆Sρ. We see that (δS− ρ ) is ρ-statistically downward continuity of f , (Sρ) is the ρ- statistical continuity, and (δSρ) is the ρ-statistically-ward continuity. It is easy to see that (δS− ρ c) implies (δS− ρ ); (δS− ρ ) does not imply (δS− ρ c); (δS− ρ ) implies (cδS− ρ ); (cδS− ρ ) does not imply (δS− ρ ); (δS− ρ c) implies (c), and (c) does not im- ply (δS− ρ c); and (c) implies (cδS− ρ ). We see that (c) can be replaced by not only ρ-statistically-continuity ([6]), but also statistically-continuity ([17]), lacunary sta- tistical continuity ([8]), Nθ-sequential continuity ([29]), I-sequential continuity ([7]), and more generally G-sequential continuity ([12], [15]). Now we give the implication (δS− ρ ) implies (δSρ), i.e. any ρ-statistically-downward continuous function is ρ-statistically-ward continuous. Theorem 2. If f is ρ-statistically-downward continuous on a subset E of R, then it is ρ-statistically-ward continuous on E. Proof 8. Let (αk) be any ρ-statistically-quasi-Cauchy sequence of points in E. Then the sequence (α1, α2, α1, α2, α3, α2, α3, ..., αn−1, αk, αn−1, αk, αn+1, αk, αn+1, ...) is also ρ-statistically quasi-Cauchy. Then it is ρ-statistically-downward quasi-Cauchy. As f is ρ-statistically-downward continuous, the sequence (f (α1), f (α2), f (α1), f (α2), f (α3), f (α2), f (α3), ..., f (αn−1), f (αk), f (αn−1), f (αk), f (αn+1), f (αk), f (αn+1), ...) A NEW VARIATION ON STATISTICAL WARD CONTINUITY 9 is ρ-statistically-downward quasi-Cauchy. It follows from this that lim n→∞ 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} = 0 for every ε > 0. This completes the proof of the theorem. We note that the converse of the preceding theorem is not always true, i.e. there are ρ-statistically-ward continuous functions which are not ρ-statistically-downward continuous. Example 3. Write ρ = (ρn) as ρn = n + 1 n for each n > 1, ρ1 = 1, and consider the sequence α = (n). Then we see that the function f defined by f (x) = −x for every x ∈ R is an ρ-statistically-ward continuous function, but not ρ-statistically- downward continuous. Now we give the implication (δS− ρ ) implies (Sρ), i.e. any ρ-statistically-downward continuous function is ρ-statistically-continuous. Corollary 5. If f is ρ-statistically-downward continuous on a subset E of R, then it is ρ-statistically-continuous on E. Proof 9. Although the proof follows from [6, Theorem 3], the preceding theorem, and the fact that ρ-statistical continuity coincides with ordinary continuity, we give a direct proof in the following for completeness. Let (αk) be any ρ-statistically convergent sequence with Sρ − limk→∞ αk = ℓ. Then (α1, ℓ, α1, ℓ, α2, ℓ, α2, ℓ, ..., αk, ℓ, αk, ℓ, ...) is also ρ-statistically convergent to ℓ. Thus it is ρ-statistically-downward quasi- Cauchy. Hence (f (α1), f (ℓ), f (α1), f (ℓ), f (α2), f (ℓ), f (α2), f (ℓ), ..., f (αk), f (ℓ), f (αk), f (ℓ), ...) is ρ-statistically-downward quasi-Cauchy. It follows from this that lim n→∞ 1 ρn {k ≤ n : f (αk) − f (ℓ) ≥ ε} = 0 for every ε > 0. This completes the proof. 10 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY Observing that ρ-statistically-continuity implies ordinary continuity, we note that it follows from Theorem 5 that ρ-statistically-downward continuity implies not only ordinary continuity, but also some other kinds of continuities, namely, lacunary statistical continuity, statistical continuity ([6]), Nθ-sequential continuity, I-continuity for any non-trivial admissible ideal I of N ([7, Theorem 4], [26]), and G-continuity for any regular subsequential method G (see [11], [12], and [15]). Theorem 3. ρ-statistically-downward continuous image of any ρ-statistically-downward compact subset of R is ρ-statistically-downward compact. Proof 10. Let E be a subset of R, f : E −→ R be an ρ-statistically-downward continuous function, and A be an ρ-statistically-downward compact subset of E. Take any sequence β = (βn) of points in f (A). Write βn = f (αk), where αk ∈ A for each n ∈ N, α = (αk). ρ-statistically-downward compactness of A implies that there is an ρ-statistically-downward quasi-Cauchy subsequence ξ of the sequence of α. Write η = (ηk) = f (ξ) = (f (ξk)). Then η is an ρ-statistically-downward quasi- Cauchy subsequence of the sequence β. This completes the proof of the theorem. Theorem 4. Any ρ-statistically-downward continuous real valued function on a ρ-statistically-downward compact subset of R is uniformly continuous. Proof 11. Let E be an ρ-statistically-downward compact subset of R and let f : E −→ R. Suppose that f is not uniformly continuous on E so that there exists an ε0 > 0 such that for any δ > 0, there are x, y ∈ E with x − y < δ but f (x) − f (y) ≥ ε0. For each positive integer n, there are αn and βn such that αn − βn < 1 n , and f (αn) − f (βn) ≥ ε0. Since E is ρ-statistically-downward compact, there exists an ρ-statistically-downward quasi-Cauchy subsequence (αnk ) of the sequence (αk). It is clear that the corresponding subsequence (βnk ) of the sequence (βn) is also ρ-statistically-downward quasi-Cauchy, since (βnk+1 − βnk ) is a sum of three ρ-statistically-downward quasi-Cauchy sequences, i.e. βnk+1 − βnk = (βnk+1 − αnk+1 ) + (αnk+1 − αnk ) + (αnk − βnk ). Then the sequence (βn1 , αn1 , βn2 , αn2 , βn3 , αn3 , ..., βnk , αnk , ...) A NEW VARIATION ON STATISTICAL WARD CONTINUITY 11 is ρ-statistically-downward quasi-Cauchy, since the sequence (βnk+1 − αnk ) is a ρ- statistically-downward quasi-Cauchy sequence which follows from the equality βnk+1 − αnk = βnk+1 − βnk + βnk − αnk . But the sequence (f (βn1 ), f (αn1 ), f (βn2 ), f (αn2), f (βn3 ), f (αn3 ), ..., f (βnk ), f (αnk ), ...) is not ρ-statistically-downward quasi-Cauchy. Thus f does not preserve ρ-statistically- downward quasi-Cauchy sequences. This contradiction completes the proof of the theorem. Theorem 5. If a function f is uniformly continuous on a subset E of R, then (f (αk)) is ρ-statistically-downward quasi Cauchy whenever (αk) is a quasi-Cauchy sequence of points in E. Proof 12. Let E be a subset of R and let f be a uniformly continuous function on E. Take any quasi-Cauchy sequence (αk) of points in A, and let ε be any positive real number in ]0, 1[. By uniform continuity of f , there exists a δ > 0 such that f (x) − f (y) < ε whenever x − y < δ and x, y ∈ E. Since (αk) is a quasi-Cauchy sequence, there exists a positive integer k0 such that αk+1 − αk < ε for k ≥ k0, therefore f (αk+1) − f (αk) < ε for k ≥ k0. Thus the number of indices k ≤ n that satisfy f (αk+1) − f (αk) ≥ ε is less than or equal to k0. Hence 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} ≤ k0 ρn . It follows from this that lim n→∞ 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} = 0. Thus (f (αk)) is a ρ-statistically-downward quasi-Cauchy sequence. This completes the proof of the theorem. Now we have the following result related to uniform convergence, namely, the uniform limit of a sequence of ρ-statistically-downward continuous functions is ρ- statistically-downward continuous. 12 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY Theorem 6. If (fn) is a sequence of ρ-statistically-downward continuous functions defined on a subset E of R and (fn) is uniformly convergent to a function f , then f is ρ-statistically-downward continuous on E. Proof 13. Let ε be a positive real number and (αk) be any ρ-statistically-downward quasi-Cauchy sequence of points in E. By the uniform convergence of (fn) there exists a positive integer N such that fn(x) − f (x) < ε 3 for all x ∈ E whenever n ≥ N . As fN is ρ-statistically-downward continuous on E, we have lim n→∞ 1 ρn {k ≤ n : fN (αk+1) − fN (αk) ≥ ε 3 } = 0. On the other hand, we have {k ≤ n : f (αk+1) − f (αk) ≥ ε} ⊆ {k ≤ n : f (αk+1) − fN (αk+1) ≥ ε 3 } 3 } ∪ {k ≤ n : fN (αk) − f (αk) ≥ ε 3 } ∪{k ≤ n : fN (αk+1) − fN (αk) ≥ ε Hence 1 ρn + 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} ≤ 1 ρn {k ≤ n : fN (αk+1) − fN (αk) ≥ ε k ≤ n : f (αk+1) − fN (αk+1) ≥ ε 3 } {k ≤ n : fN (αk) − f (αk) ≥ ε 3 } 3 } + 1 ρn Now it follows from this inequality that limn→∞ 1 ρn {k ≤ n : f (αk+1) − f (αk) ≥ ε} = 0. This completes the proof of the theorem. 4. Conclusion The results in this paper include not only the related theorems on statistical downward continuity studied in [19] as special cases, i.e. ρn = n for each n ∈ N, but also include results which are also new for statistical downward continuity. In this paper, mainly, a new type of continuity, namely the concept of ρ-statistical downward continuity of a real function, and ρ-statistical downward compactness of a subset of R are introduced and investigated. In this investigation we have obtained theorems related to ρ-statistical downward continuity, and uniform continuity. We also introduced and studied some other continuities involving ρ-statistical down- ward quasi-Cauchy sequences, and convergent sequences of points in R. It turns out that the set of ρ-statistical downward continuous functions on a above bounded subset of R is contained in the set of uniformly continuous functions. We suggest to investigate ρ-statistical downward continuity of fuzzy functions or soft functions A NEW VARIATION ON STATISTICAL WARD CONTINUITY 13 (see [31], and [49] for the definitions and related concepts in fuzzy setting, and see [2], [38] related concepts in soft setting). We also suggest to investigate ρ-statistical downward continuity via double sequences (see for example [47], [9], [32], [55] and [57] for the definitions and related concepts in the double sequences case). For another further study, we suggest to investigate ρ-statistical downward continuity in an asymmetric cone metric space since in a cone metric space the notion of an ρ-statistical downward quasi Cauchy sequence coincides with the notion of an ρ- statistically quasi Cauchy sequence, and therefore ρ-statistical downward continuity coincides with ρ-statistically-ward continuity (see [35], [33], [1], [53], [59]). 5. Acknowledgment We acknowledge that some of the results in this paper were presented at the International Conference on Recent Advances in Pure and Applied Mathematics (ICRAPAM 2017) May 11-15, 2017, Palm Wings Ephesus Resort Hotel, Kusadasi - Aydin, TURKEY. References [1] S. Akduman, C. Ozel, A.Kilicman, Some remarks on I-ward continuity, J. Math. Anal. , 6 (5), 43-49 (2015) [2] C.G. Aras, A. Sonmez, H. Cakallı, An approach to soft functions, J. Math. Anal. 8 2 (2017), 129-138. [3] H. Bor, On Generalized Absolute Cesaro Summability, An. Stiint. Univ. Al. I. Cuza Iasi. Mat. (N.S.) 57, 2, 323-328, (2011). DOI: 10.2478/v10157-011-0029-9 [4] D. Burton, J. Coleman, Quasi-Cauchy sequences, Amer. Math. Monthly 117 (2010) 328-333. [5] H. Cakalli, Lacunary statistical convergence in topological groups, Indian J. Pure Appl. Math. 26 2 (1995) 113-119. [6] H. Cakalli, A Variation on Statistical Ward Continuity, Bull. Malays. Math. Sci. Soc. DOI 10.1007/s40840-015-0195-0 [7] H. Cakalli, and B. Hazarika, Ideal quasi-Cauchy sequences, J. Inequal. Appl. 2012 (2012) Article 234, 11 pages. [8] H. Cakalli, C.G. Aras, and A. Sonmez, Lacunary statistical ward continuity , AIP Conf. Proc. 1676 (2015) Article Number: 020042 [9] H. Cakalli and R.F. Patterson, Functions preserving slowly oscillating double sequences, An. Stiint. Univ. Al. I. Cuza Iasi Mat. (N.S.) Tomul LXII, 2016, f. 2, vol. 2, 531-536. [10] J. Collins and J. Zimmer, An asymmetric Arzel-Ascoli theorem, Topology Appl. 154 (2007) 2312-2322. 14 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY [11] J. Connor and K.G. Grosse-Erdmann, Sequential definitions of continuity for real functions, Rocky Mountain J. Math. 33 1 (2003) 93-121. [12] H. C¸ akalli, Sequential definitions of compactness, Appl. Math. Lett. 21 (2008) 594-598. [13] H. C¸ akalli, Slowly oscillating continuity, Abstr. Appl. Anal. 2008 (2008) Article ID 485706, 5 pages. [14] H. C¸ akalli, δ-quasi-Cauchy sequences, Math. Comput. Modelling 53 (2011) 397-401. [15] H. C¸ akalli, On G-continuity, Comput. Math. Appl. 61 (2011) 313-318. [16] H. C¸ akalli, Statistical ward continuity. Appl. Math. Lett. 24 (2011) 1724-1728. [17] H. C¸ akalli, Statistical quasi-Cauchy sequences, Math. Comput. Modelling 54 (2011) 1620- 1624. [18] H. C¸ akalli, A variation on ward continuity, Filomat 27 (2013) 1545-1549. [19] H. C¸ akallı, Upward and downward statistical continuities, Filomat 29 10 (2015) 2265-2273. [20] H. C¸ akallı, Variations on quasi-Cauchy sequences, Filomat 29 1 (2015) 13-19. [21] H. C¸ akallı, Beyond Cauchy and quasi Cauchy sequences, Filomat, to appear. [22] H. C¸ akalli, Forward compactness, Conference on Summability and Ap- plications, Shawnee State University, November 6November 8, 2009. http://webpages.math.luc.edu/ mgb/ShawneeConference/Articles/HuseyinCakalliOhio.pdf. [23] H. C¸ akalli, Forward continuity, J. Comput. Anal. Appl. 13 (2011) 225-230. [24] H. Cakalli. A variation on arithmetic continuity. Bol. Soc. Paran. Mat., 35 3 (2017) 195-202. [25] H. Cakalli, Corrigendum to the paper entitled" A variation on arithmetic continuity" pub- lished in Boletim da Sociedade Paranaense de Matematica Volume 35, Issue 3 (2017), Bol. Soc. Paran. Mat. 37 2 (2019) 177-179. [26] H. C¸ akalli, A variation on ward continuity, Filomat 27 8 (2013), 1545-1549. [27] H. Cakalli, and H. Kaplan, A variation on lacunary statistical quasi Cauchy sequences, Com- munications Faculty of Sciences University of Ankara-Series A1 Mathematics and Statistics, 66, 2, 71-79, (2017). 10.1501/Commua1 0000000802 [28] H. Cakalli and M. Albayrak, New Type Continuities via Abel Convergence, Scientific World Journal 2014 (2014) Article ID 398379, 6 pages. http://dx.doi.org/10.1155/2014/398379 [29] H. C¸ akalli, and H. Kaplan, A study on N-theta-quasi-Cauchy sequences, Abstr. Appl. Anal. 2013 (2013), Article ID 836970, 4 pages. [30] H. Cakalli and O. Mucuk, Lacunary statistically upward and downward half quasi-Cauchy sequences, J. Math. Anal. 7 2 (2016), 12-23. [31] H. C¸ akalli, and Pratulananda Das, Fuzzy compactness via summability, Appl. Math. Lett. 22 (2009) 1665-1669. [32] H. C¸ akallı and E. Sava¸s, Statistical convergence of double sequences in topological groups, J. Comput. Anal. Appl. 12 2 (2010), 421-426. [33] H. C¸ akallı and A. Sonmez, Slowly oscillating continuity in abstract metric spaces, Filomat 27 (2013) 925-930. A NEW VARIATION ON STATISTICAL WARD CONTINUITY 15 [34] H. C¸ akallı, A. Sonmez , and C¸ .G. Aras, λ-statistically ward continuity, An. Stiint. Univ. Al. I. Cuza Iasi. Mat. Tomul LXII, 2017, Tom LXIII, f. 2, 308-3012, (2017). DOI: 10.1515/aicu- 2015-0016 [35] H. C¸ akallı, A. Sonmez, and C. Genc, On an equivalence of topological vector space valued cone metric spaces and metric spaces, Appl. Math. Lett. 25 (2012) 429-433. [36] H. C¸ akallı, and E.Iffet Taylan, On Absolutely Almost Convergence, An. Stiint. Univ. Al. I. Cuza Iasi. Mat. (N.S.), Tom LXIII, f. 1 1-6, (2017). Doi: 10.2478/aicu-2014-0032 [37] I. Canak and M. Dik, New types of continuities, Abstr. Appl. Anal. 2010 (2010) Article ID 258980, 6 pages. [38] A.E. Coskun, C.G Aras, H. Cakalli, and A. Sonmez, Soft matrices on soft multisets in an optimal decision process, AIP Conference Proceedings, 1759, 1, 020099 (2016); doi: 10.1063/1.4959713 [39] Esi, A., Statistical and lacunary statistical convergence of interval numbers in topological groups, Acta Scientiarum-Technology, 36, 3, 491-495, 2014. [40] A. Esi, Asymptotically double lacunry equivalent sequences defined by Orlicz functions, Acta Scientiarum-Technology, 36 2 (2014) 323-329. [41] A. Esi, M. Acikgoz, On almost lambda- statistical convergence of fuzzy numbers, Acta Scientiarum-Technology, 36 1 (2014) 129-133. [42] A.R. Freedman, J.J. Sember, and M.Raphael, Some Cesaro-type summability spaces, Proc. London Math. Soc., 3, 37, 508-520, 1978. MR 80c:40007. [43] Fridy, J.A. and Orhan, C., Lacunary statistical convergence, Pacific J. Math., 160 1, 43-51 (1993) [44] Fridy, J.A. and Orhan, C., Lacunary statistical summability, J. Math. Anal. Appl, 173 2, 497-504 (1993) [45] H. Kaplan, H. Cakalli, Variations on strongly lacunary quasi Cauchy sequences, AIP Conf. Proc. 1759, Article Number: 020051, (2016). doi: http://dx.doi.org/10.1063/1.4959665 [46] H. Kaplan, H. Cakalli, Variations on strong lacunary quasi-Cauchy sequences, J. Nonlinear Sci. Appl. 9, 4371-4380, (2016). [47] D. Djurcic, Ljubia D.R. Kocinac, M.R. Zizovic, Double Sequences and Selections, Abstr. Appl. Anal. (2012) Article ID:497594, 6 pp. DOI: 10.1155/2012/497594, 2012. [48] M. Keane, Understanding Ergodicity, Integers 11B (2011) 1-11. [49] Ljubisa D.R. Kocinac, Selection properties in fuzzy metric spaces, Filomat, 26 2 2012 99-106. [50] L.Leindler, Uber die de la Vallee-Pousinsche Summierbarkeit allgemeiner Orthogonalreihen, Acta Math. Acad. Sci. Hungar. 16, (1965), 375-387. [51] M. Mursaleen, λ-statistical convergence, Math. Slovaca 50 (2000) 111-115. [52] H. Seyhan Ozarslan, and S¸. Yıldız, A new study on the absolute summability factors of Fourier series, J. Math. Anal. 7, 1, 31-36, (2016). [53] S.K. Pal, E. Savas, and H. Cakalli, I-convergence on cone metric spaces, Sarajevo J. Math. 9 (2013) 85-93. 16 HUSEYIN CAKALLI, MALTEPE UNIVERSITY, ISTANBUL-TURKEY [54] F.J. Palladino, On half Cauchy sequences, Arxiv 2012 arXiv:1102.4641v1, 2012, 3 pages. [55] R.F. Patterson and H. Cakalli, Quasi Cauchy double sequences, Tbilisi Mathematical Journal 8 2 (2015) 211-219. [56] R.F. Patterson, and E. Sava¸s, Rate of P-convergence over equivalence classes of double se- quence spaces, Positivity 16 4 (2012) 739-749. [57] R.F. Patterson, and E. Savas, Asymptotic equivalence of double sequences, Hacet. J. Math. Stat. 41 (2012) 487-497. [58] I. L. Reilly, P. V. Subrahmanyam, and M. K. Vamanamurthy, Cauchy sequences in quasi pseudo metric spaces, Monatsh. Math. 93 2 (1982) 127-140. [59] A. Sonmez, and H. Cakalli, Cone normed spaces and weighted means, Math. Comput. Mod- elling 52 (2010) 1660-16660. [60] R.W. Vallin, Creating slowly oscillating sequences and slowly oscillating continuous functions (with an appendix by Vallin and H. C¸ akalli), Acta Math. Univ. Comenianae 25 (2011) 71-78. [61] P. Winkler, Mathematical Puzzles: A Connoisseurs Collection, A.K.Peters LTD, (2004) ISBN 1-56881-201-9. [62] S¸. Yıldız, A new theorem on local properties of factored Fourier series, Bull. Math. Anal. Appl., vol. 8, no 2 (2016), pp. 1-8. [63] S¸. Yıldız, On Absolute Matrix Summability Factors of Infinite Series and Fourier Series, Gazi University Journal of Science, vol. 30, no 1 (2017), pp. 363-370. [64] S. Yildiz, On Application of Matrix Summability to Fourier Series, Math. Methods Appl. Sci., DOI: 10.1002/mma.4635, (2017). [65] S¸. Yıldız, Istatistiksel bo¸sluklu delta 2 quasi Cauchy dizileri, Sakarya Univer- sity Journal of Science, 21, 6, (2017). DOI: 10.16984/saufenbilder.336128 , http://www.saujs.sakarya.edu.tr/issue/26999/336128 [66] H. C¸ akalli, Forward compactness, Conference on Summability and Ap- plications, Shawnee State University, November 6November 8, 2009. http://webpages.math.luc.edu/ mgb/ShawneeConference/Articles/HuseyinCakalliOhio.pdf. Huseyin Cakalli Maltepe University, Graduate School of Science and Engineer- ing, Department of Mathematics, Marmara EgItIm Koyu, TR 34857, Maltepe, Istanbul- Turkey Phone:(+90216)6261050 ext:2311, fax:(+90216)6261113 E-mail address: [email protected]; [email protected]
1104.4587	1	1104	2011-04-23T20:09:09	The possible shapes of numerical ranges	[ "math.FA", "math.AG" ]	Which convex subsets of the complex plane are the numerical range W(A of some matrix A? This paper gives a precise characterization of these sets. In addition to this we show that for any A there exists a symmetric matrix B of the same size such that W(A)=W(B).	math.FA	math	The possible shapes of numerical ranges J. William Helton and I.M. Spitkovsky Abstract. Which convex subsets of C are the numerical range W (A) of some matrix A? This paper gives a precise characterization of these sets. In addition to this we show that for any A there exists a symmetric B of the same size such that W (A) = W (B) thereby settling an open question from [2]. Mathematics Subject Classification (2000). Primary 47A12. Keywords. Numerical range, linear matrix inequalities. Consider Cd, the standard complex inner product space. Let h·, ·i denote its scalar product, and k·k the related norm. The numerical range W (A) of a d × d matrix A is defined as W (A) = {hAx, xi : kxk = 1}. (1) It is well known that W (A) is a compact convex subset of C containing the spec- trum of A; see, e.g., monographs [3, 6] for these and other properties, as well as for the history of the subject. In this short note we give an answer to the question of exactly which sets W actually are the numerical range of some matrix A. This question was originally raised in Kippenhahn's 1951 article [7] (see also a more accessible English translation [8]) which gave several non-trivial necessary conditions on the "geometrical shape" of a numerical range. However, a necessary and sufficient condition remained open1. One can be obtained by the observation that curves critical to the problem were effectively classified in [4]. Didier Henrion in [5] makes such a connection2 and more, and states explicitly one side (necessary) of the characterization of numerical range. While all components of our paper can easily be extracted from [5] by hose com- fortable with the theory in [4] , we think our short note will nevertheless be useful to the numerical range community, at least for expository purposes. In particular, our Theorem 2 explicitly states a necessary and sufficient condition. Research of the first author supported by NSF grants DMS-0700758, DMS-0757212, and the Ford Motor Co. 1We would like to thank P. Y. Wu for discussion of this issue during XXVII South Eastern Analysis Meeting in Gainesville, FL. 2We are especially grateful to Bernd Sturmfels for bringing [5] to our attention. 2 Helton and Spitkovsky Our characterization of numerical ranges is in terms of a type of dual convex set. For any set S ⊂ Rn its polar is defined as S∗ = {x ∈ Rn : sup y∈S hx, yi ≤ 1} (2) (see, e.g., [1, 10]). The set S∗ is closed, convex, and contains 0. Clearly (see also [10, Corollary 14.5.1]), 0 is an interior point of S∗ if and only if S is bounded. If S itself is closed, convex and contains 0, then (S∗)∗ = S (3) [10, Theorem 14.5]. The next result provides an explicit description of polar sets of numerical ranges. In some form it goes back many years, at least to §3 [7]. A different point of view (in a more general setting) is presented in [11, Section 5] (there the term dual is used in place of polar). Lemma 1. Let A ∈ Cd×d. Then W (A)∗ = {z = ξ + iη : I − ξH − ηK is positive semi-definite}. (4) Here H and K are hermitian matrices from the representation A = H + iK. (5) Proof. Directly from the definitions (1) and (2) it follows that W (A)∗ = {z : Re(hAv, viz) ≤ 1 for all v ∈ Cd with kvk = 1} = {z : h(Re(zA)v, vi) ≤ 1 for all v ∈ Cd with kvk = 1} = {z : I − Re(zA) is positive semi-definite} = {ξ + iη : I − ξH − ηK is positive semi-definite}. (cid:3) Common terminology is that (4) is a linear matrix inequality (LMI for short) representation for W (A)∗ and the lemma says that if a set W ⊂ C is a numerical range, then its polar has an LMI representation. The paper [4] describes precisely the sets C in R2, hence in C, which have an LMI representation. It characterizes them as "rigidly convex" a term we set about to define. An algebraic interior C has a defining polynomial q, namely C is the closure of the connected component of C := {z : q(z) > 0} containing 0. A minimum degree defining polynomial for C is unique (up to a constant), see Lemma 2.1 [4] and its degree we call the degree of C. A convex set C is called rigidly convex provided it is an algebraic interior and it has a defining polynomial q which satisfies the real zero (RZ) condition, namely, if µ ∈ C and q(µz) = 0, then µ ∈ R. Our main theorem is: Theorem 2. A subset W of C is the numerical range of some d × d matrix A if and only if its polar W∗ is rigidly convex of degree less than or equal to d. The possible shapes of numerical ranges Proof. Given A = H + iK, observe that p defined by p(z) = det(I − ξH − ηK) 3 (6) is an RZ polynomial, since all eigenvalues of a symmetric matrix are real. More- over, W (A)∗ coincides with the closure of the connected component of {z : p(z) > 0} containing zero. Thus the set W (A)∗ is rigidly convex. However, Theorem 3.1 of [4] says that converse also holds 3 : if V is rigidly convex, then there exist real symmetric matrices H, K such that V = {z = ξ + iη : I − ξH − ηK is positive semi-definite}. (7) Consequently, V = W (B)∗ for B = H + iK. Moreover, we can do this with an H, K whose dimension is the degree of V . (cid:3) The forward side of Theorem 2 is in [5] (stated in the language of homoge- neous coordinates, and emphasizing that numerical ranges are affine projections of semi-definite cones). The converse follows easily from ingredients there, though it is not stated explicitly. Note that the matrix B constructed in the proof of Theorem 2 is symmetric along with H, K. This yields an affirmative answer to the question stated in [2] (raised by the referee of the latter): Corollary 3. For every d × d matrix A there exists a symmetric d × d matrix B such that W (B) = W (A). Duality (3) allows us to restate Theorem 2 in the following form. Corollary 4. A subset W of C is the numerical range of some d × d matrix A if and only if it is a translation of the polar of a rigidly convex set of degree less than or equal to d. Proof. For a given d × d matrix A, pick λ ∈ W (A) and let A0 = A − λI. By Theorem 2, the polar set V of W (A0) is rigidly convex and has degree not exceeding d. But 0 ∈ W (A0), so that due to (3) we have W (A0) = V∗. Consequently, W (A) = W (A0) + λ is a translation of V∗ Conversely, if W is a translation of V∗ for some rigidly convex set V of degree not exceeding d, then W − λ = V∗ for some λ ∈ C. Applying (3) to S = V , we conclude that (W − λ)∗ = V . By Theorem 2, W − λ = W (A0) for some d × d matrix A0, so that W = W (A0 + λI). (cid:3) Remark 5. If the matrices H, K from representation (5) are linearly dependent with I, then the set V in (7) is unbounded. Moreover, V stays unbounded under translations of A. In other words, W (A) in this case has empty interior. This agrees with the fact that A in this (and only this) case has the form αR + βI for some hermitian R and α, β ∈ C, and W (A) is therefore a (closed) line segment. In all 3For perspective, [9] showed that the proof of Theorem 3.1 in [4] implies a 1958 conjecture of Peter Lax is true. In this context we might describe the characterization of numerical ranges (Theorem 2) as "polar" to the Lax Conjecture. 4 Helton and Spitkovsky other cases the interior of W (A) is non-empty, and W (A − λI)∗ is bounded for any λ lying in the interior of W (A). One such value of λ is λ = tr(A)/d. References [1] A. Ben-Tal and A. Nemirovski, Lectures on modern convex optimization, MPS/SIAM Series on Optimization, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2001, Analysis, algorithms, and engineering applications. [2] W.-S. Cheung, X. Liu, and T.-Y. Tam, Multiplicities, boundary points, and joint numerical ranges, Operators and Matrices 5 (2011), no. 1, 41 -- 52. [3] K. E. Gustafson and D. K. M. Rao, Numerical range. The field of values of linear operators and matrices, Springer, New York, 1997. [4] J. W. Helton and V. Vinnikov, Linear matrix inequality representation of sets, Comm. Pure Appl. Math. 60 (2007), no. 5, 654 -- 674. [5] D. Henrion, Semidefinite geometry of the numerical range, Electron. J. Linear Alge- bra 20 (2010), 322 -- 332. [6] R. A. Horn and C. R. Johnson, Topics in matrix analysis, Cambridge University Press, Cambridge, 1991. [7] R. Kippenhahn, Uber den Wertevorrat einer Matrix, Math. Nachr. 6 (1951), 193 -- 228. [8] , On the numerical range of a matrix, Linear Multilinear Algebra 56 (2008), no. 1-2, 185 -- 225, Translated from the German by Paul F. Zachlin and Michiel E. Hochstenbach. [9] A. S. Lewis, P. A. Parrilo, and M. V. Ramana, The Lax conjecture is true, Proc. Amer. Math. Soc. 133 (2005), no. 9, 2495 -- 2499 (electronic). [10] R. T. Rockafellar, Convex analysis, Princeton Landmarks in Mathematics, Prince- ton University Press, Princeton, NJ, 1997, Reprint of the 1970 original, Princeton Paperbacks. [11] P. Rostlaski and B. Sturmfels, Notions of duality in convex algebraic geometry, preprint. J. William Helton Department of Mathematics University of California San Diego 9500 Gilman Drive La Jolla, CA 92093-0112 USA e-mail: [email protected] I.M. Spitkovsky Department of Mathematics College of William and Mary Williamsburg, VA 23187 USA e-mail: [email protected], [email protected]
1708.01545	1	1708	2017-08-04T15:12:45	A generalized Schur complement for non-negative operators on linear space	[ "math.FA" ]	Extending the corresponding notion for matrices or bounded linear operators on a Hilbert space we define a generalized Schur complement for a non-negative linear operator mapping a linear space into its dual and derive some of its properties.	math.FA	math	A GENERALIZED SCHUR COMPLEMENT FOR NON-NEGATIVE OPERATORS ON LINEAR SPACES J. FRIEDRICH,1 M. G UNTHER,2 ∗ and L. KLOTZ2 Abstract. Extending the corresponding notion for matrices or bounded lin- ear operators on a Hilbert space we define a generalized Schur complement for a non-negative linear operator mapping a linear space into its dual and derive some of its properties. 1. Introduction In case of 2 × 2 block matrices the notion of Schur complement and its gener- alizations have a long history. We refer to [24], where it is given a comprehensive exposition of the history, theory and versatility of Schur complements. In the case of non-negative matrices the notion of generalized Schur complement can be extended to matrices whose entries are bounded operators in Hilbert spaces, cf. [17]. Moreover, generalized Schur complements are closely related to shorted operators which were introduced by M.G. Kreın [11] and have found interesting applications in electrical network theory, cf. [2]. Since there is a natural way to define non-negativity for a linear operator, which maps a linear space into its dual, one can ask for a generalized Schur complement of such an operator. A first attempt was made in [6], where non-negative bounded linear operators from a Banach space into its topological dual were discussed. The present paper deals with a generalized Schur complement and a shorted operator of a non-negative 2 × 2 block matrix, whose entries are linear operators on linear spaces. Thus, topological questions, particularly continuity problems, play a minor role. To define a generalized Schur complement of a non-negative operator on a linear space, one needs the notion of a square root. Section 3 deals with this impor- tant and useful concept, which was studied by many authors. Section 4 contains definitions and basic properties of the Schur complement and the shorted oper- ator for slightly more general than non-negative operators. In Section 5 further results on generalized Schur complements are derived. Among other things we extend the Crabtree-Haynsworth quotient formula [4]. One of the most useful results concerning 2 × 2 block matrices is Albert's non-negativity criterion [1]. A generalization to non-negative operators on linear spaces and some of its conse- quences are given in Section 6. The special class of extremal operators, which was introduced by M.G. Kreın [11] is the subject of Section 7. Date: July 30, 2018 ∗Corresponding author. 2010 Mathematics Subject Classification. 47A05, 47A07. Key words and phrases. Schur complement, shorted operator, extremal operator. 1 2 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ For bounded linear operators on Hilbert spaces many results concerning the generalized Schur complement were obtained by Yu. L. Shmulyan. A large part of them was proved independently or rediscovered later on by several mathe- maticians from western countries. The present paper is strongly influenced by Shmulyan's work and was written to illustrate his contribution to the theory of generalized Schur complement. In this way most of our assertions of Sections 4-6 are generalizations of results contained in [17] to non-negative operators on linear spaces. 2. Basic definitions and notations Any linear space of the present paper is a space over C, the field of complex numbers, and its zero element is denoted by 0. For a linear space X, let X ′ denote its dual space of all antilinear functionals on X and hx′, xiX := hx′, xi the value of x′ ∈ X ′ at x ∈ X. If X ∼ is a subspace of X ′, an arbitrary x ∈ X defines an element jx of (X ∼)′ according to where ¯α stands for the complex conjugate of α ∈ C. hjx, x∼iX ′ := hx∼, xiX, x∼ ∈ X ∼, Convention (CN): If for all x ∈ X \ {0} there exists x∼ ∈ X ∼ such that hx∼, xi 6= 0, we shall identify X and its isomorphic image under the map j and write hjx, x∼iX ′ =: hx, x∼iX ′, x ∈ X, x∼ ∈ X ∼. The linear space of all linear operators from X into a linear space Y is denoted by L (X, Y ), and I is the identity operator in case X = Y . If A ∈ L (X, Y ) and X1 is a subspace of X, the symbols ker A, ran A and A↾X1 stand for the null space, range, and restriction of A to X1, resp. Set AX1 := ran A↾X1. The dual operator A′ ∈ L (Y ′, X ′) is defined by the relation hy′, AxiY = hA′y′, xiX, x ∈ X, y′ ∈ Y ′. If Z is a linear space and A ∈ L (X, Y ), B ∈ L (Y, Z), then Examples. 1. (BA)′ = A′B′. 2. If A ∈ L (X, Y ), then A′′ ∈ L (X ′′, Y ′′) and A = A′′↾X according to (CN). 3. If A ∈ L (X, X ′), then A′ ∈ L (X ′′, X ′). Taking into account (CN), we get hx2, Ax1iX ′ = hAx1, x2iX and hx2, Ax1iX ′ = hA′x2, x1iX, hence, hAx1, x2iX = hA′x2, x1iX , x1, x2 ∈ X. (2.1) An operator A ∈ L (X, X ′) is called Hermitian, if hAx1, x2i = hAx2, x1i and non-negative if hAx1, x1i ≥ 0, x1, x2 ∈ X. The sets of all Hermitian and all non-negative operators are denoted by L h(X, X ′) and L ≥(X, X ′), resp.. The polarization identity implies that A is Hermitian if and only if hAx, xi is real for all x ∈ X. Thus L ≥(X, X ′) ⊆ L h(X, X ′) and the space L h(X, X ′) can be provided with Loewner's semi-ordering, i.e. for A, D ∈ L h(X, X ′) we shall write A ≤ D if and only if hAx, xi ≤ hDx, xi, x ∈ X. Recall Cauchy's inequality hAx1, x2i2 ≤ hAx1, x1ihAx2, x2i, x1, x2 ∈ X, (2.2) if A ∈ L ≥(X, X ′). A GENERALIZED SCHUR COMPLEMENT 3 3. Square roots Let H be a complex Hilbert space with norm k · k := k · kH and inner product (· ·) := (· ·)H, which is assumed to be antilinear with respect to the second component. Let R ∈ L (X, H). Identifying H and the space of continuous antilinear functionals on H in the common way, one has H ⊆ H ′ and (h Rx) = hR′h, xi, (3.1) Set R∗ := R′ ↾H. From (3.1) it can be concluded that ker R∗ is equal to the orthogonal complement of (ran R)c, where M c denotes the closure of a subset M of a topological space. It follows that R∗ is one-to-one if and only if ran R is dense in H and that x ∈ X, h ∈ H. ran R∗ = R∗(ran R)c (3.2) Therefore, we can define a generalized inverse R∗[−1] of R∗ by x′ ∈ ran R∗. R∗[−1]x′ := (R∗↾(ran R)c)−1x′, Lemma 3.1. Let R ∈ L (X, H). An element x′ ∈ X ′ belongs to ran R∗ if and only if the following conditions are satisfied: (i) If x ∈ ker R, then hx′, xi = 0. (ii) sup x∈X hx′, xi2 kRxk2 H < ∞ (with convention 0 0 := 0 at the left-hand side). Proof. If x′ ∈ R∗h for some h ∈ H, then hx′, xi = hR∗h, xi = (h Rx) ≤ khkkRxk, which yields (i) and (ii). Conversely, assume that (i) and (ii) are satisfied for some x′ ∈ X ′. Set ϕ(Rx) := hx′, xi, x ∈ X. Because of (i) ϕ is correctly defined and (ii) implies that ϕ is continuous, so that ϕ is a continuous antilinear functional on ran R. Thus, there exists h ∈ H such that hx′, xi = (h Rx) = hR∗h, xi for all x ∈ X, which yields x′ = R∗h ∈ ran R∗. (cid:3) Definition 3.2. Let A ∈ L (X, X ′). A pair (R, H) of a Hilbert space H and an operator R ∈ L (X, H) is called a square root of A if A = R∗R, and a minimal square root if, addionally, ran R is dense in H. Note that there exists a square root of A if and only if there exists a minimal one. The following result is basic to our considerations and generalizes the fact concerning the existence of a square root of a non-negative selfadjoint operator in a Hilbert space. Its well known short proof is recapitulated for convenience of the reader. Theorem 3.3. An operator A ∈ L (X, X ′) possesses a square root if and only if it is non-negative. Proof. Let A ∈ L ≥(X, X ′). Cauchy's inequality (2.2) implies that N := {x ∈ X : hAx, xi = 0} is a subspace of X. Define an inner product on the quotient space X/N by (x1 + N x2 + N) := hAx1, x2i, x1, x2 ∈ X, 4 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ and denote the completion of the corresponding inner product space by H. Set Rx := x + N, x ∈ X. It follows R ∈ L (X, H), (ran R)c = H, and hAx1, x2i = (Rx1 Rx2) = hR∗Rx1, x2i, x1, x2 ∈ X. Therefore, (R, H) is a minimal square root of A. The "only if"-part of the asser- tion is obvious. (cid:3) The notion of a square root of a non-negative operator acting between spaces more general than Hilbert spaces was discussed and applied by many authors. Most of them deal with a topological space X and in this case continuity prob- lems arise additionally. Some properties of square roots for operators of special type were obtained by Vaınberg and Engel'son, cf. [23]. For a Banach space X, the construction of the proof of the preceding theorem was published as an ap- pendix to [21] and attributed to Chobanyan, see also [15] and [22]. Another but related construction was proposed by Sebesty´en [16], cf. [19]. G´orniak [7] and G´orniak and Weron [8] dealt with the existence of a continuous square root if X is a topological linear space. G´orniak, Makagon and Weron [9] investigated square roots of non-negative operator-valued measures. Pusz and Woronowicz [14] ex- tended the construction of the proof of Theorem 3.3 to pairs of non-negative sequilinear forms, cf. [20] for further generalizations. Lemma 3.4. If A ∈ L (X, X ′) and (R, H) is a square root of A, then ker R = ker A = {x ∈ X : hAx, xi = 0}. Proof. The result follows from a chain of conclusions: hAx, xi = 0 ⇒ hR∗Rx, xi = 0 ⇒ (Rx Rx) = 0 ⇒ Rx = 0, and conversely Rx = 0 ⇒ R∗Rx = 0 ⇒ Ax = 0 ⇒ hAx, xi = 0. (cid:3) The preceding results can be used to derive a version of a part of Douglas' theorem [5], cf. [18]. Proposition 3.5. Let A, D ∈ L ≥(X, X ′) and (RA, HA) and (RD, HD) be square roots of A and D, resp.. The following assertions are equivalent: (i) A ≤ α2D for some α ∈ [0, ∞), (ii) there exists a bounded operator W ∈ L (HA, HD) with operator norm kW k ≤ α and such that R∗ A = R∗ DW . If (i) or (ii) are satisfied, there exists a unique W so that W ⊆ (ran RD)c. More- over, ker W = ker R∗ A for this operator W . Proof. Since R∗ (ii) it follows A = R∗ DW yields RA = R∗ A ′↾X= W ′R∗ D ′↾X= W ∗RD by (CN), from hAx, xi = kRAxk2 = kW ∗RDxk2 ≤ α2kRDxk2 = α2hDx, xi, x ∈ X, hence, (i). Let Wj ∈ L (HA, HD be such that R∗ DWj and ran Wj ⊆ (ran RD)c, j = 1, 2. Then ran(W1 − W2) ⊆ ker R∗ D and ran(W1 − W2) ⊆ (ran RD)c, which shows that W1 = W2. Now assume that (i) is true. One has A = R∗ A GENERALIZED SCHUR COMPLEMENT 5 ker RD ⊆ ker RA by Lemma 3.4, hence, ran R∗ erator W := R∗[−1] A ∈ L (HA, HD) satisfies R∗ and ran W ⊆ (ran RD)c. The inclusion ran R∗[−1] W ∗h = R∗′ D R∗ A(R∗[−1] D D )∗h = 0 if h is orthogonal to ran RD. Therefore, from A ⊆ ran R∗ DW = R∗ D by Lemma 3.1. The op- A, ker W = ker R∗ A ⊆ (ran RD)c implies that kW ∗RDxk2 = kRAxk2 = hAx, xi ≤ α2hDx, xi = α2kRDxk2, x ∈ X, one can conclude that kW k = kW ∗k ≤ α. (cid:3) As a by-product of Proposition 3.5 we obtain the following corollary. Corollary 3.6. Let A, D, (RA, HA), and (RD, HD) be as in Proposition 3.5. (i) If A ≤ α2D for some α ∈ [0, ∞), then ran R∗ A ⊆ ran R∗ D. (ii) If β2D ≤ A ≤ α2D for some α, β ∈ (0, ∞), then ran R∗ A = ran R∗ D. (iii) If (SA, GA) is a square root of A, then ran R∗ A = ran S ∗ A. Corollary 3.7. Let Hj be Hilbert spaces and Rj ∈ L (X, Hj), j = 1, 2. (R, H) is a square root of the non-negative operator A := R∗ ran R∗ = ran R∗ 1R1 + R∗ If 2R2, then 1 + ran R∗ 2. Proof. Let G be the orthogonal sum of H1 and H2 and S ∈ L (X, G) be defined 1 + 2 and that (S, G) is a square root of A. Now apply Corollary 3.6 (iii). (cid:3) 2) and S ∗S = A, we get that ran S ∗ = ran R∗ R2(cid:17). Since S ∗ = (R∗ by S = (cid:16)R1 ran R∗ 1, R∗ Lemma 3.8. Let (R, H) be a square root and (S, G) a minimal square root of A ∈ L ≥(X, X ′). There exists an isometry U ∈ L (G, H) such that US = R. Proof. By Lemma 3.4 there exists an operator U satisfying U Sx = Rx, x ∈ X. From k USxk2 = kRxk2 = hAx, xi = kSxk2 it follows that U is isometric and can be extended to an isometry U ∈ L (G, H). (cid:3) 4. Generalized Schur complements and shorted operators of operators of positive type Let X und Y be linear spaces. Definition 4.1. A pair (A, B) of an operator A ∈ L ≥(X, X ′) and B ∈ L (Y, X ′) is called a positive pair if ran B ⊆ ran R∗ for some square root (and, hence, for all square roots) (R, H) of A. The following criterion is an immediate consequence of Lemma 3.1. Lemma 4.2. Let A ∈ L ≥(X, X ′) and B ∈ L (Y, X ′). The pair (A, B) is a positive pair if and only if for all y ∈ Y the following conditions are satisfied: (i) If x ∈ ker A, then hBy, xi = 0. (ii) sup x∈X hBy, xi2 hAx, xi < ∞ (with convention 0 0 := 0). According to (CN), the space X can be considered as a subspace of the domain of B′. To abbreviate the notation we set B∼ := B′↾X . 6 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ Note that hBy, xi = hB∼x, yi, x ∈ X, y ∈ Y . Thus condition (i) of Lemma 4.2 is equivalent to the inclusion ker A ⊆ ker B∼. If (A, B) is a positive pair and (R, H) is a square root of A, the operators T := R∗[−1]B and can be defined. Note that B = R∗T . The following lemma is obvious. ω(A, B) := T ∗T = (R∗[−1])B)∗R∗[−1]B (4.1) (4.2) Lemma 4.3. If (A, B) is a positive pair and (R, H) is a square root of A, then for all y ∈ Y inf{kT y − Rxk : x ∈ X} = 0. Equivalently, ran T ⊆ (ran R)c. (cid:3) Recall that the dual space of X × Y can be written as a Cartesian product (X × Y )′ = X ′ × Y ′, where h·, ·iX×Y = h·, ·iX + h·, ·iY . Also, it should not cause confusion if we identify the subspace X × {0} of X × Y with X. An operator A of L(cid:0)X × Y, (X × Y )′(cid:1) can be represented as a 2 × 2 matrix A = (cid:16)A B A ∈ L (X, X ′), B, C ∈ L (Y, X ′), D ∈ L (Y, Y ′). It is not hard to see that A is Hermitian if and only if A and D are Hermitian and C = B∼. To abbreviate the C D(cid:17), where type if (A, B) is a positive pair. The set of operators of positive type is denoted notation we set L h(cid:0)X × Y, (X × Y )′(cid:1) =: L h and L ≥(cid:0)X × Y, (X × Y )′(cid:1) =: L ≥. Definition 4.4. An operator (cid:16) A B B∼ D(cid:17) ∈ L h is called an operator of positive by L +(cid:0)X × Y, (X × Y )′(cid:1) =: L +. Definition 4.5. Let A = (cid:16) A B B∼ D(cid:17) ∈ L +. The operator σ(A) := D − ω(A, B) S (A) := (cid:18)0 is called a generalized Schur complement of A and the operator 0 σ(A)(cid:19) 0 is called a shorted operator. The following result is a generalization of [2, Corollary 1 to Theorem 3]. Proposition 4.6. If A = (cid:16) A B Proof. Let y′ ∈ Y ′ be such that A(cid:16)x be a minimal square root of A. Since B∼ D(cid:17) ∈ L +, then ran A ∩ Y ′ ⊆ ran S (A). y′(cid:17) for some (cid:16)x y(cid:17) = (cid:16) 0 A = (cid:18)R∗R R∗T T ∗R T ∗T(cid:19) + S (A), y(cid:17) ∈ X ×Y . Let (R, H) one has R∗Rx + R∗T y = 0, hence, Rx + T y = 0 and T ∗Rx + T ∗T y = 0, which yields y′(cid:19) = S (A)(cid:18)x (cid:18) 0 y(cid:19) ∈ ran S (A). (cid:3) A GENERALIZED SCHUR COMPLEMENT 7 The next result is a simple but useful consequence of Lemma 4.3. Proposition 4.7. Let A = (cid:16) A B B∼ D(cid:17) ∈ L +. For (cid:16)x y(cid:19)(cid:29) = inf z∈X(cid:28)A(cid:18)x − z y(cid:17) ∈ X × Y , y (cid:19)(cid:29) . y (cid:19)(cid:18)x − z (cid:28)S (A)(cid:18)x y(cid:19) ,(cid:18)x (4.3) Particularly, σ(A) and S (A) do not depend on the choice of the square root of A. Proof. Since (4.3) is independent of x ∈ X, it is enough to prove it for x = 0. From Lemma 4.3 it follows z∈X(cid:28)A(cid:18)−z inf y (cid:19) ,(cid:18)−z y (cid:19)(cid:29) z∈X(cid:28)(cid:18)R∗R R∗T = inf T ∗R T ∗T(cid:19)(cid:18)−z y (cid:19) ,(cid:18)−z kT y − Rzk2 +(cid:28)S (A)(cid:18)0 y (cid:19)(cid:29) +(cid:28)S (A)(cid:18)0 y(cid:19) ,(cid:18)0 y(cid:19)(cid:29) y(cid:19) ,(cid:18)0 y(cid:19)(cid:29) = inf z∈X = (cid:28)S (A)(cid:18)0 y(cid:19) ,(cid:18)0 y(cid:19)(cid:29) . (cid:3) Corollary 4.8. (i) If A ∈ L +, then S (A) ≤ A and ker A ⊆ ker S (A). (ii) If A, A1 ∈ L + and A ≤ A1, then S (A) ≤ S (A1). Proof. The first assertion of (i) as well as (ii) are immediately clear from Propo- sition 4.7. To prove the second assertion of (i), let (cid:16)x y(cid:17) ∈ ker A. If z ∈ X, one has (cid:28)A(cid:18)x − z y (cid:19) ,(cid:18)x − z y (cid:19)(cid:29) = hAz, zi ≥ 0, which implies that the infimum at the right hand side of (4.3) is equal to 0. Since S (A) ≤ A and (cid:28)(cid:0)A − S (A)(cid:1)(cid:18)x y(cid:19)(cid:29) = 0, y(cid:17) ∈ ker(cid:0)A − S (A)(cid:1) by Lemma 3.4, hence, (cid:16)x y(cid:19) ,(cid:18)x it follows (cid:16)x Corollary 4.9. If A ∈ L ≥, then S (A) ∈ L ≥. y(cid:17) ∈ ker S (A). (cid:3) (cid:3) 5. Further applications of square roots First we express the generalized Schur complement of an operator of L ≥ with the aid of its square root and derive a range description, cf. [2, Corollary 4 to Theorem 1]. Let A ∈ L ≥ and (R, H) be a square root of A. Let L be the orthogonal complement of (RX)c and P be the orthoprojection onto L. Note that L can be characterized by L = {h ∈ H : R∗h ∈ Y ′}, which yields R∗L = ran R∗ ∩ Y ′. Proposition 5.1. If A ∈ L ≥, then S (A) = R∗P R. y(cid:19) ,(cid:18)x y(cid:17) ,(cid:16)x y(cid:19)(cid:29) = (cid:13)(cid:13)(cid:13)(cid:13) 8 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ Proof. Let (cid:16)x y(cid:17) ∈ X × Y . An application of (4.3) gives (cid:28)S (A)(cid:18)x which shows that DS (A)(cid:16)x Therefore, 2 R(cid:18)x z∈X(cid:13)(cid:13)(cid:13)(cid:13) y(cid:19)(cid:29) = inf y(cid:17)E is the squared distance of R(cid:16)x 0(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) y(cid:19) − R(cid:18)z , y(cid:17) to RX. (cid:28)S (A)(cid:18)x y(cid:19) ,(cid:18)x = (cid:28)R∗P R(cid:18)x y(cid:19) ,(cid:18)x y(cid:19)(cid:29) 2 y(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) P R(cid:18)x and the assertion follows from the polarization identity. (cid:3) Proposition 5.2. If A ∈ L ≥ and (R, H) and (S, G) are square roots of A and S (A), resp., then ran S ∗ = ran R∗ ∩ Y ′. Proof. Setting RX := R↾X and RY := R↾Y , we get A = (cid:18)R∗ X R∗ Y(cid:19) (RX RY ) = (cid:18)R∗ R∗ XRX R∗ Y RX R∗ X RY Y RY (cid:19) , hence, σ(A) = R∗ Y RY − (R∗[−1] X R∗ XRY )∗R∗[−1] X R∗ XRY = R∗ Y P RY since R∗[−1] X R∗ X = I − P . Thus S (A) = R∗P R and (P R, H) is a square root of X h = 0, y′(cid:17) for some h ∈ H, then R∗ y′(cid:17) ∈ X ′ ×Y ′ is such that R∗h = (cid:16) 0 S (A). If (cid:16) 0 hence, P h = h and (P R)∗h = R∗h = (cid:16) 0 y′(cid:17), which implies that ran R∗ ∩ Y ′ ⊆ ran(P R)∗ = ran S ∗ by Corollary 3.6 (iii). Since, obviously, ran S ∗ ⊆ Y ′ and ran S ∗ ⊆ ran R∗ by Corollaries 4.8 (i) and 3.6 (i), the assertion is proved. (cid:3) Our next result is a generalization of the Crabtree-Haynsworth quotient formula [4]. To give it a nice form let us denote σ(A) =: A/A. Proposition 5.3. Let X, Y , and Z be linear spaces, D :=   and A := (cid:16) A B A B BX B∼ D BY B∼ Y D1 X B∼   ∈ L ≥(X × Y × Z, X ′ × Y ′ × Z ′), B∼ D(cid:17). The operator A/A is the left upper corner of D/A and Proof. Let (R, H) be a minimal square root of A, RX := R↾X, RY := R↾Y , D/A. A/A = D/A. E := (R∗)−1(cid:18)BX BY(cid:19) , A GENERALIZED SCHUR COMPLEMENT 9 hence, R∗ X E = BX, R∗ Y E = BY . From R∗[−1] X R∗ X = I − P one obtains Y E Y RY R∗ D/A = (cid:18)R∗ = (cid:18)R∗ E ∗RY D1 (cid:19) −(cid:0)R∗[−1] X (R∗ E ∗P RY D1 − E ∗(I − P )E(cid:19) Y P RY Y P E R∗ XRY , R∗ X (R∗ XRY , R∗ XE) X E)(cid:1)∗R∗[−1] and A/A = R∗ Y RY −(cid:0)R∗[−1] X R∗ X RY(cid:1)∗R∗[−1] X R∗ XRY = R∗ Y P RY , which shows that A/A is the left upper corner of D/A. Since (P RY , H) is a square root of A/A, one can compute D/A. A/A = D1 − E ∗(I − P )E −(cid:0)(P RY )∗[−1]R∗ = D1 − E ∗(I − P )E − E ∗QE, Y P E(cid:1)∗(P RY )∗[−1]R∗ Y P E where Q denotes the orthoprojection onto (ran P RY )c. Comparing this with D/A = D1 −(cid:0)(R∗)−1R∗E(cid:1)∗(R∗)−1R∗E = D1 − E ∗E, we can conclude that the assertion will be proved if we can show that the restric- tion of I − P + Q to ran R is the identity. If h ∈ ran R, then h = RXx + RY y = RX x + (I − P )RY y + P RY y for some (cid:16)x y(cid:17) ∈ X × Y . Since RXx + (I − P )RY y ∈ (ran RX )c, there exists a sequence {xn}n∈N of elements of X such that limn→∞ RX xn = RX x+(I −P )RY y. For hn := RX xn + P RY y, we have (I − P + Q)hn = (I − P + Q)(RX xn + P RY y) = RX xn + P RY y = hn and therefore (I − P + Q)h = lim n→∞ = lim n→∞ (I − P + Q)hn (RXxn + P RY y) = RX x + RY y = h. (cid:3) We conclude this section with a criterion for non-negativity of operators of L h. Proposition 5.4. Let A = (cid:16) A B B∼ D(cid:17) ∈ L h. The operator A is non-negative if and only if the following two conditions are satisfied: (i) The operators A and D are non-negative. (ii) For any square roots (RA, HA) and (RD, HD) of A and D, resp., there AKR and ran K ⊆ exists a contraction K ∈ L (HD, HA) such that B = R∗ (ran RA)c. Proof. If A is non-negative, assertion (i) is trivial. To prove (ii) let (R, H) be a square root of A and RX := R↾X , RY = R↾Y , hence A = (cid:18)R∗ R∗ X RX R∗ Y RX R∗ XRY Y RY(cid:19) . Let (SA, GA) and (SD, GD) be minimal square roots of A and D, resp.. According to Lemma 3.8 there exist isometries UA ∈ L (GA, HA), VA ∈ L (GA, H), UD ∈ 10 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ L (GD, HD), VD ∈ L (GD, H) satisfying UASA = RA, VASA = RX , UDSD = RD, VDSD = RY . It follows where K := UAV ∗ D ∈ L (HD, HA) is a contraction with ran K ⊆ (ran RA)c. X RY = R∗ AUAV ∗ AVDU ∗ DRD = R∗ AKRD, B = R∗ AVDU ∗ Conversely, if (i) and (ii) are satisfied, then A = (cid:18) R∗ R∗ AKRD)∼ R∗ ARA AKRD DRD (cid:19) . (R∗ Since one obtains (R∗ AKRD)∼ = (R′ A = (cid:18)R∗ AKRD)′↾X= R′ D(cid:19)(cid:18) I K DK ′R′′ I (cid:19)(cid:18)RA A↾X= R∗ 0 RD(cid:19) , K ∗ 0 A 0 R∗ 0 DK ∗RA, which implies that A is non-negative. (cid:3) 6. Albert's theorem An application of Proposition 5.4 leads to a generalization of an important criterion for non-negativity [1], which is often called Albert's theorem in matrix theory. It should be mentioned that Shmulyan [17, Theorem 1.7] had proved a similar assertion even for bounded operators in Hilbert spaces ten years earlier. We also mention the papers [3] and [10]. it is of positive type and σ(A) is non-negative. Theorem 6.1. An operator A = (cid:16) A B Proof. If A ∈ L ≥, Proposition 5.4 implies that A is of positive type and R∗[−1] KRD for some contraction K ∈ L (HD, HA). It follows B∼ D(cid:17) ∈ L h is non-negative if and only if A B = σ(A) = R∗ DRD − (KRD)∗KRD = R∗ D(I − K ∗K)RD ≥ 0 Conversely, let A ∈ L + and σ(A) ∈ L ≥(Y, Y ′). If (RA, HA) and (RD, HD) are square roots of A and D, resp., one has kR∗[−1] A Byk2 = hω(A, B)y, yi ≤ hDy, yi = kRDyk2, y ∈ Y, which yields KRD = R∗[−1]B, hence, R∗ L (HD, HA). An application of Proposition 5.4 completes the proof. AKRD = B for some contraction K ∈ (cid:3) The preceding theorem can be used to study the set L ≥ as well as the set L + and to establish interrelations between these two sets. A first result is the inclusion L ≥ ⊆ L +. For a positive pair (A, B) set Aex := (cid:18) A B∼ ω(A, B)(cid:19) ∈ L +. B Corollary 6.2. Two operators A ∈ L (X, X ′) and B ∈ L (Y, X ′) form a positive pair if and only if the set A := nA ∈ L ≥ : A = (cid:16) A B B∼ D(cid:17) for some D ∈ L ≥(Y, Y ′)o A GENERALIZED SCHUR COMPLEMENT 11 is non-empty. If (A, B) is a positive pair, the operator Aex is the minimal element of A. (cid:3) Corollary 6.3. If A ∈ L ≥, the set is non-empty and S (A) is its maximal element. A1 := (cid:8)A1 ∈ L ≥ : A1 ≤ A and X ⊆ ker A1(cid:9) Proof. Corollaries 3.8 (i) and 3.9 imply that S (A) ∈ A1. If A = (cid:16) A B A1 ∈ A1, then A1 has representation A1 = (cid:16)0 0 Corollary 6.4. Let A = (cid:16) A B B∼ D(cid:17) and 0 D1(cid:17) and D − ω(A, B) − D1 ≥ 0, B∼ D(cid:17) ∈ L h. The operator A belongs to L + if and hence, A1 ≤ S (A) by Theorem 6.1. only if there exists an operator A1 ∈ L h satisfying X ⊆ ker A1 and A1 ≤ A. (cid:3) Proof. If A ∈ L +, the operator A1 := S (A) has all properties claimed. Con- versely, if there exists an operator A1 satisfying all conditions, it has the form A1 = (cid:16)0 0 0 D1(cid:17), where D1 ∈ L (Y, Y ′) and A − A1 = (cid:16) A follows from Theorem 6.1 that (A, B) is a positive pair, hence, A ∈ L +. B∼ D − D1(cid:17) ∈ L ≥. It B (cid:3) Another application of Theorem 6.1 gives an expression of the supremum oc- curing in Lemma 4.2. Corollary 6.5. If (A, B) is a positive pair, then hBy, xi2 hAx, xi sup x∈X = hω(A, B)y, yi, y ∈ Y. (6.1) Proof. Let y ∈ Y . Since Aex ∈ L ≥ by Corollary 6.2, one has hBy, xi2 ≤ hAx, xihω(A, B)y, yi, which yields hBy, xi2 hAx, xi ≤ hω(A, B)y, yi, x ∈ X, if one takes into account the convention 0 0 := 0. Thus, (6.1) has been proved if T y = R∗[−1]By = 0, where (R, H) is a minimal square root of A. Now assume that T y 6= 0. There exists a sequence {xn}n∈N of elements of X such that Rxn 6= 0, n ∈ N, and limn→∞ Rxn = T y. It follows lim n→∞ hBy, xni2 hAxn, xni = lim n→∞ hR∗T y, xni2 hR∗Rxn, xni = lim n→∞ (T y Rxn)2 kRxnk2 = kT yk2 = hω(A, B)y, yi. (cid:3) Corollary 6.6. Let (Aj, Bj) with Aj ∈ (X, X ′), Bj ∈ (Y, X ′), j = 1, 2, be positive pairs. Then (A1 + A2, B1 + B2) is a positive pair and ω(A1 + A2, B1 + B2) ≤ ω(A1, B1) + ω(A2, B2). (6.2) 12 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ Proof. Since the operators (Aj)ex, j = 1, 2 are non-negative, it follows (cid:18) A1 + A2 1 + B∼ hence, (6.2) by Corollary 6.2. B∼ 2 ω(A1, B1) + ω(A2, B2)(cid:19) ∈ L ≥, B1 + B2 Corollary 6.7. If Aj ∈ L +, j = 1, 2, then A1 + A2 ∈ L + and S (A1) + S (A2) ≤ S (A1 + A2). (cid:3) (cid:3) A subset A of L h is called bounded below if there exists A1 ∈ L h such that A1 ≤ A for all A ∈ A . An operator A0 ∈ L h is called an infimum of A if the following conditions are satisfied: a) A0 ≤ A for all A ∈ A , b) A1 ≤ A0 for all A1 ∈ L h such that A1 ≤ A, A ∈ A . If an infimum of A exists, it is unique. Recall that any set A , which is bounded from below and directed downwards (i.e. for all A1, A2 ∈ A there exists A ∈ A such that A ≤ A1 and A ≤ A2), possesses an infimum. Particularly, if {An}n∈N is a decreasing sequence of operators of L h, which is bounded from below, there exists an infimum A0 and hA0z1, z2i = limn→∞hAnz1, z2i for all z1, z2 ∈ X × Y . Corollary 6.8. Let A be a subset of L h, which has an infimum A0. The operator A0 belongs to L + if and only if the set S(A ) := {S (A) : A ∈ A } is bounded from below. In this case S (A0) is the infimum of S(A ). Proof. If A0 ∈ L +, the set S(A ) is bounded from below since S (A0) ≤ S (A), A ∈ A , by Corollary 4.8 (ii). Conversely, assume that there exists A1 ∈ L h such that A1 ≤ S (A) for all A ∈ A . It follows −A1 ≥ −S (A), which yields −A1 ∈ L + by Corollary 6.4 and S (−A1) ≥ S (−S (A)) = −S (A), hence, −S (−A1) ≤ S (A) ≤ A, A ∈ A , by Corollary 4.8. One obtains −S (−A1) ≤ A0 and therefore A0 ∈ L + by Corollary 6.4. Moreover, S (A0) ≤ S (A), A ∈ A , and A1 = −(−A1) ≤ −S (−A1) = S(cid:0)−S (−A1)(cid:1) ≤ S (A0) by Corollary 4.8, which implies that S (A0) is the infimum of S(A ). (cid:3) 7. Extremal operators An operator A ∈ L + was called an extremal operator by M.G. Kreın [11] if S (A) = 0. Since A = S (A) + Aex, an operator is extremal if and only if it has the form A = Aex = (cid:18) A B∼ ω(A, B)(cid:19) B for some positive pair (A, B). Particularly, any extremal operator is non-negative. Applying Proposition 4.7 we can give several criteria for an operator to be ex- tremal. Lemma 7.1. Let A ∈ L ≥. The following assertions are equivalent: (i) The operator is extremal. A GENERALIZED SCHUR COMPLEMENT 13 (ii) For all (cid:16)x y(cid:17) ∈ X × Y and arbitrary ε > 0 there exists z ∈ X such that DA(cid:16)x − z y (cid:17) ,(cid:16)x − z y (cid:17)E < ε. (iii) For any square root (R, H) of A the spaces (RX)c and (ran R)c coincide. (iv) For any square root (R, H) of A one has ran R∗ ∩ Y ′ = {0}. Proof. The equivalence of (i) and (ii) is an immediate consequence of (4.3). To prove (i) ⇔ (iii), choose a minimal square root (R, H) of A and let L and P be defined as in Proposition 5.1. Then S (A) = R∗P R = 0 if and only if P = 0 or, equivalently, L = {0}, which in turn is equivalent to (RX)c = (ran R)c. The equivalence of (iii) and (iv) follows from the equality R∗L = ran R∗ ∩ Y ′. (cid:3) Let (A, B) be a positive pair and (R, H) a square root of A. Recall the notation (4.1) of the operator T := R∗[−1]B. Moreover, let PB be the orthoprojection onto (ran T )c. Since the operator Aex is non-negative, from Theorem 6.1 one can conclude that (ω(A, B), B∼) is a positive pair as well changing the roles of X and Y . Thus, the operators T ∗[−1]B∼ and can be defined. ω(cid:0)ω(A, B), B∼(cid:1) = (cid:2)T ∗[−1]B∼(cid:3)∗T ∗[−1]B∼ Lemma 7.2. The equalities T ∗[−1]B∼ = PBR and ω(cid:0)ω(A, B), B∼(cid:1) = R∗PBR hold true. Proof. The second equality is an immediate consequence of the first one. To prove the first equality we shall show that (cid:0)T ∗[−1]B∼x h(cid:1) = (PBRx h) for all x ∈ X and h ∈ H. (7.1) Since ran T ∗[−1]B∼ ⊆ (ran T )c it is enough to prove (7.1) for x ∈ X and h ∈ ran T . If h = T y for some y ∈ Y , we get (cid:0)T ∗[−1]B∼x h(cid:1) = (cid:0)T ∗[−1]B∼x T y(cid:1) = (cid:10)T ∗T ∗[−1]B∼x, y(cid:11) = hB∼x, yi (PBRx h) = (PBRx T y) = (Rx T y) = hR∗T y, xi = hBy, xi = hB∼x, yi, and hence, (7.1). (cid:3) From Corollary 6.2 it follows that ω(cid:0)ω(A, B), B∼(cid:1) is a minimal element of the set Note also that nA1 ∈ L (X, X ′) : (cid:16) A1 B∼ ω(A, B)(cid:17) ∈ L ≥o B ω(cid:16)ω(cid:0)ω(A, B), B∼(cid:1), B(cid:17) = ω(A, B), cf. [12, Proposition 1.4 (A)]. We call an extremal operator Aex = (cid:16) A doubly extremal if ω(cid:0)ω(A, B), B∼(cid:1) = A. B B∼ ω(A, B)(cid:17) 14 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ In the case of bounded operators on Hilbert spaces the remaining results of the present section were proved by Pekarev and Shmulyan [13] and partly rediscovered by Niemiec [12]. We mention that Niemiec's proofs are based on Douglas' theorem and do not make explicit use of 2 × 2 block operators. Proposition 7.3. An operator Aex is doubly extremal if and only if ker T ∗ = ker R∗ (7.2) for any square root (R, H) of A. Proof. According to Lemma 7.2, A is doubly extremal if and only if R∗PBR = R∗R. If ker T ∗ = ker R∗ or, equivalently, (ran T )c = (ran R)c, it follows PBR = R, hence, R∗PBR = R∗R. Conversely, assume R∗PBR = R∗R, which yields kPBRxk = kRxk, x ∈ X, hence, (ran R)c ⊆ ran PB and ker PB ⊆ ker R∗. Since ran PB = (ran T )c or ker PB = ker T ∗, we get ker T ∗ ⊆ ker R∗. (7.3) On the other hand, if h ∈ ker R∗, then h is orthogonal to (ran R)c and thus, ker R∗ ⊆ ker T ∗. Taking into account (7.3) we obtain the desired equality. (cid:3) 0 = (cid:0)h T y(cid:1) = (cid:10)T ∗h, y(cid:11), y ∈ Y, The preceding assertion shows that the equality (7.2) does not depend on the choice of the square root (R, H) of A and that in the case of a minimal square root the operator Aex is doubly extremal if and only if ker T ∗ = {0}. Moreover, writing (7.2) in the equivalent form (ran T )c = (ran R)c we obtain a generalization of [13, Theorem 1.6]. It means that Aex is doubly extremal if and only if the inverse image of ran B under the map R∗ is dense in H. Corollary 7.4. If ran R∗ = ran B, the operator Aex is doubly extremal. (cid:3) To give another criterion for Aex to be doubly extremal we equip the space Y ′ with σ(Y ′, Y )-topology, i.e. the smallest topology such that for arbitrary y ∈ Y , the functional y′ 7→ hy′, yi is continuous on Y ′. Let (A, B) be a positive pair and (R, H) a square root of A. Denote by H1 the subspace of all h ∈ H such that there exists a sequence {xn}n∈N of elements of X with the following properties: a) limn→∞ Rxn = h with respect to the norm topology of H, b) limn→∞ B∼xn = 0 with respect to the σ(Y ′, Y )-topology. Lemma 7.5. The space H1 is equal to (ran R)c ∩ ker T ∗. Proof. An element h ∈ H belongs to H1 if and only if there exists a sequence {xn}n∈N of elements of X such that limn→∞ Rxn = h and for all y ∈ Y , (cid:10)T ∗h, y(cid:11) = lim n→∞(cid:0)Rxn T y(cid:1) = lim n→∞ (xn By) = lim n→∞ (B∼xn y) = 0. (cid:3) Proposition 7.6. An operator Aex is doubly extremal if and only if H1 = {0}. A GENERALIZED SCHUR COMPLEMENT 15 Proof. Since H1 = {0} if and only if ker R∗ = ker T ∗ by Lemma 7.5, the assertion follows from Proposition 7.3. (cid:3) Corollary 7.7. If an operator Aex is doubly extremal, then ker A = ker B∼. If ker A = ker B∼ and ran R is closed, then Aex is doubly extremal. Proof. If Aex is doubly extremal, then (cid:0)T ∗[−1]B∼, H(cid:1) is a square root of A, hence, ker B∼ ⊆ ker T ∗[−1]B∼ = ker A by Lemma 3.4. The first assertion of the corollary follows since ker A ⊆ ker B∼ by Lemma 4.2. Now assume that ker A = ker B∼ and ran R is closed. If h ∈ H1, there exist x ∈ X and a sequence {xn}n∈N of elements of X such that h = Rx = limn→∞ Rxn and limn→∞hB∼xn, yi = 0 for all y ∈ Y . It follows hB∼x, yi = (cid:10)R∗T y, x(cid:11) = (cid:0)Rx T y(cid:1) n→∞(cid:0)Rxn T y(cid:1) = lim = lim n→∞ hBy, xni = lim n→∞ hB∼xn, yi = 0, y ∈ Y, which implies that x ∈ ker B∼ = ker A = ker R and h = 0. An application of Proposition 7.6 completes the proof. (cid:3) References 1. Arthur Albert, Conditions for positive and nonnegative definiteness in terms of pseudoin- verses, SIAM J. Appl. Math. 17 (1969), 434 -- 440. 2. W. N. Anderson, Jr. and G. E. Trapp, Shorted operators. II, SIAM J. Appl. Math. 28 (1975), 60 -- 71. 3. T. Ando, Truncated moment problems for operators, Acta Sci. Math. (Szeged) 31 (1970), 319 -- 334. 4. Douglas E. Crabtree and Emilie V. Haynsworth, An identity for the Schur complement of a matrix, Proc. Amer. Math. Soc. 22 (1969), 364 -- 366. 5. R. G. Douglas, On majorization, factorization, and range inclusion of operators on Hilbert space, Proc. Amer. Math. Soc. 17 (1966), 413 -- 415. 6. Bernd Fritzsche, Bernd Kirstein, and Lutz Klotz, Completion of non-negative block opera- tors in Banach spaces, Positivity 3 (1999), no. 4, 389 -- 397. 7. J. G´orniak, Locally convex spaces with factorization property, Colloq. Math. 48 (1984), no. 1, 69 -- 79. 8. J. G´orniak and A. Weron, Aronszajn-Kolmogorov type theorems for positive definite kernels in locally convex spaces, Studia Math. 69 (1980/81), no. 3, 235 -- 246. 9. Janusz G´orniak, Andrzej Makagon, and Aleksander Weron, An explicit form of dilation the- orems for semispectral measures, Prediction theory and harmonic analysis, North-Holland, Amsterdam-New York, 1983, pp. 85 -- 111. 10. Hans-Peter Hoschel, Uber die Pseudoinverse eines zerlegten positiven linearen Operators, Math. Nachr. 74 (1976), 167 -- 172. 11. M. G. Kreın, The theory of self-adjoint extensions of semi-bounded Hermitian transforma- tions and its applications. I, Rec. Math. [Mat. Sbornik] N.S. 20(62) (1947), 431 -- 495. 12. Piotr Niemiec, Generalized absolute values and polar decompositions of a bounded operator, Integral Equations Operator Theory 71 (2011), no. 2, 151 -- 160. 13. `E. L. Pekarev and Yu. L. Shmulyan, Parallel addition and parallel subtraction of operators, Izv. Akad. Nauk SSSR Ser. Mat. 40 (1976), no. 2, 366 -- 387, 470. 14. W. Pusz and S. L. Woronowicz, Functional calculus for sesquilinear forms and the purifi- cation map, Rep. Mathematical Phys. 8 (1975), no. 2, 159 -- 170. 16 J. FRIEDRICH, M. G UNTHER, and L. KLOTZ 15. Laurent Schwartz, Sous-espaces hilbertiens d'espaces vectoriels topologiques et noyaux as- soci´es (noyaux reproduisants), J. Analyse Math. 13 (1964), 115 -- 256. 16. Zolt´an Sebesty´en, Operator extensions on Hilbert space, Acta Sci. Math. (Szeged) 57 (1993), no. 1-4, 233 -- 248. 17. Yu. L. Shmulyan, An operator Hellinger integral, Mat. Sb. (N.S.) 49 (91) (1959), 381 -- 430. 18. 19. Zs. Tarcsay, On the parallel sum of positive operators, forms, and functionals, Acta Math. , Two-sided division in the ring of operators, Mat. Zametki 1 (1967), 605 -- 610. Hungar. 147 (2015), no. 2, 408 -- 426. 20. T. Titkos, On means of nonnegative sesquilinear forms, Acta Math. Hungar. 143 (2014), no. 2, 515 -- 533. 21. N. N. Vakhaniya, Probabilistic distributions in linear spaces, Sakharth. SSR Mecn. Akad. Gamothvl. Centr. Srom. 10 (1971), no. 3, 155, Russian, English translation: North-Holland Publishing Co., New York-Amsterdam 1981. 22. N. N. Vakhaniya, V. I. Tarieladze, and S. A. Chobanyan, Probability Distributions in banach Spaces, Nauka, Moscow, 1985, Russian, English translation: D. Reidel Publishing Co, Dordrecht 1987. 23. M. M. Vaınberg, Variational method and method of monotone operators in the theory of nonlinear equations, Nauka, Moscow, 1972, Russian, English Translation: Halsted Press, New York-Toronto 1973. 24. Fuzhen Zhang (ed.), The Schur Complement and its Applications, Numerical Methods and Algorithms, vol. 4, Springer-Verlag, New York, 2005. 1Stauffenbergstr. 10, D-04509 Delitzsch, Germany. E-mail address: [email protected] 2Mathematisches Institut, Universitat Leipzig, PF 10 09 20, D-04009 Leipzig, Germany. E-mail address: [email protected]; [email protected]
1103.5357	2	1103	2012-03-08T13:53:16	Spaces of variable smoothness and integrability: Characterizations by local means and ball means of differences	[ "math.FA" ]	We study the spaces of Besov and Triebel-Lizorkin type with variable smoothness and integrability as introduced recently by Almeida & H\"ast\"o and Diening, H\"ast\"o & Roudenko. Both scales cover many classical spaces with fixed exponents as well as function spaces of variable smoothness and function spaces of variable integrability. These spaces have been introduced by Fourier analytical tools, as the decomposition of unity. Surprisingly, our main result states that these spaces also allow a characterization in the time-domain with the help of classical ball means of differences. To that end, we first prove a local means characterization for them with the help of the so-called Peetre maximal functions. Our results do also hold for 2-microlocal function spaces with variable integrability which are a slight generalization of generalized smoothness spaces and spaces of variable smoothness.	math.FA	math	Spaces of variable smoothness and integrability: Characterizations by local means and ball means of differences Henning Kempka∗, Jan Vyb´ıral† October 11, 2018 Abstract p(·),q(·)(Rn) and F s(·) We study the spaces B s(·) p(·),q(·)(Rn) of Besov and Triebel- Lizorkin type as introduced recently in [1] and [17]. Both scales cover many clas- sical spaces with fixed exponents as well as function spaces of variable smooth- ness and function spaces of variable integrability. The spaces B s(·) p(·),q(·)(Rn) have been introduced in [1] and [17] by Fourier analytical tools, as the decomposition of unity. Surprisingly, our main result states that these spaces also allow a characterization in the time- domain with the help of classical ball means of differences. p(·),q(·)(Rn) and F s(·) To that end, we first prove a local means characterization for B s(·) p(·),q(·)(Rn) with the help of the so-called Peetre maximal functions. Our results do also hold for 2-microlocal function spaces Bw p(·),q(·)(Rn) which are a slight generalization of generalized smoothness spaces and spaces of variable smoothness. p(·),q(·)(Rn) and F w Key words: Besov spaces, Triebel-Lizorkin spaces, variable smoothness, variable integrability, ball means of differences, Peetre maximal operator, 2-microlocal spaces. 2010 MSC: 46E35, 46E30, 42B25. 1 Introduction Function spaces of variable integrability appeared in a work by Orlicz [42] already in 1931, but the recent interest in these spaces is based on the paper of Kov´acik and R´akosnik [33] together with applications in terms of modelling electrorheological fluids [46]. A fundamental breakthrough concerning spaces of variable integrability was the observation that, under certain regularity assumptions on p(·), the Hardy- Littlewood maximal operator is also bounded on Lp(·)(Rn), see [13]. This result has been generalized to wider classes of exponents p(·) in [11], [41] and [15]. ∗Mathematical Institute, Friedrich-Schiller-University Jena, D -- 07737 Jena, Germany, email: [email protected] †Johann Radon Institute for Computational and Applied Mathematics, Austrian Academy of Sciences, Altenbergerstrasse 69, A -- 4040 Linz, Austria, email: [email protected]. 1 Besides electrorheological fluids, the spaces Lp(·)(Rn) possess interesting applications in the theory of PDE's, variational calculus, financial mathematics and image pro- cessing. A recent overview of this vastly growing field is given in [16]. Sobolev and Besov spaces with variable smoothness but fixed integrability have been introduced in the late 60's and early 70's in the works of Unterberger [58], Visik and Eskin [59], Unterberger and Bokobza [57] and in the work of Beauzamy [7]. Leopold studied in [34] Besov spaces where the smoothness is determined by a symbol a(x, ξ) of a certain class of hypoelliptic pseudodifferential operators. In the special case a(x, ξ) = (1 + ξ2)σ(x)/2 these spaces coincide with spaces of variable smoothness Bσ(x) A more general approach to spaces of variable smoothness are the so-called 2- p,q(Rn). Here the smoothness in these microlocal function spaces Bw spaces gets measured by a weight sequence w = (wj)∞ j=0. Besov spaces with such weight sequences appeared first in the works of Peetre [43] and Bony [9]. Estab- lishing a wavelet characterization for 2-microlocal Holder-Zygmund spaces in [24] it turned out that 2-microlocal spaces are well adapted in connection to regularity properties of functions ([25],[38],[36]). Spaces of variable smoothness are a special case of 2-microlocal function spaces and in [35] and [8] characterizations by differ- ences have been given for certain classes of them. p,q(Rn) and F w p,p (Rn). The theories of function spaces with fixed smoothness and variable integrability and function spaces with variable smoothness and fixed integrability finally crossed each other in [17], where the authors introduced the function spaces of Triebel-Lizorkin type with variable smoothness and simultaneously with variable integrability. It turned out that many of the spaces mentioned above are really included in this new structure, see [17] and references therein. The key point to merge both lines of investigation was the study of traces. From Theorem 3.13 in [17] trRn−1F s(·) p(·),q(·)(Rn) = F s(·)−1/p(·) p(·),p(·) (Rn−1) one immediately understands the necessity to take all exponents variable assuming p(·) or s(·) variable. So the trace embeddings may be described in a natural way in the context of these spaces. Furthermore, this was complemented in [60] by showing, that the classical Sobolev embedding theorem p0(·),q(·)(Rn) ֒→ F s1(·) F s0(·) p1(·),q(·)(Rn) holds also in this scale of function spaces if the usual condition is replaced by its point-wise analogue s0(x) − n/p0(x) = s1(x) − n/p1(x), x ∈ Rn. Finally, Almeida and Hasto managed in [1] to adapt the definition of Besov spaces to the setting of variable smoothness and integrability and proved the Sobolev and other usual embeddings in this scale. The properties of Besov and Triebel-Lizorkin spaces of variable smoothness and integrability known so far give a reasonable hope that these new scales of function 2 spaces enjoy sufficiently many properties to allow a local description of many ef- fects, which up to now could only be described in a global way. Subsequently, for the spaces F s(·) p(·),q(·)(Rn) there is a characterization by local means given in [30]. This characterization still works with Fourier analytical tools but the analyzing functions k0, k ∈ S(Rn) are compactly supported in the time-domain and we only need local values of f around x ∈ Rn to calculate the building blocks k(2−j, f )(x). This is in sharp contrast to the definition of the spaces by the decomposition of unity, cf. Def- initions 1 and 3. For the spaces Bs(·) p(·),q(·)(Rn) we will prove a local means assertion of this type in Section 3 which will be helpful later on. The main aim of this paper is to present another essential property of the func- tion spaces from [17] and [1]. We prove the surprising result that these spaces p(·),q(·)(Rn) and F s(·) Bs(·) p(·),q(·)(Rn) with variable smoothness and integrability do also allow a characterization purely in the time-domain by classical ball means of differ- ences. The paper is organized as follows. First of all we provide all necessary notation in Section 2. Since the proofs for spaces of variable smoothness and 2-microlocal function spaces work very similar (see Remark 2) we present our results for both scales. The proof for the local means characterization will be given in Section 3 in terms of 2-microlocal function spaces and we present the version for spaces of variable smoothness in Section 3.2. In Section 4 we prove the characterization by ball means of differences for Bs(·) p(·),q(·)(Rn) and the version for 2- microlocal function spaces will be given in Section 4.5. p(·),q(·)(Rn) and F s(·) 2 Notation In this section we collect all the necessary definitions. We start with the variable Lebesgue spaces Lp(·)(Rn). A measurable function p : Rn → (0, ∞] is called a vari- able exponent function if it is bounded away from zero, i.e. if p− = ess-inf x∈Rn p(x) > 0. We denote the set of all variable exponent functions by P(Rn). We put also p+ = ess-supx∈Rn p(x). The variable exponent Lebesgue space Lp(·)(Rn) consists of all measurable func- tions f for which there exist λ > 0 such that the modular is finite, where If we define Rn ∞, then the Luxemburg ∞ = {x ∈ Rn : p(x) = ∞} and Rn 0 = Rn \ Rn ϕp(x)(cid:18) f (x) λ (cid:19) dx Lp(·)(Rn)(f /λ) =ZRn ϕp(t) =  tp if p ∈ (0, ∞), if p = ∞ and t ≤ 1, 0 ∞ if p = ∞ and t > 1. 3 norm of a function f ∈ Lp(·)(Rn) is given by (cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) = inf{λ > 0 : Lp(·)(Rn)(f /λ) ≤ 1} λ (cid:19)p(x) = inf(λ > 0 :ZRn 0 (cid:18) f (x) dx < 1 and f (x) < λ for a.e. x ∈ Rn ∞) . If p(·) ≥ 1, then it is a norm otherwise it is always a quasi-norm. To define the mixed spaces ℓq(·)(Lp(·)) we have to define another modular. For p, q ∈ P(Rn) and a sequence (fν)ν∈N0 of Lp(·)(Rn) functions we define ℓq(·)(Lp(·))(fν) = inf(λν > 0 : p(·) fν λ1/q(·) ν ! ≤ 1) . ∞ Xν=0 (1) If q+ < ∞, then we can replace (1) by the simpler expression ℓq(·)(Lp(·))(fν ) =Xν (cid:13)(cid:13)(cid:13)(cid:13) fνq(·)(cid:12)(cid:12)(cid:12) L p(·) . q(·)(cid:13)(cid:13)(cid:13)(cid:13) The (quasi-)norm in the ℓq(·)(Lp(·)) spaces is defined as usual by (cid:13)(cid:13) fν ℓq(·)(Lp(·))(cid:13)(cid:13) = inf{µ > 0 : ℓq(·)(Lp(·))(fν/µ) ≤ 1}. It is known, cf. [1, 32], that ℓq(·)(Lp(·)) is a norm if q(·) ≥ 1 is constant almost everywhere (a.e.) on Rn and p(·) ≥ 1, or if 1/p(x) + 1/q(x) ≤ 1 a.e. on Rn, or if 1 ≤ q(x) ≤ p(x) ≤ ∞ a.e. on Rn. Surprisingly enough, it turned out in [32] that the condition min(p(x), q(x)) ≥ 1 a.e. on Rn is not sufficient for ℓq(·)(Lp(·)) to be a norm. Nevertheless, it was proven in [1] that it is a quasi-norm for every p, q ∈ P(Rn). For the sake of completeness, we state also the definition of the space Lp(·)(ℓq(·)), which is much more intuitive then the definition of ℓq(·)(Lp(·)). One just takes the ℓq(x) norm of (fν(x))ν∈N0 for every x ∈ Rn and then the Lp(·)-norm with respect to x ∈ Rn, i.e. (cid:13)(cid:13) fν Lp(·)(ℓq(·))(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13) fν(x) ℓq(x)(cid:13)(cid:13) Lp(·)(cid:13)(cid:13). It is easy to show ([17]) that Lp(·)(ℓq(·)) is always a quasi-normed space and it is a normed space, if min(p(x), q(x)) ≥ 1 holds point-wise. The summation in the definition of the norms of ℓq(·)(Lp(·)) and Lp(·)(ℓq(·)) can also be taken for ν ∈ Z. It always comes out of the context over which interval the summation is taken. Occasionally, we may indicate it by (cid:13)(cid:13) (fν)∞ By f = Ff and f ∨ = F −1f we denote the usual Fourier transform and its inverse on S(Rn), the Schwartz space of smooth and rapidly decreasing functions, and on S ′(Rn), the dual of the Schwartz space. ν=−∞(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13). 2.1 Spaces Bs(·) p(·),q(·)(Rn) and F s(·) p(·),q(·)(Rn) The definition of Besov and Triebel-Lizorkin spaces of variable smoothness and in- tegrability is based on the technique of decomposition of unity exactly in the same manner as in the case of constant exponents. 4 Definition 1. Let ϕ0 ∈ S(Rn) with ϕ0(x) = 1 for x ≤ 1 and supp ϕ0 ⊆ {x ∈ Rn : x ≤ 2}. For j ≥ 1 we define ϕj(x) = ϕ0(2−j x) − ϕ0(2−j+1x). One may verify easily that ϕj(x) = 1 for all x ∈ Rn. ∞ Xj=0 The following regularity classes for the exponents are necessary to make the defini- tion of the spaces independent on the chosen decomposition of unity. Definition 2. Let g ∈ C(Rn). (i) We say that g is locally log-Holder continuous, abbreviated g ∈ C log loc (Rn), if there exists clog(g) > 0 such that g(x) − g(y) ≤ clog(g) log(e + 1/x − y) (2) holds for all x, y ∈ Rn. (ii) We say that g is globally log-Holder continuous, abbreviated g ∈ C log(Rn), if g is locally log-Holder continuous and there exists g∞ ∈ R such that g(x) − g∞ ≤ clog log(e + x) holds for all x ∈ Rn. Remark 1. With (2) we obtain g(x) ≤ clog(g) + g(0), for all x ∈ Rn. This implies that all functions g ∈ C log loc (Rn) always belong to L∞(Rn) If an exponent p ∈ P(Rn) satisfies 1/p ∈ C log(Rn), then we say it belongs to the p(·),q(·)(Rn), class P log(Rn). We recall the definition of the spaces Bs(·) as given in [17] and [1]. p(·),q(·)(Rn) and F s(·) Definition 3. (i) Let p, q ∈ P log(Rn) with 0 < p− ≤ p+ < ∞, 0 < q− ≤ q+ < ∞ and let s ∈ C log loc (Rn). Then (ii) Let p, q ∈ P log(Rn) and let s ∈ C log < ∞(cid:27) , where < ∞(cid:27) , where f F s(·) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ . Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) loc (Rn). Then f Bs(·) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ . ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) f F s(·) f Bs(·) F s(·) =(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn) =(cid:26)f ∈ S ′(Rn) :(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ 2js(·)(ϕj f )∨(cid:12)(cid:12)(cid:12) p(·),q(·)(Rn) =(cid:26)f ∈ S ′(Rn) :(cid:13)(cid:13)(cid:13) 2js(·)(ϕj f )∨(cid:12)(cid:12)(cid:12) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ =(cid:13)(cid:13)(cid:13) Bs(·) 5 (cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) p(·),q(·)(Rn) and Bs(·) The subscript ϕ at the norm symbolizes that the definition formally does de- pend on the resolution of unity. From [30] and [1] we have that the definition of the spaces F s(·) p(·),q(·)(Rn) is independent of the chosen resolution of unity if p, q ∈ P log(Rn) and s ∈ C log loc (Rn). That means that different start functions ϕ0 and ϕ0 from Definition 1 induce equivalent norms in the above definition. So we will suppress the subscript ϕ in the notation of the norms. Let us comment on the conditions on p, q ∈ P log(Rn) for the Triebel-Lizorkin spaces. The condition 0 < p− ≤ p+ < ∞ is quite natural since there exists also the restric- tion p < ∞ in the case of constant exponents, see [52] and [64]. The second one, 0 < q− ≤ q+ < ∞, is a bit unnatural and comes from the use of the convolution Lemma 21 ([17, Theorem 3.2]). There is some hope that this convolution lemma can be generalized and the case q+ = ∞ can be incorporated in the definition of the F -spaces. The Triebel-Lizorkin spaces with variable smoothness have first been introduced in [17] under much more restrictive conditions on s(·). These conditions have been relaxed in [30] in the context of 2-microlocal function spaces (see the next subsec- tion). Besov spaces with variable p(·), q(·) and s(·) have been introduced in [1]. Both scales contain as special cases a lot of well known function spaces. If s, p and q are constants, then we derive the well known Besov and Triebel-Lizorkin spaces with usual Holder and Sobolev spaces included, see [52] and [53]. If the smoothness s ∈ R is a constant and p ∈ P log(Rn) with p− > 1, then F s p(·)(Rn) are the variable Bessel potential spaces from [2] and [23] with its special cases F 0 p(·)(Rn) for k ∈ N0, see [17]. p(·),2(Rn) = Lp(·)(Rn) and F k Taking s ∈ R and q ∈ (0, ∞] as constants we derive the spaces F s Bs Furthermore it holds F s(·) ∞,∞(Rn) equals the variable Holder-Zygmund space Cs(·)(Rn) introduced in [3], [4] and [45] with 0 < s− ≤ s+ ≤ 1, see [1]. p(·),q(Rn) studied by Xu in [62] and [63]. p(·),p(·)(Rn) = Bs(·) p(·),p(·)(Rn) and Bs(·) p(·),2(Rn) = W k p(·),2(Rn) = Ls p(·),q(Rn) and 2.2 2-microlocal spaces The definition of Besov and Triebel-Lizorkin spaces of variable smoothness and inte- grability is a special case of the so-called 2-microlocal spaces of variable integrability. As some of the results presented here get proved in this more general scale, we present also the definition of 2-microlocal spaces. It is based on the dyadic decomposition of unity as presented above combined with the concept of admissible weight sequences. Definition 4. Let α ≥ 0 and let α1, α2 ∈ R with α1 ≤ α2. A sequence of non- negative measurable functions w = (wj)∞ α1,α2 if and only if j=0 belongs to the class W α (i) there exists a constant C > 0 such that 0 < wj(x) ≤ Cwj(y)(cid:0)1 + 2jx − y(cid:1)α 6 for all j ∈ N0 and all x, y ∈ Rn (ii) and for all j ∈ N0 and all x ∈ Rn we have 2α1wj(x) ≤ wj+1(x) ≤ 2α2wj(x). Such a system (wj)∞ j=0 ∈ W α α1,α2 is called an admissible weight sequence. Finally, here is the definition of the spaces under consideration. Definition 5. Let w = (wj)j∈N0 ∈ W α p+, q+ < ∞ in the F -case), then we define α1,α2. Further, let p, q ∈ P log(Rn) (with f Bw (cid:13)(cid:13)(cid:13) and Bw =(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn) =(cid:26)f ∈ S ′(Rn) :(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ wj(ϕj f )∨(cid:12)(cid:12)(cid:12) p(·),q(·)(Rn) =(cid:26)f ∈ S ′(Rn) :(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ wj(ϕj f )∨(cid:12)(cid:12)(cid:12) =(cid:13)(cid:13)(cid:13) F w f F w (cid:13)(cid:13)(cid:13) < ∞(cid:27) , where < ∞(cid:27) , where f Bw p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) f F w p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13)ϕ . Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) The independence of the decomposition of unity for the 2-microlocal spaces from Definition 5 follows from the local means characterization (see [30] for the Triebel- Lizorkin and Section 3 for the Besov spaces). The 2-microlocal spaces with the special weight sequence wj(x) = 2js(1 + 2jx − x0)s′ with s, s′ ∈ R and x0 ∈ Rn (3) x0 = Bw x0 = Bw ∞,∞(Rn) and H s,s′ have first been introduced by Peetre in [43] and by Bony in [9]. Later on, Jaf- fard and Meyer gave a characterization in [24] and [25] with wavelets of the spaces C s,s′ 2,2(Rn) with the weight sequence (3). It turned out that spaces of this type are very useful to study regularity properties of functions. Subsequently, L´evy-V´ehel and Seuret developed in [36] the 2-microlocal formalism and studied the behavior of cusps, chirps and fractal functions with respect to the spaces C s,s′ x0 . A first step to a more general weight sequence w has been taken by Moritoh and Yamada in [39] and wider ranges of function spaces have been studied by Xu in [61] and by Andersson in [5]. The above definition for 2-microlocal weight sequences was presented by Besov in [8] and also in [30] by Kempka. A different line of study for spaces of variable smoothness - using different methods - are the spaces of generalized smoothness introduced by Goldman and Kalyabin in [20], [21], [26] and [27]. A systematic treatment of these spaces based on differences has been given by Goldman in [22], see also the survey [28] and references therein. Later on, spaces of generalized smoothness appeared in interpolation theory and have been investigated in [37], [10] and [40]. For further information on these spaces see the survey paper [19] where also a characterization by atoms and local means 7 for these spaces is given. From the definition of admissible weight sequences, d1σj ≤ σj+1 ≤ d2σj, it follows (σj ) p,q (Rn) of Farkas directly that the spaces of generalized smoothness B and Leopold [19] and B(s,Ψ) (Rn) from Moura [40] are a special sub- class of 2-microlocal function spaces with 2α1 = d1, 2α2 = d2 and α = 0. In a different approach Schneider in [51] studied spaces of varying smoothness. Here the smoothness at a point gets determined by a global smoothness s0 ∈ R and a local smoothness function s(·). These spaces can not be incorporated into the scale of 2-microlocal function spaces, but there exist some embeddings. (Rn) and F (s,Ψ) (σj ) p,q (Rn) and F p,q p,q Remark 2. Surprisingly, these 2-microlocal weight sequences are directly connected to variable smoothness functions s : Rn → R if we set wj(x) = 2js(x). (4) If s ∈ C log loc (Rn) (which is the standard condition on s(·)), then w = (wj(x))j∈N0 = (2js(x))j∈N0 belongs to W α α1,α2 with α1 = s− and α2 = s+. For the third index α we use Lemma 19 with m = 0 and obtain α = clog(s), where clog(s) is the constant for s(·) from (2). That means that spaces of variable smoothness from Definition 3 are a special case of 2-microlocal function spaces from Definition 5. Both types of function spaces are very closely connected and the properties used in the proofs are either 2ks(x)−s(y) ≤ c or wk(x) wk(y) ≤ c (5) loc (Rn) or from Definition 4. for x − y ≤ c2−k. This property follows directly either from the definition of s ∈ C log Nevertheless there exist examples of admissible weight sequences which can not be expressed in terms of variable smoothness functions. For example the important and well studied case of the weight sequence w from (3) can not be expressed via (4) if s′ 6= 0. Another example are the spaces of generalized smoothness which can not be identified as spaces of variable smoothness. Since spaces of variable smoothness are included in the scale of 2-microlocal function spaces all special cases of the previous subsection can be identified in the definition of 2-microlocal spaces. Although the 2-microlocal spaces include the scales of spaces of variable smoothness, we will give some of our proofs in the notation of variable smoothness, since this notation is more common. We will then reformulate the results in terms of 2- microlocal spaces, the proof works then very similar; we just have to use (5). 3 Local means characterization The main result of this section is the local means characterization of the spaces p(·),q(·)(Rn) there already exists a local means charac- Bw terization ([30, Corollary 4.7]). We shall first give the full proof for the 2-microlocal p(·),q(·)(Rn). For the spaces F w 8 spaces and later on (in Section 3.2) we restate the result also for spaces Bs(·) and F s(·) p(·),q(·)(Rn). p(·),q(·)(Rn) The crucial tool will be the Peetre maximal operator, as defined by Peetre in [43]. The operator assigns to each system (Ψk)k∈N0 ⊂ S(Rn), to each distribution f ∈ S ′(Rn) and to each number a > 0 the following quantities (Ψ∗ kf )a(x) := sup y∈Rn (Ψk ∗ f )(y) 1 + 2k(y − x)a , x ∈ Rn and k ∈ N0. (6) We start with two given functions ψ0, ψ1 ∈ S(Rn). We define ψj(x) = ψ1(2−j+1x), for x ∈ Rn and j ∈ N. Furthermore, for all j ∈ N0 we write Ψj = ψj. The main theorem of this section reads as follows. Theorem 6. Let w = (wk)k∈N0 ∈ W α with R > α2. Further, let ψ0, ψ1 belong to S(Rn) with α1,α2, p, q ∈ P log(Rn) and let a > 0, R ∈ N0 Dβψ1(0) = 0, for 0 ≤ β < R, and ψ0(x) > 0 on {x ∈ Rn : x < ε} ψ1(x) > 0 on {x ∈ Rn : ε/2 < x < 2ε} for some ε > 0. For a > n+clog(1/q) p− + α and all f ∈ S ′(Rn) we have (7) (8) (9) f Bw (cid:13)(cid:13)(cid:13) Remark 3. p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13) (Ψk ∗ f )wk ℓq(·)(Lp(·))(cid:13)(cid:13) ≈(cid:13)(cid:13) (Ψ∗ kf )awk ℓq(·)(Lp(·))(cid:13)(cid:13) . (i) The proof relies on [47] and will be shifted to the next section. Moreover, Theorem 6 shows that the definition of the 2-microlocal spaces of variable integrability is independent of the resolution of unity used in the Definition 5. (ii) The conditions (7) are usually called moment conditions while (8) and (9) are the so called Tauberian conditions. (iii) If R = 0, then there are no moment conditions (7) on ψ1. (iv) The notation clog(1/q) stands for the constant from (2) with 1/q(·). Next we reformulate the abstract Theorem 6 in the sense of classical local means (see Sections 2.4.6 and 2.5.3 in [53]). Since the proof is the same as the one from Theorem 2.4 in [30] we just state the result. Corollary 1. There exist functions k0, k ∈ S(Rn) with supp k0, supp k ⊂ {x ∈ Rn : x < 1} and Dβ k(0) = 0 for all 0 ≤ β < α2 such that for all f ∈ S ′(Rn) (cid:13)(cid:13) k0(1, f )w0 Lp(·)(Rn)(cid:13)(cid:13) +(cid:13)(cid:13) k(2−j, f )wj(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13) 9 is an equivalent norm on Bw The building blocks get calculated by p(·),q(·)(Rn). k(t, f )(x) =ZRn k(y)f (x + ty)dy = t−nZRn k(cid:18) y − x t (cid:19) f (y)dy and similarly for k0(1, f )(x). A similar characterization for F w p(·),q(·)(Rn) and details how these functions k0, k ∈ S(Rn) can be constructed can be found in [30]. 3.1 Proof of local means The proof of Theorem 6 is divided into three parts. The next section is devoted to some technical lemmas needed later. Section 3.1.2 is devoted to the proof of Theorem 12, which gives an inequality between different Peetre maximal operators. Finally, Section 3.1.3 proves the boundedness of the Peetre maximal operator in Theorem 13. These two theorems combined give immediately the proof of Theorem 6. 3.1.1 Helpful lemmas Before proving the local means characterization we recall some technical lemmas, which appeared in the paper of Rychkov [47]. For some of them we need adapted versions to our situation. The first lemma describes the use of the so called moment conditions. Lemma 7 ([47], Lemma 1). Let g, h ∈ S(Rn) and let M ∈ N0. Suppose that (Dβ g)(0) = 0 for 0 ≤ β < M. Then for each N ∈ N0 there is a constant CN such that (gt ∗ h)(z)(1 + zN ) ≤ CN tM , for 0 < t < 1, sup z∈Rn where gt(x) = t−ng(x/t). The next lemma is a discrete convolution inequality. We formulate it in a rather abstract notation and point out later on the conclusions we need. Lemma 8. Let X ⊂ {(fk)k∈Z : fk : Rn → [−∞, ∞] measurable} be a quasi-Banach space of sequences of measurable functions. Further we assume that its quasi-norm is shift-invariant, i.e. it satisfies k (fk+l)k∈Z Xk = k (fk)k∈Z Xk for every l ∈ Z and (fk)k∈Z ∈ X. For a sequence of non-negative functions (gk)k∈Z ∈ X and δ > 0 we denote Gν (x) = ∞ Xk=−∞ 2−ν−kδgk(x), x ∈ Rn, ν ∈ Z. 10 Then there exists a constant c > 0 depending only on δ and X such that for every sequence (gk)k∈Z k (Gν )ν Xk ≤ c k (gk)k Xk . Proof. Since X is a quasi-Banach space, there exists a r > 0 such that k · Xk is equivalent to some r-norm, cf. [6, 44]. We have then the following k (Gν )ν Xkr =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) .Xl∈Z r (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) X(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) r (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) X(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞ Xk=−∞ 2−ν−kδgk! ν Xl∈Z 2−lδgν+l!ν 2−lrδ k (gν+l)ν Xkr ≤ c k (gν)ν Xkr . Now taking the power 1/r yields the desired estimate. The spaces Lp(·)(ℓq(·)) and ℓq(·)(Lp(·)) are quasi-Banach spaces which fulfill the conditions of Lemma 8. Therefore, we obtain the following Lemma 9. Let p, q ∈ P(Rn) and δ > 0. Let (gk)k∈Z be a sequence of non-negative measurable functions on Rn and denote Gν(x) =Xk∈Z 2−ν−kδgk(x), x ∈ Rn, ν ∈ Z. Then there exist a constants C1, C2 > 0, depending on p(·), q(·) and δ, such that (cid:13)(cid:13) Gν ℓq(·)(Lp(·))(cid:13)(cid:13) ≤ C1(cid:13)(cid:13) gk ℓq(·)(Lp(·))(cid:13)(cid:13) (cid:13)(cid:13) Gν Lp(·)(ℓq(·))(cid:13)(cid:13) ≤ C2(cid:13)(cid:13) gk Lp(·)(ℓq(·))(cid:13)(cid:13) . and Remark 4. Of course, Lemma 9 holds true also if the indices k and ν run only over natural numbers. Since the maximal operator is in general not bounded on ℓq(·)(Lp(·)) (see [1, Example 4.1]) we need a replacement for that. It turned out that a convolution with radial decreasing functions fits very well into the scheme. A careful evaluation of the proof in [1, Lemma 4.7] together with Lemma 19 gives us the following convolution inequality. Lemma 10. Let p, q ∈ P log(Rn) with p(·) ≥ 1 and let ην,m(x) = 2nν (1 + 2ν x)−m. For all m > n + clog(1/q) there exists a constant c > 0 such that for all sequences (fj)j∈N0 ∈ ℓq(·)(Lp(·)) it holds (cid:13)(cid:13) (ην,m ∗ fν)ν∈N0 ℓq(·)(Lp(·))(cid:13)(cid:13) ≤ c(cid:13)(cid:13) (fj)j∈N0 ℓq(·)(Lp(·))(cid:13)(cid:13) . The last technical lemma is overtaken literally from [47]. Lemma 11 ([47], Lemma 3). Let 0 < r ≤ 1 and let (γν )ν∈N0, (βν )ν∈N0 be two sequences taking values in (0, ∞). Assume that for some N 0 ∈ N0, lim sup ν→∞ γν 2νN 0 < ∞. 11 (10) Furthermore, we assume that for any N ∈ N γν ≤ CN ∞ Xk=0 holds, then for any N ∈ N 2−kN βk+νγ1−r k+ν, ν ∈ N0, CN < ∞ γr ν ≤ CN ∞ Xk=0 2−kN rβk+ν, ν ∈ N0 holds with the same constants CN . 3.1.2 Comparison of different Peetre maximal operators In this subsection we present an inequality between different Peetre maximal oper- ators. Let us recall the notation given before Theorem 6. For two given functions ψ0, ψ1 ∈ S(Rn) we define ψj(x) = ψ1(2−j+1x), for x ∈ Rn and j ∈ N. Furthermore, for all j ∈ N0 we write Ψj = ψj and in an analogous manner we define Φj from two starting functions φ0, φ1 ∈ S(Rn). Using this notation we are ready to formulate the theorem. Theorem 12. Let w = (wj)j∈N0 ∈ W α R ∈ N0 with R > α2, α1,α2, p, q ∈ P(Rn) and a > 0. Moreover, let Dβψ1(0) = 0, 0 ≤ β < R φ0(x) > 0 on {x ∈ Rn : x < ε} φ1(x) > 0 on {x ∈ Rn : ε/2 < x < 2ε}, (11) (12) (13) and for some ε > 0 then holds for every f ∈ S ′(Rn). (cid:13)(cid:13) (Ψ∗ kf )awk ℓq(·)(Lp(·))(cid:13)(cid:13) ≤ c(cid:13)(cid:13) (Φ∗ kf )awk ℓq(·)(Lp(·))(cid:13)(cid:13) Remark 5. Observe that there are no restrictions on a > 0 and p, q ∈ P(Rn) in the theorem above. Proof. We have the fixed resolution of unity from Definition 1 and define the se- quence of functions (λj)j∈N0 by λj(x) = ϕj(cid:0) 2x ε (cid:1) φj(x) . 12 It follows from the Tauberian conditions (12) and (13) that they satisfy λj(x)φj (x) = 1, x ∈ Rn, ∞ Xj=0 λj(x) = λ1(2−j+1x), x ∈ Rn, j ∈ N, (14) (15) and supp λ0 ⊂ {x ∈ Rn : x ≤ ε} and supp λ1 ⊂ {x ∈ Rn : ε/2 ≤ x ≤ 2ε}. (16) Furthermore, we denote Λk = λk for k ∈ N0 and obtain together with (14) the following identities (convergence in S ′(Rn)) f = ∞ Xk=0 Λk ∗ Φk ∗ f, Ψν ∗ f = ∞ Xk=0 Ψν ∗ Λk ∗ Φk ∗ f. (17) We have (Ψν ∗ Λk ∗ Φk ∗ f )(y) ≤ZRn (Ψν ∗ Λk)(z)(Φk ∗ f )(y − z)dz ≤ (Φ∗ =: (Φ∗ kf )a(y)ZRn kf )a(y)Iν,k, (Ψν ∗ Λk)(z)(1 + 2kza)dz (18) where Iν,k :=ZRn (Ψν ∗ Λk)(z)(1 + 2kza)dz. According to Lemma 7 we get Iν,k ≤ c(2(k−ν)R, 2(ν−k)(a+1+α1), k ≤ ν, ν ≤ k. (19) Namely, we have for 1 ≤ k < ν with the change of variables 2kz 7→ z Iν,k = 2−nZRn ≤ c sup z∈Rn (Ψν−k ∗ Λ1(·/2))(z)(1 + za)dz (Ψν−k ∗ Λ1(·/2))(z)(1 + z)a+n+1 ≤ c2(k−ν)R. Similarly, we get for 1 ≤ ν < k with the substitution 2ν z 7→ z Iν,k = 2−nZRn ≤ c2(ν−k)(M −a). (Ψ1(·/2) ∗ Λk−ν)(z)(1 + 2k−ν za)dz M can be taken arbitrarily large because Λ1 has infinitely many vanishing moments. Taking M = 2a + α1 + 1 we derive (19) for the cases k, ν ≥ 1 with k 6= ν. The 13 missing cases can be treated separately in an analogous manner. The needed moment conditions are always satisfied by (11) and (16). The case k = ν = 0 is covered by the constant c in (19). Furthermore, we have (Φ∗ kf )a(y) ≤ (Φ∗ ≤ (Φ∗ kf )a(x)(1 + 2k(x − y)a) kf )a(x)(1 + 2ν (x − y)a) max(1, 2(k−ν)a). We put this into (18) and get (Ψν ∗ Λk ∗ Φk ∗ f )(y) 1 + 2ν(x − y)a sup y∈Rn ≤ c(Φ∗ kf )a(x)(2(k−ν)R, 2(ν−k)(1+α1), k ≤ ν, k ≥ ν. Multiplying both sides with wν (x) and using wν (x) ≤ wk(x)(2(k−ν)(−α2), 2(ν−k)α1 , k ≤ ν, k ≥ ν, leads us to (Ψν ∗ Λk ∗ Φk ∗ f )(y) 1 + 2ν (x − y)a sup y∈Rn wν(x) ≤ c(Φ∗ kf )a(x)wk(x)(2(k−ν)(R−α2), 2(ν−k), k ≤ ν, k ≥ ν. This inequality together with (17) gives for δ := min(1, R − α2) > 0 (Ψ∗ νf )a(x)wν (x) ≤ c ∞ Xk=0 2−k−νδ(Φ∗ kf )a(x)wk(x), x ∈ Rn. Taking the ℓq(·)(Lp(·)) norm and using Lemma 9 yields immediately the desired result. 3.1.3 Boundedness of the Peetre maximal operator We will present a theorem which describes the boundedness of the Peetre maximal operator. We use the same notation introduced at the beginning of the last subsec- tion. Especially, we have the functions ψk ∈ S(Rn) and Ψk = ψk ∈ S(Rn) for all k ∈ N0. Theorem 13. Let (wk)k∈N0 ∈ W α we assume ψ0, ψ1 ∈ S(Rn) with α1,α2, a > 0 and p, q ∈ P log(Rn). For some ε > 0 ψ0 > 0 on {x ∈ Rn : x < ε}, ψ1 > 0 on {x ∈ Rn : ε/2 < x < 2ε}. For a > n+clog(1/q) p− + α holds for all f ∈ S ′(Rn). (cid:13)(cid:13) (Ψ∗ kf )awk ℓq(·)(Lp(·))(cid:13)(cid:13) ≤ c(cid:13)(cid:13) (Ψk ∗ f )wk ℓq(·)(Lp(·))(cid:13)(cid:13) 14 Remark 6. Observe that in the theorem above no moment conditions on ψ1 are stated but this time there are restrictions on a and p(·), q(·). Proof. As in the last proof we find the functions (λj)j∈N0 with the properties (15), (16) and λk(2−νx)ψk(2−νx) = 1 for all ν ∈ N0. ∞ Xk=0 Instead of (17) we get the identity Ψν ∗ f = ∞ Xk=0 Λk,ν ∗ Ψk,ν ∗ Ψν ∗ f, (20) where Λk,ν(ξ) = [λk(2−ν ·)]∧(ξ) = 2νnΛk(2ν ξ) for all ν, k ∈ N0. The Ψk,ν are defined similarly. For k ≥ 1 and ν ∈ N0 we have Ψk,ν = Ψk+ν and with the notation σk,ν(x) =(ψ0(2−ν x) ψν (x) if k = 0, otherwise we get ψk(2−ν x)ψν(x) = σk,ν(x)ψk+ν(x). Hence, we can rewrite (20) as Ψν ∗ f = ∞ Xk=0 Λk,ν ∗ σk,ν ∗ Ψk+ν ∗ f. For k ≥ 1 we get from Lemma 7 (Λk,ν ∗ σk,ν)(z) = 2(ν−1)n(Λk ∗ Ψ1(·/2))(2ν z) ≤ CM 2νn 2−kM (1 + 2ν za) (21) (22) for all k, ν ∈ N0 and arbitrary large M ∈ N. For k = 0 we get the estimate (22) by using Lemma 7 with M = 0. This together with (21) gives us (Ψν ∗ f )(y) ≤ CM 2νn ∞ Xk=0ZRn 2−kM (1 + 2ν (y − z)a) (Ψk+ν ∗ f )(z)dz. (23) For fixed r ∈ (0, 1] we divide both sides of (23) by (1 + 2ν (x − y)a) and we take the supremum with respect to y ∈ Rn. Using the inequalities (Ψk+ν ∗ f )(z) ≤ (Ψk+ν ∗ f )(z)r(Ψ∗ (1 + 2ν (y − z)a)(1 + 2ν (x − y)a) ≥ c(1 + 2ν (x − z)a), k+νf )a(x)1−r(1 + 2k+ν(x − z)a)1−r and (1 + 2k+ν(x − z)a)1−r (1 + 2ν (x − z)a) ≤ 2ka (1 + 2k+ν(x − z)a)r , 15 we get (Ψ∗ νf )a(x) ≤ CM ∞ Xk=0 2−k(M +n−a)(Ψ∗ k+νf )a(x)1−rZRn 2(k+ν)n(Ψk+ν ∗ f )(z)r (1 + 2k+ν(x − z)a)r dz. (24) Now, we apply Lemma 11 with γν = (Ψ∗ νf )a(x), βν =ZRn 2νn(Ψν ∗ f )(z)r (1 + 2ν (x − z)a)r dz, ν ∈ N0 N = M + n − a, CN = CM + n − a and N 0 in (10) equals the order of the distribution f ∈ S ′(Rn). By Lemma 11 we obtain for every N ∈ N, x ∈ Rn and ν ∈ N0 (Ψ∗ νf )a(x)r ≤ CN ∞ Xk=0 2−kN rZRn 2(k+ν)n(Ψk+ν ∗ f )(z)r (1 + 2k+ν (x − z)a)r dz (25) provided that (Ψ∗ νf )a(x) < ∞. Since f ∈ S ′(Rn), we see that (Ψ∗ νf )a(x) < ∞ for all x ∈ Rn and all ν ∈ N0 at least if a > N 0, where N 0 is the order of the distribution. Thus we have (25) with CN independent of f ∈ S ′(Rn) for a ≥ N 0 and therefore with CN = CN,f for all a > 0 (the right side of (25) decreases as a increases). One can easily check that (25) with CN = CN,f implies that if for some a > 0 the right side of (25) is finite, then (Ψ∗ νf )a(x) < ∞. Now, repeating the above argument resurrects the independence of CN . If the right side of (25) is infinite, there is nothing to prove. More exhaustive arguments of this type have been used in [55] and [48]. We point out that (25) holds also for r > 1, where the proof is much simpler. We only have to take (23) with a + n instead of a, divide both sides by (1 + 2ν (x − y)a) and apply Holder's inequality with respect to k and then z. Multiplying (25) by wν(x)r we derive with the properties of our weight sequence (Ψ∗ νf )a(x)rwν(x)r ≤ C ′ N ∞ Xk=0 2−k(N +α1)rZRn 2(k+ν)n(Ψk+ν ∗ f )(z)rwk+ν(z)r (1 + 2k+ν(x − z)a−α)r dz, (26) for all x ∈ Rn, ν ∈ N0 and all N ∈ N. Now, we choose r = p− and we have r(a − α) > n + clog(1/q). We denote k+ν(z) = (Ψk+ν ∗ f )(z)rwk+ν(z)r then we can rewrite (26) by gr (Ψ∗ ν f )a(x)rwν(x)r ≤ C ′ N ∞ Xl=ν 2−(l−ν)(N +α1)r(cid:0)gr l ∗ ηl,r(a−α)(cid:1) (x). (27) For fixed N > 0 with δ = N + α1 > 0 we apply the ℓ q(·) r (27) (L p(·) r ) norm and derive from kf )r awr (cid:13)(cid:13) (Ψ∗ k ℓq(·)/r(Lp(·)/r)(cid:13)(cid:13) ≤ CN(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞ Xl=ν 2−(l−ν)δ(cid:0)gr l ∗ ηl,r(a−α)(cid:1)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . ℓq(·)/r(Lp(·)/r)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 16 Now application of Lemma 9 and Lemma 10 (r(a−α) > n+clog(1/q)) on the formula above give us kf )r a(·)wr (cid:13)(cid:13) (Ψ∗ k(·) ℓq(·)/r(Lp(·)/r)(cid:13)(cid:13) ≤ C ′ N(cid:13)(cid:13) (Ψν ∗ f )(·)rwν (·)r ℓq(·)/r(Lp(·)/r)(cid:13)(cid:13) which proves the theorem. 3.2 Local means characterization of Bs(·) p(·),q(·)(Rn) and F s(·) In this section we reformulate the local means characterization for Bw above and for F w If we have a variable smoothness function s ∈ C log defines an admissible weight sequence w ∈ W α α = clog(s), cf. Remark 2. Here, we denote by clog(s) the constant in (2) for s(·). p(·),q(·)(Rn) p(·),q(·)(Rn) from p(·),q(·)(Rn) from Corollary 4.7 in [30] in terms of variable smoothness. loc (Rn) given, then wj(x) = 2js(x) α1,α2 with α1 = s−, α2 = s+ and Theorem 14. Let p, q ∈ P log(Rn) (p+, q+ < ∞ in the F-case) and s ∈ C log Further let a > 0, R ∈ N0 with R > s+ and let ψ0, ψ1 belong to S(Rn) with loc (Rn). Dβψ1(0) = 0, for 0 ≤ β < R, and for some ε > 0. ψ0(x) > 0 on {x ∈ Rn : x < ε} ψ1(x) > 0 on {x ∈ Rn : ε/2 < x < 2ε} 1. For a > n+clog(1/q) p− + clog(s) and all f ∈ S ′(Rn) we have 2. For a > min(p−,q−) + clog(s) and all f ∈ S ′(Rn) we have f Bs(·) n p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13) 2ks(·)(Ψk ∗ f )(cid:12)(cid:12)(cid:12) 2ks(·)(Ψk ∗ f )(cid:12)(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13) f F s(·) (cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) 2ks(·)(Ψ∗ 2ks(·)(Ψ∗ kf )a(cid:12)(cid:12)(cid:12) kf )a(cid:12)(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) . . Remark 7. During the referee process of this work, there appeared in [18] a char- acterization by local means and a characterization by atoms for Bs(·) p(·),q(·)(Rn). The author moved the smoothness sequence 2ks(·) into the Peetre maximal operator (6) and modified it to (Ψ∗ k2ks(·)f )a(x) = sup y∈Rn 2ks(y)(Ψk ∗ f )(y) 1 + 2k(y − x)a . For this modified Peetre maximal operator he obtained in [18, Theorem 2] an equiv- alence of the norms similar to our Theorem 14 for Bs(·) p(·),q(·)(Rn). The advantage of his method is that the condition on a > 0 weakens to a > n p− . 17 s > σp = n(cid:18) s > σp,q = n(cid:18) 1 min(p, 1) 1 − 1(cid:19) − 1(cid:19) min(p, q, 1) (28) (29) 4 Ball means of differences This section is devoted to the characterization of Besov and Triebel-Lizorkin spaces p(·),q(·)(Rn) and F s(·) Bs(·) p(·),q(·)(Rn) by ball means of differences. In the case of constant indices p, q and s, this is a classical part of the theory of function spaces. We refer especially to [52, Section 2.5] and references given there. It turns out that, under the restriction in the B-case and in the F -case, Besov and Triebel-Lizorkin spaces with constant indices may be char- acterized by expressions involving only the differences of the function values without any use of Fourier analysis. This was complemented in [49] and [50] by showing that these conditions are also indispensable. Of course, we are limited by (28) and (29) also in the case of variable exponents. The characterization by (local means of) differences for 2-microlocal spaces with constant p, q > 1 was given by Besov [8] and a similar characterization for Besov spaces with p = q = ∞ and the special weight sequence from (3) was given by Seuret and Levy V´eh´el in [35]. We refer to [19] and [29] for the treatment of spaces of generalized smoothness. Our approach follows essentially [52] with some modifications described in [54]. The main obstacle on this way is the unboundedness of the maximal operator in the frame of Lp(·)(ℓq(·)) and ℓq(·)(Lp(·)) spaces, cf. [17, Section 5] and [1, Example 4.1]. This is circumvented by the use of convolution with radial functions in the sense of [17] and [1] together with a certain bootstrapping argument, which shall be described in detail below. The plan of this part of the work is as follows. First we give in Section 4.1 the necessary notation. We state the main assertions of this part in Section 4.2. Then we prove in Section 4.3 a certain preliminary version of these assertions. In Section 4.4 we prove a characterization by ball means of differences for spaces with q ∈ (0, ∞] constant (where the maximal operator is bounded) and use this together with our preliminary characterization from Section 4.3 to conclude the proof. Finally, in Section 4.5 we will present the ball means of differences characterization also for the p(·),q(·)(Rn) and in Section 4.6 we 2-microlocal function spaces Bw present separately some useful Lemmas, not to disturb the main proofs of this part. p(·),q(·)(Rn) and F w 4.1 Notation Let f be a function on Rn and let h ∈ Rn. Then we define ∆1 hf (x) = f (x + h) − f (x), x ∈ Rn. The higher order differences are defined inductively by ∆M h f (x) = ∆1 h(∆M −1 h f )(x), M = 2, 3, . . . 18 This definition also allows a direct formula ∆M h f (x) := M (−1)j(cid:18)M Xj=0 j (cid:19)f (x + (M − j)h). (30) By ball means of differences we mean the quantity dM t f (x) = t−nZh≤t ∆M h f (x)dh =ZB ∆M th f (x)dh, where B = {y ∈ Rn : y < 1} is the unit ball of Rn, t > 0 is a real number and M is a natural number. Let us now introduce the (quasi-)norms, which shall be the main subject of our study. We define kf F s(·) p(·),q(·)(Rn)k∗ := kf Lp(·)(Rn)k 0 kf F s(·) (cid:18)Z ∞ t−s(x)q(x)(cid:0)dM t f (x)(cid:1)q(x) dt and its partially discretized counterpart p(·),q(·)(Rn)k∗∗ := kf Lp(·)(Rn)k ∞ Xk=−∞ +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2ks(x)q(x)(cid:0)dM = kf Lp(·)(Rn)k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)2ks(x)dM p(·),q(·)(Rn)k∗∗ := kf Lp(·)(Rn)k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)2ks(x)dM Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) t (cid:19)1/q(x)(cid:12)(cid:12)(cid:12)(cid:12) Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2−k f (x)(cid:1)q(x)!1/q(x) (cid:12)(cid:12)(cid:12)(cid:12) k=−∞(cid:12)(cid:12)(cid:12)(cid:12) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13)(cid:13) 2−k f (x)(cid:17)∞ ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13)(cid:13) The norm kf F s(·) namely 2−k f (x)(cid:17)∞ kf Bs(·) k=−∞ . p(·),q(·)(Rn)k∗∗ admits a direct counterpart also for Besov spaces, Finally, we shall use as a technical tool also the analogues of (31) -- (32) with the integration over t restricted to 0 < t < 1. This leads to the following expressions (31) . (32) kf F s(·) p(·),q(·)(Rn)k∗ 1 := kf Lp(·)(Rn)k kf F s(·) p(·),q(·)(Rn)k∗∗ kf Bs(·) p(·),q(·)(Rn)k∗∗ , 0 (cid:18)Z 1 1 := kf Lp(·)(Rn)k t−s(x)q(x)(cid:0)dM +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2ks(x)q(x)(cid:0)dM = kf Lp(·)(Rn)k +(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)2ks(x)dM 1 := kf Lp(·)(Rn)k +(cid:13)(cid:13)(cid:13)(cid:16)2ks(x)dM t (cid:19)1/q(x)(cid:12)(cid:12)(cid:12)(cid:12) Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) t f (x)(cid:1)q(x) dt Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2−k f (x)(cid:1)q(x)!1/q(x) (cid:12)(cid:12)(cid:12)(cid:12) k=0(cid:12)(cid:12)(cid:12)(cid:12) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13)(cid:13) 2−k f (x)(cid:17)∞ 2−k f (x)(cid:17)∞ ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) ∞ Xk=0 k=0 , . 19 4.2 Main Theorem Using the notation introduced above, we may now state the main result of this section. Theorem 15. (i) Let p, q ∈ P log(Rn) with p+, q+ < ∞ and s ∈ C log M ∈ N with M > s+ and let loc (Rn). Let s− > σp−,q− ·(cid:20)1 + clog(s) n · min(p−, q−)(cid:21) . (33) Then p(·),q(·)(Rn) = {f ∈ Lp(·)(Rn) ∩ S ′(Rn) : kf F s(·) F s(·) p(·),q(·)(Rn)k and k · F s(·) p(·),q(·)(Rn)k∗∗. and k · F s(·) same holds for kf F s(·) (ii) Let p, q ∈ P log(Rn) and s ∈ C log p(·),q(·)(Rn)k∗ < ∞} p(·),q(·)(Rn)k∗ are equivalent on F s(·) p(·),q(·)(Rn). The s− > σp− ·(cid:20)1 + Then loc (Rn). Let M ∈ N with M > s+ and let clog(1/q) clog(s) n + n · p−(cid:21) . (34) p(·),q(·)(Rn) = {f ∈ Lp(·)(Rn) ∩ S ′(Rn) : kf Bs(·) Bs(·) p(·),q(·)(Rn)k and k · Bs(·) and k · Bs(·) p(·),q(·)(Rn). Remark 8. Let us comment on the rather technical conditions (33) and (34). p(·),q(·)(Rn)k∗∗ are equivalent on Bs(·) p(·),q(·)(Rn)k∗∗ < ∞} • If min(p−, q−) ≥ 1, then (33) becomes just s− > 0. Furthermore, if p, q and s are constant functions, then (33) coincides with (29). • If p− ≥ 1, then (34) reduces also to s− > 0 and in the case of constant exponents we again recover (28). As indicated already above, the proof is divided into several parts. 4.3 Preliminary version of Theorem 15 This subsection contains a preliminary version of Theorem 15 (Lemma 16). Its proof represents the heart of the proof of Theorem 15. For better lucidity, it is again divided into more parts. Lemma 16. Under the conditions of Theorem 15, the following estimates hold for all f ∈ Lp(·)(Rn) ∩ S ′(Rn): kf F s(·) kf F s(·) kf F s(·) p(·),q(·)(Rn)k∗ ≈ kf F s(·) 1 ≈ kf F s(·) p(·),q(·)(Rn)k∗ 1 . kf F s(·) p(·),q(·)(Rn)k∗∗ 1 . kf Bs(·) p(·),q(·)(Rn)k∗∗ p(·),q(·)(Rn)k∗∗, p(·),q(·)(Rn)k∗∗ 1 , p(·),q(·)(Rn)k . kf F s(·) p(·),q(·)(Rn)k . kf Bs(·) kf Bs(·) p(·),q(·)(Rn)k∗∗, p(·),q(·)(Rn)k∗∗. (35) (36) (37) (38) 20 Proof. Part I. First we prove (35) and (36). We discretize the inner part of k · k∗ and obtain (cid:20)Z ∞ 0 t−s(x)q(x)(cid:18)ZB =(cid:20)Z ∞ =(cid:20) ∞ Xk=−∞Z 2−k 0 2−k−1 ∆M th f (x)dh(cid:19)q(x) dt t−s(x)q(x)(cid:18)t−nZtB ∆M t (cid:21)1/q(x) κ f (x)dκ(cid:19)q(x) dt t−s(x)q(x)(cid:18)t−nZtB ∆M t (cid:21)1/q(x) κ f (x)dκ(cid:19)q(x) dt t (cid:21)1/q(x) (39) . If 2−k−1 ≤ t ≤ 2−k, then 2ks(x)q(x) ≤ t−s(x)q(x) ≤ 2(k+1)s(x)q(x) and 2knZ2−(k+1)B ∆M κ f (x)dκ . t−nZtB ∆M κ f (x)dκ . 2(k+1)nZ2−kB ∆M κ f (x)dκ. Plugging these estimates into (39), we may further estimate t (cid:21)1/q(x) (cid:20)Z ∞ 0 ∆M t−s(x)q(x)(cid:18)ZB .(cid:20) ∞ Xk=−∞ .(cid:20) ∞ Xk=−∞ th f (x)dh(cid:19)q(x) dt 2(k+1)s(x)q(x)(cid:18)2knZ2−kB 2ks(x)q(x)(cid:18)ZB ∆M 2−k κf (x)dκ(cid:19)q(x)(cid:21)1/q(x) . ∆M κ f (x)dκ(cid:19)q(x)(cid:21)1/q(x) The estimate from below follows in the same manner. Finally, the proof of (36) is almost the same. Part II. This part is devoted to the proof of the left hand side of (37). It is divided into several steps to make the presentation clearer. Step 1. First, we point out that the estimate (cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) .(cid:13)(cid:13)(cid:13) f Bs(·) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) follows from the characterization of Bs(·) p(·),q(·)(Rn) in terms of Nikol'skij representa- tions (cf. Theorem 8.1 of [1]). We refer also to Remark 2.5.3/1 in [52]. The extension to F -spaces is then given by the simple embedding (cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) .(cid:13)(cid:13)(cid:13) f Bs(·)−ε p(·),p(·)(Rn)(cid:13)(cid:13)(cid:13) .(cid:13)(cid:13)(cid:13) f F s(·) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) with ε > 0 chosen small enough. Step 2. Let (ϕj )j∈N0 be the functions used in Definition 3. We use the decom- position f = ∞ Xl=−∞ f(k+l), k ∈ Z, 21 where f(k+l) = (ϕk+l f )∨, or = 0 if k + l < 0 and get (♣) := ∞ 2ks(x)q(x)(cid:18)ZB Xk=0 2ks(x)q(x) ZB Xk=0 ∞ ∆M 2−khf (x)dh(cid:19)q(x) 2−kh ∞ Xl=−∞ ∆M f(k+l)! (x)dh!q(x) . = If q(x) ≤ 1 then we proceed further (♣) ≤ ≤ ∞ ∞ Xk=0 Xk=0 ∞ 2ks(x)q(x) ZB Xl=−∞ Xl=−∞ 2ks(x)q(x)(cid:18)ZB ∞ If q(x) > 1, we use Minkowski's inequality ∞ ∞ (♣)1/q(x) ≤ 2ks(x)q(x) ZB Xk=0  Xl=−∞ ∞ Xk=0 Xl=−∞ 2ks(x)q(x)(cid:18)ZB ≤ ∞ ∆M ∆M 2−khf(k+l)(x)dh!q(x) 2−khf(k+l)(x)dh(cid:19)q(x) . ∆M 1/q(x) 2−khf(k+l)(x)dh!q(x)  2−khf(k+l)(x)dh(cid:19)q(x)!1/q(x) ∆M . (40) We split in both cases ∞ Xl=−∞ · · · = I + II = 0 Xl=−∞ · · · + . . . ∞ Xl=1 Step 3. We estimate the first summand with l ≤ 0. We use Lemma 22 in the form ∆M h f(k+l)(x) ≤ C max(1, bha) · min(1, bhM )Pb,af(k+l)(x), where a > 0 is arbitrary, b = 2k+l and Pb,af (x) = sup z∈Rn f (x − z) 1 + bza . Furthermore, we use this estimate with 2−kh instead of h. We obtain ZB ∆M 2−khf(k+l)(x)dh .ZB max(1, b2−kha) · min(1, b2−khM )Pb,af(k+l)(x)dh . 2lM P2k+l,af(k+l)(x). (41) The last inequality follows from max(1, b2−kha) ≤ 1 (recall that l ≤ 0 and h ≤ 1) and min(1, b2−khM ) ≤ 2lM . 22 If q(x) ≤ 1, we estimate the first sum in (40) ∞ ∞ Xk=0 Xk=0 ∆M 2−khf(k+l)(x)dh(cid:19)q(x) 2ks(x)q(x)(cid:18)ZB 2ks(x)q(x)(cid:16)2lM P2k+l,af(k+l)(x)(cid:17)q(x) 2(k+l)s(x)q(x)P q(x) 2k+l,af(k+l)(x) I ≤ . = ≈ 0 0 Xl=−∞ Xl=−∞ Xl=−∞ Xk=0 ∞ 0 ∞ 2l(M −s(x))q(x) Xk=0 2ks(x)q(x)P q(x) 2k,a f(k)(x), where the last estimate makes use of M > s+, q− > 0 and the fact that f(k+l) = 0 for k + l < 0. If q(x) > 1, we proceed in a similar way to obtain I 1/q(x) ≤ 0 0 . ∆M 2−khf(k+l)(x)dh(cid:19)q(x)!1/q(x) 2ks(x)q(x)(cid:18)ZB 2ks(x)q(x)(cid:16)2lM P2k+l,af(k+l)(x)(cid:17)q(x)!1/q(x) 2k+l,af(k+l)(x)!1/q(x) Xl=−∞ ∞ Xk=0 Xl=−∞ ∞ Xk=0 2l(M −s(x)) ∞ Xl=−∞ Xk=0 2k,a f(k)(x)!1/q(x) . ∞ Xk=0 2(k+l)s(x)q(x)P q(x) 2ks(x)q(x)P q(x) = 0 . We have used in the last estimate again M > s+ and the definition of f(k+l). Hence, I 1/q(x) . ∞ Xk=0 holds for all x ∈ Rn. Finally, we obtain 2ks(x)q(x)P q(x) 2k,a f(k)(x)!1/q(x) 2ks(x)q(x)P q(x) kI 1/q(·)Lp(·)(Rn)k .(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞ Xk=0 =(cid:13)(cid:13)(cid:13)(cid:16)2ks(x)P2k,af(k)(x)(cid:17)∞ .(cid:13)(cid:13)(cid:13)(cid:16)2ks(·)f(k)(cid:17)∞ k=0 Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2k,a f(k)(x)!1/q(x) (cid:12)(cid:12)(cid:12)(cid:12) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) = kf F s(·) k=0 Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)k, where we used the boundedness of Peetre maximal operator as described in Theorem 14 for a > 0 large enough. (42) 23 Step 4. We estimate the second summand in (40) with l > 0. If min(p−, q−) > 1, then we put λ = 1. Otherwise we choose real parameters 0 < λ < min(p−, q−) and a > 0 such that n a > min(p−, q−) + clog(s) and a(1 − λ) < s−. Due to (33), this is always possible. We start again with estimates of the ball means of differences. We use Lemma 22 and (30) to obtain ∆M ZB · ∆M ∆M 2−khf(k+l)(x)λ · ∆M 2−khf(k+l)(x)1−λdh 2−khf(k+l)(x)dh =ZB .ZB(cid:16)max(1, 2k+l2−kha) min(1, 2k+l2−khM )P2k+l,af(k+l)(x)(cid:17)1−λ ≤(cid:16)2laP2k+l,af(k+l)(x)(cid:17)1−λZB ≤(cid:16)2laP2k+l,af(k+l)(x)(cid:17)1−λ Xj=0 f(k+l)(x + j2−kh)λdh, 2−khf(k+l)(x)λdh 2−khf(k+l)(x)λdh cj,MZB ∆M M · (43) where the constants cj,M are given by (30). We shall deal in detail only with the term with j = 1. The term with j = 0 is much simpler to handle (as there the integration over h ∈ B immediately disappears) and this case reduces essentially to Holder's inequality and boundedness of the Peetre maximal operator. The terms with 2 ≤ j ≤ M may be handled in the same way as the one with j = 1. We use Lemma 20 with r = λ in the form f(k+l)(y)λ . (ηk+l,2m ∗ f(k+l)λ)(y), with m > max(n, clog(s)), Lemma 23 and Lemma 19 to get 2ks(x)λZB f(k+l)(x + 2−kh)λdh . 2ks(x)λZB (ηk+l,2m ∗ f(k+l)λ)(x + 2−kh)dh = 2ks(x)λ([2knχ2−kB] ∗ ηk+l,2m ∗ f(k+l)λ)(x) . 2ks(x)λ(ηk,2m ∗ f(k+l)λ)(x) . (ηk,m ∗ 2ks(·)f(k+l)λ)(x) ≤ 2−ls−λ(ηk,m ∗ 2(k+l)s(·)f(k+l)λ)(x). (44) (45) We insert (44) into (43) and arrive at 2ks(x)ZB ∆M 2−khf(k+l)(x)dh . 2la(1−λ)−ls−(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)1−λ (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)(x). 24 If q(x) > 1, we proceed further with the use of Holder's inequality ∞ II 1/q(x) . 2la(1−λ)−ls− Xl=1 ∞ Xk=0(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)(1−λ)q(x) Xl=1 ≤ ∞ 2la(1−λ)−ls− ∞ · ∞ Xk=0 Xk=0(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)q(x)!(1−λ)/q(x) (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ/q(x) · (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)(x)!1/q(x) Xk=0(cid:16)2ks(x)P2k,af(k)(x)(cid:17)q(x)!(1−λ)/q(x) = ∞ · ∞ · Xl=1 (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ/q(x) 2la(1−λ)−ls− ∞ Xk=0 . (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)(x) If q(x) ≤ 1, we obtain in a similar way ∞ II . ≤ ∞ ∞ 2(la(1−λ)−ls−)q(x) Xl=1 Xk=0(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)(1−λ)q(x) 2(la(1−λ)−ls−)q(x) ∞ Xl=1 (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ · ∞ Xk=0 Xk=0(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)q(x)!1−λ · Xk=0(cid:16)2ks(x)P2k,af(k)(x)(cid:17)q(x)!1−λ = ∞ (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ 2(la(1−λ)−ls−)q(x) ∞ Xk=0 Xl=1 ∞ · · 25 and further (with use of Lemma 24) II 1/q(x) . ∞ ∞ Xk=0(cid:16)2ks(x)P2k,af(k)(x)(cid:17)q(x)!(1−λ)/q(x) 2(la(1−λ)−ls−)q(x) ∞ Xk=0 Xk=0(cid:16)2ks(x)P2k,af(k)(x)(cid:17)q(x)!(1−λ)/q(x) . ∞ ·  Xl=1 · · (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ  1/q(x) · Xl=1 If we denote ∞ (ηk,m ∗ 2(k+l)s(·)f(k+l)λ)q(x)/λ(x)!λ/q(x) 21/2·(la(1−λ)−ls−) ∞ Xk=0 . F (x) := ∞ Xk=0(cid:16)2ks(x)P2k,af(k)(x)(cid:17)q(x)!1/q(x) , x ∈ Rn and Bk+l(x) := 2(k+l)s(x)f(k+l)(x), x ∈ Rn we get for δ := −1/2 · (a(1 − λ) − s−) > 0 II 1/q(x) . F (x)1−λ · ∞ Xl=1 2−lδ ∞ Xk=0 (ηk,m ∗ Bλ k+l)q(x)/λ(x)!λ/q(x) . (46) 1 F λ 2 kp(·) ≤ 2kF1k1−λ We use kF 1−λ p(·), cf. [16, Lemma 3.2.20], and suppose that the Lp(·)− (quasi-)norm is equivalent to an r-norm with 0 < r ≤ 1. Together with Lemma 21 we arrive at p(·) kF2kλ kII 1/q(x)kr p(·) . kF (x)k(1−λ)r p(·) . kF (x)k(1−λ)r p(·) (ηk,m ∗ Bλ (ηk,m ∗ Bλ · ∞ ∞ ∞ ·(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2−lδ ∞ Xl=1 Xk=0 2−lδr(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞ Xl=1 Xk=0 2−lδr(cid:13)(cid:13)(cid:13) Xl=1 2−lδr(cid:13)(cid:13)(cid:13) Xl=1 2−lδr(cid:13)(cid:13)(cid:13) Xl=1 ∞ ∞ 26 p(·) λr p(·) λr k+l)q(x)/λ(x)!1/q(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k+l)q(x)/λ(x)!1/q(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k=0(cid:13)(cid:13)(cid:13) k=0(cid:13)(cid:13)(cid:13) k=0(cid:13)(cid:13)(cid:13) k+l(x))∞ Lp(·)/λ(ℓq(·)/λ) r r r Lp(·)/λ(ℓq(·)/λ) Lp(·)/λ(ℓq(·)/λ) (47) . kf F s(·) p(·),q(·)(Rn)k(1−λ)r · (ηk,m ∗ Bλ . kf F s(·) p(·),q(·)(Rn)k(1−λ)r · (Bλ k+l(x))∞ . kf F s(·) p(·),q(·)(Rn)k(1−λ)r · (Bλ k (x))∞ . kf F s(·) . kf F s(·) p(·),q(·)(Rn)k(1−λ)r · kBk(x)kλr p(·),q(·)(Rn)kr, Lp(·)(ℓq(·)) which finishes the proof. Part III.: We prove the right hand side of (37). We follow again essentially [52, Section 2.5.9] with some modifications as presented in [54]. Roughly speaking, compared to the case of constant exponents, only minor modifications are necessary. Let ψ ∈ C ∞ 0 (Rn) with ψ(x) = 1, x ≤ 1 and ψ(x) = 0, x > 3/2. We define M −1 ϕ0(x) = (−1)M +1 Xµ=0 (−1)µ(cid:18)M µ(cid:19)ψ((M − µ)x). It follows that ϕ0 ∈ C ∞ 0 (Rn) with ϕ(x) = 0, x > 3/2 and ϕ(x) = 1, x < 1/M . We also put ϕj(x) = ϕ0(2−jx) − ϕ0(2−j+1x) for j ≥ 1. This is the decomposition of unity we used in the definition of kf F s(·) p(·),q(·)(Rn)k, cf. Definition 3. Recall that due to [17] and [30], this (quasi-)norm of kf F s(·) p(·),q(·)(Rn)k does not depend on the choice of the decomposition of unity. We observe that and ϕ0(x) = (−1)M +1(∆M x ψ(0) − (−1)M ), (F −1ϕjFf )(x) =((F −1∆M (F −1(∆M ξ ψ(0)Ff )(x) + (−1)M +1f (x), 2−j ξψ(0) − ∆M 2−j+1ξψ(0))Ff )(x), Furthermore, a straightforward calculation shows that j = 0, j ≥ 1. (48) (49) (F −1(∆M M M (−1)uF −1[ψ((M − u)2−j ·)Ff ](x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2−j ξψ(0))Ff )(x) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xu=0 (−1)uF −1[ψ((M − u)2−j ·)] ∗ f (x)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≈(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xu=0 ≈(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (−1)uZRn Xu=0 =(cid:12)(cid:12)(cid:12)(cid:12) ZRn F −1ψ(h)f (x − (M − u)2−jh)dh(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 2−j hf (x)dh(cid:12)(cid:12)(cid:12)(cid:12) ≤ZRn ψ(h) · ∆M ψ(h)∆M M 2−j hf (x)dh holds for every j ∈ N0. We denote g = ψ ∈ S(Rn) and obtain kf F s(·) p(·),q(·)(Rn)k ≈ k2js(x)(F −1ϕjFf )(x)Lp(·)(ℓq(·)k (50) .(cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:13) 2js(x)ZRn g(h) · ∆M 2−j hf (x)dhLp(·)(ℓq(·))(cid:13)(cid:13)(cid:13)(cid:13) . The rest of this part consists essentially of using the property of g ∈ S(Rn) to come from (50) to k · k∗∗. 27 We denote I0 := B, Iu := 2uB \ 2u−1B, u ∈ N and use g(h) ≤ c2−ur, h ∈ Iu with r taken large enough (recall that g ∈ S(Rn)) and estimate g(h) · ∆M 2−j hf (x)dh = ZRn . = 2−j hf (x)dh ∞ g(h) · ∆M Xu=0ZIu 2−ur2jnZ2u−j B Xu=0 2u(n−r)2−(u−j)nZ2u−j B Xu=0 ∆M ∞ ∞ h f (x)dh (51) ∆M h f (x)dh. We put and Gj(x) := 2js(x)(F −1(∆M 2−jξψ(0))Ff )(x), j ∈ N0 gk(x) := 2ks(x)2knZ2−kB ∆M h f (x)dh, k ∈ Z. Using (48), (49) and (51), we obtain the estimate Gj(x) . 2js(x) = = j Xk=−∞ Xk=−∞ j ∞ ∆M 2u(n−r)2−(u−j)nZ2u−j B Xu=0 2(j−k)s(x)2(j−k)(n−r)2ks(x)2knZ2−kB h f (x)dh ∆M h f (x)dh (52) 2(j−k)(s(x)+n−r)gk(x) ≤ ∞ Xk=−∞ 2j−k·(s(x)+n−r)gk(x). Choosing r > s+ + n and applying Lemma 9 then finishes the proof. Part IV. The proof of the left hand side of (38) follows in the same manner as in Part II. We shall describe the necessary modifications. First, let us mention, that the condition q+ < ∞ was used only in the application of Lemma 21. In the rest of the arguments also the case q(x) = ∞ may be incorporated with only slight change of notation. Let us put f (k)(x) := 2ks(x)ZB ∆M 2−khf (x)dh, x ∈ Rn. We obtain (in analogue to (40)) f (k) ≤ f (k),I + f (k),II := 0 Xl=−∞ 2ks(x)ZB ∆M 2−khf(k+l)(x)dh + ∞ Xl=1 2ks(x)ZB ∆M 2−khf(k+l)(x)dh. 28 We estimate the first sum using (41) and get f (k),I . 0 Xl=−∞ 2l(M −s+)g1 k+l = k Xu=0 2(u−k)(M −s+)g1 u ≤ 2−u−k(M −s+)g1 u, ∞ Xu=0 where g1 a > 0 large enough gives u := 2us(x)P2u,af(u)(x). The application of Lemma 9 and Theorem 14 with k(cid:16)f (k),I(cid:17)∞ k=0 ℓq(·)(Lp(·))k . k (gu)∞ u=0 ℓq(·)(Lp(·))k . kf Bs(·) p(·),q(·)(Rn)k. To estimate f (k),II, we proceed as in the Step 4 of Part II. If p− > 1, we choose again λ = 1, otherwise we take 0 < λ < p− and a > n + clog(1/q) p− + clog(s) such that a(1 − λ) < s−. This is possible due to (34). We use (44) with m > max(n + clog(1/q), clog(s)) to get f (k),II . ∞ Xl=1 2la(1−λ)−ls−(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)1−λ · (53) · (ηk,m ∗ 2(k+l)s(·)f(k+l)(·)λ)(x) 2la(1−λ)−ls− = ∞ Xl=1 (g1 k+l(x))1−λ · (ηk,m ∗ (g2 k+l)λ)(x), k+l(x) := 2(k+l)s(x)f(k+l)(x). We take the ℓq(·)(Lp(·)) (quasi-)norm of the where g2 last expression - and assume that it is equivalent to some r-norm. This gives for δ := s− − a(1 − λ) > 0 the following estimate kf (k),II ℓq(·)(Lp(·))kr . ∞ ∞ Xl=1 2−lδrk(g1 k+l(x))1−λ · (ηk,m ∗ (g2 k+l)λ)(x)ℓq(·)(Lp(·))kr 2−lδrkg1 k+lℓq(·)(Lp(·))k(1−λ)r · k[ηk,m ∗ (g2 k+l)λ]1/λℓq(·)(Lp(·))kλr 2−lδrkg1 kℓq(·)(Lp(·))k(1−λ)r · kηk,m ∗ (g2 k+l)λℓq(·)/λ(Lp(·)/λ)kr 2−lδrkg1 kℓq(·)(Lp(·))k(1−λ)r · k(g2 k+l)λℓq(·)/λ(Lp(·)/λ)kr (54) 2−lδrkg1 kℓq(·)(Lp(·))k(1−λ)r · kg2 k+lℓq(·)(Lp(·))kλr . . . . ∞ ∞ Xl=1 Xl=1 Xl=1 Xl=1 ∞ . kf Bs(·) p(·),q(·)(Rn)kr. We have used Lemma 10 and Lemma 25. Part V. The right hand side inequality of (38) follows also along the same line as in Part III. We just combine (52) with the choice r > s+ + n and apply Lemma 9. 29 4.4 Proof of Theorem 15 This section is devoted to the proof of Theorem 15. We start with the case of constant q. In that case, the usual Hardy-Littlewood maximal operator M f (x) = sup r>0 1 B(x, r)ZB(x,r) f (y)dy is bounded on ℓq(Lp(·)) and Lp(·)(ℓq). Indeed, the following lemma is a consequence of [12] and [16, Theorem 4.3.8]. Lemma 17. (i) Let p ∈ P log(Rn) with 1 < p− ≤ p+ < ∞ and 1 < q < ∞. Then for all (fj)∞ j=−∞ ∈ Lp(·)(ℓq). (ii) Let p ∈ P log(Rn) with p− > 1 and 0 < q ≤ ∞. Then (cid:13)(cid:13) (M fj)∞ (cid:13)(cid:13) (M fj)∞ j=−∞(cid:12)(cid:12) Lp(·)(ℓq)(cid:13)(cid:13) .(cid:13)(cid:13) (fj)∞ j=−∞(cid:12)(cid:12) ℓq(Lp(·))(cid:13)(cid:13) .(cid:13)(cid:13) (fj)∞ j=−∞(cid:12)(cid:12) Lp(·)(ℓq)(cid:13)(cid:13) j=−∞(cid:12)(cid:12) ℓq(Lp(·))(cid:13)(cid:13) for all (fj)∞ j=−∞ ∈ ℓq(Lp(·)). Proof of Theorem 15. With the help of Lemma 17, we prove Theorem 15 for q constant. In view of Lemma 16, it is enough to prove p(·),q(·)(Rn)k∗∗ . kf F s(·) kf F s(·) p(·),q(·)(Rn)k and a corresponding analogue for the B-spaces. (55) Part I. In this part we point out the necessary modifications in the proof of p(·),q(Rn) p(·),q(Rn). The proof follows the scheme of Part II of the proof of Lemma 16. Lemma 16 to obtain a characterization by ball means of differences for Bs(·) and F s(·) We start with instead of . With this modification the Steps 1-3 go through ∞ Xk=−∞ ∞ Xk=0 without any other changes and we obtain (42) again (just recall that f(k) = 0 if k < 0). Due to the boundedness of the maximal operator there is no need for the use of r-trick and convolution with ην,m. The analogue of (43), (44) and (45) now reads as follows: 2ks(x)ZB ∆M 2−khf(k+l)(x)dh M cj,MZB .(cid:16)2ks(x)2laP2k+l,af(k+l)(x)(cid:17)1−λ ≤ 2la(1−λ)−ls−(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)1−λ Xj=0 M · 2ks(x+j2−kh)λf(k+l)(x + j2−kh)λdh cj,MZB · Xj=0 2(k+l)s(x+j2−kh)λf(k+l)(x + j2−kh)λdh . 2la(1−λ)−ls−(cid:16)2(k+l)s(x)P2k+l,af(k+l)(x)(cid:17)1−λ 30 cj,M M (2(k+l)s(·)f(k+l)(·)λ)(x), M Xj=0 where we used Holder's regularity of s(·), see (5). As a consequence, we obtain II 1/q(x) . F (x)1−λ · ∞ Xl=1 2−lδ ∞ Xk=−l (M Bλ k+l)q(x)/λ(x)!λ/q(x) . instead of (46). The rest then follows in the same manner with the help of Lemma 17 and the proof of (55) is finished. The proof of kf Bs(·) p(·),q(·)(Rn)k∗∗ . kf Bs(·) p(·),q(·)(Rn)k follows along the same lines. Especially, we get f (k),II . ∞ Xl=1 2la(1−λ)−ls− (g1 k+l(x))1−λ · (M (g2 k+l)λ)(x) instead of (53). The rest follows again by Lemma 17. Part II. Finally, we present how the characterization for q constant can help us to improve on the case of variable exponent q(·). In view of Lemma 16, it is enough to show that 2ks(x)q(x)(cid:0)dM But this is a consequence of . kf F s(·) p(·),q(·)(Rn)k. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 0 Xk=−∞ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 0 Xk=−∞ 2−k f (x)(cid:1)q(x)!1/q(x) (cid:12)(cid:12)(cid:12)(cid:12) 2−k f (x)(cid:1)q(x)!1/q(x) 2ks(x)q(x)(cid:0)dM .(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 0 2k(s(x)−ε)q−(cid:0)dM Xk=−∞ p(·),q−(Rn)k . kf F s(·) Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:12)(cid:12)(cid:12)(cid:12) 2−k f (x)(cid:1)q−!1/q− p(·),q(·)(Rn)k, . kf F s(·)−ε Lp(·)(Rn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:12)(cid:12)(cid:12)(cid:12) where ε > 0 is small enough and we used the differences characterization for fixed q and a trivial embedding theorem. The same arguments apply for the Besov spaces and the proof is finished. Remark 9. The somewhat complicated proof of Theorem 15 would work more direct and simpler if we could use versions of Lemmas 10 and 21 in (47) and (54) where the ℓq(·) summation runs over ν ∈ Z. For Triebel-Lizorkin spaces there seems to exist such an extension [14], but for Besov spaces the proof of Lemma 10 in [1] seems to be to customized to the situation ν ∈ N0. 31 4.5 Ball means of differences for 2-microlocal spaces As already remarked in Section 2.2 all the proofs for spaces of variable smoothness do also serve for 2-microlocal spaces. One just has to use the definition of admissible weight sequences and the property (5), see Remark 2. First of all we give the notation for the (quasi-)norms. For simplicity we just use the discrete versions, although it is also possible to give continuous versions of 2- microlocal weights, see [56, Definition 4.1]. In analogy to the spaces of variable smoothness we introduce the following norms and ∗∗ ∗∗ p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) p(·),q(·)(Rn)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:0)wk(x)dM =(cid:13)(cid:13) f Lp(·)(Rn)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)(cid:0)wk(x)dM k=−∞(cid:12)(cid:12)(cid:12) 2−k f (x)(cid:1)∞ k=−∞(cid:12)(cid:12)(cid:12) 2−k f (x)(cid:1)∞ ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) Lp(·)(ℓq(·))(cid:13)(cid:13)(cid:13) . Finally, the preceding calculations show that the following theorem is true. f Bw f F w (cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) Theorem 18. (i) Let p, q ∈ P log(Rn) with p+, q+ < ∞ and w ∈ W α M > α2 and α1,α2. Let (56) (57) α1 > σp−,q− ·h1 + α n · min(p−, q−)i . Then p(·),q(·)(Rn) = {f ∈ Lp(·)(Rn) : kf F w F w p(·),q(·)(Rn)k∗∗ < ∞} and k · F w p(·),q(·)(Rn)k and k · F w p(·),q(·)(Rn)k∗∗ are equivalent on F w p(·),q(·)(Rn). (ii) Let p, q ∈ P log(Rn) and w ∈ W α α1,α2. Let M > α2 and α1 > σp− ·(cid:20)1 + clog(1/q) n + α n · p−(cid:21) . Then Bw p(·),q(·)(Rn) = {f ∈ Lp(·)(Rn) : kf Bw p(·),q(·)(Rn)k∗∗ < ∞} and k · Bw p(·),q(·)(Rn)k and k · Bw p(·),q(·)(Rn)k∗∗ are equivalent on Bw p(·),q(·)(Rn). Remark 10. Again, if min(p−, q−) ≥ 1 in the F-case, or p− ≥ 1 in the B-case, then the conditions (56) and (57) simplify to α1 > 0. In the case of constant exponents p, q we obtain similar results to [8] and [35]. 4.6 Lemmas The following lemma is a variant of Lemma 6.1 from [17]. Lemma 19. Let s ∈ C log from (2) for s(·). Then loc (Rn) and let R ≥ clog(s) , where clog(s) is the constant 2νs(x)ην,m+R(x − y) ≤ c 2νs(y)ην,m(x − y) holds for all x, y ∈ Rn and m ∈ N0. 32 Lemma 20. Let r > 0, ν ≥ 0 and m > n. Then there exists c > 0, which depends only on m, n and r, such that for all g ∈ S′(Rn) with supp g ⊂ {ξ ∈ Rn : ξ ≤ 2ν+1}, we have g(x) ≤ c(ην,m ∗ gr(x))1/r, x ∈ Rn. The following lemma is the counterpart to Lemma 10 for Triebel-Lizorkin spaces. Lemma 21 ([17], Theorem 3.2). Let p, q ∈ P log(Rn) with 1 < p− ≤ p+ < ∞ and 1 < q− ≤ q+ < ∞. Then the inequality holds for every sequence (fν)ν∈N0 of Lloc 1 (Rn) functions and m > n. (cid:13)(cid:13) (ην,m ∗ f )∞ ν=0 Lp(·)(ℓq(·))(cid:13)(cid:13) ≤ c(cid:13)(cid:13) (fν)∞ ν=0 Lp(·)(ℓq(·))(cid:13)(cid:13) The following lemma is well known (cf. [52]). We sketch its proof for the sake of completeness. Lemma 22. Let a, b > 0, M ∈ N and h ∈ Rn. Let f ∈ S′(Rn) with supp f ⊂ {ξ ∈ Rn : ξ ≤ b}. Then there is a constant C > 0 independent of f, b and h, such that ∆M h f (x) ≤ C max(1, bha) · min(1, bhM )Pb,af (x) holds for every x ∈ Rn. Proof. The estimate f (x + jh) = f (x + jh) 1 + jbha · (1 + jbha) ≤ (1 + M bha) sup z∈Rn f (x − z) 1 + bza . max(1, bha)Pb,af (x), j = 0, . . . , M, holds for all the admissible parameters even without the assumption on f . Hence we need to prove only ∆M h f (x) ≤ C max(1, bha) · bhM · Pb,af (x). (58) Using the Taylor formula for the (analytic) function f , we obtain by direct calcula- tion ∆M h f (x) ≤ chM sup α=M sup y≤M h (Dαf )(x − y) 1 + bya · (1 + bya) ≤ c′hM max(1, bha) · sup α=M sup y≤M h (Dαf )(x − y) 1 + bya . If supp g ⊂ {ξ ∈ Rn : ξ ≤ 1}, then this may be combined with the Nikol'skij inequality, cf. [52, Section 1.3.1], in the form sup α=M sup z∈Rn (Dαg)(x − z) 1 + za . sup z∈Rn g(x − z) 1 + za to obtain ∆M h g(x) ≤ c′′hM max(1, ha) · sup z∈Rn g(x − z) 1 + za . (59) 33 If supp f ⊂ {ξ ∈ Rn : ξ ≤ b}, we define g(x) = f (x/b), apply (59) together with ∆M bhg(bx) and obtain h f (x) = ∆M ∆M h f (x) . bhM max(1, bha) sup z∈Rn g(bx − z) 1 + za . From this (58) follows and the proof is then complete. The following lemma resembles Lemma A.3 of [17]. Lemma 23. Let k ∈ Z, l ∈ N0 and m > n. Then ηk+l,m ∗ [2knχ2−kB] . ηk,m. Proof. Using dilations, we may suppose that k = 0. If x ≤ 2, then Z{y:x−y≤1} 2nl(1 + 2ly)−mdy ≤Zy∈Rn 2nl(1 + 2ly)−mdy . (1 + x)−m. If x > 2 and x − y ≤ 1, we obtain 1 + 2ly & 1 + 2lx and 2nl(1 + 2lx)−m . (1 + x)−m. This immediately implies that Z{y:x−y≤1} 2nl(1 + 2ly)−mdy .Z{y:x−y≤1} (1 + x)−mdy . (1 + x)−m. Remark 11. Another way, how to prove Lemma 23 is to use the inequality χB(x) ≤ 2mη0,m(x) and apply Lemma A.3 of [17]. The following Lemma is quite simple and we leave out its proof. Lemma 24. Let 0 < q < ∞, δ > 0 and let (al)l∈N be a sequence of non-negative real numbers. Then 2−lδqal!1/q . ∞ Xl=1 2−lδ/2a1/q l , ∞ Xl=1 where the constant involved depends only on δ and q. Finally, we shall need a certain version of Holder's inequality for ℓq(·)(Lp(·)) spaces. Lemma 25. Let p, q ∈ P(Rn) and let 0 < λ < 1. Then kfk · gkℓq(·)(Lp(·))k ≤ 21/q− kf 1/(1−λ) k ℓq(·)(Lp(·))k1−λ · kg1/λ k ℓq(·)(Lp(·))kλ (60) holds for all sequences of non-negative functions (fk)k∈N0 and (gk)k∈N0. 34 Proof. Due to the homogeneity, we may assume that f 1/(1−λ) k (cid:13)(cid:13)(cid:13) (cid:12)(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13) g1/λ k (cid:12)(cid:12)(cid:12) ℓq(·)(Lp(·))(cid:13)(cid:13)(cid:13) = 1. Then for every ε > 0, there exist two sequences of positive real numbers (λk)k∈N0 and (µk)k∈N0, such that and We put λk < 1 + ε, ∞ Xk=0 p(·) f 1/(1−λ) k λ1/q(·) k ! ≤ 1, ∞ µk < 1 + ε Xk=0 p(·) g1/λ k µ1/q(·) k ! ≤ 1. c := 21/q− and γk := λk + µk 2 ≥ λk + µk cq(x) and use the Young inequality in the form [fk(x)gk(x)]p(x) ≤ (1 − λ)fk(x)p(x)/(1−λ) + λgk(x)p(x)/λ to obtain ZRn fk(x)gk(x) cγ1/q(·) k !p(x) dx ≤ (1 − λ)ZRn ≤ (1 − λ)ZRn fk(x)p(x)/(1−λ) cp(x)γp(x)/q(x) fk(x)p(x)/(1−λ) k λp(x)/q(x) k dx + λZRn dx + λZRn k gk(x)p(x)/λ cp(x)γp(x)/q(x) gk(x)p(x)/λ µp(x)/q(x) k dx dx ≤ 1. Furthermore, the estimate γk < 2(1 + ε) 2 = 1 + ε ∞ Xk=0 finishes the proof of (60) with the constant c = 21/q− . Acknowledgement: The first author acknowledges the financial support pro- vided by the DFG project HA 2794/5-1 "Wavelets and function spaces on domains". Furthermore, the first author thanks the RICAM for his hospitality and support dur- ing a short term visit in Linz. The second author acknowledges the financial support provided by the FWF project Y 432-N15 START-Preis "Sparse Approximation and Optimization in High Dimen- sions". We thank the anonymous referee for pointing the reference [18] out to us. 35 References [1] A. Almeida, P. Hasto: Besov spaces with variable smoothness and integrability, J. Funct. Anal. 258 (2010), no. 5, 1628 -- 1655. [2] A. Almeida, S. Samko: Characterization of Riesz and Bessel potentials on vari- able Lebesgue spaces, J. Function Spaces Appl. 4 (2006), no. 2, 113 -- 144. [3] A. Almeida, S. Samko: Pointwise inequalities in variable Sobolev spaces and applications, Z. Anal. Anwend. 26 (2007), no. 2, 179 -- 193. [4] A. Almeida, S. Samko: Embeddings of variable Haj lasz-Sobolev spaces into Holder spaces of variable order, J. Math. Anal. Appl. 353 (2009), no. 2, 489 -- 496. [5] P. Andersson: Two-microlocal spaces, local norms and weighted spaces, Paper 2 in PhD Thesis (1997), 35 -- 58. [6] T. Aoki: Locally bounded linear topological spaces, Proc. Imp. Acad. Tokyo 18 (1942), 588 -- 594. [7] B. Beauzamy: Espaces de Sobolev et de Besov d'ordre variable d´efinis sur Lp, C.R. Acad. Sci. Paris (Ser. A) 274 (1972), 1935 -- 1938. [8] O. V. Besov: Equivalent normings of spaces of functions of variable smoothness, Proc. Steklov Inst. Math. 243 (2003), no. 4, 80 -- 88. [9] J.-M. Bony: Second microlocalization and propagation of singularities for semi- linear hyperbolic equations, Taniguchi Symp. HERT. Katata (1984), 11 -- 49. [10] F. Cobos, D. L. Fernandez: Hardy-Sobolev spaces and Besov spaces with a function parameter Proc. Lund Conf. 1986, Lect. Notes Math. 1302 (1986), Berlin: Springer, 158 -- 170. [11] D. Cruz-Uribe, A. Fiorenza, C. J. Neugebauer: The maximal function on vari- able Lp spaces, Ann. Acad. Sci. Fenn. Math. 28 (2003), 223 -- 238. [12] D. Cruz-Uribe, A. Fiorenza, J. M. Martell, C. P´erez: The boundedness of clas- sical operators in variable Lp-spaces, Ann. Acad. Sci. Fenn. Math. 31 (2006), 239 -- 264. [13] L. Diening: Maximal function on generalized Lebesgue spaces Lp(·), Math. In- equal. Appl. 7 (2004), no. 2, 245 -- 254. [14] L. Diening: private communication. [15] L. Diening, P. Harjulehto, P. Hasto, Y. Mizuta, T. Shimomura: Maximal func- tions in variable exponent spaces: limiting cases of the exponent, Ann. Acad. Sci. Fenn. Math. 34 (2009), (2), 503 -- 522. 36 [16] L. Diening, P. Harjulehto, P. Hasto, M. Ruzicka: Lebesgue and Sobolev Spaces with Variable Exponents, Springer, Lecture Notes in Mathematics 2017, Springer (2011). [17] L. Diening, P. Hasto, S. Roudenko: Function spaces of variable smoothness and integrability, J. Funct. Anal. 256 (2009), (6), 1731 -- 1768. [18] D. Drihem: Atomic decomposition of Besov spaces with variable smoothness and integrability, J. Math. Anal. Appl. 389 (2012), 15 -- 31. [19] W. Farkas, H.-G. Leopold: Characterisations of function spaces of generalised smoothness, Ann. Mat. Pura Appl. 185 (2006), no. 1, 1 -- 62. [20] M. L. Goldman: A description of the traces of some function spaces, Trudy Mat. Inst. Steklov 150 (1979), 99 -- 127; English transl.: Proc. Steklov Inst. Math. 150 (1981), no. 4. [21] M. L. Goldman: A method of coverings for describing general spaces of Besov type, Trudy Mat. Inst. Steklov 156 (1980), 47 -- 81; English transl.: Proc. Steklov Inst. Math. 156 (1983), no. 2. [22] M. L. Goldman: Imbedding theorems for anisotropic Nikolskij-Besov spaces with moduli of continuity of general type, Trudy Mat. Inst. Steklov 170 (1984), 86 -- 104; English transl.: Proc. Steklov Inst. Math. 170 (1987), no. 1. [23] P. Gurka, P. Harjulehto, A. Nekvinda: Bessel potential spaces with variable exponent, Math. Inequal. Appl. 10 (2007), no. 3, 661 -- 676. [24] S. Jaffard: Pointwise smoothness, two-microlocalisation and wavelet coeffi- cients, Publications Mathematiques 35, (1991), 155 -- 168. [25] S. Jaffard, Y. Meyer: Wavelet methods for pointwise regularity and local oscil- lations of functions, Memoirs of the AMS, vol. 123, (1996). [26] G. A. Kalyabin: Characterization of spaces of generalized Liouville differentia- tion, Mat. Sb. Nov. Ser. 104 (1977), 42 -- 48. [27] G. A. Kalyabin: Description of functions in classes of Besov-Lizorkin-Triebel type, Trudy Mat. Inst. Steklov 156 (1980), 82 -- 109; English transl.: Proc. Steklov Institut Math. 156 (1983), no. 2. [28] G. A. Kalyabin, P. I. Lizorkin: Spaces of functions of generalized smoothness, Math. Nachr. 133 (1987), 7 -- 32. [29] G. A. Kalyabin, Characterization of spaces of Besov-Lizorkin and Triebel type by means of generalized differences, Trudy Mat. Inst. Steklov 181, (1988), 95 -- 116; English transl.: Proc. Steklov Inst. Math. 181 (1989), no. 4. [30] H. Kempka: 2-microlocal Besov and Triebel-Lizorkin spaces of variable integra- bility, Rev. Mat. Complut. 22 (2009), no. 1, 227 -- 251. 37 [31] H. Kempka: Atomic, molecular and wavelet decomposition of 2-microlocal Besov and Triebel-Lizorkin spaces with variable integrability, Funct. Approx. 43 (2010), (2), 171 -- 208. [32] H. Kempka, J. Vyb´ıral: A note on the spaces of variable integrability and summability of Almeida and Hasto, submitted. [33] O. Kov´acik, J. R´akosn´ık: On spaces Lp(x) and W 1,p(x), Czechoslovak Math. J. 41 (1991), no. 4, 592 -- 618. [34] H.-G. Leopold: On function spaces of variable order of differentiation, Forum Math. 3 (1991), 1 -- 21. [35] J. L´evy V´ehel, S. Seuret: A time domain characterization of 2-microlocal Spaces, J. Fourier Analysis and Appl. 9 (2003), (5), 473 -- 495. [36] J. L´evy V´ehel, S. Seuret: The 2-microlocal formalism, Fractal Geometry and Applications: A Jubilee of Benoit Mandelbrot, Proceedings of Symposia in Pure Mathematics, PSPUM, vol. 72 (2004), part2, 153 -- 215. [37] C. Merucci: Applications of interpolation with a function parameter to Lorentz Sobolev and Besov spaces, Proc. Lund Conf. 1983, Lect. Notes Math. 1070 (1983), Berlin: Springer, 183 -- 201. [38] Y. Meyer: Wavelets, vibrations and scalings, CRM Monograph Series 9, AMS (1998). [39] S. Moritoh, T. Yamada: Two-microlocal Besov spaces and wavelets, Rev. Mat. Iberoamericana 20, (2004), 277 -- 283. [40] S. Moura: Function spaces of generalised smoothness, Diss. Math. 398 (2001), 1 -- 87. [41] A. Nekvinda: Hardy-Littlewood maximal operator on Lp(x)(Rn), Math. Inequal. Appl. 7 (2004), no. 2, 255 -- 266. [42] W. Orlicz: Uber konjugierte Exponentenfolgen, Studia Math. 3, (1931), 200 -- 212. [43] J. Peetre: On spaces of Triebel-Lizorkin type, Ark. Math. 13, (1975), 123 -- 130. [44] S. Rolewicz: On a certain class of linear metric spaces, Bull. Acad. Polon. Sci. S´er. Sci. Math. Astrono. Phys. 5, (1957), 471 -- 473. [45] B. Ross, S. Samko: Fractional integration operator of variable order in the spaces H λ, Int. J. Math. Sci. 18 (1995), no. 4, 777 -- 788. [46] M. Ruzicka: Electrorheological fluids: modeling and mathematical theory, Lec- ture Notes in Mathematics 1748, Springer, Berlin, (2000). [47] V. S. Rychkov: On a theorem of Bui, Paluszy´nski and Taibleson, Steklov Insti- tute of Mathematics 227, (1999), 280 -- 292. 38 [48] B. Scharf: Atomare Charakterisierungen vektorwertiger Funktionenraume, Diploma Thesis, Jena (2009). [49] C. Schneider: On dilation operators in Besov spaces, Rev. Mat. Complut. 22 (2009), no. 1, 111 -- 128. [50] C. Schneider, J. Vyb´ıral: On dilation operators in Triebel-Lizorkin spaces, Funct. Approx. Comment. Math. 41 (2009), part 2, 139 -- 162. [51] J. Schneider: Function spaces of varying smoothness I, Math. Nachr. 280 (2007), no. 16, 1801 -- 1826. [52] H. Triebel: Theory of function spaces, Basel, Birkhauser, (1983). [53] H. Triebel: Theory of function spaces II, Basel, Birkhauser, (1992). [54] T. Ullrich, Function spaces with dominating mixed smoothness, characteriza- tion by differences, Technical report, Jenaer Schriften zur Math. und Inform., Math/Inf/05/06, (2006). [55] T. Ullrich: Continuous characterizations of Besov-Lizorkin-Triebel spaces and new interpretations as coorbits, to appear in J. Funct. Spaces Appl. [56] T. Ullrich, H. Rauhut: Generalized coorbit space theory and inhomogeneous function spaces of Besov-Lizorkin-Triebel type, to appear in J. Funct. Anal. [57] A. Unterberger, J. Bokobza: Les op´erateurs pseudodiff´erentiels d'ordre variable, C.R. Acad. Sci. Paris 261 (1965), 2271 -- 2273. [58] A. Unterberger: Sobolev Spaces of Variable Order and Problems of Convexity for Partial Differential Operators with Constant Coefficients, Ast´erisque 2 et 3 SOC. Math. France (1973), 325 -- 341. [59] M. I. Visik, G. I. Eskin: Convolution equations of variable order (russ.), Trudy Moskov Mat. Obsc. 16 (1967), 26 -- 49. [60] J. Vyb´ıral: Sobolev and Jawerth embeddings for spaces with variable smoothness and integrability, Ann. Acad. Sci. Fenn. Math. 34 (2009), no. 2, 529 -- 544. [61] H. Xu: G´en´eralisation de la th´eorie des chirps `a divers cadres fonctionnels et application `a leur analyse par ondelettes, Ph. D. thesis, Universit´e Paris IX Dauphine (1996). [62] J.-S. Xu: Variable Besov and Triebel-Lizorkin spaces, Ann. Acad. Sci. Fenn. Math. 33 (2008), no. 2, 511 -- 522. [63] J.-S. Xu: An atomic decomposition of variable Besov and Triebel-Lizorkin spaces, Armenian J. Math. 2 (2009), no. 1, 1 -- 12. [64] W. Yuan, W. Sickel, D. Yang: Morrey and Campanato meet Besov, Lizorkin and Triebel, Lecture Notes in Mathematics 2005, Springer (2010). 39
1602.05557	2	1602	2016-06-23T15:48:17	Equiangular tight frames from hyperovals	[ "math.FA", "math.CO" ]	An equiangular tight frame (ETF) is a set of equal norm vectors in a Euclidean space whose coherence is as small as possible, equaling the Welch bound. Also known as Welch-bound-equality sequences, such frames arise in various applications, such as waveform design, quantum information theory, compressed sensing and algebraic coding theory. ETFs seem to be rare, and only a few methods of constructing them are known. In this paper, we present a new infinite family of complex ETFs that arises from hyperovals in finite projective planes. In particular, we give the first ever construction of a complex ETF of 76 vectors in a space of dimension 19. Recently, a computer-assisted approach was used to show that real ETFs of this size do not exist, resolving a longstanding open problem in this field. Our construction is a modification of a previously known technique for constructing ETFs from balanced incomplete block designs.	math.FA	math	Equiangular Tight Frames from Hyperovals Matthew Fickus Member, IEEE, Dustin G. Mixon and John Jasper 1 6 1 0 2 n u J 3 2 ] . A F h t a m [ 2 v 7 5 5 5 0 . 2 0 6 1 : v i X r a Abstract-An equiangular tight frame (ETF) is a set of equal norm vectors in a Euclidean space whose coherence is as small as possible, equaling the Welch bound. Also known as Welch-bound-equality sequences, such frames arise in various applications, such as waveform design, quantum information theory, compressed sensing and algebraic coding theory. ETFs seem to be rare, and only a few methods of constructing them are known. In this paper, we present a new infinite family of complex ETFs that arises from hyperovals in finite projective planes. In particular, we give the first ever construction of a complex ETF of 76 vectors in a space of dimension 19. Recently, a computer- assisted approach was used to show that real ETFs of this size do not exist, resolving a longstanding open problem in this field. Our construction is a modification of a previously known technique for constructing ETFs from balanced incomplete block designs. Index Terms-equiangular tight frame, Welch bound I. INTRODUCTION An equiangular tight frame is a type of optimal packing of lines in Euclidean space. To be precise, let d ≤ n be positive integers, and let Hd be a real or complex Hilbert space of dimension d. Welch [31] gives a lower bound on the coherence of any sequence {ϕi}n i=1 of n equal norm vectors in Hd: max i6=j hϕi, ϕji d(n − 1)(cid:21) kϕikkϕjk ≥ (cid:20) n − d 1 2 . (1)    It is well-known [25] that this lower bound is achieved if and i=1 is an equiangular tight frame (ETF) for Hd. only if {ϕi}n For example, when d = 6 and n = 16, the Welch bound (1) is 1 3 , and it is achieved by certain special choices of 16 vectors in R6, such as the columns of the following matrix: + − + − + − + − 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 + − + − + − + − + + − − 0 0 0 0 + + − − 0 0 0 0 0 0 0 0 + + − − 0 0 0 0 + + − − + − − + 0 0 0 0 0 0 0 0 + − − + 0 0 0 0 + − − + + − − + 0 0 0 0 Here, "+" and "−" denote 1 and −1, respectively. Having minimal coherence, ETFs are useful in a number of real-world applications, including waveform design for wireless communication [25], compressed sensing [2], [3], quantum information theory [22], [34] and algebraic coding theory [16]. Unfortunately, ETFs also seem to be rare. In particular, when Hd is a real Hilbert space, n and d necessarily satisfy certain strong integrality conditions [26]. Moreover,  . (2) M. Fickus and D. G. Mixon are with the Department of Mathematics and Statistics, Air Force Institute of Technology, Wright-Patterson Air Force Base, OH 45433, USA, e-mail: [email protected]. J. Jasper is with the Department of Mathematical Sciences, University of Cincinnati, Cincinnati, OH 45221, USA. This paper was presented in part at the 2016 Spring Central Sectional Meeting of the American Mathematical Society. only a few methods for constructing infinite families of ETFs are known, and with the exception of orthonormal bases and regular simplices, each of these methods involve some type of combinatorial design. Real ETFs in particular are equivalent to special types of strongly regular graphs (SRGs) [15], [29]. Such graphs have a rich literature [5], [6], [7]. The interrelated concepts of conference matrices, Hadamard matrices, Paley tournaments, and quadratic residues have all been used to construct various infinite families of ETFs in which n is either nearly or exactly 2d [25], [15], [21], [24]. Other constructions offer much more flexibility regarding the size of n d . These include harmonic ETFs, which are obtained by restricting the Fourier basis on a finite abelian group to a difference set for that group [25], [32], [9], and Steiner ETFs, which arise from a tensor-like product of a simplex and the incidence matrix of a certain type of balanced incomplete block design [12], [16]. To be clear, many of these ideas are rediscoveries or reimaginings of more classical results. In particular, see [20], [28] and [13] for precursors of the Welch bound, harmonic ETFs and Steiner ETFs, respectively. In this paper, we generalize the Steiner ETF construction of [12], [16] to construct a new infinite family of complex ETFs. In particular for any e ≥ 1, we construct an n-vector ETF for a complex d-dimensional Hilbert space where d = 22e + 2e − 1, n = 2e(22e + 2e − 1). (3) From our perspective, this construction is significant for two reasons. First, as noted above, ETFs seem to be rare, and few infinite families of them are known. Comparing against known constructions [11], it appears that with the exception of e = 1 case, no real or complex ETF with parameters (3) has been discovered before. In fact, taking e = 2 in (3) gives d = 19 and n = 76, and the existence of a real ETF of this size appears to have been recently ruled out with a computer-assisted search [1], resolving a longstanding open problem [33]. Second, though the construction itself is a generalization of Steiner ETFs [12], it exploits a new realization: like previous work, we construct n-vector ETFs for Hd as the columns of an m×n matrix; unlike previous work, we realize that sometimes, the most natural choice of m is neither d nor n. For example, the columns of the following 6 × 10 matrix form a 10-vector ETF for a 5-dimensional Hilbert space, namely the orthogonal complement of the all-ones vector 1 in R6:   + + + + + + + + + + + + + + − − − − − − + − − − + + + − − − − + − − + − − + + − − − + − − + − + − + − − − + − − + − + +   . (4) This can be verified by noting that all the columns of this matrix are orthogonal to 1 and that together they achieve equality in (1) for n = 10 and d = 5. Inspired by problems in experimental design, this example was produced by taking all (cid:0)6 3(cid:1) = 20 {±1}-valued vectors in R6 that are orthogonal to 1, and discarding one vector from each pair of antipodes. Our new construction is a generalization of this example: we form ETFs of size (3) as the columns of an m × n matrix where m = d + 1 = 22e + 2e. The construction only happens to be real in the e = 1 case, namely (4). In the next section, we introduce the background material we need for the rest of the paper. In Section III we show how to construct complex ETFs of size (3). The construction relies on special collections of lines in finite affine planes that arise from hyperovals in corresponding projective planes. In the fourth and final section, we discuss how these ETFs can be made flat. This leads to several interesting open problems. II. BACKGROUND A. Steiner equiangular tight frames Throughout, let the field of scalars F be either R or C, and let Hd be a d-dimensional inner product space over F. The synthesis operator of a finite sequence of vectors {ϕi}n i=1 in Hd is Φ : Fn → Hd, Φy := Pn i=1 y(i)ϕi. The adjoint of Φ is the analysis operator Φ∗ : Hd → Fn, (Φ∗x)(n) = hϕn, xi. Composing these two operators gives the frame operator ΦΦ∗ : Hd → Hd, ΦΦ∗x = Pn i=1hϕi, xiϕi and the n × n Gram matrix Φ∗Φ whose (i, j)th entry is hϕi, ϕji. A sequence of vectors {ϕi}n i=1 is a tight frame for Hd if there exists a > 0 such that ΦΦ∗ = aI. When a is specified, it is an a-tight frame. It is equal norm if there exists r > 0 such that kϕik2 = r for all i, and is equiangular if it is equal norm and there exists w ≥ 0 such that hϕi, ϕji = w for all i 6= j. When {ϕi}n i=1 is both equiangular and a tight frame, it is an equiangular tight frame (ETF). It is well known [25] i=1 in Hd achieve equality in (1) that equal norm vectors {ϕi}n if and only if they form an ETF for Hd; see Lemma 1 below for a generalized version of this result. In the special case where Hd is Fd, the synthesis operator of i=1 is simply the d×n matrix Φ whose ith column is ϕi. {ϕi}n In this case, {ϕi}n i=1 is an ETF for Fd if and only if the rows of Φ are orthogonal and have constant norm (tightness) and the columns of Φ are equiangular. As we now discuss, Steiner ETFs [12] are one way to directly construct such matrices; in Section III, we generalize this construction. A balanced incomplete block design (BIBD) is a finite set V-whose elements are called vertices-along with any set B of subsets of V-whose elements are called blocks-for which there exists positive integers λ, k, r such that: (i) every block contains exactly k vertices, (ii) every vertex is contained in exactly r blocks, (iii) any pair of vertices is contained in exactly λ blocks. Letting v and b denote the desired number of vertices and blocks, respectively, a {0, 1}-valued b × v matrix X is an incidence matrix of a BIBD precisely when X1 = k1, XTX = (r − λ)I + λJ, (5) 2 for some integers k, r and λ. Here and throughout, 1 and J denote all-ones vectors and matrices, respectively. Note the second condition in (5) implies 1TX = r1T. Moreover, v, k and λ determine r and b according to bk = vr, λ(v − 1) = r(k − 1). (6) To see this, note multiplying X on the left and right by 1T and 1, respectively, gives the first identity, while doing the same to the equation XTX = (r− λ)I + λJ gives the second. Because of this, such incidence structures are often called a 2-(v, k, λ) design, and are denoted as a BIBD(v, k, λ). Steiner ETFs arise from BIBDs with λ = 1, which are also known as (2, k, v)-Steiner systems. Having λ = 1 means that any two distinct vertices determine a unique block, and thus such BIBDs are a type of finite geometry. Indeed the canonical examples of such BIBDs are finite affine and projective planes of order q ≥ 2, namely BIBD(q2, q, 1) and BIBD(q2 + q + 1, q + 1, 1), respectively. As detailed later on in this section, such designs are known to exist whenever q is the power of a prime. When q = 2, we have the following incidence matrices X, for example:   1 1 0 0 0 0 1 1 1 0 1 0 0 1 0 1 1 0 0 1 0 1 1 0 ,   1 1 0 0 1 0 0 0 0 1 1 1 0 0 1 0 1 0 0 1 0 0 1 0 1 0 1 0 1 0 0 1 0 0 1 0 1 1 0 0 0 1 0 0 0 0 1 1 1     . (7) A Steiner ETF is formed by taking a tensor-like product of the b × v incidence matrix of a BIBD(v, k, 1) and a (possibly complex) Hadamard matrix of size r + 1. To define them rigorously, we borrow the following concept from [10]: Definition 1. Let X be a b × v incidence matrix of a BIBD(v, k, 1). For any j = 1, . . . , v, a corresponding em- bedding is an operator Ej : Fr → Fb that maps the standard basis of Fr to the standard basis of the r-dimensional subspace of Fb that consists of vectors supported on {i : X(i, j) = 1}. For example, for the 6 × 4 incidence matrix in (7) we can take E1, E2, E3 and E4 to be     1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0  ,   1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 ,    0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 1 ,     0 0 0 1 0 0 0 0 0 0 1 0 0 0 1 0 0 0   respectively. A Steiner ETF is formed by using these operators to embed v copies of a unimodular simplex for Fr into Fb. To be precise, let {sl}r+1 l=1 be the columns of an r × (r + 1) submatrix S of an (r + 1) × (r + 1) Hadamard matrix whose entries lie in F. For example, the 6× 4 incidence matrix in (7) has r = 3 and we can form S by removing the first row from the canonical 4 × 4 Hadamard matrix:  + − + − + + − −  . + − − + S =   (9) , (8) For any BIBD(v, k, 1) and unimodular simplex {sl}r+1 l=1 , r+1 the corresponding Steiner ETF is {Ejsl}v l=1 . As shown in [12], [10], these v(r + 1) vectors form an ETF for Fb, and their synthesis operator is the block matrix j=1, Φ = (cid:2) E1S ··· EvS (cid:3) . For example, using (8) to embed 4 copies of the unimodular simplex in (9) gives the 16-vector ETF for R6 given in (2). The main idea of the construction is that hEj sl, Ej ′ sl′i has unit modulus whenever (j, l) 6= (j′, l′): if j 6= j′, this follows from the fact that any two distinct blocks in the BIBD have exactly one vertex in common; if j = j′ but l 6= l′, this follows from the fact that the inner product of any two distinct columns of S has unit modulus. A Steiner ETF can be real whenever there exists a real Hadamard matrix of size r + 1. The famously-unresolved Hadamard conjecture proposes that this happens precisely when r + 1 is divisible by 4. B. Finite affine and projective planes In the next section, we generalize the Steiner ETF construc- tion in a way that yields complex ETFs with parameters (3). This construction relies on some unusually nice properties of finite affine and projective planes of order q, namely of BIBD(q2, q, 1) and BIBD(q2 + q + 1, q + 1, 1), respectively. (For the sake of brevity, we omit the standard, axiomatic definitions of these geometries). We will need the fact that any affine plane of order q is a subincidence structure of a projective plane of order q. This follows from a type of parallel postulate that any BIBD(v, k, 1) must satisfy. To be precise, let X be the incidence matrix of such a BIBD, and note the (i, i′)th entry of (k − 1)I + J − XXT is 1 when the ith and i′th blocks are disjoint, and is otherwise 0. The number of blocks disjoint from block i that contain vertex j is thus given by the (i, j)th entry of [(k − 1)I + J − XXT]X = (r − k)(J − X); here we have used (6) and (5) to simplify. As such, if a block does not contain a vertex, there are exactly r − k blocks disjoint from it that do contain that vertex. In particular, for a BIBD(q2, q, 1) we have r−k = 1 and so any vertex not in a block is contained in exactly one disjoint block. This implies that if blocks i and i′ are disjoint, and blocks i′ and i′′ are disjoint, then blocks i and i′′ are either equal or disjoint, or else any point in their intersection is not contained in exactly one block disjoint from block i′. Saying two blocks are parallel when they are either equal or disjoint, we thus have that parallelism is an equivalence relation on the blocks of an affine plane. As a consequence, any affine plane is resolvable, meaning its b = q(q + 1) blocks can be arranged as r = q + 1 groupings of v k = q blocks, each forming a partition for the set of vertices. For example, the incidence matrix of the affine plane of order 2 given in (7) is arranged so that its first and second blocks partition its vertices, as does its third and fourth blocks, and its fifth and sixth blocks. The fact that any affine plane is resolvable allows it to be extended to a projective plane of the same order. To do this, we append each block in any of the q + 1 parallel classes with a new "vertex at infinity" that is unique to that class, 3 and create a single new "block at infinity" that contains the q + 1 new vertices. For example, extending the 6× 4 incidence matrix of (7) in this manner yields the 7 × 7 matrix to its right. Conversely, any projective plane contains many affine planes: choose any block, and remove it as well as all vertices it contains. Projective planes are symmetric BIBDs, meaning v = b, or equivalently that k = r. In general, the dual of a BIBD(v, k, λ) is the incidence structure obtained by inter- changing the roles of vertices and blocks, or equivalently, taking the transpose of its incidence matrix X. Here, (6) implies r2 − bλ = r2(v−k) k(v−1) ≥ 0, at which point (5) implies the columns of (r − λ)− 1 2 ]J} are orthonormal. When the BIBD is symmetric, this matrix is square and so its rows are also necessarily orthonormal, implying XXT = (k−λ)I+λJ. Thus, the dual of a symmetric BIBD(v, k, λ) is another BIBD(v, k, λ). In particular, the dual of any projective plane is another projective plane. b [r ± (r2 − bλ) 2 {X − 1 1 Whenever q is the power of a prime, there is a canonical construction of affine and projective planes of order q. In particular, letting Fq be the field of q elements, we can form an affine plane by letting V = F2 q and letting B be the collection of all affine lines in this vector space, namely sets of the form {(x, y) ∈ F2 q : dx + ey + f = 0} (10) for some d, e, f ∈ Fq where (d, e) 6= (0, 0). Though this manner of parametrizing these lines is redundant-multiplying (d, e, f ) by a nonzero scalar yields the same line-it gives a simple way to produce parallel classes: fix (d, e) 6= (0, 0) and vary f . It also facilitates the embedding of this affine plane in a projective plane where vertices and blocks correspond to one- and two-dimensional subspaces of F3 q, respectively. To be precise, let [x, y, z] denote the set of all nonzero q. To form a scalar multiples of a nonzero vector (x, y, z) ∈ F3 projective plane, let V be the set of all such [x, y, z], and let B be all sets of the form {[x, y, z] : dx + ey + f z = 0} (11) for some [d, e, f ]. The mapping (x, y) 7→ [x, y, 1] embeds our canonical affine geometry into this projective geometry: for any [d, e, f ] with (d, e) 6= (0, 0), (11) is the image of (10) under this mapping, along with the "vertex at infinity" [−e, d, 0]. Note that as expected, each new vertex corresponds to a unique parallel class in the affine plane, and taken together they constitute a new "block at infinity," namely (11) where [d, e, f ] = [0, 0, 1]. Here, we can identify F3 q with the additive group of the field Fq3. To be precise, let α be a generator of the cyclic multiplicative group F×q3, and let (x, y, z) := x + yα + zα2 for all x, y, z ∈ Fq. Since each vertex [x, y, z] consists of all nonzero scalar multiples of the nonzero vector (x, y, z), it is a unique element of the quotient group F×q3 /F×q . That is, V = F×q3 /F×q = hαi/hαq2+q+1i ∼= Zq2+q+1. Moreover, each two-dimensional subspace of Fq3 is the kernel of a mapping of the form β 7→ tr(α−lβ) where tr : Fq3 → Fq, tr(β) := β + βq + βq2, is the field trace. As such, all blocks correspond to translates of the subset of Zq2+q+1 which correspond to members of F×q3 /F×q whose coset representatives have trace 0. The resulting incidence matrix is circulant, and corresponds to a Singer difference set. These explicit constructions show that affine and projective planes of order q exist whenever q is a power of a prime. Conversely, it is famously conjectured that if there exists an affine or projective plane of order q then q is necessarily a prime power. This conjecture remains open for many values of q, such as 12. We also note that for some prime powers q, there are constructions of projective planes of order q that are provably not Desarguesian, meaning they are not isomorphic to the canonical construction given above. Our new method for constructing complex ETFs can be applied to any projective plane of order q that contains a hyperoval, defined below. Applying this construction to the canonical projective plane of order q = 2e yields complex ETFs with parameters (3). If non-prime-power-order examples of such projective planes are discovered in the future, this same construction will yield yet more complex ETFs. III. NEW ETFS FROM HYPEROVALS In this section, we show how to construct ETFs with parameters (3). As detailed below, the main idea is that any projective plane whose dual contains a hyperoval also contains an affine plane that contains the dual of a BIBD(q + 1, 2, 1). This permits a new Steiner-like ETF construction involving Hadamard matrices of two distinct sizes. Some of the novelty here is that these ETFs most naturally arise in non-obvious subspaces of Fm. The following result characterizes such frames in general. Lemma 1. For any vectors {ϕi}n i=1 in Fm, let Φ be the m×n matrix whose ith column is ϕi for all i. For any a > 0, the following are equivalent: (i) {ϕi}n (ii) ΦΦ∗Φ = aΦ, (iii) (ΦΦ∗)2 = aΦΦ∗, (iv) (Φ∗Φ)2 = aΦ∗Φ. Also, if {ϕi}n i=1 is contained in a d-dimensional subspace Hd of Fm, then it forms an a-tight frame for Hd if and only if ΦΦ∗ = aΠ where Π is the m × m orthogonal projection matrix onto Hd. i=1 forms an ETF for its span if and only if As such, {ϕi}n i=1 is equiangular. In this case, letting (ii)–(iv) hold and {ϕi}n r = kϕik2, the dimension d can be computed from either the tight frame constant or the equiangularity constant: i=1 forms an a-tight frame for its span, a = rn d d(n − 1)(cid:21) , w = r(cid:20) n − d Alternatively, equal norm vectors {ϕi}n i=1 in a subspace Hd of Fm of dimension d form an ETF for Hd if and only if they achieve equality in (1). 1 2 . Proof: Fix {ϕi}n i=1 in Fm and a > 0, and let Hd be any d-dimensional subspace of Fm that contains {ϕi}n To show (i) is equivalent to (ii), note the codomain of the synthesis operator y 7→ Φy can either be regarded as Fm or i=1. 4 i=1 forms an a-tight frame Hd, and that by definition, {ϕi}n for Hd if and only if the restricted version of this operator, namely Φ : Fn → Hd, satisfies ΦΦ∗ = aI. When regarding Φ as an m × n matrix, this is equivalent to having ΦΦ∗x = ax for all x ∈ Hd. In the special case where Hd is taken to be i=1 = {Φy : y ∈ Fn}, this tells us (i) is equivalent span{ϕi}n to having ΦΦ∗Φy = aΦy for all y ∈ Fn, namely (ii). More generally, writing Hd = {Πx : x ∈ Fm}, we see that {ϕi}n i=1 forms an a-tight frame for Hd if and only if ΦΦ∗Πx = aΠx for all x ∈ Fm. To simplify this further, note that since Πϕi = ϕi for all i, we have ΠΦ = Φ and so Φ∗Π = (ΠΦ)∗ = Φ∗. Thus, {ϕi}n i=1 forms an a-tight frame for Hd if and only if ΦΦ∗ = aΠ, as claimed. We emphasize that the above argument is delicate, and fails if we do not assume that each ϕi lies in Hd. Indeed, without this i=1 to not form an a-tight assumption, it is possible for {ϕi}n frame for Hd and yet still have ΦΦ∗x = ax for all x ∈ Hd, namely tight pseudoframes [18]. Continuing, note that multiplying (ii) by Φ∗ on the right or left gives (iii) and (iv), respectively. Conversely, taking the singular value decomposition Φ = UΣV∗, if either (iii) or (iv) hold then the diagonal entries of the diagonal block of Σ are all either 0 or √a, implying ΣΣ∗Σ = Σ and so (ii). To summarize, we have proven that (i)–(iv) are equivalent as well as the second claim. For the remaining conclusions, we further assume that kϕik2 = r for all i, and generalize an argument of [16]. If {ϕi}n i=1 is an a-tight frame for Hd, cycling a trace gives ad = Tr(aΠ) = Tr(ΦΦ∗) = Tr(Φ∗Φ) = rn, and so a = rn d . As such, one way to measure the tightness of d Π. We {ϕi}n now simplify and bound this quantity using trace properties: i=1 is to take the Frobenius norm of ΦΦ∗ − rn d Tr(Φ∗Φ) + ( rn d )2 Tr(Π) − r2n( n d − 1) n n 0 ≤ Tr[(ΦΦ∗ − rn d Π)2] = Tr(Φ∗ΦΦ∗Φ) − 2 rn Xj=1 Xi=1 = Xj=1 Xi=1 hϕi, ϕji2 − r2n2 kϕikkϕjk(cid:17)2 (cid:16) hϕi,ϕji = r2 n n d j6=i ≤ r2n(n − 1)(cid:16)max i6=j hϕi,ϕji kϕikkϕjk(cid:17)2 − r2 n(n−d) d . (12) Solving for the coherence here gives the Welch bound (1). Moreover, if {ϕi}n i=1 is an ETF for Hd with hϕi, ϕji = w for all i 6= j, the two above inequalities become equalities, giving equality in (1) and w = r[ n−d 2 . Conversely, if (1) holds with equality, the final quantity in (12) is 0, implying i=1 is tight and both inequalities are equalities and so {ϕi}n equiangular. d(n−1) ] 1 These characterizations in hand, we turn to the construction itself. A hyperoval in a projective plane of order q is a set of q + 2 vertices where no three of these vertices lie in a common block. For example, the first four vertices (columns) of the projective plane of order 2 given in (7) are a hyperoval since no block (row) contains three of them. As we now detail, hyperovals decompose a projective plane into several other incidence structures. Lemma 2. A projective plane of order q contains a hyperoval if and only if it has an incidence matrix of form X = (cid:20) X1,1 X1,2 0 X2,2 (cid:21) (13) where X1,1 is the incidence matrix of a BIBD(q + 2, 2, 1). Here, the columns of X1,1 correspond to the vertices of the hyperoval. Moreover, in this case, q is even and XT 2,2 is the incidence matrix of a BIBD( 1 2 q(q − 1), 1 2 q, 1). Proof: Consider a BIBD(q2 +q +1, q +1, 1) that contains u vertices that are special in the sense that no three of them lie in a common block. Without loss of generality, the first u columns of the incidence matrix X correspond to these special vertices. The total number of special vertex-block pairs j=1(q + 1) = u(q + 1). At the 2(cid:1) pairs of distinct special vertices X(i, j) ≥ (cid:18)u is Pu same time, each of the (cid:0)u i=1 X(i, j) = Pu 2(cid:19)2 = u(u − 1), determine a unique block, and so j=1Pb u(q + 1) = u b Xi=1 Xj=1 implying u ≤ q + 2. Having u = q + 2 gives equality above, meaning all special vertex-block pairs occur in blocks contain- ing exactly two special vertices. That is, when the projective plane contains a hyperoval, there is an enumeration of its vertices and blocks so that its incidence matrix is of form (13) where X1,1 is the incidence matrix of a BIBD(q + 2, 2, 1). Conversely, when the projective plane has such an incidence matrix, its first q + 2 vertices form a hyperoval since each block contains either 2 or 0 of them. For the remaining conclusions, assume X1,1 is the incidence matrix of a BIBD(q + 2, 2, 1). Here, (6) implies X1,1 is of size 1 2 (q + 1)(q + 2) × (q + 2), implying the size of X2,2 is 1 2 q(q− 1)× (q2− 1). Also, recall the dual of a projective plane is another projective plane, implying XXT = qI + J and so 2,2 is the incidence matrix X2,2XT of BIBD( 1 2 q1. To see this, note X1,11 = 21. Also, the upper-right block of the equation XTX = qI + J gives XT 1,1X1,2 = J, where J is (q + 2) × (q2 − 1). Together, these facts imply 2 q, 1) it suffices to show XT 2,2 = qI+J. Thus, to show XT 2 q(q − 1), 1 2,21 = 1 (q + 2)1T = 1TJ = 1TXT 1,1X1,2 = 2(1TX1,2). Thus, 1TX1,2 = 1 Also, since 1TX = (q + 1)1T then 2 (q + 2)1T, which implies that q is even. (q + 1)1T = 1TX1,2 + 1TX2,2 = 1 2 q1. Thus, we indeed have XT 2,21 = 1 2 (q + 2)1T + 1TX2,2. The partition given in Lemma 2 has a geometric meaning: the column partition indicates whether a vertex is in the hyperoval or not, while the row partition indicates whether a block intersects the hyperoval in two or zero vertices, namely whether the block is secant or exterior to the hyperoval. The matrix XT 2,2 given in Lemma 2 is the incidence matrix of a type of Denniston design [8]. As an example of this lemma, note the incidence matrix of the projective plane of 5 order 2 given in (7) is already in this form. There, X2,2 is a 1 × 3 matrix of ones, corresponding to a degenerate BIBD with v = k = λ = 1 and b = r = 3. To prove our main result, we need an infinite family of matrices of form (13), and thus need a general construction of hyperovals. Recall that projective planes of order q are only known to exist when q is the power of a prime. When coupled with the requirement that q be even, this means that all known constructions of hyperovals lie in projective planes of order q = 2e for some e ≥ 1. For the canonical projective plane of order q = 2e, the canonical hyperoval is set: {[t, t2, 1] : t ∈ Fq} ∪ {[0, 1, 0]} ∪ {[1, 0, 0]}. (14) No three of these vertices lie in a common block, since no three of the corresponding vectors are linearly dependent, a fact that follows from the corresponding 3 × 3 determinants. We form new ETFs from the duals of projective planes containing hyperovals, as well as special affine planes that they contain. The requisite designs are produced by the following: Lemma 3. If a projective plane of order q contains a hyper- oval then its dual has an incidence matrix of the form 0 Y2,2 (cid:21) Y = (cid:20) Y1,1 Y1,2 where Y1,1 and YT 2,2 are the incidence matrices of a BIBD( 1 2 q, 1) and BIBD(q + 2, 2, 1), respectively. Moreover, removing any one of the last q + 2 rows of Y along with the q + 1 columns it indicates produces an affine plane of order q with an incidence matrix of the form 2 q(q − 1), 1 (15) Z2,2 (cid:21) Z = (cid:20) Y1,1 Z1,2 0 (16) where ZT 2,2 is the incidence matrix of a BIBD(q + 1, 2, 1). Proof: For (15), take the transpose of the decomposition given in Lemma 2 and then permute rows and columns. For the remaining conclusions, recall from Section II that removing any single block (row) of Y along with the q + 1 vertices (columns) it contains produces an affine plane of order q. Choosing one of the last q+2 blocks in particular removes one row and q + 2 columns from Y2,2, removes one row from 0, removes q + 1 columns from Y1,2 and leaves Y1,1 untouched, resulting in a matrix of form (16). Moreover, since the columns of Y2,2 indicate all pairs of q + 2 rows, the columns of Z2,2 indicate all pairs of the remaining q + 1 rows, meaning ZT 2,2 is the incidence matrix of a BIBD(q + 1, 2, 1). For example, the projective plane of order 2 given in (7) contains a hyperoval since it has form (13). Taking its trans- pose and then permuting rows and columns gives a 7 × 7 incidence matrix of a projective plane that has form (15):   1 0 0 1 1 0 0 1 0 1 0 0 1 0 1 1 0 0 0 0 1 0 1 1 1 0 0 0 0 1 0 0 1 1 0 0 0 1 0 1 0 1 0 0 0 1 0 1 1 ,   1 1 0 0 1 0 1 0 1 0 0 1 0 1 1 0 0 1 0 1 0 0 1 1   .   (17) The second matrix here is an example of an incidence matrix of an affine plane that has form (16). It is obtained by removing the fourth row and columns 2, 3 and 4 from the first matrix. We form Steiner-like ETFs by using such BIBDs to embed the rows of two distinct (possibly complex) Hadamard matri- ces, one of size q + 2 and the other of size q. As with Steiner ETFs, we remove a row from a Hadamard matrix of size q + 2 to obtain a (q + 1)× (q + 2) matrix S whose columns {sl}q+2 l=1 form a unimodular simplex for Fq+1. See (9) for an example of S for q = 2. We also take the negative of the last row of a q × q Hadamard matrix and attach it to its bottom to produce a (q + 1) × q matrix C. For example, when q = 2, C =   + + + − − +   . (18) . (19)    The columns of such a matrix C have the following properties: Definition 2. For any r ≥ 3, a corresponding unimodular cosimplex is a sequence of vectors {cl}r−1 l=1 in Fr with the property that each cl has unimodular entries, the last two entries of any cl sum to zero, and hcl, cl′i = 1 for all l 6= l′. To form new ETFs, let Z be the incidence matrix of an affine plane of order q of form (16), use the first 1 2 q(q − 1) of the operators {Ej}q2 j=1 (cf. Definition 1) to embed S, and use the remaining operators to embed C. For example, using the first column of the 6 × 4 affine plane in (17) to embed (9) and using the remaining columns to embed (18) gives (cid:2) E1S E2C E3C E4C (cid:3), namely  + − − + + + 0 0 0 0 + + − − 0 0 + + 0 0 + − + − 0 0 0 0 + + 0 0 0 0 + − + − 0 0 0 0 0 0 − + 0 0 + − 0 0 0 0 0 0 − + − + Note that by inspection, the 10 columns of this matrix are equiangular with coherence 1 3 . However, these vectors do not form an ETF for R6: the rows of (19) are not orthogonal, and moreover letting n = 10 and d = 6 in (1) does not yield 1 3 . Rather, by the final statement of Lemma 1, they form an ETF for a 5-dimensional orthogonal complement of (0, 0, 0, 1, 1, 1). When applied to the affine plane produced by Lemma 3 from the canonical hyperoval and projective plane of order n = 4, this same approach yields the 76-vector ETF for a 19- dimensional subspace of C20, cf. Figure 1. It can also be ap- plied to projective planes of form (15). For example, using the 7 × 7 projective plane in (17) to embed (9) and (18) produces (cid:2) E1S E2C E3C E4C E5C E6C E7C (cid:3), a 16- vector ETF for a 6-dimensional orthogonal complement of (0, 0, 0, 1, 1, 1, 1) in R7. We now formally prove these facts. Theorem 1. For a projective plane of order q that contains a hyperoval, let {Ej}q2 j=1 be embeddings arising from an affine plane of form (16), cf. Definition 1. Let {sl}q+2 l=1 and {cl}q l=1 be a unimodular simplex and cosimplex for Fq+1, respectively, cf. Definition 2. Then {Ejsl}q+2 l=1, 1 2 q(q−1) j=1 ∪ {Ejcl}q l=1, q2 j= 1 2 q(q−1)+1 (20) 6 is a q(q2 + q− 1)-vector ETF for the (q2 + q− 1)-dimensional subspace of Fq(q+1) that consists of those vectors whose last q + 1 entries sum to zero. Moreover, if we instead let {Ej}q2+q+1 arising from a projective plane of form (15), then be embeddings j=1 {Ejsl}q+2 l=1, 1 2 q(q−1) j=1 ∪ {Ejcl}q l=1, q2+q+1 j= 1 2 q(q−1)+1 (21) is a q2(q+2)-vector ETF for the q(q+1)-dimensional subspace of Fq2+q+1 of vectors whose last q + 2 entries sum to zero. 1 Proof: We prove this result in the affine case; the proof in the projective case is similar. When n = q(q2 + q − 1) q+1 . Moreover, and d = q2 + q − 1, the Welch bound (1) is by definition, each embedding Ej maps an orthonormal basis of its domain to one of its range. This means each Ej is an isometry, namely E∗j Ej = I. Since each sl and cl has q + 1 entries, all unimodular, this implies kEjslk2 = kslk2 = q + 1 and similarly kEjclk2 = q + 1. As such, to prove the first claim using Lemma 1, it thus suffices to show that (i) each of the vectors in (20) lies in a hyperplane of Fq(q+1) and that (ii) the inner product of any two of these vectors has unit modulus. For (i), we claim all vectors in (20) lie in d⊥ where d ∈ Fq(q+1), d(i) = (cid:26) 0, 1 ≤ i ≤ q2 − 1, 1, q2 ≤ i ≤ q(q + 1). Indeed, for any j with 1 ≤ j ≤ 1 2 q(q − 1), the form (16) of our affine plane gives (Ej sl)(i) = 0 for all l whenever q2 ≤ i ≤ q(q + 1), and so hd, Ej sli = 0. In the remaining case where 1 2 q(q − 1) + 1 ≤ j ≤ q2 + q + 1, the form of our affine plane along with a property of the cosimplex gives hd, Ejcli = q(q+1) Xi=q2 (Ej cl)(i) = cl(q) + cl(q + 1) = 0. For (ii), since each Ej is an isometry and {sl}q+2 l=1 is a uni- modular simplex, hEj sl, Ejsl′i = hsl, sl′i = 1 whenever l 6= l′. Similarly, since {cl}q+2 l=1 is a unimodular cosimplex, hEj cl, Ejcl′i = hcl, cl′i = 1 whenever l 6= l′. A different argument is required for inner products of the form hEj sl, Ej ′ sl′i, hEj sl, Ej ′ cl′i, hEj cl, Ej ′ cl′i, (22) when j 6= n′. There, the fact that any two distinct blocks in our BIBD have exactly one vertex in common implies that exactly one column of Ej appears as a column of Ej ′ while all other columns have disjoint support. This implies E∗j Ej ′ is the outer product δiδ∗i′ of two standard basis elements of Fr, cf. Lemma 2.1 of [10]. In particular, hEj sl, Ej ′ sl′i = hsl, E∗j Ej ′ sl′i = hsl, δiδ∗i′ sl′i = sl(i)sl′(i′). Similarly, the other two inner products in (22) are sl(i)cl′ (i′) and cl(i)cl′ (i′). Since the entries of every sl and cl are unimodular, these inner products have unit modulus. We have some remarks about this result. First, recall that all known constructions of hyperovals lie in projective planes of order q = 2e for some e ≥ 1. Further recall that for any such q, the canonical projective plane of order q contains the hyperoval (14). In this case, we construct the requisite unimodular simplex and cosimplex from Hadamard matrices a a a a a a b c d e b b c c d e d e b c d e b c d e b c d e f g h i j f g h i j f g h i j Φ =   f f f f g g h i j g h i j h i j f g h g h i j i j f g h i j ,   f g h i j a b c d e f g h i j     7 ω = exp( 2πi 6 ), = =         1 ω2 ω4 1 ω2 ω4 1 ω4 ω2 1 ω4 ω2 1 ω3 ω3 ω3 1 1 1 ω2 ω4 ω3 ω5 ω1 1 ω4 ω2 ω3 ω1 ω5 ,   1 1 1 −1 1 1 −1 −1 −1 1 1 1 −1 1 −1 −1 1 1 −1 1   . Fig. 1. A complex ETF of n = 76 vectors for the 19-dimensional subspace of C20 that consists of vectors whose last 5 entries sum to zero. It was recently shown that no real ETF with parameters n = 76, d = 19 exists [1], [33]. To increase readability, we denote the rows of a 5 × 6 and 5 × 4 unimodular simplex and cosimplex by the first 10 letters of the alphabet. Blank entries denote rows of zeros of the appropriate size. This ETF is formed in the manner of Theorem 1: a hyperoval in a projective plane of order 4 decomposes a certain affine plane, cf. Lemma 3; this affine plane is then used to embed 6 copies of the simplex and 14 copies of the cosimplex into C20. The first 36 vectors form a known Steiner ETF for C15 arising from a Denniston design [12]. The last 40 of these vectors are real. In detail, we form a projective plane of order q = 22 = 4 by identifying F3 q3 = F64 = Z2(α), where α is a root of the primitive polynomial β6 + β + 1 over Z2 [14]. For the sake of simplicity, we performed field calculations in MATLAB using an isomorphism between 4 = hαi/hα21i ∼= Z21, and all blocks correspond to F64 and certain 6 × 6 matrices over Z2 [30]. The set of vertices in our projective plane is V = F× cyclic shifts of the Singer difference set {i ∈ Z21 : 0 = tr(αi) = αi + α4i + α16i} = {3, 6, 7, 12, 14}. Here, the canonical hyperoval {0, 1, 2, 3, 5, 14} is obtained by taking the logarithm base α of the vertices {t + t2α + α2 : t ∈ F4} ∪ {1} ∪ {α} given in (14). This hyperoval provides a 21 × 21 incidence matrix of form (13). Permuting its dual (transpose) gives a matrix of form (15). Removing a "hyperoval" row and its corresponding columns produces the 20 × 16 affine plane of form (16) that we used above. q with the additive group of F 64/F× of size q + 2 = 2e + 2 and q = 2e, respectively. The second of these two Hadamard matrices can always be chosen to be real, obtained for example by taking the Kronecker product of the canonical 2×2 Hadamard matrix with itself e times. However, q+2 = 2e+2 is not divisible by 4 except when e = 1, meaning our first Hadamard matrix is necessarily complex except when q = 2. That is, we obtain our (complex) unimodular simplex by removing a row from a discrete Fourier transform of size q + 2, for example. Together, these facts imply: Corollary 1. For any e ≥ 1, applying Theorem 1 to the canonical projective plane of order q = 2e gives an n-vector ETF for Cd where n and d are given by (3). This ETF is real when e = 1, and is otherwise complex. A second remark: while the ETF (20) arising from an affine plane of order q = 2e is new, the ETF (21) arising from a projective plane may not be. To be precise, two n-vector ETFs for Hd are equivalent if their synthesis operators satisfy Φ1 = UΦ2DP for some unitary operator U on Hd and some n×n matrices D and P where D is diagonal with unimodular diagonal entries and P is a permutation. Two ETFs that are constructed in different ways may, in fact, be equivalent. For example, in [16], it is shown that every harmonic ETF arising from a McFarland difference set [9] is equivalent to a special type of Steiner ETF arising from an affine geometry [12]. The ETF (21) has n = q2(q+2) and d = q(q+1), and so it might be equivalent to Steiner ETFs from affine planes. We do not know: determining whether two ETFs are equivalent is similar to determining whether two graphs are isomorphic, and we leave a deeper investigation of (21) for future work. For any q = 2e where e > 1, the ETF (20) is new [11]. Indeed, strongly regular graphs corresponding to real ETFs with these parameters are not known to exist [5], [6], and have been shown to not exist when q = 4 [1], [33]. Moreover, no ETF with these parameters can be harmonic since d(d − 1) n − 1 = q + 1 − 2q + 3 (q + 1)2 is not an integer; for harmonic ETFs, this quantity is the number of ways a nonzero element of the group can be written as a difference of members of the difference set. Also, no ETF with these parameters is a Steiner ETF, since by Theorem 2 of [12], the corresponding BIBD(v, k, 1) would have b = q2 + q − 1 and r = q + 1 and so k = r(r − 1) r2 − b which is not an integer. = q − 1 + 2 q + 2 1 l=1, 2 q(q−1) j=1 2 q(q − 1), 1 Another remark about Theorem 1: note (16) contains the incidence matrix Y1,1 of a BIBD( 1 2 q, 1), and is also contained in (15). This means the Steiner ETF {Ejsl}q+2 arising from this Denniston design is con- tained in the ETF (20), which in turn is contained in the ETF (21). To our knowledge, this is the first time three nontrivial nested ETFs have been discovered. This possibility suggests a new program for discovering ETFs: given an existing ETF, find other ETFs it contains as well as other ETFs that contain it. A final remark: note that by Lemma 1, the frame operator of (20) is necessarily a scalar multiple of a projection onto the subspace of vectors in Fq(q+1) whose last q + 1 entries sum to zero. In fact, this condition implies the construction of Theorem 1 does not generalize to any BIBDs that are not affine or projective planes. To be precise, let X = (cid:20) X1,1 X1,2 0 X2,2 (cid:21) (23) be the incidence matrix of any BIBD(v, k, 1) where X1,1 and 2,2 are the incidence matrices of some BIBD(v0, k0, 1) and XT BIBD(b0, 2, 1), respectively. Here, (5) and (6) imply k(k + 1) , 2r v0 = r(k0 − 1) + 1, b0 = k + 1. k0 = k − Use the first v0 columns of X to embed an r × (r + 1) unimodular simplex S, and use the remaining columns to embed an r×(r−1) unimodular cosimplex C `a la Theorem 1. Let Φ be the resulting b× [v(r − 1) + 2v0] synthesis operator. The columns of Φ are equiangular; we want to know when they form a tight frame for their span. Since SS∗ = (r + 1)I, CC∗ = (r − 1)  I 0 0 −1 0 0 1 −1 1 ,   the structure of X implies the frame operator is 0 0 ΦΦ∗ = (cid:20) k(r + 1 − k+1 r )I (r − 1)[(k + 1)I − J] (cid:21) . Computing the spectrum of ΦΦ∗, we see it is a multiple of an orthogonal projection if and only if k(r + 1 − k+1 r ) = (k + 1)(r − 1), or equivalently, 0 = (r − k)[r − (k + 1)]. By Lemma 1, we thus see these equiangular vectors form an ETF for their span if and only if either r = k or r = k + 1, namely when the underlying BIBD is either a projective plane or affine plane of form (23). IV. FLAT ETFS A matrix Φ is flat if all of its entries have constant modulus. Flat waveforms are often used in real-world applications such as radar and wireless communication since they allow a transmitted waveform to have the maximal amount of energy subject to the transmitter's power limit. That is, mathemati- cally speaking, flat vectors provide the largest possible ratio of 2-norm to ∞-norm. In light of this, it is natural to investigate ETFs that have flat synthesis operators. Previous work on this topic has focused on flat m × n matrices whose columns form an ETF for Fm. In particular, 8 the synthesis operator of a harmonic ETFs is flat, being a submatrix of a character table (discrete Fourier transform) of a finite abelian group. Some of these ETFs have recently been used to construct minimally-coherent vectors in a regime where no ETF can exist [4]. Steiner ETFs, on the other hand, are very sparse. Nevertheless, whenever the underlying BIBD is resolvable, one can rotate the synthesis operator of a Steiner ETF so as to make it flat [16]. We begin this section by generalizing the method of [16] so as to apply it to the ETF in (20). This produces a flat m × n matrix Φ whose columns form an ETF for a proper subspace Hd of Fm. We then consider such ETFs in general, discussing a connection between them and supersaturated designs, as well as how some of them imply the existence of other ETFs. From Section II, recall that parallelism is an equivalence relation on the blocks of an affine plane, and that this implies its blocks can be arranged as q + 1 groupings of q disjoint blocks. In particular, two rows of (16) lie in the same parallel class precisely when they have disjoint support. Let P be a permutation matrix that rearranges the rows of (16) into these parallel classes. For example, for the affine plane of order 2 in (17), take P to be the 6 × 6 permutation matrix such that P   1 1 0 0 1 0 1 0 1 0 0 1 0 1 1 0 0 1 0 1 0 0 1 1   = 0 0 1 1 1 1 0 0 0 1 0 1 1 0 1 0 0 1 1 0 1 0 0 1   .   (24) Since a permutation is unitary, applying it to the synthesis operator Φ of the ETF (20) that arises from (16) yields another ETF. For example, recall the columns of the 6 × 10 matrix Φ given in (19) form an ETF for the orthogonal complement of (0, 0, 0, 1, 1, 1). Taking P from (24), the columns of PΦ =   0 0 0 0 0 0 − + − + + − − + + + 0 0 0 0 0 0 0 0 − + 0 0 + − + + − − 0 0 + + 0 0 0 0 0 0 + − + − 0 0 + − + − 0 0 0 0 + +   , (25) thus form a 10-vector ETF for the orthogonal complement of (1, 0, 1, 0, 1, 0). Note here that we have chosen the first member of every parallel class from the bottom q + 1 rows of (16). This is always possible since ZT 2,2 is the incidence matrix of a BIBD(q + 1, 2, 1), meaning none of those q + 1 rows have disjoint support. Doing so ensures that the columns of PΦ are orthogonal to 1 ⊗ δ1 in general (the concatenation of q + 1 copies of the first standard basis element in Fq). With PΦ in this form, we now multiply it on the left by a block-diagonal matrix I ⊗ H whose diagonal blocks are all some Hadamard matrix H of size q whose first column is 1. Since H is a scalar multiple of a unitary operator, the columns of (I ⊗ H)PΦ form an ETF for the orthogonal complement of (I ⊗ H)(1 ⊗ δ1) = 1 ⊗ 1 = 1. For example, multiplying (25) by I ⊗ H where H is the canonical 2 × 2 Hadamard matrix gives a 10-vector ETF for the orthogonal complement of (1, 1, 1, 1, 1, 1), namely for the space of vectors in R6 whose entries sum to zero:   + − − + + + − + − + − + + − − − − + − + + + − − − + + + + − − − + + − + − − + − + − + − + − + − + + − + − + + − + − − −   . As seen in this example, (I⊗ H)PΦ is flat since each column of PΦ has only one index of support in each parallel class. Note here that whenever q = 2e for some e ≥ 1 (the only case where projective planes that contain hyperovals are known to exist), both the requisite cosimplex and Hadamard matrix can be chosen to be real, implying the last 1 2 q2(q + 1) vectors in the ETF are {±1}-valued and also have the property that their entries sum to zero. For example, permuting the rows of the 20× 76 in Figure 1 and then multiplying it by 5 copies of the canonical 4 × 4 Hadamard matrix gives a flat 20 × 76 matrix whose columns form a 76-vector ETF for the space of vectors in C20 whose entries sum to zero. The last 40 of these vectors are {±1}-valued while the first 36 take values from the sixth roots of unity. In summary: Theorem 2. For a projective plane of order q that contains a hyperoval, let Φ be the synthesis operator of the ETF (20). There exists a permutation matrix P and a Hadamard matrix H of size q such that (I ⊗ H)PΦ is flat (unimodular) and its columns form an ETF for the space of vectors in Fq(q+1) whose entries sum to zero. In particular, when q = 2e, e ≥ 1, choosing the simplex, cosimplex and Hadamard matrix appropriately yields an ETF of this type where each vector's entries are (q + 2)th roots of unity, with the last 1 2 q2(q + 1) vectors being {±1}-valued. A. Flat ETFs and supersaturated designs The flat ETFs given in Theorem 2 are closely related to supersaturated designs in the field of statistics known as design of experiments. There, one seeks {±1}-valued m× n matrices i=1 are orthogonal to 1 and are also Φ whose columns {ϕi}n maximally orthogonal in some sense. To relate that theory to our work here, recall the proof of Lemma 1: for any {ϕi}n in a d-dimensional subspace Hd of Fm with kϕik2 = r for all i, rewriting (12) gives: i=1 r2(n−d) d(n−1) ≤ 1 n(n−1) n Xi,j=1 i6=j hϕi, ϕji2 ≤ max i6=j hϕi, ϕji2, (27) where the first and second inequalities hold with equality pre- cisely when {ϕi}n i=1 is a tight frame for Hd and equiangular, respectively. In supersaturated designs, each ϕi is restricted to be {±1}-valued and lie in the orthogonal complement of 1 ∈ Rm. Under these assumptions, the first half of (27) becomes m2(n − m + 1) (m − 1)(n − 1) ≤ E(s2) := 1 n(n − 1) n Xi6=j hϕi, ϕji2. 9 In the supersaturated design literature, this inequality is well- known [19], [27], and {±1}-valued tight frames for 1⊥ are called E(s2)-optimal designs. In light of the second half of (27), we pose the following problem: (26) Problem 1. For what values of m, n does there exist a i=1 of {±1}-valued vectors that forms an ETF sequence {ϕi}n for the orthogonal complement of 1 in Rm, and so has (m − 1)(n − 1)(cid:21) hϕi, ϕji = m(cid:20) (n − m + 1) 1 2 , ∀i 6= j? Such ETFs are optimal in minimax sense, cf. [23]. It is unclear when they exist: reviewing both the ETF and supersaturated design literature, the only example of such an ETF we could find is when m = 6 and n = 10, namely (26) or the equivalent ETF (4). In the case where m = q(q + 1), n = q(q2 + q − 1) for some q = 2e, e > 1, the ETFs of Theorem 2 almost work. The only issue is that the entries of their first 1 2 q(q2 + q − 2) vectors are (q + 2)th roots of unity, not necessarily ±1. As a partial solution to Problem 1, we have the following necessary condition on m and n: Theorem 3. If m > 2 and there are n vectors in Rm with entries in {±1} that form an ETF for 1⊥, then there exists an even integer q ≥ 2 such that m = q(q + 1), n = q(q2 + q − 1). (28) Proof: We exploit known integrality conditions on the existence of real ETFs: if 1 < d < n − 1 and n 6= 2d, and there exists an n-vector ETF for a d-dimensional real Hilbert space, then both 1 2 , n − d (cid:21) (cid:20) d(n − 1) (cid:20) (n − d)(n − 1) d (cid:21) 1 2 , (29) are odd integers [26]. These quantities are the reciprocals of the Welch bounds for the ETF and its Naimark complements, respectively. Here, our Hilbert space 1⊥ ⊂ Rm has dimension d = m − 1 > 1. Since our vectors are {±1}-valued, their inner products are integers. Since they are also orthogonal to 1, m = d + 1 is necessarily even. Since they form an ETF for 1⊥, they achieve the Welch bound (1), meaning the absolute value of their inner products is d(n − 1)(cid:21) (d + 1)(cid:20) n − d 1 2 . (30) Together, these facts imply (30) is an integer. This immediately rules out d = n − 1, since in this case (30) becomes 1 + 1 d . Moreover, in the case n = 2d, a polynomial long division reveals the square of (30) to be (d + 1)2 2d − 1 = 1 4(cid:18)2d + 5 + 9 2d − 1(cid:19). Since d is odd, this is only an integer when d = 5, in which case (m, n) = (6, 10) is of form (28) with q = 2; an example of such an ETF is given in (4). Knowing 1 < d < n − 1 and having handled the case where n = 2d, we now assume n 6= 2d, implying the numbers in (29) are odd integers. As such, products of (29) and (30) are integers. In particular, d(n − 1)(cid:21) (d + 1)(cid:20) n − d 1 2(cid:20) (n − d)(n − 1) d (cid:21) 1 2 is an integer, implying the redundancy q := n Squaring the integer (30) and writing n = qd gives = n − d − 1 + n d d is an integer. (q − 1)(d + 1)2 (qd − 1) = r2 (cid:20)qd + 2q + 1 + r − 1 (q + 1)2 qd − 1 (cid:21), implying qd−1 divides (q−1)(q +1)2. As such, there exists a positive integer j such that (q−1)(q+1)2 = j(qd−1). Modulo q, this equation becomes 1 ≡ j mod q. At the same time, the Gerzon bound [17] on real ETFs states that n ≤ 1 2 d(d + 1). Thus, d2 > n and so d > n (q − 1)(q + 1)2 (q − 1)(q + 1)2 d = q, implying = q + 1. j = < qd − 1 q2 − 1 Since 0 < j < q + 1 and j ≡ 1 mod q, we have j = 1 and so (q − 1)(q + 1)2 = qd − 1. Solving for d then gives d = q2 + q − 1. Noting m = d + 1 and n = qd gives (28). To show q is even, recall the first quantity in (29) is odd; since d = q2 + q − 1 and n = qd, this quantity is q + 1. Here, we emphasize that the parameters m and n of the complex flat ETFs produced by Theorem 2 are identical to those given in (29). That is, when q = 2e, e ≥ 1, defining m and n by (29), Theorem 2 produces an m × n complex flat matrix whose columns form an ETF for 1⊥. We also note that in light of Theorem 3, Problem 1 is equivalent to: i=1 Problem 2. Given q even, does there exist a sequence {ϕi}q(q2+q−1) of {±1}-valued vectors that form an ETF for 1⊥ in Rq(q+1), and so have hϕi, ϕji = q for all i 6= j? The answer to Problem 2 is "yes" when q = 2, e.g. (4), and "no" when q = 4, since real ETFs with (d, n) = (19, 76) do not exist [1], [33]. To our knowledge, the q = 6 case is open. Here, (m, n) = (42, 246) and so computer-assisted searches might be impractical. Cases like this where n m = q ≡ 2 mod 4 are particularly interesting since they might be related to a fact from [16]: if there exist m×n real flat matrices whose columns form an ETF for the entire space Rm, and if these matrices arise from BIBD(v, k, 1) in the manner of [12], [16], then m ≈ k ≡ 2 mod 4. n B. Extending flat ETFs to larger ETFs We conclude this paper by discussing how we can append vectors to certain flat ETFs to produce other ETFs. The inspiration for this work was the realization that both the columns and the rows of the matrices discussed in Theorems 2 and 3 form ETFs for their respective spans, and that the Welch bounds for these "paired" ETFs are closely related. To be precise, if q is any positive integer and Φ is any q(q + 1) × q(q2 + q − 1) matrix whose columns form an ETF for 1⊥ in Fm, then Lemma 1 implies ΦΦ∗ = aΠ where a = q2(q + 1) and Π = I − 1 q(q+1) J is the orthogonal projection operator onto 1⊥. That is, ΦΦ∗ = q2(q + 1)I − qJ. 10 This means the columns of Φ∗ are equiangular, with the inner product of any two of them being −q. Remarkably, scaling the Welch bound of its columns by their norms, we find their inner products have the same magnitude: 1 1 2 2 q − 1 = q(q + 1)(cid:20) q3 + q2 − q − 1(cid:21) d(n − 1)(cid:21) m(cid:20) n − d Even more remarkably, we found that concatenating such Φ √q+2−1 with qI + q−1 J led to even larger ETFs. For example, for the 10-vector flat ETF for 1⊥ ⊆ R6 given in (4), we can append 6 vectors to it to form a 16-vector ETF for R6: = q.   3 + + + + + + + + + + 7 3 + + + + − − − − − − 1 + − − − + + + − − − 1 − + − − + − − + + − 1 − − + − − + − + − + 1 − − − + − − + − + + 1 3 3 3 3 1 3 7 3 1 3 1 3 1 3 1 3 1 3 1 3 7 3 1 3 1 3 1 3 1 3 1 3 1 3 7 3 1 3 1 3 1 3 1 3 1 3 1 3 7 3 1 3 1 3 1 3 1 3 1 3 1 3 7 3   . In general, we do not know whether these ETFs are equivalent to other known families of ETFs with these same parameters, such as harmonic ETFs arising from certain McFarland dif- ference sets [9], Steiner ETFs arising from affine planes [12], or those instances of (21) in Theorem 1 arising from a known hyperoval in a projective plane of order q = 2e. One reason such a construction is possible is that the columns of Φ are orthogonal to those of J, and so only interact with the "I" component of the columns of matrices of the form f I + gJ. Generalizing this construction gives the following result: Theorem 4. Suppose Φ is an m × n non-square matrix with unimodular entries and that the columns of Φ and Φ∗ both form ETFs for their d-dimensional spans with 1 d(n − 1)(cid:21) n(cid:20) n − d 1 2 = 1 d(m − 1)(cid:21) m(cid:20) m − d 1 2 . (31) Then there exist real scalars f, g such that the columns of Ψ = (cid:2) Φ f ΦΦ∗ + gI (cid:3) form an (m + n)-vector ETF for Fm, namely f = − g = (m + n) (m + n) 1 2 . 1 (m + n − 1) 2 (m + n − 1) 1 2 , 1 2 ± (mn) 1 2 (32) Proof: We begin by expressing d in terms of m and n. Squaring (31) and simplifying gives = m2(m − 1) − n2(n − 1). mn[m(m − 1) − n(n − 1)] Since Φ is not square by assumption, we can divide both sides of this equation by m − n to obtain 1 d mn(m + n − 1) = (m + n)(m + n − 1) − mn. That is, d = rank(Φ) = rank(Φ∗) necessarily satisfies 1 d 1 d = 1 m + 1 n − 1 m + n − 1 . (33) This fact in hand, note that since the columns {ϕi}n i=1 of Φ form an ETF for their span with kϕik2 = m for all i, Lemma 1 gives ΦΦ∗Φ = aΦ where 1 m mn d (34) a = mn 1 n + = + . Taking adjoints gives Φ∗ΦΦ∗ = aΦ∗ as well. As such, the frame operator of Ψ is m + n − 1 ΨΨ∗ = ΦΦ∗ + (f ΦΦ∗ + gI)(f ΦΦ∗ + gI)∗ = (af 2 + 2f g + 1)ΦΦ∗ + g2I. Since the f and g given in (32) satisfy af 2 + 2f g + 1 = 0, (35) we have ΨΨ∗ = (m + n)I. Thus, the columns of Ψ form a tight frame for Fm. Next, the Gram matrix of Ψ is Ψ∗Ψ = (cid:20) = (cid:20) = (cid:20) Φ∗ Φ∗Φ f ΦΦ∗ + gI (cid:21)(cid:2) Φ f ΦΦ∗ + gI (cid:3) (af + g)Φ (af 2 + 2f g)ΦΦ∗ + g2I (cid:21) (af + g)Φ g2I − ΦΦ∗ (cid:21) . (af + g)Φ∗ (af + g)Φ∗ Φ∗Φ (36) Since Φ is an m × n matrix with unimodular entries and g2 = m + n, all of the diagonal entries of g2I − ΦΦ∗ have value (m+n)−n = m, which equals the value of the diagonal entries of Φ∗Φ. Thus, the columns of Ψ form an equal norm tight frame for Fm. To show they form an ETF, note we can rewrite (31) as 1 2 d(n − 1)(cid:21) m(cid:20) n − d = n(cid:20) m − d d(m − 1)(cid:21) 1 2 , (37) namely that the off-diagonal entries of Φ∗Φ and ΦΦ∗ have the same modulus. We further note that the values in (37) are equal to af + g. Indeed, (35), (32) and (34) give mn (af + g)2 = a(af 2 + 2f g) + g2 = , m + n − 1 which is the same value obtained by substituting (33) into the squares of the quantities in (37), e.g., m2 n − d d(n − 1) = = m2 m n − 1(cid:16) n m + n − 1 mn + 1 − . n m + n − 1 − 1(cid:17) Recalling Φ has unimodular entries, this means all the off- diagonal entries of (36) have the same modulus, and so the columns of Ψ form an ETF for Fm. In the special case where d = m− 1, (33) reduces to having n2 + (m− 1)n− m(m− 1)2 = 0, whose only positive solution is n = 1 2 − 1]. Here, since 4m + 1 is an odd square, it can be written as 2 (m − 1)[(4m + 1) 1 4m + 1 = (2q + 1)2 = 4q2 + 4q + 1 for some integer q, meaning m = q(q +1), which in turn gives n = 1 2 (q2 + q − 1)[(2q + 1) − 1] = q(q2 + q − 1). For this choice of m and n, if there exists an m × n flat matrix with unimodular entries whose columns form an ETF for 1⊥, then 11 Theorem 4 produces an ETF of q2(q + 2) vectors for Fq(q+1). In the real-variable setting, this leads to the following result, which is an avenue for future research on the nexus of ETFs and supersaturated designs: Corollary 2. If there exists a solution to Problem 2 for a given q, then there exists a real ETF of q2(q + 2) vectors for Fq(q+1). In particular, when q = 6, if there exists a {±1}-valued matrix of size 42 × 246 whose columns form an ETF for 1⊥, then there is an ETF of 288 vectors in R42, resolving an open problem in the theory of strongly regular graphs. Interestingly, there are choices of (m, n, d) that satisfy (31) but do not have d = m − 1, like (d, m, n) = (8, 10, 15). So far, we have been unable to find an ETF with these parameters that meets the hypotheses of Theorem 4. We leave a deeper investigation of such ETFs with d < m−1 for future research. We finish with a tantalizing connection between Theorem 4 and difference sets. To be precise, for a finite abelian group G of order v, a k-element subset D of G is a difference set of G if the cardinality of {(δ, ε) ∈ G × G : γ = δ − ε} is constant over all γ 6= 0. Letting H denote the v × v character table of G, it is well-known that the columns of a k × v submatrix of H form an ETF for Fk if and only if the k rows correspond to a difference set of G [32], [9]. Now consider a k × k′ submatrix Φ of H corresponding to two subsets D and D′ of G which indicate rows and columns, respectively. If both D and D′ are difference sets of G, then Φ is a matrix with unimodular entries and equiangular rows and columns. It is possible that Φ has rank d where d satisfies (31). For example, 2, numerical experimentation reveals we can take when G = Z4 D to be a McFarland difference set, and take D′ to be the complement of a distinct McFarland difference set, such as: D = {0000, 0010, 1000, 1001, 1100, 1111}, D′ = {0000, 0001, 0010, 0100, 1000, 1001, 1011, 1100, 1110, 1111}. The resulting 6 × 10 matrix has unimodular entries and its columns and rows form an ETF for their 5-dimensional spans in R6 and R10 respectfully. Remarkably, numerical experimen- tation reveals other such "paired" difference sets exist when 4, but not when G = Z2 × Z8 or G = Z2 × Z2 × Z4, G = Z2 despite the fact that difference sets of cardinality 6 and 10 exist in them all. That is, the existence of these sets seems very sensitive to the group structure of G itself. We summarize this train of thought with the following open problem: Problem 3. Given a finite abelian group G, find all pairs of difference sets D, D′ of G so that the corresponding submatrix Φ of the character table of G satisfies the hypotheses of Theorem 4, namely when d = rank(Φ), m = D and n = D′ satisfy (31). ACKNOWLEDGMENTS This work was partially supported by NSF DMS 1321779, AFOSR F4FGA05076J002 and an AFOSR Young Investigator Research Program award. The views expressed in this article are those of the authors and do not reflect the official policy or position of the United States Air Force, Department of Defense, or the U.S. Government. REFERENCES [1] J. Azarija, T. Marc, There is no (75,32,10,16) strongly regular graph, preprint, arXiv:1509.05933. [2] W. U. Bajwa, R. Calderbank, D. G. Mixon, Two are better than one: fundamental parameters of frame coherence, Appl. Comput. Harmon. Anal. 33 (2012) 58-78. 12 [32] P. Xia, S. Zhou, G. B. Giannakis, Achieving the Welch bound with difference sets, IEEE Trans. Inform. Theory 51 (2005) 1900–1907. [33] W.-H. Yu, There are no 76 equiangular lines in R19, preprint, arXiv:1511.08569. [34] G. Zauner, Quantum designs: Foundations of a noncommutative design theory, PhD thesis, University of Vienna, 1999. [3] A. S. Bandeira, M. Fickus, D. G. Mixon, P. Wong, The road to deterministic matrices with the Restricted Isometry Property, J. Fourier Anal. Appl. 19 (2013) 1123–1149. [4] B. G. Bodmann, J. Hass, Achieving the orthoplex bound and con- structing weighted complex projective 2-designs with Singer sets, submitted, arXiv:1509.05333. [5] A. E. Brouwer, Strongly regular graphs, in: C. J. Colbourn, J. H. Dinitz (Eds.), CRC Handbook of Combinatorial Designs (2007) 852-868. [6] A. E. Brouwer, Parameters of Strongly Regular Graphs, http://www.win.tue.nl/∼aeb/graphs/srg/ [7] D. Corneil, R. Mathon, eds., Geometry and combinatorics: Selected works of J. J. Seidel, Academic Press, 1991. [8] R. H. F. Denniston, Some maximal arcs in finite projective planes, J. Combin. Theory 6 (1969) 317–319. [9] C. Ding, T. Feng, A generic construction of complex codebooks meeting the Welch bound, IEEE Trans. Inform. Theory 53 (2007) 4245–4250. [10] M. Fickus, J. Jasper, D. G. Mixon, J. Peterson, Tremain equiangular tight frames, arXiv:1602.03490 (2016). [11] M. Fickus, D. G. Mixon, Tables of the existence of equiangular tight frames, arXiv:1504.00253 (2015). [12] M. Fickus, D. G. Mixon, J. C. Tremain, Steiner equiangular tight frames, Linear Algebra Appl. 436 (2012) 1014–1027. [13] J. M. Goethals, J. J. Seidel, Strongly regular graphs derived from combinatorial designs, Can. J. Math. 22 (1970) 597–614. [14] T. Hansen, G. L. Mullen, Primitive polynomials over finite fields, Math. Comp. 59 (1992) 639–643. [15] R. B. Holmes, V. I. Paulsen, Optimal frames for erasures, Linear Algebra Appl. 377 (2004) 31–51. [16] J. Jasper, D. G. Mixon, M. Fickus, Kirkman equiangular tight frames and codes, IEEE Trans. Inform. Theory. 60 (2014) 170–181. [17] P. W. H. Lemmens, J. J. Seidel, Equiangular lines, J. Algebra 24 (1973) 494–512. [18] S. Li, H. Ogawa, Pseudoframes for subspaces with applications, J. Fourier Anal. Appl. 10 (2004) 409–431. [19] N. K. Nguyen, An algorithmic approach to constructing supersaturated designs, Technometrics 38 (1996) 69–73. [20] R. A. Rankin, On the minimal points of positive definite quadratic forms, Mathematika 3 (1956) 15–24. [21] J. M. Renes, Equiangular tight frames from Paley tournaments, Linear Algebra Appl. 426 (2007) 497–501. [22] J. M. Renes, R. Blume-Kohout, A. J. Scott, C. M. Caves, Symmetric informationally complete quantum measurements, J. Math. Phys. 45 (2004) 2171–2180. [23] K. J. Ryan, D. A. Bulutoglu, E(s2)-optimal supersaturated designs with good minimax properties, J. Statist. Plann. Inference 137 (2007) 2250–2262. [24] T. Strohmer, A note on equiangular tight frames, Linear Algebra Appl. 429 (2008) 326-330. [25] T. Strohmer, R. W. Heath, Grassmannian frames with applications to coding and communication, Appl. Comput. Harmon. Anal. 14 (2003) 257–275. [26] M. A. Sustik, J. A. Tropp, I. S. Dhillon, R. W. Heath, On the existence of equiangular tight frames, Linear Algebra Appl. 426 (2007) 619– 635. Matthew Fickus (M'08) received a Ph.D. in Mathematics from the University of Maryland in 2001. In 2004, he joined the Department of Mathematics and Statistics at the Air Force Institute of Technology, where he is currently a Professor of Mathematics. His research focuses on applying harmonic analysis and combinatorial design to problems of signal and image processing. Dustin G. Mixon received a Ph.D. in Applied and Computational Mathemat- ics from Princeton University in 2012. He is currently an Assistant Professor of Mathematics in the Department of Mathematics and Statistics at the Air Force Institute of Technology. His research interests include frame theory, compressed sensing, signal and image processing, and machine learning. John Jasper received a Ph.D. in Mathematics from the University of Oregon in 2011. He is currently a Visiting Assistant Professor in the Department of Mathematical Sciences at the University of Cincinnati. His research focuses on operator theory and combinatorial design. [27] B. Tang, C. F. J. Wu, A method for constructing supersaturated designs and its Es2 optimality, Canad. J. Statist. 25 (1997) 191–201. [28] R. J. Turyn, Character sums and difference sets, Pacific J. Math. 15 (1965), 319–346. [29] S. Waldron, On the construction of equiangular frames from graphs, Linear Algebra Appl. 431 (2009) 2228–2242. [30] W. P. Wardlaw, Matrix representation of finite fields, Math. Mag. 67 (1994) 289–293. [31] L. R. Welch, Lower bounds on the maximum cross correlation of signals, IEEE Trans. Inform. Theory 20 (1974) 397-399.
1109.0305	3	1109	2012-05-17T14:57:45	Compactness characterization of operators in the Toeplitz algebra of the Fock space $F_\alpha ^p$	[ "math.FA", "math.CV" ]	For $1 < p < \infty$ let $\mathcal{T}_p ^\alpha$ be the norm closure of the algebra generated by Toeplitz operators with bounded symbols acting on the standard weighted Fock space $F_\alpha ^p$. In this paper, we will show that an operator $A$ is compact on $F_\alpha ^p$ if and only if $A \in \mathcal{T}_p ^\alpha$ and the Berezin transform $B_\alpha (A)$ of $A$ vanishes at infinity.	math.FA	math	Compactness characterization of operators in the Toeplitz algebra of the Fock space F p α. Wolfram Bauer and Joshua Isralowitz1 Mathematisches Institut Georg-August Universitat Gottingen Bunsenstrasse 3-5 D-37073, Gottingen Germany Abstract For 1 < p < ∞ let T α p be the norm closure of the algebra generated by Toeplitz operators with bounded symbols acting on the standard weighted Fock space F p α. In this paper, we will show that an operator A is compact on F p p and the Berezin transform Bα(A) of A vanishes at infinity. α if and only if A ∈ T α Keywords: Toeplitz algebra, Toeplitz operators, Fock spaces 2010 MSC: 47B35, 47B38, 47L80 1. Introduction and preliminaries For any α > 0, consider the Gaussian measure π(cid:17)n dµα(z) :=(cid:16)α e−αz2 dv(z), where dv denotes the usual Lebesgue measure on Cn ∼= R2n. Let 1 ≤ p < ∞, and write Lp α for the space of (equivalence classes) of measurable complex Email address: [email protected], [email protected] (Wolfram Bauer and Joshua Isralowitz ) 1Both authors were supported by an Emmy-Noether grant of Deutsche Forschungsge- meinschaft Preprint submitted to Elsevier October 26, 2018 valued function f on Cn such that kfkα,p :=(cid:20)(cid:16) pα 2π(cid:17)nZCn(cid:12)(cid:12)(cid:12)f (z)e−αz2/2(cid:12)(cid:12)(cid:12) p dv(z)(cid:21)1/p < ∞. (1.1) If H(Cn) denotes the space of all entire functions on Cn, then the Fock space F p α is the Banach space defined by F p α with the norm k · kα,p (cf. [17]). Recall that F 2 α is a Hilbert space with the natural inner product h·,·iα induced by (1.1) and is sometimes called the Segal-Bargmann space, cf. [1]. In the case of p = ∞, we define the Banach space F ∞ α := H(Cn) ∩ Lp α by F ∞ α :=nf ∈ H(Cn) : kfkα,∞ := kf e− α 2 ·2 k∞ < ∞o. Let Pα be the orthogonal projection from L2 α onto F 2 α given by Pαf (z) =ZCn eα(z·w)f (w) dµα(w). It is well known [17] that as an integral operator, Pα is a bounded projection If g ∈ L∞, then we can define the from Lp bounded Toeplitz operator T α α for all 1 ≤ p ≤ ∞. g on F p α by the formula α onto F p T α g := PαMg where Mg is "multiplication by g." Now let be the reproducing kernel of F 2 reproducing kernel given by K α(w, z) = eα(w·z) α and let kα z be the corresponding normalized z (w) = eα(w·z)− α kα 2 z2 . Given any bounded operator A on F p Berezin transform of A defined by α for 1 ≤ p < ∞, let Bα(A) be the Bα(A)(z) = hAkα z , kα z iα. (1.2) z kα,p = 1 for all z ∈ Cn and 1 ≤ p < ∞, it is easy to see Bα(A) is Since kkα well defined and in fact bounded. Furthermore, it is well known and easy to show that the map A 7→ Bα(A) is one-to-one if A is bounded on F p α (see [10], 2 p. 42 for a proof of the special case α = 1 and p = 2 that easily extends to general case α > 0 and 1 < p < ∞.) Moreover, for a "nice enough" function f (for example, f ∈ L∞), we define the Berezin transform of f to be z iα z , kα Bα(f )(z) := hf kα π(cid:17)nZCn =(cid:16)α f (w)e−αz−w2 dv(w). Note that an easy application of Fubini's theorem gives us that Bα(T α f ) = Bα(f ). Also note that when f is positive and measurable, we can define the (possibility infinite) function Bα(f ) without any other assumptions on f . z = 0 weakly on F p kα α (see [4] for examples on F 2 Information regarding the operator A can be often described in terms of properties of the function Bα(A), and this point of view has been especially successful when dealing with the boundedness, compactness, and Schatten α if 1 < p < ∞, so class membership of A. Note that lim z→∞ that Bα(A) vanishes at infinity if A is compact. Unfortunately, the converse in general is not true for bounded operators on the Fock space F p α, and is even not true for certain Toeplitz operators on F p α). However, it was proven in [8] that A is compact on F 2 α if and only if Bα(A) : f ∈ L∞} vanishes at infinity when A is in the algebra generated by {T α (that is, A is the finite sum of finite products of Toeplitz operators T α f with bounded symbols f .) Moreover, it was proved in [21] that if 1 < p < ∞ and A is any bounded operator on the standard unweighted Bergman space a(Bn, dv) of the unit ball Bn, then A is compact if and only if A is in the norm Lp closure of the algebra generated by Toeplitz operators with bounded symbols a(Bn, dv) vanishes on and the Berezin transform B(A) of A associated to L2 the boundary ∂Bn (see also [19] where the results of [21] are extended to the a(Bn, dvγ) with standard weights dvγ for γ > −1.) weighted Bergman space Lp In this paper, we will show that this result also holds for the Fock space F p α when 1 < p < ∞. In particular, if T α is the norm closure of the algebra generated by Toelitz operators with bounded symbols acting on F p α, then we will prove the following, which is the main result of this paper: Theorem 1.1. If 1 < p < ∞ and A is a bounded operator on F p compact if and only if A ∈ T α p and Bα(A) vanishes at infinity. α, then A is f p Moreover, in Section 3, we will show that T α p is in fact the closed algebra generated by Toeplitz operators T α ν where ν is a complex Borel measure on Cn such that the total variation measure ν is Fock-Carleson, which greatly 3 widens the scope of Theorem 1.1 (see Section 2 for a discussion of Fock- Carleson measures and Toeplitz operators with measure symbols.) Note that a version of this result was proven in [21] for the unweighted Bergman space of the ball and was proven in [19] for the weighted Bergman space of the ball. The basic strategy used to prove Theorem 1.1 is similar to the strategy used in [21] to prove the corresponding result for the Bergman space, and in particular relies on obtaining quantitative estimates for the essential norm kAke of operators A ∈ T α p for 1 < p < ∞. However, the details of the proofs involved in implementing this strategy will often be considerably different than details found in [21]. Note that this is often the case when one is trying to prove a result for the Fock space that is already known to be true for the Bergman space (see [3, 8] for example). We now give a short outline of the rest of the paper. Sections 2 − 5 will consist of preliminary lemmas that will be used to prove Theorem 1.1 and Section 6 will contain a proof of Theorem 1.1. More specifically, Section 2 will discuss Fock-Carleson measures and prove an important lemma (Lemma 2.6) that will be used in Section 3. Section 3 will discuss various approxima- tion results that will be needed to prove Theorem 1.1. Section 4 will prove two important lemmas (Lemmas 4.1 and 4.3) related to sampling and inter- polation in Fock spaces. Section 5 will introduce a useful uniform algebra A and will extend the Berezin transform and other related objects that are defined on Cn to the maximal ideal space MA of A. Section 6 will tie all these ideas and results together to prove Theorem 1.1, and finally in Section 7 we will discuss improvements to our results when p = 2 that are similar to the ones in [21], and we will briefly discuss some open problems. α and since A ∈ T α p α)∗ = F q if and only if A∗ ∈ T α We will close this introduction with a short comment about the proof of Theorem 1.1 and Section 6. Since (F p α under the natural pairing induced by F 2 q where q is the dual exponent of p, it is easy to see that we only need to prove Theorem 1.1 for 2 ≤ p < ∞. More generally, if A ∈ T α p , then it will be seen later (using the two above mentioned facts) that many of the necessary estimates for kAke when A ∈ T α p only need to be obtained for the case 2 ≤ p < ∞. This simple observation will be crucial to the proof of Theorem 1.1 since many of the needed preliminary estimates are only directly obtainable for p = 2 and (in much weaker form) for p = ∞, and will subsequently follow for all 2 ≤ p < ∞ either by duality or by complex interpolation. Along these lines, we will often use the following consequence of complex interpolation in Fock spaces (see [17]): 4 Lemma 1.2. Let 2 < p < ∞. If A is an operator A : F 2 such that A maps F 2 A : F p α boundedly and maps F ∞ α boundedly. More precisely, we have that α to L2 α to L∞ α + F ∞ α → L2 α + L∞ α α boundedly, then α → Lp kAkF p α→Lp α ≤ kAk 2 p F 2 α→L2 αkAk 1− 2 p F ∞ α →L∞ α . 2. Fock-Carleson measures and related operators Let 1 < p < ∞. A positive Borel measure ν on Cn will be called a Fock-Carleson measure if dν(z) ≤ Ckfkp α,p p ZCn(cid:12)(cid:12)(cid:12)f (z)e−αz2/2(cid:12)(cid:12)(cid:12) ν f (z) =ZCn α by for all f ∈ F p ν : F p operator T α α with C independent of f . In this case, define the Toeplitz α → F p T α f (w)eα(z·w)−αw2 dν(w). For any r > 0, and z ∈ Cn, let B(z, r) be a Euclidean ball centered at z with radius r. If ν is a Fock-Carleson measure, then we let ıν be the canonical imbedding from F p α(dν) is the space of ν measurable functions f where α(dν) where Lp α into Lp p ZCn(cid:12)(cid:12)(cid:12)f (z)e−αz2/2(cid:12)(cid:12)(cid:12) dν(z) < ∞. It turns out that the property of ν being a Fock-Carleson measure is inde- pendent of both p and α, as the following result in [14] shows: Lemma 2.1. For any 1 < p < ∞ and any α, r > 0, the following quantities are equivalent, where the constants of equivalence only depend on p, n, α and r: 2 z−w2 dν(w), z∈CnRCn e− α (a) kνk∗ := sup (b) kıνkp α(dν), F p α→Lp ν(B(z, r)), (c) sup z∈Cn ν kp (d) kT α α→F p F p α . 5 Before we state and prove the main result of this section (Lemma 2.6), we will need some preliminary results related to Fock-Carleson measures. Throughout the paper, we will let C denote a constant that might change from estimate to estimate, or even from line to line in a single estimate. We will indicate the parameters that C depends on only when it is important to do so. Lemma 2.2. If ν is a Fock-Carleson measure and F ⊂ Cn is compact, then there exists C > 0 independent of f, F, and ν such that kT α χF νfkF 2 α→F 2 α ≤ CkıνkF 2 α→L2 α(dν)(cid:18)ZF f (z)2e−αz2 2 dν(z)(cid:19) 1 where χF is the characteristic function of F . Proof. Let g ∈ F 2 ν is a Fock-Carleson measure, we have that α with kgkα,2 = 1. Using Fubini's theorem and the fact that π(cid:17)n(cid:12)(cid:12)(cid:12)(cid:12)ZCn χF νf, giα =(cid:16) α hT α π(cid:17)n ≤(cid:16)α kıνkF 2 α→L2 χF (z)f (z)g(z)e−αz2 dν(z)(cid:12)(cid:12)(cid:12)(cid:12) dν(z)(cid:19) 1 α(dν)(cid:18)ZF f (z)2e−αz2 2 . Lemma 2.3. For any α > 0 and s real, we have that ZCn(cid:12)(cid:12)es(z·w)(cid:12)(cid:12) dµα(w) = es2z2/4α. Proof. See [7]. Lemma 2.4. For any r, α, p > 0, any z ∈ Cn, and any entire f , there exists C independent of f and z where p (cid:12)(cid:12)(cid:12)f (z)e− α 2 z2(cid:12)(cid:12)(cid:12) Proof. See [16]. ≤ CZB(z,r)(cid:12)(cid:12)(cid:12)f (w)e− α 2 w2(cid:12)(cid:12)(cid:12) p dv(w). For the rest of the paper, we will canonically treat Z2n as a lattice in Cn. The following is the main technical result that is needed to prove the Lemma 2.6. 6 Lemma 2.5. Let ν be a Fock-Carleson measure and let Fj, Kj ⊂ Cn be Borel sets where {Fj} are pairwise disjoint and d(Fj, Kj) > δ ≥ 1 for each j. Then [χFj (z)χKj (w)]eα(z·w)e− α 2 w2 dν(w) ≤ o(δ)kνk∗e α 2 z2 ZCnXj where o(δ) only depends on δ (and α, n) and lim δ→∞ o(δ) = 0. Proof. Clearly Kj ⊆ Cn\B(z, δ) if z ∈ Fj, which means that χFj (z)χCn\B(z,δ)(w). Thus, where χFj (z)χKj (w) ≤Xj Xj ZCnXj [χFj (z)χKj (w)]eα(z·w)e− α χFj (z)ZCn\B(z,δ) eα(z·w)e− α ≤Xj =Xj Jz =ZB(z,δ)c eα(z·w)e− α χFj (z)Jz 2 w2 dν(w) 2 w2 dν(w) (2.1) 2 w2 dν(w) and B(z, δ)c = Cn\B(z, δ). We will now estimate Jz using Lemma 2.1 and Lemma 2.4 with respect to the entire function w 7→ eα(w·z) : 7 B(σ, 1 10 (2n)−1/2Z2n 10 )∩B(z,δ)c6=∅ Jz ≤ Xσ∈ 1 ≤ C Xσ∈ 1 ≤ C Xσ∈ 1 ≤ Ckνk∗ Xσ∈ 1 10 (2n)−1/2Z2n 10 )∩B(z,δ)c6=∅ 10 (2n)−1/2Z2n 10 )∩B(z,δ)c6=∅ B(σ, 1 B(σ, 1 2 w2 dν(w) ZB(σ, 1 10 ) eα(z·w)e− α 10 )ZB(w, 1 ZB(σ, 1 10 ) eα(z·u)e− α ))ZB(σ, 1 5 ) eα(z·u)e− α ZB(σ, 1 ν(B(σ, 1 10 2 u2 10 (2n)−1/2Z2n 10 )∩B(z,δ)c6=∅ B(σ, 1 5 ) eα(z·u)e− α 2 u2 dv(u) dv(u). 2 u2 dv(u)dν(w) Since δ ≥ 1 we have that B(σ, 1 10 )\B(z, δ) 6= ∅ =⇒ B(σ, 1 5 ) ∩ B(z, δ 2 ) = ∅ and since there exists M > 0 such that every z ∈ Cn belongs to at most M of the sets B(σ, 1 5), we get that Jz ≤ Ckνk∗ZCn\B(z, δ = Ckνk∗ZCn\B(0, δ = Ckνk∗e α 2 z2ZCn\B(0, δ 2 ) dv(u) 2 u2 2 ) eα(z·u)e− α 2 ) eα(z·(z−u))e− α 2 z−u2 dv(u) e− α 2 u2 dv(u). Finally, since the sets Fj are pairwise disjoint, we get that where (2.1) ≤ Ckνk∗e ≤ Ckνk∗e α 2 z2 α 2 z2 o(δ)Xj o(δ) χFj (z) o(δ) :=ZCn\B(0, δ 2 ) e− α 2 u2 dv(u). 8 Lemma 2.6. Suppose that ν is a Fock-Carleson measure and that Fj, Kj ⊂ Cn. Moreover, assume that aj ∈ L∞(dv) and bj ∈ L∞(dν) both with norms ≤ 1, and assume that (a) d(Fj, Kj) > δ ≥ 1, (b) supp aj ⊆ Fj and supp bj ⊆ Kj, (c) every z ∈ Cn belongs to at most N ∈ N of the sets Fj. Then for 2 ≤ p < ∞, we have thatPj Maj T α and ν Mbj is bounded from F p α to Lp α Maj T α ν MbjkF p α→Lp α ≤ Nkνk∗o(δ) (2.2) kXj where o(δ) only depends on δ (and p, α, n) and lim δ→∞ every f ∈ F p α with norm 1, we have that ν Mbj fkp kMaj T α α,p ≤ Nkνkp ∗o(δ)p. o(δ) = 0. Moreover, for (2.3) Proof. We will first prove the lemma for the special case N = 1. Note that Xj ν Mbj f(cid:1) (z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤Xj e− α 2 z2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xj (cid:0)Maj T α aj(z)ZCn bj(w)f (w)e− α 2 w2 e− α 2 z−w2 dν(w) ≤ kνk∗kfkα,∞. Thus, by Lemma 1.2, the lemma will be proved for the special case N = 1 if we can show that (2.2) holds (which is equivalent to (2.3) when N = 1) when p = 2. Moreover, by Lemma 2.1, it is enough to prove that and kXj Xj Maj T α ν MbjkL2 1 2 ∗ o(δ) α(dν)→L2 α ≤ kνk kMaj T α ν Mbj fk2 α,2 ≤ kνk∗o(δ)2 9 (2.2') (2.3') for every f ∈ L2 α(dν) of norm ≤ 1. Let so that χFj (z)χKj (w)eα(z·w) aj(z)ZCn Φ(z, w)f (w)e−αw2 Φ(z, w) =Xj =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xj (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xj (cid:0)Maj T α ν Mbj f(cid:1) (z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ZCn ZCn Φ(z, w)h(z)2dµα(z) ≤ CZCn eα(z·w)e− α 4 z2, then Lemma 2.5 tells us that Φ(z, w)h(w)2e−αw2 ZCn Moreover, Lemma 2.3 tells us that If h(z) = e α bj(w)f (w)eα(z·w)e−αw2 dν(w). dν(w)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) dν(w) ≤ o(δ)kνk∗h(z)2. 2 z2 dv(z) = Ch(w)2. The Schur test now proves (2.2′), (2.3′), and consequently proves the lemma for the special case N = 1. As in [21], the general case N > 1 follows easily from the special case j}N j = {z ∈ j := ∅ if i > card Λ(z)) so that when N = 1 by writing {Fj} as the union of the family of sets {Ai where Λ(z) = {j : z ∈ Fj}, ordered in the natural way, and Ai Fj : j is the ith element of Λ(z)} (where Ai j ∩ Ai Ai k = ∅ for any j 6= k. i=1 3. Approximation results for Fock space operators In this section, we will prove various approximation results that will be needed for the proof of Theorem 1.1. The first such result, Lemma 3.3, will allow us to approximate operators of the form ST α ν by simple sums of "truncations" of ST α p and ν is a Fock-Carleson measure. For convenience, we will use the canonical identification Cn ∼= R2n and we will use the norm z∞ = max{z1, . . . ,z2n}. For some δ > 0, enumerate ν , where here S ∈ T α 10 the disjoint family of sets {[−δ, δ)2n + σ}σ∈2δZ2n as {Bj}∞ j=1 and let Ωδ(Bj) = {z ∈ Cn : dist∞(z, Bj) ≤ δ} where dist∞(z, Bj) is the distance between z and Bj in the · ∞ norm. The following result follows easily from the above definitions: Lemma 3.1. For any δ > 0, the Borel sets Bj ⊂ Cn above satisfy the followig conditions: (a) Bj ∩ Bk = ∅ if j 6= k, (b) Every z ∈ Cn belongs to at most 42n of the sets Ωδ(Bj), (c) diam(Bj) ≤ 2δ√2n where diam(Bj) is the Euclidean diameter of Bj. Now let δ > 0 and k be a non-negative integer. Let {Bj}∞ j=1 be a covering of Cn satisfying the conditions of the above lemma for (k + 1)δ instead of δ. For 1 ≤ i ≤ k and j ≥ 1, write F0,j = Bj, and Fi+1,j = {z ∈ Cn : dist∞(z, Fi,j) ≤ δ}. The next result is now easy to prove. Lemma 3.2. Let δ > 0 and k be a non-negative integer. For each 1 ≤ i ≤ k + 1 the family F i = {Fi,j : j ≥ 1} forms a covering of Cn such that (a) F0,j1 ∩ F0,j2 = ∅ if j1 6= j2, (b) F0,j ⊂ F1,j ⊂ · · · ⊂ Fk+1,j for all j ≥ 1, (c) dist∞(Fi,j, F c i+1,j) ≥ δ for all 0 ≤ i ≤ k and j ≥ 1, (d) Every point belongs to at most 42n elements of F i, (e) diam(Fi,j) ≤ δ(3k + 1)√2n for each i and j. Lemma 3.3. Let 2 ≤ p < ∞, S ∈ T α let ǫ > 0. Then there are Borel sets Fj ⊂ Gj ⊂ Cn with j ≥ 1 such that: p , ν be a Fock-Carleson measure, and (a) Cn =Sj≥1 Fj, (b) Fj ∩ Fk = ∅ if j 6= k, (c) each point of Cn belongs to at most 42n of the sets Gj, 11 (d) diam Gj ≤ d = d(S, ǫ) and ν −Xj kST α MχFj ST α χGj νkF p α→Lp α ≤ ǫ. Proof. The proof is a combination of Lemmas 2.6 and 3.2 and Lemmas 4.1 and 4.2 of [21], and is identical to the proof of Theorem 4.3 in [21]. In particular, Fj := F0,j and Gj := Fk+1,j where {Fi,j} are the sets that come from Lemma 3.2 with δ = δ(S, ǫ) set large enough to invoke the conclusion of Lemma 2.6. p We will now show that a Toeplitz operator T α ν where the total variation measure ν of ν is Fock-Carleson measure can be approximated in the F p α f with f ∈ C ∞ operator norm for any 1 < p < ∞ by Toeplitz operators T α b , is the space of C valued smooth where as stated in the introduction, C ∞ b functions where f and all of its derivatives are bounded. In particular, this will imply that T α for any 1 < p < ∞ is the closed algebra generated by {T α ν : ν is Fock-Carleson}. First, however, we will need some preliminary definitions and results. If ν is a complex Borel measure on Cn where ν is Fock-Carleson, then define the "heat transform"eν(t) of ν at "time" t > 0 to be If f is a function on Cn such that f dv is Fock-Carleson, then define ef (t) :=eν(t) where dν := f dv. A simple computation using Lemma 2.1, Fubini's theorem, and the reproducing property gives us that (4πt)nZCn eν(t)(z) := e− w−z2 dν(w). 1 4t α(cid:17)n ν ) =(cid:16) π 4α ). Bα(T α Similarly, one can easily show that the semi-group property eν( 1 =eν(s+t) holds for s, t > 0. Since Lemma 2.1 says that fν it follows easily from the semi-group property that eν(t) is smooth and all of its derivatives are bounded. is bounded for all t > 0, {eν(s)} (t) (t) 12 Now for any z ∈ Cn and any complex Borel measure ν, let νz be the complex Borel measure defined by νz(E) := ν(z − E) for any Borel set E ⊂ Cn. Note that ZCn f (z − w) dν(w) =ZCn f (w) dνz(w) for any z ∈ Cn and f where f (z − ·) ∈ L1(Cn, dν). Lemma 3.4. If ν is a complex Borel measure such that ν is Fock-Carleson and νβ :=eν( 1 β ) dv − ν, then lim β→∞ z∈Cn kBα(T α sup (νβ)z )k∞ = 0 Proof. From the discussion above, it is enough to show that lim β→∞ sup z∈Cn k](νβ)z ( 1 4α ) k∞ = 0. ( 1 8α ) and note that Lemma 2.1 gives us that supz∈Cn kGzk∞ < C for some C > 0. Also note that (νβ)z = (νz)β To that end, let Gz := g(νz) for any z ∈ Cn and β > 0. w ∈ Cn ](νβ)z ( 1 β + 1 ( 1 4α ) 8α ) (cid:12)(cid:12)(cid:12)(cid:12) Using the semi-group property and the above equality, we have that for (w)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)fGz πn(cid:12)(cid:12)(cid:12)(cid:12)ZCn ≤ kGzk∞ πn = 1 ( 1 8α ) (w)(cid:12)(cid:12)(cid:12)(cid:12) (w) −fGz β + 8α(cid:19)n Gz(u)(cid:20)(cid:18) 2βα ZCn(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) 2βα β + 8α(cid:19)n e− 2βα β+8α w−u2 e− 2βα β+8α u2 − (2α)ne−2αw−u2(cid:21) dv(u)(cid:12)(cid:12)(cid:12)(cid:12) − (2α)ne−2αu2(cid:12)(cid:12)(cid:12)(cid:12) dv(u). The result now follows immediately by an application of the dominated con- vergence theorem. The proof of the following lemma is a slight variation of the proof of Lemma 3.4 in [20]. 13 Lemma 3.5. If ν is a complex Borel measure such that ν is Fock-Carleson then z∈Cn T α sup lim β→∞ (νβ)z 1(w) = 0 where the convergence is pointwise for any w ∈ Cn. Proof. First note that Lemma 2.1 and the semi-group property tells us that T α α norm with respect to both β and z ∈ Cn. Thus, by the reproducing property and an easy approximation argument, it is enough to show that is uniformly bounded in the F 2 (νβ )z lim β→∞ z∈Cn hT α sup for each fixed multiindex k ∈ Nn 0 . into the definition of the Berezin transform, we have that To that end, writing kα w(u) = e− α (νβ)z 1, ukiα = 0 2 w2Pγ(γ!)−1αγuγwγ and plugging this Bα(T α (νβ)z )(w) = e−αw2Xγ,γ′ αγαγ′ γ!γ′! hT α (νβ )z uγ, uγ′ iαwγwγ′ . Then for any fixed multiindex k and any 0 < r < 1 we have (νβ )z uγ, uγ′ wk+γwγ′ dv(w) (νβ )z )(w)wk dv(w) iαZB(0,r) (νβ )z 1, ukiα + eαw2 Bα(T α αγαγ′ γ!γ′! hT α ZB(0,r) =Xγ,γ′ = r2n+2k n!αk (νβ)z 1, ukiα(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)hT α Since T α (νβ )z (n + k)!hT α ∞Xγ=1 n!αγαk+γ (n + k + γ)!γ!hT α (νβ )z uγ, uγ+kiαr2γ . is uniformly bounded in the F 2 α operator norm, we have ≤ Cr−2n−2kkBα(T α (νβ )z )k∞ZB(0,r) eαw2wk dv(w) + C r2γ ∞Xγ=1 14 for some C > 0 independent of z and β. Lemma 3.4 then gives us that lim sup β→∞ sup z∈Cn(cid:12)(cid:12)(cid:12)hT α (νβ)z 1, ukiα(cid:12)(cid:12)(cid:12) ≤ C r2γ ∞Xγ=1 and letting r → 0+ completes the proof. Lemma 3.6. If ν is a complex Borel measure where ν is Fock-Carleson and νβ is defined as in Lemma 3.4, then there exists C > 0 independent of β such that Proof. Note that Lemma 2.1 says that sup z∈Cn(cid:12)(cid:12)(cid:12)T α (νβ )z 1(w)(cid:12)(cid:12)(cid:12) ≤ Ce α 4 w2 kι(νβ )zkF 2 α→L2 α(dν) < C for some C independent of β and z ∈ Cn where ινz is the canonical imbedding α(dνz). Thus, there exists C > 0 independent of z ∈ Cn such from F 2 that α into L2 νz1(w)(cid:12)(cid:12) ≤(cid:16)α (cid:12)(cid:12)T α π(cid:17)nZCn eα(w·u)e−αu2 ≤ CZCn eα(w·u)e−αu2 dv(u) = Ce α 4 w2 dνz(u) where the last equality comes from Lemma 2.3. Finally, we can now prove Theorem 3.7. If ν is a complex Borel measure where ν is Fock-Carleson and 1 < p < ∞, then lim α(cid:17)n β→∞k(cid:16) π T α ( 1 β eν ) − T α ν kF p α→F p α = 0. In particular, T α {T α ν : ν is Fock-Carleson}. p for any 1 < p < ∞ is the closed algebra generated by 15 Proof. As before, let νβ =eν( 1 T α A direct calculation shows that β ) dv − ν, so that νβ =(cid:16) π α(cid:17)n β kT α sup νβkF ∞ T α ( 1 β eν α →F ∞ α < ∞. ) − T α ν . Thus, by an easy duality argument and Lemma 1.2, it is enough to prove the theorem for p = 2. To that end, we will proceed in a manner that is similar to the proof of Theorem 1 in [6]. First note that νβ K α(·, u)(cid:17) (w)e−αu2 dv(u) νβ f, K α(·, w)iα f (w)(cid:16)T α νβ f (w) = hT α T α π(cid:17)nZCn =(cid:16) α 4 w2 and let Φβ(u, w) :=(cid:12)(cid:12)(cid:12)(cid:16)T α α Let h(w) := e change of variables we have that which means that T α We will now use the Schur test to complete the proof. νβ is an integral operator on F 2 νβ K α(·, u)(cid:17) (w). α with kernel(cid:16)T α νβ K α(·, u)(cid:17) (w)(cid:12)(cid:12)(cid:12) . By a simple so that ZCn Φβ(u, w)h(u)2 e−αu2 (νβ)u1(u − w) Φβ(u, w) = eα(w·u)T α dv(u) =ZCn eα(w·u)T α =ZCn eα(w·(u+w))T α 2 w2ZCn T α 2 w2ZCn ≤ Ce e− α 4 u2 = e α α (νβ)(u+w) (νβ)u1(u − w)e− α 2 u2 dv(u) (νβ)(u+w) 1(u)e− α 2 u+w2 dv(u) 1(u)e− α 2 u2 dv(u) dv(u) where C > 0 comes from Lemma 3.6. Thus, we have that π(cid:17)nZCn (cid:16)α Φβ(u, w)h(u)2 e−αu2 dv(u) ≤ Ch(w)2 16 for some C > 0 independent of w and β. Furthermore, an application of the Cauchy-Schwarz inequality gives us that dv(u) Φβ(w, u)h(u)2 e−αu2 ZCn =ZCn eα(u·w)T α =ZCn eα((w−u)·w)T α ≤(cid:18)ZCn T α (νβ)w 1(u)2e− 4α (νβ)w1(w − u)e− α (νβ)w 1(u)e− α dv(u)(cid:19) 1 5 u2 2 u2 dv(u) 2 w−u2 dv(u) 2(cid:18)ZCn 4α 5 u2 e e2α((w−u)·w)e−αw−u2 2 dv(u)(cid:19) 1 By a simple computation, we have that (cid:18)ZCn e 4α 5 u2e2α((w−u)·w)e−αw−u2 2 dv(u)(cid:19) 1 = e α 2 w2(cid:18)ZCn e− α 5 u2 2 dv(u)(cid:19) 1 . However, Lemmas 3.5 and 3.6 give us that lim β→∞ w∈Cn T α sup (νβ)w 1(u)2 = 0 pointwise in u and w∈Cn T α sup (νβ )w1(u)2e− 4α 5 u2 ≤ Ce− 3α 10 u2 for some C > 0 independent of β. Thus, the dominated convergence theorem gives us that ZCn Φβ(w, u)h(u)2 e−αu2 dv(u) ≤ c(β)h(w)2 where c(β) is independent of w and limβ→∞ c(β) = 0. An application of the Schur test now completes the proof. Next we will show that all compact operators A on F p α for 1 < p < ∞ are contained in the Toeplitz algebra T α p . Note that this was first proved in [9] (using completely different methods) for the special case p = 2 and α = 1 2 . 17 then qβ(z) :=(cid:18) β π(cid:19)n β→∞kT α lim Proof. Note that (1⊗1)h = T α α = 0. α→F p qβ − 1 ⊗ 1kF p δ0h = h(0) if h ∈ F p exp(cid:8)−βz2(cid:9) , For any f ∈ F p the standard tensor product operator on F p α and g ∈ F q α where q is the dual exponent of p, let f ⊗ g be α defined by f ⊗ g = h·, giαf. Since all Lp spaces have the bounded approximation property (see [22], p. 69-70), this will be proved (by linearity) if we can show that each finite rank operator f ⊗ g on F p α can be approximated in the operator norm by a Toeplitz operator with symbol in C ∞ b . First we will show that 1 ⊗ 1 can be approximated in the operator norm by a Toeplitz operator with symbol in C ∞ b . Lemma 3.8. Let β > 0 and let 1 < p < ∞. If α where δ0 is the usual point- ( 1 4β ) = qβ for each β > 0. The result now immediately follows from Theorem 3.7 since δ0 is a Fock-Carleson measure. mass measure at 0 ∈ Cn. Also, note that by defintion we have eδ0 Given w ∈ Cn, define the "weighted shift" operator Cα(w) on Lp α by [Cα(w)f ] (z) := f (z − w)eα(z·w)− α 2 w2 . It is known (see [17]) that Cα(w) is an isometry of F p α) onto itself for 1 ≤ p ≤ ∞, and it is easy to check that Cα(w)−1 = Cα(−w). Moreover, for w1, w2 ∈ Cn, one has that α (and Lp Cα(w1)Cα(w2) = e−iαIm(w1·w2)Cα(w1 + w2). Note that the operators Cα(w) are in fact Toeplitz operators with bounded 2w2 + 2iα Im(z · w)}, then we have symbols. To see this, if sw(z) := exp{ α 18 that α swf(cid:3) (u) = e (cid:2)T α α = e 2 w2Df Kα(·, w)Kα(·,−w), Kα(·, u)Eα 2 w2Df Kα(·, w), Kα(·, u − w)Eα =hCα(w)fi(u) = f (u − w)Kα(u − w, w)e α 2 w2 α ∩ F 2 Since Cα(v)T α for all f ∈ F p can be shown that Cα(v)T α "product" ♯α is defined by α, which shows that T α qβ(·−u) can be written as the Toeplitz product T α sw = Cα(w) on F p α. qβ(·−u), it where fβ := sv ♯α qβ(· − u) and the sv T α qβ(·−u) = T α fβ ψ ♯α ϕ := Xγ∈Nn 0 1 (−α)γγ! ∂γψ ∂zγ · ∂γϕ ∂zγ for suitable smooth functions ψ and ϕ on Cn (see [2] for more details.) Using this formula, one can directly compute that α (cid:19)n fβ(z) =(cid:18)β + α expnα 2 v2 + β(z − u) · v + 2iα Im(z · v) − βz − u2o. qβ(·−u) = T α Note that one could also directly verify the equality Cα(v)T α fβ where fβ is defined as above by comparing the Berezin transforms of both sides. Using these shift operators and their properties, we can now prove α, and g ∈ F q Theorem 3.9. If 1 < p < ∞, f ∈ F p Proof. Since span{K(·, w) : w ∈ Cn} is dense in F p that each K(·, v) ⊗ K(·, w) is in T α Cα(v)(cid:0)1 ⊗ 1(cid:1)Cα(−w)g = hCα(−w)g, 1iα Cα(v)1 p . Furthermore, if g ∈ F p α, then f ⊗ g ∈ T α p . α, it is enough to show α, then = hg, Cα(w)1iα Cα(v)1 = e− α 2 (w2+v2)Dg, Kα(·, w)Eα Kα(·, v) 2 (w2+v2) (Kα(·, v) ⊗ Kα(·, w)) g. = e− α 19 Thus, we only need to show that operators of the form Cα(v)(cid:0)1 ⊗ 1(cid:1)Cα(−w) can be approximated by Toeplitz operators with symbols in C ∞ b . Moreover, since Cα(v)T α qβ Cα(−w) = Cα(v)Cα(−w)Cα(w)T α = eiα·Im(v·w)Cα(v − w)T α qβ Cα(−w) qβ(·−w), we can write Cα(v)T α (depending on v and w) Fβ ∈ C ∞ qβ Cα(−w) as a single Toeplitz operator T α Fβ b . Finally, this fact tells us that with symbol lim β→∞(cid:13)(cid:13)(cid:13)T α α→F p α = lim Fβ − Cα(v)(cid:0)1 ⊗ 1(cid:1)Cα(−w)(cid:13)(cid:13)F p qβ Cα(−w) − Cα(v)(cid:0)1 ⊗ 1(cid:1)Cα(−w)(cid:13)(cid:13)(cid:13)F p β→∞(cid:13)(cid:13)(cid:13)Cα(v)T α qβ − 1 ⊗ 1(cid:13)(cid:13)(cid:13)F p β→∞(cid:13)(cid:13)(cid:13)T α ≤ lim = 0 α→F p α α→F p α where the last equality follows from Lemma 3.8. 4. Sampling and interpolation results for the Fock space The proofs of the following two lemmas (Lemmas 4.1 and 4.3) borrow deep ideas from the theory of sampling and interpolation in Fock spaces. In particular, the proof of Lemma 4.1 is similar to the proof of Theorem 5.1 in [18]. On the other hand, Lemma 4.3 is a "folklore" result in sampling theory and follows from the machinery developed in [11] for abstract coorbit spaces. However, since Lemma 4.3 is not explicitly stated in [11], we will provide a short and direct proof. Before we state and prove Lemma 4.1, we need to briefly discuss the pseudo-hyperbolic metric ρ on Bn. Given any z ∈ Bn, let φz be the involutive automorphism of Bn that interchanges 0 and z. The pseudo-hyperbolic metric ρ on Bn is then defined by the formula ρ(z, w) = φz(w). It is well known (see [24]) that ρ is indeed a metric on Bn and that ρ satisfies the identity 1 − (ρ(z, w))2 = (1 − z2)(1 − w2) . 1 − z · w 20 Lemma 4.1. Suppose that 1 < p ≤ 2, r > 1, and wk ∈ B(0, r) for k = 1, . . . , m are points where wk − wj ≥ ǫ > 0 if j 6= k. Then for any 1 ≤ k0 ≤ m, there exists gk0 ∈ F p α and a constant C = C(ǫ, r) > 0 (which is assumed to also depend on n, α, and p, but does not depend on the sequence {wk}k itself ) such that gk0(wk) = δk0,k and kgk0kα,p ≤ C. Remark. Since we require that C = C(ǫ, r) does not depend on the actual se- quence {wk}k itself, Lemma 4.1 does not immediately follow from the results in [18]. Proof. First note that if 1 < p < 2, α > 0, and g ∈ Lp α, then a direct application of Holder's inequality tells us that kgkα,p ≤ Cα,α′kgkα′,2 for any 0 < α′ < α. Thus, it is enough to prove the lemma for p = 2 and arbitrary α > 0. For the rest of the proof, C will denote a positive constant that may depend on ǫ, r, n, p and α, but not on the actual sequence {wk}k itself. Now if wk − wj ≥ ǫ when j 6= k, then clearly {B(wk, ǫ 2)}k is a pairwise disjoint sequence of balls with B(cid:16)wk, [k m ≤(cid:18) 2r 2(cid:17) ⊂ B(cid:16)0, r + + 1(cid:19)2n =: Mr,ǫ. ǫ 2(cid:17) ǫ ǫ which means that Thus, since inf k mYj6=k wj − wk r + ǫ >(cid:18) ǫ r + ǫ(cid:19)Mr,ǫ−1 for k ≤ m, it follows from the discussion preceeding the statement of Lemma 4.1 that inf k mYj6=k ρ(cid:18) wj r + ǫ , wk r + ǫ(cid:19) > C 21 for k ≤ m. Now, it is easy to construct a bounded function ϕk0 that is holomorphic on B(0, r + ǫ) with ϕk0(wk) = δk0,k and sup z∈B(0,r+ǫ)ϕk0(z) ≤ C. In particular, let mYj6=k0 φwj (wk0)φwj (z) φwj(wk0)2 Let C ∞ fϕk0(z) := r+ǫ(cid:1). and set ϕk0(z) := fϕk0(cid:0) z η ≡ 1 on B(cid:16)0, c (Cn) denote the space of all smooth, compactly supported complex valued functions on Cn. Pick any η ∈ C ∞ c (Cn) where ǫ 2(cid:17) and η ≡ 0 on Cn\B(cid:18)0, 2ǫ 3(cid:19) . If we define ψ ∈ C ∞ c (Cn) by ψ(z) := mXk=1 η(z − wk), then ψ satisfies ψ ≡ 1 on m[k=1 B(cid:16)wk, ǫ 2(cid:17) and ψ ≡ 0 on Cn\ m[k=1 B(cid:18)wk, 2ǫ 3(cid:19) . If we extend ϕk0(z) to z ≥ r + ǫ by setting ϕk0(z) ≡ 0 for z ≥ r + ǫ and c (Cn) satisfies let fFk0(z) = ψ(z)ϕk0(z), then fFk0 ∈ C ∞ (i) fFk0(wk) = δk0,k, (ii) kfFk0kL∞ ≤ C, (iii) ∂fFk0 is supported onSm (iv) fFk0 is supported onSm 22 k=1 B(wk, ǫ)\B(wk, ǫ 2), k=1 B(wk, ǫ). Now for large R > r, let v(z) be the negative function v(z) := n Xk:z−wk<R"log(cid:12)(cid:12)(cid:12)(cid:12) z − wk R (cid:12)(cid:12)(cid:12)(cid:12) 2 + 1 −(cid:12)(cid:12)(cid:12)(cid:12) z − wk 2# . R (cid:12)(cid:12)(cid:12)(cid:12) It is easy to see that φ(z) := v(z) + α 2z2 is plurisubharmonic for R large enough (depending on ǫ and r), and so Hormander's Theorem (Theorem loc(Cn) to the equation 4.4.2 in [13]) gives us a (distributional) solution u ∈ L2 ∂u = ∂fFk0 where ZCn u(z)2(1 + z2)−2e− α 2 z2 However, since α (cid:12)(cid:12)(cid:12)φ(z) − log z − wk2n − we get that φ(z) ≤ C for all z ∈Sm ∂fFk0 is supported onSm ZCn u(z)2e−αz2 dv(z) ≤ZCn u(z)2(1 + z2)−2e−φ(z) dv(z) (4.1) k=1 B(wk, ǫ)\B(wk, ǫ ≤ZCn ∂fFk0(z)2e−φ(z) dv(z). 2z2(cid:12)(cid:12)(cid:12) ≤ C when z − wk < ǫ, dv(z) ≤ CZCn u(z)2(1 + z2)−2e− α ≤ CZCn ∂fFk0(z)2e−φ(z) dv(z) 2 )∂fFk0(z)2 k=1 B(wk, ǫ)\B(wk, ǫ k=1 B(wk,ǫ)\B(wk, ǫ z∈Sm sup ≤ C ≤ C 2 z2 dv(z) 2 ). Moreover, since 2), we get from (4.1) that (4.2) where the last inequality follows from the product rule combined with the Cauchy estimates applied to fFk0 Now note that if Fk0 := u − fFk0, then Fk0 is entire, so that u ∈ C ∞(Cn) and kFk0kα,2 ≤ C. Finally, (4.2) and the fact that e−φ(z) ≈ z − wk−2n for z near wk tells us that u(wk) = 0, so that Fk0(wk) = δk0,k, which completes the proof. 23 We need to set up some simple notation and machinery before we state and prove Lemma 4.3. Let Hn = Cn×∂D be the quotient of the n dimensional complex Heisenberg group by 2πZ, with group law (z1, t1)(z2, t2) = (z1 + z2, t1t2e−iαIm z1·z2) and with Haar measure m being the Lebesgue measure m = dvdθ on Cn×∂D where dθ is the ordinary (normalized) arc length measure on ∂D. Let T : α → Lp(Hn) be the isometry given by F p T f (z, t) = te− α 2 z2 f (z). For f ∈ Lp(Hn) and g ∈ Lq(Hn) where q is the dual exponent of p, let f ∗ g be the convolution product defined by f ∗ g(h) =ZHn f (y)g(hy−1) dm(y) 2 z2, then the reproducing property of F p for h ∈ Hn. If G(z, u) = ue− α us that F ∗ G = F for any F ∈ T (F p α). Now, enumerate ǫZ2n for fixed ǫ as {zj}j. For any fixed integer Nǫ > ǫ−1 and any integer 0 ≤ k < Nǫ, let uk = exp( 2πik ). Let Uǫ = [0, ǫ)2n × {e2πiθ : 0 ≤ θ < 1 Nǫ} ⊂ Hn and (for any integer 0 ≤ k < Nǫ) let Gjk be the set Uǫ translated on the right by (zj, uk), so that Gjk = Uǫ(zj, uk). Clearly we then have that: α tells Nǫ (a) Hn =Sj,k Gjk, (b) Gjk ∩ Gj′k′ = ∅ if (j, k) 6= (j′, k′). Note that m(Gjk) only depends on ǫ and not on j or k. Thus, if cǫ = m(Gjk), then we can define an operator Rǫ on T (F p α) ⊂ Lp(Hn) by RǫF (z, u) := cǫXj,k F (zj, uk)G((z, u)(zj, uk)−1). By a direct calculation we have that Rǫ : T (F p particular, if f ∈ F p α, then α) → T (F p α) boundedly. In RǫT f (z, u) = c′ ǫue− α f (zj)eα(z·zj)−αzj 2 = c′ ǫue− α 2 z2 Tνǫf (z) 2 z2Xj 24 ǫ = v([0, ǫ)2n), v is the ordinary Lebesgue volume measure on Cn, where c′ and νǫ is the measure νǫ = Xσ∈ǫZ2n δσ where δσ is the point-mass measure at σ. But since νǫ is a Fock-Carleson measure, it is clear that Rǫ : T (F p α) boundedly. Let χjk be the α) → Lp(Hn) characteristic function of Gjk and define the operator Sǫ : T (F p by α) → T (F p SǫF =Xj,k F (zj, uk)χjk ∗ G. Finally, define the sharp maximal function G♯ Uǫ on Hn by G♯ Uǫ(h) = sup u∈UǫG(u−1h) − G(h) and define Ge♯ Uǫ on Hn by Ge♯ Uǫ(h) = sup u∈UǫG(hu) − G(h). Lemma 4.2. Given F ∈ T (F p α), we have that kF − SǫFkLp(Hn) ≤ o(ǫ)kFkLp(Hn) where lim ǫ→0+ o(ǫ) = 0. Proof. Since F ∗ G = F for F ∈ T (F p us that α), Young's convolution inequality gives kF − SǫFkLp(Hn) = k(F −Xj,k ≤ kF −Xj,k F (zj, uk)χjk) ∗ GkLp(Hn) F (zj, uk)χjkkLp(Hn)kGkL1(Hn). 25 erty as follows: F (zj, uk)χjk(z, u) We now estimate F −Pj,k F (zj, uk)χjk pointwise using the reproducing prop- F (z, u) −Xj,k ≤Xj,k =Xj,k (cid:12)(cid:12)(cid:12)(cid:12)ZHn [G((z, u)y−1) − G((zj, uk)y−1)]F (y) dm(y)(cid:12)(cid:12)(cid:12)(cid:12) χjk(z, u). F (z, u) − F (zj, uk)χjk(z, u) (4.3) Fix any j and k. Since (z, u) ∈ Gjk for each summand in (4.3), we can write (z, u) = (z′, u′)(zj, uk) where (z′, u′) ∈ Uǫ, so that (zj, uk) = (z′, u′)−1(z, u). Plugging this into (4.3) and recalling the definition of G♯ Uǫ, we get that F (zj, uk)χjk(z, u) F (z, u) −Xj,k ≤Xj,k (cid:12)(cid:12)(cid:12)(cid:12)ZHn ≤Xj,k (cid:18)ZHn [G((z, u)y−1) − G((z′, u′)−1(z, u)y−1)]F (y) dm(y)(cid:12)(cid:12)(cid:12)(cid:12) χjk(z, u) Uǫ((z, u)y−1)F (y) dm(y)(cid:19) χjk(z, u) G♯ = F ∗ G♯ Uǫ(z, u). The proof is now completed by another application of Young's convolution inequality and the easily checked fact that ǫ→0+ kG♯ lim UǫkL1(Hn) = 0. Now we will state and prove Lemma 4.3. Note that in the language of sampling theory, Lemma 4.3 states that the "frame operator" f 7→ Xσ∈ǫZ2nhf, kα σiαkα σ on F p α associated to the frame {kα σ}σ∈ǫZ2n for small enough ǫ > 0 is invertible. 26 Lemma 4.3. Let 1 < p < ∞ and let νǫ be the measure νǫ = Xσ∈ǫZ2n δσ where δσ is the point-mass measure at σ. Then Tνǫ is invertible on F p small enough ǫ > 0. α for Proof. Note that by definition, we have Tνǫf = c′ enough to show that Rǫ : T (F p will be proved if we can show that α) → T (F p ǫT −1RǫT f . Thus, it is α) is invertible. By Lemma 4.2, this kRǫF − SǫFkLp(Hn) ≤ o(ǫ)kFkLp(Hn) where F ∈ T (F p α) and lim ǫ→0+ o(ǫ) = 0. To that end, we now pointwise estimate RǫF − SǫF as follows: RǫF (z, u) − SǫF (z, u) ≤Xj,k ≤Xj,k (m(Gjk)F (zj, uk)G((z, u)(zj, uk)−1) − F (zj, uk)ZGjk G((z, u)y−1) dm(y)) F (zj, uk)ZGjk G((z, u)(zj, uk)−1) − G((z, u)y−1) dm(y). (4.4) However, since y ∈ Gjk we can write y = y′(zj, uk) for some y′ ∈ Uǫ so that (zj, uk)−1 = y−1y′, and plugging this into (4.4) gives us that RǫF (z, u)−SǫF (z, u) F (zj, uk)ZGjk G((z, u)y−1y′) − G((z, u)y−1) dm(y) F (zj, uk)ZGjk Ge♯ F (zj, uk)χjk! ∗ Ge♯ Uǫ(z, u)y−1) dm(y) Uǫ. ≤Xj,k ≤Xj,k = Xj,k 27 Again, it is easy to see that lim ǫ→0+ kGe♯ UǫkL1(Hn) = 0 so by Young's inequality, we only need to show that kXj,k F (zj, uk)χjkkLp(Hn) ≤ CkFkLp(Hn) which follows easily from Lemma 2.4. For the rest of the paper, ν will denote the Fock-Carleson measure νǫ0 ν is invertible. from Lemma 4.3 where ǫ0 is fixed and small enough so that T α 5. A uniform algebra A and its maximal ideal space Let A ⊂ L∞ be the unital C ∗-algebra of all bounded and uniformly continuous functions on Cn. Since C ∞ b ⊂ A, it follows from the Theorem 3.7 that T α p for 1 < p < ∞ is the closed algebra generated by Toeplitz operators with symbols in A. In this section, we will extend the Berezin transform and other related objects defined on Cn to MA, where MA denotes the space of non-zero multiplicative functionals on A equipped with the weak∗ topology. Note that A is not separable and hence the space MA is not metrizable. Since A is a commutative, unital C ∗-algebra, the Gelfand-transform ∧ : A → C(MA) defined by a(ϕ) = ϕ(a) for a ∈ A and ϕ ∈ MA gives us an isomorphism between A and C(MA). In the following we will often write a(ϕ) instead of a(ϕ). For x ∈ Cn, let δx ∈ MA be the point evaluation at x defined by δx(f ) = f (x). It is not difficult to see that the map x 7→ δx induces a dense embedding of Cn into MA. For w ∈ Cn, let τw be the usual translation function τw(z) := z − w. More generally, if x ∈ MA and w ∈ Cn, then define τx ∈ Qw∈Cn MA by τx(w)(a) := x(a ◦ τw) where a ∈ A. We will write a ◦ τx(w) instead of τx(w)(a) since τx naturally extends the translation by elements in Cn to a "translation" by elements in MA. Let ǫ > 0, w1, w2 ∈ Cn and a ∈ A, then we have a ◦ τx(w1) − a ◦ τx(w2) ≤ ka ◦ τw1 − a ◦ τw2k∞ < ǫ if w1 − w2 < δ where δ > 0 is chosen suitably according to the uniformly continuity of a. Therefore we have shown that the map τx : Cn → MA is continuous. Next, prove: 28 Lemma 5.1. If (zβ)β is a net in Cn converging to x ∈ MA, then a◦τzβ (w) → a ◦ τx(w) for all a ∈ A and w ∈ Cn where the convergence is uniform on compact subsets of Cn. Proof. Since a ◦ τw ∈ A for all w ∈ Cn, it follows by definition of the conver- gence zβ → x in MA that a ◦ τzβ (w) = δzβ (a◦ τw) → x(a ◦ τw) = a◦ τx(w) for all a ∈ A and w ∈ Cn. Assume that the above convergence is not uniform on compact subsets of Cn. Then there is ǫ > 0, a function a ∈ A, and a compact set K ∈ Cn such that for all γ, there is β > γ and ξβ ∈ K with (5.1) (cid:12)(cid:12)(a ◦ τzβ )(ξβ) − (a ◦ τx)(ξβ)(cid:12)(cid:12) > ε. By passing to a subnet, we can assume that ξβ → ξ ∈ K. Now, we have (a ◦ τzβ )(ξβ) − (a ◦ τx)(ξβ) ≤(cid:12)(cid:12)(a ◦ τzβ )(ξβ) − (a ◦ τzβ )(ξ)(cid:12)(cid:12) +(cid:12)(cid:12)(a ◦ τzβ )(ξ) − (a ◦ τx)(ξ)(cid:12)(cid:12) + (a ◦ τx)(ξ) − (a ◦ τx)(ξβ) . Since τzβ (ξβ) − τzβ (ξ) = ξβ − ξ it follows that the first term on the right hand side tends to zero. The third term on the right tends to zero by the continuity of τx : Cn → MA, and the second term tends to zero by what was said at the beginning of the proof. We obtain a contradiction to (5.1) and the lemma is proven. For w ∈ Cn, let Cα(w) be the "weighted shifts" defined in Section 3. α, then we write α) defined by If 1 ≤ p ≤ ∞ and A is a fixed bounded operator on F p Aw := Cα(w)ACα(−w), which induces a map ΨA : Cn → L(F p ΨA(w) := Aw. Since (cid:2)Cα(−w)Kα(·, ξ)(cid:3)(z) = Kα(z, ξ − w)Kα(w, ξ)e− α 2 w2 , we have that empty) multi-valued function on MA defined by Bα ◦ ΨA(w) = Bα(Aw) = Bα(A) ◦ τw. (5.2) Let E be a metric space and let f : Cn → E. Consider the (possibly F (x0) :=(cid:8)λ : f (zβ) → λ : for some net zβ → x0 and zβ ∈ Cn(cid:9) where x0 ∈ MA. We will say that F is "single valued" if for any x0 ∈ MA and any convergent net zβ → x0 where zβ ⊂ Cn, the net (f (zβ))β converges in E. 29 Lemma 5.2. Assume that F is single valued for all x0 ∈ MA. Then F : MA → E is well-defined and continuous. Proof. Since Cn is dense in A and E (as a metric space) is regular, the result follows immediately by our hypotheses and Bourbaki's extension theorem (Theorem 1, p. 81 in [5]). α) denote the unit ball (in the norm topology) of the space of If "SOT" refers to the strong operator topol- α), SOT) is a complete metric space since F p α is p and without loss of Let B1(F p bounded operators on F p α. ogy, then recall that (B1(F p separable. In the next proposition we will fix A ∈ T α generality assume that A ∈ B1(F p α). Proposition 5.3. If 1 ≤ p < ∞ and A ∈ T α extends continuously to MA. p , then ΨA : Cn → (B1(F p α), SOT) Proof. Consider the multi-valued function ΨA(x) :=(cid:8)λ : ΨA(zγ) → λ for some net zγ → x and zγ ∈ Cn(cid:9) where x ∈ MA. According to Lemma 5.2 we need to show that ΨA(x) is single valued. Since (B1(F p α), SOT) is a complete metric space, it is sufficient to show that {ΨA(zγ)}γ is a Cauchy net whenever {zγ}γ ⊂ Cn is a net converging to some x ∈ MA. To that end, let {zγ}γ ⊂ Cn be a net that converges to x ∈ MA. Let A ∈ T α p and pick ǫ > 0. Choose R in the (non-closed) algebra generated by {T α f : f ∈ A} with kRkF p α < ǫ. Then for all f ∈ F p α ≤ 1 such that kA− RkF p α, we have that α→F p α→F p p →F α p (cid:13)(cid:13)(cid:2)ΨA(zγ) − ΨA(zβ)(cid:3)f(cid:13)(cid:13)F α ≤hkAzγ − RzγkF p ≤ 2ǫkfkα,p +(cid:13)(cid:13)(cid:2)ΨR(zγ) − ΨR(zβ)(cid:3)f(cid:13)(cid:13)α,p. α + kRzβ − AzβkF p α→F p α→F p αikfkα,p + k[Rzγ − Rzβ ]fkα,p Therefore, it is sufficient to show that {ΨR(zγ)}γ is a Cauchy net with re- spect to the SOT for all R in the algebra generated by {T α f : f ∈ A} with kRkF p α ≤ 1. Moreover, by linearity and Proposition 3.6, we can assume that R = T α am is a finite product of Toeplitz operators with sym- bols in aj ∈ A where kajk∞ ≤ 1. Since the product of convergent nets in a2 · · · T α a1T α α→F p 30 α), SOT) is convergent, is it sufficient to show that {ΨT α (B1(F p a ∈ A with kak∞ ≤ 1 has a limit in (B1(F p α), SOT). Now if w ∈ Cn, then a (zγ)}γ for all ( 1 4α ) a◦τw(cid:1)(z) = ^(a ◦ τw) B(cid:0)T α = B(cid:0)T α (z) =ea( 1 a(cid:1) ◦ τw(z) = B(cid:0)(T α 4α ) ◦ τw(z) a )w(cid:1)(z) where we have used standard properties of the heat transform together with (5.2) in the last equality. Since the Berezin transform is one-to-one on oper- ators, it follows that ΨT α a (zγ) = (T α a )zγ = T α a◦τzγ . Let f ∈ A and let (fβ)β ⊂ A be a net that converges to f uniformly on compact subsets of Cn. Then it is easy to check that T α in SOT. According to Lemma 5.1, we have the uniform compact convergence a◦τzγ → a ◦ τx, and therefore ΨT α a◦τzγ → T α fβ → T α a◦τx in SOT. a (zγ) = T α f The proof of the following corollary follows precisely from the proof of Proposition 5.3, though it will be useful to record it for future use. p . If x ∈ MA\Cn and Corollary 5.4. Suppose that 1 < p < ∞ and S ∈ T α {zγ}γ is a net converging to x, then Sx = limγ Szγ where the limit is taken in the strong operator topology. 6. Proof of Theorem 1.1 Finally in this section we will give a proof of Theorem 1.1. As in [21], the proof will follow easily from quantitative estimates on the essential norm kAke for operators A ∈ T α p . We will first prove the following simple lemma and then state some definitions that are needed for the proof of Theorem 1.1. p , then (a) and (b) below are equivalent: Lemma 6.1. Let A ∈ T α (a) Bα(A)(z) → 0 as z → ∞. (b) Ax = 0 for all x ∈ MA \ Cn. 31 Proof. Obviously we may assume that kAkF p (a) ⇒ (b): Let x ∈ MA \ Cn and (zγ)γ be a net with zγ → x. According to Corollary 5.4, we have for all fixed ξ ∈ Cn that: α ≤ 1. α→F p −→ 0. (6.1) Combining (6.1) with (5.2) tells us that ξ , kα (cid:12)(cid:12)Bα(Azγ )(ξ) − Bα(Ax)(ξ)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:10)(cid:2)Azγ − Ax(cid:3)kα ξ(cid:11)α(cid:12)(cid:12)(cid:12) ≤(cid:13)(cid:13)(cid:2)Azγ − Ax(cid:3) kα ξ(cid:13)(cid:13)α,p Bα(A)(ξ − zγ) =(cid:12)(cid:12)Bα(Azγ )(ξ)(cid:12)(cid:12) −→ Bα(Ax)(ξ) . γ γ α it follows that Ax = 0. Since x ∈ MA \ Cn, we can assume that zγ → ∞ and from the condition limz→∞ Bα(A)(z) = 0 we conclude that Bα(Ax)(ξ) = 0 for all ξ ∈ Cn. Since Bα is one-to-one on the bounded operators on F p (b) ⇒ (a): Assume that there is a sequence (zk)k ⊂ Cn such that zk → ∞ and (6.2) Since MA is compact there is a subnet (zγ)γ of (zk)k and x ∈ MA such that zγ → x ∈ MA. From (6.1) with ξ = 0 and (6.2), we get that Bα(Ax)(0) ≥ δ > 0, which says that Ax 6= 0 as desired. Now for any bounded operator S on F p Bα(Azk )(0) = Bα(A)(−zk) ≥ δ > 0. α and any r > 0, let αS(r) := lim sup z→∞ sup{kSfkα,p : f ∈ T α χB(z,r)ν(F p α),kfkα,p ≤ 1}. An easy application of Lemma 4.1 tells us that T α when r1 < r2, which means that αS(r) is an increasing function of r. particular, recall from the end of Section 4 that χB(z,r1)ν(F p χB(z,r2)ν(F p α) In α) ⊂ T α ν = Xσ∈ǫ0Z2n δσ for ǫ0 fixed so that T α ν is invertible. Now if h ∈ T α χB(z,r1)ν(F p α) then h(w) = Xσ∈ǫ0Z2n∩B(z,r1) g(σ)eα(w·σ)−ασ2 32 for some g ∈ F p some gσ ∈ F p α. For each σ ∈ ǫ0Z2n ∩ B(z, r1), Lemma 4.1 allows us to pick α where gσ(σ′) = δσ,σ′ for any σ′ ∈ ǫ0Z2n ∩ B(z, r2). Thus, if g(σ)gσ, eg := Xσ∈ǫ0Z2n∩B(z,r1) χB(z,r2)νeg. Note that since αS(r) ≤ kSk for all αS := lim r→∞ αS(r) = sup r>0 αS(r) ≤ kSk. α and clearly h = T α theneg ∈ F p r, we have The proof of Theorem 1.1 will follow easily from Lemma 6.1 and the following result, Theorem 6.2. In the statement and proof of this result, we will use the symbol " ≈ " to indicate that two quantities are equivalent with constants only depending on α, ǫ0, p, and n. Moreover, C will denote a constant depending only on α, ǫ0, p, and n, and can possibly change from line to line. Theorem 6.2. Let 2 ≤ p < ∞ and let A ∈ T α p . If kAke denotes the essential norm of A, then kAke is equivalent to the following quantities (with constants depending on only α, ǫ0, p, and n) (i) αA, (ii) βA := supd>0 lim supz→∞ kMχB(z,d)AkF p (iii) γA := limr→∞ kMχB(0,r)c AkF p α→Lp α, α→Lp α where B(0, r)c = Cn\B(0, r). Moreover, for all 1 < p < ∞, we have that kAke ≈ sup x∈MA\CnkAxkF p α. α→F p Remark. Note that the proof of Theorem 6.2 is similar to the proofs of The- orems 5.2 and 9.3 in [21]. Thus, we will sometimes only outline arguments of the proof and refer the reader to [21] for the full details. Proof. First note that in the last statement of Theorem 6.2, it is enough to prove kAke ≈ sup x∈MA\CnkAxkF p α→F p α 33 if A ∈ T α A ∈ T α p p , then for 2 ≤ p < ∞, since if 1 < p ≤ 2, q is the dual exponent, and kAke = kA∗ke ≈ sup = sup α→F q α α→F q α x∈MA\Cnk(A∗)xkF q x∈MA\Cnk(Ax)∗kF q x∈MA\CnkAxkF p α→F p α = sup where we have used the equality (A∗)x = (Ax)∗ for all x ∈ MA, which follows from the WOT continuity of taking adjoints combined with Corollary 5.4. As in [21], we will use the notation k · ke and k · kex to distinguish the α to α. Moreover, if R is a bounded operator from F p α to itself, then it is easy essential norms of an operator from F p Lp to see that α to itself and an operator from F p kRkex ≤ kRke ≤ kPαkLp α→Lp αkRkex so that kRkex and kRke are equivalent. where Now pick ǫ > 0 and choose Borel sets Fj ⊂ Gj ⊂ Cn as in Lemma 3.3 (6.3) AT α MχFj χGj νkF p α→Lp α ≤ ǫ ν −Xj kAT α and write Am =Pj≥m MχFj any m ≥ 1, (6.3) tells us that j=1 MχFj AT α χGj ν. SincePm kAT α ν − Amkex < ǫ AT α χGj ν is compact for (6.4) for any m ≥ 1. However, by Lemma 3.3, the monotonicity of the function r 7→ αA(r), and the arguments in [21] p. 2209-2210, we have that m→∞ kAmkF p lim sup which combined with (6.4) tells us that α→Lp α ≤ CαA kAT α ν kex ≤ lim sup m→∞ kAmkF p α→Lp α + ǫ ≤ CαA + ǫ. But since Lemma 4.3 tells us that T α inequality tells us that ν is invertible, letting ǫ ↓ 0 in the previous (6.5) kAke ≤ CαA. 34 α→Lp If Q is some compact operator from F p Now again by the arguments in [21], p. 2210 - 2211, we have that βA, γA, α are equivalent. Moreover (6.5) will give us that and lim supm→∞ kAmkF p kAke is equivalent to (i), (ii), and (iii) if we can show that αA ≤ CkAke. α to itself, then it follows from Theorem 3.9 that Q ∈ T α p , and thus it follows easily from Lemma 6.1 that Qx = 0 for all x ∈ MA\Cn. Fix x ∈ MA\Cn and let {zγ}γ ⊂ Cn be a net converging to x. Since SOT limits (except for a multiplicative constant) do not increase the norm, Proposition 5.3 and the SOT convergence Azγ +Qzγ → Ax + Qx = Ax gives us that kAxkF p α→F p α ≤ C lim inf γ kAzγ + QzγkF p α. α→F p Since this holds for all x ∈ MA\Cn and all compact Q, we easily get that x∈MA\CnkAxkF p sup α→F p α ≤ CkAke. (6.6) Finally, by (6.5) and (6.6), the proof will be completed if we can show that αA ≤ C sup x∈MA\CnkAxkF p α. α→F p To that end, let r > 0. Pick a sequence {zj}j tending to ∞ as j → ∞ and a normalized sequence fj ∈ T α α) such that kAfjkα,p → αA(r). Thus, there are hj ∈ F p χB(zj ,r)ν(F p α where fj(w) = T α χB(zj ,r)νhj(w) = Xσ∈ǫ0Z2n∩B(zj ,r) hj(σ)eα(w·σ)−ασ2 . For each σ ∈ ǫ0Z2n ∩ B(zj, r), let σ(j) = σ − zj so that each σ(j) ∈ B(0, r). A direction computation now tells us that Cα(zj)∗fj(w) = Xσ(j)∈B(0,r) aσ(j)kα σ(j)(w) where aσ(j) = hj(σ)e− α 2 σ2eiα Im σ·zj . Let q be the dual exponent of p, so that 1 < q ≤ 2. Now for each fixed j and fixed σ0(j) ∈ B(0, r), pick g = gj,σ0(j) according to Lemma 4.1 where g(σ(j)) =(cid:26) 1 0 if σ = σ0 if σ 6= σ0 35 and kgj,σ0(j)kα,q ≤ C where C = C(α, ǫ0, p, n, r). Then by the reproducing property, we get that hCα(−zj)fj, giF 2 aσ(j)e− α 2 σ(j)2 g(σ(j)) α = Xσ(j)∈B(0,r) = aσ0(j)e− α 2 σ0(j)2 which says that each aσ(j) ≤ C for some C = C(α, ǫ0, p, n, r). Now pick some M = M(ǫ0, n, r) where M ′(j) := card ǫ0Z2n∩B(zj, r) ≤ M and enumerate ǫ0Z2n ∩ B(zj, r) as σ1, . . . , σM . Clearly we may choose a subsequence of {zj}j such that M ′(j) ≡ M0 ≤ M for some M0 ∈ N that is independent of j, and so without loss of generality we can assume that {(σ1(j), . . . , σM0(j), aσ1(j), . . . , aσM0 (j))}∞ j=1 ⊂ C(n+1)M0. Since the sequence {(σ1(j), . . . , σM0(j), aσ1(j), . . . , aσM0 (j))}∞ j=1 is bounded, we can (passing to another subsequence if necessary) assume that this sequence converges to a point (σ1, . . . , σM0, aσ1, . . . , aσM0 ) ∈ C(n+1)M0 where each σi ≤ C = C(α, ǫ0, p, n, r). An easy application of the Lebesgue dominated convergence theorem then gives us that Cα(zj)∗fj → h := M0Xi=1 aσikσi in F p α (6.7) which says that αA(r) = lim j kAfjkα,p = lim = lim j kA−zj Cα(−zj)fjkα,p j kA−zj hkα,p. (6.8) Now since MA is compact, we can choose a subnet {−zγ} of the sequence {−zj} converging to some x ∈ MA\Cn, which means that limγ kA−zγ hkα,p = limγ kA−xhkα,p. Combining this with (6.8) and Proposition 5.3 finally gives us that αA(r) = lim γ kA−zγ hkα,p = kAxhkα,p ≤ sup u∈MA\CnkAukF p α→F p α since (6.7) tells us that khkα,p = 1. 36 Corollary 6.3. Let 1 < p < ∞ and A ∈ T α z→∞ kAzfkα,p. lim sup kAke ≈ sup p . Then kf kα,p=1 Proof. The proof is identical to the proof of Corollary 9.4 in [21]. Proof of Theorem 1.1. It has already been proven that A is compact on F p α only if A ∈ T α p and it was remarked in the introduction that Bα(A) vanishes at infinity if A is compact. Now if Bα(A) vanishes at infinity, then Lemma 6.1 tells us that Ax = 0 for all x ∈ MA\Cn. Theorem 6.2 then says that kAke = 0, which means that A is compact. (cid:3) 7. The case p = 2 and open problems As in [21], we can significantly improve Theorem 6.2 when p = 2. If A ∈ F 2 α, let σ(A) be the spectrum of A and let r(A) = max{λ : λ ∈ σ(A)}. Moreover, let σe(σ) be the essential spectrum of A and let re(A) = max{λ : λ ∈ σe(A)}. Theorem 7.1. If A ∈ T α 2 , then kAke = sup x∈MA\CnkAxkF 2 α. α→F 2 Moreover, sup x∈MA\Cn k→∞ sup x∈MA\CnkAk xk r(Ax) ≤ lim 1 k F 2 α→F 2 α! = re(A) with equality if A is essentially normal. Theorem 7.2. If A ∈ T α 2 then following are equivalent: (i) λ /∈ σe(A), (ii) λ /∈Sx∈MA\Cn σ(Ax) and (iii) there is γ > 0 depending only on λ, such that x∈MA\Cnk(Sx − λI)−1kF 2 sup α→F 2 α < ∞ k(Sx − λI)fkα,2 ≥ γkfkα,2 and k(S∗ x − λI)fkα,2 ≥ γkfkα,2 for all f ∈ F 2 α and x ∈ MA\Cn. 37 Theorem 7.3. If A ∈ T α 2 , then [x∈MA\Cn σ(Ax) ⊂ σe(A) The proofs of these results are identical to the proofs of the corresponding results in [21], and in particular depend on the well known fact that r(A) = lim k→∞kAkk 1 k F 2 α→F 2 α = kAkF 2 α→F 2 α whenever A is self adjoint. We will close this paper with a discussion of some open problems. By using ideas in [11, 18], it is very likely that the results in Section 4 hold for more general weighted Fock spaces F p φ for suitable weight functions φ : Cn → R+, where F p φ := {f entire : f (·)e−φ(·) ∈ Lp(Cn, dv)}. Thus, it would be interesting to know if our results hold for other weighted Fock spaces. For the space F 2 α, one should notice that the reproducing kernel and the Gaussian weight behave extremely "nicely" together. One should also notice that this simple fact was crucial in proving many of the results in Sections 2 and 3 (and is in fact crucial for proving compactness results for individual Toeplitz operators or finite sums of finite products of Toeplitz operators on the Fock space, see [3, 15].) Unfortunately, such nice behavior between the reproducing kernel and more general weights rarely holds, and overcoming this would most likely be the most challenging obstacle in extending the results of this paper to more general weighted Fock spaces. is not compact on F 2 Note that one can easily find examples of functions f ∈ L1 α where Bα(f ) vanishes at infinity but T α α (see [4, 3]). It would be f interesting to know if other such examples can be found for F p α when p 6= 2. However, it would be far more interesting to know if one could come up with similar examples for the Bergman space of the ball, weighted or unweighted α is replaced by f ∈ L1(Bn, dvγ) in the weighted (where the condition f ∈ L1 case.) Finally, it would be interesting to know whether the results of this paper hold for the p = 1 or p = ∞ case. While some incomplete results are known in the Bergman space setting when p = 1 or p = ∞ (see [12, 23]), it appears that there are no known results for the Fock space when p = 1 or p = ∞. 38 Acknowledgements The authors would like to thank Karlheinz Grochenig and Kristian Seip for their useful information regarding Section 4. The authors would also like to thank the reviewer for his or her helpful comments that greatly improved the exposition, and would like to thank Brett Wick for interesting comments regarding this paper and for providing us with the preprint [19]. References [1] V. Bargmann, On a Hilbert space of analytic functions and an asso- ciated integral transform. Comm. Pure Appl. Math. 14 (1961), 187-214. [2] W. Bauer, Berezin-Toeplitz quantization and composition formulas. J. Funct. Anal. 256 (2009), no. 10, 3107-3142. [3] W. Bauer, L.A. Coburn, J. Isralowitz, Heat flow, BMO, and the compactness of Toeplitz operators. J. Funct. Anal. 259 (2010), no. 1, 57-78. [4] C.A. Berger, L.A. Coburn, Heat flow and Berezin-Toeplitz esti- mates. Amer. J. Math. 116 (1994), no. 3, 563-590. [5] N. Bourbaki, Elements of mathematics: general topology chapters 1− 4. Springer, 1989 [6] L.A. Coburn, J. Isralowitz, B. Li, Toeplitz operators with BMO symbols on the Segal-Bargmann space. Trans. Amer. Math. Soc. 363 (2011), no. 6, 3015-3030 [7] M. Dostani´c, K. Zhu, Integral operators induced by the Fock kernel. Integral Equations Operator Theory 60 (2008), no. 2, 217-236. [8] M. Englis, Compact Toeplitz operators via the Berezin transform on bounded symmetric domains. Integral Equations Operator Theory 33 (1999), no. 4, 426-455. [9] M. Englis, Density of algebras generated by Toeplitz operator on Bergman spaces. Ark. Mat. 30 (1992), no. 2, 227-243. [10] G. Folland, Harmonic analysis in phase space. The annals of mathe- matics studies, Princeton University press, Princeton, 1989. 39 [11] K. Grochenig, Describing functions: atomic decomposition versus frames. Monatsh. Math. 112 (1991), no. 1, 1-42. [12] P. Galindo, M. Lindstrom, Essential norm of operators on weighted Bergman spaces of infinite order. J. Operator Theory 64 (2010), no. 2, 387399. [13] L. Hormander, An introduction to complex analysis in several vari- ables. Third edition. North-Holland Mathematical Library, 7. North- Holland Publishing Co., Amsterdam, 1990. xii+254 pp. [14] Z. Hu, X. Lv, Toeplitz operators from one Fock space to another. Inte- gral Equations Operator Theory 70 (2011), no. 4, 541-559. [15] J. Isralowitz, Compact Toeplitz operators on the Segal-Bargmann space. J. Math. Anal. Appl. 374 (2011), no. 2, 554-557. [16] J. Isralowitz, K. Zhu, Toeplitz Operators on the Fock space. Integral Equations Operator Theory 66 (2010), no. 4, 593-611. [17] S. Janson, J. Peetre, R. Rochberg, Hankel forms and the Fock space. Rev. Mat. Iberoamericana 3 (1987), no. 1, 61-138. [18] X. Massaneda, P. Thomas, Interpolating sequences for Bargmann- Fock spaces in Cn. Indag. Math. (N.S.) 11 (2000), no. 1, 115-127. [19] M. Mitkovski, B. Wick, The essential norm of operators on Ap α(Bn). Preprint [20] K. Nam, D. Zheng, C. Zhong, m-Berezin transform and compact operators. Rev. Mat. Iberoam. 22 (2006), no. 3, 867-892. [21] D. Su´arez, The essential norm of operators in the Toeplitz algebra on Ap(Bn). Indiana Univ. Math. J. 56 (2007), no. 5, 2185-2232. [22] P. Wojtaszczyk, Banach spaces for analysts. Cambridge Studies in Advanced Mathematics, 25. Cambridge University Press, Cambridge, 1991. xiv+382 pp. [23] T. Yu, Compact operators on the weighted Bergman space A1(ψ). Studia Math. 177 (2006), no. 3, 277-284. 40 [24] K. Zhu, Spaces of holomorphic functions in the unit ball. Graduate Texts in Mathematics, 226. Springer-Verlag, New York, 2005. 41
1207.3428	1	1207	2012-07-14T14:52:04	Similarity of perturbations of the shift and a different product of rational functions	[ "math.FA" ]	Necessary and sufficient conditions are given for the similarity between two perturbations of the (backward) shift by rank one operators, under certain assumptions on the perturbations. The proof of similarity is based on an explicit construction of intertwiners between the perturbations. These intertwiners, in turn, are parametrized by the elements of a certain algebra, with the group of "circle invertible" elements of this algebra giving rise to invertible intertwiners.	math.FA	math	SIMILARITY OF PERTURBATIONS OF THE SHIFT AND A DIFFERENT PRODUCT OF RATIONAL FUNCTIONS LEONEL ROBERT Abstract. Necessary and sufficient conditions are given for the similarity between two perturbations of the (backward) shift by rank one operators, under certain assumptions on the perturbations. The proof of similarity is based on an explicit construction of intertwiners between the perturbations. These intertwiners, in turn, are parametrized by the elements of a certain algebra, with the group of "circle invertible" elements of this algebra giving rise to invertible intertwiners. 1. Introduction Let H 2 denote the Hardy space of the circle. Let U : H 2 → H 2 denote the backward shift operator. The perturbations of U (or U ∗) by a rank one operator have been occasionally studied and shown to have a rich theory (see [Cla72], [Nak93], [CT04]). It is shown in [Rob05] that a large class of perturbations of U by small and/or compact operators are in fact similar to perturbations of U by rank one operators. This motivates the question addressed in this paper: when are two perturbations of U by rank one operators similar? Theorem 1.1 below gives necessary and sufficient conditions for the operators U + r ⊗ φ and U + s ⊗ φ to be similar, under the assumption that r and s are rational functions in H 2 (i.e., with poles outside of the closed unit disc). Although assuming that r and s are rational certainly simplifies the analysis, we will see how even in this case an interesting algebraic structure remains. The unraveling of this structure leads to the solution of the similarity problem. Let us introduce some notation. Let D denote the closed unit disc. Let R(D) denote the rational functions with poles outside D. Let φ ∈ H 2 and r ∈ R(D). For each w < 1, define Γ+(w; r) = whr, φ 1 − wz i, Γ−(w; r) = h φ w − z , ri. These functions are analytic in w. It will be shown below that Γ+(·, r) is a rational function with poles outside D; in particular, it extends analytically to a neighborhood of D. Let ordw(f ) denote the order of the zero of f on w, where f is analytic in a neighborhood of w. Theorem 1.1. Let r, s ∈ R(D). The following propositions are equivalent: (i) The operators U + r ⊗ φ and U + s ⊗ φ are similar. (ii) (a) For each w 6 1, ordw(1 − Γ+(w; r)) = ordw(1 − Γ+(w; s)), and (b) for each w < 1, min(ordw(φ), ordw(1 − Γ−(w; r))) = min(ordw(φ), ordw(1 − Γ−(w; s))). 1 2 LEONEL ROBERT The proof of (ii) ⇒ (i) relies on the construction of certain intertwiners between the oper- ators U + r ⊗ φ, with r varying over R(D) and φ ∈ H 2 fixed. In turn, these intertwiners are described in terms of a "twisted" multiplication on the rational functions. More specifically, define on R(D) the binary operation r × s = zrTφ(s) + zsTφr − Tφ(zrs). Here z : T → T denotes the identity function and Tφ is the co-analytic Toeplitz operator with symbol φ. It is easy to check that r×s is again an element of R(D). It is not at all clear that the operation × is associative, but it will be shown below that this is the case (Section 2). Thus, R(D) becomes an algebra under the multiplication × (and standard addition and scalar multiplication). For each r ∈ R(D), define Kr : H 2 → H 2 by Krf = r × f . The operators I − Kr are intertwiners between perturbations of U : (I − Kr)(U + s ⊗ φ) = (U + (r ◦ s) ⊗ φ)(I − Kr). Here r ◦ s := r + s − r × s is the operation of circle composition in the algebra (R(D), ×). The proof of Theorem 1.1 (ii) ⇒ (i) passes through an analysis of the algebra (R(D), ×) and in particular of its circle invertible elements (Theorem 3.3 and Corollary 3.4). On the other hand, the implication (i) ⇒ (ii) follows from a rather straightforward spectral analysis (Section 4). 2. Intertwiners Let us start by fixing some notation. For each w < 1, let kw = 1/(1 − wz). If f ∈ L2(T), we shall always understand by f (w) the evaluation on w of the harmonic extension of f to D, i.e., f (w) = hf kw, kwi. Let P+ : L2(T) → H 2 denote the orthogonal projection. We denote by Tf the Toeplitz operator on H 2 with symbol f ∈ L2(T), i.e., Tf g = P+(f g). If f is unbounded then Tf is only densely defined (say, on R(D)). However, in this case we will find it useful to regard Tf as a continuous operator from H 2 to the space H(D) of analytic functions in the interior of D, endowed with the topology of uniform convergence on compact sets (see [Sar94, (IV- 12)]). The operator P+ : L2(T) → H(D) is then taken to mean P+(f )(w) := hf, kwi, for w < 1. For w < 1 and n = 0, 1, . . . , let k(n) w = n!zn (1 − wz)n+1 . w = dn dwn kw, i.e., k(n) Observe that k(n) parameter w. Let Sw denote the linear span of {k(0) rational function into simple fractions implies that R(D) = M w is the n-th derivative of kw with respect to the w , . . . }. The decomposition of a w , k(1) Sw. w<1 Let φ ∈ H 2 and w < 1. We have the following formula for evaluating the Toeplitz operator Tφ on k(n) w : (2.1) Tφ(k(n) w ) = dn dwn (φ(w)kw). SIMILARITY OF PERTURBATIONS OF THE SHIFT... 3 This formula is deduced from the case n = 0 -- which is well known -- by repeatedly differenti- ating with respect to w. From this formula we see that Sw and R(D) are both invariant by Tφ. It follows that if r, s ∈ R(D) then r × s := zrTφ(s) + zsTφ(r) − Tφ(zrs) is also in R(D). Let r ∈ R(D). Define Krf = r × f , with f ∈ H 2. Observe that we can make sense of r × f as a function in H(D), bearing in mind the convention stated above for the evaluation of Toeplitz operators with unbounded symbol. Nevertheless, we will show shortly that Kr is in fact a bounded operator on H 2. For each w < 1 and f ∈ H 2, let Γ−(w; f ) := h φ w−z , f i. Lemma 2.1. Let w < 1, n ∈ N, and f ∈ H 2. Then k(n) w × f = zTφ(k(n) w )f + dm dwm (Γ−(w; f ) · kw). Proof. We have w × f = zTφ(k(n) k(n) w )f + zk(n) w Tφ(f ) − Tφ(zf k(n) w ). The first term on the right hand side is already present in the desired formula. Thus, we must deal with the other two terms. We have zk(n) w Tφ(f ) − Tφ(zf k(n) w ) = zk(n) w P+(φf ) − P+(φzk(n) w f ) = P+(k(n) w z(P+ − I)(φf )). Set z(P+−I)(φf ) = f , so that zk(n) w f ). Observe that f ⊥ zH 2. So, the harmonic extension of f to the unit disc is conjugate analytic, i.e., analytic in w. By the same argument used in the derivation of (2.1), we have P+(k(n) dwn ( f (w)kw). On the other hand, w Tφ(f )−Tφ(zf k(n) w ) = P+(k(n) w f ) = dn f (w) = h f , 1 1 − zw i = hz(P+ − I)(φf ), 1 1 − zw i = hf, φ w − z i = Γ−(w; f ). This proves the lemma. (cid:3) Proposition 2.2. Kr : H 2 → H 2 is a bounded operator which is a perturbation of the analytic Toeplitz operator with symbol zTφr by a finite rank operator. Proof. We may reduce ourselves to the case that r = k(n) w for some w < 1 and n ∈ N (since these functions span R(D)). In this case, the proposition follows from the previous lemma. Indeed, observe that f 7→ zTφ(k(n) w ) and that f 7→ di w )f is a Toeplitz operator with symbol zTφ(k(n) dwi Γ−(w; f ) is a bounded linear functional for all i = 0, 1, 2, . . . . (cid:3) Proposition 2.3. Let r ∈ R(D). Then (2.2) (2.3) U Kr − KrU = r ⊗ φ, K ∗ r (φ) = 0. Furthermore, these two equations determine Kr uniquely for given r ∈ R(D) and φ ∈ H 2. 4 LEONEL ROBERT Proof. The verification of (2.2) and (2.3) is straightforward, although somewhat cumber- some. We will sketch the computations here and leave the details to the reader: The following formula is well known and easily established: U Tl − TlU = (U l) ⊗ 1 for all l ∈ H 2. Thus, U TzrTφ − TzrTφU = (U Tzr − TzrU )Tφ = r ⊗ φ. It can be shown by a similar computation that the operator TzTφr − Tzrφ commutes with U . Since Kr = TzrTφ + (TzTφr − Tzrφ), we get (2.2). In order to prove (2.3), we first compute that K ∗ r f = φP+(zrf ) − P+(zrφ)f , for all f ∈ H 2. Then Krφ = φP+(zrφ) − P+(zrφ)φ = 0. Finally, let us show that (2.2) and (2.3) determine Kr uniquely. Suppose that K ′ is a bounded operator that satisfies these equations. Then C := K ′ − Kr commutes with U and satisfies C ∗φ = 0. Since C commutes with U , we have C = Tl, with l ∈ H ∞. So C ∗ = Tl is multiplication by l. But then we cannot have C ∗φ = 0 unless l = 0 (since φ 6= 0). We conclude that C = 0, i.e., K ′ = Kr. (cid:3) Proposition 2.4. R(D) is a commutative algebra under the multiplication × and standard addition and scalar multiplication. The map r 7→ Kr is a representation of this algebra by operators acting on H 2. Proof. It is clear that × is bilinear and commutative. Let us show that it is associative. It is easily verified, using (2.2) and (2.3), that the operators KrKs and Kr×s have the same commutator with U (equal to (r × s) ⊗ φ) and that their adjoints vanish at φ. We conclude by the previous proposition that KrKs = Kr×s. This means that r × (s × f ) = (r × s) × f for all f ∈ H 2. In particular, × is associative. Thus, (R(D), ×) is a commutative algebra over C. Since Kr depends linearly on r, r 7→ Kr is an algebra homomorphism. (cid:3) We will use the notation R×(D) to refer to R(D) regarded as an algebra under ×. Observe that for φ = 1 we get r × s = zrs, and so r 7→ zr is an isomorphism from R×(D) to zR(D) (where the latter is endowed with the standard multiplication). In the next section we will elucidate the structure of R×(D) for an arbitrary φ. Consider on R×(D) the binary operation Proposition 2.5. We have r ◦ s = r + s − r × s. (I − Kr)(U + s ⊗ φ) = (U + (r ◦ s) ⊗ φ)(I − Kr). Proof. This follows at once from (2.2) and (2.3). (cid:3) The preceding proposition implies that if I−Kr is invertible then U +r⊗φ and U +(r◦s)⊗φ are similar. We have (I − Kr)(I − Ks) = I − Kr◦s and I − K0 = I. So, if r is an invertible element of R×(D) with respect to ◦ (where 0 is the neutral element) -- i.e., r ◦ s = 0 for some s ∈ R(D) -- then (I − Kr)(I − Ks) = I, and so I − Kr is invertible. In general, given a ring R the operation a ◦ b 7→ a + b − ab is called circle composition and the collection of invertible elements with respect to this operation is called the circle group of the ring. We arrive to the following corollary: Corollary 2.6. Let r, s ∈ R(D). If there exists a circle invertible element t ∈ R×(D) such that r ◦ t = s then U + r ⊗ φ and U + s ⊗ φ are similar. SIMILARITY OF PERTURBATIONS OF THE SHIFT... 5 3. The algebra R×(D) In this section we elucidate the structure of the algebra R×(D). We then show that r ◦ t = s, for some circle invertible t, if and only if the conditions of Theorem 1.1 (ii) hold. Together with Corollary 2.6, this proves Theorem 1.1 (ii)⇒(i). Lemma 3.1. The map γ+ : R×(D) → R(D) defined by γ+(r) = zTφ(r) is an algebra homomorphism (where R(D) is endowed with the standard multiplication). Proof. The map γ+ may be viewed as the composition r 7→ Kr 7→ zTφ(r). As shown in Proposition 2.4, r 7→ Kr is an algebra homomorphism. On the other hand, Kr is an operator in the Toeplitz algebra and has symbol zTφ(r) (by Proposition 2.2). Thus, Kr 7→ zTφ(r) is simply the symbol map. It follows that γ+ is an algebra homomorphism. (cid:3) Observe that Γ+(w; r) -- as defined in the introduction -- is simply the analytic extension of γ+(r) -- as defined in the proposition above -- to the interior of the unit disc. The range of the homomorphism γ+ is zR(D) (because Tφ maps R(D) onto itself, which in turn can be deduced from (2.1)). So we get a short exact sequence (3.1) We will show below that this short exact sequence splits, and so R×(D) ∼= ker γ+ ⊕ zR(D). But first, let us investigate the ideal ker γ+ further. 0 −→ ker γ+ −→ R×(D) γ+−→ zR(D) −→ 0. We have ker γ+ = (ker Tφ) ∩ R(D). From (2.1), we see that k(n) is a zero of φ of order larger than n. For each a < 1 and N ∈ N, let S N span of k(j) a , with j = 0, . . . , N − 1. Then a ∈ ker Tφ if and only if a a denote the linear (3.2) ker γ+ = ker Tφ ∩ R(D) = M {a<1φ(a)=0} SNa a , a × r ∈ S m where the direct sum is taken over the zeros of φ and Na denotes the order of the zero. By Lemma 2.1, k(m−1) a if a is a zero of φ and m 6 Na. It follows that, in this a is an ideal of R×(D). Thus, the direct sum in (3.2) holds in the ring theoretic case, S m sense, i.e., the different summands are orthogonal to each other with respect to × (indeed, a × SNb SNa b = {0} if a 6= b). Let a be a zero of φ. Let ua = ( z−a 1−az )Na. Let ψ ∈ H 2 denote the function such that φ = uaψ. Observe that ua and ψ are relatively prime, since a is a zero of order Na of φ. Thus, there exists α, β ∈ H 2 such that αψ − uaβ = 1. In fact, we can choose α a rational function in (uaH 2)⊥. Define ea ∈ S Na a by ea = P+(zαua). b ⊆ SNa a ∩ SNb Lemma 3.2. Let a be a zero of φ. (i) If Na > 2 the map k(Na−2) a C[x]/(xNa ). 7→ x extends to an algebra isomorphism from S Na−1 a to (ii) The map ea 7→ 1, k(Na−2) unitization of C[x]/(xNa ). a 7→ x extends to an algebra isomorphism from S Na a to the Proof. (i) It suffices to show that the elements ×m S Na−1 that and that ×Na i=1k(Na−2) a a , with m = 1, 2, . . . , Na − 1, span = 0. These assertions, in turn, will follow once we have shown i=1k(Na−2) a 6 LEONEL ROBERT (1) k(Na−2) (2) k(Na−2) a a a ∈ Sm−1 a × k(m) × ka = 0. From Lemma 2.1, we have \S m−2 a for all m = 1, 2, . . . , Na − 1, and k(Na−2) a d dw Both (1) and (2) above follow from Γ−(a; k(Na−2) a = Γ−(a; k(Na−2) × k(m) a + )k(m) a (ii) Since S Na a /S Na−1 a )w=a 6= 0. is one dimensional, it suffices to show that ea is a unit of S Na dw Γ−(w; k(Na−2) ) = 0 and d a a (and a Γ−(w; k(Na−2) a )w=a · k(m−1) a + . . . . use (i)). We must show that ea × k(m) to Γ−(a; ea) = 1 and dm that a is a zero of Γ−(w; ea) − 1 of order at least Na. Let us compute Γ−(w; ea): for all m 6 Na. By Lemma 2.1, this is equivalent dwm Γ−(w; ea)w=a = 0 for all m = 1, 2, . . . , Na. Thus, we must show a = k(m) a Γ−(w; ea) = h φ w − z , eai = h φ w − z , zαuai = h ψ w − z , zαi = hαψ, 1 1 − zw i = h1 + uaβ, 1 1 − zw i = 1 + ua(w)β(w) = 1 + (cid:16) w − a 1 − aw(cid:17)Na β(w). (cid:3) Recall the exact sequence (3.1). We can now conclude that this sequence splits, since the map γ− : R×(D) → ker γ+ defined by γ−(r) = X {a<1φ(a)=0} r × ea is a right inverse of the inclusion ker γ+ ֒→ R×(D). Thus, R×(D) ∼= ker γ+ ⊕ zR(D). We summarize our findings in the following theorem: Theorem 3.3. The following propositions are true. (γ−,γ+) −→ ker γ+ ⊕ zR(D) is an isomorphism. (i) R×(D) (ii) ker γ+ = L{a<1φ(a)=0} SNa (iii) For each a < 1, zero of φ, there is an isomorphism from SNa r 7→ r × ea. a , where the projection onto the a-th summand is given by to (C[x]/(xNa ))∼ 7→ x. Here (C[x]/(xNa ))∼ denotes the unitization of a such that ea 7→ 1 and k(Na−2) C[x]/(xNa ). a We are now ready to describe when two elements of R×(D) are in the same orbit of the action of the group of circle invertible elements. Corollary 3.4. (i) The element t ∈ R×(D) is circle invertible if and only if 1 − γ+(t) is invertible in R(D) and Γ−(a; t) 6= 1 for any a < 1, zero of φ. (ii) Let r, s ∈ R×(D). There exists a circle invertible element t ∈ R×(D) such that r◦t = s if and only if r and s satisfy conditions (a) and (b) of Theorem 1.1. Proof. (i) By the previous theorem, t ∈ R×(D) is circle invertible if γ−(t) ∈ ker γ+ and γ+(t) ∈ zR(D) are circle invertible. The latter condition is equivalent to 1 − γ+(t) being invertible in R(D), while the former is equivalent to ea − ea × t ∈ S Na a being invertible for all a < 1, zero of φ. An element of (C[x]/(xNa ))∼ is invertible if and only if it is not in the nil ideal C[x]/(xNa ). Applied to S Na in kNa−1 a × t being different from 1. By Lemma 2.1, this is the same as Γ−(a; t) 6= 1. a , this is equivalent to the coefficient of k(Na−1) a SIMILARITY OF PERTURBATIONS OF THE SHIFT... 7 (ii) In order for r and s to be related by circle invertible elements, we must have that (1) γ+(r) and γ+(s) are related by a circle invertible element of zR(D), (2) for each a < 1, zero if φ, ea × r and ea × t are related by a circle invertible element of SNa a . The first condition is equivalent to 1 − γ+(r) and 1 − γ+(s) differing by an invertible factor of R(D). This is equivalent to condition (a) of Theorem 1.1 (ii). The second condition is equivalent to ea − ea × r and ea − ea × s being both invertible or having the same order of nilpotency for each a zero of φ. (This criterion is easily verified in (C[x]/(xNa ))∼.) In turn, this is equivalent to k(Na−1) a − k(Na−1) a × r ∈ S m a ⇔ k(Na−1) a − k(Na−1) a × s ∈ S m a , for all m = 1, 2, . . . , Na. By Lemma 2.1, this is equivalent to condition (b) of Theorem 1.1 (ii). (cid:3) Proof of Theorem 1.1 (ii)⇒(i). This follows at once from Corollary 2.6 (ii) and Corollary 3.4. (cid:3) 4. Proof of (i) implies (ii) In this section we prove the implication (i)⇒(i) of Theorem 1.1. We start with a lemma. Lemma 4.1. Let A be a bounded operator on a Hilbert space and let B be a left inverse for A, i.e., BA = I. Let A = A + f ⊗ g. (i) We have ker A 6= 0 if and only if ABf = f and 1 + hBf, gi = 0. In this case ker A = span{Bf }. (ii) Assume that ker A 6= 0. For k > 1 we have that ker Ak 6= ker Ak−1 if and only In this case if ABif = Bi−1f for 1 < i 6 k and hBif, gi = 0 for 1 < i 6 k. ker Ak = span{Bif 1 6 i 6 k}. Proof. (i) This is a straightforward computation (left to the reader). (ii) Since ker A has dimension 1 (by (i)), the dimension of ker Ak, for k = 1, 2, . . . , grows by 1 and then becomes stationary. So dim ker Ak 6 k. If hBif, gi = 0 for 1 < i 6 k (and -1 for i = 1) and ABif = Bi−1f then we easily verify that span{Bif 1 6 i 6 k} ⊂ ker Ak. Also, the vectors on the left side are linearly independent (they form a Jordan chain). So we must have equality. This also shows that ker Ak−1 6= ker Ak. We will prove the other implication by induction on k. Assume it is true for k. Suppose that ker Ak+1 6= ker Ak. Since A maps ker Ak+1 surjectively onto ker Ak, there exists x such that Ax = Bkf . That is, Ax + f hx, gi = Bkf . Multiplying by B we get x + Bf hx, gi = Bk+1f . It follows that ABk+1f = Bkf . This in turn implies that ABk+1f + f hBk+1f, gi = Bkf . Multiplying by B and using that Bf 6= 0 we get hBk+1f, gi = 0. Then ABk+1f = Bkf . (cid:3) Proposition 4.2. Let r ∈ R(D) and φ ∈ H 2. Set U + r ⊗ φ = Ur. (i) Let w 6 1. Then dim ker(1 − wUr)k = min(k, ordw(1 − Γ+(w; r)). (ii) Let w < 1. Then dim ker(U ∗ r − w)k = min(k, ordw(φ), ordw(1 − Γ−(w; r))). 8 LEONEL ROBERT Proof. (i) We have that 1 − wUr = (1 − wU ) − r ⊗ wφ. Assume first that w < 1. We can apply the previous lemma with A = 1 − wU and B = (1 − wU )−1. We get that dim ker(1 − cU ρ)k = k if and only if h(1 − wU )−1r, wφi = 1 and h(1 − wU )−ir, wφi = 0 for 1 < i 6 k. This leads to ordw(1 − Γ+(w; r)) > k. The case w = 1 can be handled similarly. In this case we set A = 1−wU and B = T 1 1−wz . Observe that, although B is not bounded, it maps R(D) surjectively onto itself. Also, A maps R(D) into itself, and BAf = f for all f R(D). This makes the computations of the previous lemma still applicable, since it is easy to check that in this case ker Ak ⊆ R(D), for all k. So we may restrict our computations to R(D) from the outset. (ii) We have that U ∗ r − wI = (U ∗ − wI) + φ ⊗ r. Thus, we can apply the previous lemma with A = U ∗ − wI and B = T 1 . We get that dim ker(U ∗ r − w)k = k if and only if z−w (1) (U ∗ − wI)(T 1 z−w )iφ = (T 1 z−w (2) hT 1 z−w φ, ri = 1, and h(T 1 z−w )i−1φ for 1 6 i 6 k, )iφ, ri = 0 for 1 < i 6 k. The first condition is satisfied if and only if ordw(φ) > k, and the second if and only if ordw(1 − Γ−(w; r)) > k. This proves (ii). (cid:3) Proof of Theorem 1.1(i) ⇒ (ii). For each k = 1, 2, . . . , the quantities dim ker(1−wUr)k and dim ker(U ∗ r − w)k are similarity invariants. This, combined with the previous proposition, proves the implication (i) ⇒ (ii) in Theorem 1.1. (cid:3) References [CT04] Gilles Cassier and Dan Timotin, Power boundedness and similarity to contractions for some per- turbations of isometries, J. Math. Anal. Appl. 293 (2004), no. 1, 160 -- 180. [Cla72] Douglas N. Clark, One dimensional perturbations of restricted shifts, J. Analyse Math. 25 (1972), 169 -- 191. [Nak93] Yoshihiro Nakamura, One-dimensional perturbations of the shift, Integral Equations Operator The- ory 17 (1993), no. 3, 373 -- 403. [Rob05] Leonel Robert, Similarity of perturbations of Hessenberg matrices, J. Operator Theory 54 (2005), no. 1, 125 -- 136. [Sar94] Donald Sarason, Sub-Hardy Hilbert spaces in the unit disk, University of Arkansas Lecture Notes in the Mathematical Sciences, 10, John Wiley & Sons Inc., New York, 1994. A Wiley-Interscience Publication.
1606.04924	1	1606	2016-06-15T19:34:22	From Frazier-Jawerth characterizations of Besov spaces to Wavelets and Decomposition spaces	[ "math.FA" ]	This article describes how the ideas promoted by the fundamental papers published by M. Frazier and B. Jawerth in the eighties have influenced subsequent developments related to the theory of atomic decompositions and Banach frames for function spaces such as the modulation spaces and Besov-Triebel-Lizorkin spaces. Both of these classes of spaces arise as special cases of two different, general constructions of function spaces: coorbit spaces and decomposition spaces. Coorbit spaces are defined by imposing certain decay conditions on the so-called voice transform of the function/distribution under consideration. As a concrete example, one might think of the wavelet transform, leading to the theory of Besov-Triebel-Lizorkin spaces. Decomposition spaces, on the other hand, are defined using certain decompositions in the Fourier domain. For Besov-Triebel-Lizorkin spaces, one uses a dyadic decomposition, while a uniform decomposition yields modulation spaces. Only recently, the second author has established a fruitful connection between modern variants of wavelet theory with respect to general dilation groups (which can be treated in the context of coorbit theory) and a particular family of decomposition spaces. In this way, optimal inclusion results and invariance properties for a variety of smoothness spaces can be established. We will present an outline of these connections and comment on the basic results arising in this context.	math.FA	math	From Frazier-Jawerth characterizations of Besov spaces to Wavelets and Decomposition spaces H. G. Feichtinger and F. Voigtlaender Abstract. This article describes how the ideas promoted by the fundamental papers published by M. Frazier and B. Jawerth in the eighties have influenced subsequent developments related to the theory of atomic decompositions and Banach frames for function spaces such as the modulation spaces and Besov- Triebel-Lizorkin spaces. Both of these classes of spaces arise as special cases of two different, general constructions of function spaces: coorbit spaces and decomposition spaces. Coorbit spaces are defined by imposing certain decay conditions on the so-called voice transform of the function/distribution under consideration. As a concrete example, one might think of the wavelet transform, leading to the theory of Besov-Triebel-Lizorkin spaces. Decomposition spaces, on the other hand, are defined using certain de- compositions in the Fourier domain. For Besov-Triebel-Lizorkin spaces, one uses a dyadic decomposition, while a uniform decomposition yields modulation spaces. Only recently, the second author has established a fruitful connection between modern variants of wavelet theory with respect to general dilation groups (which can be treated in the context of coorbit theory) and a particular family of decomposition spaces. In this way, optimal inclusion results and invariance properties for a variety of smoothness spaces can be established. We will present an outline of these connections and comment on the basic results arising in this context. 1. The Frazier-Jawerth Atomic Characterizations 1.1. Introductory statement. This article is part of a series of notes (see e.g. [28, 27, 26]), which describe the role of different function spaces, their various characterizations and their possible applications from a "postmodern viewpoint", emphasizing the importance of the concept of Banach frames (first defined in [46]). For this development, a series of papers by Michael Frazier and Bjoern Jawerth ([36, 37, 38]) have been of great relevance. Hence we will try to demonstrate how the ideas of these papers lived on and expanded in the subsequent decades. 2010 Mathematics Subject Classification. Primary 42C15, 46E35, 42C40. 1 1.2. Notation. We will employ the following notation: For a group G and any function f : G → S, for some set S, we define Lxf : G → S, y (cid:55)→ f (x−1y), Rxf : G → S, y (cid:55)→ f (yx), x, y ∈ G. In the special case of the (abelian) group Rd, we also write Tx := Lx. Furthermore, for f : Rd → C and ξ ∈ Rd, we define the modulation of f by ξ as Mξf : Rd → C, x (cid:55)→ e2πix·ξ · f (x). Finally, for h ∈ GL(Rd), we define the (L2 normalized) dilation of f by h as Dhf : Rd → C, x (cid:55)→ det h −1/2 · f (h−1x). For the special case h = a · id with a ∈ R \ {0}, we also write Da := Da·id. For the Fourier transform, we use the normalization f (x)e−2πix·ξdx, f ∈ L1(Rd). Rd (cid:90) Ff (ξ) := (cid:98)f (ξ) := (cid:90) −1h(x) = F It is well-known that the Fourier transform extends to a unitary automorphism of L2(Rd), where the inverse is the unique extension of the operator F −1, given by h(ξ)e2πix·ξdξ, Rd h ∈ L1(Rd). λ(M ) denotes the d-dimensional Lebesgue measure of a (measurable) set M ⊂ Rd. Finally, for x ∈ R, we write x+ := max{x, 0}. 1.3. The Essence of the work of Frazier-Jawerth. To the best of our knowledge, the influential papers [36, 37, 38] have been the first to fully charac- terize two families of Banach spaces of tempered distributions, namely the Besov spaces (Bs ), by corresponding growth and summability conditions on a suitably defined sequence of coefficients. These coefficients depend linearly on the function/distribution under consideration. Similar atomic representation theorems had been realized only a few years earlier in the context of harmonic function spaces (see e.g. [9, 64]) ) and the Triebel-Lizorkin spaces (F s p,q(Rd),(cid:107)·(cid:107)Bs p,q(Rd),(cid:107)·(cid:107)F s p,q p,q In a more modern terminology (going back to the work of K. Grochenig [46]), one could say that Frazier and Jawerth established specific, but rather concrete Banach frame expansions for these two families of function spaces, starting from the characterization of these spaces via dyadic partitions of unity on the Fourier transform side. This characterization in turn is based on the description of Besov- Triebel-Lizorkin spaces via dyadic decompositions (see [62, 68, 69]). lowing situation: In this general theory of Banach frames, one has-roughly speaking-the fol- rich) atoms (gi)i∈I , one obtains atomic representations of the form f =(cid:80) (i) Starting from a fixed family of (possibly non-orthogonal, but sufficiently i∈I cigi, typically with convergence in the norm of (B, (cid:107)·(cid:107)B) (or at least w∗-convergence linear coefficient mappings f (cid:55)→ ci = ci(f ) satisfying f =(cid:80) (ii) The coefficients are obtained in a linear way, i.e., there are suitable bounded i∈I cigi. These map- pings are often realized as scalar products with respect to a suitable family of atoms (hi)i∈I , forming a so-called dual frame. Thus, ci = (cid:104)f, hi(cid:105), where the "scalar product" can be viewed as an extension of the Hilbert-space scalar product. for the case of dual spaces). 2 (iii) Although the representation is by far not unique, there is a high degree of compatibility between the membership of f in one of the spaces under consideration and membership of the coefficients (ci)i∈I in a suitable Banach space of sequences, in fact, in a suitable solid BK space. These spaces have the property that sequences which are smaller in terms of absolute values-in a coordinate-wise sense-also have a smaller norm. Precisely, it is claimed that for a function space X (from a certain class), one can find a BK space Y ≤ CI , a bounded coefficient map C : X → Y , f (cid:55)→ (ci(f ))i∈I , and a synthesis mapping R : Y → X, c (cid:55)→ i∈I cigi which is a left inverse for C. More concretely, the ϕ-transform as introduced by Frazier and Jawerth in [36, 37, 38] is a fixed, linear transformation Sϕ : f (cid:55)→ ((Sϕf )Q)Q∈Q with the benefit that a variety of function spaces can be characterized in terms of the size of Sϕf , i.e. by (cid:107)Sϕf(cid:107)Y for suitable solid sequence spaces Y . Solidity of Y formally encodes the requirement that only the size of the coefficients should be important, so that there is no cancellation between the different coefficients. (cid:80) The class of spaces for which a characterization in terms of the ϕ-transform is possible includes the classes of Besov spaces Bs p,q and the Triebel-Lizorkin spaces F s p,q. Note that both of these classes of spaces come in two variants: homogeneous and inhomogeneous spaces. The theory developed by Frazier and Jawerth applies to both of these subclasses. For concreteness, we will concentrate in the sequel on the inhomogeneous spaces, which have the advantage of being modulation invariant. To describe the ϕ-transform and the resulting decomposition results more pre- cisely, we begin with the index set Q, which is the set of all dyadic cubes with side-length ≤ 1. Here, by definition (cf. [38]), a dyadic cube is a set of the form Qν,k =(cid:8)x ∈ Rd(cid:12)(cid:12)∀i ∈ {1, . . . , d} : 2−νki ≤ xi < 2−ν(ki + 1)(cid:9) for arbitrary ν ∈ Z and k ∈ Zd. Given such a cube Q = Qν,k, we define the lower left corner of Q as xQ := 2−νk and the side length of Q as (cid:96)(Q) := 2−ν. To distinguish cubes with different side lengths, we also define for ν ∈ Z the set Qν :=(cid:8)Q ∈ Q (cid:12)(cid:12) (cid:96)(Q) = 2−ν(cid:9) . Finally, for an arbitrary function ϕ : Rd → C and Q = Qν,k, we let ϕQ(x) := 2νd/2 · ϕ(2νx − k) = 2−νd/2 · ϕν(x − xQ), where ϕν(x) := 2νd · ϕ(2νx). Note that if ϕ is "concentrated" in [0, 1]d, then ϕQ is "concentrated" on Q. Furthermore, we have (cid:107)ϕQ(cid:107)L2 = (cid:107)ϕ(cid:107)L2. Now, the ϕ-transform is defined using two analyzing windows ϕ, ϕ0 ∈ S(Rd). The basic assumption regarding these windows is that there are "dual windows" ψ, ψ0 ∈ S(Rd) satisfying the following (cf. [38, eqs. (12.1)–(12.2) and (2.1)–(2.3)]) supp(cid:99)φ0,(cid:99)ψ0 ⊂ supp(cid:98)ϕ,(cid:98)ψ ⊂ (cid:98)ϕ(2−νξ) · (cid:98)ψ(2−νξ) = 1 (cid:8)ξ ∈ Rd(cid:12)(cid:12)ξ ≤ 2(cid:9) , (cid:8)ξ ∈ Rd(cid:12)(cid:12) 1/2 ≤ ξ ≤ 2(cid:9) , ∀ξ ∈ Rd. (1.1) (1.2) (1.3) (cid:99)ϕ0(ξ) ·(cid:99)ψ0(ξ) + ∞(cid:88) ν=1 Under this assumption, the (inhomogeneous) ϕ-transform Sϕ is the analysis operator of the frame, so that the ϕ-transform Sϕf = ((Sϕf )Q)Q∈Q of a tempered 3 (cid:40) distribution f ∈ S(cid:48)(Rd) is defined by (Sϕf )Q := Finally, the (formal) inverse of Sϕ-the corresponding synthesis operator-is (cid:104)f, ϕQ(cid:105), (cid:104)f, ϕ0 Q(cid:105), (cid:88) Q∈Q0 if (cid:96)(Q) < 1, if (cid:96)(Q) = 1. ∞(cid:88) (cid:88) ν=1 Q∈Qν Tψ[(sQ)Q∈Q] := sQψ0 Q + sQψQ. Now, as shown in [38, equation (12.4)], we have f = TψSϕf for all f ∈ S(cid:48)(Rd), with convergence of the series defining Tψ(Sϕf ) in the topology of S(cid:48)(Rd). The main result of Frazier and Jawerth concerning inhomogeneous Triebel- Lizorkin spaces using the ϕ-transform reads as follows (cf. [38, Theorem 12.2]): Theorem 1. For s ∈ R, 0 < p < ∞ and 0 < q ≤ ∞, the operators are bounded, with Tψ ◦ Sϕ = id on F s retract of f s and F s p,q is a p,q, i.e., it can be identified with a complemented subspace of f s p,q. p,q and Tψ : f s p,q. Hence, (cid:107)f(cid:107)F s p,q → F s p,q (cid:16) (cid:107)Sϕf(cid:107)f s p,q Sϕ : F s p,q → f s p,q (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:16) p,q is defined as the set of all se- In the above theorem, the solid BK-space f s quences c = (cQ)Q∈Q ∈ CQ for which the (quasi)-norm is finite. Here, we used the symbol (cid:102)χE := (λ(E))−1/2 · χE for the L2-normalized version of χE, for any measurable set E ⊂ Rd of finite, positive measure. −s · cQ · (cid:102)χQ In addition to the above result for Triebel-Lizorkin spaces, [36] provides the (cid:107)c(cid:107)f s [(cid:96) (Q)] Q∈Q := p,q (cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q(Q) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Lp( dx) (cid:17) following analogue characterization of Besov spaces. Theorem 2. For 0 < p, q ≤ ∞ and s ∈ R, the operators are bounded, with Tψ ◦ Sϕ = id on Bs retract of bs p,q, and can be identified with a complemented subspace of bs p,q. and Bs p,q is a and Tψ : bs p,q. Hence, (cid:107)f(cid:107)Bs p,q → Bs p,q (cid:16) (cid:107)Sϕf(cid:107)bs p,q p,q Sϕ : Bs p,q → bs p,q The solid BK-space bs ing quasi-norm is finite: p,q consists of all c = (cQ)Q∈Q ∈ CQ for which the follow- (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:32)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16) (cid:107)c(cid:107)bs p,q := [(cid:96) (Q)]d( 1 p− 1 2 )−s cQ Q∈Qν (cid:17) (cid:33) (cid:13)(cid:13)(cid:13)(cid:13)(cid:96)p(Qν ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q(N0) . ν∈N0 Similar results have later been described in the books of H. Triebel, who showed that multivariate wavelet orthonormal bases are not just bases for the Hilbert space L2(Rd), but also (in a modern terminology) Riesz projection bases for the corresponding solid BK-spaces. This is another way of saying that the unitary isomorphism between L2(Rd) and (cid:96)2(I)-which is induced by the orthonormal ba- sis (ϕi)i∈I -extends to an isomorphism between Banach spaces and their discrete counterparts, for a whole family of spaces and with uniform bounds. But the exis- tence of orthonormal wavelet bases was established (in [56, 57, 17]) only after the appearance of the paper [36] of Frazier and Jawerth. 4 In summary, these "decomposition" results of Frazier and Jawerth imply a variety of statements: (1) A consistency statement: Any two (admissible) window-families ϕ1, ϕ0 ϕ2, ϕ0 ϕ-transform, i.e., we have for i = 1, 2: 1 and 2 yield the same space by imposing certain decay conditions on the associated (cid:8)f ∈ S(cid:48)(Rd)(cid:12)(cid:12) Sϕif ∈ f s p,q (cid:9) = F s p,q. (2) A decomposition (or representation) statement: (cid:88) Q∈Q f = sQ(f ) · ψQ, f ∈ F s p,q, with (sQ(f ))Q∈Q = Sϕf ∈ f s p,q and atoms ψQ of the special form ψQ(x) = 2−νd/2 · ψν(x − xQ) = [π(xQ, 2−ν)ψ](x) for (cid:96)(Q) = 2−ν, where π(x, a) := TxDa is the quasi-regular representation on L2(Rd) of the ax + b group. This was already observed in [38, Remark 3.2]. (3) The two parts of the representation are all valid for a certain range of spaces, not just for a single Hilbert space or one individual Banach space. It is also worth noting that, if one restricts the construction of Frazier and Jaw- erth to the setting of Hilbert spaces, one finds that they are implicitly establishing that the set of analyzing atoms forms a frame; but the elements used to perform the reconstruction are not the elements of the (canonical) dual frame. (4) The concrete BK-spaces f s p,q arising in this context are not just abstract BK-spaces (Banach spaces of numerical sequences, whose coordinates depend con- tinuously on the sequence). In addition, they are Banach lattices (following the terminology of Luxemburg-Zaanen) resp. so-called solid BK-spaces: With each se- quence d = (dQ)Q∈Q ∈ f s p,q, also any sequence b = (bQ)Q∈Q with bQ ≤ dQ for all Q ∈ Q belongs to f s p,q ≤ (cid:107)d(cid:107)f s The same statements remain true with Bs p,q. The above observations are similar to the main properties of coorbit spaces, as p,q, with (cid:107)b(cid:107)f s p,q and bs p,q instead of F s p,q and f s . p,q introduced by Feichtinger and Grochenig in [30, 31, 32], see also Section 2. Remark. While the ϕ-transform and the resulting atomic decompositions for Besov (resp. Triebel-Lizorkin) spaces grew out of the characterization of these spaces using Littlewood-Paley theory-essentially established by the earlier work of J. Peetre and H. Triebel-it turned out that they were also fore-runners for the characterization of these function spaces via wavelet frames. 1.4. Modulation Spaces. One of the earliest variations of the atomic de- composition method proposed by Frazier and Jawerth arose in connection with so-called modulation spaces. These spaces have been introduced with the idea of describing smoothness for functions over locally compact Abelian groups (LCA groups) via uniform (instead of dyadic) decompositions on the Fourier transform side. The first report [20] issued in 1983 (see also [24] for an expanded form) came too early to be appreciated at the time, certainly also because it was formulated in the most general setting. A large number of references concerning these spaces can be found in [25]. Nowadays, modulation spaces are viewed as natural domains for time-frequency analysis and certain families of pseudo-differential operators. 5 Although designed from the very beginning at the most general level, a specific subfamily of modulation spaces, namely the modulation spaces(cid:0)M s (cid:1), p,q(Rd), (cid:107)·(cid:107)M s with 1 ≤ p, q ≤ ∞, s ∈ R, are most similar to the corresponding family of Besov spaces (Bs ), with the same parameters. In fact (only) for the case p = q = 2 these spaces coincide, while otherwise they are different (cf. the PhD thesis of P. Grobner, [45]). By now, there is an extensive literature concerning the inclusion relation between different types of modulation spaces resp. more general decomposition spaces, see e.g. [58, 66, 65, 52, 70]. p,q(Rd),(cid:107)·(cid:107)Bs p,q p,q A Frazier-Jawerth type atomic characterization of modulation spaces has first been given in [22] (received by the editors in Oct. 1986). These results are quite similar to those of Frazier-Jawerth, although no direct reference is made to their papers, obviously because their results have not been used. Instead, the argument is based on a combination of Shannon's theorem and techniques from the theory of Wiener amalgam spaces developed in the early 80s ([19, 23]). These techniques provide refined variants of Poisson's formula, which are used to obtain what is nowadays called a Gabor characterization of modulation spaces. To emphasize the analogy between [36, 37] and [22], we note the following: In each case, suitable partitions of unity are used, which are bounded families within (cid:1). Further, the building blocks are carefully the Fourier algebra(cid:0) FL1(Rd), (cid:107)·(cid:107)F L1 chosen by the authors in a specific way, to achieve the goal of atomic representation. In contrast to the last point, the typical question treated e.g. in the context of Gabor analysis-or more generally coorbit theory-is the following: Given some (structured) family of atoms (so without chance of the user to match the atoms with the function spaces), can one still verify frame properties and compute a dual frame, or at least guarantee the existence of such a dual frame? Compared to the situation in wavelet theory, one can say: The possibility of generating orthonormal wavelet bases came certainly as a surprise, especially to Yves Meyer, who tried to prove the converse, but ended up with his first con- struction of such a basis. His negative expectations were probably based on his knowledge of the Balian-Low Theorem which excludes the existence of Riesz bases (and thus also of orthonormal bases) of Gaborian type with "good atoms" which are well concentrated in the time-frequency sense ([2, 3]). The rapid development of wavelet theory, starting with [56], is in large parts due to the possibility of having orthonormal wavelet bases which may be difficult to construct, but which are easy to use. The construction of compactly supported orthonormal wavelet bases with a given amount of smoothness by Ingrid Daubechies ([17]) was one of the main reasons why the Frazier-Jawerth decomposition method was not pursued too much for a while. A crucial property of both orthonormal wavelets as well as the Schwartz build- ing blocks for the Frazier-Jawerth decompositions was the fact that from the very beginning, the canonical isomorphism between the Hilbert space (cid:0)L2(Rd), (cid:107)·(cid:107)2 (cid:1) (cid:1) extends automatically to an isomorphism between the classical and (cid:0)(cid:96)2, (cid:107)·(cid:107)2 function spaces (those from the family of Besov-Triebel-Lizorkin type, including H 1(Rd) and BM O) and (subspaces of) the corresponding solid BK-spaces (typi- cally weighted mixed-norm sequence spaces). It is interesting to note that the sequence spaces used for the Frazier-Jawerth characterizations and the orthonormal wavelet systems are more or less the same, once one identifies the collection of cubes with different centers with the unique 6 affine transformation (from the ax + b-group) needed to obtain it from a symmetric standard cube. The common structure of all the different types of atomic charac- terizations going back to Frazier-Jawerth is the identification of a family of Banach spaces (of functions or distributions) via some coefficient mapping with suitable closed and complemented subspace of the corresponding family of solid BK-spaces. 2. Coorbit Theory Group representation theory finally provided such a link: 2.1. Unifying Approach to the CWT and the STFT. Once the theory of wavelets started to gain momentum, with the classical function spaces being characterized via the Frazier-Jawerth decompositions or via suitable (orthonormal) wavelet systems and once it was understood that there are similar characterizations for modulation spaces via Gabor expansions, it became a natural question to ask whether these two signal representation methods have anything further in common. in both (as well as other) cases there is a (square) integrable group representation of some locally compact group, acting in an irreducible way (more or less) on a given Hilbert space. In the language of group representation theory, the relevant group for the theory of wavelets is the affine group (also called "ax + b"-group), acting by dilations and translations on L2(Rd). On the other hand, for the theory of modulation spaces or Gabor expansions, the group acting in the background is the (reduced) Heisenberg group Hd, acting on L2(Rd) by translations and modulations. 2.2. The Essence of Coorbit Theory. The core of coorbit theory, as out- lined in [31, 32], is to generate-from a given unitary group representation π of some locally compact group G on a Hilbert space H-a whole family of spaces Co(Y ), the so-called coorbit spaces (for (G , π)). For large subfamilies of these spaces, coorbit theory provides atomic decompositions which allow to characterize elements of the space Co(Y ) by the fact that they can be described using coefficients in the solid BK-spaces Y d which is naturally associated with Co(Y ). The first step for describing Co(Y ) is to define the voice transform Vg, by (Vgf )(x) := (cid:104)f, π(x)g(cid:105), maps H isometrically[48] into(cid:0)L2(G ), (cid:107)·(cid:107)2 If g is suitably chosen (i.e. satisfying (cid:107)Vgg(cid:107)L2 = (cid:107)g(cid:107)H), the mapping f (cid:55)→ Vgf invariant subspace of L2(G ). Note that this crucially uses that the representation π is (square) integrable and irreducible. (cid:1), so that H0 := Vg(H) is a closed, left- x ∈ G , f, g ∈ H. The name "voice transform" goes back to the paper [50], where this name is used synonymously with the term "cycle-octave transform" for what later became the continuous wavelet transform (CWT). Since in that paper both the Gabor and the wavelet case have been addressed, this terminology was taken into the general coorbit theory. It was also used later on in the context of the Blaschke group by Margit Pap and Ferenc Schipp (see e.g. [61, 59, 60]). Given the voice transform Vg and a translation invariant, solid BF-space1 (Y , (cid:107)·(cid:107)Y ) on G , we would like to define the associated coorbit space Co(Y ) as Co(Y ) := {f ∈ H Vgf ∈ Y }, with the norm (cid:107)f(cid:107)Co(Y ) := (cid:107)Vgf(cid:107)Y . But this will in general not yield a complete space. 1This means that (Y , (cid:107) · (cid:107)Y ) is a space of measurable functions on G such that if f ∈ Y and if g : G → C is measurable with g ≤ f almost everywhere, then g ∈ Y with (cid:107)g(cid:107)Y ≤ (cid:107)f(cid:107)Y . 7 Recalling that the classical function spaces (like Besov spaces) are defined as subsets of the space S(cid:48)(Rd) of tempered distributions and not of L2(Rd), we thus have to find a suitable replacement for the class of Schwartz functions in the present generality. In this setting, however, there will (in general) be no "universal" class of test functions like the Schwartz space. Instead, given the solid BF-space (Y , (cid:107)·(cid:107)Y ), one has to choose a suitable class of test functions specific to (Y , (cid:107)·(cid:107)Y ). For this, one first needs to choose a so-called control weight w : G → (0,∞) for the space (Y , (cid:107)·(cid:107)Y ). This weight depends in a certain way on the operator norms of left- and right translation on (Y , (cid:107)·(cid:107)Y ). For the sake of brevity we omit these details. Note, however, that w is assumed to be submultiplicative (i.e., w(xy) ≤ w(x) · w(y)). Given such a control weight w, we introduce the space w := {f ∈ H Vgf ∈ L1 H1 w}, where w = {f f w ∈ L1(G )}, L1 ∠ (2.1) of H1 w. w- w is independent of 0 (cid:54)= g ∈ Aw(G ) :=(cid:8)g ∈ H Of course, the "analyzing window" g-which is used to define the space H1 which will play the role of "test functions" in the present setting; the coorbit spaces will thus be subspaces of the (anti)dual space (H1 w) needs to be chosen carefully. As shown in [30, 31], the space H1 the precise choice of g, as long as (cid:9) . (2.2) The space Aw is called the space of analyzing windows. A crucial assumption for the applicability of coorbit theory is that Aw (cid:41) {0} is nontrivial. Later on we will need the class of better vectors Bw ⊂ Aw, whose precise definition we omit. For the initiated reader, we note that g ∈ Bw needs to satisfy Vgg ∈ W R(L∞, L1 Before defining coorbit spaces in full generality, we need to extend the voice ∠ transform to the reservoir (H1 w, the (generalized) w) voice transform w), where W R denotes a (right-sided) Wiener amalgam space. : Due to the π-invariance of H1 (cid:12)(cid:12) Vgg ∈ L1 w (Vgf )(x) = (cid:104)f, π(x)g(cid:105), x ∈ G , ∠ of f ∈ (H1 w) as (cid:104)f, g(cid:105) = f (g), which extends the usual scalar product on H. with respect to g ∈ H1 With all this, we define the corresponding coorbit space Vgf ∈ Y }, Co(Y ) := {f ∈ (H1 w) ∠ w is well-defined. Here, the pairing is understood with the usual natural norm. Again one can show that the space is independent of the analyzing window 0 (cid:54)= g ∈ Aw. Another main result of coorbit theory concerns the atomic decomposition claim: Theorem 3. For any 0 (cid:54)= g ∈ Bw there exists a neighborhood U of the identity e ∈ G such that the following is true: For any well-separated2 family (xi)i∈I in G which is U -dense, the family (π(xi)g)i∈I defines a Banach frame for Co(Y ). More precisely: There exists a solid BK-space Y d = Y d((xi)i∈I ) ≤ CI and a bounded linear mapping C : f (cid:55)→ (ci(f ))i∈I from Co(Y ) to Y d, such that f = R(Cf ) = ∀ f ∈ Co(Y ), with (unconditional) convergence in (Co(Y ),(cid:107)·(cid:107)Co(Y )), if H1 (and in the w∗-sense otherwise). Furthermore, R : Y d → Co(Y ) is bounded. w is dense in Co(Y ) ci(f )π(xi)g, (cid:88) i∈I 2This means that (xi)i∈I is the finite union of uniformly separated sets in G . 8 In comparison with the original Frazier-Jawerth approach we note the following: (a) In the FJ-approach both the analyzing and the synthesizing vector are specified by the construction; although there is some freedom in the construction, the two elements have to be chosen in a very specific way. (b) In contrast, coorbit theory provides a much larger reservoir of atoms; in fact, coorbit theory shows that one can reconstruct the complete voice transform Vgf (and hence the element f ∈ Co(Y )) from its samples (Vgf (xi))i∈I , as long as g ∈ Bw and as long as the samples are taken densely enough. (c) Both theories provide a retraction between a family of function spaces and the corresponding solid BK-spaces. Since the ax+b-group and the (reduced) Heisen- berg group both have two unbounded variables, it is natural to use mixed-norm conditions on the variables. The corresponding discrete variants are discrete mixed-norm spaces, as long as the sampling points (xi)i∈I form a product set. For a rearrangement invariant space Y , the sequence space Yd is just the "natural" discrete variant; e.g. (Lp)d = (cid:96)p. (d) In both the Frazier-Jawerth-theory and the coorbit approach, one has the choice to describe them in full generality or for a family of spaces. The published versions of Frazier-Jawerth-theory describe the decompositions in the setting of tempered distributions; consequently they are able to characterize Besov spaces for all real parameters s. If one relaxes the conditions on the building blocks (e.g. if one assumes only a certain number of vanishing moments instead of requiring the building blocks to be compactly supported away from the origin), then one would have a valid statement only for a certain range of spaces Bs In the coorbit setting, the family of spaces Co(Y ) for which a certain window g is suitable is determined by the weight w: An analyzing window g ∈ Aw is only guaranteed to characterize Co(Y ), i.e., to satisfy Co(Y ) = {f Vgf ∈ Y } , if w is a control weight for Y . Likewise, Theorem 3 only guarantees that g ∈ Bw generates a Banach frame for Co(Y ) if w is a control weight for Y . p,q(Rd), with s ≤ s0. 2.3. Shearlets and other constructions. The first aim of the theory of coorbit spaces as proposed in [30, 31, 32, 33] was certainly to provide a unified treatment of the two most important situations (see Subsection 2.1) where painless non-orthogonal expansions arose ([18]). But it also paved the way to consider other examples through the lens of the general theory. Already early on, it was clear that certain Moebius invariant Banach spaces of analytic functions on the unit disk (see the work of Arazy-Fisher-Peetre [1]) fit into the framework of coorbit theory. These spaces are described by the be- haviour of their members, typically by imposing integrability properties, expressed by weighted mixed norm spaces, with a radial and a circular component. In a way, they can be compared with Fock spaces, which consist of analytic functions over phase space, but can be identified with the set of voice transforms (i.e. STFTs) with Gaussian window. After shearlets were first introduced by Kutyniok, Labate, Lim, Guo and Weiss[55, 51], their group theoretic nature was realized in [13], see also [54]. Build- ing upon that group-theoretic background of the shearlet transform, the theory of shearlet spaces was investigated in [14, 15, 10, 16, 11, 12]. In [14], applicability of coorbit theory (i.e. Bw (cid:54)= {0}) in dimension d = 2 was established and the associ- ated Banach frames (as provided by Theorem 3) were written down explicitly. The 9 generalization to higher dimensions (including the definition of a possible shearlet group in dimensions d > 2) was obtained in [15]. Another generalization of the shearlet group to higher dimensions was considered in [12]. Moreover, the relation of shearlet coorbit spaces to more classical smoothness spaces (namely to (sums of) homogeneous Besov spaces) was investigated in [16] for d = 2 and in [11] for higher dimensions. Related results will be discussed in Section 4. Yet another group, the so-called Blaschke group, is in the background of a series of papers by M. Pap and her coauthors ([61, 59, 60, 34]). 3. Decomposition Spaces imposing certain decay conditions on the sequence of Lp norms(cid:0) A natural starting point for the definition of decomposition spaces is the ob- servation that modulation spaces, as well as Besov spaces, can be described by i∈I for suitable families of functions (ϕi)i∈I . In the case of (inhomogeneous) Besov spaces, the (ϕn)n∈N0 form a dyadic partition of unity, while for modulation spaces, a uniform partition of unity (ϕk)k∈Zd is used. −1(ϕi(cid:98)g)(cid:107)Lp(cid:1) (cid:107)F By utilizing this observation, decomposition spaces, as introduced by Feichtinger and Grobner in [29, 21], provide a common framework for Besov- and modulation spaces, as well as many other smoothness spaces occurring in harmonic analysis. Indeed, we will see that decomposition spaces can be used to describe the α-modulation spaces-a family of spaces intermediate to Besov- and modulation spaces. Furthermore, they provide an alternative view on a large class of wavelet type coorbit spaces. We will see that the decomposition space viewpoint makes many properties of these spaces transparent, while it is highly nontrivial (if not impossible) to obtain these properties directly using the coorbit viewpoint. 3.1. Basic properties of decomposition spaces. The original-very gen- eral-definition of decomposition spaces[29, 21] starts with a covering Q = (Qi)i∈I of a given space X. Most of the time, one may think of X as (a subset of) the frequency space Rd; but in principle, it could be some manifold or other domain. Now, given a suitable function space B, the idea is to define the decomposi- tion space norm of a function/distribution f by measuring the local behaviour of f in the B-norm, i.e. by localizing f to the different sets Qi (using a suitable par- tition of unity (ϕi)i∈I ) and by measuring the individual pieces using the B-norm. The global properties of this (generalized) sequence of B norms are then restricted using a suitable sequence space Y . Formally, we define (cid:107)f(cid:107)D(Q,B,Y ) :=(cid:13)(cid:13)((cid:107)ϕi · f(cid:107)B)i∈I (cid:13)(cid:13)Y . (3.1) Of course, one has to impose certain restrictions on the covering Q, on the sequence space Y and on the family (ϕi)i∈I to ensure that equation (3.1) yields a well-defined norm/space, independent of the partition of unity (ϕi)i∈I . The most important assumptions are that the covering Q has the finite overlap property (described below) and that the (ϕi)i∈I are uniformly bounded in the pointwise multiplier algebra of B. In this paper, however, we will not use the general framework of decomposition spaces from [29, 21]. Instead, we restrict ourselves to (a slight modification of) the more specialized setting from [6], which we describe now. We start with an open subset O of the frequency space Rd and a covering Q = (Qi)i∈I of O, which we assume to be of a certain regular form: 10 Definition 4. (cf. [29, Def. 2.1 and 2.3], [6, Def. 7] and [70, Def. 3.2.8]) (1) For J ⊂ I we define the set of neighbors of J as J∗ := {i ∈ I ∃j ∈ J : Qi ∩ Qj (cid:54)= ∅} . Inductively, we set J 0∗ := J and We also define ik∗ := {i}k∗ with Qi (cid:54)= ∅ for all i ∈ I and if the following constant is finite: (2) We say that Q is an admissible covering of O , if Q is a covering of O ∗ J (n+1)∗ := (J n∗) for i ∈ I and k ∈ N0. for n ∈ N0. NQ := sup i∈I i∗ . (3) We say that Q is an almost structured covering (of O) if it is of i + bi)i∈I for certain Ti ∈ GL(Rd), bi ∈ Rd and (b) The set(cid:83) If NQ < ∞, we say that Q has the finite overlap property. the form (Qi)i∈I = (TiQ(cid:48) certain open sets Q(cid:48) i ⊂ Rd, satisfying the following properties: (a) Q is an admissible covering of O, i∈I Q(cid:48) i ⊂ Rd is bounded. (c) The following expression (then a constant) is finite: j∈i∗ (cid:107)T −1 CQ := sup i Tj(cid:107) . i∈I i )i∈I of open sets P (cid:48) (d) There is a family (P (cid:48) i for all i ∈ I and O =(cid:83) i i ∈ I} and {Q(cid:48) i ⊂ Q(cid:48) i ⊂ Rd such that: i i ∈ I} are both finite. i∈I (TiP (cid:48) i + bi). (i) The sets {P (cid:48) (ii) We have P (cid:48) sup In addition to these assumptions on Q we require Φ = (ϕi)i∈I to be uniformly bounded as pointwise multipliers of B, i.e., we impose (cid:107)ϕif(cid:107)B ≤ CΦ(cid:107)f(cid:107)B, for all f ∈ B, i ∈ I. For B = FLp, this boils down to the condition supi∈I (cid:107) ϕi(cid:107)L1 < ∞. The precise definition reads as follows: Definition 5. (cf. [29, Definition 2.2] and [6, Definition 2]) Let Q = (Qi)i∈I be an almost structured covering of O. A family Φ = (ϕi)i∈I is called a bounded admissible partition of unity (BAPU) (subordinate to Q), if the following hold: c (O) with supp ϕi ⊂ Qi for all i ∈ I, (2) (cid:80) (1) ϕi ∈ C∞ (3) supi∈I (cid:107)F Remark. In most concrete cases, the requirement ϕi ∈ C∞ −1ϕi(cid:107)L1 < ∞. i∈I ϕi ≡ 1 on O, substantially, as long as all the involved expressions make sense. however, the reservoirs D 9 below-have to be replaced by suitable substitutes, cf. [29, Definition 2.4]. c (O) can be relaxed In this case, (cid:48)(O) and Z(cid:48)(O)-which will be used in Definitions 7 and The only remaining assumption which we need to define decomposition spaces pertains to the sequence space Y from equation (3.1). For the sake of simplicity, we restrict ourselves to the case of weighted sequence spaces3, i.e. Y = (cid:96)q u(I) for a certain weight u = (ui)i∈I . Now, the general theory of decomposition spaces requires Y to be a so-called Q-regular sequence space, cf. [29, Definition 2.5]. In our setting, this leads to the following condition: 3We use the convention (cid:96)q u(I) = {(ci)i∈I (uici)i∈I ∈ (cid:96)q(I)}, with the obvious norm. 11 Definition 6. Let Q = (Qi)i∈I be an admissible covering of O. We say that a weight u = (ui)i∈I is Q-moderate, if the constant Cu,Q := sup i∈I sup (cid:96)∈i∗ ui u(cid:96) is finite, i.e. if ui (cid:16) u(cid:96) for Qi ∩ Q(cid:96) (cid:54)= ∅ (uniformly w.r.t. i, (cid:96)). −1(ϕi(cid:98)f ), it is often more convenient to directly work on the Fourier side, as Now that we have clarified all of our assumptions, we can finally give a formal definition of the decomposition spaces that we will consider in this paper. Since the norm on these spaces requires to compute frequency localizations of the form fi = F in the following definition. The usual space sided version of decomposition spaces will be introduced below. Definition 7. Assume that Φ = (ϕi)i∈I is a BAPU subordinate to the almost structured covering Q = (Qi)i∈I of O, let p, q ∈ [1,∞], and assume that u = (ui)i∈I is Q-moderate. Then we define for f ∈ D u) := −1 (ϕif )(cid:13)(cid:13)Lp (cid:48) (O) (cid:1) u) := (cid:107)f(cid:107)DF ,Φ(Q,Lp,(cid:96)q (cid:13)(cid:13)(cid:13)(cid:0)(cid:13)(cid:13)F (cid:107)f(cid:107)DF (Q,Lp,(cid:96)q u ∈ [0,∞] , (cid:13)(cid:13)(cid:13)(cid:96)q i∈I u (cid:110) u(I). with the convention that for a family c = (ci)i∈I with ci ∈ [0,∞], the expression is to be read as ∞ if ci = ∞ for some i ∈ I or if c /∈ (cid:96)q (cid:107)c(cid:107)(cid:96)q Define the Fourier-side decomposition space DF (Q, Lp, (cid:96)q (cid:111) u) with respect to the covering Q, integrability exponent p and global component (cid:96)q u as . (cid:12)(cid:12)(cid:12)(cid:107)f(cid:107)DF (Q,Lp,(cid:96)q (cid:48) (O) (cid:48)(cid:0)Rd(cid:1) is given by (integration against) a Remark. We observe that ϕif is a distribution on O with compact support −1 (ϕif )(cid:13)(cid:13)Lp , with the caveat in O, which thus extends to a (tempered) distribution on Rd. By the Paley-Wiener theorem, this implies that F that this expression could be infinite. smooth function. Thus, it makes sense to write (cid:13)(cid:13)F −1 (ϕif ) ∈ S DF (Q, Lp, (cid:96)q u) < ∞ f ∈ D u) := We finally observe that the notations (cid:107)·(cid:107)DF (Q,Lp,(cid:96)q u) sup- press the family (ϕi)i∈I used to define the norm above. But the general theory of decomposition spaces from [29] implies that different choices yield the same spaces with equivalent norms. u) and DF (Q, Lp, (cid:96)q Since it is more common to work with the proper, "space-sided" objects, rather than with their Fourier-side versions, we will now define the usual, space-sided version of decomposition spaces. To this end, we first introduce the reservoir Z(cid:48) (O) which will be used for these spaces. Our notation is inspired by Triebel's book [68]. Z (O) := F (C∞ c (O)) := {(cid:98)f f ∈ C∞ Definition 8. For ∅ (cid:54)= O ⊂ Rd open, we define F : C∞ c (O) → Z (O) c (O)} ≤ S (Rd) and endow this space with the unique topology that makes the Fourier transform a homeomorphism. (cid:48) We equip the topological dual space Z(cid:48) (O) := [Z (O)] ∗-topology, i.e., with the topology of pointwise convergence on Z (O). of Z (O) with the weak- 12 to Z(cid:48) (O), i.e. we define Finally, as on the Schwartz space, we extend the Fourier transform by duality (cid:48) (O) , f (cid:55)→ (cid:98)f := f ◦ F. F : Z(cid:48) (O) → D (3.2) c (O) → Z (O) is a linear homeomorphism, the Fourier Remark. Since F : C∞ transform as defined in equation (3.2) is a linear homeomorphism as well. (cid:107)f(cid:107)D(Q,Lp,(cid:96)q Finally, we can define the space-side decomposition spaces. Definition 9. In the setting of Definition 7 and for f ∈ Z(cid:48) (O), we define (cid:13)(cid:13)(cid:13)(cid:96)q i∈I (cid:111) and define the (space-side) decomposition space D (Q, Lp, (cid:96)q the covering Q, integrability exponent p and global component (cid:96)q u) < ∞ (cid:48) (O) is an isomor- Remark 10. Since the Fourier transform F : Z(cid:48) (O) → D u) := (cid:107)(cid:98)f(cid:107)DF (Q,Lp,(cid:96)q (cid:110) f ∈ Z(cid:48) (O) (cid:13)(cid:13)(cid:13)(cid:0)(cid:13)(cid:13)F −1 (ϕi(cid:98)f )(cid:13)(cid:13)Lp (cid:12)(cid:12)(cid:12)(cid:107)f(cid:107)D(Q,Lp,(cid:96)q u ∈ [0,∞] u) with respect to u by phism, it is clear that the Fourier transform restricts to an (isometric) isomorphism D (Q, Lp, (cid:96)q u) := u) = (cid:1) . F : D (Q, Lp, (cid:96)q u) → DF (Q, Lp, (cid:96)q u) . Hence, it does not really matter whether one considers the "space-side" or the "Fourier-side" version of these spaces. (cid:48)(O) and Z(cid:48)(O) at all, At this point, one might ask why we use the spaces D instead of resorting to the more common reservoir S(cid:48)(Rd), which is for example used to define Besov spaces. The main reasons for this are the following: (a) We want to allow the case O (cid:40) Rd. If we were to take the space S(cid:48)(Rd), the decomposition space norm would not be positive definite, or we would have to factor out a certain subspace of S(cid:48)(Rd). This is for example done in the usual definition of homogeneous Besov spaces, which are subspaces of S(cid:48)(Rd)/P, where (cid:48) (O). P is the space of polynomials. Here, it seems more natural to use the space D (b) In case of O = Rd, one could use S(cid:48)(Rd) as the reservoir, as e.g. in [6]. But as shown in [70, Example 3.4.14], this does in general not yield a complete space, even for the case Y = (cid:96)1 u with a Q-moderate weight u. The following theorem settles the issue of completeness: Theorem 11. (cf. [70, Theorem 1.4.13]) Under the assumptions of Definition 7, the (Fourier-side) decomposition space DF (Q, Lp, (cid:96)q u) is a Banach space which embeds continuously into D u) is also complete and embeds continuously into Z(cid:48)(O). Likewise, D (Q, Lp, (cid:96)q Before we close this section on the basic properties of decomposition spaces, we note that we have always assumed that we are given some BAPU Φ = (ϕi)i∈I subordinate to the covering Q. It is thus important to know whether such a BAPU actually exists. The next theorem shows that this is the case for every almost struc- tured covering. We remark that the result itself, and also the proof, are inspired heavily by the construction used in [6, Proposition 1]. (cid:48) (O). Theorem 12. (cf. [71, Theorem 2.8]) Let Q be an almost structured covering. Then there is a BAPU Φ = (ϕi)i∈I subordinate to Q. 13 All in all, we now know how to obtain well-defined decomposition spaces with respect to reasonable coverings. In the next subsection, we consider a special ex- ample of decomposition spaces-the α modulation spaces-in greater detail. 3.2. α-modulation spaces. The starting point for the original definition of α-modulation spaces in Grobner's thesis[45] were the two parallel worlds of Besov- Triebel-Lizorkin spaces and modulation spaces. Given these two types of spaces, it was natural to ask whether there is a way to connect these two families in a "continuous way". Although one could of course apply complex interpolation in order to construct spaces which are (for fixed parameters p, q, s) "in between" (Bs ally successful, because it seems to be very difficult to provide constructive charac- terizations of the members of such spaces which could be used in practice. (cid:1), this approach was-so far-not re- ) and (cid:0)M s p,q(Rd), (cid:107)·(cid:107)M s p,q(Rd),(cid:107)·(cid:107)Bs p,q p,q In contrast, it appeared-after some reflection-quite natural to try to interpo- late the two types of spaces in a geometric sense, i.e. to consider decompositions of the Fourier domain which are "in between" the dyadic partitions of unity, playing a crucial role in the description of Besov-Triebel-Lizorkin-spaces, and the uniform partitions of unity, which are relevant for the description of modulation spaces. The basic observation for this "geometric interpolation" approach is that the uniform covering (Qk)k∈Zd and the dyadic covering (Pn)n∈N0 satisfy [λ(Qk)]1/d (cid:16) (cid:104)ξ(cid:105)0 [λ(Pn)]1/d (cid:16) (cid:104)ξ(cid:105)1 ∀ξ ∈ Qk ∀ξ ∈ Pn ∀k ∈ Zd, ∀n ∈ N0, with (cid:104)ξ(cid:105) := 1 + ξ. (Qi)i∈I of Rd which satisfies Thus, for α ∈ [0, 1], an α-covering should be an (open, admissible) covering [λ(Qi)]1/d (cid:16) (cid:104)ξ(cid:105)α ∀ξ ∈ Qi ∀i ∈ I, with the implied constant independent of the precise choice of ξ and i. Apart from this natural condition, a certain further assumption (cf. [5, Definition 3.1]) is imposed to rule out pathological coverings. For brevity, we omit this condition. Given this definition, one might wonder (at least for α ∈ (0, 1)), whether there exist α-coverings. This is indeed the case. As shown in [5, Proposition A.1], there is an α covering Q(α) of the form Q(α) = (Br k :=(cid:8)ξ ∈ Rd(cid:12)(cid:12)ξ − kα0k < r · kα0(cid:9) , k)k∈Zd\{0}, with where r = rα > 0 is chosen suitably and where α0 := α 1−α . Furthermore, as shown in [70, Lemma 6.1.2 and Theorem 6.1.3], Q(α) is an (almost) structured covering of Rd and thus admits a BAPU (ϕi)i∈I . Finally, in [5, Lemma B.2], it was shown that any two α-coverings of Rd are equivalent (in a certain sense). Br Given the covering Q(α), we need suitable Q(α)-moderate weights, in order to define the α-modulation spaces. By [70, Lemma 6.1.2], moderateness holds-for arbitrary γ ∈ R-for the weight w(γ) := ((cid:104)k(cid:105)γ)k∈Zd\{0} , so that the α modulation spaces p,q (Rd) := D(Q(α), Lp, (cid:96)q M s,α w(s/(1−α))) 14 are well-defined Banach spaces. Note that the weight w(s/(1−α)) satisfies k (cid:104)ξ(cid:105)s (cid:16) w(s/(1−α)) ∀k ∈ Zd \ {0}∀ξ ∈ Bk r , similar to the case of Besov spaces and modulation spaces. p,q(Rd) = M s the "limit cases" α = 0 and α = 1, we have M s,0 M s,1 p,q(Rd) = Bs Since the α modulation spaces are obtained using "geometric interpolation" p,q (Rd) between Besov spaces and modulation spaces, one could expect that M s,α can also be obtained by "ordinary" means of interpolation (like complex interpo- lation) from Besov- and modulation spaces. But at least for the case of complex interpolation, this is false: In [49, Theorem 1.1], it is shown that In particular, for p,q(Rd), as well as p,q(Rd). (cid:2)M s1,α1 p1,q1 (Rd)(cid:3) (Rd), M s2,α2 p2,q2 = M s,α p,q (Rd) θ for certain p, p1, p2, q, q1, q2 ∈ [1,∞] and s, s1, s2 ∈ R, as well as α, α1, α2 ∈ [0, 1] and θ ∈ (0, 1) can only hold if or p2 = q2 = 2. p1 = q1 = 2 α1 = α2 or If one of these conditions hold, then interpolation works as expected. This might be surprising for α1 (cid:54)= α2; but in this case, we have p1 = q1 = 2 or p2 = q2 = 2, which implies M s1,α1 , respectively. p1,q1 But also apart from interpolation, it is natural to ask how the different α mod- ulation spaces are related. Concretely, one might wonder under which conditions an embedding of the form = H s2 = M s2,α1 p2,q2 = H s1 = M s1,α2 p1,q1 or M s2,α2 p2,q2 M s1,α1 p1,q1 (Rd) (cid:44)→ M s2,α2 p2,q2 (Rd) (3.3) holds. For α1, α2 ∈ {0, 1}, this question was solved by Kobayashi and Sugimoto[53]. For general α, it was considered by Toft and Wahlberg[67], shortly before it was solved completely-for (p1, q1) = (p2, q2)-by Han and Wang[52]. Based in part on their ideas, the second named author of the present paper developed a general theory of embeddings between decomposition spaces (cf. [70]), which we will outline in Subsection 3.3. Using this theory, the above question can be solved easily-even for (p1, q1) (cid:54)= (p2, q2)-cf. [70, Theorems 6.1.7 and 6.2.8]: Theorem 13. Let 0 ≤ α ≤ β ≤ 1, p1, p2, q1, q2 ∈ [1,∞] and s1, s2 ∈ R. Define (cid:79) with p := min{p, p(cid:48) We have if and only if p1 ≤ q1 and s(1) := α 1 p1 p2 − + (α − β) }, as well as p(cid:52) := max{p, p(cid:48) }. (Rd) (cid:44)→ M s2,β (Rd) M s1,α p1,q1 p2,q2 (cid:40) s2 ≤ s1 + d · s(0), s2 < s1 + d · (cid:0)s(0) + (1 − β)(cid:0)q−1 (cid:1)(cid:1) , 2 1 − q−1 (Rd) M s1,β p1,q1 (Rd) (cid:44)→ M s2,α p2,q2 15 Conversely, we have + if q1 ≤ q2, if q1 > q2. (cid:18) 1 (cid:18) 1 (cid:19) (cid:19) s(0) := α p2 − 1 p1 + (α − β) (cid:19) (cid:33) + (cid:18) 1 (cid:32) (cid:79) 2 − p 1 q1 1 q2 − 1 (cid:52) 1 p , , if and only if p1 ≤ q1 and (cid:40) s2 ≤ s1 + d · s(1), s2 < s1 + d · (cid:0)s(1) + (1 − β)(cid:0)q−1 1 − q−1 2 (cid:1)(cid:1) , if q1 ≤ q2, if q1 > q2. Using the theory of embeddings between decomposition spaces from Subsection 3.3, one can explain the main geometric properties of the α coverings Q(α) which lead to the preceding theorem: The main point is that the covering Q(α) is almost subordinate to ("finer than", cf. equation (3.5)) the covering Q(β) for α ≤ β. Fur- thermore, the precise conditions depend on the number of "smaller sets" that are needed to cover the "bigger" sets, cf. [70, Lemma 6.1.5] and Theorems 15 and 16 from below. We finally remark that the theorems in [70] apply for the full range (0,∞] of the exponents. But in the present paper, we restrict ourselves to the range [1,∞] for simplicity. 3.3. Embeddings between different decomposition spaces. In this sub- section, we consider embeddings between decomposition spaces with respect to dif- ferent coverings, i.e. of the form D(Q, Lp1, (cid:96)q1 u ) (cid:44)→ D(P, Lp2, (cid:96)q2 v ) (3.4) for p1, p2, q1, q2 ∈ [1,∞] and two almost structured coverings Q = (Qi)i∈I = (TiQ(cid:48) i + bi)i∈I and (cid:48) of the frequency space Rd. We assume the of two (possibly different) subsets O,O weights u = (ui)i∈I and v = (vj)j∈J to be moderate w.r.t. Q and P, respectively. As seen in the previous subsection, our setting includes embeddings between α modulation spaces for different values of α. j + cj)j∈J P = (Pj)j∈J = (SjP (cid:48) As the main standing requirement for this subsection, we assume that Q is almost subordinate to P. Very roughly, this means that the sets Qi are "smaller" than the sets Pj, or that Q is "finer" then P. Rigorously, it means that there is some fixed n ∈ N0 such that for every i ∈ I, there is some ji ∈ J satisfying Qi ⊂ P n∗ (cid:48). Even more importantly, this assumption Note that this assumption implies O ⊂ O destroys the symmetry between Q,P in equation (3.4), so that in addition to (3.4), we will also consider the "reverse" embedding (cid:91) (cid:96)∈jn∗ (3.5) P(cid:96). := ji i (3.6) Under the assumption that Q is almost subordinate to P, the object which describes those features of the coverings Q,P which are relevant for us is the family of intersection sets given by v ) (cid:44)→ D(Q, Lp1 , (cid:96)q1 u ). D(P, Lp2, (cid:96)q2 Ij := {i ∈ I Qi ∩ Pj (cid:54)= ∅} for j ∈ J. Of course, Ij = ∅ if and only if Pj ∩ O = ∅. Since these sets will be irrelevant for our purposes, we additionally define JO := {j ∈ J Pj ∩ O (cid:54)= ∅} . In the remainder of this subsection, we will state sufficient conditions and nec- essary conditions for the existence of the embeddings (3.4) and (3.6), respectively. 16 In general, these two conditions will only coincide for a certain range of p1 or p2, while there is a gap between the two conditions outside of this range. Under suitable additional assumptions, however, more strict necessary con- ditions can be derived; in fact, a complete characterization of the existence of the embeddings can be achieved. For this to hold, we will (occasionally, but not always) assume that the following properties are fulfilled: Definition 14. (1) We say that the weight u = (ui)i∈I is relatively P-moderate, if sup j∈J sup i,(cid:96)∈Ij ui u(cid:96) < ∞. (2) The (almost structured) covering Q = (TiQ(cid:48) i + bi)i∈I is called relatively P-moderate, if the weight ( det Ti)i∈I is relatively P-moderate. sentially the same measure if they intersect the same (large) set Pj. Roughly speaking, this means that two (small) sets Qi, Q(cid:96) have es- Although these assumptions might appear rather restrictive, they are fulfilled in many practical cases; in particular if Q and P are coverings associated to α- modulation spaces, and if u, v are the usual weights for these spaces. We can now analyze existence of the embedding (3.6): Theorem 15. (cf. [70, Theorem 5.4.1]) Define p 2} and for r ∈ [1,∞], let (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:32)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  uij (cid:17) (cid:52) 2 := max{p2, p(cid:48) − 1 p1 · ui det Ti (cid:18) 1 p2 − 1 p2 1 q1 vj · det Tij i∈Ij − 1 (cid:52) p 2 (cid:33) (cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q1·(r/q1)(cid:48) (cid:30) (cid:18) − 1 p1 · det Sj (cid:19) (Ij ) + vj j∈J 1 q1 − 1 (cid:52) 2 p (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q1·(q2/q1)(cid:48) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  (cid:19) j∈JO + C (r) 1 := C2 := (cid:96)q1·(q2 /q1)(cid:48) where for each j ∈ JO some ij ∈ I with Qij ∩ Pj (cid:54)= ∅ can be chosen arbitrarily. (JO) (J) , , Then the following hold: (1) If C (p 1 (cid:52) 2 ) is well-defined and bounded. < ∞ and p2 ≤ p1, then the map v ) → D(Q, Lp1 , (cid:96)q1 (cid:17) ι : D(P, Lp2, (cid:96)q2 (cid:16) (2) Conversely, if u ), f (cid:55)→ fF (C∞ c (O)) (3.7) (3.8) −1(C∞ θ : F c (O)),(cid:107)·(cid:107)D(P,Lp2 ,(cid:96)q2 v ) → D(Q, Lp1, (cid:96)q1 u ), f (cid:55)→ f is bounded, then C (p2) 1 < ∞ and p2 ≤ p1. (3) Finally, if Pj ⊂ O holds4 for all j ∈ JO and if additionally the covering Q and the weight u = (ui)i∈I are both relatively P-moderate, we have the following equivalence: ι is bounded ⇐⇒ θ is bounded (cid:16) (cid:52) 2 ) C (p 1 < ∞ and p2 ≤ p1 ⇐⇒ ⇐⇒ (C2 < ∞ and p2 ≤ p1) . (cid:17) 4The main case in which this holds is if O = O(cid:48). 17 Remark. (1) We achieve a complete characterization of the existence of the embedding (3.6) if Q and u are relatively P-moderate, but also in case of p2 ∈ [2,∞], since in this case, p (2) Even for well-understood special cases like α modulation spaces, the above theorem yields new results, since even in the most general previous result [52], only the case (p1, q1) = (p2, q2) was studied. (cid:52) 2 = p2 and hence C (p = C (p2) (cid:52) 2 ) 1 1 . (3) Considering ι as an embedding is not always justified. For example, if \ O)) satisfies \ O has nonempty interior, then every f ∈ F (cid:48) O ιf = 0, although f (cid:54)= 0 is possible. −1(C∞ c (O (cid:48) 2} and for r ∈ [1,∞], let , (J) (cid:19) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:96)q2·(q1/q2)(cid:48) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  −1(ϕi · (cid:98)f ) j∈JO (cid:88) + , (3.9) The result for the embedding (3.4) is similar: Theorem 16. (cf. [70, Theorem 5.4.4]) (cid:79) Let (ϕi)i∈I be a BAPU for Q. Define p (cid:13)(cid:13)(cid:13)(cid:13)(cid:16) u−1 i (cid:17) (cid:19) + (cid:33) 2 := min{p2, p(cid:48) (cid:13)(cid:13)(cid:13)(cid:13)(cid:96)r·(q1/r)(cid:48) (Ij ) j∈J (cid:18) − 1 p2 · det Sj − 1 q1 (cid:79) 1 p 2 − 1 p2 1 p1 i∈Ij · det Ti (cid:18) − 1 p1 − 1 q1 (cid:79) 1 p 2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:32) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  vj vj · C (r) 1 := C2 := uij · det Tij (cid:96)q2·(q1/q2)(cid:48) where for each j ∈ JO some ij ∈ I with Qij ∩ Pj (cid:54)= ∅ can be chosen arbitrarily. (JO) Then the following hold: (1) If C (p (cid:79) 2 ) 1 < ∞ and if p1 ≤ p2, then the map ι = ιΦ : D(Q, Lp1, (cid:96)q1 u ) (cid:44)→ D(P, Lp2, (cid:96)q2 F is well-defined and bounded. Here, ιf acts as follows: v ), f (cid:55)→ i∈I (cid:104)ιf, γ(cid:105) = (cid:104)f,F(ϕi · F −1γ)(cid:105) for γ ∈ F(C∞ c (O (cid:48))), (cid:88) i∈I (2) If the map v ) ⊂ Z(cid:48)(O with absolute convergence of the series for f ∈ D(Q, Lp1 , (cid:96)q1 u ). Furthermore, (cid:104)ιf, γ(cid:105) = (cid:104)f, γ(cid:105) holds for all γ ∈ F(C∞ c (O)), so that (cid:16) ιf ∈ D(P, Lp2, (cid:96)q2 −1(C∞ (cid:48)) extends f ∈ D(Q, Lp1, (cid:96)q1 → D(P, Lp2, (cid:96)q2 c (O)),(cid:107)·(cid:107)D(Q,Lp1 ,(cid:96)q1 is bounded, then p1 ≤ p2 and C (p2) (3) Finally, if Pj ⊂ O holds for all j ∈ JO and if additionally the covering Q and the weight u = (ui)i∈I are both relatively P-moderate, we have the following equivalence: u ) ⊂ Z(cid:48)(O). v ), f (cid:55)→ f 1 < ∞. (3.10) (cid:17) F θ : u ) (cid:16) ι is bounded ⇐⇒ θ is bounded. (cid:79) C (p 2 ) 1 < ∞ and p1 ≤ p2 ⇐⇒ ⇐⇒ (C2 < ∞ and p1 ≤ p2) . (cid:17) Remark. (1) We achieve a complete characterization of the existence of the embedding (3.4) for p2 ∈ [1, 2]. If Q and u are relatively P-moderate, we get a complete characterization for arbitrary p2 ∈ [1,∞]. 18 (2) In contrast to Theorem 15, ι is always injective in the present setting. (3) Note that the definition of ι is independent of p1, p2, q1, q2 and u, v. In (cid:48)). (cid:48), then ιf = f for all f ∈ D(Q, Lp1 , (cid:96)q1 u ) ⊂ Z(cid:48)(O) = Z(cid:48)(O fact, if O = O For one concrete application of Theorems 15 and 16, we refer the reader to the characterization of embeddings between different α modulation spaces in Theorem 13. Further applications will be given in Section 4. There are also results which apply if neither Q is almost subordinate to P, nor (cid:48) = A∪ B, such that Q is almost vice versa. In fact, it suffices if one can write O∩O subordinate to P near A and vice versa near B. For a precise formulation of this condition, and the resulting embedding results, we refer to [70, Theorem 5.4.5]. Finally, there are also results for embeddings of the form D(Q, Lp, Y ) (cid:44)→ W k,q(Rd) for the classical Sobolev spaces W k,q. These results are easy to apply, since no subordinateness is required. As shown in [71], existence of the embedding can be completely characterized for q ∈ [1, 2]∪{∞}, while for q ∈ (2,∞), certain sufficient and certain necessary criteria are given; but in general, these do not coincide. 3.4. Banach frames for Decomposition spaces. Borup and Nielsen[6] gave a construction of Banach frames for decomposition spaces which applies in a very general setting. In particular, their construction applies to (α)-modulation spaces and Besov spaces. Since this construction makes the power of Banach frames available for decomposition spaces, we could not resist discussing their results. Furthermore, their results fit well into the present context: As Borup and Nielsen write themselves: "[our] frame expansion should perhaps be considered an adaptable variant of the ϕ-transform of Frazier and Jawerth" (cf. [6, Section 3.2]). To describe their construction, we first introduce structured coverings. An i = Q for all i ∈ I, i.e., if all Qi are affine images of a fixed set. almost structured covering Q = (Qi)i∈I = (TiQ(cid:48) structured if Q(cid:48) i + bi)i∈I of O = Rd is called The idea is to transfer the orthonormal basis (e2πi(cid:104)k,·(cid:105))k∈Zd of L2([−1/2, 1/2]d) to each of the sets Q(a) i = Ti[−a, a]d+bi ⊃ Qi for certain a > 0. Then, one truncates these periodic functions using a certain (quadratic) partition of unity subordinate to Q. Thus, one obtains a tight frame for L2(Rd). The nontrivial part is to show that one also obtains Banach frames for the whole range of decomposition spaces. The construction proceeds as follows: By [6, Proposition 1] on finds a family (θi)i∈I such that: (2) (cid:80) (1) supp θi ⊂ Qi for all i ∈ I, i ≡ 1 on O = Rd, i∈I θ2 −1θi(cid:107)L1 < ∞, (3) supi∈I (cid:107)F (4) supi∈I (cid:107)∂α[θi(Ti · +bi)](cid:107)sup < ∞ for all α ∈ Nd 0. Given such a family (θi)i∈I , we choose a cube Qa ⊂ Rd of side-length 2a satisfying Q ⊂ Qa. Finally, for i ∈ I and n ∈ Zd, define en,i : Rd → C by (ξ−bi) for ξ ∈ Rd, en,i(ξ) := (2a)−d/2 · det Ti −1/2 · χQa (T −1 (ξ − bi)) · ei π a n·T −1 i and set ηn,i := F −1(θi · en,i). 19 i cause of(cid:80) i∈I θ2 forms a tight frame for L2(Rd). Since the family (en,i)n∈Zd forms an orthonormal basis of L2(TiQa +bi) and be- i ≡ 1, it follows (cf. [6, Proposition 2]) that the family (ηn,i)n∈Zd,i∈I Of course, we are not simply interested in (tight) frames for L2(Rd) with a given form of frequency localization-we want to obtain a Banach frame for the decomposition space D(Q, Lp, (cid:96)q η(p) n,i := det Ti for i ∈ I and n ∈ Zd. Then, Borup and Nielsen showed (cf. [6, Proposition 3, Definition 8, Lemma 4 and Theorem 2]) that there is a suitable solid BK space d(Q, (cid:96)p, (cid:96)q u) such that the coefficient operator u). Thus, we define the Lp-normalized version 2− 1 p · ηn,i 1 C : D(Q, Lp, (cid:96)q u) → d(Q, (cid:96)p, (cid:96)q u), f (cid:55)→ ((cid:104)f, η(p) n,i(cid:105))n∈Zd,i∈I is bounded. As familiar by now, there is also a bounded reconstruction operator R : d(Q, (cid:96)p, (cid:96)q u). Thus, the family (η(p) u) which satisfies R ◦ C = idD(Q,Lp,(cid:96)q n,i )i∈I,n∈Zd forms a Banach frame for D(Q, Lp, (cid:96)q u). u) → D(Q, Lp, (cid:96)q 4. Abstract and Concrete Wavelet Theory As noted in Section 2, the description of the spaces Bs p,q via the ϕ- transform-or via wavelets-can be viewed (at least in part) as an application of the general theory of coorbit spaces to the affine group which acts on L2(Rd) via translations and isotropic dilations. To be precise, this group action yields the s homogeneous Besov- and Triebel-Lizorkin spaces B p,q, not the inhomo- geneous ones. p,q and F s p,q and F s One well known characterization of (homogeneous) Besov spaces shows that these spaces are obtained by placing certain integrability conditions on the (con- tinuous) wavelet transform Wϕf : Rd × R∗ → C, (x, a) (cid:55)→ (cid:104)f, TxDaϕ(cid:105) of a function or distribution f and a certain analyzing window ϕ. Other applica- tions of the wavelet transform include the characterization of the wave-front set of a distribution using the decay of the transform [63] However, due to the isotropic nature of the dilations Daf (x) = a−d/2 · f (a−1x), such a single wavelet characterization is only valid in dimension d = 1 (cf. [7] or [35, Lemma 4.4, Lemma 4.10]), where smoothness is an "undirected property". Even beyond this specific problem of characterizing the wave-front set, it was noted in recent years that the isotropic, directionless nature of the wavelet transform is a limitation for many applications. To overcome this problem, a vast number of "directional" variants of wavelets were invented: In particular, curvelets[7, 8] and shearlets[54, 14]. Among these two systems, shearlets have the special property that there is-as in the case of wavelets-an underlying dilation group through which the family of shearlets can be generated from a single "mother wavelet", see also Section 2.3. In view of these two very different dilation groups-the affine group and the shearlet group-it becomes natural to consider the bigger picture: Given any (closed) subgroup H ⊂ GL(Rd), one can form the group G = Rd (cid:111) H of all 20 affine mappings generated by arbitrary translations and all dilations in H. The multiplication on G is given by (x, h)(y, g) = (x + hy, hg) and the Haar measure is d(x, h) = det h −1dx dh, (4.1) where dh is the Haar measure of H. The group G from above acts unitarily on L2(Rd) via translations and dilations, i.e., by the quasi-regular representation π : G → U(L2(Rd)), (x, h) (cid:55)→ TxDh. (4.2) This representation comes with an associated (generalized) wavelet transform Wϕf : G → C, (x, h) (cid:55)→ (cid:104)f, π(x, h)ϕ(cid:105) for f, ϕ ∈ L2(Rd), where the (fixed) function ϕ is called the analyzing window. description of coorbit theory in Section 2, this was called the voice transform. In the general Given this transform, it is natural to ask which properties of f can be easily read off from Wϕf . As for wavelets (for d = 1 [63]) or for shearlets (for d = 2 [54, 47]), it turns out[35] that for large classes of dilation groups, the wave-front set of a (tempered) distribution f can be characterized via the decay of Wϕf . Another important property of a generalized wavelet system (like shearlets) are its approximation theoretic properties. Here, the question is: Which classes of functions can be approximated well by linear combinations of only a few elements of the wavelet system? For "ordinary" wavelets, this question leads to the theory of Besov spaces and their atomic decompositions, as explored by Frazier and Jawerth. For a general dilation group, these approximation theoretic properties are (at least in part) encoded by the associated wavelet type coorbit spaces, which we will now discuss in greater detail. One particular problem which is of interest to us is the following: If a function/signal f can be well approximated by one wavelet system, can it also be well approximated using a different wavelet system? Of course, the answer to this question will depend on the precise nature of the two wavelet systems and on the way in which the statement "f can be well approximated by . . . " is made mathematically precise. 4.1. General wavelet type coorbit spaces. As long as π acts irreducibly on L2 (and if the representation is (square) integrable), we can apply the general coorbit theory as described in Section 2 to form the coorbit spaces Co(G, Y ) = {f ∈ R Wϕf ∈ Y } , (4.3) Besov- or Triebel-Lizorkin spaces. where R = RY is a suitable reservoir, which plays the role of S(cid:48)(Rd) for the usual Furthermore, ϕ ∈ L2(Rd) has to be a suitable analyzing window. Formally, this means that ϕ must fulfill ϕ ∈ Av0 (cf. equation (2.2)) for a so-called control weight v0 : G → (0,∞). As we will see (cf. Definition 18), this condition is closely related to the usual "vanishing moments condition" for ordinary wavelets. In this section, instead of the general coorbit spaces Co(Y ) = Co(G, Y ), we will consider the more restrictive case of the (weighted) mixed Lebesgue space Y = Lp,q m (G) for p, q ∈ [1,∞] and a weight m : H → (0,∞). Precisely, the space m (G) is the space of all measurable functions f : G → C for which the norm Lp,q m :=(cid:13)(cid:13)h (cid:55)→ m(h) · (cid:107)f (x, h)(cid:107)Lp(Rd, dx) (cid:13)(cid:13)Lq(H, dh/ det h) (cid:107)f(cid:107)Lp,q 21 m (G) = Lp m(G), cf. equation (4.1). is finite. This normalization-in particular the measure dh/ det h on H-is chosen such that we have Lp,p Recall from Section 2 that the space Y needs to be a solid BF space on G which is invariant under left- and right translations. Clearly, Y = Lp,q m (G) satisfies all of these properties, except possibly for invariance under left- and right translations. To ensure this, we assume that m is v-moderate for some (measurable, locally bounded, submultiplicative) weight v : H → (0,∞), i.e., we assume ∀x, y, z ∈ H. m(xyz) ≤ v(x)m(y)v(z) Under these assumptions, it is shown in [44, Lemma 1 and Lemma 4] that there is a control weight v0 : G → (0,∞) for Y = Lp,q m (G) which is (by abuse of notation) of the form v0(x, h) = v0(h) and measurable, submultiplicative and locally bounded. In the present setting, coorbit theory can be seen as a theory of nice wavelets and nice signals, cf. [42]. Nice wavelets are those belonging to the class Av0 of analyzing windows, while a nice signal f (with respect to an analyzing wavelet ϕ) is one for which Wϕf ∈ Lp,q m ). Now, coorbit theory-if it is applicable-yields two main properties: m , i.e., for which f ∈ Co(Lp,q • A consistency statement: Nice wavelets agree upon nice signals, i.e. if ϕ, ψ ∈ Av0 \ {0} (with v0 depending on p, q, m), then Wψf ∈ Lp,q m . Wϕf ∈ Lp,q ⇐⇒ m m (G) is well-defined. We even get a norm-equivalence, so that the coorbit space from equation (4.3) with Y = Lp,q atomic decompositions of the form f =(cid:80) • An atomic decomposition result: As seen in Theorem 3, we can guarantee αg(f ) · π(g)ψ for elements f ∈ Co(Lp,q m ) and coefficients (αg(f ))g∈G0 lying in a suitable sequence space, if ψ is a better vector (in comparison to just being an analyzing vector), i.e. if ψ ∈ Bv0. spaces raises several questions: Despite these pleasant features, the theory of (generalized) wavelet type coorbit g∈G0 (Q1) For which dilation groups H is the quasi-regular representation from equa- tion (4.2) irreducible and (square)-integrable, so that coorbit theory is applicable in principle? (Q3) How are the resulting coorbit spaces Co(Rd (cid:111) H, Lp,q (Q2) Is coorbit theory applicable, i.e. are there "nice wavelets"? Precisely, do we have Av0 (cid:54)= {0} and Bv0 (cid:54)= {0} and are there convenient sufficient criteria for a function ϕ ∈ L2(Rd) to belong to Av0 or to Bv0? m ) related to classical function spaces like Bs p,q (or their homogeneous counterparts)? Furthermore, how are coorbit spaces with respect to different dilation groups related to each other? p,q of F s This connects to the question posed above: If a given function/signal can be well approximated using one wavelet system, does the same also hold for a different system? Even for the special case of the shearlet dilation group, these questions are non- trivial and triggered several papers [14, 10, 16, 11]. Nevertheless, they also admit satisfactory answers in the present generality: As we will see, each of these questions is linked to the dual action : H × Rd → Rd, (h, ξ) (cid:55)→ h−T ξ 22 of the dilation group H on the frequency space Rd. To see the relevance of the dual action, note that the Fourier transform of Wϕf (·, h) is given by (F[Wϕf (·, h)]) (ξ) = det h1/2 · (cid:98)f (ξ) · (cid:98)ψ(hT ξ). Thus, if (cid:98)ψ has support in U ⊂ Rd, then Wϕf (·, h) is bandlimited to h−T U = (h, U ). The remaining subsections deal with the three questions listed above. (4.4) 4.2. Question 1: Irreducibility and square-integrability of π. As shown in [40, 39, 4], the quasi-regular representation π is irreducible and square-integrable if and only if the following two properties are satisfied: (1) There is some ξ0 ∈ Rd \ {0} such that the orbit O := H T ξ0 =(cid:8)hT ξ0 (2) The stabilizer Hξ0 :=(cid:8)h ∈ H(cid:12)(cid:12) hT ξ0 = ξ0 (cid:12)(cid:12) h ∈ H(cid:9) (cid:9) is compact. is open and of full measure (i.e. Oc is a Lebesgue null set). In this case, we call O the (open) dual orbit of the dilation group H, whereas the null-set Oc is called the blind spot of H. Finally, a dilation group H fulfilling the two properties above is called admissible. In the following, we fix ξ0 ∈ O. To give the reader an idea of the richness of admissible dilation groups, we mention the following admissible dilation groups in dimension d = 2: (1) The diagonal group with dual orbit O = (R \ {0})2. (2) The similitude group H2 := (0,∞) · SO(R2), with O = R2 \ {0}. (3) The family of shearlet type groups H1 := {diag(a, b) a, b ∈ R \ {0}} (cid:18)a (cid:19)(cid:12)(cid:12)(cid:12)(cid:12) a ∈ (0,∞), b ∈ R, ε ∈ {1,−1} b 0 ac ε · (cid:26) (cid:27) H (c) 3 := . (4.5) Here, the anisotropy parameter c ∈ R can be chosen arbitrarily. Re- gardless of this choice, the dual orbit is always O = (R \ {0}) × R. 4.3. Question 2: Existence of "nice" wavelets. Here, we are given an admissible dilation group H and we are interested in conditions which guarantee Av0 (cid:54)= {0} or Bv0 (cid:54)= {0}, where v0 : H → (0,∞) (interpreted by abuse of notation as a weight on G = Rd (cid:111) H) is a (locally bounded) control weight for Y = Lp,q m . As we saw above, such a control weight always exists under our general assumptions on the weight m : H → (0,∞). In the present setting, "nice wavelets" and "better wavelets" exist in abundance: Theorem 17. (cf. [44, Theorem 9]) Let v0 : H → (0,∞) be measurable and locally bounded. Then, every function c (O)) ⊂ S(Rd) satisfies ϕ ∈ Bv0 ⊂ Av0 and the map −1(C∞ ϕ ∈ F : C∞ is well-defined and continuous. c (O) → L1 v0 (G), g (cid:55)→ Wϕ(F −1g) (4.6) wavelets". This theorem, however, does not yield existence of compactly supported "nice In the case of "traditional" wavelets, it is well-known that a certain amount of "vanishing moments" (i.e. ∂α(cid:98)ϕ(0) = 0 for all α ≤ r) makes a wavelet "nice". This generalizes to the present setting: 23 (∂α(cid:98)ϕ)Oc ≡ 0 vanishing moments on Oc of order r if (cid:98)ϕ ∈ C r(Rd) and if Definition 18. ([43, Definition 1.4]) Let r ∈ N. We say that ϕ ∈ L1(Rd) has ψ ∈ L1(Rd) with (cid:107)(cid:98)ψ(cid:107)(cid:96),(cid:96) < ∞ and with vanishing moments of order (cid:96) on Oc satisfies As shown in [43, Theorems 1.5(c), 2.12], given the weight v0 : H → (0,∞), one can explicitly compute a number (cid:96) = (cid:96)(H, v0) ∈ N such that every function ψ ∈ Bv0 ⊂ Av0 . Here, we employed the usual Schwartz-type norm (1 + x)(cid:96) · ∂αf (x) ∈ [0,∞]. (cid:107)f(cid:107)(cid:96),(cid:96) := max for α < r. α∈Nd sup x∈Rd 0 ,α≤(cid:96) Finally, in the setting of ordinary wavelets, one can obtain a wavelet with enough vanishing moments by taking derivatives of a given smooth function of sufficient decay. In the present general setting, [41, Lemmas 3.1 and 4.1] yield a similar "algorithm" for obtaining functions with suitably many vanishing moments on Oc. For the sake of brevity, we refrain from giving more details and instead refer the interested reader to [41, 43]. m1 , Rd (cid:111) H1) and Co(Lp2,q2 4.4. Question 3: Relation of generalized wavelet coorbit spaces to other spaces. Here, we are interested in the relation of the coorbit space Co(Lp,q m ) to the classical Besov- or Triebel-Lizorkin spaces, but also in the relation between Co(Lp1,q1 , Rd (cid:111) H2) for two different admissible dilation groups H1, H2 ≤ GL(Rd). In view of the embedding theory for decomposition spaces (cf. Subsection 3.3), this question would be solved (at least to a significant extent) if we knew that the coorbit space Co(Lp,q m ) is (canonically isomorphic to) a decomposition space. But the main result of [44] is precisely such an isomorphism. In more detail, [44] shows m2 Co(Lp,q m (Rd (cid:111) H)) ∼= D(QH , Lp, (cid:96)q u) (4.7) for a so-called induced covering QH and a so-called decomposition weight u. Below, we provide a concrete example indicating how this isomorphism and the above embedding results can be used to obtain novel embedding results for shearlet coorbit spaces. given by QH = (h−T construction is described in the following paragraph. But once Q = (h−T In the above isomorphism, the induced covering QH of the dual orbit O is i Q)i∈I for certain hi ∈ H and a suitable Q ⊂ O. Its precise i Q)i∈I is known, the decomposition weight u = (ui)i∈I from above (cf. [44, Lemma 35 and the ensuing remark]) is given by (4.8) Let us reconsider the induced covering QH . The family (hi)i∈I has to be well spread in H, i.e. there are compact unit neighborhoods K1, K2 ⊂ H such that ui = det hi q · m(hi) ∀i ∈ I. 1 2− 1 hiK1 and hiK2 ∩ hjK2 = ∅ for i, j ∈ I with i (cid:54)= j. H = (cid:91) P ⊂ Q and O =(cid:83) i∈I Finally, the set Q ⊂ O is an arbitrary open, bounded set such that the closure Q ⊂ Rd is contained in O and such that there is a smaller open set P ⊂ Rd with i P . As shown in [44, Theorem 20], such sets P, Q always exist if (hi)i∈I is well-spread in H. The same theorem also shows that the resulting i∈I h−T 24 covering QH = (h−T i Q)i∈I is an (almost) structured admissible covering of O. In particular, there is a BAPU (ϕi)i∈I subordinate to QH , cf. Theorem 12. Finally, [44, Lemmas 22 and 23] show that the decomposition weight from equa- tion (4.8) is QH -moderate. This implies (cf. Subsection 3.1) that the decomposition space D(QH , Lp, (cid:96)q As an illustration of the concept of an induced covering, we consider the shearlet type group H (c) from equation (4.5). A picture of (a part of) the associated induced covering for different values of the anisotropy parameter c ∈ R is shown in Fig. 1. u) on the right-hand side of equation (4.7) is well-defined. 3 3 for c = − 1 Figure 1. The figure shows (a part of) the induced covering for the group H (c) 2 . These choices show the qualitatively different behaviour of the covering for different values of c. For c1 < c2, the covering S (c1) is "larger/coarser" near the y-axis, whereas the covering S (c2) is "larger/coarser" away from the y-axis. Note: The images are taken from [70, Figure 1]. 2 , c = 0 and c = 1 For the explicit description of the isomorphism from equation (4.7), we note that it is a (relatively) easy consequence of equation (4.6) that c (O)) → H1 v0 , f (cid:55)→ f , v0 with H1 this implies that ι : Z(O) = F(C∞ (cid:1)∠ ιT :(cid:0) subspace of the reservoir R = (cid:0) H1 H1 v0 v0 25 as in equation (2.1) is well-defined, antilinear and continuous. By duality, is well-defined, continuous and linear. Recall from Section 2 that Co(Lp,q → Z(cid:48)(O), θ (cid:55)→ θ ◦ ι (cid:1)∠ (4.9) m ) is a for a certain control weight v0. Thus, c=−0.5020212223012345c=0.0020212223012345c=0.5020212223012345ιT f ∈ Z(cid:48)(O) is well-defined for every f ∈ Co(Lp,q precisely that ιT : Co(Lp,q A proof of this fact can be found in [44, Theorem 43]. m ) → D(Q, Lp, (cid:96)q This representation of Co(Lp,q m ). The claim of equation (4.7) is u) is an isomorphism of Banach spaces. m (G)) as a decomposition space has several im- portant consequences, both mathematically and conceptually: • As we saw above, the coorbit space Co(Lp,q m (G)) with its original definition is heavily tied to the group G = Rd × H and thus to the dilation group H. In particular, Co(Lp,q ∠ . (Rd(cid:111)H1)) as an element of another coorbit space Co(Lp2,q2 (Rd (cid:111) H2)), or as an m2 element of more classical function spaces like Bs p,q. This makes it difficult to consider an element f ∈ Co(Lp1,q1 m (G)) is a subspace of the reservoir R = (H1 (G)) m1 v0 In contrast, as we saw in Section 3, it is (at least in principle) possible v ) which ⊂ Rd. to compare decomposition spaces D(Q, Lp1, (cid:96)q1 are defined using two different coverings Q,P of the sets O,O compare wavelet coorbit spaces defined by different dilation groups. Thus, using the decomposition space view, it becomes possible to u ) and D(P, Lp2, (cid:96)q2 (cid:48) • Even ignoring the issue of the different reservoirs (e.g. by restricting to Schwartz functions), it is not at all obvious how the decay or integrabil- (Rd (cid:111) H1) relates to another decay condition ity condition Wϕf ∈ Lp1,q1 (Rd(cid:111)H2), even if the same analyzing window is used in both Wϕf ∈ Lp2,q2 cases. One of the reasons is that it is difficult to compare the two actions of the dilation groups on ϕ, as well as the two distinct Haar measures. m2 m1 In comparison, the decomposition space point of view translates these two elusive properties into (more or less) transparent quantities, namely i Q)i∈I for some well-spread fam- (1) The induced covering QH = (h−T ily (hi)i∈I in H and a suitable set Q ⊂ O, (2) The decomposition weight ui = det hi Using the methods from Subsection 3.3, it is then (comparatively) easy to establish embeddings D(QH1, Lp1 , (cid:96)q1 ) be- tween the associated decomposition spaces and thus of the two coorbit spaces Co(Lp1,q1 ) (cid:44)→ D(QH2, Lp2 , (cid:96)q2 (Rd (cid:111) H1)) and Co(Lp2,q2 (Rd (cid:111) H2)). q · m(hi). 2− 1 u1 u2 1 m2 m1 • Similarly, one can use the methods from Subsection 3.3 to establish embed- dings between generalized wavelet coorbit spaces and classical smoothness spaces like Sobolev- and Besov spaces. • Conceptually, all these considerations show that the approximation the- oretic properties of the wavelet system generated by a dilation group H are completely determined by the way in which (the dual action of ) H covers/partitions the frequencies. Theorem 19 below is an example of results that can be obtained by combining the embedding results from Subsection 3.3 with the isomorphism Co(Lp,q m (Rd (cid:111) H)) ∼= D(QH , Lp, (cid:96)q u). Precisely, we consider embeddings between shearlet coorbit spaces and inhomoge- neous Besov spaces. For the sake of brevity, we only consider embeddings of the shearlet coorbit space into inhomogeneous Besov spaces. Results for the reverse direction are also available (cf. [70, Theorem 6.3.14]), but are omitted here. We only consider the case c ∈ (0, 1]. This ensures that the induced covering is almost subordinate to the inhomogeneous dyadic covering. See [70, S (c) = QH (c) 3 26 Lemma 6.3.10] for a formal proof and Figure 1 for a graphical illustration. Note, however, that S (c) is not relatively moderate with respect to the inhomogeneous dyadic covering. This limits sharpness of our results to a certain range of p2. Theorem 19. ([70, Theorem 6.3.14]) Let c ∈ (0, 1], p1, p2, q1, q2 ∈ [1,∞] and (cid:79) α, β, γ ∈ R. Set p and set α(1) := 2 := min{p2, p(cid:48) u(α,β) : H (c) (cid:18) 1 p1 − (cid:18) 1 1 + c c 2}, define the weight 3 → (0,∞), h (cid:55)→ (cid:107)h−1(cid:107)α · det hβ (cid:18) 1 (cid:19) 1 p2 − + β + , 1 q1 1 p2 − 1 2 1 q1 + (c − 1) · (cid:79) 2 − p (cid:19) 1 q1 . + p1 − γ(1) := −(1 + c) · (cid:40) If p1 ≤ p2 as well as γ ≤ α + γ(1), γ < α + γ(1), if q1 ≤ q2, if q1 > q2 and hold, then (cid:19) (cid:110) 1 + + β 1 2 (cid:111) (cid:40) , α max max{0, α} ≤ α(1), 2 − 1 q1 (cid:79) p < α(1), (cid:79) if q1 > p 2 , (cid:79) if q1 ≤ p 2 (4.10) (cid:79) 2 Co(Lp1,q1 u(α,β) (R2 (cid:111) H (c) 3 )) (cid:44)→ Bγ p2,q2 (R2). A necessary condition for existence of this embedding is obtained by replacing p by p2 everywhere (also in the definition of γ(1)). Remark. (1) Existence of the embedding (4.10) has to be interpreted suitably. Precisely, (4.10) means that there is a bounded linear map ι : Co(Lp1,q1 3 )) → Bγ u(α,β)(R2 (cid:111) H (c) (R2) which satisfies ιf = f for all f ∈ L2(R2) ∩ Co(Lp1,q1 u(α,β)(R2 (cid:111) H (c) of the strict subspace SCCp,r (cid:12) Co(Lp,p p,p(R2) + B (2) The preceding theorem is superficially similar to [16, Theorem 4.7]. But the two results are very different, since Dahlke et al. consider embeddings )) into a sum of p,p(R2) for certain σ1, σ2. u(0,−2r/3) (R2 (cid:111) H (1/2) homogeneous Besov spaces B 3 )). p2,q2 σ1 σ2 3 In contrast, the preceding theorem investigates embeddings of the whole shearlet coorbit space into a single, inhomogeneous Besov space. (3) The preceding theorem achieves a complete characterization of the em- (cid:79) 2 = p2 in this range. bedding (4.10) for p2 ∈ [1, 2], since we have p As a conclusion, we remark that the embedding results and the isomorphism between generalized wavelet coorbit spaces and decomposition spaces can also be used to derive embeddings between the coorbit spaces Co(Lp,q 3 )) for different values of c, see [70, Theorem 6.3.9]. m (R2 (cid:111) H (c) They can also be used to derive (non)boundedness of certain operators-e.g. dilation operators-acting on coorbit spaces. For example, in [70, Theorem 6.5.9], the set of matrices which act boundedly by dilation simultaneously on all coorbit spaces of the shearlet type group H (c) (for a fixed c ∈ (0, 1)) is characterized completely. 3 27 Acknowledgments Both authors want to thank HIM-the Hausdorff Institute of Mathematics- where we both spent some time during the preparation of this manuscript. FV was funded by the Excellence Initiative of the German federal and state governments, and by the German Research Foundation (DFG), under the contract FU 402/5-1. References [1] J. Arazy, S. Fisher, and J. Peetre. Mobius invariant function spaces. J. Reine Angew. Math., 363:110–145, 1985. [2] R. Balian. Un principe d'incertitude fort en th´eorie du signal ou en m´ecanique quantique. C. R. Acad. Sci. Paris S´er. II M´ec. Phys. Chim. Sci. Univers Sci. Terre, 292(20):1357–1362, 1981. [3] J. J. Benedetto, C. Heil, and D. F. Walnut. Differentiation and the Balian-Low theorem. J. Fourier Anal. Appl., 1(4):355–402, 1995. [4] D. Bernier and K. F. Taylor. Wavelets from square-integrable representations. SIAM J. Math. Anal., 27(2):594–608, 1996. [5] L. Borup and M. Nielsen. Banach frames for multivariate α-modulation spaces. J. Math. Anal. Appl., 321(2):880–895, 2006. [6] L. Borup and M. Nielsen. Frame decomposition of decomposition spaces. J. Fourier Anal. Appl., 13(1):39–70, 2007. [7] E. J. Cand`es and D. L. Donoho. Continuous curvelet transform. I: Resolution of the wavefront set. Appl. Comput. Harmon. Anal., 19(2):162–197, 2005. [8] E. J. Cand`es and D. L. Donoho. Continuous curvelet transform. II: Discretization and frames. Appl. Comput. Harmon. Anal., 19(2):198–222, 2005. [9] R. R. Coifman and R. Rochberg. Representation theorems for holomorphic and harmonic functions in Lp. Ast'erisque, 77:11–66, 1980. [10] S. Dahlke, G. Steidl, and G. Teschke. Multivariate shearlet transform, shearlet coorbit spaces and their structural properties. In Shearlets. Multiscale analysis for multivariate data., pages 105–144. Boston, MA: Birkhauser, 2012. [11] S. Dahlke, S. Hauser, G. Steidl, and G. Teschke. Shearlet coorbit spaces: traces and embed- dings in higher dimensions. Monatsh. Math., 169(1):15–32, 2013. [12] S. Dahlke, S. Hauser, and G. Teschke. Coorbit space theory for the Toeplitz shearlet trans- form. Int. J. Wavelets Multiresolut. Inf. Process., 10(04):1250037, 13 p., 2012. [13] S. Dahlke, G. Kutyniok, P. Maass, C. Sagiv, H.-G. Stark, and G. Teschke. The uncertainty principle associated with the continuous shearlet transform. Int. J. Wavelets Multiresolut. Inf. Process., 6(2):157–181, 2008. [14] S. Dahlke, G. Kutyniok, G. Steidl, and G. Teschke. Shearlet coorbit spaces and associated Banach frames. Appl. Comput. Harmon. Anal., 27(2):195–214, 2009. [15] S. Dahlke, G. Steidl, and G. Teschke. The continuous shearlet transform in arbitrary space dimensions. J. Fourier Anal. Appl., 16(3):340–364,, 2010. [16] S. Dahlke, G. Steidl, and G. Teschke. Shearlet coorbit spaces: Compactly supported analyzing shearlets, traces and embeddings. J. Fourier Anal. Appl., 17(6):1232–1255, 2011. [17] I. Daubechies. Orthonormal bases of compactly supported wavelets. Comm. Pure Appl. Math., 41(7):909–996, 1988. [18] I. Daubechies, A. Grossmann, and Y. Meyer. Painless nonorthogonal expansions. J. Math. Phys., 27(5):1271–1283, May 1986. [19] H. G. Feichtinger. Banach convolution algebras of Wiener type. In Proc. Conf. on Functions, Series, Operators, Budapest 1980, volume 35 of Colloq. Math. Soc. Janos Bolyai, pages 509– 524. North-Holland, Amsterdam, Eds. B. Sz.-Nagy and J. Szabados. edition, 1983. [20] H. G. Feichtinger. Modulation spaces on locally compact Abelian groups. Technical report, January 1983. [21] H. G. Feichtinger. Banach spaces of distributions defined by decomposition methods. II. Math. Nachr., 132:207–237, 1987. [22] H. G. Feichtinger. Atomic characterizations of modulation spaces through Gabor-type repre- sentations. In Proc. Conf. Constructive Function Theory, volume 19 of Rocky Mountain J. Math., pages 113–126, 1989. 28 [23] H. G. Feichtinger. New results on regular and irregular sampling based on Wiener amalgams. In K. Jarosz, editor, Function Spaces, Proc Conf, Edwardsville/IL (USA) 1990, volume 136 of Lect. Notes Pure Appl. Math., pages 107–121, New York, 1992. Marcel Dekker. [24] H. G. Feichtinger. Modulation spaces of locally compact Abelian groups. In R. Radha, M. Kr- ishna, and S. Thangavelu, editors, Proc. Internat. Conf. on Wavelets and Applications, pages 1–56, Chennai, January 2002, 2003. New Delhi Allied Publishers. [25] H. G. Feichtinger. Modulation Spaces: Looking Back and Ahead. Sampl. Theory Signal Image Process., 5(2):109–140, 2006. [26] H. G. Feichtinger. Banach Gelfand triples for applications in physics and engineering. volume 1146 of AIP Conf. Proc., pages 189–228. Amer. Inst. Phys., 2009. [27] H. G. Feichtinger. Choosing Function Spaces in Harmonic Analysis, volume 4 of Excur- sions in Harmonic Analysis . The February Fourier Talks at the Norbert Wiener Center. Birkhauser, 2015. [28] 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, pages 77–105. Cham: Springer, 2015. [29] H. G. Feichtinger and P. Grobner. Banach spaces of distributions defined by decomposition methods. I. Math. Nachr., 123:97–120, 1985. [30] H. G. Feichtinger and K. Grochenig. A unified approach to atomic decompositions via inte- grable group representations. Lect. Notes in Math., 1302:52–73, 1988. [31] H. G. Feichtinger and K. Grochenig. Banach spaces related to integrable group representations and their atomic decompositions, I. J. Funct. Anal., 86(2):307–340, 1989. [32] H. G. Feichtinger and K. Grochenig. Banach spaces related to integrable group representations and their atomic decompositions, II. Monatsh. Math., 108(2-3):129–148, 1989. [33] H. G. Feichtinger and K. Grochenig. Irregular sampling theorems and series expansions of band-limited functions. J. Math. Anal. Appl., 167(2):530–556, 1992. [34] H. G. Feichtinger and M. Pap. Hyperbolic wavelets and multiresolution in the Hardy space of the upper half plane. Blaschke Products and Their Applications: Fields Institute Commu- nications, 65:193–208, 2013. [35] J. Fell, H. Fuhr, and F. Voigtlaender. Resolution of the wavefront set using general continuous wavelet transforms. J. Fourier Anal. Appl., pages 1–62, 2015. [36] M. Frazier and B. Jawerth. Decomposition of Besov spaces. Indiana Univ. Math. J., 34:777– 799, 1985. [37] M. Frazier and B. Jawerth. The ϕ-transform and applications to distribution spaces. In Function Spaces and Applications, Proc. US-Swed. Seminar, Lund/Swed, Lect. Notes Math. 1302, pages 223–246. 1988. [38] M. Frazier and B. Jawerth. A discrete transform and decompositions of distribution spaces. J. Funct. Anal., 93(1):34–170, 1990. [39] H. Fuhr. Wavelet frames and admissibility in higher dimensions. J. Math. Phys., 37(12):6353– 6366, 1996. [40] H. Fuhr. Generalized Calderon conditions and regular orbit spaces. Colloq. Math., 120(1):103– 126, 2010. [41] H. Fuhr. Coorbit spaces and wavelet coefficient decay over general dilation groups. Transac- tions of the American Mathematical Society, 367(10):7373–7401, 2015. [42] H. Fuhr. Wavelet coorbit theory in higher dimensions: An overview. In Sampling Theory and Applications (SampTA), 2015 International Conference on, pages 63–67. IEEE, 2015. [43] H. Fuhr and R. Raisi-Tousi. Simplified vanishing moment criteria for wavelets over general dilation groups, with applications to abelian and shearlet dilation groups. arXiv preprints, 2015. arxiv.org/abs/1407.0824. [44] H. Fuhr and F. Voigtlaender. Wavelet coorbit spaces viewed as decomposition spaces. J. Funct. Anal., (1):80–154, 2015. [45] P. Grobner. Banachraume glatter Funktionen und Zerlegungsmethoden. PhD thesis, Univer- sity of Vienna, 1992. [46] K. Grochenig. Describing functions: atomic decompositions versus frames. Monatsh. Math., 112(3):1–41, 1991. [47] P. Grohs. Continuous shearlet frames and resolution of the wavefront set. Monatsh. Math., 164(4):393–426, 2011. 29 [48] A. Grossmann, J. Morlet, and T. Paul. Transforms associated to square integrable group representations. I: General results. J. Math. Phys., 26:2473–2479, 1985. [49] W. Guo, D. Fan, H. Wu, and G. Zhao. Sharpness of complex interpolation on α-modulation spaces. J. Fourier Anal. Appl., pages 1–35, 2015. [50] P. Goupillaud, A. Grossmann, and J. Morlet. Cycle-octave and related transforms in seismic signal analysis. Geoexploration, 23(1):85 – 102, 1984. [51] K. Guo and D. Labate. Optimally sparse multidimensional representation using shearlets. SIAM J. Math. Anal., 39(1):298–318, 2007. [52] J. Han and B. Wang. α-modulation spaces (I) scaling, embedding and algebraic properties. J. Math. Soc. Japan, 66(4):1315–1373, 2014. [53] M. Kobayashi and M. Sugimoto. The inclusion relation between Sobolev and modulation spaces. J. Funct. Anal., 260(11):3189 – 3208, June 2011. [54] G. Kutyniok and D. Labate. Resolution of the wavefront set using continuous shearlets. Trans. Amer. Math. Soc., 361(5):2719–2754, 2009. [55] G. Kutyniok, D. Labate, W. Q. Lim, and G. Weiss. Sparse multidimensional representation using shearlets. In Wavelets XI (San Diego, CA, 2005), volume 5914, pages 254–262. SPIE, 2005. [56] P. G. Lemari´e and Y. Meyer. Ondelettes et bases hilbertiennes. (Wavelets and Hilbert bases). Rev. Mat. Iberoam., 2:1–18, 1986. [57] Y. Meyer. Constructions de bases orthonorm´ees d'ondelettes. (Construction of orthonormal bases of wavelets). Rev. Mat. Iberoam., 4(1):31–39, 1988. [58] K. A. Okoudjou. Embedding of some classical Banach spaces into modulation spaces. Proc. Amer. Math. Soc., 132(6):1639–1647 (electronic), 2004. [59] M. Pap. The voice transform generated by a representation of the Blaschke group on the weighted Bergman spaces. Annales Univ. Sci. Budapest., Sect. Comp., 33:321–342, 2010. [60] M. Pap. Properties of the voice transform of the Blaschke group and connections with atomic decomposition results in the weighted Bergman spaces. J. Math. Anal. Appl., 389(1):340–350, 2012. [61] M. Pap and F. Schipp. The voice transform on the Blaschke group I. Pure Math. Appl. (PU.M.A.), 17(3-4):387–395, 2006. [62] J. Peetre. New thoughts on Besov spaces. Duke University Mathematics Series, No. 1. Math- ematics Department, Duke University, 1976. [63] S. Pilipovi´c and M. Vuleti´c. Characterization of wave front sets by wavelet transforms. Tohoku Math. J. (2), 58(3):369–391, 2006. [64] F. Ricci and M. H. Taibleson. Boundary values of harmonic functions in mixed norm spaces and their atomic structure. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5), 10:1–54, 1983. [65] M. Sugimoto and N. Tomita. The dilation property of modulation spaces and their inclusion relation with Besov spaces. J. Funct. Anal., 248(1):79–106, 2007. [66] J. Toft. Convolutions and embeddings for weighted modulation spaces. In Advances in Pseudo-differential Operators, volume 155 of Oper. Theory Adv. Appl., pages 165–186. Birkhauser, Basel, 2004. [67] J. Toft and P. Wahlberg. Embeddings of α-modulation spaces. Pliska Stud. Math. Bulgar., 21:25–46, 2012. [68] H. Triebel. Fourier Analysis and Function Spaces. Selected topics. B. G. Teubner, Leipzig, 1977. [69] H. Triebel. Theory of Function Spaces., volume 78 of Monographs in Mathematics. Birkhauser, Basel, 1983. [70] F. Voigtlaender. Embedding Theorems for Decomposition Spaces with Applica- 2015. to Wavelet Coorbit Spaces. PhD thesis, RWTH Aachen University, tions http://publications.rwth-aachen.de/record/564979. [71] F. Voigtlaender. Embeddings of decomposition spaces into Sobolev and BV spaces. ArXiv e-prints, Jan 2016. http://arxiv.org/abs/1405.2782. Faculty of Mathematics, University of Vienna E-mail address: [email protected] Lehrstuhl A fur Mathematik, RWTH Aachen University E-mail address: [email protected] 30
1804.08718	1	1804	2018-04-23T20:20:49	A characterization for fractional integral and its commutators in Orlicz and generalized Orlicz-Morrey spaces on spaces of homogeneous type	[ "math.FA" ]	In this paper, we investigate the boundedness of maximal operator and its commutators in generalized Orlicz-Morrey spaces on the spaces of homogeneous type. As an application of this boundedness, we give necessary and sufficient condition for the Adams type boundedness of fractional integral and its commutators in these spaces. We also discuss criteria for the boundedness of these operators in Orlicz spaces.	math.FA	math	A characterization for fractional integral and its commutators in Orlicz and generalized Orlicz-Morrey spaces on spaces of homogeneous type Vagif S. Guliyev Department of Mathematics, Ahi Evran University, Kirsehir, Turkey Institute of Mathematics and Mechanics, Baku, Azerbaijan [email protected] Fatih Deringoz Department of Mathematics, Ahi Evran University, Kirsehir, Turkey [email protected] Abstract In this paper, we investigate the boundedness of maximal operator and its commutators in generalized Orlicz-Morrey spaces on the spaces of ho- mogeneous type. As an application of this boundedness, we give necessary and sufficient condition for the Adams type boundedness of fractional in- tegral and its commutators in these spaces. We also discuss criteria for the boundedness of these operators in Orlicz spaces. AMS Mathematics Subject Classification: Key words: Orlicz space; Generalized Orlicz-Morrey space; Maximal operator; Fractional integral; Commutator; Spaces of homogeneous type 42B20, 42B25, 42B35 1 Introduction In the 1970s, in order to extend the theory of Calder´on-Zygmund singular in- tegrals to a more general setting, R. Coifman and G. Weiss introduced certain topological measure spaces which are equipped with a metric which is compatible with the given measure in a sense. These spaces are called spaces of homogeneous type. In this work, we find necessary and sufficient conditions for the boundedness of fractional integral and its commutators in Orlicz and generalized Orlicz-Morrey spaces on spaces of homogeneous type. 1 As a generalization of Lp(Rn), the Orlicz spaces were introduced by Birnbaum- Orlicz in [2] and Orlicz in [29], since then, the theory of the Orlicz spaces them- selves has been well developed and the spaces have been widely used in probabil- ity, statistics, potential theory, partial differential equations, as well as harmonic analysis and some other fields of analysis. They have been thoroughly investi- gated, and two excellent monographs [22] and [31] are available on this subject. Also [3] provides a good overview on the subject. The spaces Mp,ϕ(Rn) defined by the norm kf kMp,ϕ := sup x∈Rn, r>0 ϕ(r)−1 B(x, r)− 1 p kf kLp(B(x,r)) with a function ϕ positive and measurable on Rn×(0, ∞) are known as generalized Morrey spaces. For certain functions ϕ, the spaces Mp,ϕ(Rn) reduce to some p , where 0 ≤ λ ≤ n, then Mp,ϕ is the classical spaces. For instance, if ϕ(r) = r classical Morrey space Mp,λ. λ−n The classical result by Hardy-Littlewood-Sobolev states that if 1 < p < q < ∞, then the fractional integral (also known as Riesz potential) Iα (0 < α < n) is bounded from Lp(Rn) to Lq(Rn) if and only if α = n(cid:16) 1 Littlewood-Sobolev theorem is an important result in the fractional integral the- ory and the potential theory. Later then, this result has been extended from Lebesgue spaces to various function spaces. q(cid:17) The Hardy- p − 1 Around the 1970's, the Hardy-Littlewood-Sobolev inequality is extended from Lebesgue spaces to Morrey spaces. As stated in [30], Spanne proved the following result. Theorem 1.1. (Spanne, but published by Peetre [30]) Let 0 < α < n, 1 < p < n α, 0 < λ < n − αp. Moreover, let 1 p − 1 q . Then the operator Iα is bounded from Mp,λ(Rn) to Mq,µ(Rn). n and λ p = µ q = α Later on, a stronger result was obtained by Adams [1], and reproved by Chiarenza and Frasca [6]. Theorem 1.2. (Adams [1]) Let 0 < α < n, 1 < p < n p − 1 1 n−λ. Then the operator Iα is bounded from Mp,λ(Rn) to Mq,λ(Rn). α, 0 < λ < n − αp and q = α For the boundedness of Iα on generalized Morrey spaces see [17, 18, 23, 32] and references therein. The fractional integral in Orlicz spaces was studied in [7, 24, 28, 33]. For more details we refer to survey paper [27]. Commutators of classical operators of harmonic analysis play an important role in various topics of analysis and PDE, see for instance [4, 5, 10, 11], where in particular in [5] it was shown that the commutator [b, Iα] is bounded from Lp(Rn) to Lq(Rn) for 1 < p < n n and b ∈ BMO(Rn). p − α α, 1 q = 1 2 In order to extend the traditional Euclidean space to build a general underly- ing structure for the real harmonic analysis, the notion of spaces of homogeneous type was introduced by Coifman and Weiss [8]. Let X = (X, d, µ) be a space of homogeneous type, i.e. X is a topological space endowed with a quasi-distance d and a positive measure µ such that d(x, y) ≥ 0 and d(x, y) = 0 if and only if x = y, d(x, y) = d(y, x), d(x, y) ≤ K1(d(x, z) + d(z, y)), (1.1) the balls B(x, r) = {y ∈ X : d(x, y) < r}, r > 0, form a basis of neighborhoods of the point x, µ is defined on a σ-algebra of subsets of X which contains the balls, and 0 < µ(B(x, 2r)) ≤ K2 µ(B(x, r)) < ∞, (1.2) where Ki ≥ 1 (i = 1, 2) are constants independent of x, y, z ∈ X and r > 0. As usual, the dilation of a ball B = B(x, r) will be denoted by λB = B(x, λr) for every λ > 0. Note that (1.2) implies that µ(λB) ≤ C(µ, λ) µ(B), (1.3) for all λ ≥ 1. In the sequel, we always assume that µ(X) = ∞, the space of compactly supported continuous function is dense in L1(X, µ) and that X is Q-homogeneous (Q > 0), i.e. K −1 3 rQ ≤ µ(B(x, r)) ≤ K3rQ, (1.4) where K3 ≥ 1 is a constant independent of x and r. The n-dimensional Euclidean space Rn is n-homogeneous. In proving the boundedness of the fractional integral operators on various spaces, some researchers find that the translation invariance and the doubling properties of the Lebesgue measure play an important role. This is also true in studying other operators such as maximal operators and various types of singular integral operators. Thus, inspired by this fact, they studied the operators in the homogeneous setting. We refer to [12, 16, 19, 25, 26] and references therein. The authors introduced generalized Orlicz-Morrey spaces in [13] to investigate the boundedness of maximal and singular operators. Generalized Orlicz-Morrey spaces unify Orlicz and generalized Morrey spaces. Also, in [14] the authors extended the Adams type boundedness of Riesz potential and its commutators to the generalized Orlicz-Morrey spaces on the n-dimensional Euclidean space Rn. Moreover, the authors find criteria for the boundedness of Riesz potential and its commutators on Orlicz spaces on the n-dimensional Euclidean space Rn 3 in [20]. The purpose of this paper is to extend these results to the spaces of homogeneous type. Before describing the characterization for fractional integral and its commu- tators in Orlicz and generalized Orlicz-Morrey spaces on spaces of homogeneous type, we give several examples of spaces of homogeneous type ([8, 9, 12, 16]). (1) X = Rn, ρ(x, y) = x − y =(cid:16) nPj=1 (xj − yj)2(cid:17) 1 measure. 2 and µ equals Lebesgue (2) X = Rn, ρ(x, y) = (xj − yj)αj , where α1, α2, . . . , αn are positive nPj=1 numbers, not necessarily equal, and equals Lebesgue measure (this distance is called nonisotropic). (3) X = [0, 1), ρ(x, y) is the length of the smallest dyadic interval containing x and y, and µ is Lebesgue measure. (4) Any C ∞ compact Riemannian manifold with the Riemannian metric and volume. (5) Let G be a nilpotent Lie group with a left-invariant Riemannian metric and µ is the induced measure. (6) When X is the boundary of a smooth and bounded pseudo-convex domain in Cn one can introduce a nonisotropic quasi-distance that is related to the complex structure in such a way that we obtain a space of homogeneous type by using Lebesgue surface measure. For example, if X is the surface of the unit sphere σ2n−1 =nz ∈ Cn : z · z = nXj=1 zjzj = 1o, the nonisotropic distance is given by d(z, w) = 1 − z · w 1 2 . By A . B we mean that A ≤ CB with some positive constant C independent of appropriate quantities. If A . B and B . A, we write A ≈ B and say that A and B are equivalent. 2 Preliminaries The Morrey spaces and weak Morrey spaces on spaces of homogeneous type are defined as follows. Definition 2.1. Let 1 ≤ p < ∞ and 0 ≤ λ ≤ Q, Mp,λ(X) =(cid:26)f ∈ Lloc W Mp,λ(X) =(cid:26)f ∈ Lloc p (X) : kf kMp,λ := sup x∈X, r>0 p (X) : kf kW Mp,λ := sup x∈X, r>0 r− λ p kf kLp(B(x,r)) < ∞(cid:27) , p kf kW Lp(B(x,r)) < ∞(cid:27) , r− λ 4 where kf kLp(B(x,r)) =(cid:18)ZB(x,r) f (y)pdµ(y)(cid:19) 1 p and W Lp(B(x, r)) denotes the weak Lp-space of measurable functions f for which kf kW Lp(B(x,r)) = sup τ >0 τ µ(cid:0){y ∈ B(x, r) : f (y) > τ }(cid:1) 1 p . We recall the definition of Young functions. Definition 2.2. A function Φ : [0, ∞) → [0, ∞] is called a Young function if Φ is convex, left-continuous, Φ(r) = ∞. lim r→+0 Φ(r) = Φ(0) = 0 and lim r→∞ From the convexity and Φ(0) = 0 it follows that any Young function is in- If there exists s ∈ (0, ∞) such that Φ(s) = ∞, then Φ(r) = ∞ for creasing. r ≥ s. Let Y be the set of all Young functions Φ such that 0 < Φ(r) < ∞ for 0 < r < ∞ If Φ ∈ Y, then Φ is absolutely continuous on every closed interval in [0, ∞) and bijective from [0, ∞) to itself. The Orlicz spaces and weak Orlicz spaces on spaces of homogeneous type are defined as follows. Definition 2.3. For a Young function Φ, LΦ(X) =(cid:26)f ∈ Lloc 1 (X) :ZX Φ(ǫf (x))dµ(x) < ∞ for some ǫ > 0(cid:27) , kf kLΦ = inf(cid:26)λ > 0 :ZX W LΦ(X) :=(cid:26)f ∈ L1 loc(X) : sup r>0 λ (cid:17)dµ(x) ≤ 1(cid:27) , Φ(cid:16) f (x) Φ(r)m(cid:16)r, ǫf(cid:17) < ∞ for some ǫ > 0(cid:27) , kf kW LΦ = inf(cid:26)λ > 0 : sup t>0 Φ( t λ )df (t) ≤ 1(cid:27) , where df (t) = {x ∈ Rn : f (x) > t}. We note that, kχB kW LΦ = kχB kLΦ = 1 Φ−1 (µ(B)−1) , (2.1) 5 where B is a µ-measurable set in X with µ(B) < ∞ and χB is the characteristic function of B, that ZX kf kLΦ(cid:17)dµ(x) ≤ 1 Φ(cid:16) f (x) Φ( sup t>0 t kf kW LΦ )df (t) ≤ 1. (2.2) (2.3) and that For a Young function Φ and 0 ≤ s ≤ ∞, let Φ−1(s) = inf{r ≥ 0 : Φ(r) > s} (inf ∅ = ∞). If Φ ∈ Y, then Φ−1 is the usual inverse function of Φ. We note that Φ(Φ−1(r)) ≤ r ≤ Φ−1(Φ(r)) for 0 ≤ r < ∞. A Young function Φ is said to satisfy the ∆2-condition, denoted by Φ ∈ ∆2, if Φ(2r) ≤ kΦ(r) for r > 0 for some k > 1. If Φ ∈ ∆2, then Φ ∈ Y. A Young function Φ is said to satisfy the ∇2-condition, denoted also by Φ ∈ ∇2, if Φ(r) ≤ 1 2k Φ(kr), r ≥ 0, for some k > 1. ∞ For a Young function Φ, the complementary function eΦ(r) is defined by eΦ(r) =(cid:26) sup{rs − Φ(s) : s ∈ [0, ∞)} , r ∈ [0, ∞), The complementary function eΦ is also a Young function and eeΦ = Φ. If Φ(r) = r, then eΦ(r) = 0 for 0 ≤ r ≤ 1 and eΦ(r) = ∞ for r > 1. 1/p + 1/p′ = 1 and Φ(r) = rp/p, then eΦ(r) = rp′/p′. If Φ(r) = er − r − 1, then eΦ(r) = (1 + r) log(1 + r) − r. Note that Φ ∈ ∇2 if and only if eΦ ∈ ∆2. It is If 1 < p < ∞, , r = ∞. known that for r ≥ 0. (2.4) Note that by the convexity of Φ and concavity of Φ−1 we have the following r ≤ Φ−1(r)eΦ−1(r) ≤ 2r properties (cid:26) Φ(αt) ≤ αΦ(t), Φ(αt) ≥ αΦ(t), if if 0 ≤ α ≤ 1 α > 1 and (cid:26) Φ−1(αt) ≥ αΦ−1(t), Φ−1(αt) ≤ αΦ−1(t), if if 0 ≤ α ≤ 1 α > 1. (2.5) 6 The following analogue of the Holder inequality is known, ZX f (x)g(x)dµ(x) ≤ 2kf kLΦkgkL eΦ . (2.6) In the next sections where we prove our main estimates, we use the following lemma, which follows from Holder inequality, (2.1) and (2.4). Lemma 2.4. For a Young function Φ and B = B(x, r), the following inequality is valid where kf kLΦ(B) = kf χBkLΦ. kf kL1(B) ≤ 2µ(B)Φ−1(cid:0)µ(B)−1(cid:1) kf kLΦ(B), 3 Generalized Orlicz-Morrey spaces The generalized Orlicz-Morrey spaces and the weak generalized Orlicz-Morrey spaces on spaces of homogeneous type are defined as follows. Definition 3.1. Let (X, d, µ) be Q−homogeneous, ϕ(r) be a positive measurable function on (0, ∞) and Φ any Young function. We denote by MΦ,ϕ(X) the gen- eralized Orlicz-Morrey space, the space of all functions f ∈ Lloc Φ (X) with finite quasinorm kf kMΦ,ϕ ≡ kf kMΦ,ϕ(Rn) = sup x∈X,r>0 ϕ(r)−1Φ−1(µ(B(x, r))−1)kf kLΦ(B(x,r)), where Lloc all balls B ⊂ X. Φ (X) is defined as the set of all functions f such that f χB ∈ LΦ(X) for Also by W MΦ,ϕ(X) we denote the weak generalized Orlicz-Morrey space of all functions f ∈ W Lloc Φ (X) for which kf kW MΦ,ϕ ≡ kf kW MΦ,ϕ(Rn) = sup x∈X,r>0 ϕ(r)−1Φ−1(µ(B(x, r))−1)kf kW LΦ(B(x,r)) < ∞, where W Lloc for all balls B ⊂ X. Φ (X) is defined as the set of all functions f such that f χB ∈ W LΦ(X) Remark 3.2. Thanks to (1.4) and (2.5) we have Φ−1(µ(B(x, r))−1) ≈ Φ−1(r−Q). Therefore we can also write kf kMΦ,ϕ ≡ sup x∈X,r>0 ϕ(r)−1Φ−1(r−Q)kf kLΦ(B(x,r)), kf kW MΦ,ϕ ≡ sup x∈X,r>0 ϕ(r)−1Φ−1(r−Q)kf kW LΦ(B(x,r)), and respectively. 7 According to this definition, we recover the generalized Morrey space Mp,ϕ(X) and weak generalized Morrey space W Mp,ϕ(X) under the choice Φ(r) = rp, 1 ≤ p < ∞. If Φ(r) = rp, 1 ≤ p < ∞ and ϕ(r) = r , 0 ≤ λ ≤ Q, then MΦ,ϕ(X) and W MΦ,ϕ(X) coincide with Mp,λ(X) and W Mp,λ(X), respectively and if ϕ(r) = Φ−1(r−Q), then MΦ,ϕ(X) and W MΦ,ϕ(X) coincide with the LΦ(X) and W LΦ(X), respectively. λ−Q p A function ϕ : (0, ∞) → (0, ∞) is said to be almost increasing (resp. almost decreasing) if there exists a constant C > 0 such that ϕ(r) ≤ Cϕ(s) (resp. ϕ(r) ≥ Cϕ(s)) for r ≤ s. For a Young function Φ, we denote by GΦ the set of all almost decreasing functions ϕ : (0, ∞) → (0, ∞) such that t ∈ (0, ∞) 7→ ϕ(t) Φ−1(t−Q) is almost increasing. Lemma 3.3. Let B0 := B(x0, r0). If ϕ ∈ GΦ, then there exist C > 0 such that 1 ϕ(r0) ≤ kχB0kMΦ,ϕ ≤ C ϕ(r0) . Proof. Let B = B(x, r) denote an arbitrary ball in X. By the definition and (2.1), it is easy to see that kχB0kMΦ,ϕ = sup x∈X,r>0 ϕ(r)−1Φ−1(µ(B)−1) 1 Φ−1(µ(B ∩ B0)−1) ≥ ϕ(r0)−1Φ−1(µ(B0)−1) 1 Φ−1(µ(B0 ∩ B0)−1) = 1 ϕ(r0) . Now if r ≤ r0, then ϕ(r0) ≤ Cϕ(r) and ϕ(r)−1Φ−1(µ(B)−1)kχB0kLΦ(B) ≤ 1 ϕ(r) ≤ C ϕ(r0) . On the other hand if r ≥ r0, then ϕ(r0) Φ−1(µ(B0)−1) ≤ C ϕ(r) Φ−1(µ(B)−1) and ϕ(r)−1Φ−1(µ(B)−1)kχB0kLΦ(B) ≤ C ϕ(r0) . This completes the proof. 4 Maximal operator and its commutators in gen- eralized Orlicz-Morrey spaces Let Mf (x) be the maximal function, i.e. Mf (x) = sup r>0 1 µ(B(x, r))ZB(x,r) 8 f (y)dµ(y). The known boundedness statement for M in Orlicz spaces on spaces of ho- mogeneous type runs as follows. Theorem 4.1. [16] Let Φ ∈ Y. Then M is bounded from LΦ(X) to W LΦ(X). Moreover, if Φ ∈ ∇2, then M is bounded from LΦ(X) to LΦ(X). Lemma 4.2. Let Φ ∈ Y, f ∈ Lloc Φ (X) and B = B(x, r). Then for any Young function Φ ∈ ∇2 and kMf kLΦ(B) . kMf kW LΦ(B) . 1 t>r Φ−1(cid:0)r−Q(cid:1) sup Φ−1(cid:0)r−Q(cid:1) sup t>r 1 Φ−1(cid:0)t−Q(cid:1) kf kLΦ(B(x,t)) Φ−1(cid:0)t−Q(cid:1) kf kLΦ(B(x,t)) for any Young function Φ. (4.1) (4.2) Proof. Let Φ ∈ ∇2. We put f = f1 + f2, where f1 = f χB(x,2kr) and f2 = f χ ∁ B(x,2kr), where k is the constant from the triangle inequality (1.1). Estimation of Mf1: By Theorem 4.1 we have kMf1kLΦ(B) ≤ kMf1kLΦ(X) . kf1kLΦ(X) = kf kLΦ(B(x,2kr)). By using the monotonicity of the functions kf kLΦ(B(x,t)), Φ−1(cid:0)t(cid:1) with respect to t and (2.5) we get, Consequently we have 1 Φ−1(cid:0)r−Q(cid:1) sup t>2kr ≥ Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x,t)) kf kLΦ(B(x,2kr)) Φ−1(cid:0)r−Q(cid:1) sup t>2kr kMf1kLΦ(B) . 1 Φ−1(cid:0)r−Q(cid:1) sup t>r Φ−1(cid:0)t−Q(cid:1) & kf kLΦ(B(x,2kr)). Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x,t)) (4.3) (4.4) ∁ (B(x, 2kr)) 6= Estimation of Mf2: Let y be an arbitrary point from B. If B(y, t)∩ if z ∈ B(y, t) ∩ ∁ (B(x, 2kr)), then t > d(y, z) ≥ ∅, then t > r. 1 k d(x, z) − d(x, y) > 2r − r = r. On the other hand, B(y, t)∩ Indeed, ∁ (B(x, 2kr)) ⊂ B(x, 2kt). Indeed, if z ∈ B(y, t)∩ ∁ (B(x, 2kr)), then we get d(x, z) ≤ kd(y, z) + kd(x, y) < kt + kr < 2kt. 9 Therefore, Mf2(y) = sup t>0 ≤ sup t>r ≤ sup t>r = sup t>2kr 1 1 µ(B(y, t))ZB(y,t)∩ µ(B(y, t))ZB(x,2kt) µ(B(y, 2kt))ZB(x,2kt) µ(B(y, t))ZB(x,t) C C f (z)dµ(z) ∁ (B(x,2kr)) f (z)dµ(z) f (z)dµ(z) f (z)dµ(z) by the doubling condition (1.3). Hence by Lemma 2.4 and (1.4) Mf2(y) . sup t>2kr µ(B(x, t)) µ(B(y, t)) Φ−1(cid:0)µ(B(x, t))−1(cid:1)kf kLΦ(B(x,t)) . sup t>r Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x,t)) (4.5) Thus the function Mf2(y), with fixed x and r, is dominated by the expression not depending on y. Then we integrate the obtained estimate for Mf2(y) in y over B, we get kMf2kLΦ(B) . 1 Φ−1(cid:0)r−Q(cid:1) sup t>r Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x,t)) (4.6) Gathering the estimates (4.4) and (4.6) we arrive at (4.1). Let now Φ be an arbitrary Young function. It is obvious that kMf kW LΦ(B) ≤ kMf1kW LΦ(B) + kMf2kW LΦ(B). By the boundedness of the operator M from LΦ(X) to W LΦ(X), provided by Theorem 4.1, we have kMf1kW LΦ(B) . kf kLΦ(B(x,2kr)). By using (4.3), (4.5) and (2.1) we arrive at (4.2). Theorem 4.3. Let Φ ∈ Y, the functions ϕ1, ϕ2 and Φ satisfy the condition sup r<t<∞ Φ−1(cid:0)t−Q(cid:1) ess inf t<s<∞ ϕ1(s) Φ−1(cid:0)s−Q(cid:1) ≤ C ϕ2(r), (4.7) where C does not depend on r. Then the maximal operator M is bounded from MΦ,ϕ1(X) to W MΦ,ϕ2(X) and for Φ ∈ ∇2, the operator M is bounded from MΦ,ϕ1(X) to MΦ,ϕ2(X). 10 Proof. Note that (cid:16) ess inf x∈A f (x)(cid:17)−1 = ess sup x∈A 1 f (x) is true for any real-valued nonnegative function f and measurable on A and the fact that kf kLΦ(B(x,t)) is a nondecreasing function of t kf kLΦ(B(x,t)) ϕ1(s) ess inf 0<t<s<∞ Φ−1(cid:0)s−Q(cid:1) = ess sup 0<t<s<∞ ≤ sup s>0,x∈X ϕ1(s) Φ−1(cid:0)s−Q(cid:1)kf kLΦ(B(x,t)) Φ−1(cid:0)s−Q(cid:1)kf kLΦ(B(x,s)) ϕ1(s) = kf kMΦ,ϕ1 . Since (ϕ1, ϕ2) and Φ satisfy the condition (4.7), sup r<t<∞ ≤ sup r<t<∞ kf kLΦ(B(x,t)) ϕ1(s) kf kLΦ(B(x,t))Φ−1(cid:0)t−Q(cid:1) Φ−1(cid:0)s−Q(cid:1) r<t<∞(cid:16) ess inf ess inf t<s<∞ t<s<∞ sup ≤ Ckf kMΦ,ϕ1 ≤ Cϕ2(r)kf kMΦ,ϕ1 Then by (4.1) kMf kMΦ,ϕ2 . sup x∈X,r>0 . sup x∈X,r>0 ϕ1(s) ϕ1(s) ess inf t<s<∞ Φ−1(cid:0)s−Q(cid:1) Φ−1(cid:0)t−Q(cid:1) Φ−1(cid:0)s−Q(cid:1)(cid:17) Φ−1(cid:0)t−Q(cid:1) Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x,t)) ϕ1(r)−1Φ−1(cid:0)r−Q(cid:1) kf kLΦ(B(x,r)) sup t>r ϕ2(r) 1 = kf kMΦ,ϕ1 The estimate kMf kW MΦ,ϕ2 local estimate (4.2). . kf kMΦ,ϕ1 can be proved similarly by the help of The commutators generated by b ∈ L1 loc(X) and the maximal operator M is defined by Mb(f )(x) = sup t>0 µ(B(x, t))−1 ZB(x,t) b(x) − b(y)f (y)dµ(y). We recall that the space BMO(X) = {b ∈ L1 loc(X) : kbk∗ < ∞} is defined by the seminorm kbk∗ := sup x∈X,r>0 1 µ(B(x, r))ZB(x,r) 11 b(y) − bB(x,r)dµ(y) < ∞, where bB(x,r) = BMO-functions: 1 kbk∗ ≈ sup where 1 ≤ p < ∞, and 1 µ(B(x,r))RB(x,r) b(y)dµ(y). We will need the following properties of x∈X,r>0(cid:18) µ(B(x, r))ZB(x,r) (cid:12)(cid:12)bB(x,r) − bB(x,t)(cid:12)(cid:12) ≤ Ckbk∗ ln b(y) − bB(x,r)pdµ(y)(cid:19) 1 for 0 < 2r < t, (4.8) (4.9) p , t r where C does not depend on b, x, r and t. Next, we recall the notion of weights. Let w be a locally integrable and positive function on X. The function w is said to be a Muckenhoupt A1 weight if there exists a positive constant C such that for any ball B 1 µ(B)ZB w(x)dµ(x) ≤ C ess inf x∈B w(x). Lemma 4.4. [16, Chapter 1] Let ω ∈ A1, then the reverse Holder inequality holds, that is, there exist q > 1 such that (cid:18) 1 µ(B)ZB w(x)qdµ(x)(cid:19) 1 q . 1 µ(B)ZB w(x)dµ(x) for all balls B. Lemma 4.5. Let Φ be a Young function with Φ ∈ ∆2. Then we have 1 2µ(B)ZB f (x)dµ(x) ≤ Φ−1(cid:0)µ(B)−1(cid:1) kf kLΦ(B) ≤ C(cid:18) 1 µ(B)ZB for some 1 < p < ∞. f (x)pdµ(x)(cid:19) 1 p Proof. The left-hand side inequality is just Lemma 2.4. Next we prove the right-hand side inequality. Our idea is based on [21]. Take g ∈ L eΦ with kgkL eΦ on L eΦ(X) from Theorem 4.1. Let Q := kMkL eΦ→L eΦ ≤ 1. Note thateΦ ∈ ∇2 since Φ ∈ ∆2, therefore M is bounded and define a function Rg(x) := M kg(x) (2Q)k , ∞Xk=0 where For every g ∈ L eΦ with kgkL eΦ properties: M kg := g k = 0, k = 1, Mg M(M k−1g) k ≥ 2. ≤ 1, the function Rg satisfies the following 12 • g(x) ≤ Rg(x) for almost every x ∈ X; • kRgkL eΦ ≤ 2kgkL eΦ • M(Rg)(x) ≤ 2QRg(x), that is, Rg is a Muckenhoupt A1 weight with the A1 constant less than or equal to 2Q. By Lemma 4.4, there exist positive constants q > 1 and C independent of g such that for all balls B, (cid:18) 1 µ(B)ZB Rg(x)qdµ(x)(cid:19) 1 q ≤ C µ(B)ZB Rg(x)dµ(x). By Theorem 2.6 and Lemma 2.4, we obtain kRgkLq(B) = µ(B)1/q(cid:18) 1 ≤ Cµ(B)−1/q′ ≤ µ(B)1/q C q Rg(x)qdµ(x)(cid:19) 1 µ(B)ZB Φ−1(cid:0)µ(B)−1(cid:1) ≤ Cµ(B)−1/q′ kRgkL eΦ µ(B)ZB Φ−1(cid:0)µ(B)−1(cid:1) . 1 Thus we have Rg(x)dµ(x) Since the Luxemburg-Nakano norm is equivalent to the Orlicz norm we get ZB f (x)g(x)dµ(x) ≤ZB f (x)Rg(x)dµ(x) ≤ kf kLq′ (B)kRgkLq(B) q′ 1 f (x)q′ dµ(x)(cid:19) 1 ≤ C(cid:18) 1 µ(B)ZB Φ−1(cid:0)µ(B)−1(cid:1). kf kLΦ(B) ≤ sup(cid:26)(cid:12)(cid:12)(cid:12)(cid:12)ZB f (x)g(x)dµ(x)(cid:12)(cid:12)(cid:12)(cid:12) : g ∈ L eΦ, kgkL eΦ ≤ 1(cid:27) ≤ C(cid:18) 1 µ(B)ZB Φ−1(cid:0)µ(B)−1(cid:1) . dµ(x)(cid:19) 1 f (x)q′ 1 q′ Consequently, the right-hand side inequality follows with p = q. We have the following result from (4.8) and Lemma 4.5. Lemma 4.6. Let b ∈ BMO(X) and Φ be a Young function with Φ ∈ ∆2. Then kbk∗ ≈ sup x∈X,r>0 Φ−1(cid:0)r−Q(cid:1)(cid:13)(cid:13)b(·) − bB(x,r)(cid:13)(cid:13)LΦ(B(x,r)) . The known boundedness statements for the commutator operator Mb on Or- licz spaces run as follows, see [15, Theorem 1.9 and Corollary 2.3]. Note that in [15] a more general case of multi-linear commutators was studied. 13 Theorem 4.7. Let Φ be a Young function with Φ ∈ ∆2 ∩ ∇2 and b ∈ BMO(X). Then Mb is bounded on LΦ(X) and the inequality kMbf kLΦ ≤ C0kbk∗kf kLΦ (4.10) holds with constant C0 independent of f . Lemma 4.8. Let Φ be a Young function with Φ ∈ ∆2 ∩ ∇2, b ∈ BMO(X), then the inequality kMbf kLΦ(B(x0,r)) . kbk∗ Φ−1(cid:0)r−Q(cid:1) sup t>r(cid:16)1 + ln t r(cid:17)Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x0,t)) holds for any ball B(x0, r) and for all f ∈ Lloc Φ (X). Proof. For B = B(x0, r), write f = f1 + f2 with f1 = f χ2kB and f2 = f χ ∁ where k is the constant from the triangle inequality (1.1), so that (2kB) , kMbf kLΦ(B) ≤ kMbf1kLΦ(B) + kMbf2kLΦ(B) . By the boundedness of the operator Mb in the space LΦ(X) provided by Theorem 4.7, we obtain kMbf1kLΦ(B) ≤ kMbf1kLΦ(X) . kbk∗ kf1kLΦ(X) = kbk∗ kf kLΦ(2B). (4.11) As we proceed in the proof of Lemma 4.2, we have for x ∈ B Mb(f2)(x) . sup t>2kr b(y) − b(x)f (y)dµ(y). 1 µ(B(x0, t))ZB(x0,t) µ(B(x0, t))ZB(x0,t) µ(B(x0, t))ZB(x0,t) 1 1 Then t>r kMbf2kLΦ(B) .(cid:13)(cid:13)(cid:13)(cid:13)sup . J1 + J2 =(cid:13)(cid:13)(cid:13)(cid:13)sup +(cid:13)(cid:13)(cid:13)(cid:13)sup t>2r t>r 1 µ(B(x0, t))ZB(x0,t) For the term J1 by (1.4) and (2.1) we obtain b(y) − b(·)f (y)dµ(y)(cid:13)(cid:13)(cid:13)(cid:13)LΦ(B) b(y) − bBf (y)dµ(y)(cid:13)(cid:13)(cid:13)(cid:13)LΦ(B) . b(·) − bBf (y)dµ(y)(cid:13)(cid:13)(cid:13)(cid:13)LΦ(B) µ(B(x0, t))ZB(x0,t) 1 b(y) − bBf (y)dµ(y) J1 ≈ 1 Φ−1(cid:0)r−Q(cid:1) sup t>r 14 and split it as follows: J1 . 1 1 t>r Φ−1(cid:0)r−Q(cid:1) sup Φ−1(cid:0)r−Q(cid:1) sup t>r + 1 µ(B(x0, t))ZB(x0,t) 1 µ(B(x0, t)) b(y) − bB(x0,t)f (y)dµ(y) bB(x0,r) − bB(x0,t)ZB(x0,t) f (y)dµ(y). Applying Holder's inequality, by Lemmas 2.4 and 4.6 and (4.9) we get 1 J1 . . + 1 t>r Φ−1(cid:0)r−Q(cid:1) sup Φ−1(cid:0)r−Q(cid:1) sup Φ−1(cid:0)r−Q(cid:1) sup kbk∗ t>2r t>r For J2 we obtain 1 1 µ(B(x0, t))(cid:13)(cid:13)b(·) − bB(x0,t)(cid:13)(cid:13)L eΦ(B(x0,t)) kf kLΦ(B(x0,t)) Φ−1(cid:0)t−Q(cid:1)(cid:16)1 + ln bB(x0,r) − bB(x0,t)µ(B(x0, t))Φ−1(cid:0)t−Q(cid:1) kf kLΦ(B(x0,t)) µ(B(x0, t)) J2 ≈ kb(·) − bBkLΦ(B) sup t>r f (y)dµ(y) t 1 r(cid:17)kf kLΦ(B(x0,t)). µ(B(x0, t))ZB(x0,t) Φ−1(cid:0)t−Q(cid:1)kf kLΦ(B(x0,t)) Φ−1(cid:0)t−Q(cid:1)(cid:16)1 + ln gathering the estimates for J1 and J2, we get . kbk∗ t>r Φ−1(cid:0)r−Q(cid:1) sup Φ−1(cid:0)r−Q(cid:1) sup kbk∗ t>r kMbf2kLΦ(B) . t r(cid:17)kf kLΦ(B(x0,t)). (4.12) By using (4.3) we unite (4.12) with (4.11), which completes the proof. Theorem 4.9. Let Φ be a Young function with Φ ∈ ∆2 ∩ ∇2, b ∈ BMO(X) and the functions ϕ1, ϕ2 and Φ satisfy the condition sup r<t<∞(cid:16)1 + ln t r(cid:17)Φ−1(cid:0)t−Q(cid:1) ess inf t<s<∞ ϕ1(s) Φ−1(cid:0)s−Q(cid:1) ≤ C ϕ2(r), where C does not depend on r. Then the operator Mb is bounded from MΦ,ϕ1(X) to MΦ,ϕ2(X). Proof. The proof is similar to the proof of Theorem 4.3 thanks to Lemma 4.8. (4.13) 5 Fractional integral and its commutators in Or- licz spaces For a Q-homogeneous space (X, d, µ), let Iαf (x) =ZX f (y) d(x, y)Q−α dµ(y), 0 < α < Q. For proving our main results, we need the following estimate. 15 Lemma 5.1. If B0 := B(x0, r0), then rα 0 ≤ CIαχB0(x) for every x ∈ B0. Proof. If x, y ∈ B0, then d(x, y) ≤ k(d(x, x0)+d(y, x0)) < 2kr0. Since 0 < α < Q, we get rα−Q 0 ≤ Cd(x, y)α−Q. Therefore IαχB0(x) =ZB0 d(x, y)α−Qdµ(y) ≥ Crα−Q 0 µ(B0) = Crα 0 . The known boundedness statement for Iα in Orlicz spaces on spaces of homo- geneous type runs as follows. Theorem 5.2. [25] Let (X, d, µ) be Q−homogeneous and Φ, Ψ ∈ Y. Assume that there exist constants A, A′ > 0 such that Z ∞ r tα−1Φ−1(cid:0)t−Q(cid:1) dt ≤ ArαΦ−1(cid:0)r−Q(cid:1) rαΦ−1(cid:0)r−Q(cid:1) ≤ A′Ψ−1(cid:0)r−Q(cid:1) for 0 < r < ∞, (5.1) for 0 < r < ∞. (5.2) Then Iα is bounded from LΦ(X) to W LΨ(X). Moreover, if Φ ∈ ∇2, then Iα is bounded from LΦ(X) to LΨ(X). Theorem 5.3. Let (X, d, µ) be Q−homogeneous and Φ, Ψ ∈ Y. Assume that Iα is bounded from LΦ(X) to W LΨ(X) then condition (5.2) holds. Proof. Let B0 = B(x0, r0) and x ∈ B0. By (1.4) and Lemmas 5.1 and 2.1, we have 0 . Ψ−1(r−Q rα 0 )kIαχB0kW LΨ(B0) . Ψ−1(r−Q 0 )kIαχB0kW LΨ . Ψ−1(r−Q 0 )kχB0kLΦ . Ψ−1(r−Q Φ−1(r−Q 0 0 ) ) . Since this is true for every r0 > 0, we are done. Combining Theorems 5.2 and 5.3 we have the following result. Theorem 5.4. Let (X, d, µ) be Q−homogeneous and Φ, Ψ ∈ Y. If (5.1) holds, then the condition (5.2) is necessary and sufficient for the boundedness of Iα from LΦ(X) to W LΨ(X). Moreover, if Φ ∈ ∇2, the condition (5.2) is necessary and sufficient for the boundedness of Iα from LΦ(X) to LΨ(X). The commutators generated by b ∈ L1 loc(X) and the operator Iα are defined by [b, Iα]f (x) =ZX b(x) − b(y) d(x, y)Q−α f (y)dµ(y), 0 < α < Q. 16 The operator b, Iα is defined by b, Iαf (x) =ZX b(x) − b(y) d(x, y)Q−α f (y)dµ(y), 0 < α < Q. The following lemma is the analogue of the Hedberg's trick for [b, Iα]. Lemma 5.5. If (X, d, µ) be Q−homogeneous, 0 < α < Q and f, b ∈ L1 then for all x ∈ X and r > 0 we get loc(X), ZB(x,r) f (y) d(x, y)Q−α b(x) − b(y)dµ(y) . rαMbf (x). Proof. ZB(x,r) ∞Xj=0 . f (y) d(x, y)Q−α b(x) − b(y)dµ(y) = (2−jr)α(2−jr)−QZd(x,y)<2−j r ∞Xj=0Z2−j−1r≤d(x,y)<2−j r f (y) d(x, y)Q−α b(x) − b(y)dµ(y) f (y)b(x) − b(y)dµ(y) . rαMbf (x). Lemma 5.6. If b ∈ L1 loc(X) and B0 := B(x0, r0), then rα 0 b(x) − bB0 ≤ Cb, IαχB0(x) for every x ∈ B0. Proof. The proof is similar to the proof of Theorem 5.1. Theorem 5.7. Let (X, d, µ) be Q−homogeneous, 0 < α < Q, b ∈ BMO(X) and Φ, Ψ ∈ Y. 1. If Φ ∈ ∇2 and Ψ ∈ ∆2, then the condition rαΦ−1(cid:0)r−Q(cid:1) +Z ∞ r (cid:16)1 + ln t r(cid:17)Φ−1(cid:0)t−Q(cid:1)tα dt t ≤ CΨ−1(cid:0)r−Q(cid:1) (5.3) for all r > 0, where C > 0 does not depend on r, is sufficient for the boundedness of [b, Iα] from LΦ(X) to LΨ(X). 2. If Ψ ∈ ∆2, then the condition (5.2) is necessary for the boundedness of b, Iα from LΦ(X) to LΨ(X). 3. Let Φ ∈ ∇2 and Ψ ∈ ∆2. If the condition Z ∞ r (cid:16)1 + ln t r(cid:17)Φ−1(cid:0)t−Q(cid:1)tα dt t ≤ CrαΦ−1(cid:0)r−Q(cid:1) (5.4) holds for all r > 0, where C > 0 does not depend on r, then the condition (5.2) is necessary and sufficient for the boundedness of b, Iα from LΦ(X) to LΨ(X). 17 Proof. (1) For arbitrary x0 ∈ X, set B = B(x0, r) for the ball centered at x0 and of radius r. Write f = f1 + f2 with f1 = f χ2kB and f2 = f χ ∁ , where k is the constant from the triangle inequality (1.1). (2kB) For x ∈ B we have [b, Iα]f2(x) .ZX .Z ∁ b(y) − b(x) d(x, y)Q−α f2(y)dµ(y) ≈Z ∁ d(x0, y)Q−α f (y)dµ(y) +Z ∁ b(y) − bB (2kB) (2kB) (2kB) b(y) − b(x) d(x0, y)Q−α f (y)dµ(y) b(x) − bB d(x0, y)Q−α f (y)dµ(y) = J1 + J2(x), since x ∈ B and y ∈ ∁ (2kB) implies 1 2k d(x0, y) ≤ d(x, y) ≤ (k + 1 2 )d(x0, y). Let us estimate J1. b(y) − bB d(x0, y)Q−α f (y)dµ(y) ≈Z ∁ (2kB) b(y) − bBf (y)Z ∞ d(x0,y) dt dt tQ+1−α dµ(y) b(y) − bBf (y)dµ(y) tQ+1−α (2kB) J1 =Z ∁ ≈Z ∞ 2krZ2kr≤d(x0,y)≤t .Z ∞ 2krZB(x0,t) b(y) − bBf (y)dµ(y) dt tQ+1−α. Applying Holder's inequality, by (2.4), (4.9), (4.6) and Lemma 2.4 we get dt tQ+1−α tQ+1−α f (y)dµ(y) b(y) − bB(x0,t)f (y)dµ(y) bB(x0,r) − bB(x0,t)ZB(x0,t) J1 .Z ∞ 2r ZB(x0,t) +Z ∞ .Z ∞ 2r (cid:13)(cid:13)b(·) − bB(x0,t)(cid:13)(cid:13)L eΦ(B(x0,t)) kf kLΦ(B(x0,t)) +Z ∞ bB(x0,r) − bB(x0,t)kf kLΦ(B(x0,t))Φ−1(cid:0)µ(B(x0, t))−1(cid:1) dt . kbk∗ Z ∞ r(cid:17)kf kLΦ(B(x0,t))Φ−1(cid:0)µ(B(x0, t))−1(cid:1) dt 2r (cid:16)1 + ln . kbk∗ kf kLΦZ ∞ r(cid:17)Φ−1(cid:0)t−Q(cid:1)tα dt 2r (cid:16)1 + ln t1−α . tQ+1−α . t dt dt 2r 2r t t t1−α A geometric observation shows 2kB ⊂ B(x, δ) for all x ∈ B, where δ = 18 (2k + 1)kr. Using Lemma 5.5, we get J0(x) := [b, Iα]f1(x) .Z2kB .ZB(x,δ) b(y) − b(x) d(x, y)Q−α f (y)dµ(y) b(y) − b(x) d(x, y)Q−α f (y)dµ(y) . rαMbf (x). Consequently, we have J0(x) + J1 . kbk∗rαMbf (x) + kbk∗kf kLΦZ ∞ Thus, by (5.3) we obtain J0(x) + J1 . kbk∗(cid:18)Mbf (x) Ψ−1(r−Q) Φ−1(r−Q) . t t r(cid:17)Φ−1(cid:0)t−Q(cid:1)tα dt 2r (cid:16)1 + ln + Ψ−1(r−Q)kf kLΦ(cid:19) . Choose r > 0 so that Φ−1(r−Q) = Mbf (x) C0kbk∗kf kLΦ . Then Ψ−1(r−Q) Φ−1(r−Q) = (Ψ−1 ◦ Φ)( Mbf (x) C0kbk∗kf kLΦ Mbf (x) C0kbk∗kf kLΦ ) . Therefore, we get J0(x) + J1 ≤ C1kbk∗kf kLΦ(Ψ−1 ◦ Φ)( Mbf (x) C0kbk∗kf kLΦ ). Let C0 be as in (4.10). Consequently by Theorem 4.7 we have ZB C1kbk∗kf kLΦ(cid:19) dµ(x) ≤ZB Ψ(cid:18) J0(x) + J1 ≤ZX Φ(cid:18) Mbf (x) C0kbk∗kf kLΦ(cid:19) dµ(x) Φ(cid:18) Mbf (x) kMbf kLΦ(cid:19) dµ(x) ≤ 1, i.e. kJ0(·) + J1kLΨ(B) . kbk∗kf kLΦ. (5.5) 19 In order to estimate J2, by (4.6), Lemma 2.4 and condition (5.3), we also get kJ2kLΨ(B) =(cid:13)(cid:13)(cid:13)(cid:13)Z ∁ (2kB) f (y) (2kB) (2kB) . kbk∗ ≈ kbk∗ ≈ kbk∗ b(·) − bB d(x0, y)Q−α dµ(y) f (y) d(x0, y)Q−α dµ(y) dt d(x0, y)Q−α f (y)dµ(y)(cid:13)(cid:13)(cid:13)(cid:13)LΨ(B) ≈ kb(·) − bBkLΨ(B)Z ∁ Ψ−1(cid:0)r−Q(cid:1)Z ∁ f (y)Z ∞ Ψ−1(cid:0)r−Q(cid:1)Z ∁ Ψ−1(cid:0)r−Q(cid:1)Z ∞ 2krZ2kr≤d(x0,y)<t Ψ−1(cid:0)r−Q(cid:1)Z ∞ 2r ZB(x0,t) Ψ−1(cid:0)r−Q(cid:1)Z ∞ kf kLΦ(B(x0,t))Φ−1(cid:0)t−Q(cid:1)tα−1dt Ψ−1(cid:0)r−Q(cid:1)kf kLΦ Z ∞ tαΦ−1(cid:0)t−Q(cid:1) dt t tQ+1−α dµ(y) dt . kbk∗ kf kLΦ. f (y)dµ(y) f (y)dµ(y) 1 1 1 1 1 1 (2kB) d(x0,y) dt tQ+1−α . kbk∗ . kbk∗ . kbk∗ tQ+1−α 2r 2r Consequently, we have kJ2kLΨ(B) . kbk∗ kf kLΦ. Combining (5.5) and (5.6), we get k[b, Iα]f kLΨ(B) . kbk∗kf kLΦ. By taking supremum over B in (5.7), we get k[b, Iα]f kLΨ . kbk∗kf kLΦ, (5.6) (5.7) since the constants in (5.7) do not depend on x0 and r. (2) We shall now prove the second part. Let B0 = B(x0, r0) and x ∈ B0. By Lemmas 5.6, 4.6 and 2.1 we have rα 0 . kb, IαχB0kLΨ(B0) kb(·) − bB0kLΨ(B0) . Ψ−1(r−Q 0 )kb, IαχB0kLΨ(B0) . Ψ−1(r−Q 0 )kb, IαχB0kLΨ . Ψ−1(r−Q 0 )kχB0kLΦ . Ψ−1(r−Q Φ−1(r−Q 0 0 ) ) . Since this is true for every r0 > 0, we are done. (3) The third statement of the theorem follows from the first and second parts of the theorem. 20 6 Fractional integral and its commutators in gen- eralized Orlicz-Morrey spaces The following theorem is one of our main results. Theorem 6.1. Let 0 < α < Q, Φ ∈ Y, β ∈ (0, 1) and η(t) ≡ ϕ(t)β and Ψ(t) ≡ Φ(t1/β). 1. If Φ ∈ ∇2 and ϕ(t) satisfies (4.7), then the condition tαϕ(t) +Z ∞ t rα ϕ(r) dr r ≤ Cϕ(t)β, (6.1) for all t > 0, where C > 0 does not depend t, is sufficient for boundedness of Iα from MΦ,ϕ(X) to MΨ,η(X). 2. If ϕ ∈ GΦ, then the condition tαϕ(t) ≤ Cϕ(t)β, (6.2) for all t > 0, where C > 0 does not depend t, is necessary for boundedness of Iα from MΦ,ϕ(X) to MΨ,η(X). 3. Let Φ ∈ ∇2. If ϕ ∈ GΦ satisfies the regularity condition Z ∞ t rα ϕ(r) dr r ≤ Ctαϕ(t), (6.3) for all t > 0, where C > 0 does not depend t, then the condition (6.2) is necessary and sufficient for boundedness of Iα from MΦ,ϕ(X) to MΨ,η(X). Proof. Proof of the first part of the theorem: For arbitrary ball B = B(x, t) we represent f as f = f1 + f2, f1(y) = f (y)χB(y), f2(y) = f (y)χ ∁ (B)(y), and have Iαf (x) = Iαf1(x) + Iαf2(x). For Iαf1(x), following Hedberg's trick, we obtain Iαf1(x) ≤ C1tαMf (x). For Iαf2(x) by Lemma 2.4 we have Z ∁ (B) (B) d(x,y) f (y) f (y)Z ∞ d(x, y)Q−α dµ(y) ≈Z ∁ ≈Z ∞ t Zt≤d(x,y)<r ≤ C2Z ∞ t 21 dr rQ+1−α dµ(y) dr f (y)dµ(y) rQ+1−α Φ−1(r−Q)rα−1kf kLΦ(B(x,r))dr. Consequently we have Iαf (x) . tαMf (x) +Z ∞ t . tαMf (x) + kf kMΦ,ϕZ ∞ t rαϕ(r) dr r . Φ−1(r−Q)rα−1kf kLΦ(B(x,r))dr From (6.1) we obtain Iαf (x) . min{ϕ(t)β−1Mf (x), ϕ(t)βkf kMΦ,ϕ} . sup s>0 min{sβ−1Mf (x), sβkf kMΦ,ϕ} = (Mf (x))β kf k1−β MΦ,ϕ , where we have used that the supremum is achieved when the minimum parts are balanced. Hence for every x ∈ X we have Iαf (x) . (Mf (x))β kf k1−β MΦ,ϕ . (6.4) By using the inequality (6.4) we have kIαf kLΨ(B) . k(Mf )βkLΨ(B) kf k1−β MΦ,ϕ . Note that from (2.2) we get ZB Ψ (Mf (x))β LΦ(B)! dµ(x) =ZB kMf kβ kMf kLΦ(B)(cid:19) dµ(x) ≤ 1. Φ(cid:18) Mf (x) Thus k(Mf )βkLΨ(B) ≤ kMf kβ have LΦ(B). Consequently by using this inequality we kIαf kLΨ(B) . kMf kβ LΦ(B) kf k1−β MΦ,ϕ . (6.5) From Theorem 4.3 and (6.5), we get kIαf kMΨ,η = sup η(t)−1Ψ−1(t−Q)kIαf kLΨ(B) sup x∈X,t>0 η(t)−1Ψ−1(t−Q)kMf kβ LΦ(B) = kf k1−β MΦ,ϕ (cid:18) sup x∈X,t>0 ϕ(t)−1Φ−1(t−Q)kMf kLΦ(B)(cid:19)β x∈X,t>0 . kf k1−β MΦ,ϕ . kf kMΦ,ϕ. Proof of the second part of the theorem: 22 Let B0 = B(x0, t0) and x ∈ B0. By Lemma 5.1 we have tα 0 ≤ CIαχB0(x). Therefore, by (2.1) and Lemma 3.3 we have tα 0 ≤ CΨ−1(µ(B0)−1)kIαχB0kLΨ(B0) ≤ Cη(t0)kIαχB0kMΨ,η ≤ Cη(t0)kχB0kMΦ,ϕ ≤ C η(t0) ϕ(t0) = Cϕ(t0)β−1. Since this is true for every t0 > 0, we are done. The third statement of the theorem follows from first and second parts of the theorem. Theorem 6.2. Let 0 < α < Q, Φ ∈ Y, β ∈ (0, 1) and η(t) ≡ ϕ(t)β and Ψ(t) ≡ Φ(t1/β). 1. If ϕ(t) satisfies (4.7), then the condition (6.1) is sufficient for boundedness of Iα from MΦ,ϕ(X) to W MΨ,η(X). 2. If ϕ ∈ GΦ, then the condition (6.2) is necessary for boundedness of Iα from MΦ,ϕ(X) to W MΨ,η(X). 3. If ϕ ∈ GΦ satisfies the regularity condition (6.3), then the condition (6.2) is necessary and sufficient for boundedness of Iα from MΦ,ϕ(X) to W MΨ,η(X). Proof. Proof of the first part of the theorem: By using the inequality (6.4) we have kIαf kW LΨ(B) . k(Mf )βkW LΨ(B) kf k1−β MΦ,ϕ , where B = B(x, t). Note that from (2.3) we get Ψ sup t>0 tβ kMf kβ W LΦ(B)! d(M f )β (tβ) = sup t>0 Φ(cid:18) t kMf kW LΦ(B)(cid:19) dM f (t) ≤ 1. Thus k(Mf )βkW LΨ(B) ≤ kMf kβ have W LΦ(B). Consequently by using this inequality we kIαf kW LΨ(B) . kMf kβ W LΦ(B) kf k1−β MΦ,ϕ . (6.6) From Theorem 4.3 and (6.6), we get kIαf kW MΨ,η = sup η(t)−1Ψ−1(t−Q)kIαf kW LΨ(B) x∈X,t>0 . kf k1−β MΦ,ϕ sup x∈X,t>0 η(t)−1Ψ−1(t−Q)kMf kβ W LΦ(B) = kf k1−β MΦ,ϕ (cid:18) sup x∈X,t>0 ϕ(t)−1Φ−1(t−Q)kMf kW LΦ(B)(cid:19)β . kf kMΦ,ϕ. 23 Proof of the second part of the theorem: Let B0 = B(x0, t0) and x ∈ B0. By Lemma 5.1 we have tα 0 ≤ CIαχB0(x). Therefore, by (2.1) and Lemma 3.3 tα 0 ≤ CΨ−1(µ(B0)−1)kIαχB0kW LΨ(B0) ≤ Cη(t0)kIαχB0kW MΨ,η ≤ Cη(t0)kχB0kMΦ,ϕ ≤ C η(t0) ϕ(t0) = Cϕ(t0)β−1. Since this is true for every t0 > 0, we are done. The third statement of the theorem follows from first and second parts of the theorem. The following theorem is one of our main results. Theorem 6.3. Let 0 < α < n, Φ ∈ Y, b ∈ BMO(X), β ∈ (0, 1) and η(t) ≡ ϕ(t)β and Ψ(t) ≡ Φ(t1/β). 1. If Φ ∈ ∆2 ∩ ∇2 and ϕ satisfies (4.13), then the condition rαϕ(r) +Z ∞ r (cid:16)1 + ln t r(cid:17)ϕ(t)tα dt t ≤ Cϕ(r)β, (6.7) for all r > 0, where C > 0 does not depend on r, is sufficient for the boundedness of [b, Iα] from MΦ,ϕ(X) to MΨ,η(X). 2. If Φ ∈ ∆2 and ϕ ∈ GΦ, then the condition (6.2) is necessary for the boundedness of b, Iα from MΦ,ϕ(X) to MΨ,η(X). 3. Let Φ ∈ ∆2 ∩ ∇2. If ϕ ∈ GΦ satisfies the conditions and t sup r<t<∞(cid:16)1 + ln Z ∞ r (cid:16)1 + ln r(cid:17)ϕ(t) ≤ C ϕ(r), r(cid:17)ϕ(t)tα dt t t ≤ Crαϕ(r), for all r > 0, where C > 0 does not depend on r, then the condition (6.2) is necessary and sufficient for the boundedness of b, Iα from MΦ,ϕ(X) to MΨ,η(X). Proof. For arbitrary x0 ∈ X, set B = B(x0, r) for the ball centered at x0 and of radius r. Write f = f1 + f2 with f1 = f χ2kB and f2 = f χ ∁ , where k is the constant from the triangle inequality (1.1). (2kB) If we use the same notation and proceed as in the proof of Theorem 5.7 for x ∈ B we have J0(x) + J1 . kbk∗rαMbf (x) + kbk∗ Z ∞ 2r (cid:16)1 + ln . kbk∗(cid:16)rαMbf (x) + kf kMΦ,ϕ Z ∞ 2r (cid:16)1 + ln t r(cid:17)kf kLΦ(B(x0,t))Φ−1(cid:0)t−Q(cid:1) dt t1−α t r(cid:17)ϕ(t) dt t1−α(cid:17). 24 Thus, by (6.7) we obtain J0(x) + J1 . kbk∗ min{ϕ(r)β−1Mbf (x), ϕ(r)βkf kMΦ,ϕ} . kbk∗ sup s>0 min{sβ−1Mbf (x), sβkf kMΦ,ϕ} = kbk∗(Mbf (x))β kf k1−β MΦ,ϕ. Consequently for every x ∈ B we have J0(x) + J1 . kbk∗(Mbf (x))β kf k1−β MΦ,ϕ. (6.8) By using the inequality (6.8) we have kJ0(·) + J1kLΨ(B) . kbk∗k(Mbf )βkLΨ(B) kf k1−β MΦ,ϕ. Note that from (2.2) we get ZB Ψ (Mbf (x))β LΦ(B)! dµ(x) =ZB kMbf kβ Φ(cid:18) Mbf (x) kMbf kLΦ(B)(cid:19) dµ(x) ≤ 1. Thus k(Mbf )βkLΨ(B) ≤ kMbf kβ LΦ(B). Therefore, we have kJ0(·) + J1kLΨ(B) . kbk∗kMbf kβ LΦ(B) kf k1−β MΦ,ϕ. If we also use the same notation and proceed as in the proof of Theorem 5.7, we also get kJ2kLΨ(B) . kbk∗ From this estimate and condition (6.7) we have kf kLΦ(B(x0,t))Φ−1(cid:0)t−Q(cid:1)tα−1dt. 2r 1 Ψ−1(cid:0)r−Q(cid:1)Z ∞ Ψ−1(cid:0)r−Q(cid:1) kf kMΦ,ϕ Z ∞ Ψ−1(cid:0)r−Q(cid:1) kf kMΦ,ϕϕ(r)β. kbk∗ kbk∗ . 2r kJ2kLΨ(B) . tαϕ(t) dt t Consequently by using Theorem 4.9, we get k[b, Iα]f kMΨ,η = sup η(r)−1Ψ−1(r−Q)k[b, Iα]f kLΨ(B) x0∈X,r>0 MΦ,ϕ (cid:18) sup x0∈X,r>0 . kbk∗kf k1−β . kbk∗kf kMΦ,ϕ. ϕ(r)−1Φ−1(r−Q)kMbf kLΦ(B)(cid:19)β + kbk∗kf kMΦ,ϕ 25 We shall now prove the second part. Let B0 = B(x0, r0) and x ∈ B0. By 0 b(x) − bB0 ≤ Cb, IαχB0(x). Therefore, by Lemma 4.6 Lemma 5.6 we have rα and Lemma 3.3 rα 0 ≤ C kb, IαχB0kLΨ(B0) kb(·) − bB0kLΨ(B0) C kbk∗ ≤ C kbk∗ kb, IαχB0kLΨ(B0)Ψ−1(r−Q) ≤ η(r0)kb, IαχB0kMΨ,η ≤ Cϕ2(r0)kχB0kMΦ,ϕ ≤ C η(r0) ϕ(r0) ≤ Cϕ(r0)β−1. Since this is true for every r0 > 0, we are done. The third statement of the theorem follows from the first and second parts of the theorem. References [1] D.R. Adams, A note on Riesz potentials, Duke Math. J. 42 (1975), 765-778. [2] Z. Birnbaum, W. Orlicz, ber die verallgemeinerung des begriffes der zueinan- der konjugierten potenzen. Studia Math. 3, 11767 (1931). [3] C. Bennett and R. Sharpley, Interpolation of operators, Academic Press, Boston, 1988. [4] A.P. Calder´on, Commutators of singular integral operators. Proc. Nat. Acad. Sci. U.S.A. 53, 1092-1099 (1965) [5] S. Chanillo, A note on commutators. Indiana Univ. Math. J. 31 (1), 7-16 (1982) [6] F. Chiarenza, M. Frasca, Morrey spaces and Hardy-Littlewood maximal function, Rend. Math. Appl. 1987, 7, 7, 273-279. [7] A. Cianchi, Strong and weak type inequalities for some classical operators in Orlicz spaces, J. London Math. Soc. 60 (2) (1999), no. 1, 187-202. [8] R.R. Coifman, G. Weiss, Analyse harmonique non-commutative sur certain espaces homogenes, in "Lecture Notes in Math.," No. 242, Springer-Verlag, Berlin, 1971 [9] R.R. Coifman, G. Weiss, Extensions of Hardy Spaces and their use in anal- ysis, Bull. Amer. Math. Soc. 831 (1977), 569-645. [10] R.R. Coifman, R. Rochberg, G. Weiss, Factorization theorems for Hardy spaces in several variables. Ann. of Math.(2) vol 103 (3), 611-635 (1976) 26 [11] R. Coifman, P. Lions, Y. Meyer, S. Semmes, Compensated compactness and Hardy spaces. J. Math. Pures Appl.72, 247-286 (1993) [12] D. Deng, Y. Han, Harmonic analysis on spaces of homogeneous type, Springer-Verlag, Berlin, 2009. xii+154 pp. [13] F. Deringoz, V.S. Guliyev and S.G. Samko, Boundedness of maximal and singular operators on generalized Orlicz-Morrey spaces, Operator Theory, Operator Algebras and Applications, Series: Operator Theory: Advances and Applications Vol. 242 (2014), 1-24. [14] F. Deringoz, V.S. Guliyev, S.G. Hasanov, Characterizations for the Riesz po- tential and its commutators on generalized Orlicz-Morrey spaces. J. Inequal. Appl. 2016, Paper No. 248, 22 pp. [15] X. Fu, D. Yang, W. Yuan, Boundedness of multilinear commutators of Calder´on-Zygmund operators on Orlicz spaces over non-homogeneous spaces. Taiwanese J. Math. 16, 2203-2238 (2012) [16] I. Genebashvili, A. Gogatishvili, V. Kokilashvili, M. Krbec, Weight theory for integral transforms on spaces of homogeneous type. Longman, Harlow, (1998) [17] V.S. Guliyev, Boundedness of the maximal, potential and singular operators in the generalized Morrey spaces, J. Inequal. Appl. 2009, Art. ID 503948, 20 pp. [18] V.S. Guliyev, S.S. Aliyev, T. Karaman, P. S. Shukurov, Boundedness of sublinear operators and commutators on generalized Morrey Space, Int. Eq. Op. Theory. 71 (3) (2011), 327-355. [19] V.S. Guliyev, R. Ch. Mustafayev, Fractional integrals in spaces of functions defined on spaces of homogeneous type, (Russian) Anal. Math. 24 (1998), no. 3, 181-200. [20] V.S. Guliyev, F. Deringoz, S.G. Hasanov, Riesz potential and its commuta- tors on Orlicz spaces. J. Inequal. Appl. 2017, Paper No. 75, 18 pp. [21] M. Izuki, Y. Sawano, Characterization of BMO via ball Banach function spaces. Vestn. St.-Peterbg. Univ. Mat. Mekh. Astron. 4(62) (2017), no. 1, 78-86. [22] M.A. Krasnoselskii and Ya. B. Rutickii, Convex Functions and Orlicz Spaces, English translation P. Noordhoff Ltd., Groningen, 1961. 27 [23] E. Nakai, Hardy -- Littlewood maximal operator, singular integral operators and Riesz potentials on generalized Morrey spaces, Math. Nachr. 166 (1994), 95-103. [24] E. Nakai, On generalized fractional integrals, Taiwanese J. Math. 5 (2001), no. 3, 587-602. [25] E. Nakai, On generalized fractional integrals in the Orlicz spaces on spaces of homogeneous type, Scientiae Mathematicae Japonicae, Vol.54 (2001), 473- 487. [26] E. Nakai, The Campanato, Morrey and Holder spaces on spaces of homoge- neous type. Studia Math. 176 (2006), no. 1, 1-19. [27] E. Nakai, Recent topics of fractional integrals, Sugaku Expositions, Ameri- can Mathematical Society, Vol.20, No.2 (2007), 215 -- 235. [28] R. O'Neil, Fractional integration in Orlicz spaces, Trans. Amer. Math. Soc. 115 (1965), 300-328. [29] W. Orlicz, Uber eine gewisse Klasse von Raumen vom Typus B, Bull. Acad. Polon. A (1932), 207-220; reprinted in: Collected Papers, PWN, Warszawa, 1988, 217-230. [30] J. Peetre, On the theory of Lp,λ, J. Funct. Anal. 4 (1969), 71-87. [31] M.M. Rao and Z.D. Ren, Theory of Orlicz Spaces, M. Dekker, Inc., New York, 1991. [32] S. Sugano, H. Tanaka, Boundedness of fractional integral operators on gen- eralized Morrey spaces. Sci. Math. Jpn. 58 (2003), no. 3, 531540. [33] A. Torchinsky, Interpolation of operators and Orlicz classes, Studia Math. 59 (1976), 177-207. 28
1102.2615	2	1102	2011-02-21T04:58:16	Guaranteeing Convergence of Iterative Skewed Voting Algorithms for Image Segmentation	[ "math.FA", "cs.CV", "nlin.CG" ]	In this paper we provide rigorous proof for the convergence of an iterative voting-based image segmentation algorithm called Active Masks. Active Masks (AM) was proposed to solve the challenging task of delineating punctate patterns of cells from fluorescence microscope images. Each iteration of AM consists of a linear convolution composed with a nonlinear thresholding; what makes this process special in our case is the presence of additive terms whose role is to "skew" the voting when prior information is available. In real-world implementation, the AM algorithm always converges to a fixed point. We study the behavior of AM rigorously and present a proof of this convergence. The key idea is to formulate AM as a generalized (parallel) majority cellular automaton, adapting proof techniques from discrete dynamical systems.	math.FA	math	Guaranteeing Convergence of Iterative Skewed Voting Algorithms for Image Segmentation Doru C. Balcana,∗, Gowri Srinivasab, Matthew Fickusc, Jelena Kovacevi´cd aSchool of Interactive Computing, Georgia Institute of Technology, Atlanta, USA bDept. of Information Science and Engineering, and Center for Pattern Recognition, PES School of Engineering, Bangalore, India cDept. of Mathematics and Statistics, Air Force Institute of Technology, Wright-Patterson AFB, USA dDept. of Biomedical Eng., Electrical and Computer Eng. and Center for Bioimage Informatics, Carnegie Mellon University, Pittsburgh, USA 1 1 0 2 b e F 1 2 ] . A F h t a m [ 2 v 5 1 6 2 . 2 0 1 1 : v i X r a Abstract In this paper we provide rigorous proof for the convergence of an iterative voting-based image segmentation algorithm called Active Masks. Active Masks (AM) was proposed to solve the challenging task of delineating punctate patterns of cells from fluorescence microscope images. Each iteration of AM consists of a linear convolution composed with a nonlinear thresholding; what makes this process special in our case is the presence of additive terms whose role is to "skew" the voting when prior information is available. In real-world implementation, the AM algorithm always converges to a fixed point. We study the behavior of AM rigorously and present a proof of this convergence. The key idea is to formulate AM as a generalized (parallel) majority cellular automaton, adapting proof techniques from discrete dynamical systems. Keywords: active masks, cellular automata, convergence, segmentation. 1. Introduction Let the "image" f be any real-valued function over the domain Ω := QD Recently, a new algorithm called Active Masks (AM) was proposed for the segmentation of biological images [14]. d=1 ZNd and refer to the N := N1N2. . .ND elements of Ω as pixels; here, ZNd denotes the finite group of integers modulo Nd. A segmentation of f assigns one of M possible labels to each of the N pixels in Ω. For the fluorescence microscope image depicted in Figure 1(a), one example of a successful segmentation is to label all of the background pixels as "1," assign label "2" to every pixel in the largest cell, "3" to every pixel in the second largest cell, and so on. Formally, a segmentation is a label function ψ : Ω → {1, 2, . . . , M}, or, equivalently, a collection of M binary masks µm : Ω → {0, 1} where, at any given n ∈ Ω, we have µm(n) = 1 if and only if ψ(n) = m. That is, µm at any iteration i can be defined as m := ( 1, ψi(n) = m, µ(i) 0, ψi(n) , m, In AM, these masks actively evolve according to a given rule. To understand this evolution, it helps to first discuss iterative voting: in each iteration, at any given pixel, one counts how often a given label appears in the neighborhood of that pixel -- weighting nearby neighbors more than distant ones -- and assigns the most frequent label to that pixel in the next iteration. For example, if a pixel labeled "1" in the current iteration is completely surrounded by pixels labeled "2", its label will likely change to "2" in the next iteration. Formally speaking, iterative voting is the repeated application of the rule: Iterative Voting: ψi(n) = argmax 1≤m≤M (cid:2)(µ(i−1) m ∗ g)(n)(cid:3), (1) ∗Corresponding author Email addresses: [email protected] (Doru C. Balcan), [email protected] (Gowri Srinivasa), [email protected] (Matthew Fickus), [email protected] (Jelena Kovacevi´c) 1Email: [email protected]. Mailing address: College of Computing Building, room 218, Georgia Institute of Technology, 801 Atlantic Drive, Atlanta, GA 30332-0280. Phone: (404) 385-8547, Fax: (404)894-0673. Preprint submitted to Elsevier October 2, 2018 where i is the index of the iteration, g : Ω → R is some arbitrarily chosen fixed weighting function and "∗" denotes circular convolution over Ω. Iterative voting is referred to as a convolution-threshold scheme since it simplifies to (i) 1 in the special case M = 2. Experimentation reveals that for typical low-pass filters rounding the filtered version of µ g, repeatedly applying (1) to a given initial ψ0 results in a progressive smoothing of the contours between distinctly labeled regions of Ω. Despite this nice property, note that taken by itself, iterative voting is useless as a segmentation scheme, as (1) evolves masks in a manner that is independent of any image under consideration. The AM algorithm is a generalization of (1) that contains additional image-based terms whose purpose is to drive the iteration towards a meaningful segmentation. To be precise, the AM iteration is: Active Masks: ψi(n) = argmax 1≤m≤M (cid:2)(µ(i−1) m ∗ g)(n) + Rm(n)(cid:3), (2) where the region-based distributing functions {Rm}M m=1 can be any image-dependent real-valued functions over Ω. These will be referred to as skew functions in this paper, due to their role to bias the voting. Essentially, at any given pixel n, these additional terms skew the voting towards labels m whose Rm(n) values are large. For good segmentation, one should define the Rm's in terms of features in the image that distinguish regions of interest from each other. For example, for the fluorescence microscope image given in Figure 1(a), the cells appear noticeably brighter than the background. As such, we choose R1 to be a soft-thresholded version of the image's local average brightness, and choose the remaining Rm's to be identically zero. When (2) is applied, such a choice in Rm's forces pixels which lie outside the cells towards label "1," while pixels that lie inside a cell can assume any other label. Intuitively, repeated (i) applications of (2) will cause the mask µ to converge to an indicator function of the background, while each of 1 m=2 converges either to a smooth blob contained within the foreground or to the empty set. the other masks {µ Experimentation reveals that the AM algorithm indeed often converges to a ψ which assigns a unique label to each cell provided the scale of the window g is chosen appropriately [14]; see Figure 1 for examples. (i) m }M The purpose of this paper is to provide a rigorous investigation of the convergence behavior of the AM algorithm. To be precise, we note that in a real-world implementation the AM algorithm occasionally fails to converge to a ψ which is biologically meaningful. However, even in these cases, the algorithm always seems to converge to something. Indeed, when ψ0 and the Rm's are chosen at random, experimentation reveals that the repeated application of (2) always seems to eventually produce ψi such that ψi+1 = ψi. At the same time, a simple example tempers one's expectations: taking Ω = Z4, M = 2, w = δ−1 + δ0 + δ1 and R1 = R2 = 0, we see that AM will not always converge, as repeatedly applying (2) to ψ0 = δ0 + δ2 produces the endless 2-cycle δ0 + δ2 7→ δ1 + δ3 7→ δ0 + δ2. In summary, even though random experimentation indicates that AM will almost certainly converge, there exist trivial examples which show that it will not always do so. The central question that this paper seeks to address is therefore: Under what conditions on g and {Rm}M m=1 will the AM algorithm always converge to a fixed point of (2) ? We show that when g is an even function, AM will either converge to a fixed point or will get stuck in a 2-cycle; no higher-order cycles are possible. We can further rule out 2-cycles whenever g is taken so that the convolutional operator f 7→ f ∗ g is positive semidefinite. The following is a compilation of these results: Theorem 1.1. Given any Ω := QD d=1 ZNd , initial segmentation ψ0 : Ω → {1, . . . , M} and any real-valued functions m=1 over Ω, the Active Mask algorithm, namely the repeated application of (2), will always converge to a fixed {Rm}M point of (2) provided the discrete Fourier transform of g is nonnegative and even. A preliminary version of the results in this paper appears in the conference proceeding [2]. Though the specific AM algorithm was introduced in [14], iterative lowpass filtering has long been a subject of interest in applied harmonic analysis, having deep connections to continuous-domain ideas such as diffusion and the maximum principle [8]. For instance, [9] gives an edge detection application of a discretized version of these ideas. Meanwhile, [6] gives diffusion-inspired conditions under which lowpass filtering is guaranteed to produce a coarse version of a given image. One way to prove the convergence of iterative convolution-thresholding schemes is to show that lowpass filtering decreases the number of zero-crossings in a signal; such a condition is equivalent to a version of the maximum principle [11]. More recently, the continuous-domain version of (1) has been used to model the motion of interfaces between media [12, 13]; in that setting, (1) is known to converge if M = 2. Since the AM algorithm is iterative, many of the proof techniques we use here were adapted from majority cellular automata (MCA), a well-studied class of 2 (a) Original image (b) i=0, M=256 scale= 4 scale= 16 scale= 32 Gaussian Filter i = 2 i = 8 i = 14 Final state (c) Segmentation outcomes for various scales Figure 1: Active mask segmentation of punctate patterns of proteins [14]. (a) Original image (courtesy of Linstedt Lab [1]) (b) Initialization using M=256 random masks. After one iteration of (2), the background is separated from the foreground by the region-based skew function R1. (c) Segmentation results using various scales of the voting filter at iterations 2, 8, 14 and at convergence. The first row shows a cross section of three Gaussian filters (scale=4, 16 and 32 respectively). The second row shows the segmentation result after two iterations of the algorithm. When scale=4, we observe a large number of small regions in the foreground. This is contrasted by the fewer number of regions when scale=16 and scale=32. Subsequent rows represent the state of the system at various stages of evolution. The last row represents the convergence states. Note that the algorithm converges regardless of the scale chosen, but the segmentation is only biologically meaningful at the proper scale of 16; at scales=4, 32 the cells are oversegmented or undersegmented, respectively. 3 discrete dynamical systems. Indeed, theoretical guarantees on the convergence of a symmetric class of MCA have been known for several decades; see [3, 10], and references therein. Such results were recently generalized to a quasi- symmetric class via the use of Lyapunov functionals [7]. Whereas much of traditional MCA theory focuses on the convergence of repeated applications of (1), our work differs due to the presence of the additive Rm terms in (2). The paper is organized as follows. In the next section, we use an MCA formulation of AM to prove our main convergence results. In Section 3, we then briefly discuss the generalization of our main results to a less elegant yet more realistic version of (2) involving noncircular convolution. We conclude in Section 4 with some examples illustrating our main results, as well as some experimental results indicating the AM algorithm's rate of convergence. 2. Active Masks as a Majority Cellular Automaton Cellular automata are self-evolving discrete dynamical systems [3]. They have been applied in various fields such as statistical physics, computational biology, and the social sciences. A tremendous amount of work in this area has focused on studying the convergence behavior of various types of automata. In this section, we formulate the AM algorithm (2) as an MCA in order to facilitate our understanding of its convergence behavior. To be precise, we consider a generalization of (2) in which the convolutional operator f 7→ f ∗ g may more broadly be taken to be any linear operator A from ℓ(Ω) := { f f : Ω → R} into itself: ψi(n) := min(cid:16)argmax m (cid:2)(Aµ(i−1) m )(n) + Rm(n)(cid:3)(cid:17), µ(i−1) m := ( 1, ψi−1(n) = m, 0, ψi−1(n) , m. (3) (i−1) Here, the contribution of mask m in deciding the outcome at location n at iteration i is (Aµ m )(n), and any ties are broken by choosing the smallest m corresponding to a maximal element. Note that given any initial segmentation ψ0, i=0. However, as there are only MN distinct possible configurations applying (3) ad infinitum produces a sequence {ψi}∞ for ψ : Ω → {1, . . . , M}, this sequence must eventually repeat itself. Indeed, taking minimal indices i0 and K > 0 such that ψi0+K = ψi0, the deterministic nature of (3) implies that ψi+K = ψi for all i ≥ i0. The finite sequence {ψi} is i=0 converges if and only if K = 1, which happens precisely when ψi0 called a cycle of (3) of length K. Note that {ψi}∞ is a fixed point of (3). i0 +K−1 i=i0 Thus, from this perspective, proving that (3) always converges is equivalent to proving that K = 1 regardless of one's choice of ψ0. The following result goes a long way towards this goal, showing that if A is self-adjoint, then for any ψ0 we have that the resulting K is necessarily 1 or 2. That is, if A is self-adjoint, then for any ψ0, the sequence i=0 will either converge in a finite number of iterations, or it will eventually come to a point where it forever {ψi}∞ oscillates between two distinct configurations ψi0 and ψi0+1. Theorem 2.1. If A is self-adjoint, then for any ψ0, the cycle length K of (3) is either 1 or 2. Proof. As we are not presently concerned with the rate of convergence of (3), but rather the question of whether it i=0 has already entered its cycle. That is, we reindex does converge, we may assume without loss of generality that {ψi}∞ so that ψ0 is the beginning of the K-cycle, and heretofore regard all iteration indices as members of the cyclic group ZK. We argue by contrapositive, assuming K > 2 and concluding that A is not self-adjoint. For any i = 1, . . . , K, (3) is equivalent to the system of inequalities: (Aµ (Aµ (i−1) ψi(n))(n) + Rψi(n)(n) > (Aµ(i−1) ψi(n))(n) + Rψi(n)(n) ≥ (Aµ(i−1) m )(n) + Rm(n) m )(n) + Rm(n) (i−1) if 1 ≤ m < ψi(n), if ψi(n) ≤ m ≤ M. (4a) (4b) (i−1) m )(n) + Rm(n). Moreover, in the event Here, (4b) follows from the fact that ψi(n) is a value of m that maximizes (Aµ of a tie, ψi(n) is chosen to be the least of all such maximizing m, yielding the strict inequality in (4a). For any i and n, picking m = ψi−2(n) in (4a) and (4b) leads to the subsystem of inequalities: (Aµ (Aµ (i−1) ψi(n))(n) − (Aµ (i−1) ψi(n))(n) − (Aµ (i−1) ψi−2(n))(n) + Rψi(n)(n) − Rψi−2(n)(n) > 0 (i−1) ψi−2(n))(n) + Rψi(n)(n) − Rψi−2(n)(n) ≥ 0 if ψi−2(n) < ψi(n), if ψi−2(n) ≥ ψi(n). (5a) (5b) 4 (Aµ (i−1) ψi(n))(n) − Xi∈ZK Xn∈Ω Rψi(n)(n) = Xi∈ZK ψi(n))(n) − Xi∈ZK Xn∈Ω (i−1) (Aµ Now since K > 2, there exists a pixel n for which {ψ0(n), ψ1(n), . . . , ψK−1(n)} is not of the form {a, a, . . . , a} nor of the form {a, b, a, b, . . . a, b}. At such an n, there must exist an i such that ψi−2(n) < ψi(n). Consequently, at least one inequality in (5) is strict. Thus, summing (5) over all pixels n and all cycle indices i yields: 0 < Xi∈ZK Xn∈Ω (Aµ (i−1) ψi−2(n))(n) + Xi∈ZK Xn∈Ω Rψi(n)(n) − Xi∈ZK Xn∈Ω Rψi−2(n)(n). Since ZK is shift-invariant, Xi∈ZK Rψi−2(n)(n) for any n ∈ Ω, reducing the previous equation to: 0 < Xi∈ZK Xn∈Ω (Aµ (i−1) ψi−2(n))(n) = Xi∈ZK Xn∈Ω (Aµ (i−1) ψi(n))(n) − Xi∈ZK Xn∈Ω (Aµ (i) ψi−1(n))(n), (6) where the final equality also follows from the shift-invariance of ZK. To continue, note that for any i, j ∈ ZK we have ( j) m = 1 if and only if ψ j(n) = m and so: µ (Aµ Xn∈Ω (i) ψ j(n))(n) = Xn∈Ω M Xm=1 (Aµ(i) m )(n)µ ( j) m (n) = M Xm=1 hAµ(i) m , µ ( j) m i, (7) where h f , gi := Xn∈Ω f (n)g(n) is the standard real inner product over Ω. Using (7) in (6) gives: 0 < Xi∈ZK M Xm=1 hAµ(i−1) m , µ(i) m i − Xi∈ZK M Xm=1 hAµ(i) m , µ(i−1) m i = Xi∈ZK M Xm=1 h(A − A∗)µ(i−1) m , µ(i) m i, implying A − A∗ , 0, and so A is not self-adjoint. Theorem 2.1 has strong implications for the AM algorithm (2). Indeed, it is well-known that if g is real-valued, then the adjoint of the convolutional operator A f = f ∗ g is A∗ f = f ∗ g where g(n) = g(−n) is the reversal of g. As such, if g is an even function, Theorem 2.1 guarantees that AM will always either converge or enter a 2-cycle. We now build on the techniques of the previous proof to find additional restrictions on A which suffice to guarantee convergence: Theorem 2.2. If A is self-adjoint and hA f , f i ≥ 0 for all f : Ω → {0, ±1}, then (3) always converges. Proof. In light of Theorem 2.1, our goal is to rule out cycles of length K = 2. We argue by contrapositive. That is, we assume that there exist two distinct configurations ψ0 and ψ1 which are successors of each other, and will use this fact to produce f : Ω → {0, ±1} such that hA f , f i < 0. Substituting i = 0 and m = ψ1(n) into (4a) and (4b) yields: (Aµ (Aµ (1) ψ0(n))(n) − (Aµ (1) ψ0(n))(n) − (Aµ (1) ψ1(n))(n) + Rψ0(n)(n) − Rψ1(n)(n) > 0 (1) ψ1(n))(n) + Rψ0(n)(n) − Rψ1(n)(n) ≥ 0 Similarly, letting i = 1 and m = ψ0(n) into (4a) and (4b) yields: (Aµ (Aµ (0) ψ1(n))(n) − (Aµ (0) ψ1(n))(n) − (Aµ (0) ψ0(n))(n) + Rψ1(n)(n) − Rψ0(n)(n) > 0 (0) ψ0(n))(n) + Rψ1(n)(n) − Rψ0(n)(n) ≥ 0 if ψ1(n) < ψ0(n), if ψ1(n) ≥ ψ0(n). if ψ0(n) < ψ1(n), if ψ0(n) ≥ ψ1(n). (8a) (8b) (9a) (9b) Since ψ0 and ψ1 are distinct, there exists n0 ∈ Ω such that ψ0(n0) , ψ1(n0). If ψ0(n0) < ψ1(n0), we sum (8b) and (9a) over all n ∈ Ω. If on the other hand ψ0(n0) > ψ1(n0), we sum (8a) and (9b) over all n ∈ Ω. Either way, we obtain: Applying (7) four times then gives: 0 < N Xn=1(cid:2)(Aµ (1) ψ0(n))(n) − (Aµ (1) ψ1(n))(n) + (Aµ (0) ψ1(n))(n) − (Aµ (0) ψ0(n))(n)(cid:3). M 0 < M Xm=1(cid:2)hAµ(1) m , µ(0) m i − hAµ(1) m , µ(1) m i + hAµ(0) m , µ(1) m i − hAµ(0) m , µ(0) As such, there exists at least one index m0 such that 0 > (cid:10)A(µ 5 m i(cid:3) = − (0) m0 ), (µ (1) m0 − µ m − µ(0) m ), (µ(1) Xm=1(cid:10)A(µ(1) m0)(cid:11); choose f to be µ m − µ(0) m )(cid:11). (0) m0. (1) m0 − µ (0) (1) m0 − µ The most obvious way to ensure that hA f , f i ≥ 0 for all f : Ω → {0, ±1} is for A to be positive semidefinite, that is, hA f , f i ≥ 0 for all f : Ω → R. This in turn can be guaranteed by taking A to be diagonally dominant with nonnegative diagonal entries, via the Gershgorin circle Theorem [5]. Note that in fact strict diagonal dominance guarantees that iterative voting (1) always converges in one iteration. More interesting examples can be found in the special case where A is a convolutional operator A f = f ∗ g. Indeed, letting F be the standard non-normalized discrete Fourier transform (DFT) over Ω, we have: hA f , f i = h f ∗ g, f i = 1 N(cid:10)F( f ∗ g), F f(cid:11) = 1 N(cid:10)(F f )(Fg), F f(cid:11) = 1 N Xn∈Ω 2 . (10) (Fg)(n)(cid:12)(cid:12)(cid:12) (F f )(n)(cid:12)(cid:12)(cid:12) As such, if g : Ω → R is even and (Fg)(n) ≥ 0 for all n ∈ Ω, then A is self-adjoint positive semidefinite. Moreover, it is well-known that g is real-valued and even if and only if Fg is also real-valued and even. Thus, A is self-adjoint positive semidefinite provided Fg is nonnegative and even. For such g, Theorem 2.2 guarantees that the AM algorithm (2) will always converge. These facts are summarized in Theorem 1.1, which is stated in the introduction. Examples of g that satisfy these hypotheses are given in Section 4. We emphasize that Theorem 2.2 does not require A to be positive semidefinite, but rather only that hA f , f i ≥ 0 for all f : Ω → {0, ±1}. In the case of convolutional operators, this means we truly only need (10) to hold for such f 's. As such, it may be overly harsh to require that (Fg)(n) ≥ 0 for all n ∈ Ω. Unfortunately, the problem of characterizing such g's appears difficult, as we could find no useful frequency-domain characterizations of {0, ±1}-valued functions. A spatial domain approach is more encouraging: when Ω = ZN, writing f : Ω → {0, ±1} as the difference of two characteristic functions χI1 , χI2 : Ω → {0, 1}, we have: hA f , f i = (cid:10)A(χI1 − χI2), (χI1 − χI2)(cid:11) = sum(A1,1) + sum(A2,2) − sum(A1,2) − sum(A2,1), where sum(Ai, j) denotes the sum of all entries of the submatrix of A consisting of rows from Ii and columns from I j. As such, the condition of Theorem 2.2 reduces to showing that 0 ≤ sum(A1,1) + sum(A2,2) − sum(A1,2) − sum(A2,1) for all choices of subsets Ii and I j of ZN. We conclude this section by noting that (3) is similar to threshold cellular automata (TCA) [3, 4]. In fact, (3) is (i−1) equivalent to TCA in the special case of M = 2; in this case, µ 0 (i−1) (n) = 1 − µ 1 (n) for all n ∈ Ω, implying: (i−1) (Aµ 1 (i−1) )(n) + R1(n) > (Aµ 0 )(n) + R0(n) ⇐⇒ (cid:2)A(µ ⇐⇒ nA(cid:2)µ (i−1) 1 (i−1) 1 (i−1) ⇐⇒ (Aµ 1 (i−1) ⇐⇒ (Aµ 1 (i−1) − µ 0 )(cid:3)(n) + (R1 − R0)(n) > 0 2 (R1 − R0 − A1)(n) > 0 )(cid:3)o(n) + (R1 − R0)(n) > 0 (i−1) − (1 − µ 1 )(n) + 1 )(n) + b(n) > 0, where b(n) := 1 2 (R1 − R0 − A1)(n). That is, when M = 2, the AM algorithm is equivalent to a threshold-like decision. But whereas the traditional method for proving the convergence of TCA involves associated quadratic Lyapunov functionals [4], our method for proving the convergence of AM is more direct, being closer in spirit to that of [10]. 3. Beyond symmetry Up to this point, we have focused on the convergence of (3) in the special case where A is self-adjoint. In this section, we discuss how Theorems 2.1 and 2.2 generalize to the case of quasi-self-adjoint operators, which arise in real-world implementation of the AM algorithm. To clarify, up to this point, we have let the image f and weights g d=1 ZNd and have taken the convolutions in (2) and (3) to be circular. In real-world implementation, the use of such circular convolutions can result in poor segmentation, as values at one edge of the image are used to influence the segmentation at the unrelated opposite edge. be functions over the finite abelian group Ω = QD One solution to this problem -- implemented in [14] -- is to redefine the set of pixels as a subset Ω := QD d=1[0, Nd) of the D-dimensional integer lattice ZD, and regard our image f as a member of ℓ(Ω) := { f : ZD → R f (n) = 0 ∀n < Ω}. Here, the label function ψ and masks µm are regarded as {1, . . . , M}- and {0, 1}-valued members of ℓ(Ω), respectively, and the (noncommutative) convolution of any f , g ∈ ℓ(Ω) with g ∈ ℓ2(ZD) is defined as f ⋆ g ∈ ℓ(Ω), ( f ⋆ g)(n) := , ∀n ∈ Ω, ( f ∗ g)(n) (χΩ ∗ g)(n) 6 (11) where χΩ is the characteristic function of Ω, and ∗ denotes standard (noncircular) convolution in ℓ2(ZD). For the theory below, we need to place additional restrictions on g, namely that it belongs to the class: In this setting, for a given g ∈ G(Ω), the AM algorithm (2) becomes: G(Ω) := {g ∈ ℓ2(ZD) : (χΩ ∗ g)(n) > 0 ∀n ∈ Ω}. Noncircular Active Masks: ψi(n) = argmax 1≤m≤M (cid:2)(µ(i−1) m ⋆ g)(n) + Rm(n)(cid:3), µ(i−1) m := ( 1, ψi−1(n) = m, 0, ψi−1(n) , m. (12) Note that the use of the ⋆-convolution in (12) ensures that any "missing votes" are not counted in favor of any label m. Moreover, the denominator of (11) ensures that when n is close to an edge of Ω, the weights in the g-neighborhood (i) m ⋆ g)(n) = 1 for all n ∈ Ω, avoiding m=1(µ of n are rescaled so as to always sum to one. This rescaling ensures that PM any need to modify the skew functions Rm near the boundary. We then ask the question: for what g will (12) always converge? The key to answering this question is to realize that the ⋆-filtering operation A f = f ⋆ g can be factored as A = DB, where B is the standard filtering operator B f = f ∗ g and (D f )(n) = λn f (n), where λn = [(χΩ ∗ g)(n)]−1. Here, A, B and D are all regarded as linear operators from ℓ(Ω) into itself. More generally, we inquire into the convergence of: ψi(n) = argmax 1≤m≤M (cid:2)(Aµ(i−1) m )(n) + Rm(n)(cid:3), µ(i−1) m := ( 1, ψi−1(n) = m, 0, ψi−1(n) , m, (13) where A = DB and D is positive-multiplicative, that is, (D f )(n) = λn f (n) where λn > 0 for all n ∈ Ω. In particular, we follow [7] in saying that A is quasi-self-adjoint if there exists a positive-multiplicative operator D and a self-adjoint operator B such that A = DB. This definition in hand, we have the following generalization of Theorems 2.1 and 2.2: Theorem 3.1. Let A be quasi-self-adjoint: A = DB where D is positive-multiplicative and B is self-adjoint. Then for any ψ0, the cycle length K of (13) is either 1 or 2. Moreover, if B is positive-semidefinite, then (13) always converges. Proof. We only outline the proof, as it closely follows those of Theorems 2.1 and 2.2. Let (D f )(n) = λn f (n) with λn > 0 for all n ∈ Ω. We prove the first conclusion by contrapositive, assuming K > 2. Rather than summing (5) over all n and i directly, we instead first divide each instance of (5) by the corresponding λn, and then sum. The resulting quantity is analogous to (6): 0 < Xi∈ZK Xn∈Ω 1 λn (Aµ (i−1) ψi(n))(n) − Xi∈ZK Xn∈Ω 1 λn (Aµ (i) ψi−1(n))(n) = Xi∈ZK Xn∈Ω (Bµ (i−1) ψi(n))(n) − Xi∈ZK Xn∈Ω (Bµ (i) ψi−1(n))(n). (14) Simplifying the right-hand side of (14) with (7) quickly reveals that B cannot be self-adjoint, completing this part of the proof. For the second conclusion, we again prove by contrapositive, assuming K = 2. Dividing (8a), (8b), (9a) and (9b) by λn and then summing either (8a) and (9b) over all n or (8b) and (9a) over all n gives: 0 < N Xn=1 1 λn(cid:2)(Aµ (1) ψ0(n))(n) − (Aµ (1) ψ1(n))(n) + (Aµ (0) ψ1(n))(n) − (Aµ M (0) ψ0(n))(n)(cid:3) = − Xm=1(cid:10)B(µ(1) m − µ(0) m ), (µ(1) m − µ(0) m )(cid:11), implying B is not positive semidefinite. For a result about the convergence of (12), we apply Theorem 3.1 to A = DB where λn = [(χΩ ∗ g)(n)]−1 and B f = f ∗ g. Note that we must have g ∈ G(Ω) in order to guarantee that D is positive. Moreover, B is self-adjoint if g ∈ ℓ2(ZD) is even; since g is real-valued, this is equivalent to having its classical Fourier series g ∈ L2(TD) be real-valued and even. Meanwhile, since: then B is positive semidefinite if g(x) ≥ 0 for almost every x ∈ TD. To summarize, we have: hB f , f i = h f ∗ g, f i = h f g, f i = ZTd 2 dx, g(x)(cid:12)(cid:12)(cid:12) f (x)(cid:12)(cid:12)(cid:12) Corollary 3.2. If the Fourier series of g ∈ G(Ω) is nonnegative and even, then (12) will always converge. In the next section, we discuss how to construct such windows g, along with other implementation-related issues. 7 4. Examples of Active Masks in practice In this section we present a few representative and interesting examples of filter-based cellular automata, and discuss their behavior in relation with the results we proved in the previous sections. We also present some preliminary experimental findings on the rate of convergence of AM. For ease of understanding, let us for the moment restrict ourselves to circulant iterative voting (1), namely the version of AM (2) in which all the skew functions Rm are identically zero. The simplest nonzero filter is g = δ0. The DFT of δ0 has constant value 1, and is therefore nonnegative and even. As such, Theorem 1.1 guarantees that (1) will always converge. Of course, we already knew that: since f ∗ δ0 = f for all f ∈ ℓ(Ω), (1) will always converge in one step; as noted above, the same holds true for any g whose convolutional operator is strictly diagonally dominant with a nonnegative diagonal: g(0) ≥ Pn,0 g(n). More interesting examples arise from box filters: symmetric cubes of Dirac δ's. For instance, fix N ≥ 3 and consider (1) over Ω = ZN where g = δ−1 + δ0 + δ1. Since g is symmetric, Theorem 2.1 guarantees that (1) will either always converge or will enter a 2-cycle. However, if N is even, then (1) will not always converge, since ψ0 = δ0 + δ2 + · · · + δN−2 generates a 2-cycle. This phenomenon is depicted in Figure 2(a). This simple example shows that symmetry alone does not suffice to guarantee convergence; one truly needs additional hypotheses on g, such as the requirement in Theorem 1.1 that its DFT is nonnegative. This hypothesis does not hold for g = δ−1 + δ0 + δ1, since (Fg)(n) = 1 + 2 cos( 2πn N ). Similar issues arise in the two-dimensional setting Ω = ZN1 × ZN2: both the 3 × 3 box filter (Moore's automaton, see Figure 2(b)) and the "plus" filter (von Neumann's automaton, see Figure 2(c)) are symmetric, meaning their cycle lengths are either 1 or 2, but neither are positive semidefinite, having DFTs of [1 + 2 cos( 2πn1 ), respectively. Indeed, when N1 and N2 are even, N1 alternating stripes generate a 2-cycle for the box filter, while the checkerboard generates a 2-cycle for the plus filter. Of course, it is not difficult to find filters g which do satisfy the hypotheses of Theorem 1.1: one may simply let g be the inverse DFT of any nonnegative even function. More concrete examples, such as a discrete Gaussian over ZN, can be found using the following process. Let h : R → R be an even Schwartz function whose Fourier transform is nonnegative; an example of such a function is a continuous Gaussian. Let g be the N-periodization of the integer n′=−∞ h(n + Nn′). Then g is even, and moreover, by the Poisson summation formula: )] and 1 + 2 cos( 2πn1 N1 )][1 + 2 cos( 2πn2 N2 ) + 2 cos( 2πn2 N2 samples of h, namely g(n) := P∞ g(n′)e− 2πinn′ (Fg)(n) = N−1 N = Xn′=0 N−1 Xn′=0 ∞ Xn′′=−∞ h(n′ + Nn′′)e− 2πinn′ N = ∞ Xk=−∞ h(k)e− 2πink N = ∞ Xk=−∞ h(k + n N ) ≥ 0. In particular, if g is chosen as a periodized version of the integer samples of any zero-mean Gaussian, then Theorem 1.1 gives that the AM algorithm (2) necessarily converges. This construction method immediately generalizes to higher- dimensional settings where D > 1. It also generalizes to the noncircular convolution setting considered in Section 3. There, we further restrict h to be strictly positive, and let g be the integer samples of h. The positivity of h implies (χΩ ∗ g)(n) > 0 for all n ∈ Ω, implying g ∈ G(Ω) as needed. Moreover, g is even and the Poisson summation formula h(k + x) ≥ 0. Any g constructed in this manner satisfies the gives that its Fourier series is nonnegative: g(x) = P∞ k=−∞ hypotheses of Corollary 3.2, implying the corresponding noncirculant AM (12) necessarily converges. 4.1. The rate of convergence of the AM algorithm Up to this point, we have focused on the question of whether or not the AM algorithm (2) converges. Having settled that question to some degree, our focus now turns to another question of primary importance in real-world implementation: at what rate does AM converge? Experimentation reveals that this rate highly depends on the con- figuration of the boundary between two distinctly labeled regions of Ω. This led us to postulate that the number of boundary crossings (see Figure 3) should monotonically decrease with each iteration. Experimentation reveals that this number indeed often decreases extremely rapidly, regardless of the scale of g. Figure 4 depicts such an experiment for the fluorescence microscope image shown in Figure 1(a). Starting from a random initial configura- tion of 64 masks, we used a Gaussian filter under three different scales, with each plot depicting the evolution of 5 independently-initialized runs of the algorithm. We emphasize the algorithm's fast rate of convergence: the vertical axis represents a nested four-fold application of the natural logarithm to the number of boundary crossings. We leave a more rigorous investigation of the AM algorithm's rate of convergence for future work. 8 (a) 3-tap filter automaton (b) Moore's automaton (c) von Neumann's automaton Figure 2: An illustration of oscillating states produced by various automata: (a) The 3-tap box filter δ−1 + δ0 + δ1 over Ω = Z4. Using this g in (1) with ψ0 = δ0 + δ2 results in the endless 2-cycle δ0 + δ2 7→ δ1 + δ3 7→ δ0 + δ2. This is because at each iteration, each pixel's two neighbors will outvote him in deciding his label in the next iteration. (b) Convergence is also an issue in two dimensions, as illustrated by Moore's automaton -- a 3 × 3 box filter -- over Ω = Z4 × Z4. (c) Two-cycles persist in two dimensions even when the box filter is replaced by the smoother "plus" filter of von Neumann's automaton. In all three cases, these filters are even and so Theorem 2.1 ensures that the cycle length K of (1) is either 1 or 2. However, none of them are positive semidefinite, as their DFTs attain negative values. As such, the convergence guarantees of Theorem 2.2 do not hold. Figure 3: An illustration of the zero-crossings in an image with M=3 masks. 9 Figure 4: The rate of decrease of the AM algorithm in terms of the number of boundary crossings. Acknowledgments We thank Prof. Adam D. Linstedt and Dr. Yusong Guo for providing the biological images which were the original inspiration for the AM algorithm and this work. Fickus and Kovacevi´c were jointly supported by NSF CCF 1017278. Fickus received additional support from NSF DMS 1042701 and AFOSR F1ATA00183G003, F1ATA00083G004 and F1ATA0035J001. Kovacevi´c also received support from NIH R03-EB008870. The views expressed in this article are those of the authors and do not reflect the official policy or position of the United States Air Force, Department of Defense, or the U.S. Government. References [1] Linstedt Lab at CMU. http://www.cmu.edu/bio/faculty/linstedt.html. [2] D.C. Balcan, G. Srinivasa, M. Fickus, and J. Kovacevi´c. Convergence behavior of the Active Mask segmentation algorithm. In IEEE ICASSP, 2010. [3] E. Golés and S. Martínez. Neural and Automata Networks. Dynamical Behavior and Applications. Kluwer, 1990. [4] E. Golés-Chacc, Fr. Fogelman-Soulie, and D. Pellegrin. Decreasing energy functions as a tool for studying threshold networks. Discrete Applied Mathematics, 12:261 -- 277, 1985. [5] G. Golub and Ch. L. Van Loan. Matrix Computations. Johns Hopkins University Press, London, 3rd edition, 1996. [6] J.J. Koenderink. The structure of images. Journ. Bio. Cybern., 50(5):363 -- 370, 1984. [7] P. Melatagia Yonta and R. Ndoundam. Opinion dynamics using majority functions. Mathematical Social Sciences, 57:223 -- 244, 2009. [8] L. Nirenbarg. A strong maximum principle for parabolic equations. Commun. Pure and Appl. Math., 6:167 -- 177, 1953. [9] P. Perona and J. Malik. Scale-space and edge detection using anisotropic diffusion. IEEE Trans. Patt. Anal. and Mach. Intelligence, 12(7):629 -- 639, 1990. [10] S. Poljak and M. Sura. On periodical behaviour in societies with symmetric influences. Combinatorica, 3(1):119 -- 121, 1983. [11] R. A. Hummel. Readings in computer vision: issues, problems, principles, and paradigms, chapter Representations based on zero-crossings in scale-space, pages 753 -- 758. Morgan Kaufmann Readings Series archive. Morgan Kaufmann Publishers, Inc., CA, USA, 1987. [12] S. J. Ruuth. Efficient algorithms for diffusion-generated motion by mean curvature. Journ. Comp. Phys., 144:603 -- 625, 1998. [13] S. J. Ruuth and B. Merriman. Convolution-thresholding methods for interface motion. Journ. Comp. Phys., 169:678 -- 707, 2001. [14] G. Srinivasa, M. Fickus, Y. Guo, A. Linstedt, and J. Kovacevi´c. Active mask segmentation of fluorescence microscope images. IEEE Trans. Imag. Proc., 18(8):1817 -- 1829, 2009. 10
1609.05998	2	1609	2017-05-02T12:53:11	Duality and Geodesics for Probabilistic Frames	[ "math.FA" ]	Probabilistic frames are a generalization of finite frames into the Wasserstein space of probability measures with finite second moment. We introduce new probabilistic definitions of duality, analysis, and synthesis and investigate their properties. In particular, we formulate a theory of transport duals for probabilistic frames and prove certain properties of this class. We also investigate paths of probabilistic frames, identifying conditions under which geodesic paths between two such measures are themselves probabilistic frames. In the discrete case, this is related to ranks of convex combinations of matrices, while, in the continuous case, this is related to the continuity of the optimal transport plan.	math.FA	math	Duality and Geodesics for Probabilistic Frames Clare Wickman Johns Hopkins University Applied Physics Laboratory Kasso Okoudjou Department of Mathematics, University of Maryland, College Park Keywords: frames, probabilistic frames, optimal transport, Wasserstein metric, duality Abstract Probabilistic frames are a generalization of finite frames into the Wasserstein space of proba- bility measures with finite second moment. We introduce new probabilistic definitions of duality, analysis, and synthesis and investigate their properties. In particular, we formulate a theory of transport duals for probabilistic frames and prove certain properties of this class. We also inves- tigate paths of probabilistic frames, identifying conditions under which geodesic paths between two such measures are themselves probabilistic frames. In the discrete case this is related to ranks of convex combinations of matrices, while in the continuous case this is related to the continuity of the optimal transport plan. 1 Introduction 1.1 Probabilistic frames in the Wasserstein metric Frames are redundant spanning sets of vectors or functions that can be used to represent signals in a faithful but nonunique way and that provide an intuitive framework for describing and solving problems in coding theory and sparse representation. We refer to [5, 4, 19] for more details on frames and their applications. To set the notations for this paper, we recall that a set of column vectors Φ " tϕiuN i"1 Ă Rd is a frame if and only if there exist 0 ă A ď B ă 8 such that @x P Rd, Akxk2 ď N ÿi"1xx , ϕi y2 ď Bkxk2. Throughout this paper we abuse notation by also using Φ to denote rϕ1 . . . ϕNsJ, the analysis operator of the frame. The (optimal) bounds in the above inequality are the smallest and largest eigenvalues of the frame operator SΦ " ΦJΦ. 1 Continuous frames are natural generalization of frames and were introduced by Ali, Antoine, and Gazeau [1] (see also, [11]). Specifically, let X be a metrizable, locally compact space and ν be positive, inner regular Borel measure for X supported on all of X. Let H be a Hilbert space. Then a set of vectors ηi the vectors ηi that @f P H, x, i P t1, , nu, x P X( Ă H is a rank-n (continuous) frame if, for each x P X, x, i P t1, , nu( are linearly independent, and if there exist 0 ă A ď B ă 8 such Akf k2 ď n ÿi"1żX xηi x , f y2dνpxq ď Bkf k2. In this paper, we are concerned with a different generalization of frames called probabilistic frames. Developed in a series of papers [8, 10, 9], probabilistic frames are an intuitive way to gener- alize finite frames to the space of probability measures with finite second moment. The probabilistic setting is particularly compelling, given recent interest in probabilistic approaches to optimal coding, such as [15, 20]. In the new setting, the defining characteristics of a frame amount to a restriction on the mean and covariance matrix of the probability measure. Because of this characterization, a natural space to explore probabilistic frames is the Wasserstein space of probability measures with finite second moment, a metric space with distance defined by an optimal transport problem. Before we give the definitions and the concepts needed to state our results we first observe that in the simplest example, each finite frame can be associated with a probabilistic frame. Indeed, let Φ " tϕiuN i"1 αi " 1. Then the canonical α-weighted probabilistic frame associated with Φ is the probability measure µΦ,α given by i"1 be a frame and let tαiuN i"1 Ă p0, 1q be such that řN dµΦ,αpxq " N ÿi"1 αiδϕipxq. More generally, a probabilistic frame µ for Rd is a probability measure on Rd for which there exist 0 ă A ď B ă 8 such that for all x P Rd, Akxk2 ďżRdxx , y y2dµpyq ď Bkxk2. Tight (finite, continuous, or probabilistic) frames are those for which the frame bounds are equal. While the work of this paper is limited to probabilistic frames on Rd, of interest is also the possible extension of these ideas to probabilistic frames on infinite dimensional spaces, as outlined in [17]. Probabilistic frames form a subclass of the continuous frames defined above. Indeed, defining the support of a probability measure µ on Rd as the set: supppµq :"!x P Rd s.t. for all open sets Ux containing x, µpUxq ą 0) , it is not difficult to prove that the support of any probabilistic frame is canonically associated with a rank-one continuous frame. And conversely, certain continuous frames can be rewritten as probabilistic frames. However, despite this equivalence, there is a strong difference in the tools 2 available in the different settings. We shall investigate probabilistic frames in the setting of the Wassertein metric defined on P2pRdq, the set of probability measures µ on Rd with finite second moment: M 2 2pµq :"żRd kxk2dµpxq ă 8. By [10, Theorem 5], µ is a probabilistic frame if and only if it has finite second moment and the linear span of its support is Rd. This characterization can be restated in terms of the probabilistic frame operator for µ, Sµ, which for all y P Rd satisfies: Sµy "żRdxx , y y x dµpxq. µ span Rd is equivalent to this matrix being positive definite. Equating Sµ with its matrix representation şRd xxJdµpxq, the requirement that the support of The (2-)Wasserstein distance, W2 between two probability measures µ and ν in P2pRdq is: W 2 2 pµ, νq :" inf γ $'& ijRdRd '% kx ´ yk2dγpx, yq : γ P Γpµ, νq,/. /- , where Γpµ, νq is the set of all joint probability measures γ on Rd Rd such that for all A, B Ă BpRdq, γpA Rdq " µpAq and γpRdBq " νpBq. The Monge-Kantorovich optimal transport problem is the search for the set of joint measures which induce the infimum; any such joint distribution is called an optimal transport plan. A special subclass of transport plans are those given by deterministic transport maps (or deterministic couplings), where ν can be written as the pushforward of µ by a map T, denoted ν " T#µ. That is, for all ν-integrable functions φ, żRd φpyqdνpyq "żRd φpTpxqqdµpxq. When µ is absolutely continuous with respect to Lebesgue measure [2, p. 150], then W 2 2 pµ, νq :" inf T "żRd kx ´ Tpxqk2dµpxq : T#µ " ν* . Equipped with the 2-Wasserstein distance, P2pRd, W2q is a complete, separable metric space. Con- vergence in P2pRdq is the usual weak convergence of probability measures, combined with conver- gence of the second moments. A few structural statements can be made about probabilistic frames as a subset of P2pRdq. For brevity, the probabilistic frames for Rd are denoted by PFpRdq, and PFpA, B, Rdq denotes the set of probabilistic frames in PFpRdq with optimal upper frame bound less than or equal to B and optimal lower frame bound greater than or equal to A, with 0 ă A, B ă 8. Then PFpA, B, Rdq is 3 a nonempty, convex, closed subset of P2pRdq. The nonemptiness and convexity are trivial to show. With respect to closedness, let tµnu be a sequence in PFpA, B, Rdq converging to µ P P2pRdq. Let Because y0 " argminyPSd´1żRdxx , y y2dµpxq. xx , y0 y2 ď kxk2ky0k2 ď ky0k2p1 ` kxk2q, it follows by definition of weak convergence in P2pRdq that żRdxx , y0 y2dµnpxq ÑżRdxx , y0 y2dµpxq. Since for all n, the values of şRdxx , y0 y2dµnpxq are bounded above and below by B and A, re- spectively, µ is an element of PFpA, B, Rdq. Taking A " B, this also shows the closedness of P FpA, Rdq " P FpA, A, Rdq, the set of tight probabilistic frames with frame bound A. However, the set of probabilistic frames itself is not closed, since one can construct a sequence of probabilistic frames whose lower frame bounds converge to zero: for example, a sequence of zero-mean, Gaussian measures with covariances 1 n I, n P N. 1.2 Our contributions The goal of this paper is to investigate two main topics on probabilistic frames in the setting of the Wasserstein space. The first topic is the notion of duality. For a finite frame, Φ " tϕiuN i"1 Ă Rd, a set Ψ " tψiuN i"1 Ă Rd is said to be a dual frame to it if for every x P Rd, N x " ÿi"1xx , ϕi yψi. It is known that the redundancy of frames implies among other things the existence of many dual frames. While much attention has been paid to the so-called canonical dual frame, certain recent investigations have focused on alternate duals. For example, Sobolev duals were considered in [3, 14] in relation to Σ ´ ∆ quantization. Another example is the construction of dual frames for reconstruction of signals in the presence of erasures [16]. In this paper, we introduce two other type of dualities, one dictated by the optimal transport problem, and the other grounded in the probabilistic setting we are working in. These two approaches will be developed in Section 2. The second goal of the paper is to investigate paths of probabilistic frames. Indeed, looking at the geodesic between any two probabilistic frames, it is natural to ask if the all probability measures along this path are probabilistic frames. This will be developed in Section 3. 4 2 Duality, Analysis, and Synthesis in the Set of Probabilistic Frames 2.1 Transport Duals Duality, analysis, and synthesis are well-studied objects in finite frame theory. Sobolev duals have been proposed for use in reducing error in Σ∆ quantization [3], and the authors of [15] have found optimal dual frames for random erasures. Through the lens of optimal transport, extra nuances can be found in the probabilistic setting. Given a frame Φ " tϕiuN i"1 as above, any possible dual frame to Φ can be written as: tψiuN i"1 " tS´1 Φ ϕi ` βi ´ N Φ ϕi , ϕk yβkuN i"1 ÿk"1 xS´1 (1) where tβiuN i"1 Ă Rd and SΦ is the frame operator for Φ [5, Theorem 5.6.5]. When βi " 0 for all i, we have the canonical dual to Φ, which consists of the columns of the Moore-Penrose pseudoinverse of its analysis operator. Inspired by the definition of duality above and this enumeration of the set of all possible duals to finite frames, we introduce a new notion of duality in the probabilistic context in this section. Definition 1. Let µ be a probabilistic frame on Rd. We say that a probability measure ν P P2pRdq is a transport dual to µ if there exists γ P Γpµ, νq such that ijRdRd xyJdγpx, yq " I. We denote the set of transport duals to µ by Dµ :"$'& '% ν P P2pRdq Dγ P Γpµ, νq with ijRdRd xyJdγpx, yq " I,/. /- . the duality. We recall that the canonical dual to a probabilistic frame µ defined in [8, 10, 9], was given by We let ΓDµ Ă Γpµ, νq be the set of joint distributions on RdRd with first marginal µ (π1 #γ " µ) for which ťRdRd xyJdγpx, yq " I. This is the set of couplings (joint distributions) which induce µ :" pS´1 µ q#µ, yielding the reconstruction formula x "şRdxx , y ySµydµpyq. It is easily seen that the canonical dual µ is an example of transport dual to µ. Indeed, it is clear that γ " pι S´1 µ q#µ P Γpµ, µq, where ι signifies the identity, and ijRdRd xyJdγpx, yq "żRd xpS´1 µ xqJdµpxq " SµS´1 µ " I. Therefore, for a given probabilistic frame µ, µ P Dµ. 5 In fact, for a given probabilistic frame µ, there are other transport duals corresponding to similar deterministic couplings. Generalizing the set of duals for discrete frames outlined in (1) leads to the following construction: Theorem 2. Let µ be a probabilistic frame for Rd, and let h : Rd Ñ Rd be any function in L2pRd, µq :" tf : Rd Ñ Rdşkfpxqk2 by ψhpxq " S´1 Proof. Consider µ, ψh#µ as above. Define γ :" pι, ψhq#µ P Γpµ, ψh#µq. µ x ` hpxq ´şRdxS´1 2dµpxq ă 8u. Then ψh#µ P Dµ, where ψh : Rd Ñ Rd is defined µ x , y yhpyqdµpyq. Then ijRdRd xyJdγpx, yq "żRd x„S´1 µ x ` hpxq ´żRdxS´1 xhpxqJdµpxq ´ ijRdRd " I `żRd µ x , z yhpzqdµpzqJ xpS´1 µ xqJzhpzqJdµpxqdµpzq dµpxq " I The restriction of the set of transport duals Dµ to lie inside P2pRdq is necessary, unlike in the i"1 Ă Rd denote the ?i2iei, i P t1, , du, and let ϕd`1 " 0. finite frame case. One might consider the following simple example. Let teiud standard orthonormal basis. Let tϕiud`1 Take the weights αi " 1 2i , i P N. Define i"1 be given by ϕi " Let tψiu8i"1 be given by ψi " b 2i but i e1`rpi´1q mod ds, αiδψi . Then µ1 P P2pRdq, 8 ři"1 µ1 " 2´dδ0 ` d αiδϕi . ÿi"1 i P N. Let µ2 " M 2 2pµ2q " 1 2i kψik2 " 8 ÿi"1 1 2i 2i i " 8. 8 ÿi"1 Hence, µ2 R P2pRdq. However, letting γ P PpRd Rdq be given by γ " d ÿi"1 αiδpϕi,ψiq ` 8 ÿi"d`1 αiδp0,ψiq, it is clear that γ P Γpµ1, µ2q, and ijRdRd xyJdγpx, yq " 1 2i ?i2ic 2i i d ÿi"1 eieJi " I. This example shows that the Bessel-like restriction in the definition of transport duals, requiring 6 them to lie in P2pRdq, is necessary. Given this restriction, we can assert the following theorem: Theorem 3. Let µ be a probabilistic frame. Then: (i) Each ν P Dµ is also a probabilistic frame. (ii) Dµ is a compact subset of P2pRdq with respect to the weak topology. In particular, Dµ is a closed subset of PFpRdq with respect to the weak topology on P2pRdq. Proof. (i) Suppose ν P Dµ Ă P2pRdq. Since Dµ Ă P2pRdq by definition, it is sufficient to show that supppνq spans Rd. Let us assume, on the contrary, that there exists z P Rd, z ‰ 0, such that z K w for all w P spanpsupppνqq. Pick γ P Γpµ, νq such that ť xyJdγpx, yq " I. Because for all x P supppνq, zJx " 0, kzk2 "ij xz , xyxz , y ydγpx, yq "ij xz , xyxz , y y1rsupppνqRdspx, yqdγpx, yq " 0 which is a contradiction. (ii) Consider the lifting of the dual set, ΓDµ :" tγ P Γpµ, νq s.t. ť xyJdγpx, yq " Iu. It can be shown by Prokhorov's Theorem that ΓDµ is precompact [22, Chapter 4]. That is, given tγnu Ă ΓDµ, there exists a subsequence tγnku converging weakly to a limit γ P P2pRd Rdq. With this in mind, if tνnu is a sequence in Dµ, we can choose the corresponding tγnu P ΓDµ, and let tνnku be the second marginals of a subsequence tγnku. For all ϕ P CpRdRdq satisfying for some C ą 0 ϕpx, yq ď Cp1 ` kxk2 ` kyk2q, ij ϕpx, yqdγnkpx, yq ÝÑij ϕpx, yqdγpx, yq. In particular, for all such ϕ " ϕpxq, ij ϕpxqdγnkpx, yq "ż ϕpxqdνnkpxq ÝÑij ϕpxqdγpx, yq "ż ϕpxqdpπ2 #γqpxq. #γ ": ν, so that tνnu contains a weakly convergent Thus νnk converges weakly in P2pRdq to π2 subsequence. Therefore Dµ is precompact. Now let tνnu be any convergent sequence in Dµ which has a limit ν and which forms the second marginals of tγnu Ă ΓDµ. Take again a convergent subsequence tγnku with limit γ necessarily in Γpµ, νq. Since xiyj ď 1 2pkxk2 ` kyk2q, it follows that ij xiyjdγnkpx, yq ÝÑij xiyjdγpx, yq. Then, since for each nk, ť xiyjdγnkpx, yq " δi,j , it follows that ť xiyjdγpx, yq " δi,j, and 7 therefore ν P Dµ. This shows that Dµ is also closed, and is therefore compact. The closedness in P FpRdq then follows naturally. From the definition of transport duals, it is clear that their construction depends on the creation of a probability distribution on the product space which has a predetermined second-moments matrix and first and second marginals. This is, in general, a very difficult problem, which becomes a bit more tractable for probabilistic frames supported on finite, discrete sets by appealing to tools from linear algebra. Suppose we have two frames Φ " tϕiuN i"1 and tβjuM tαiuN case, any joint distribution γ for µΦ,α and µΨ,β satisfies j"1, summing to unity. Let µΦ,α :"řN i"1 and Ψ " tψjuM j"1, and two sets of positive weights, j"1 βj δψj . In this i"1 αiδϕi, and let µΨ,β :"řM dγpx, yq " N ÿi"1 M ÿj"1 Ai,j δϕipxqδψjpyq where A P RNM with N ÿi"1 Ai,j " βj , M ÿj"1 Ai,j " αi, Ai,j ě 0 @i, j, and N N ÿi"1 ÿj"1 Ai,j " 1. That is, there is a one-to-one correspondence between ΓpµΦ,α, µΨ,βq and this set of "doubly stochas- tic" matrices, which we denote by DSpα, βq. Thus, to show that µΦ,α P DµΨ,β , one must construct a matrix A P DSpα, βq solving ΦJAΨ " I. Regarding this question, we have the following result: Theorem 4. Given frames tϕiuN j"1 for Rd with analysis operators Φ and Ψ, there exists A P DSpα, βq with ΦJAΨ " I if and only if there is no triplet pB, u, vq with B P Rdd, u P RM , v P RN such that i"1 and tψjuM # ϕJi Bψj ` ui ` vj ě 0 tracepBq ` uJα ` vJβ ă 0 Proof. Recall that we must solve the system ΦJAΨ " I, Ai,j ě 0 M ÿj"1 Ai,j " αi, N ÿi"1 Ai,j " βj (2) Defining, for a matrix B, vecpBq to be the vector formed by stacking the columns of B, we may rewrite the problem in terms of the Kronecker product. Using the following variables, K " ΨJbΦJ, 8 a " vecpAq, zN " r1 . . . 1sJ P RN , zM " r1 . . . 1sJ P RM , and t " vecpIq, I P Rdd, we have: We can combine the equations above, letting $'''''& '''''% K1 "» -- -- Ka " t pzJN b IpMMqqa " β pIpNNq b zJMqa " α ai ě 0 @i P t1, . . . , M Nu K pzJN b IpMMqq pIpNNq b zJMq fi ffifl and t1 "» -- -- t β α fi ffifl Then the problem simplifies to solving K1a " t1 such that ai ě 0 @i P t1, . . . , M Nu. By Farkas' Lemma [12, Lemma 1], either this system has a solution or there exists y P Rd2`M`N such that # yJK1 ě 0 yJt1 ă 0 (3) (4) Now write any such y as y " » -- -- B P Rdd. Then Equations (3) and (4) hold if and only if , with b P Rd2 fi ffifl b u v , u P RM , and v P RN , and let b " vecpBq with bJK ` uJpzJN b IpMMqq ` vJpIpNNq b zJMq ě 0 bJt ` uJβ ` vJα ă 0 vecpΦBΨJq ` vecpzN uJIpMMqqJ ` vecpIpNNqvzJMqJ ě 0 bJt ` uJβ ` vJα ă 0 That is, or, equivalently, ϕJi Bψj ` ui ` vj ě 0 @i, j tracepBq ` uJβ ` vJα ă 0 The simplicity of the following corollary is alluring because it connects the weighted averages of the frame vectors to the existence of the plan yielding the duality, but the condition is difficult to 9 show because of its scope. Corollary 5. If there does not exist B P Rdd such that αJΦBΨJβ ´ tracepBq ą 0 then by Farkas' Lemma the system of (2) (and its equivalents) is not solvable and the desired matrix A P DSpα, βq exists. A true converse has proven elusive. However, we can identify a few related conditions under which no transport duals whatsoever can be constructed. In particular, in the case that the frames are uniformly weighted, we have the following zero-centroid condition. d. N zN i"1 ψi " 0, then µΨ, 1 i"1 Ă Rd is a frame such has no equal-weight transport dual supported on a set of of cardinality Theorem 6. Again, take zN :" r1 . . . 1sJ P RN . Suppose that Ψ " tψiuN thatřN j"1 Ă RN denote the columns of the analysis operator Ψ, and Proof. Given Ψ as above, let tvjud let tuiud N zNq. Ψ will have a transport dual of cardinality d if and only if for some A, AΨ " rrxui , vj yss is invertible. (Here, Q " rrqi,jss denotes the entrywise definition of Q.) Each ui " s ` λi, where s " r 1 i"1 Ă RN denote the rows of some A P DSp 1 N ds P Rd, and N d 1 d zd, 1 N d ÿk"1 ÿi"1 λi k " 0 for each i P t1, ..., du λi k " 0 for each k P t1, ..., Nu so that tλiud Then i"1 has zero centroid as well and is therefore linearly dependent. Let Λ " rλ1 . . . λdsJ. d d d detpAΦq " źi"1 xλi , vi y " detpΛΨq " 0 because vi K s for all i P t1, . . . , du and since rankpΛq ď d ´ 1. źi"1 xs ` λi , vi y " źi"1 xui , vi y " As a consequence, Theorem 6 implies that no equiangular tight frame in R2 has a transport dual of cardinality 2. Remark 7. One interesting aspect of the transport duals in the context of finite discrete probabilistic frames, i.e., finite frames, is the existence of pairs of dual frames with different cardinalities. For example, one can consider the probabilistic frame given by dµ " 1 3 δϕ3 with ϕ1 " ?3 r1 0sJ, ϕ2 " r 2 δψ2 4´?3sJ is a transport dual for µ with support of with ψ1 " r 18 4´?3 different cardinality. The role of transport duals in problems such as reconstruction in the presence 2sJ, and ϕ3 " r0 1sJ. Then the probabilistic frame ν given by dν " 1 6p2`?3 4´?3 sJ and ψ2 " r ´12 4´?3 2 δϕ1 ` 1 6 δϕ2 ` 1 2 δψ1` 1 24 2 1 of erasure will be the object of future investigations. 10 2.2 Analysis and Synthesis in the Probabilistic Context In [8, 10, 9], the analysis and synthesis operators for probabilistic frames were defined analogously to those of continuous frames. Given a probabilistic frame µ, the analysis operator was defined [10, 2.2] as Its synthesis operator was Aµ : Rd Ñ L2pRd, µq given by x ÞÑ xx ,y. Aµ : L2pRd, µq Ñ Rd given by f ÞÑżRd xfpxqdµpxq. The foregoing construction of transport duals, on the other hand, begs a more probability-theoretic definition of analysis and synthesis. As defined above, the analysis operator Aµ is independent of the measure µ. Indeed, it is not clear from this definition how one could do "analysis" with one probabilistic frame followed by "synthesis" with another. However, finite frame theory itself gives us a clue about how to think about analysis and synthesis in the probabilistic context. Example 1. Consider two frames for Rd, tϕiuN basis for RN . Then the analysis operator for Φ, AΦ : Rd Ñ RN is given by i"1. Let teiuN i"1 and tψiuN i"1 Ă RN be an orthonormal ÿi"1xx , ϕi yei The synthesis operator for Ψ, AΨ : RN Ñ Rd, is given by AΦpxq " Φx " N for x P Rd. AΨpyq " ΨJy " N ÿi"1 xy , ei yψi for y P RN . Then we can compose the operators simply by writing AΨAΦpxq " choose some σ and π in ΠN , the set of permutations on N-element sets, and instead choose to do analysis and synthesis with the two frames as ři"1xx , ϕi yψi. If, however, we N AΨAΦpxq " N ÿi"1 xx , ϕσpiq yψπpiq, then it will be as if we had chosen two different finite frames to work with. This is because the ordering of the frame vectors is implicitly tied to the ordering of the reference basis teiuN i"1. Order matters! From the example, it is clear that even given the fixed reference basis, we cannot truly speak of a single analysis operator for the set tϕiuN i"1, without imposing an order on it relating it to the fixed reference basis. Similarly, for a probabilistic frame µ, there must be a reference measure η playing the role of the reference basis, and this will still lead to a family of analysis operators, each corresponding to a joint distribution γ P Γpµ, ηq. The orthogonality of the 11 reference basis in the above example turns out not to be necessary; its function is to match up frame coefficients with the appropriate vectors. What is key is that transport plans exist between the probabilistic frame and the reference measure and that the support of the reference measure is sufficient to "glue" together arbitrary probabilistic frames through analysis and synthesis. To make this idea of coefficient-matching rigorous, some technicalities about conditional prob- abilities are necessary. Conditional probabilities can be defined via the Rokhlin Disintegration Theorem [2, Theorem 5.3.1]. If µ P PpRM RNq and ν " µ1 " π1 #µ, then one can find a Borel family of probability measures tµxuxPRM Ă PpRNq which is µ1-a.e. uniquely determined such that µ "şRM µxdµ1pxq. In the language of conditional probability, for any f P CbpRM RNq, it is then meaningful to write ijRMRN fpx, yqdµpx, yq "żRN żRM fpx, yqdµpyxqdµ1pxq, with the understanding that µpxq is defined µ1-a.e. Gluings can then be constructed, which allow us to use conditional probabilities with respect to a common reference measure to construct a joint distribution between previously unrelated measures. #γ13 " µ1. Then there exists µ P PpRK RM RNq such that π1,2 Lemma 8. Gluing Lemma [2, Lemma 5.3.2] Let γ12 P PpRK RMq, γ13 P PpRK RNq such that π1 # µ " γ12 and π1,3 x1 dµ1, and µ " şRK µx1dµ1 are the disintegrations of γ12, γ13, and µ with respect to µ1, then the first statement is equivalent to µx1 P Γpγ12 #γ12 " π1 # µ " γ13. Moreover, if γ12 " şRK γ12 x1 , γ13 x1q Ă PpRM RNq for µ1-a.e. x1 P RK. x1 dµ1, γ13 " şRK γ13 Now let us consider a probabilistic frame µ and another probability measure η and take γ P Γpµ, ηq. From Lemma 8, there is a set of conditional probability measures tγpwquwPRd that are uniquely defined η-a.e. To proceed with the construction of analysis and synthesis in the probabilistic context, we will first establish a useful fact. Recall that L2pRN RM , RK , γq :" tf : RN RM Ñ RK ij kfpx, yqk2 2dγpx, yq ă 8u. Then, by condition Jensen's inequality, if f f P L2pRdRd, Rd, γq, it follows that gpwq :"şRd fpy, wqdγpywq is in L2pRd, Rd, ηq. Finally, since hpz, wq :" kzk2 P L2pRd Rd, R, γq for any γ P Γpµ, ηq provided that µ P P2pRdq, it follows that the vector-valued function ş zdγpzwq lies in L2pRd, Rd, ηq. To define analysis and synthesis operators which are more closely tied to their probabilistic frames, a reference measure must be chosen; take an absolutely continuous η P P2pRdq whose support is Rd. Given µ P PFpRdq, we define families of analysis and synthesis operators for µ with respect to η. Definition 9. tAγ µuγPΓpµ,ηq is the family of analysis operators, and for each γ P Γpµ, ηq we have: 12 Aγ µ : Rd Ñ L2pRd, Rd, ηq, is given by Aγ µpxqpwq "żRdxx , y ydγpywq. Similarly, the family of synthesis operators, tZ γ L2pRd, R, ηq Ñ Rd, given by µuγPΓpµ,ηq is defined for each γ P Γpµ, ηq by Z γ µ : Z γ µpfq " ijRdRd zfpwqdγpzwqdηpwq The class of reference measure η was chosen such that, for any probabilistic frame µ, the probabilistic analysis and synthesis operators can be constructed using deterministic couplings between η and µ. There are several interesting ways to pair disparate types of probabilistic frames with one an- other. A useful technique is the transport of an absolutely continuous measure to a discrete measure using power (Voronoi) cells. Following [18], we define maps which can be used for these pairings. It is an interesting fact due to Brenier that the Voronoi mapping we will describe, T w P , is in fact an optimal map between the two measures it couples, µ and T w P #µ, when µ is absolutely continuous with respect to Lebesgue measure [18, Theorem 1]. Definition 10. Given a probability measure µ on Rd, a finite set P of points in Rd and w : P Ñ R` a weight vector, the power diagram or weighted Voronoi diagram of pP, wq is a decomposition of Rd into cells corresponding to each member of P . Given p P P , a point x P Rd belongs to Vorw Pppq if and only if for every q P P , kx ´ pk2 ´ wppq ď kx ´ qk2 ´ wpqq. P be the map that assigns to each x in a power cell Vorw Let T w power cell. We call T w P the weighted Voronoi mapping. T w P #µ " ÿpPP µpVorw Pppqqδp. Pppq to p, the "center" of that dηpxq. P ppq Let η be an absolutely continuous measure in P2pRdq, and let ν " řpPP λpδp be a discrete measure in P2pRdq supported on a finite set of points P with weights tλpu summing to unity. Then we say that a vector weight w : P Ñ R` is adapted to pη, νq if for all p P P , λp " ηpVorw Pppqq " şVorw Example 2. Now given discrete frames Φ " tϕiuM measure in Definition 9, choose γ1 " pι, T w1 are adapted to pµΦ, ηq and pµΨ, ηq, respectively. Then µΦpxqq "ż xx , T w1 j"1 for Rd, and η a reference Ψ q#η, where the weights w1 and w2 Φ q#η and γ2 " pι, T w2 Φ pyqyT w2 Ψ pyqdηpyq. i"1 and Ψ " tψjuN µΨpAγ1 Z γ2 13 Example 3. Recovering the old definitions of analysis and synthesis In the special case M " N , we could choose P " tpiuN i"1 Ă Rd and w0 adapted to pµP , ηq. Then let fΨ : P Ñ Ψ be given by fΨppiq " ψi, and let fΦ : P Ñ Φ be similarly defined. Then if γ1 " pι, fΦ T w0 P q#η and γ2 " pι, fΨ T w0 P q#η, it follows that µΨpAγ1 Z γ2 µΦpxqq "ż xx , fΦ T w0 P pyqyfΨ T w0 P pyqdηpyq " N ÿi"1 xx , ϕi yψi. Hence, we have recovered the analysis and synthesis operation of finite frames. Example 4. Discrete dual to absolutely continuous probabilistic frame Finally, choose a frame contained in the support of η, say tψiuN map between η and µΨ, as constructed above. Choose tϕiuN f : Ψ Ñ Φ be given by fpψiq " ϕi. Then γ " pι, f T w in Γpη, µΨq such that ť xyJdγpx, yq "ş xT w in PFpRdq. i"1. Let T w Ψ be the transport i"1, and let Ψq#η P P2pRd Rdq is a joint transport plan Ψpxqdηpxq " I, so that η and µΨ are dual to one another i"1 to be any dual to tψiuN 3 Paths of Frames: Geodesics for the Wasserstein Space A number of important questions in finite frame theory involve determining distances between frames and constructing new frames. In this section we consider geodesics in P2pRdq and investi- gate conditions under which probability measures on these paths are probabilistic frames. As we shall prove, in the case of discrete probabilistic frames, this question is equivalent to one of ranks of convex combinations of matrices. Furthermore, for probabilistic frames with density, a sufficient condition for geodesic measures to be probabilistic frames is the continuity of the optimal deter- ministic coupling. This question has ramifications for constructions of paths of frames in general, for frame optimization problems, and for our understanding of the geometry of P FpRdq. 3.1 Wasserstein Geodesics In constructing paths of probabilistic frames, minimal paths between frames in P2pRdq are a natural place to start since PFpRdq is not closed. We follow the construction of geodesics in the Wasserstein space given in [13]. To this end, given t P r0, 1s define Πt : Rd Rd Ñ Rd as Πtpx, yq " px,p1 ´ tqx ` tyq. For µ0, µ1 P P2pRdq, take γ0 P Γpµ0, µ1q to be an optimal transport plan for µ0 and µ1 with respect to the 2-Wasserstein distance. Then let the interpolating joint probability measure be γt on Rd Rd, given by: ijRdRd Fpx, yqdγtpx, yq " ijRdRd FpΠtpx, yqqdγ0px, yq 14 for all F P CbpRd Rdq. In particular, for F P CbpRdq, ijRdRd Fpxqdγtpx, yq " ijRdRd Fpxqdγ0px, yq "żRd Fpxqdµ0pxq. Given t P r0, 1s let µt be the probability measure such that for all G P CbpRdq: żRd Gpyqdµtpyq " ijRdRd Gpyqdγtpx, yq " ijRdRd Gpp1 ´ tqx ` tyqdγ0px, yq, (8) we call µt a geodesic measure with respect to µ0 and µ1. Indeed, the mapping t Ñ µt is truly a geodesic of the 2-Wasserstein distance in the sense that W2pµ0, µtq ` W2pµt, µ1q " W2pµ0, µ1q. Recall that a probability measure µ on Rd is a probabilistic frame if it is an element of P2pRdq and if Sµ is positive definite. It is easy to show that µt, as constructed by the method above, always meets the first requirement. Lemma 11. For any measure µt, t P r0, 1s, on the geodesic between two probabilistic frames µ0 and µ1, M 2 2pµtq ă 8. Showing that Sµt is positive definite, or, equivalently, that the support of µt spans Rd depends on the characteristics of the support of the measures at the endpoints. For this reason, it is natural to divide the analysis into two parts: the discrete case and the absolutely continuous case. In both, a monotonicity property that characterizes optimal transport plans will play a key role. 3.2 Probabilistic Frames with Discrete Support For the canonical discrete probabilistic frames with uniform weights, we have: Lemma 12. [2, Theorem 6.0.1] Given µ0 " µΦ and µ1 " µΨ, discrete probabilistic frames with supports of equal cardinality N, uniformly weighted, the Monge-Kantorovich problem simplifies, and denoting by Γp 1 N q the set of matrices with row and column sums identically 1 N : W 2 2 pµ0, µ1q " min APΓp 1 N q N ÿi"1 N ÿj"1 ai,jkϕi ´ ψjk2 and, by the Birkhoff-von Neumann Theorem, the optimal transport matrix A is a permutation matrix corresponding to some σ P ΠN , i.e.: W 2 2 pµ0, µ1q " min σPΠN 1 N N ÿi"1 kϕi ´ ψσpiqk2 15 In this case, for some optimal σ P ΠN , ÿi"1 Sµt :" 1 N N rp1 ´ tqϕi ` tψσpiqsrp1 ´ tqϕi ` tψσpiqsJ. (9) N The optimality of σ implies that σ maximizes ři"1xϕi , ψσpiq y among all elements of ΠN , and this crucial fact motivates the definition of a monotonicity condition. Definition 13. A set S Ă Rd Rd is said to be cyclically monotone if, given any finite subset tpx1, y1q, ...,pxN , yNqu Ă S, for every σ P SN holds the inequality: ÿi"1xxi , yσpiq y. ÿi"1xxi , yi y ě N N With this definition in hand, the main result of this section can be stated: Theorem 14. Let tϕiuN and tpϕi, ψiquN probabilistic frames µΦ and µΨ is a probabilistic frame. If Ψ:Φ has no negative eigenvalues i"1 is cyclically monotone, then every measure on the geodesic between the canonical i"1 be frames for Rd. i"1 and tψiuN The proof of this theorem will follow from Lemma 11 and Proposition 16, proven below. To prove Proposition 16, the following lemma from matrix theory is necessary: Lemma 15. [21, Theorem 2] Let A and B be m n complex matrices, m ě n. Let rankpAq " rankpBq " n. If B:A has no nonnegative eigenvalues, then every matrix in hpA, Bq :" tp1 ´ tqA ` tB, t P r0, 1su has rank n. Similarly, if A and B are n n complex matrices with rank n, we can define in rpA, Bq :" tpI ´ TqA ` T Bu, where T is a real diagonal matrix with diagonal entries in r0, 1s. Then, if B´1A is such that all its principal minors are positive, then every matrix in rpA, Bq will have rank n. Combining the cyclical monotonicity condition with Lemma 15, we can state the following result which gives sufficient conditions for a geodesic between discrete probability measures in P2pRdq to be a path of frames. We note that little can be claimed about the spectra of the frame operators along the path (i.e., the frame bounds of the probabilistic frames along the geodesic) in general, other than their boundedness away from zero. Proposition 16. Let tϕiuN i"1 be frames for Rd with analysis operators Φ and Ψ. Denoting by P si: the Moore-Penrose pseudoinverse of Ψ, if Ψ:Φ has no negative eigenvalues, and if tpϕi, ψiquN i"1 is a cyclically monotone set, then every measure µt on the geodesic between µΦ and µΨ has support which spans Rd. i"1 and tψiuN 16 Proof. Each measure on the geodesic µt will be supported on a new set of vectors, namely tp1 ´ tqϕi ` tψσpiquN i"1, and will be a probabilistic frame provided this set of vectors spans Rd. Equivalently, µt will be a probabilistic frame if the probabilistic frame operator Sµt is positive definite. Let Pσ be the N N permutation matrix corresponding to σ P ΠN , where now σ is the optimal permutation for the Wasserstein distance. Let Ψσ " PσΨ. A quick calculation shows: 1 Sµt " N `p1 ´ tqΦJ ` tΨJσpp1 ´ tqΦ ` tΨσq . Ψ and Ψσ have rank d, and to show that Sµt is positive definite, it remains to prove that every matrix in the set hpΦ, Ψσq :" tp1´ tqΦ` tΨσutPr0,1s has rank d. By Lemma 15, a sufficient condition for this to be true is that Ψ:σΦ be positive semi-definite. Finally, we note that if tpϕi, ψiquN i"1 is a cyclically monotone set, then Pσ " I, the identity, is an optimal permutation, and then Ψ:σΦ " Ψ:Φ is positive definite by assumption. Proof. Proof of Theorem 14 With Lemma 11 showing that measures on the geodesic have finite second moment and Propo- sition 16 showing that the support of these measures spans Rd, Theorem 14 is now proved. Certain dual frame pairs immediately satisfy the conditions laid out in Theorem 14. Proposition 17. If tϕiuN monotone. i"1 is the canonical dual frame to tψiuN i"1, then tpϕi, ψiquN i"1 is cyclically Proof. Let S " ΨJΨ. Then suppose that ΦJ " S´1ΨJ. For any permutation σ P ΠN , let Pσ denote the matrix such that for @x " rx1 . . . xNsJ P RN , Pσx " rxσp1q . . . xσpNqsJ. Then, N N N ÿi"1xS´1ψi , ψi ´ ψσpiq y ÿi"1xϕi , ψi ´ ψσpiq y " ÿi"1 pψi ´ ψσpiqqJS´1ψi " " TrppΨ ´ PσΨqS´1ΨJq " TrppIN ´ PσqΨS´1ΨJq ě 0 Here we use the fact that ΨS´1ΨJ " I d N , the N N diagonal matrix with d leading ones on the diagonal and zeros else, because S´1ΨJ is the Moore-Penrose pseudoinverse of Ψ. Therefore, the identity is an optimal permutation, i.e., the set tpϕi, ψiquN i"1 is cyclically monotone. 17 Proposition 18. Let tβiuN tβiuN cyclically monotone. i"1 to define tϕiuN i"1 Ă Rd be such that tpβi, ψiquN i"1, one of the dual frames to tψiuN i"d`1 is cyclically monotone. Then use i"1 as given in (1). Then tpϕi, ψiquN i"1 is Proof. Take tϕiuN rows are the tβiuN i"1 to be a dual of the form given in Equation (1). Let W be the matrix whose i"1. Then, noting that ΦJ " pS´1ΨJ ` W JpIN ´ ΨS´1ΨJqq, ÿi"1xψi ´ ψσpiq , ϕi y " TrppIN ´ PσqΨΦJq N " TrppIN ´ PσqΨpS´1ΨJ ` W JpIN ´ ΨS´1ΨJqqq " TrppIN ´ PσqI d " TrppIN ´ PσqI d N ` pIN ´ PσqΨW JpIN ´ I d Nq ` xψi ´ ψσpiq , βi y Nqq N ÿi"d`1 ě 0 i"1 is cyclically monotone. i"1 is the canonical dual frame to tψiuN Therefore, under these conditions, tpϕi, ψiquN Proposition 19. If tϕiuN i"1 is a dual frame to tψiuN i"d`1 is cyclically monotone, then Ψ:σΦ is positive definite, where σ is the optimal permutation for the Wasserstein distance. Consequently, any path on the geodesic joining tψiuN i"1 is a probabilistic frame. Proof. By definition, i"1, or if tϕiuN i"1 ordered so that tphi, ψiquN i"1 of the form given in (1), with the thiuN i"1 and tϕiuN Ψ:σ " pPσΨq: " pΨJP Jσ PσΨq´1ΨJP Jσ " pΨJΨq´1ΨJP Jσ This is a permutation of the matrix whose columns are canonically dual to the rows of Ψσ. If tϕiuN i"1 i"1, then ΨJΦ " Id. Therefore, if σ is the identity, then Ψ:σΦ " pΨJΨq´1ΨJΦ " is any dual of tψiuN pΨJΨq´1, which is positive definite. It remains to show that the optimal permutation is the iden- tity. But this is clear: Proposition 17 shows that if tϕiuN i"1, then tpϕi, ψiquN i"1 is any dual to tψiuN i"1 which meets the above condition, then tpϕi, ψiquN i"1 is cyclically monotone, and Proposition 18 shows that if tϕiuN i"1 is cyclically monotone. i"1 is the canonical dual to tψiuN There are other frame and dual-frame pairs which can easily be shown to meet the above i"1 Ă Rd with respective analysis conditions. Consider the finite sequences tϕiuN operators Φ and Ψ. Then the finite sequences are disjoint if ΦpRdqŞ ΨpRdq " t0u. i"1 are disjoint frames for Rd, associated canonically with Proposition 20. If tϕiuN the probabilistic frames µΦ and µΨ, then every measure on the geodesic between µΦ and µΨ is a probabilistic frame. i"1 Ă Rd and tψiuN i"1 and tψiuN 18 Proof. Given v P Rd, consider: N N ÿi"1 xv ,p1 ´ tqϕi ` tψi y2 " N ÿi"1 xv ,p1 ´ tqΦJei ` tΨJei y2 ÿi"1xp1 ´ tqΦv ` tΨv , ei y2 " " kp1 ´ tqΦv ` tΨvk2 ě Crp1 ´ tq2kΦvk2 ` t2kΨvk2s RN for some C ą 0, since the frames are disjoint. Since the two sequences in question are finite frames, choosing the minimum of the two lower frame bounds, say A0, the last quantity can be bounded below by p1 ´ 2t ` 2t2qC A0kvk2, yielding the result. Finally, in the following result control of the distance between the elements of a one frame and those of the canonical dual of the other by a coherence-like quantity guarantees the frame properties for the frames on the geodesic. Proposition 21. Let tψiuN operator. For each i, let zi " ψi´ S´1 then the optimal σ for the mass transport problem is the identity. i"1 be a dual frame to a frame tϕiuN Φ ϕi, and let a :" mini‰jxϕi , S´1 i"1 Ă Sd´1. Let SΦ denote the frame Φ pϕi ´ ϕjqy. If maxjkzjk ď a N , Proof. First, we note that a ě 0. If a " 0, then our hypothesis guarantees that kzik " kψi´S´1 Φ ϕik " 0 for all i, so that Ψ is the canonical dual to Φ, and in this case our result holds by Proposition 17. Therefore, it only remains to consider the case when a ą 0. N For all u, v P Rd, not fixed by σ. Then if σ is the identity, nσ " 0 and ři"1xu , zi yxv , ϕi y " 0. Then given σ P SN , let nσ be the number of elements N ÿi"1 xϕi , S´1 Φ xσpiq ` zσpiq y " T rpΨJPσΦq " d If σ is not the identity, then N ÿi"1 xϕi , S´1 Φ ϕσpiq ` zσpiq y " ÿi‰σpiq Φ pϕσpiq ´ ϕiq ` zσpiq ´ zi y ` d xϕi , zσpiq ´ zi y xϕi , S´1 ď d ´ anσ ` ÿi‰σpiq 2 ď d ´ p1 ´ N qanσ Since, given the hypothesis, for all i, j, xϕi , zj y ď kϕikkzjk " kzjk ď a d ´ p1 ´ 2 map for the Wasserstein metric. N . Thus TrpΨJPσΦq ď N qanσ ď d " T rpΨJΦq for all σ, and it follows that the identity is the optimal transport 19 3.3 Absolutely Continuous Probabilistic Frames The question of the nature of the optimal transport plan for the 2-Wasserstein distance is simpler for absolutely continuous measures. From [2, Theorem 6.2.10 and Proposition 6.2.13], which gather together a long list of characteristics, two key facts about this plan can be extracted, which are collected in the following lemma. Lemma 22. [2, Chapter 6.2.3] If µ0 and µ1 are absolutely continuous probability measures in P2pRdq, then there exists a unique optimal transport plan for the 2-Wasserstein distance which is induced by a transport map r. This transport map is defined (and injective) µ0-a.e. Indeed, there exists a µ0-negligible set N Ă Rd such that xrpx1q ´ rpx2q , x1 ´ x2 y ą 0 for all x1, x2 P RdzN . Then we have the following result for absolutely continuous probabilistic frames: Proposition 23. If µ0 and µ1 are absolutely continuous (with respect to Lebesgue measure) prob- abilistic frames for which there exists a linear, positive semi-definite deterministic coupling which minimizes the Wasserstein distance, then all measures on the geodesic between these frames have support which spans Rd and will therefore be probabilistic frames. Proof. Given the assumptions, let rpxq denote the linear transformation which induces the coupling µ1 " r#µ0. Defining htpxq " p1 ´ tqx ` trpxq µ0-a.e., the geodesic measure is given by µt :" ht#µ0. (10) Then Sµt "şRd htpxqhtpxqJdµ0pxq. If rpxq " Ax for some A P Add, then: Sµt "żRdpp1 ´ tqIx ` tAxqpp1 ´ tqIx ` tAxqJdµ0pxq " pp1 ´ tqI ` tAqSµ0pp1 ´ tqI ` tAqJ Since A must be nonsingular -- recall that Sµ1 " ASµ0 AJ, which is certainly of rank d -- by Lemma 15, p1 ´ tqI ` tA will also nonsingular for all t P r0, 1s provided that A has no negative eigenvalues, as we assumed. Example 5. An example in which the assumptions of the above proposition hold is the case of non- degenerate Gaussian measures on Rd. Let µ0 and µ1 be zero-mean Gaussians. Let rpxq " S µ0 x. According to a result in [7], if X and Y are two zero-mean random vectors with covariances ΣX and ΣY , respectively, then a lower bound for EpkX ´ Y k2q is TrrΣX ` ΣY ´ 2pΣX ΣY q 2s, and the bound is attained, for nonsingular ΣX, when Y " Σ´ 1 Y X, so that the coupling r is an optimal X Σ µ1 S´ 1 2 1 2 2 1 2 1 positive definite linear deterministic coupling of µ0 and µ1. 20 Now, given absolutely continuous probabilistic frames µ, ν for Rd, take rpxq to be the optimal transport map pushing µ to ν guaranteed by Lemma 22. Define htpxq " p1 ´ tqx ` trpxq for t P r0, 1s; then Sµt "ş htpxq b htpxqdµpxq, with µt " phtq#µ. Then we can state the following: Proposition 24. Given two such probabilistic frames, there exists a set N with µpNq " 0 such that ht is injective for all t P r0, 1s on supppµqzN . Proof. Given x, y P supppµqzN , with N as defined in Lemma 22, suppose htpxq " htpyq for some t P r0, 1s. Then, since: 0 " xhtpxq ´ htpyq , x ´ y y " xp1 ´ tqpx ´ yq ` tprpxq ´ rpyqq , x ´ y y " p1 ´ tqkx ´ yk2 ` txrpxq ´ rpyq , x ´ y y xrpxq ´ rpyq , x ´ y y " t ´ 1 t kx ´ yk2. it follows that This implies that xrpxq´ rpyq , x´ y y ď 0. However, from the proposition above, we also know that xrpxq ´ rpyq , x ´ y y ě 0. Therefore kx ´ yk " 0, and ht is injective on supppµqzN . This injectivity claim is crucial for the main result of this section: 2 pRdq, and let r be the unique optimal transport map for the 2 ´ Theorem 25. Let µ, ν P P r W asserstein distance. Let N be the set of measure zero define in Proposition 24. If r is continuous, and if supppµqzN contains an open set, then every geodesic measure µt is a probabilistic frame. Proof. Since r is continuous and, by Proposition 24, injective outside a set N of measure zero, so is ht for each t. Let x0 P supppµqzN . First, we show that for any ǫ ą 0, h´1 t pBǫphtpx0qqq contains an open set containing x0. Since ht is continuous at any such x0, given ǫ ą 0, there exists δ ą 0 such that @x P Bδpx0q, t pBǫphtpx0qqq. khtpxq´htpx0qk ă ǫ. Hence for any x P Bδpx0q, x P h´1 t pBǫphtpx0qqq -- i.e., Bδpx0q Ă h´1 Then @x0 P supppµqzN , consider B 1 µtpB 1 kphtpx0qq: k phtpx0qqq "ż 1„B 1 k phtpx0qqphtpyqqdµpyq "ż 1„h´1 k phtpx0qqqpyqdµpyq pB 1 " µph´1 kphtpx0qqqq t pB 1 ą 0 t 21 t pB 1 k where the last inequality holds since x0 P supppµq and, as shown above, h´1 kphtpx0qqqq contains an open set containing x0. Thus, we have shown that for any k P N, the open ball of radius 1 around htpx0q has positive µt-measure, and therefore htpx0q lies in supppµtq. Thus htpsupppµqzNq Ă supppµtq. Therefore, since ht is injective by Proposition 24 above and continuous on supppµqzN and by assumption, there exists open set U Ă supppµqzN , by invariance of domain, htpUq Ă supppµtq is open, we conclude that ht#µ has support which spans Rd. The question of when r is continuous is the subject of ongoing research. One example is when µ and ν are supported on a bounded convex subset of Rd [6]. Acknowledgment Clare Wickman would like to thank the Norbert Wiener Center for Harmonic Analysis and Appli- cations for its support during this research. Kasso Okoudjou was partially supported by a grant from the Simons Foundation (#319197 to Kasso Okoudjou), and ARO grant W911NF1610008. References [1] S. Ali, J-P. Antoine, and J.-P. Gazeau, Coherent states, wavelets, and their generalizations, ch. Positive Operator-Valued Measures and Frames, Springer, New York, 2000. [2] L. Ambrosio, N. Gigli, and G. Savare, Gradient flows in metric spaces and in the space of probability measures, Birkhauser, Boston, 2005. [3] J. Blum, M. Lammers, A. M. Powell, and O. Ylmaz, Sobolev duals in frame theory and sigma- delta quantization, J. Fourier Anal. Appl. 16 (2010), no. 3, 365 -- 381. [4] P. G. Casazza and G. Kutyniok (eds.), Finite Frames: Theory and Applications, Springer- Birkhauser, New York, 2013. [5] O. Christensen, An introduction to frames and Riesz bases, Birkhauser, Boston, 2003. [6] G. De Philippis and A. Figalli, The Monge-Amp`ere equations and its links to optimal trans- portation, Bulletins of the American Mathematical Association 51 (2014), no. 4, 527 -- 580. [7] D. C. Dowson and B. V. Landau, The Fr´echet distance between multivariate normal distribu- tions, Journal of Multivariate Analysis 12 (1982), 450 -- 455. [8] M. Ehler, Random tight frames, Journal of Fourier Analysis and Applications 18 (2012), no. 1, 1 -- 20. [9] M. Ehler and K. A. Okoudjou, Minimization of the probabilistic p-frame potential, Journal of Statistical Planning and Inference 142 (2012), no. 3, 645 -- 659. 22 [10] , Finite frames, ed. G. Kutyniok and P. G. Casazza, ch. Probabilistic frames: An overview, Springer, New York, 2013. [11] M. Fornasier and H. Rauhut, Continuous frames, function spaces, and the discretization prob- lem, Journal of Fourier Analysis and Applications 11 (2005), no. 3, 245 -- 287. [12] D. Gale, H.W. Kuhn, and A.W. Tucker, Linear programming and the theory of games, Activity Analysis of Production and Allocations, vol. 13, J. Wiley and Sons, New York, 1951, pp. 317 -- 335. [13] W. Gangbo, An introduction to the mass transportation theory and its applications, Lecture Notes, 2004. [14] C. S. Gunturk, M. Lammers, A. M. Powell, R. Saab, and O. Ylmaz, Sobolev duals for random frames and Σ ´ ∆ quantization of compressed sensing measurements, Found. Comput. Math. 13 (2013), no. 1, 1 -- 36. [15] D. Han, J. Leng, and T. Huang, Optimal dual frames for communication coding with proba- bilistic erasures, IEEE Transactions on Signal Processing 59 (2011), 5380 -- 5389. [16] D. Han and W. Sun, Reconstruction of signals from frame coefficients with erasures at unknown locations, IEEE Trans. Inform. Theory 60 (2014), no. 7, 4013 -- 4025. [17] P.E. Jorgensen and M.S. Song, Infinite-dimensional measure spaces and frame analysis, arXiv preprint arXiv:1606.04866 (2016). [18] Q. Merigot, A multiscale approach to optimal transport, Computer Graphics Forum 30 (2011), no. 5, 1584 -- 1592. [19] K. A. Okoudjou (ed.), Finite Frame Theory: A Complete Introduction to Overcompleteness, Proceedings of Symposia in Applied Mathematics, vol. 73, AMS, Providence, RI, 2016. [20] A. Powell and J.T. Whitehouse, Error bounds for consistent reconstruction: random polytopes and coverage processes, Foundations of Computational Mathematics 16 (2016), 395 -- 423. [21] T. Szulc, Ranks of convex combinations of matrices, Reliable Computing 2 (1996), no. 2, 181 -- 185. [22] C. Villani, Optimal transport, old and new, Grundlehren der mathematischen Wissenschaften, no. 338, Springer-Verlag, New York, 2009. 23
1112.3813	2	1112	2012-03-07T14:25:31	Sigma-porosity is separably determined	[ "math.FA" ]	We prove a separable reduction theorem for sigma-porosity of Suslin sets. In particular, if A is a Suslin subset in a Banach space X, then each separable subspace of X can be enlarged to a separable subspace V such that A is sigma-porous in X if and only if the intersection of A and V is sigma-porous in V. Such a result is proved for several types of sigma-porosity. The proof is done using the method of elementary submodels, hence the results can be combined with other separable reduction theorems. As an application we extend a theorem of L.Zajicek on differentiability of Lipschitz functions on separable Asplund spaces to the nonseparable setting.	math.FA	math	σ-POROSITY IS SEPARABLY DETERMINED MAREK C ´UTH, MARTIN RMOUTIL Abstract. We prove a separable reduction theorem for σ-porosity of Suslin sets. In par- ticular, if A is a Suslin subset in a Banach space X, then each separable subspace of X can be enlarged to a separable subspace V such that A is σ-porous in X if and only if A ∩ V is σ-porous in V . Such a result is proved for several types of σ-porosity. The proof is done using the method of elementary submodels, hence the results can be combined with other separable reduction theorems. As an application we extend a theorem of L.Zaj´ıcek on differentiability of Lipschitz functions on separable Asplund spaces to the nonseparable setting. 1. Introduction The aim of this article is to obtain separable reduction theorems for some classes of σ- porous sets by employing the method of elementary submodels. This is a set-theoretical method which can be used in various branches of mathematics. A. Dow in [2] illustrated the use of this method in topology, W. Kubi´s in [4] used it in functional analysis, namely to construct projections on Banach spaces. In this article we shall use the method of elementary submodels to prove Theorem 5.1 and Theorem 5.4 which have as a consequence for example the following: 1.1. Theorem. Let hX, k·ki be a Banach space and let A ⊂ X be a Suslin set. Then for every separable subspace V0 ⊂ X there exists a closed separable space V ⊂ X such that V0 ⊂ V and (i) A is σ-upper porous if and only if A ∩ V is σ-upper porous in the space V , (ii) A is σ-lower porous if and only if A ∩ V is σ-lower porous in the space V . As a consequence of Theorem 5.1 and [1, Theorem 5.7] we get the following: 1.2. Theorem. Let X, Y be Banach spaces, G ⊂ X an open subset and f : G → Y be a function. Then for every separable subspace V0 ⊂ X there exists a closed separable space V ⊂ X such that V0 ⊂ V and that the following two conditions are equivalent: (i) the set of the points where f is not Fr´echet differentiable is σ-upper porous, (ii) the set of the points where f ↾ V is not Fr´echet differentiable is σ-upper porous in V . The first result is in a certain sense an improvement of the result of J. Lindenstrauss, D.Preiss and J.Tiser [6, Corrolary 3.6.7], from where only one implication follows. Moreover, we are able to easily extend results concerning points of non-differentiability from separable Banach spaces to the non-separable case. An example of such a result is Theorem 5.5 which has been proved in the article [9] -- the generalization is in Theorem 5.6. Let us recall the most relevant notions, definitions and notations: 2010 Mathematics Subject Classification. 28A05, 54E35, 58C20. Key words and phrases. Elementary submodel, separable reduction, porous set, σ-porous set. M.C´uth was supported by the Grant No. 282511/B-MAT/MFF of the Grant Agency of the Charles Uni- versity in Prague. M. Rmoutil was supported by the Grant SVV-2011-263316. 1 2 MAREK C ´UTH, MARTIN RMOUTIL Notation. We denote by ω the set of all natural numbers (including 0), by N the set ω \ {0}, by R+ the interval (0, ∞) and Q+ stands for R+ ∩ Q. Whenever we say that a set is countable, we mean that the set is either finite or infinite and countable. If f is a mapping then we denote by Rng f the range of f and by Dom f the domain of f . By writing f : X → Y we mean that f is a mapping with Dom f = X and Rng f ⊂ Y . By the symbol f ↾Z we denote the restriction of the mapping f to the set Z. If hX, ρi is a metric space, we denote by U (x, r) the open ball (i.e. the set {y ∈ X : ρ(x, y) < r}) and by d(x, A) the distance function from a set A ⊂ X (i.e. d(x, A) = inf{ρ(x, a); a ∈ A}). We shall consider normed linear spaces over the field of real numbers (but many results hold for complex spaces as well). If X is a normed linear space, X ∗ stands for the (continuous) dual space of X. 2. Elementary Submodels The method of elementary submodels enables us to find specific separable subspaces (of Banach spaces) which can be used for proofs of separable reduction theorems. In this section we briefly describe this method and recall some basic notions. More information can be found in [1] where this method is described in greater detail. First, let us recall some definitions: Let N be a fixed set and φ a formula in the language of ZF C. Then the relativization of φ to N is the formula φN which is obtained from φ by replacing each quantifier of the form "∀x" by "∀x ∈ N " and each quantifier of the form "∃x" by "∃x ∈ N ". For example, if and N = {a, b}, then the relativization of φ to N is φ = ∀x ∀y ∃z ((x ∈ z) ∧ (y ∈ z)) φN = ∀x ∈ N ∀y ∈ N ∃z ∈ N ((x ∈ z) ∧ (y ∈ z)). It is clear that φ is satisfied, but φN is not. If φ(x1, . . . , xn) is a formula with all free variables shown (i.e. a formula whose free variables are exactly x1, . . . , xn) then φ is absolute for N if and only if ∀a1, . . . , an ∈ N (φN (a1, . . . , an) ↔ φ(a1, . . . , an)). The method is based mainly on the following set-theoretical theorem (a proof can be found in [5, Chapter IV, Theorem 7.8]). 2.1. Theorem. Let φ1, . . . , φn be any formulas and X any set. Then there exists a set M ⊃ X such, that (φ1, . . . , φn are absolute for M ) ∧ (M ≤ max(ω, X)). Since the previous theorem will often be used throughout the paper, the following notation is useful. Definition. Let φ1, . . . , φn be any formulas and let X be any countable set. Let M ⊃ X be a countable set satisfying that φ1, . . . , φn are absolute for M . Then we say that M is an elementary submodel for φ1, . . . , φn containing X. This is denoted by M ≺ (φ1, ..., φn; X). Let φ(x1, . . . , xn) be a formula with all free variables shown and let M be some elementary submodel for φ. To use the absoluteness of φ for M efficiently, we need to know that many sets are elements of M . The reason is that for a1, . . . , an ∈ M we have φ(a1, . . . , an) if and only if φM (a1, . . . , an). Using the following lemma we can force the elementary submodel M σ-POROSITY IS SEPARABLY DETERMINED 3 to contain all the required objects created (uniquely) from elements of M (for a proof see [1, Lemma 2.5]). 2.2. Lemma. Let φ(y, x1, . . . , xn) be a formula with all free variables shown and let X be a countable set. Let M be a fixed set, M ≺ (φ, ∃y φ(y, x1, . . . , xn); X) and let a1, . . . , an ∈ M be such that there exists only one set u satisfying φ(u, a1, . . . , an). Then u ∈ M . It would be very laborious and pointless to use only the basic language of the set theory. For example, we often write x < y and we know that this is in fact a shortcut for the formula ϕ(x, y, <) with all free variables shown. Therefore, in the following text we use this extended language of the set theory as we are used to. We shall also use the following convention. 2.3. Convention. Whenever we say for any suitable elementary submodel M (the following holds...), we mean that there exists a list of formulas φ1, . . . , φn and a countable set Y such that for every M ≺ (φ1, . . . , φn; Y ) (the following holds...). By using this new terminology we lose the information about the formulas φ1, . . . , φn and the set Y . This is, however, not important in applications. 2.4. Remark. Let us have sentences T1(a), . . . , Tn(a). Assume that whenever an i ∈ {1, . . . , n} is given, then for any suitable elementary submodel Mi the sentence Ti(Mi) is satisfied. Then it is easy to verify that for any suitable model M the sentence T1(M ) ∧ . . . ∧ Tn(M ) is satisfied (it suffices to combine all the lists of formulas and all the sets from the definition above). In other words, we are able to combine any finite number of results we have proved using the technique of elementary submodels. This includes all the theorems starting with "For any suitable elementary submodel M the following holds:" Let us recall several more results about suitable elementary submodels (proofs can be found in [1, Chapters 2 and 3]): 2.5. Proposition. For any suitable elementary submodel M the following holds: (i) If A, B ∈ M , then A ∩ B ∈ M , B \ A ∈ M and A ∪ B ∈ M . (ii) Let f be a function such that f ∈ M . Then Dom f ∈ M , Rng f ∈ M and for every x ∈ Dom f ∩ M , f (x) ∈ M . (iii) Let S be a finite set. Then S ∈ M if and only if S ⊂ M . (iv) Let S ∈ M be a countable set. Then S ⊂ M . (v) For every natural number n > 0 and for arbitrary (n + 1) sets a0, . . . , an it is true,that a0, . . . , an ∈ M ↔ ha0, . . . , ani ∈ M. 2.6. Notation. • If A is a set, then by saying that an elementary model M contains A we mean that A ∈ M . 4 MAREK C ´UTH, MARTIN RMOUTIL • If hX, ρi is a metric space (resp. hX, +, ·, k·ki is a normed linear space) and M an elementary submodel, then by saying M contains X (or by writing X ∈ M ) we mean that hX, ρi ∈ M (resp. hX, +, ·, k·ki ∈ M ). • If X is a topological space and M an elementary submodel, then we denote by XM the set X ∩ M . 2.7. Proposition. For any suitable elementary submodel M the following holds: (i) If X is a metric space then whenever M contains X, it is true that ∀r ∈ R+ ∩ M ∀x ∈ X ∩ M U (x, r) ∈ M. (ii) If X is a normed linear space then whenever M contains X, it is true that XM is closed separable suspace of X. 2.8. Convention. The proofs in the following text often begin in the same way. To avoid unnecessary repetitions, by saying "Let us fix a (∗)-elementary submodel M [containing A1, . . . , An]" we will understand the following: Let us have formulas ϕ1, . . . , ϕm and a countable set Y such that the elementary submodel M ≺ (ϕ1, . . . , ϕm; Y ) is suitable for all the propositions from [1]. Add to them formulas marked with (∗) in all the preceding proofs from this paper and formulas marked with (∗) in the proof below (and all their subformulas). Denote such a list of formulas by φ1, . . . , φk. Let us fix a countable set X containing the sets Y , ω, Z, Q, Q+, R, R+ and all the common operations and relations on real numbers (+, −, ·, :, <). Fix an elementary submodel M for formulas φ1, . . . , φk containing X [such that A1, . . . , An ∈ M ]. Thus, any (∗)-elementary submodel M is suitable for the results from [1] and all the preceding theorems and propositions from this paper, making it possible to use all of these results for M . In order to demonstrate how this technique works, we prove the following two easy lemmas which we use later (the proof of the second lemma is also contained in the proof of Proposition 4.1 in [1]). 2.9. Lemma. For any suitable elementary submodel M the following holds: Whenever A ∈ M is a nonempty set, then A ∩ M is nonempty. Proof. Let us fix a (∗)-elementary submodel M and fix some nonempty set A ∈ M . Then (∗) ∃x (x ∈ A). This formula has only one free variable A and the set A is contained in M . Thus, due to the absoluteness of the formula above, there exists an x ∈ M such that x ∈ A. (cid:3) 2.10. Lemma. For any suitable elementary submodel M the following holds: Let hX, ρi be a metric space, B ⊂ X. Then whenever M contains X, B and a set D ⊂ B, it is true that D is dense in B → D ∩ M is dense in B ∩ XM . Proof. Let us fix a (∗)-elementary submodel M containing X such that B, D ∈ M . If the set B is empty then the proposition is obvious. Otherwise fix b ∈ B ∩ XM and r > 0. Choose some bo ∈ U (b, r 2 ). Then U (b0, q) ⊂ U (b, r) and (∗) 2 ) ∩ M and a rational number q ∈ (ρ(b, b0), r ∃d ∈ D (d ∈ U (b0, q)). σ-POROSITY IS SEPARABLY DETERMINED 5 In the preceding formula we use the shortcut d ∈ U (b0, q) which stands for d ∈ X ∧ ρ(d, b0) < q. Free variables in this formula are X, ρ, <, D, b0, q. Those are contained in M and thus we can use the absoluteness to find a d ∈ D∩M such that (d ∈ U (b0, q))M . Using the absoluteness again we obtain that d is an element of U (b0, q). Consequently, and so the set D ∩ M is dense in B ∩ XM . (cid:3) U (b, r) ∩ D ∩ M ⊃ U (b0, q) ∩ D ∩ M 6= ∅ 3. σ-porous sets In this section we compile several known results concerning different notions of σ-porous sets. The usefulness of these facts for our needs will be apparent later; for more information about properties and applications of different types of porosity we refer the reader to survey articles [10] and [12]. On some occasions we shall also refer to the paper [8]. Let us begin by stating several basic definitions. Definition. Let hX, ρi be a metric space, A ⊂ X, x ∈ X and R > 0. Then we denote by γ(x, R, A) the supremum of all r ≥ 0 for which there exists z ∈ X such that U (z, r) ⊂ U (x, R) \ A. The set A is called upper porous at x in the space X if lim sup R→0+ γ(x, R, A) R > 0. In most cases it is clear which space X we have in mind. Therefore we often omit the words "in the space X". (We shall apply this convention to other notions as well.) Let g be a strictly increasing real-valued function defined on [0, h) (where h > 0) with g(0) = 0. Such a function is called porosity function. We say that A is hgi-porous at x (in the space X) if there exists a sequence of open balls {U (cn, rn)} such that cn → x, U (ck, rk) ∩ A = ∅ and x ∈ U (ck, g(rk)) for each k. We say the set A is hgi-porous (upper-porous) if it is hgi-porous (upper-porous) at each of its points and σ-hgi-porous (σ-upper porous) if it is a countable union of hgi-porous (upper- porous) sets. Definition. Let hX, ρi be a topologically complete metric space and let g be a porosity function. We say that F is a Foran system for hgi-porosity in X if the following conditions hold: (i) F is a nonempty family of nonempty Gδ subsets of X. (ii) For each S ∈ F and each open set G ⊂ X with S ∩ G 6= ∅ there exists S ∗ ∈ F such that S ∗ ⊂ S ∩ G and S is hgi-porous at no point of S ∗. 3.1. Proposition (Foran Lemma). Let hX, ρi be a topologically complete metric space and let F be a Foran system for hgi-porosity in X. Then no member of F is σ-hgi-porous. This is a special case of the general Foran Lemma which works for any porosity-like relation (see [11, Proposition 1]). Our definition of Foran system is, therefore, accordingly simplified as well. We also need the following. 3.2. Notation. By 3-porosity we mean hgi-porosity where g(x) = 3x for x ∈ R. 3.3. Lemma ([11, Lemma E]). Let hX, ρi be a metric space and let A ⊂ X. Then A is σ-upper porous if and only if it is σ-3-porous. 6 MAREK C ´UTH, MARTIN RMOUTIL Another result from [11] which we shall use is the following partial converse of the Foran Lemma. For ordinary σ-upper porosity we can extend its validity from Gδ sets to Suslin sets using the inscribing Theorem 3.5 of J. Pelant and M. Zelen´y from the work [15]. It could be interesting to note that in case our metric space X is locally compact, we can use a different inscribing theorem due to L. Zaj´ıcek and M. Zelen´y [14, Theorem 5.2] and obtain an extension of 3.4 to analytic sets for general σ-hgi-porosity. 3.4. Lemma ([11, Corollary 1]). Let hX, ρi be a topologically complete metric space, let ∅ 6= A ⊂ X be Gδ and let g be a porosity function. Then A is not σ-hgi-porous if and only if it contains a member of a Foran system for hgi-porosity. 3.5. Theorem ([15, Theorem 3.1]). Let hX, ρi be a topologically complete metric space and let S ⊆ X be a non-σ-upper porous Suslin set. Then there exists a closed non-σ-upper porous set F ⊆ S. Definition. Let hX, ρi be a metric space, A ⊂ X and x ∈ X. We say that A is lower porous at x if γ(x, R, A) > 0. lim inf R→0+ R The set A is lower porous if it is lower porous at each of its points and σ-lower porous if it is a countable union of lower porous sets. Even though the Foran Lemma can be used for any notion of porosity, we have to use a different approach in the case of lower porosity. The reason is that unlike in the case of upper porosity, we were unable to separably reduce the property of not being lower porous at a point. Therefore, we use the following proposition. 3.6. Proposition ([7, Proposition 2.11]). Let hX, ρi be a topologically complete metric space and let A ⊆ X be a Suslin set. Then the following propositions are equivalent: (i) A is not σ-lower porous. (ii) There exists a closed set F ⊆ A and a set D ⊆ F dense in F such that F is lower porous at no point of D. 4. Auxiliary results In this section we prove some preliminary statements which will be of use later. In general, for a space X and a set A ⊂ X, we are trying to find a separable subspace XM ⊂ X with certain special properties. The first desired property is: Whenever A is a member of a Foran system in X then A ∩ XM is a member of a Foran system in XM . Together with Lemma 3.4 this will be essential to the proof of Theorem 5.1 about σ-upper porosity. Also, in order to prove a result similar to 5.1 for σ-lower porosity, two auxiliary propositions (based on the ideas from [1]) are collected. 4.1. Proposition. For any suitable elementary submodel M the following holds: Let hX, ρi be a metric space and g a porosity function. Then whenever M contains X and a set A ⊂ X, it is true that for every x ∈ XM A is not hgi-porous at x → A ∩ XM is not hgi-porous at x in the space XM . If M contains also g, then A is not hgi-porous → A ∩ XM is not hgi-porous in the space XM . σ-POROSITY IS SEPARABLY DETERMINED 7 Proof. Let us fix a (∗)-elementary submodel M containing X and A and fix some x ∈ XM such that A is not hgi-porous at x. Take sequences {cn}n∈N ⊂ XM and {rn}n∈N ⊂ (0, ∞) such that cn → x and x ∈ U (cn, g(rn)) for all n ∈ N. It is sufficient to show that there exists an n ∈ N satisfying U (cn, rn) ∩ A ∩ XM 6= ∅. Since A is not hgi-porous, we can fix some n ∈ N such that U (cn, rn) ∩ A 6= ∅. Take some a ∈ A ∩ U (cn, rn) and choose an ε > 0 such that ρ(a, cn) + 2ε < rn. Then take a point c ∈ X ∩ M ∩ U (cn, ε) and qn ∈ Q ∩ (ρ(a, cn) + ε, rn − ε). Hence, (∗) ∃a ∈ A (ρ(a, c) < qn). Thus, by the absoluteness, there exists an a ∈ A ∩ M such that ρ(a, cn) ≤ ρ(a, c) + ε < qn + ε < rn. Consequently, a ∈ A ∩ U (cn, rn) ∩ M and thus the set A ∩ XM is not hgi-porous at x in the space XM . If A is not hgi-porous then (∗) ∃x ∈ A (A is not hgi-porous at x). Using the absoluteness and the already proved part we obtain an x ∈ A∩M such that A∩XM is not hgi-porous at x in the space XM . (cid:3) 4.2. Proposition. For any suitable elementary submodel M the following holds: Let hX, ρi be a topologically complete metric space and g a porosity function. Then whenever M contains X, g and a set A ⊂ X, it is true that if A is a member of a Foran system for hgi-porosity in X, then A ∩ XM is a member of a Foran system for hgi-porosity in XM . Proof. Let us fix a (∗)-elementary submodel M containing X such that A ∈ M and let the following formula be true (∗) ∃F (F is a Foran system for hgi-porosity in X such that A ∈ F ). Notice that the preceding is a formula with free variables from M . Thus, by the absoluteness, there exists an F ∈ M which is a Foran system for hgi-porosity in X with A ∈ F. Set F ′ := {S ∩ XM : S ∈ F ∩ M, S ∩ XM 6= ∅}. First we notice that, by Lemma 2.9, the set A ∩ M is nonempty; it follows that A ∩ XM ∈ F ′. Thus it suffices to establish that F ′ is a Foran system for hgi-porosity in XM . Clearly, F ′ is a nonempty family of nonempty Gδ subsets of XM so there only remains to be verified the second condition from the definition of Foran system. To that end, take some S ∈ F ∩ M such that S ∩ XM 6= ∅ (denote by SM the set S ∩ XM ). Then take an arbitrary open set G ⊂ X with SM ∩ G 6= ∅ and fix some x ∈ SM ∩ G and r ∈ Q+ such that U (x, r) ⊂ G. Choose x0 ∈ U (x, r 2 ) ⊂ U (x, r); thus, S ∩ U (x0, r 2 ) ∈ M . 2 ) ∩ M . Then x ∈ U (x0, r 2 ) 6= ∅. Using Propositions 2.5 and 2.7 we obtain that S ∩ U (x0, r Now, as F is a Foran system (in X), the following formula is true: (∗) ∃S ∗ ∈ F : (S ∗ ⊂ S ∩ U (x0, r 2 ), S is hgi-porous at no point of S ∗) By the absoluteness, there exists an S ∗ ∈ M satisfying the formula above. Using Lemma 2.9 we can see that S ∗ ∩ M 6= ∅. Thus, S ∗ is a member of F ′, S ∗ ∩ XM ⊂ SM ∩ U (x0, r 2 ) ⊂ SM ∩ G and by Proposition 4.1 above, SM is hgi-porous at no point of S ∗ ∩XM . Consequently, A∩XM is, indeed, a member of a Foran system for hgi-porosity in XM -- the system F ′. (cid:3) 8 MAREK C ´UTH, MARTIN RMOUTIL 4.3. Remark. Note that the last proof depends solely on our ability to separably reduce hgi- porosity of a set at a point. It would work for any other type of porosity which fulfils this condition, e.g., the (g)-porosity (for the definition see [8] or [10]). Before proceeding to the last section where we use the propositions above, let us briefly turn our attention to the matter of lower porosity and formulate two related facts: 4.4. Lemma. For any suitable elementary submodel M the following holds: Let hX, ρi be a metric space, A ⊂ X and d(·, A) : X → R the function defined by the formula d(·, A)(x) := d(x, A). Then whenever M contains X and A then d(·, A) is an element of M . Proof. Let us fix a (∗)-elementary submodel M containing X such that A ∈ M . Then the lemma follows immediately from Lemma 2.2 and from the absoluteness of the following formula and its subformulas (∗) ∃d(·, A) (d(·, A) is a function which maps every x ∈ X to the real number inf{ρ(x, a); a ∈ A}). (cid:3) Finally, we present the following proposition (its proof is contained in the proof of Propo- sition 4.10 in [1]). 4.5. Proposition. For any suitable elementary submodel M the following holds: Let hX, ρi be a metric space and A ⊂ X. Then whenever M contains X and A, it is true that for every x ∈ A ∩ M A is not lower porous at x → A ∩ XM is not lower porous at x in the space XM . Note, that this is exactly the moment, where we were unable to reduce the property of not being lower porous at a point. However, thanks to Proposition 3.6, this proposition will be sufficient. 5. Main results In the main part of this article we show that the set properties "to be σ-upper porous" and "to be σ-lower porous" are separably determined. We formulate the related theorems in the language of elementary submodels (which is useful when we want to combine several results concerning elementary submodels together). However, we also formulate a corollary of these results in such a setting that no knowledge of elementary submodels is required (see Theorem 1.1). Next, we show that these results may be useful for proving that some results concerning separable spaces hold in a nonseparable setting as well. This is demonstrated on Theorem 5.6. First, let us show that σ-upper porosity is a separably determined notion. 5.1. Theorem. For any suitable elementary submodel M the following holds: Let hX, ρi be a topologically complete metric space, g a porosity function and A ⊂ X a Suslin set. Then whenever M contains X and A, it is true that A is σ-upper porous in X ↔ A ∩ XM is σ-upper porous in XM . Moreover, if A is Gδ and M contains also g, then A is not σ-hgi-porous in X → A ∩ XM is not σ-hgi-porous in XM . σ-POROSITY IS SEPARABLY DETERMINED 9 Proof. Let us fix a (∗)-elementary submodel M containing X such that g, A ∈ M . Assume the set A is of the type Gδ; we shall prove the second part of the proposition first. Due to Lemma 3.4 and the absoluteness of the formula (and its subformulas) (∗) ∃B (B ⊂ A and B is a member of a Foran system for hgi-porosity), we can assume that the set A is a member of a Foran system F for hgi-porosity. Hence the set A ∩ XM is a member of a Foran system F ′ for hgi-porosity in XM (Proposition 4.2) and thus is not σ-hgi-porous in XM (Proposition 3.1). The implication from the left to the right for σ-upper porosity follows immediately from Lemma 4.4 and [1, Corollary 4.13]. We shall prove the other implication indirectly; owing to Theorem 3.5 we can assume that A is Gδ again (even closed). The result now follows from the already proved part and Lemma 3.3, using the absoluteness of the formula (and its subformulas) (∗) ∃g (g : R → R is a function such that for all x ∈ R is g(x) = 3x). (cid:3) 5.2. Remark. It is not known to the authors whether the other implication for σ-hgi-porosity holds. However, under the assumptions of the preceding theorem, it is true that whenever A is σ-hgi-porous then A ∩ XM is σ-hdgi-porous in XM for any d > 2. This may be established in the following way: First, using the ideas presented in [1] (mainly Proposition 4.12 and Corrolary 4.13), we are able to see that if A is σ-(g, c)-porous in X (where c > 0; the definition is natural -- see [8]), then A ∩ XM is σ-(g, c/2)-porous in the space XM . Now let us assume the set A is σ-hgi-porous in X. Then [8, Lemma 3.1(ii)] implies it is σ- (g, 1/2)-porous in X and thus A ∩ XM is σ-(g, 1/4)-porous in the space XM . In the nontrivial case when there exists a δ > 0 such that g(x) > x for all x ∈ (0, δ) (if that is not the case, then the notion of hgi-porosity is usually not very interesting) it is not difficult to prove that g satisfies the assumption from [8, Proposition 4.4]. Thus A ∩ XM is σ-(g, c)-porous for any c ∈ (0, 1/2). To pass back to h·i-porosity, we use a slightly refined version of [8, Lemma 3.1(i)] which for any d > 1 states that (f, d)-porosity of a given set N at a given point x implies hf i-porosity of N at x. We easily obtain that the set A ∩ XM is σ-hdgi-porous for any d > 2. Moreover, under the additional assumption that there exists a d > 2 and a δ > 0 such that g(x) > dx for any x ∈ (0, δ), we are able to prove (similarly as above) that whenever A is σ-hgi-porous then A ∩ XM is σ-hgi-porous in XM . 5.3. Remark. Under the assumptions of Theorem 5.1 the following holds: If g is a porosity function such that for some c > 0 there is a δ > 0 such that cg(x) > x for all x ∈ (0, δ), then A is σ-(g)-porous in X ↔ A ∩ XM is σ-(g)-porous in XM . This can be established as follows: Let d = 12c and let A be non-σ-(g)-porous in X. Then it is non-σ-(dg, 1)-porous and thus A is non-σ-(cid:10) d 2 g(cid:11)-porous in X. Theorem 5.1 asserts that the same holds also for A ∩ XM in XM . Hence, A ∩ XM is non-σ-(cid:0) d 2 g, 2(cid:1)-porous ([8, Lemma 3.1(i)]), i.e., it is non σ-(cid:0) d 12 g = cg satisfies the assumption of [8, Proposition 4.4], we obtain that A ∩ XM is not σ-(g)-porous in XM . 3(cid:1)-porous (in XM ). Now, since the function d 12 g, 1 For the proof of the other implication assume that the set A is σ-(g)-porous. It is easy to see n=1 An. 2 (cid:1)-porous in XM that there exist (g, cn)-porous sets An (with cn > 0 for each n ∈ N) such that A = S∞ In the same way as in the previous remark we obtain that An ∩ XM is (cid:0)g, cn for each n. Hence, A ∩ XM is σ-g-porous in XM . 10 MAREK C ´UTH, MARTIN RMOUTIL We shall now turn our attention to σ-lower porosity and show it is separably determined. 5.4. Theorem. For any suitable elementary submodel M the following holds: Let hX, ρi be a topologically complete metric space and let A ⊂ X be a Suslin set. Then whenever M contains X and A, it is true that A is σ-lower porous in X ↔ A ∩ XM is σ-lower porous in XM . Proof. Let us fix a (∗)-elementary submodel M containing X such that A ∈ M . Then the implication from the left to the right follows from [1, Corollary 4.13] and from Lemma 4.4. To prove the opposite implication we use Proposition 3.6. Let us assume that the set A is not σ-lower porous in X. Then (∗) ∃F ∃D (F ⊂ A is a nonempty closed set such that D ⊂ F is dense in F and F is not lower porous at any point of D). By the absoluteness of this formula (and its subformulas) above, we are able to find sets F, D ∈ M satisfying the conditions above. Using Lemma 2.9 and Lemma 2.10 we can see that F ∩ M 6= ∅ and D ∩ M is dense in F ∩ XM . Moreover, by Proposition 4.5, F ∩ XM is not lower porous at any point of D ∩ M . Thus, from Proposition 3.6 it follows that the set A ∩ XM is not σ-lower porous in the space XM . (cid:3) Theorem 1.1 from the introduction is just an easy consequence of Theorem 5.1, Theorem 5.4 and Proposition 2.7 since Convention 2.3 allows us to combine these three results; by doing that we obtain a theorem in the setting of Banach spaces which concerns both types of porosity. In a similar way, Theorem 1.2 follows from the Theorem 5.1, Theorem 5.7 in [1] and Proposition 2.7 (because the set of the points where a function is Fr´echet differentiable is a Fσδ set - see for example [1, Theorem 5.8]) ). Finally, we give the following application of our results. In [9] is proved the following theorem (we use the more common terminology from [3]). Definition. Let hX, k·ki be a Banach space and let f be a real function defined on X. We say that f is Fr´echet superdifferentiable at x ∈ X if and only if there exists a x∗ ∈ X ∗ such that (f (x + h) − f (x) − x∗(h)) ≤ 0. lim sup h→0 khk 5.5. Theorem ([9, Theorem 2]). Let hX, k·ki be a Banach space with separable dual space and let G ⊂ X be an open set. Let f be a Lipschitz function on G and let A be the set of all the points x ∈ G such that f is Fr´echet superdifferentiable at x and f is not Fr´echet differentiable at x. Then A is σ-upper porous. Using the method of elementary submodels, it is now easy to extend the validity of this result to general Asplund spaces. 5.6. Theorem. Let hX, k·ki be an Asplund space and let G ⊂ X be an open set. Let f be a Lipschitz function on G and let A be the set of all the points x ∈ G such that f is Fr´echet superdifferentiable at x and f is not Fr´echet differentiable at x. Then A is σ-upper porous. Proof. Let us denote by D(f ) the set of points where f is Fr´echet differentiable and by S(f ) the set of points where f is Fr´echet superdifferentiable. Then, using Theorem 5.1, Proposition 2.7 and [1, Theorem 5.7], take an elementary submodel M satisfying: • XM is a separable subspace of X, σ-POROSITY IS SEPARABLY DETERMINED 11 • D(f ) ∩ XM = D(f ↾XM ), • A is σ-upper porous if and only if A ∩ XM is σ-upper porous in the space XM . Note that A ∩ XM ⊂ {x ∈ XM ; x ∈ S(f ↾XM ) \ D(f ↾XM )} and that the set on the right side is σ-upper porous (because XM is a separable space with separable dual); thus the set A is σ-upper porous. (cid:3) References [1] M.C´uth, Separable reduction theorems by the method of elementary submodels, preprint [2] A.Dow: An introduction to applications of elementary submodels to topology, Topology Proc. 13 (1988), 17-72. [3] A.Y.Kruger: On Fr´echet subdifferentials, J. Math. Sci. (N. Y.) 116 (2003), no. 3, 3325-3358. [4] W.Kubi´s: Banach spaces with projectional skeletons, J. Math. Anal. Appl. 350 (2009), no. 2, 758-776. [5] K.Kunen: Set Theory, Stud. Logic Found. Math., vol. 102, North-Holland Publishing Co., Amsterdam, 1983. [6] J.Lindenstrauss, D.Preiss, J. Tiser: Fr´echet differentiability of Lipschitz functions and porous sets in Banach spaces, Princeton Math. Ser., 2012 [7] M. Rmoutil, Products of non-σ-lower porous sets, preprint [8] L. Zaj´ıcek, Sets of σ-porosity and sets of σ-porosity (q), Casopis Pest. Mat. 101 (1976), 350-359. [9] L. Zaj´ıcek, A Generalization of an Ekeland-Lebourg theorem and the differentiability of distance functions, Rend. Circ. Mat. Palermo 2 (1984), Suppl. No. 3, 403-410. [10] L. Zaj´ıcek, Porosity and σ-porosity, Real Anal. Exchange 13 (1987/88), 314-350. [11] L. Zaj´ıcek, Products of non-σ-porous sets and Foran systems, Atti Semin. Mat. Fis. Univ. Modena Reggio Emilia, 44 (1996), 497-505. [12] L. Zaj´ıcek: On sigma-porous sets in abstract spaces, Abstr. Appl. Anal. 5 (2005), 509-534. [13] L. Zaj´ıcek, M. Zelen´y, Inscribing closed non-σ-lower porous sets into Suslin non-σ-lower porous sets, Abstr. Appl. Anal. 3 (2005), 221-227. [14] L. Zaj´ıcek, M. Zelen´y, Inscribing compact non-σ-porous sets into analytic non-σ-porous sets, Fund. Math. 185 (2005), 19-39. [15] M. Zelen´y, J. Pelant, The structure of the σ-ideal of σ-porous sets, Comment. Math. Univ. Carolin. 45 (2004), 37-72. E-mail address: [email protected], [email protected] Charles University, Faculty of Mathematics and Physics, Sokolovsk´a 83, 186 75 Praha 8 Karl´ın, Czech Republic

Subsets and Splits