On free bases of Banach spaces
Abstract.
We call a closed subset of a Banach space a free basis of if it contains the null vector and every Lipschitz map from to a Banach space , which preserves the null vectors can be uniquely extended to a bounded linear map from to . We then say that two complete metric spaces and are -equivalent if they admit bi-Lipschitz copies and , respectively that are free bases of a common Banach space satisfying .
In this note, we compare -equivalence with some other natural equivalences on the class of complete metric spaces. The main result states that -equivalent spaces must have the same Δech-Lebesgue covering dimension. In combination with the work of Godard, this implies that two complete metric spaces with isomorphic Lipschitz-free spaces need not be -equivalent. Also, there exist non-homeomorphic -equivalent metric spaces, and, in contrast with the covering dimension, the metric Assouad dimension is not preserved by -equivalence.
Key words and phrases:
metric space, covering dimension, free basis, Lipschitz map, Lipschitz-free space2020 Mathematics Subject Classification:
46B03, 54E35, 54F45.1. Introduction
Vector spaces throughout this paper are considered over the field of real numbers. The zero vector of a vector space is denoted by . We say that two normed spaces are isomorphic provided that there exists a linear bijection between them which is simultaneously a homeomorphism. Similarly, we call two metric spaces bi-Lipschitz isomorphic if there exists a bijection between them that is Lipschitz and whose inverse is also Lipschitz. By covering dimension of a metric space we mean the Δech-Lebesgue covering dimension (or equivalently the large inductive dimension) of its underlining topological space.
We will now introduce two notions that will be central to our work.
Definition 1.1.
Given a Banach space , we say that its closed subset is a free basis of provided that and for every Banach space and every Lipschitz map which preserves zero vectors it is possible to uniquely extend to a bounded linear map .
We note that every complete metric space is a free basis of some Banach space. This is thanks to the existence of free Banach spaces recalled at the beginning of Section 2 (see Remark 2.2 (i)).
Definition 1.2.
Given two complete metric spaces and , we say that and are -equivalent provided that there are free bases and of common Banach space such that is bi-Lipschitz isomorphic to and to , and that .
The notation of -equivalence emphasizes the fact that and are -equivalent if and only if their corresponding normed spaces of molecules constructed by Arens and Eells in [6] are isomorphic. This follows from the universal property of the completion of the space of molecules established by Weaver in [20, Theorem 3.6]. The Banach space obtained by such completion will be referred to as Lipschitz-free space and described in Section 2.
We will study the relations of -equivalence to some other natural equivalences on the class of complete metric spaces. More specifically, we will deal with equivalences given by: bi-Lipschitz isomorphisms, homeomorphisms, covering dimension, and isomorphisms of the corresponding Lipschitz-free spaces (which we will refer to as -equivalence).
Obviously, a bi-Lipschitz isomorphism is a homeomorphism, and homeomorphisms preserve covering dimension. Also, bi-Lipschitz isomorphic spaces are -equivalent and, because Lipschitz-free spaces are completions of spaces of molecules, -equivalence implies -equivalence. It is proved in [20, Theorem 3.55] that if two complete metric spaces that are uniformly concave (see [20, Definition 3.33]) have linearly isometric Lipschitz-free spaces, then there exists a bijection between such metric spaces and a positive constant such that the distance of any two images is just the multiple of the distance of their preimages. In particular, the metric spaces are bi-Lipschitz isomorphic and thus -equivalent. However, as we will see in the sequel, mere -equivalence does not ensure -equivalence, and the latter does not even imply that the metric spaces are homeomorphic.
No other implication can be added to the following diagram (here dim stands for the relation of having the same covering dimension):
The fact that the implication (a) does not hold is implicitly contained in [3], it is discussed in detail in Remark 3.2 and follows also from the non-existence of the implication (c). Non-trivial examples of homeomorphic spaces that are not -equivalent, and hence witness that the implication (b) does not hold, were constructed in [16, 11]. Naor and Schechtman showed in [16] that the homeomorphic discrete spaces and with their Euclidean metrics do not have isomorphic Lipschitz-free spaces, while Godard proved in [11] that the homeomorphic Cantor sets one of zero and the other of non-zero Lebesgue measure have non-isomorphic Lipschitz-free spaces. We will address the failure of the implication (c) in Section 3, where we recall a construction of -equivalent spaces that can be found for instance in [1] (see also [14]) and leads to examples of non-homeomorphic -equivalent spaces such as in Example 3.6. Note that there are many examples of non-homeomorphic -equivalent spaces. To mention a few, see [2, 10, 11, 14, 20]. Nevertheless, these results are based on (techniques similar to) the PeΕczyΕskiβs decomposition method and therefore it is not clear, whether they can give examples of -equivalent spaces. We also point out on two classical spaces that -equivalence does not preserve Assouad dimension (Example 3.7). This should be compared with Corollary 4.4 of the last section.
In the main result of our note, Theorem 4.3 and Corollary 4.4, we prove that -equivalent spaces must have the same covering dimension, i.e. the implication (d). On the other hand, by a result due to Godard, [11, Corollary 3.4], the closed unit interval and a Cantor set of a positive measure have isomorphic Lipschitz-free spaces and are therefore -equivalent while having different covering dimensions. This shows the non-existence of the implication (e). So in particular, -equivalence does not imply -equivalence. Let us also mention in this context a classical open problem in the area of Lipschitz-free spaces, which is to determine whether the Lipschitz-free spaces over Euclidean spaces of different dimensions strictly greater than one might be isomorphic. Corollary 4.4 implies that, should there exist an isomorphism, it cannot preserve the finiteness of supports introduced in [4, Definition 2.5].
2. Some basic properties of free bases and -equivalence
A pointed metric space is a metric space with a distinguished base point . We consider a base point of a normed space to be its null vector. Given a Banach space , we denote by the Banach space of all -valued Lipschitz maps that preserve the base points, endowed with the norm which assigns to each its Lipschitz number . As usually, we write instead of .
Recall that given a pointed metric space , the Lipschitz-free space is the unique (up to a linear isometry) Banach space characterized by the following universal property (see [20, Theorem 3.6]): There is an isometry which preserves the base points and for every Banach space and every , there is a unique bounded linear map making the following diagram commutative
(1) | Β Β Β |
and such that , where is the Lipschitz number of . We will write instead of when we need to emphasize the underlying metric space . Moreover, for the map that assigns to each the bounded linear functional as in (1) is a linear isometry from onto (see for instance [20, Theorem 3.3]). The weakβ topology of induced by agrees on bounded sets with the topology of point-wise convergence of functions. It is a simple and folklore fact, that does not depend on the choice of a base point of .
Now we relate the notions of free bases and -equivalence to Lipschitz-free spaces. It is immediate in view of the extension property of the free bases and the universal property (1) of the Lipschitz-free spaces.
Observation 2.1.
-
(i)
If is a free basis of a Banach space , then is isomorphic to .
-
(ii)
Two complete metric spaces are -equivalent if and only if there exists an isomorphism between and that maps onto .
In the following remark, we summarize some basic properties of free bases and -equivalent spaces that follow from the definitions and Observation 2.1.
Remark 2.2.
-
(i)
For a metric space , is a free basis of . Since exists for every metric space , it follows that every complete metric space can be isometrically embedded into some Banach space as its free basis.
-
(ii)
Complete metric spaces are -equivalent if and only if they admit bi-Lipschitz copies , respectively, that are free bases of a common Banach space.
-
(iii)
By [20, Corollary 3.46], infinite-dimensional reflexive Banach spaces do not admit free bases.
-
(iv)
Note that if is a free basis of , then there is such that for every Lipschitz map from to a Banach space that preserves the base point, its bounded linear extension satisfies . Indeed, the extension of the isometric embedding is an isomorphism of and . From (1) we get . Since , it suffices to put . Thus the definition of a free basis can be expressed formally stronger by requiring the existence of such a .
-
(v)
If is a free basis of a Banach space , then must be linearly independent and is dense in . Indeed, the first property follows from the existence of linear extensions of Lipschitz maps from to the whole (or just ), and the second follows from its uniqueness, both required by the definition of the free basis.
We proceed by characterizing the -equivalence of metric spaces in terms of the corresponding spaces of Lipschitz functions.
Lemma 2.3.
Let and be metric spaces such that there exists a linear isomorphism from onto . Then if and only if the adjoint is continuous in the topology of point-wise convergence.
Proof.
First assume that and consider a net (not necessarily bounded) that converges point-wise to some . Fix . Then there exist , , such that and clearly
Hence is continuous with respect to the topologies of point-wise convergence.
For the converse, assume that there exists such that . For any finite , find by the Hahn-Banach theorem an such that and . Then , directed by the set inclusion, is a net in that converges point-wise to the zero function in . Indeed, for any and any finite such that we have . However, for every finite we have and does not converge point-wise to in . So we conclude that if is continuous in the topology of point-wise convergence, then . β
The following is a direct consequence of Lemma 2.3, Observation 2.1 and the description of weakβ topology on bounded sets of .
Proposition 2.4.
Complete metric spaces are -equivalent if and only if there exists an isomorphism between and that is also a homeomorphism in the topology of point-wise convergence.
Remark 2.5.
Given a Tychonoff space with a distinguished base point, denote by the free abelian topological group over in the sense of Graev (see [12] or [5] for the definition). It is an abelian topological group with the property, that there is a homeomorphic embedding which preserves the base points and such, that for every continuous map from to an abelian topological group that preserves the base points, there is a unique continuous homomorphism making the following diagram commutative:
Β Β Β |
During his construction of the free abelian topological group Graev extended in a specific way every continuous pseudometric on to a translation-invariant continuous pseudometric on . As in [8], we denote with such a pseudometric . If is a metric on , then becomes a metric group and is an isometric embedding. It was noticed by Chasco et al. in [8, Theorem 2.11] that the metric group generated by in is (isometrically isomorphic to) (i.e. there is a group isomorphism of the group generated by onto which is an isometry). The metric group uniquely determines111alternatively, similarly as in [17], we may say that the metric group has invariant linear span in the class of normed spaces - that is, whenever isometrically and algebraically embeds into a normed space, then it produces the same linear span (up to a linear isometry). the normed space of molecules on constructed by Arens and Eells in [6], and, consequently, its completion β the Lipschitz-free space β by a standard construction. Namely, using positive homogeneity of norms, we can extend the distance of each element of from zero to every finite rational linear combination of elements of because appropriate integer multiple of such a linear combination is an element of . Since the rational span of is dense in , this distance can be uniquely extended to the whole . It follows that if for two complete metric spaces there is a group isomorphism of the corresponding metric groups which is at the same time bi-Lipschitz, then and are -equivalent.
3. Construction of -equivalent spaces
In this section, we present two constructions of -equivalent spaces that provide important examples.
The following fact is a part of [3, Theorem 3.9 and Proposition 4.2] translated to the language of free bases.
Fact 3.1.
Let be a free basis of a Banach space . Define by
If is an isolated point of , then
-
(1)
is a free basis of (in particular, is complete);
-
(2)
is a homeomorphic embedding.
In particular, if is an isolated point of , then and are -equivalent and homeomorphic metric spaces, and is bounded.
Remark 3.2.
Fact 3.1 and Remark 2.2 (i) show that every unbounded complete metric space with an isolated point is -equivalent to a homeomorphic bounded complete metric space. Since bounded and unbounded metric spaces are not bi-Lipschitz isomorphic, this provides a source of examples of -equivalent spaces that are not bi-Lipschitz isomorphic.
Now we proceed to building examples of -equivalent spaces that are not even homeomorphic.
Similarly as in [20], given two pointed metric spaces , their sum is their disjoint union with glued base points, and the metric remains the same on and and we define the distance of each and as . Following [20], if is a closed subset of a complete metric space , then stands for the equivalence relation on such that if or if . We define the distance of two equivalence classes as . By [20, Proposition 1.26] this distance is a complete metric on the set of the equivalence classes. We denote this quotient metric space by . If is a pointed metric space, we consider to be a pointed metric space with a base point being the equivalence class which contains the base point of .
The quotient metric spaces can be used to produce -equivalent metric spaces as follows. Fact 3.3 can be extracted from [1, Theorem 2.13]. We sketch the proof to show the exact relation.
Fact 3.3.
Let be a complete metric space and a Lipschitz retraction. Pick a base point arbitrarily. Then is -equivalent to .
Proof.
Denote the natural quotient map. It is not difficult to verify that the mapping defined by is an isomorphism with the inverse given for a function by
Moreover, it is a homeomorphism in the topology of point-wise convergence. We can now apply Proposition 2.4 to conclude the -equivalence. β
A counterpart of this construction for free bases of Banach spaces takes the following form.
Theorem 3.4.
Let be a free basis of a Banach space and a bounded linear projection. Denote by the corresponding projection defined as . If , then is a free basis of . In particular, and are -equivalent.
Before proving the theorem, we show that the set is closed.
Lemma 3.5.
The sets and as in Theorem 3.4 are closed in .
Proof.
For this argument, we consider endowed with the equivalent norm inherited from . Then the following well-known fact can be derived from [12] (see Remark 2.5, for its proof see also [6, 8, 20]): Given , we have
(2) |
The free basis is closed by definition and so is because is continuous. Since , it follows that is closed in .
To show that is closed in , we pick a Cauchy sequence of elements of arbitrarily and find its subsequence which converges to an element of .
If , then and we are done. Otherwise, there is and a subsequence of such that For each pick such that . Thus we have
(3) |
As is Cauchy, we may fix such that
(4) |
Because is Cauchy, so is by (4). Since is complete, has limit . Since is bounded, has limit . Hence has limit . This finishes the proof. β
Now we can continue to prove the theorem above.
Proof of Theorem 3.4.
Pick a Banach space and arbitrarily. Define by
We claim that . Indeed, for we have
Thus can be extended to a bounded linear map on .
Let us show, that the restriction of to is .
Indeed, for we have and consequently
Similarly, given , there is such that . Thus
To summarize: is a bounded linear map which extends , and this extension is unique because is linearly dense in . Since is a closed subset of by Lemma 3.5, this shows that is a free basis of .
Obviously, . Hence and are -equivalent. β
Using the quotient spaces, we can find examples demonstrating that -equivalent spaces need not be homeomorphic.
Example 3.6.
There are non-homeomorphic -equivalent spaces. Indeed, since is a Lipschitz retract of , the spaces and are -equivalent by Fact 3.3.
Note that, by Corollary 4.4, two -equivalent spaces must have the same covering dimension. Nevertheless, the next example shows that Assouad dimension of a metric space is not preserved by -equivalence. This dimension was introduced in [7]. For the definition and its properties we refer the reader to [15], where it is explained, that a metric space is doubling if and only if it has a finite Assouad dimension [15, Facts 3.3]. We shall also need the trivial facts, that every metric space with infinite bounded uniformly discrete subset has infinite Assouad dimension and that the Assouad dimension of natural numbers (with the usual metric) is equal to 1.
Example 3.7.
There are -equivalent spaces such that one has Assouad dimension 1 and the other has Assouad dimension . In particular, there are two -equivalent spaces one of which has the doubling property and the other one does not.
Indeed, if is an arbitrary bounded separable uniformly discrete space, then is isomorphic to via the isomorphism given by . So we can use Observation 2.1 (ii) to conclude that and are -equivalent.
4. -equivalence preserves the covering dimension
Given topological spaces and ,we denote the fact that is a mapping from to the power set of by . We say that is finite-valued, if is a finite subset of for all . Further, we say that is lower semi-continuous (abbreviated by lsc), if for every open the set
is open in .
As in [19, Definition 3.3], given a linear basis of a vector space we define the finite-valued support function by letting be the unique set such that
Here we use the convention that sum over an empty set is the null vector.
The following relatively simple fact is a direct consequence of [19, Proposition 5.6].
Fact 4.1.
Let be a linear basis of a vector space and some topology on . Assume that for every -closed subset of and all there is a continuous linear functional such that and . Then is lsc.
[18, Corollary 5.4] provides us with the following fact, where stands for the covering dimension of the metric space .
Fact 4.2.
For suppose that is a perfectly normal space, and is a finite-valued lower semi-continuous mapping. If for every there exists with and , then .
Combining the above two facts, we can derive the following result about the correspondence of dimensions.
Theorem 4.3.
Let be a vector space endowed with a topology . For let be a linear basis of which is perfectly normal with respect to the topology . Assume that for every -closed subset of and all there is a continuous linear functional such that and . Then .
Proof.
Finally, we show that Theorem 4.3 applies to -equivalent complete metric spaces.
Corollary 4.4.
If and are -equivalent complete metric spaces, then .
Proof.
Since the covering dimension is preserved by bi-Lipschitz maps, we may and will assume that and are free bases of a common Banach space and .
Put , , and consider with the norm topology inherited from . We claim that the conditions of Theorem 4.3 are satisfied. Indeed, for both , is linearly independent by Remark 2.2 (v) and thus it is a linear basis of . Moreover, given a closed subset and , there is a function such that and . Since is a free basis, such a function can be extended to a bounded linear functional on and the equality follows.
To complete the proof it suffices to observe that given a point of a perfectly normal space (of a metric space in particular), we have . This folklore fact follows from the Countable Sum Theorem [9, 3.1.8] and from the Subspace Theorem for the Δech-Lebesgue dimension of perfectly normal spaces [9, 2.1.3 and 3.1.19], and from the fact that open subsets of a perfectly normal space are -sets. β
Remark 4.5.
Let us note, that for bounded separable metric spaces, Corollary 4.4 can be proved by a technique developed in [13]. Indeed, for -equivalent spaces , by Proposition 2.4, the spaces and are linearly homeomorphic (hence uniformly homeomorphic) in the topology of point-wise convergence. Consequently, the spaces and of all Lipschitz functions on and respectively are uniformly homeomorphic in the topology of point-wise convergence. Since the space of Lipschitz functions on a metric space is closed under point-wise addition and scalar multiplication and for bounded metric space also under multiplication (see for instance [20, Propositions 1.29 and 1.30]), and Lipschitz functions separate points and closed sets, and form so called QS-algebras (see [13, 1.1 Definition]). To derive Corollary 4.4 for bounded separable metric spaces it suffices now to apply [13, 1.4. Proposition].
References
- [1] F. Albiac, J. L. Ansorena, M. CΓΊth, and M. Doucha, Embeddability of and bases in Lipschitz free -spaces for . J. Funct. Anal. 278 (2020), art. 108354.
- [2] F. Albiac, J. L. Ansorena, M. CΓΊth, and M. Doucha, Lipschitz free spaces isomorphic to their infinite sums and geometric applications. Trans. Amer. Math. Soc. (2021) 7281β7312.
- [3] F. Albiac, J. L. Ansorena, M. CΓΊth, and M. Doucha, Lipschitz algebras and Lipschitz-free spaces over unbounded metric spaces . Int. Math. Res. Not. (2021).
- [4] R. J. Aliaga, E. PerneckΓ‘, C. Petitjean and A. ProchΓ‘zka, Supports in Lipschitz-free spaces and applications to extremal structure, J. Math. Anal. Appl. 489 (2020), art. 124128.
- [5] A. V. Arkhangelskii and M. Tkachenko, Topological groups and related structures, Atlantis Press, Paris; World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ (2008).
- [6] R. F. Arens and J. Eells, Jr., On embedding uniform and topological spaces, Paciο¬c J. Math. 6 (1956), 397β403.
- [7] P. Assouad, Espaces mΓ©triques, plongements, facteurs, ThΓ¨se de doctorat dβΓtat, Publ. Math. Orsay 223β7769, Univ. Paris XI, Orsay, (1977).
- [8] M. J. Chasco, X. DomΓnguez and M. G. Tkachenko, Topological groups of Lipschitz functions and Graev metrics J. Math. Anal. Appl. 508 (2022), art. 125882.
- [9] R. Engelking, Theory of Dimensions, Finite and Infinite, Heldermann Verlag, Berlin, (1995).
- [10] L. C. GarcΓa-Lirola and A. ProchΓ‘zka, PeΕczyΕski space is isomorphic to the Lipschitz free space over a compact set, Proc. Amer. Math. Soc. 147 (2019), 3057β3060.
- [11] A.Β Godard, Tree metrics and their Lipschitz-free spaces, Proc. Amer. Math. Soc., 138 (2010), no. 12, 4311β4320.
- [12] M. I. Graev, Free topological groups, in: Topology and Topological Algebra, in: Translations Series 1, vol. 8, American Mathematical Society, 1962, pp. 305β364, Russian original in: Izv. Akad. Nauk SSSR, Ser. Mat. 12 (3) (1948) 279β324.
- [13] S. P. Gulβko, On uniform homeomorphisms of spaces of continuous functions, (Russian) Trudy Mat. Inst. Steklov. 193 (1992), 82-88; English translation: Proc. Steklov Inst. Math. 193 (1993), 87β93.
- [14] P. L. Kaufmann, Products of Lipschitz-free spaces and applications, Studia Math. 226 (2015), no. 3, 213β227.
- [15] J. Luukkainen, Assouad dimension: antifractal metrization, porous sets, and homogeneous measures, J. Korean Math. Soc. 35, (1998), 23β76.
- [16] A. Naor and G. Schechtman, Planar earthmover is not in , SIAM J. Comput. 37, (2007), 804-826.
- [17] E. PerneckΓ‘ and J. SpΔvΓ‘k, Topological groups with invariant linear spans. Rev. Mat. Complut. 35, 219β226 (2022).
- [18] J. SpΔvΓ‘k, Finite-valued mappings preserving dimension, Houston J. Math. 37 (2011), 327β348.
- [19] J. SpΔvΓ‘k, Module-valued functors preserving the covering dimension, Comment. Math. Univ. Carolin. 56 (3) (2015), 377β399.
- [20] N. Weaver, Lipschitz algebras, 2nd ed., River Edge: World Scientific Publishing Co., (2018).