Characterizing Lipschitz images of injective metric spaces

The implications (5)\Leftrightarrow(1)\Leftrightarrow(2) are proved in [BL, Proposition 1.2]; (3)\Rightarrow(2) in [AP, Theorem 8], and (2)\Leftrightarrow(4) in [BL, Proposition 1.4]. To see that (4)\Rightarrow(3), take any indexed family \big(B[x_i,r_i]\big)_{i\in I} of closed balls in X such that d_X(x_i,x_j)\leq r_i+r_j for every i,j\in I. The metric convexity of X ensures that for any distinct indices i,j\in I, the closed balls B[x_i,r_i] and B[x_j,r_j] have a common point. The binary intersection property guarantees that \bigcap_{i\in I}B[x_i,r_i]\neq\emptyset.

The implications (1)\Rightarrow(2)\Rightarrow(3) are obvious. To prove that (3)\Rightarrow(1), assume that a metric space X satisfies the condition (3). Given any points x,y\in X, we need to show that the metric segment [x,y] is isometric to the interval [0,d_X(x,y)] in the real line. By the condition (3), there exists a unique 1-Lipschitz map f:[0,d_X(x,y)]\to X such that f(0)=x and f(d_X(x,y))=y. The triangle inequality implies that the image I\coloneqq f\big[[0,d_X(x,y)]\big] is a subset of the metric segment [x,y]. Assuming that [x,y]\ne I, we can find a point z\in [x,y]\setminus I. Then d_X(x,y)=d_X(x,z)+d_X(z,y). By condition (3), there exist 1-Lipschitz maps \alpha:[0,d_X(x,z)]\to X and \beta:[0,d_X(z,y)]\to X such that \alpha(0)=x, \alpha(d_X(x,z))=z=\beta(0) and \beta(d_X(z,y))=y. The 1-Lipschitz maps \alpha,\beta compose a 1-Lipschitz map \gamma:[0,d_X(x,y)]\to X defined by

Since \gamma(0)=x=f(0), \gamma(d_X(x,y))=y=f(d_X(x,y)), and \gamma\ne f, we obtain a contradiction with the uniqueness of the 1-Lipschitz path f. This contradiction shows that [x,y]=I. The triangle inequality implies that the 1-Lipschitz map f:[0,d_X(x,y)]\to [x,y] is an isometry, witnessing that the metric space X is \mathbb R-convex. Indeed, for every u,z such that 0 \leq u\leq z \leq d_X(x,y) we have |u-z| \geq d_X(f(u),f(z)) \geq d_X(x,y)-d_X(x,f(u))-d_X(f(z),y) \geq d_X(x,y)-|0-u|-|z-d_X(x,y)|=|u-z|.

Since the metric space X is \mathbb R-convex, there exist isometries \alpha:[0,d_X(x,y)]\to[x,y] and \beta:[0,d_X(x,z)]\to [x,z] such that \alpha(0)=x=\beta(0), \alpha(d_X(x,y))=y, and \beta(d_X(x,z))=z. By the compactness of [x,z], the closed set \alpha^{-1}[x,z] in [0,d_X(x,y)] has the largest element s. Since \alpha is an isometry, s=d_X(\alpha(0),\alpha(s))=d_X(x,\alpha(s)). Let u=\alpha(s) and t=\beta^{-1}(u). Since \beta is an isometry, t=d_X(\beta(0),\beta(t))=d_X(x,u)=s. Observe that \alpha'\coloneqq \alpha{\restriction}_{[0,s]}:[0,s]\to X and \beta'\coloneqq \beta{\restriction}_{[0,t]}:[0,t]\to X are two 1-Lipschitz maps from the interval [0,s]=[0,t] to X such that such that \alpha'(0)=x=\beta'(0) and \alpha'(s)=u=\beta'(t). Proposition 2.2 ensures that \alpha'=\beta' and hence [x,y]\cap[x,z]=\alpha[[0,s]]=[x,u]. The uniqueness of the point u implies from the uniqueness of the end-points of the segment [x,u]=[x,y]\cap[x,z].

Given any point x\in \overline X_I with \overline \operatorname{id}(x)\in X, we need to prove that x\in X_I. Since x\in\overline X_I, there exists a sequence (x_n)_{n\in\mathbb{N}} in X_I such that d_{IX}(x,x_n)<\frac1{2^{n+2}} for all n\in\mathbb{N}. For every n\in\mathbb{N} we have

So, we can find a 1-Lipschitz path \gamma_n:[\tfrac1{2^{n+1}},\tfrac1{2^n}]\to X such that \gamma_n(\tfrac1{2^{n+1}})=x_{n+1} and \gamma_n(\tfrac1{2^n})=x_n. The paths \gamma_n, n\in\mathbb{N}, compose a 1-Lipschitz path \gamma:[0,1]\to X defined by \gamma(0)=\overline \operatorname{id}(x) and \gamma{\restriction}_{[\tfrac1{2^{n+1}},\tfrac1{2^n}]}=\gamma_n for all n\in\mathbb{N}. The definition of \gamma guarantees that

and hence x=\operatorname{id}^{-1}(\overline \operatorname{id}(x))\in \operatorname{id}^{-1}[X]=X_I.

(1) Suppose that the metric d_X is complete and take a Cauchy sequence (x_n)_{n \in \mathbb{N}} in the metric space (X,d_{IX}). Since d_X\le d_{IX}, the sequence (x_n)_{n \in \mathbb{N}} remains Cauchy in the metric space (X,d_{X}). Since (X,d_X) is complete, this sequence converges to some point x \in X in the metric space (X,d_X). Lemma 3.3 guarantees that \lim_{n \to \infty} x_n=x in (X,d_{IX}) too, witnessing that the intrinsic metric d_{IX} is complete.

(2) If the metric d_{IX} is analytic, then d_{IX} is separable and d_X is analytic, because of the continuity of the identity map (X,d_{IX})\to (X,d_X).

Next, assume that the metric d_{IX} is separable and d_X is analytic. Let \overline X be the completion of the metric space (X,d_X) and \overline X_I be the completion of the metric space X_I\coloneqq(X,d_{IX}). Let \bar I:\overline X_I\to\overline X be the continuous extension of the (1-Lipschitz) identity map I:X_I\to X. By Lemma 3.3, X_I=\bar I^{-1}[X]. Since the spaces \overline X_I and \overline X are Polish, the preimage X_I=\bar I^{-1}[X] is an analytic subspace of \overline X_I, by [Ke, 14.4].

The statement (3) follows from the 1-Lipschitz property of the identity map (X,d_{IX})\to(X,d_{IX}).

To prove the “if” part, assume that the metric space T_w is totally bounded. Then for every \varepsilon>0, there exists a finite set F''\subseteq T_w such that for every x''\in T_w, there exists y''\in F'' with d_w(x'',y'')<\varepsilon. Since F'' is finite, there exists a finite set F\subseteq T such that F''\subseteq\bigcup\{{\downarrow}(t,w_t):t\in F\}\subseteq T_w. Replacing F by the finite set {\downarrow}F\coloneqq\bigcup_{t\in F}{\downarrow}t, we can assume that F={\downarrow}F. We claim that \sum_{x\in L}w_x<\varepsilon for every finite linearly ordered set L\subseteq T\setminus F. Since L is finite and linearly ordered, there exists an element u\in L such that L\subseteq{\downarrow}u. By the choice of the set F'', there exists a point y''\in F'' such that d_w((u,w_u),y'')<\varepsilon. Write the element y'' as (y,y') for some y\in T and y'\in [0,w_y]. The choice of the set F ensures that y\in F and hence L\subseteq {\downarrow}u\setminus F\subseteq {\downarrow}u\setminus{\downarrow}y. Then

witnessing that the weighted tree (T,w) is precompact.

To prove the “only if” part, assume that the weighted tree (T,w) is precompact. To prove that the metric space T_w is totally bounded, fix any \varepsilon>0. By the precompactness of the weighted tree (T,w), there exists a nonempty finite set F\subseteq T such that \sum_{x\in L}w_x<\varepsilon for every finite linearly ordered set L\subseteq T\setminus F. Replacing F with the finite set {\downarrow}F\coloneqq\bigcup_{t\in F}{\downarrow}t, we can assume that F={\downarrow}F. Since the set F is finite, the set

is compact and hence totally bounded in the metric space T_w. Then there exists a finite set F''\subseteq K such that for every x''\in K there exists y''\in F'' with d_w(x'',y'')<\varepsilon. Replacing F'' by the set F''\cup \{(t,w_t):t\in F\}, we can assume that \{(t,w_t):t\in F\}\subseteq F''. We claim that for every x''\in T_w, there exists y''\in F'' such that d_w(x'',y'')<\varepsilon.

Indeed, if x''\in K, then the existence of a point y''\in F'' with d_w(x'',y'')<\varepsilon follows from the choice of the set F''. So, we assume that x''\notin K. In this case, x''=(x,x') for some x\in T\setminus F and x'\in (0,w_x]. Since r_T\in {\downarrow}F=F, there exists a point y\in F\cap{\downarrow}x such that F\cap{\downarrow}x={\downarrow}y. The choice of the set F'' ensures that it contains the point y''\coloneqq (y,w_y). The choice of F guarantees that \sum_{t\in{\downarrow}x\setminus {\downarrow}y}w_t=\sum_{t\in{\downarrow}x\setminus F}w_t<\varepsilon and hence d_w(x'',y'')=x'-w_x+\sum_{t\in {\downarrow}x\setminus {\downarrow}y}w_t\le \sum_{t\in {\downarrow}x\setminus {\downarrow}y}w_t<\varepsilon, by the definition of the metric d_w. This shows that the metric space T_w is totally bounded.

If the tree T is countable, then T_w\cap(T\times\mathbb Q) is a countable dense set in the \mathbb R-tree, witnessing that the metric space T_w is separable.

Next, assume that the tree T is uncountable. The definition of metric d_w ensures that for every node t\in T, the set \{t\} \times (0,w_t) is open in the metric space T_w. Therefore, the metric space T_w contains the uncountable family of pairwise disjoint open sets \big\{\{t\}\times(0,w_t):t\in T\big\} and hence T_w cannot be separable, according to [E, Corollary 4.1.16].

The definition of the metric on \overline T_w implies that the metric space T_w is a dense subspace of the metric space \overline T_w. To prove that \overline T_w is a completion of T_w, it suffices to show that every Cauchy sequence (s''_n)_{n\in\omega} in the metric space T_w has a limit in the space \overline T_w. If for some element t of T_w, the set \{n\in\omega:s''_n=t\} is infinite, then t is the limit of the Cauchy sequence (s''_n)_{n\in\omega} and we are done. So, we can assume that all elements of the Cauchy sequence (s''_n)_{n\in\omega} are pairwise distinct and moreover all of them are distinct from the root (r_T,0) of the \mathbb R-tree T_w. For every n\in\omega, write the element s''_n of T_w as a pair (s_n,s_n') where s_n\in T and s_n'\in(0,w_{s_n}]. Let \check s_n\coloneqq \max({\downarrow}s_n\setminus\{s_n\}) be the immediate predecessor of s_n in the tree T.

In the tree T, consider the set

and observe that it contains the root r_T of the tree T and hence is not empty. Two cases are possible.

1. First assume that the set \Lambda has no maximal elements. In this case \Lambda contains an increasing sequence (\lambda_n)_{n\in\omega}. For this sequence the set b\coloneqq \bigcup_{n\in\omega}{\downarrow}\lambda_n is an infinite branch of the tree T such that for every t\in b the set \{n\in\omega:t\preceq \check s_n\} is infinite. Write the branch b as b=\{b_n\}_{n\in\omega}, where b_n is a unique element of b such that |{\downarrow}b_n|=n+1. We claim that b\in\partial T_w. Since the sequence (s''_n)_{n\in\omega} is Cauchy, for every \varepsilon>0, there exists m\in\mathbb N such that d_w(s''_n,s''_m)<\varepsilon for every n\ge m.

Since the sequence (b_n)_{n\in\omega} is unbounded in the tree T, there exists a number i\in\omega such that b_i\not\preceq \check s_m. Given any number j\ge i, find a number n\ge m such that b_j\preceq \check s_n. Observe that b_i,b_j,\check s_n and \check s_n\wedge\check s_m are elements of the linearly ordered set {\downarrow}\check s_n. Assuming that \check s_m\wedge\check s_n\not\preceq b_i, we conclude that b_i\preceq\check s_m\wedge\check s_n\preceq \check s_m, which contradicts the choice of the number i. This contradiction shows that \check s_m\wedge\check s_n\preceq b_i.

In this case, the definition of the metric d_w ensures that

and hence \sum_{k=i+1}^\infty w_{b_k}\le\varepsilon<\infty, witnessing that b\in\partial T_w.

Since the sequence (s''_n)_{n\in\omega} is Cauchy and \varepsilon>0 was arbitrary, the equality b=\lim_{n\to\infty}s''_n will follow as soon as we check that \overline d_w(b,s''_n)<2\varepsilon for infinitely many numbers n. Indeed, for every number n\ge m in the infinite set \{n\in\omega:b_i\preceq \check s_n\} we have \check s_n\wedge \check s_m\preceq b_i and

2. Next, assume that the set \Lambda has a maximal element \lambda. Consider the infinite set I\coloneqq\{n\in\omega:\lambda\preceq \check s_n\}. If for some s\in T, the set J\coloneqq\{n\in I:s=s_n\} is infinite, then the maximality of \lambda ensures that \lambda is the immediate predecessor of s in the tree T. Then the subsequence (s_n'')_{n\in J} of the sequence (s''_n)_{n\in\omega} is contained in the compact set \{(\lambda,w_\lambda)\}\cup(\{s\}\times(0,w_s]) and hence the Cauchy sequence (s_n'')_{n\in\omega} converges to some points of this compact set. So, assume that for every s\in T, the set \{n\in I:s_n=s\} is finite. In this case we shall show that the Cauchy sequence (s_n'')_{n\in\omega} converges to the element \lambda''\coloneqq (\lambda,w_\lambda) of the \mathbb R-tree T_w. Since the sequence (s_n'')_{n\in\omega} is Cauchy and consists of pairwise distinct elements, it suffices to show that for every \varepsilon>0 there exists n\in\omega such that d_w(\lambda'',s_n'')<\varepsilon. Since the sequence (s''_n)_{n\in\omega} is Cauchy, there exists m\in\omega such that d_w(s_n'',s_k'')<\varepsilon for all n,k\ge m. We can assume that \lambda\preceq \check s_m\prec s_m. Thus the element \lambda has an immediate successor \hat\lambda in the tree T such that \lambda\prec\hat\lambda\preceq s_m. The maximality of \lambda ensures that \hat\lambda\notin\Lambda and hence the set \{n\in\omega:\hat \lambda\preceq \check s_n\} is finite. Then we can choose a number n\in I such that n\ge m and \hat\lambda\not\preceq\check s_n. Then s_n\wedge s_m=\check s_n=\lambda and