Correspondence between Topological and Discrete Connectivities in Hausdorff Discretization

Abstract We consider Hausdorff discretization from a metric space E to a discrete subspace D, which associates to a closed subset F of E any subset S of D minimizing the Hausdorff distance between F and S; this minimum distance, called the Hausdorff radius of F and written rH(F), is bounded by the resolution of D. We call a closed set F separated if it can be partitioned into two non-empty closed subsets F1 and F2 whose mutual distances have a strictly positive lower bound. Assuming some minimal topological properties of E and D (satisfied in ℝn and ℤn), we show that given a non-separated closed subset F of E, for any r > rH(F), every Hausdorff discretization of F is connected for the graph with edges linking pairs of points of D at distance at most 2r. When F is connected, this holds for r = rH(F), and its greatest Hausdorff discretization belongs to the partial connection generated by the traces on D of the balls of radius rH(F). However, when the closed set F is separated, the Hausdorff discretizations are disconnected whenever the resolution of D is small enough. In the particular case where E = ℝn and D = ℤn with norm-based distances, we generalize our previous results for n = 2. For a norm invariant under changes of signs of coordinates, the greatest Hausdorff discretization of a connected closed set is axially connected. For the so-called coordinate-homogeneous norms, which include the Lp norms, we give an adjacency graph for which all Hausdorff discretizations of a connected closed set are connected.


Introduction
Hausdorff discretization is a metric approach to the problem of discretizing Euclidean sets. It was introduced in [12,13] in the general setting of an arbitrary metric space (E, d) and a subspace D of E which is "discrete" in the sense that every bounded subset of D is finite (we say then that D is boundedly finite).
A Hausdorff discretization of a non-empty compact subset K of E, is a subset S of the discrete space D that minimizes the Hausdorff distance between K and any subset of D. This minimum Hausdorff distance is given by the Hausdorff radius of K: r H (K) = max x∈K d(x, D). See Figures 1 and 2.
In our framework, the role played by the grid step h in the common discretization pair E = R n , D = (hZ) n , is held by the covering radius rc: the least r > 0 such that E is covered by the union of all balls Br(p) = {q ∈ Figure 1: Here E = R 2 , D is a discrete set of irregularly spaced points, and we take the Euclidean distance d 2 . A compact set K (in grey), and its greatest Hausdorff discretization (shown with filled dots), which consists of all points of D whose distance to K is at most r H (K). Here E = R 2 , D = Z 2 , and we take the Euclidean distance d 2 . Left: A compact set K = A ∪ B ∪ C overlayed with discrete points p, q, r, s and their digital cells. Right: We show the balls of Hausdorff radius r H (K) centered about p, q, r, s; since these balls are all those that intersect K, the greatest Hausdorff discretization of K is {p, q, r, s}; now since the balls centered about p, q, s cover K, {p, q, s} is also a Hausdorff discretization of K, and there is no other one than {p, q, r, s} and {p, q, s}. E | d(p, q) ≤ r} with p ∈ D. Equivalently, rc = sup x∈E d (x, D). Then, we have r H (K) ≤ rc, so the Hausdorff distance between K and its Hausdorff discretizations is always bounded by the resolution of D.
The set M H (K) of Hausdorff discretizations of K is non-empty, finite and closed under non-empty unions. It has thus a greatest element, called the greatest Hausdorff discretization of K and written ∆ H (K). Actually, ∆ H (K) consists of all points of D whose distance to K is at most r H (K), see Figures 1 and 2 again.
Both M H (K) and ∆ H (K) were characterized in [12], and their relation with other types of discretization was analysed in [13]. For instance, we showed that when digital cells constitute the Voronoi diagram of D, every cover discretization is a Hausdorff discretization. On the other hand, in morphological discretization by dilation, provided that the dilation of D is E (every point of E belongs to the structuring element associated to a point of D), the Hausdorff distance between a compact and its discretization is bounded by the radius of the structuring element.
In [14,15] we extended our theoretical framework to the discretization of non-empty closed subsets of E instead of compact ones.
In Section 5 of [22], we analysed the preservation of connectivity by Hausdorff discretization in the case where E = R 2 , D = Z 2 and the metric d is induced by a norm N such that for (x 1 , x 2 ) ∈ R 2 and ε 1 , ε 2 = ±1, N(ε 1 x 1 , ε 2 x 2 ) = N(x 1 , x 2 ) = N(x 2 , x 1 ), for example the Lp norm (1 ≤ p ≤ ∞), see (1) next page. We showed then that for a non-empty connected closed subset F of E, 1. for N ≠ L 1 , every Hausdorff discretization of F is 8-connected; 2. the greatest Hausdorff discretization of F is 4-connected.
There was an error in [22]: we overlooked the condition N ≠ L 1 in item 1; it was later pointed out by D. Wagner (private communication), and indeed for N = L 1 we show a counterexample in Section 4 ( Figure 8).
Since Section 5 of [22] gives the starting point of the present paper, we can look closely at the proof of the above result: it depends only on the properties of closed balls. Indeed, items 1 and 2 follow from the corresponding two facts: 1. when N ≠ L 1 , for any x, y ∈ D such that Br c (x) ∩ Br c (y) ≠ ∅, x is 8-adjacent to y; 2. for any x ∈ E and r > 0, Br(x) ∩ D is 4-connected.
The goal of this paper is to extend these connectivity preservation properties of Hausdorff discretization to the most general situation: we consider the discretizations of any closed subset of a space E into a "discrete" subspace D of "bounded resolution", in other words D is boundedly finite and 0 < rc < ∞. The space E will be assumed to be "continuous" in some sense, in other words some results may require additional conditions, such as E being boundedly compact (that is, every bounded and closed subset of D is compact) and having the middle point property (for any p, q ∈ E, there is some x ∈ E such that d(p, x) = d(x, q) = 1 2 d(p, q)). Note that all conditions considered here are satisfied by R n and Z n with a norm-based metric.
This extension of connectivity preservation has been made possible by the development of a very broad framework for the notion of connectivity, namely the concept of connection [10,18] and its recent generalization to that of partial connection [11].
We will obtain several general results. Let F be a non-empty closed subset of the boundedly compact space (E, d). For any s ≥ 0, we say that F is s-separated if F can be partitioned into two non-empty closed subsets F 1 and F 2 such that d(x 1 , x 2 ) > s for all x 1 ∈ F 1 and x 2 ∈ F 2 . Then we say that F is separated if F is s-separated for some s > 0 or, equivalently, F can be partitioned into two non-empty closed subsets F 1 and F 2 such that inf{d(x 1 , x 2 ) | x 1 ∈ F 1 , x 2 ∈ F 2 } > 0. Finally, F is non-separated if it is not separated, that is, for every s > 0, F is not s-separated. If F is connected, then it is non-separated but the converse is false, see for instance Example 2 (in Subsection 3.1) and Figure 5 (in Subsection 3.2). With this definition, we get the following: -If F is connected, then (see Theorem 18 in Subsection 3.2): 1. every Hausdorff discretization of F is connected in the graph on D where we join by an edge any two points whose distance apart is at most 2r H (F) (twice the Hausdorff radius); 2. the greatest Hausdorff discretization ∆ H (F) of F belongs to the partial connection generated by all B r H (F) (x) ∩ D, x ∈ F.
-If F is disconnected but non-separated, then (see Theorem 19 in Subsection 3.2), for any r > r H (F): 1. every Hausdorff discretization of F is connected in the graph on D where we join by an edge any two points whose distance apart is at most 2r; 2. the greatest Hausdorff discretization ∆ H (F) of F belongs to the partial connection generated by all Br(x) ∩ ∆ H (F), x ∈ F.
-If F is s-separated for some s > 0, then (see Proposition 21 in Subsection 3.3), for any r ≤ s/4, every Hausdorff discretization of F will be disconnected in the graph on D where we join by an edge any two points whose distance apart is at most 2r. In particular, if rc < s/4, this will be the case for some r > r H (F).
We can interpret the above results in the framework of discretizations in multiple resolutions: the Hausdorff discretizations of a non empty closed subset F of the boundedly compact space E remain connected when the resolution of D tends to zero iff F is non-separated. See Theorem 23 in Subsection 3.3.
In the second part of our paper, we consider the particular case where E = R n , D = Z n and d is induced by a norm invariant under any change of sign or permutation of the basis vectors of R n , for example the Lp norm (1 ≤ p ≤ ∞): We will then see that every Hausdorff discretization of F will be connected according to an adjacency graph depending on the norm, while the greatest Hausdorff discretization of F will be connected for the 2n-adjacency relation, linking any two points of Z n that differ by 1 in exactly one coordinate. For n = 2, this will give the result of [22] mentioned above. The paper is organized as follows. Our notation is summarized in Table 1. Section 2 recalls the theoretical background: first partial connections and connections [11], then some special families of closed sets [14,15], and finally Hausdorff discretization [12,13]. It also discusses related works by other authors. Section 3 gives our general results about the connectivity of discretizations of connected, non-separated or separated sets. Then Section 4 considers the particular case of E = R n and D = Z n with a coordinate-homogeneous norm. Finally Section 5 concludes.
Throughout our paper, the statement of a well-known or published fact will be designated "Property", while the words "Lemma", "Proposition" and "Theorem" will be reserved for new results.

Theoretical Background
We will recall some known concepts and results, first about connections and partial connections [11], then about some families of closed subsets in a metric space [14,15], and finally in the theory of Hausdorff discretization [12,13]. We also summarize related works.
Let E be any non-empty set. We write P(E) for the family of all subsets of E. Given three sets X, Y , Z, we

Connections and Partial Connections
Connections were defined by Serra [18] and further analysed by Ronse [10], who then generalized them to partial connections [11]: We call the partial connection C a connection on P(E) if it satisfies the following third condition: 3. for all p ∈ E, {p} ∈ C.
Note that condition 2 remains valid for B = ∅. Indeed, ⋂︀ B = E ≠ ∅ and ⋃︀ B = ∅ ∈ C thanks to condition 1. Elements of a partial connection C are said to be connected; those of P(E) \ C are said to be disconnected. For X ∈ P(E), a connected component of X is any element of P(X) ∩ C \ {∅} that is maximal for inclusion. When C is a connection, the connected components of X constitute a partition of X; when C is a partial connection, they constitute a partition of the subset ⋃︀(︀ P(X) ∩ C )︀ , the union of all connected subsets of X.
The above expressions "connected set" and "connected component" must always be understood in the context of a given (partial) connection. Let us give three well-known examples of connections. In a topological space E, a subset X is connected if it cannot be partitioned by two open sets G 1 and G 2 ; equivalently, it cannot be partitioned by two closed sets F 1 and F 2 . Then the set of connected subsets of E constitutes a connection [18]. It is easily seen that the closure of a connected set is connected, hence the connected components of a closed set are closed. Another connection is made of path-connected sets, that is, sets X such that for any x, y ∈ X, there is a path joining x to y, i.e., a continuous map f : [0, 1] → X with f (0) = x and f (1) = y. This second connection contains the first one, since any path-connected set is connected. A third example is given by connectivity in a graph; we will describe it precisely after Property 2. Other examples of connections and partial connections can be found in [10,11,20]. Property 1. An intersection of connections on P(E) is a connection on P(E), an intersection of partial connections on P(E) is a partial connection on P(E), and P(E) is the greatest (partial) connection on P(E). Thus for any family B of subsets of E, there is a least partial connection including B and a least connection including B.
The least partial connection (resp., connection) including B is called the partial connection (resp., connection) generated by B and it is written Con * (B) (resp., Con(B)). Obviously ∅ (as set of points of E) is chained by ∅ (as set of non-empty subsets of E). In a non-empty subset A of E, a chain between two points has always a strictly positive length. A point x is chained to itself if x ∈ B for some B ∈ B; thus, the binary relation on E linking two points when they are chained by B, is generally not reflexive. However, it is symmetrical and transitive; hence this relation induces an equivalence relation on ⋃︀ B (not on E).

Property 2.
Given a family B of non-empty subsets of E, Con * (B) is the set of all X ∈ P(E) chained by P(X) ∩ B; thus X ∈ Con * (B) if and only if for any y, z ∈ X, there are B 0 , . . . , Bn ∈ B (n ≥ 0) such that y ∈ B 0 , z ∈ Bn, B i ⊆ X for 0 ≤ i ≤ n and B i ∩ B i+1 ≠ ∅ for 0 ≤ i < n. Now, Con(B) is obtained by adding to Con * (B) all singletons {x}, x ∈ E.
For example, in a graph with vertex set V, let B be the set of pairs of distinct vertices that are linked by an edge. Then Con(B) is the set of all parts X of V such that the graph induced on X is connected [10]. On the other hand, Con * (B) consists of such parts X of size at least 2, or singletons X = {p} with a loop on p; in other words, isolated vertices are excluded from Con * (B). Indeed, graph connectivity is based on chains of edges, which can possibly be of length 0 (isolated vertices), while Con * (B) is based on chains of length at least 1.
; in other words, Y is the set of all points in X that are covered by some B ∈ B such that B ⊆ X. Then Y ⊆ X, and the binary relation on Y linking x, y ∈ Y iff x and y are chained by P(X) ∩ B, is an equivalence relation, whose equivalence classes are the connected components of X according to the partial connection Con * (B); thus for x ∈ Y, the connected component of X containing x is the set of all y ∈ Y such that x and y are chained by P(X) ∩ B, while for x ∈ X \ Y, x belongs to no connected component; in particular, when Y = ∅, X has no connected component. Now, the connected components of X according to Con(B) are those according to Con * (B), plus the singletons of X \ Y.
Let A and B be two non-empty sets. For any map ψ : P(A) → P(B) and x ∈ A, write ψ(x) for ψ({x}). When B = A, the map ψ is said "on P(A)". A map δ : P(A) → P(B) such that for any Z ∈ P(A) we have We finally recall what is called a connection by dilation [10,11,18]: Let C be a partial connection on P(E) and let δ be a dilation on P(E) such that for any x ∈ E, δ(x) ≠ ∅. Then C δ = {X ∈ P(E) | δ(X) ∈ C} is a partial connection on P(E). Furthermore, if for any x ∈ E, x ∈ δ(x) and δ(x) ∈ C, then C δ is a connection and C ⊆ C δ .

Some families of closed sets in a metric space
The reader is assumed to be familiar with basic topological and metric notions, such as a metric (or distance function), a metric space, a bounded set, an open set, a closed set, the relative topology on a subset, a compact space and a compact set, see [6]. We recall some more advanced definitions and results [14,15]. Let (E, d) be a metric space, where E ≠ ∅ and d is a metric on E. For X ⊆ E and p ∈ E we define the distance between p and X as Note that |d(p, Note that for r = 0 we have B 0 (x) = {x} and B ∘ 0 (x) = ∅. The open balls of radius > 0 constitute the basis of the metric topology on E.
Given X ⊆ E, (X, d) is a metric space, and the metric topology of (X, d) coincides with the relative topology on X induced by the metric topology of (E, d). We will consider several topological or metric properties on a metric space (E, d) (compact, boundedly compact, boundedly finite, ...), and we will speak of such a property for a subset X of E to mean that the metric space (X, d) has that property.
Given a family S(E) (defined from some property) of subsets of the space E, we will write S ′ (E) for S(E)\{∅}, the family of non-empty elements of S(E). For any X ⊆ E, we will write S(X) (resp., S ′ (X)) for the corresponding family in the relative topology or metric of X.
Let us write: F(E) for the family of closed subsets of E, K(E) for that of compact subsets of E, and Fin(E) for that of finite subsets of E; then P ′ (E), F ′ (E), K ′ (E) and Fin ′ (E) will designate the restrictions of these families to non-empty subsets of E.
The following definition and property were given in [14,15]: A subset X of E is called proximinal if either X = ∅, or for every y / ∈ X, there is some x ∈ X minimizing the distance to y, that is, d(y, x) = d(y, X).
For instance, for the metrics induced by Lp norms, R n is boundedly compact and Z n is boundedly finite. Write F bc (E), F bf (E) and Fp(E) respectively, for the family of boundedly compact subsets of E, of boundedly finite subsets of E, and proximinal subsets of E.

Property 4.
In any metric space (E, d):

If E is boundedly finite, then every subset of E is boundedly finite, that is
We will now see the relation of these properties with dilations by balls. For r ≥ 0 we define three operators P(E) → P(E), two of which are dilations: first δr, the dilation of radius r, given by then δ ∘ r , the open dilation of radius r, given by and finally δ + r , the upper extension of radius r (which is generally not a dilation), given by We have Note that δ ∘ r (X), being a union of open balls, is open. On the other hand, δ + r (X), being the inverse image of the closed interval [0, r] by the continuous function x ↦ → d(x, X), is closed.
For r = 0 we have δ 0 (X) = X, δ ∘ 0 (X) = ∅ and δ + 0 (X) = X. Now, δr(X), δ ∘ r (X) and δ + r (X) increase with r, and they satisfy the inclusions In [14] we characterized proximinal sets and boundedly compact spaces in terms of properties of δr: In a metric space (E, d), a set X is proximinal if and only if for every r ≥ 0 we have δr(X) = δ + r (X). In particular, when X is proximinal, δr(X) is closed. Property 6. The following properties are equivalent in a metric space (E, d): 1. E is boundedly compact. 2. For every non-empty compact K and r > 0, δr(K) is compact. 3. For every non-empty closed F and r > 0, δr(F) is boundedly compact.

Hausdorff Discretization
Let us first recall from [12,14,15] some basic concepts and facts about the Hausdorff metric. We start with non-empty compact sets. We remind first that for any K ∈ K ′ (E), every continuous function K → R has a compact image, in particular it attains a maximum and a minimum.
which we call the oriented Hausdorff distance from X to Y; thus there exist x * ∈ X and y * ∈ Y such that . We define then the Hausdorff distance between X and Y as Now, H d is a metric on the space K ′ (E), it is thus called the Hausdorff metric. For X, Y ∈ K ′ (E) and r ≥ 0, we One can extend the Hausdorff metric from K ′ (E) to F ′ (E). Given two non-empty closed sets X and Y, we set now we define the Hausdorff distance H d (X, Y) as in (9). Then H d is a generalized metric on F; by this we mean that H d satisfies the axioms of a metric, with the only difference that it can take infinite values. Using (6,7), and the convention that the infimum (in R + ) of an empty set is equal to ∞, we get: In particular: As in R + the supremum and infimum of an empty set are equal to 0 and ∞ respectively, we can extend h d and H d to the empty set. We get then We can now recall the theory of Hausdorff discretization [12,14,15]. We have a metric space (E, d), and let D ⊂ E, D ≠ ∅. Here E will be the "continuous" space and D will be the "discrete" space. We define the covering radius (of D for the metric d) as the positive number The covering radius rc is in some way a measure of the resolution of D. We assume the following:  (7). Hence rc > 0 and rc is the least r > 0 such that δr(D) = E. Note that every subset of D is closed and that a subset of D is compact iff it is finite; thus, by analogy with the corresponding subsets in E, we will write F ′ (D) for the family of all non-empty subsets of D, and K ′ (D) for the family of all non-empty finite subsets of D.
Then D is boundedly finite, rc = 1/2, but E is not boundedly compact, because it has infinite dimension: covering a ball of radius 1/2 requires an infinity of balls of radius 1/4.
For any F ∈ F ′ (E), define the Hausdorff radius of F: Then r H (F) ≤ rc, and r H (F) is the least r ≥ 0 such that F ⊆ δr(D). The particular case F = E gives r H (E) = rc.
For K ∈ K ′ (E), the above supremum is attained, thus r H (K) = max x∈K d(x, D).
For r ≥ 0, the discretization of radius r is the map ∆r : P(E) → P(D) defined by setting ∀ X ⊆ E: Then ∆r is a dilation P(E) → P(D), and ∆r(X) is finite for every bounded X. For x ∈ E, we write ∆r(x) for ∆r ({x}). Next, the upper discretization of radius r is the map ∆ + r : P(E) → P(D) given by setting ∀ X ⊆ E: The last two equalities follow from (6) and (7). In general, ∆ + r is not a dilation. By Property 5, when F is proximinal, we have ∆ + r (F) = ∆r(F). We will now examine the Hausdorff distance between non-empty closed subsets of E and non-empty subsets of D, and between non-empty compacts subsets of E and non-empty finite subsets of D. Indeed, for In [12,13], we said a Hausdorff discretizing set of K. The family of Hausdorff discretizations of F is written M H (F).

Property 9. For F ∈ F ′ (E), r H (F) minimizes the Hausdorff distance between F and elements of F ′ (D)
: : every Hausdorff discretization of a compact set is finite. It is easily seen that M H (F) is closed under non-empty unions, and in [14] we showed that it is down-continuous: for a decreasing sequence (Sn) n∈N of elements of M H (K), In [12,13], we called it the maximal Hausdorff discretization of K. Following the above remark that H d (∅, ∅) = 0 and H d (F, ∅) = ∞ for F ∈ F ′ (E), we deduce that the only Hausdorff discretization of the empty set (in E) is the empty set (in D).

Related Works
Though Hausdorff discretization is a broad discretization scheme [12,13,22], several other approaches are described in the literature, primarily for the space E = R n and the discrete subspace D = Z n . Many of them were investigated from the topological point of view, specially about the connectedness preservation property. Write o for the origin of R n and Z n , and for any p ∈ Z n , let C(p) = B ∞ 1/2 (p), the ball of radius 1/2 for the L∞ norm (see Figure 3 (c) for n = 2); then, C(p) is the digital cell around p.
Let us first clarify the terminology concerning digital adjacency and connectivity. There are essentially two notations. On the one hand, from a combinatorial point of view, k-adjacency designates the one where each digital point has k neighbours, for instance the 4-and 8-adjacencies in Z 2 , then the 6-, 18-and 26adjacencies in Z 3 . On the other hand, in a coordinate-based approach, for 0 ≤ k ≤ n − 1, two points of Z n are said to be k-adjacent if they differ by 1 in at least 1 and at most n − k coordinates, the other coordinates being equal; to avoid confusion with the combinatorial notation, we will say that they are k/n-adjacent. Two particular adjacencies stand out, the axial and diametral ones. Two digital points p and q are axially adjacent if ‖p − q‖ 1 = 1, that is, p and q are (n − 1)/n-adjacent (or 2n-adjacent in the combinatorial notation). Now p and q are diametrally adjacent if ‖p − q‖∞ = 1, that is, p and q are 0/n-adjacent (or (2 n − 1)-adjacent in the combinatorial notation). The graph-theoretical connectivity corresponding to the k-adjacency is called k-connectivity.
In discretization by dilation [12], one chooses a structuring element A ⊂ R n , then the discretization of a subset X of R n is ∆ A , one obtains the supercover discretization, made of all digital points whose cell intersects the Euclidean set. It is well-known that the supercover discretization of a connected subset of R n is axially connected.
For n = 2, another example is the Freeman/Bresenham grid-intersect discretization, which uses the unit cross structuring element of Figure 3 (a). Then the discretization of a connected set will be diametrally connected. Sekiya and Sugimoto [16] considered the discretization by dilation of connected curves in R 2 , using the two structuring elements of Figure 3 (b) and (c); the latter gives the supercover, which is axially connected. On the other hand the former (b) gives a discretization which is intermediate between the grid-intersect and the supercover; they showed that the discretization of a path-connected curve will be diametrally connected, and it is easy to check that the result also holds for any connected subset of R 2 . Note that both structuring elements (a) and (b) are not covering [12] (i.e., the union of translates of the structuring element by points of D does not cover E), so a non-empty subset of R 2 can have an empty discretization. , the unit cell or ball of radius 1/2 for the L∞ norm. Bottom: the discretization by dilation of a curve, for each of the 3 structuring elements, will be diametrally connected; for the supercover discretization (f), it will be axially connected.
Given the high computational cost of discretization by dilation, Sekiya and Sugimoto suggest an approximation that leads to a slightly smaller discretization. Assuming that a curve C is given by an analytical equation f (x) = 0 (x ∈ R 2 ), where f is a continuous function R 2 → R, the approximated discretization will consist of all digital points p such that Ap has two corners x, y ∈ R 2 with f (x) > 0 and f (y) < 0; by continuity, there will be some z ∈ Ap with f (z) = 0, so this discretization is included in the one by dilation by A. They showed that with the structuring elements of Figure 3 (b) and (c), the approximated discretization will still be diametrally connected for (b) and axially connected for (c).
A particular case of discretization by dilation is the Gauss discretization, where the structuring element is reduced to the singleton made of the origin: A = {o}; in other words, a set X ⊂ R n is discretized as X ∩ Z n . Pavlidis [9] and Serra [17] give local geometric conditions (see Figure 4) on the boundary points of X ⊂ R 2 that ensure a homeomorphic reconstruction of X from its discretization (the set X is reconstructed as the union of the unit squares centerered in X ∩ Z 2 ). In particular, under these conditions, the discretization of a subset X of R 2 yields an axially connected graph iff X is connected. This result was later extended by Gross et al. in [5,8] to threshold area discretizations where an integer point p of Z 2 belongs to the dicretization set if the area of the intersection between X and the unit square centered in p is above some fixed threshold. The minimal cover of a set X ⊂ R n was studied by Brimkov et al. in [3]. It is the smallest union of unit cells of R n , centered on integer points, that includes X. The integer points of the minimal cover of a simply connected n-dimensional manifold of R n form an axially connected graph [3]. When X is an h-dimensional surface (the image of R h into R n by a continous function), another kind of minimal cover, the k-discretization (where 0 ≤ k ≤ n − 1), is described in the same article. It is a subset of the supercover that is minimal in the sense that the deletion of any of its unit cells will change the topology of the k/n-adjacency graph. Moreover, it is required that this cover includes some h-dimensional surface, not necessarily X itself. In the paper, there is another requirement but it not used in the proof of the following theorem: this k-discretization is (h − 1)/nconnected.
Later, Brimkov et al. [4] showed that the discretization by dilation by an Euclidean ball of radius r (there called r-offset discretization) of a bounded path-connected subset of R n is axially connected if r ≥ √ n/2 and is diametrally connected if r ≥ √ n − 1/2. Furthermore, the values √ n/2 and √ n − 1/2 are minimal for such a statement. This result is first extended to connected sets in [2] provided, for the first claim, that r is greater than or equal to the Hausdorff radius of the set. Then, in [1] it is extended to disconnected sets whose closure is connected provided the conditions on the radius r are strict inequalities.
Incidentally, Brimkov et al. [2] obtain the following result about Hausdorff discretizations (for the Euclidean distance): for any connected subset of R n , its greatest Hausdorff discretization is axially connected, while any other Hausdorff discretization is diametrally connected if the Hausdorff radius is less than 1 and may be diametrally disconnected otherwise.
In [21], the authors consider the Euclidean space E = R 2 , the discrete space D(ρ) = ρZ 2 of grid step ρ > 0, and a metric based on a norm N such that N(±x 1 , ±x 2 ) = N(±x 2 , ±x 1 ) = N(x 1 , x 2 ) and the only point at distance N(1/2, 1/2) from both (0, 0) and (1, 1) is (1/2, 1/2) (this excludes in particular the L 1 norm). For X ⊆ D(ρ), let Rρ(X) be its Euclidean reconstruction given by the union of closed digital cells (relatively to D(ρ)) of points of X: Rρ(X) = ⋃︀ p∈X B ∞ ρ/2 (p). For r > 0, we say that F is r-convex if for any point x / ∈ F at distance at most r from F, this distance is attained by a unique point in F, and for any r ′ ≤ r, the closed ball of radius r ′ centered about x has a connected intersection with F.
Given a non-empty closed subset F of E and a Hausdorff discretization M ρ of F in D(ρ), several topological relations between F and the reconstruction Rρ(M ρ ) are announced (without proof), which hold when ρ is sufficiently small. First, any r-convex non-empty closed subset F of E is homotopically equivalent to Rρ(M ρ ) for any Hausdorff discretization M ρ of F. Next, when both F and the closure of its complement are r-convex and form a union of closed balls of radius r, a Hausdorff discretization M ρ of F will not contain a singular configuration [7] (a 2 × 2 digital square with one diagonal in the figure and the other in the complement), and Rρ(M ρ ) will be a bordered 2D manifold (that is, locally homeomorphic to a closed half-plane); in particular, if F is a compact bordered 2D manifold, then F and Rρ(M ρ ) will be homeomorphic.

General Results
Our goal is to show that every Hausdorff discretization of a non-empty "connected" closed subset of E is "connected", and that conversely every Hausdorff discretization of a non-empty "disconnected" set will be "disconnected" when the resolution of the discrete space D is small enough. Here the connectivity in D will be in the graph with edges linking points distant by at most twice the Hausdorff radius. The "connectedness" of a closed subset F of E is more complicated. When F is compact, it is ordinary topological connectivity. However, a disconnected non-compact closed set can have a connected discretization when its connected components are asymptotic to each other (cf. Example 2). We will thus introduce the notion of a s-separated closed set, which is partitioned into two closed sets with mutual distances always > s. Then a non-empty closed set F will have its Hausdorff discretizations connected at all resolutions if and only if F is not s-separated for any s > 0. Such sets, called non-separated, constitute a connection on F(E), which contains topologically connected sets.
Our method of proof is similar to the one used in Section 5 of [22] to show the reciprocal link between discrete and continuous connectivities for compact sets, in the particular case where E = R 2 , D = Z 2 and the distance is based on a homogeneous norm. Subsection 3.1 introduces some needed notions: the middle point and interval properties in a metric space, and s-separated closed sets. Subsection 3.2 shows how, under some general conditions, the Hausdorff discretization of a non-empty non-separated closed set F gives a discrete set which is connected in the graph with edges linking points of D that are "close enough" (see Definition 7). Conversely, Subsection 3.3 shows how this fails for s-separated closed sets; in particular it considers discretization in a sequence of discrete spaces whose resolution tends to zero, and conditions under which the discretizations are connected at all resolutions.

Further Continuity Properties in a Metric Space
Let (E, d) be a metric space. For r, s ≥ 0, the triangular inequality implies that for any X ∈ P(E), δr(δs(X)) ⊆ δr+s(X). The equality holds if for any p, q ∈ E such that d(p, q) = r + s, there exists x ∈ E such that d(p, x) = r and d(x, q) = s. We will in fact consider two conditions, a first one weaker and a second one stronger; they will be equivalent in a boundedly compact space.
Every metric induced by a norm on a vector space satisfies the interval property. When d has the interval property, we obtain δr(δs(X)) = δr+s(X) for any r, s ≥ 0 and X ∈ P(E). Proof. 1. Assume the middle point property. Let U = { k 2 n | k, n ∈ N, 0 ≤ k ≤ 2 n }. We construct f : U → E by induction on the exponent n. For n = 0, f (0) = p and f (1) = q; for n ≥ 0 and 0 ≤ k ≤ 2 n − 1, given (p, q). Let α, β ∈ U such that α < β; then there are a, b, n ∈ N with 0 ≤ a < b ≤ 2 n such that α = a 2 n and β = b 2 n . We get then: In particular, f (α) ∈ Kp,q for all α ∈ U. Since f is a continuous function U → Kp,q and the compact metric space Kp,q is complete, there is a continuous extension of f to a function [0, 1] → Kp,q. More precisely, given α ∈ [0, 1], for any n ∈ N we set αn = ⌊2 n α⌋/2 n ; then αn ≤ α < αn + 1 2 n , so limn→∞ αn = α; then the f (αn), n ∈ N constitute a Cauchy sequence in Kp,q, which converges in Kp,q to some point that we define as f (α). For 0 ≤ α < β ≤ 1, by continuity we have Therefore (E, d) has the interval property. Proof. For p, q ∈ E, let r = d(p, q); then for any x ∈ E with d(p, x) + d(x, q) = d(p, q) = r, we have x ∈ Br(p), which is compact; we can thus apply Lemma 10 with Kp,q = Br(p).
The connectedness of closed balls is useful in relation to Property 3: if all Br(p) are connected, then for a connected set X, δr(X) will be connected.

Definition 5.
Let F ∈ F ′ (E) and s ≥ 0. We says that F is s-separated if F can be partitioned into F 1 , F 2 ∈ F ′ (E) such that for any x 1 ∈ F 1 and x 2 ∈ F 2 , d(x 1 , x 2 ) > s. We say that F is separated if for some s > 0, F is s-separated, and that F is non-separated if for every s > 0, F is not s-separated.
Note that F is separated iff it can be partitioned into two non-empty closed subsets F 1 and F 2 such that inf{d(x 1 , x 2 ) | x 1 ∈ F 1 , x 2 ∈ F 2 } > 0. For s > s ′ ≥ 0, if F is s-separated, then it is s ′ -separated. Moreover, F is 0-separated iff it is disconnected, that is, partitioned into two non-empty closed sets. Thus a connected set is not 0-separated, hence it is non-separated.

Lemma 12. A non-empty compact set K is non-separated if and only if it is connected.
Proof. If K is connected, it is not 0-separated, thus it is not s-separated for any s > 0. If K is disconnected, then it is partitioned by two closed sets F 1 and F 2 , so it is partitioned into two compact sets K 1 and K 2 . By the compactness of K 1 , K 2 and the continuity of the distance function, the set of all d(x 1 , x 2 ) for x 1 ∈ K 1 and x 2 ∈ K 2 reaches a minimum a; since K 1 and K 2 are disjoint, a > 0, so K is s-separated for any s > 0 such that s < a.

Example 2.
A non-compact closed set can be both disconnected (0-separated) and non-separated. For instance (cf. Section 5 of [22]), in R 2 with the Euclidean distance, the closed set F = , tending to 0 for x → 0. Note that δr(F) is connected for any r > 0.

Proposition 13. Let (E, d) be boundedly compact and satisfying the middle point property. Then for any F ∈ F ′ (E) and s > 0, F is s-separated if and only if δ s/2 (F) is disconnected. In particular, F is non-separated if and only if for any r > 0, δr(F) is connected.
Proof. Every closed set X is proximinal by Property 4, so δ s/2 (X) is closed by Property 5.
If δ s/2 (F) is disconnected, it is partitioned into A 1 , A 2 ∈ F ′ (E). As F ⊆ δ s/2 (F), F is the disjoint union of the two closed sets F 1 = A 1 ∩ F and F 2 = A 1 ∩ F. For i = 1, 2 and x i ∈ F i , B s/2 (x i ) is connected by Corollary 11, so it cannot be partitioned by A 1 and A 2 ; we deduce that B s/2 (x i ) ⊆ A i . If we had F = F i , we would get B s/2 (x) ⊆ A i for all x ∈ F, so δ s/2 (F) ⊆ A i , a contradiction; therefore {F 1 , F 2 } is a partition of F. If we had x 1 ∈ F 1 and x 2 ∈ F 2 with d(x 1 , x 2 ) ≤ s, then by the middle point property there would be p ∈ E such that d(x 1 , p) = d(x 2 , p) = d(x 1 , x 2 )/2 ≤ s/2, so p ∈ B s/2 (x 1 ) ∩ B s/2 (x 2 ); but B s/2 (x i ) ⊆ A i (i = 1, 2), so we would have p ∈ A 1 ∩ A 2 , a contradiction. Therefore for all x 1 ∈ F 1 and x 2 ∈ F 2 we have d(x 1 , x 2 ) > s, thus F is s-separated.
Now F is non-separated iff for all r > 0 it is not 2r-separated, that is, δr(F) is connected for any r > 0.
Serra [19] generalized connections from P(E) to complete lattices. Now F(E) is a complete lattice, where the infimum and supremum of a family are respectively its intersection and the closure of its union, and it comprises all singletons (as closed balls of radius 0); thus Serra's definition takes here the following form: The partial connection C is a connection on F(E) if it satisfies the following third condition: 3. for all p ∈ E, {p} ∈ C.
Then Property 1 is also valid with F(E) instead of P(E): an intersection of connections (resp., partial connections) on F(E) is a connection (resp., partial connection) on F(E).

Proposition 14. For any s ≥ 0, the family of closed sets that are not s-separated constitutes a connection on F(E). The family of non-separated closed sets is also a connection on F(E).
Proof. Obviously the empty set and the singletons are closed and cannot be partitioned by two sets, so they are not s-separated. Let B be a family of closed sets that are not s-separated, such that ⋂︀ B ≠ ∅; let F = ⋃︀ B and let z ∈ ⋂︀ B. Suppose that F is s-separated: F is partitioned into F 1 , F 2 ∈ F ′ (E) such that for any x 1 ∈ F 1 and x 2 ∈ F 2 , d(x 1 , x 2 ) > s. As z ∈ F, for some i = 1, 2 we have z ∈ F i . For any B ∈ B, as z ∈ B, B ∩ F i ≠ ∅; as B is not s-separated, it cannot be partitioned by F 1 and F 2 , so we must have B ⊆ F i . We deduce that ⋃︀ B ⊆ F i , and as F i is closed, F = ⋃︀ B ⊆ F i , which contradicts the partitioning of F. Therefore F is not s-separated. The family of non-separated closed sets is the intersection, for all s > 0, of the families of closed sets that are not s-separated. It is thus an intersection of connections, hence it is a connection on F(E).

Discretizing Non-Separated Sets
Let (E, d) be a metric space and D ⊂ E, D ≠ ∅, satisfying Axiom 1.

Definition 7.
Let r > 0, S ∈ F ′ (D) and Z ∈ P ′ (E). Then write Gr(S) and G Z r (S) for the two graphs both with vertex set S, and with an edge joining any two distinct p, q ∈ S such that respectively: Note that the all edges of G Z r (S) are edges of Gr(S), and that when r grows, the sets of edges of Gr(S) and G Z r (S) will be growing. Thus connectedness extends from G Z r (S) to Gr(S), and for r ′ > r, from Gr(S) to G r ′ (S) and from G Z r (S) to G Z r ′ (S). For p, q ∈ E, Br(p) ∩ Br(q) ≠ ∅ implies that d(p, q) ≤ 2r. If (E, d) has the middle point property, then for p, q ∈ E, we have Br(p) ∩ Br(q) ≠ ∅ ⇔ d(p, q) ≤ 2r, which simplifies the definition of Gr(S). This fact intervenes in the following result, which will be used below.

Then, (A) implies (B) and (B) implies (C).
Proof. Let X = {d(p, q) | p, q ∈ D}. If (A) holds, then for p, q ∈ D, d(p, q) = N(p − q), so X = {N(x) | x ∈ D}. Given the origin o in E, for any r > 0, Hence X is boundedly finite. Now for any p, q ∈ D, x = 1 2 (p + q) satisfies Hence Gr(S) has the same edges as G R (S), so the two graphs are equal. Therefore (C) holds.
We give below a relation between graph connectivity and chaining by sets, which will be used in our main results:

Proposition 16. Let A be a non-empty subset of E and let B be a non-empty family of non-empty subsets of A. Let G be the undirected graph with vertex set A and with edges joining all pairs {x, y} of distinct elements of A such that there is B ∈ B with x, y ∈ B. Then G is connected if and only if A is chained by B.
Proof. If |A| = 1, we have A = {p}, G is connected, and we must have B = {A}, so A is chained by B. Suppose now that |A| ≥ 2 and consider two distinct p, q ∈ A. Now, p and q are chained in B iff there are n ≥ 0 and B 0 , . . . , Bn ∈ B such p ∈ B 0 , q ∈ Bn and for t = 1, . . . , n, B t−1 ∩ B t ≠ ∅. Writing p = u 0 and q = u n+1 , and choosing any u t ∈ B t−1 ∩ B t (t = 1, . . . , n), the statement becomes: there are the two sequences p = u 0 , . . . , u n+1 = q in A and B 0 , . . . , Bn in B such that for  t = 0, . . . , n, we have u t , u t+1 ∈ B t . In other words, we have the sequence p = u 0 , . . . , u n+1 = q in A such that for t = 0, . . . , n, u t and u t+1 are joined by an edge. Hence it is equivalent to: p and q are joined by a chain of edges of G.
We have shown that two distinct vertices p, q ∈ A are joined by a chain of edges in G iff they are chained in B. Assuming this property, p is chained to q and q is chained back to p (both for edges and for B), so p is chained to itself.
In the above proof, we had to distinguish the case where |A| = 1 and the chaining of a point p to itself. Indeed, as said in Subsection 2.1, Con * (B), the partial connection generated by B, is obtained through chaining by B, with chains of length at least 1, and excluding isolated points (that is, not belonging to any element of B); on the other hand, in a graph we take the connection generated by edges, based on chains of edges of length at least 0, including thus isolated vertices (that is, not incident to any edge).
We can now analyse the connectedness of Hausdorff discretizations of non-separated closed sets:
Consider a partition S into S 1 and S 2 . Now Y ⊆ δr+s(S) = δr+s(S 1 ∪ S 2 ) = δr+s(S 1 ) ∪ δr+s(S 2 ). For i = 1, 2, every p ∈ S i satisfies Br+s(p)∩X ≠ ∅, hence δr+s(S i )∩X ≠ ∅; as X ⊆ Y, δr+s(S i )∩Y ≠ ∅. Next, S i is proximinal by Property 4, so δr+s(S i ) is closed by Property 5. Thus Y is the union of the two non-empty closed sets δr+s(S i )∩Y for i = 1, 2. If d(y 1 , y 2 ) > t for all y i ∈ δr+s(S i )∩Y (i = 1, 2), then Y is t-separated, a contradiction. Hence there are y i ∈ δr+s(S i ) ∩ Y (i = 1, 2) such that d(y 1 , y 2 ) ≤ t; now y i ∈ Br+s(p i ) for some p i ∈ S i ; then y 2 ∈ B r+s+t (p 1 ), so y 2 ∈ B r+s+t (p 1 ) ∩ B r+s+t (p 2 ) ∩ Y; in particular, y 2 ∈ B r+s+t (p 1 ) ∩ B r+s+t (p 2 ). As y 2 ∈ Y ⊆ δs(X), there is some x ∈ X with d(x, y 2 ) ≤ s, so x ∈ B r+2s+t (p 1 ) ∩ B r+2s+t (p 2 ) ∩ X. Thus for any partition {S 1 , S 2 } of S, there are p 1 ∈ S 1 and p 2 ∈ S 2 joined by an edge in each of the graphs G Y r+s+t (S), G X r+2s+t (S) and G r+s+t (S), therefore these three graphs are connected.
For two distinct p, q ∈ S, p and q are joined by an edge of , in other words p, q ∈ B r+s+t (y). Hence: two distinct p, q ∈ S are joined by an edge of G Y r+s+t (S) iff there is some y ∈ Y such that p, q ∈ B r+s+t (y) ∩ S. As B r+s+t (y) ∩ S ≠ ∅ for all y ∈ Y, we can apply Proposition 16: S is chained by {B r+s+t (y) ∩ S | y ∈ Y}. The same argument with G X r+2s+t (S) gives that S is chained by Our first application of this result is for the Hausdorff discretization of non-empty connected proximinal sets: Theorem 18. Let F ∈ F ′ p (E) be connected and let r ≥ r H (F). Then:

In particular, for any S ∈ M H (F), G r H (F) (S) and G F r H (F) (S) are both connected, and ∆ H (F) is chained by {∆
Proof. We apply Lemma 17 with s = t = 0 and X = Y = F, and combine it with Property 8, where ∆ + r (F) = ∆r(F) because F is proximinal. Then G F r (S) and Gr(S) are connected and S is chained by {Br(x) ∩ S | x ∈ F}. Thus item 1 holds. Now F ⊆ δr(∆r(F)), so we apply item 1 with S = ∆r(F). For x ∈ F, Br(x) ⊆ δr(F), so (13) gives We get thus item 2. By Property 9, for any S ∈ M H (F), S ⊆ ∆ + r H (F) (F) and F ⊆ δ r H (F) (S); since ∆ H (F) = ∆ + r H (F) (F) = ∆ r H (F) (F), the last sentence of the above statement follows.
In view of Property 2, we can write: For any r > 0, the set of all S ⊆ D such that Gr(S) is connected, that is, such that S spans a connected subgraph of Gr(D), is a connection. Thus, all Hausdorff discretizations of all connected proximinal subsets of E (including ∅), belong to the connection of all connected subsets of the graph Gr c (D). We now consider the discretization of non-separated sets. The result will be slightly weaker than for connected ones: Item 2 follows.

Remark 1. For an arbitrary S ∈ M H (F), the above result is optimal: in general we cannot get connectivity for a graph G F r (S) with r < r H (F). As we will see in Section 4 (see in particular
has two connected components, S 1 and S 2 , and the same connected components are obtained in the chaining by the ∆ r H (F) (x) ∩ ∆ H (F) for x ∈ F. For any s > 0 and y ∈ N large enough (in absolute value) to have 2 −|y| ≤ s, the distance between each (︀ y, 1 2 ± 2 −|y| )︀ and each of (y, 0) and (y, 1) is ≤ 1 2 +2 −|y| ≤ r H (F)+ s, so (y, 0) and (y, 1) will be joined together by an edge of G F r H (F)+s (∆ H (F)),

Discretizing Separated Sets in Multiple Resolutions
We will now give in some way the converse of the results of the previous subsection, namely that every Hausdorff discretization of a s-separated closed set (s > 0) is disconnected (for the graphs in Definition 7) when the Hausdforff radius rc is small compared to s. Next, we will consider discretization at varying resolutions, and we will see that a non-empty closed set has its Hausdorff discretizations connected at all resolutions if and only if it is non-separated.

If G F r (S)
is connected for some r > r H (F), then s < 2r; if it is for all r > r H (F), then s ≤ 2rc.

If Gr(S) is connected for some r > r H (F), then s < 4r; if it is for all r > r H (F), then s ≤ 4rc.
Proof. Take any r > r H (F). By Properties 8 and 9, F ⊆ δ r H (F) (S) ⊆ δr(S) and If one of the graphs G F r (S) or Gr(S) is connected, then S 1 ∩ S 2 ≠ ∅ or that connected graph has an edge with extremities p 1 ∈ S 1 and p 2 ∈ S 2 .
If S 1 ∩ S 2 ≠ ∅, then for p ∈ S 1 ∩ S 2 and i = 1, 2 we have some x i ∈ Br(p) ∩ F i , so d(p, x i ) ≤ r. Then If G F r (S) is connected, then it has an edge with extremities p i ∈ S i (i = 1, 2), so Br(p 1 ) ∩ Br(p 2 ) ∩ F ≠ ∅; now for x ∈ Br(p 1 ) ∩ Br(p 2 ) ∩ F, either x ∈ F 1 and then p 2 ∈ S 1 , or x ∈ F 2 and then p 1 ∈ S 2 . Thus one of p 1 , p 2 belongs to both S 1 and S 2 , hence S 1 ∩ S 2 ≠ ∅, and s < 2r. If this inequality holds for any r > r H (F), we get then s ≤ 2r H (F) ≤ 2rc. Thus item 1 holds.
We will now consider Hausdorff discretization at multiple resolutions. We suppose a set R of arbitrarily small "grid steps" ρ > 0, and for each ρ ∈ R, a subset D(ρ) of E, where ∅ ≠ D(ρ) ≠ E, for which we define the covering radius for ρ: We extend then Axiom 1 as follows: There is a set R ⊂ {r ∈ R | r > 0} such that inf R = 0 and: Then, given ρ ∈ R, for every F ∈ F ′ (E), the definitions and notations given for D in Subsection 2.3 extend to D(ρ): the Hausdorff radius r H (F, ρ), the discretization and upper discretization of radius r ≥ 0, ∆r(F, ρ) and ∆ + r (F, ρ), the family M H (F, ρ) of Hausdorff discretizations of F, and the greatest Hausdorff discretization of F, ∆ H (F, ρ).
From now on, we will assume that (E, d) is boundedly compact and it has the middle point property. Then (E, d) has the interval property, every closed ball Br(p) is connected (Corollary 11), every closed set F is proximinal (Property 4), then for all r > 0, δr(F) = δ + r (F), which is closed (Property 5), and for any s > 0, F is s-separated if and only if δ s/2 (F) is disconnected (Proposition 13).
We first show that these assumptions guarantee the existence of D(ρ), ρ ∈ R, satisfying Axiom 2. We define the diameter of E as diam(E) = sup{d(x, y) | x, y ∈ E}.
Proof. For 0 < ρ < diam(E) there exist p, q ∈ E such that d(p, q) > ρ. By the interval property, there exists p 0 ∈ E such that d(p, p 0 ) = ρ. Choose σ ∈ R such that 0 < σ < ρ. For every n ∈ N we define the two sets Then p 0 ∈ C 0 and Cn ⊆ Rn for all n ∈ N; hence R 0 ≠ ∅. For any x ∈ E, either d(p, x) ≤ ρ and x ∈ Bρ(p), or d(p, x) > ρ and there is some n ∈ N such that nσ < d(p, x) − ρ ≤ (n + 1)σ, hence x ∈ Rn. Thus E = Bρ(p) ∪ ⋃︀ n∈N Rn. Since Rn is the inverse image by the continuous function d(p, ·) of the closed interval [ρ + nσ, ρ +(n +1)σ], it is closed; now Rn is bounded, and as E is boundedly compact, Rn is compact. For any x ∈ Rn, by the interval property there exists y ∈ E such that d(p, y) = ρ + nσ and d(y, x) = d(p, x) − d(p, y) ≤ σ < ρ; in other words, y ∈ Cn and x ∈ B ∘ ρ (y). Hence the union of open balls B ∘ ρ (y), y ∈ Cn, covers Rn, and as Rn is compact, there is a finite subset Gn of Cn such that the union of B ∘ ρ (y), y ∈ Gn, covers Rn; hence Rn ⊆ δρ(Gn). Since R 0 ≠ ∅, G 0 ≠ ∅.
We can now summarize Theorems 18 and 19 and Proposition 21 in the multi-resolution framework of Axiom 2:  (F, ρ), the two graphs G F r (S) and Gr(S) are connected.
3. If F is s-separated for some s > 0, then there is some ρ 0 ∈ R such that for every ρ ∈ R with ρ ≤ ρ 0 , there is some r > r H (F, ρ) such that for every S ∈ M H (F, ρ) and every r ′ ≤ r, the two graphs G F r ′ (S) and G r ′ (S) are disconnected.
Let us now briefly consider the convergence to the original closed set of the discretization in resolution ρ when ρ tends to 0. If for every ρ ∈ R we choose some Sρ ∈ M H (F, ρ), as H d (F, Sρ) = r H (F, ρ) ≤ rc(ρ), then we will have lim Moreover, the argument of Proposition 21 can be extended to show that if we take a constant c > 1 and for every ρ ∈ R we choose some Sρ ⊆ D(ρ) (not necessarily a discretization of F) such that G c·rc(ρ) (Sρ) is connected, if we have lim ρ→0 H d (F, Sρ) = 0, then F must be non-separated.

Coordinate-Homogeneous Norms in R n and Z n
From now on, we assume that E = R n , D = Z n and the metric d is induced by a norm N: for x, y ∈ E, d(x, y) = N(x − y). Write o for the origin (0, . . . , 0) of R n . Then B 1 (o) is a symmetrical convex compact subset of E with non-empty interior, and for x ∈ E and r > 0, we have Br(x) = {x + ry | y ∈ B 1 (o)}. In particular, the metric d is topologically equivalent to the Euclidean metric d 2 , in other words, there exist β > α > 0 such that for any x, y ∈ E, αd 2 (x, y) ≤ d(x, y) ≤ βd 2 (x, y).
Clearly, (E, d) is boundedly compact and satisfies the interval property, see Definition 4: in item 2 we take . Moreover, D is boundedly finite. Note also that condition (A) of Lemma 15 is satisfied.
Everything that we will say here can easily be extended to the case where D = ρZ n for a resolution ρ > 0; in fact, for S ∈ F ′ (ρZ n ) and For p = (p 1 , . . . , pn) ∈ D, let C(p) be the square/cubic/hypercubic unit cell centered about p, that is, is the set of all x = (x 1 , . . . , xn) ∈ E such that ‖x − p‖∞ ≤ 1/2, that is, |x i − p i | ≤ 1/2 for each i = 1, . . . , n. For any x ∈ E, there is some p ∈ D such that x ∈ C(p); we have then d(x, D) ≤ d(x, p) = N(x − p), with x − p ∈ C(o), from which we deduce that rc ≤ sup x∈C(o) N(x). Therefore Axiom 1 is satisfied.
We say that the norm N is coordinate-symmetrical [13]  It follows then that a coordinate-symmetrical norm N satisfies We refer the reader to the beginning of Subsection 2.4 for the terminology on digital adjacency and connectivity, in particular the axial adjacency and connectivity.

Proposition 25. If N is coordinate-symmetrical, then for any x ∈ E and r > 0, Br(x) ∩ D is axially connected.
Proof. Write x = (x 1 , . . . , xn). Let p = (p 1 , . . . , pn) and q = (q 1 , . . . , qn) two points of Br(x) ∩ D. We show that they are joined in Br(x) ∩ D by a path for the axial adjacency. We use induction on the number k of coordinates on which p and q differ, that is, the number of i ∈ {1, . . . , n} such that p i ≠ q i . For k = 0, p = q and the result is obvious. Suppose now that the result is true for k, and let p and q differ on k + 1 coordinates. Take i such that p i ≠ q i ; without loss of generality, we can assume that |p i − x i | ≤ |q i − x i |, otherwise we exchange p and q in the following argument. Consider the sequence of integers m between p i and q i : For each such m, let z(m) = (q 1 , . . . , q i−1 , m, q i+1 , . . . , qn) be the point in D whose i-th coordinate is m, and whose j-th coordinate is q j for j ≠ i; by Property 24, Hence the z(m) constitute a path for the axial adjacency joining z(p i ) to z(q i ) = q inside Br(x) ∩ D; on the other hand, z(p i ) and p differ in k coordinates, so by induction hypothesis they are joined inside Br(x) ∩ D by a path for the axial adjacency. It follows thus that p and q are joined by such a path, and the result is thus satisfied for k + 1.
Combining this result with Theorem 18 (item 2), we deduce the following: Corollary 26. Let N be coordinate-symmetrical, let F ∈ F ′ (E) be connected and let r ≥ r H (F). Then ∆r(F) is axially connected. In particular, ∆ H (F) is axially connected.
Indeed, F will be chained by the axially connected sets ∆r(x) = Br(x) ∩ D for x ∈ F. For n = 2 (E = R 2 and D = Z 2 ) we obtain the result of [22]: for a connected F ∈ F ′ (E), the greatest Hausdorff discretization ∆ H (F) is 4-connected.
In other words there is a group of permutations of the coordinates, that acts transitively on them, under which the norm N is invariant. For example a coordinate-symmetrical norm N satisfying N(x 1 , . . . , xn) = N(x 2 , . . . , xn , x 1 ) is coordinate-homogeneous, because the permutation i ↦ → i + 1 (i < n), n ↦ → 1 generates the cyclic permutation group on {1, . . . , n}.
In [22] we said that the norm N is homogeneous if it is coordinate-symmetrical and invariant under any permutation of coordinates; this corresponds to the case where the group G is the symmetrical group consisting of all permutations of {1, . . . , n}. For instance, the Lp norm (1 ≤ p ≤ ∞) is homogeneous, so it is coordinate-homogeneous.
In [13] we showed that a coordinate-homogeneous norm N satisfies the following counterpart of (16): Since ‖(1, . . . , 1)‖ 1 = n and ‖(1, . . . , 1)‖∞ = 1, while the covering radius is given by the norm of ( 1 2 , . . . , 1 2 )︀ , we get: Property 27. [13] If N is coordinate-homogeneous, then for any x ∈ E, Then for any x ∈ E and r > 0, in particular for r = rc, In other words, the closed ball of radius rc for d is intermediate between the one of radius rc[∞] = 1 2 for the L∞ norm and the one of radius rc [1] = n 2 for the L 1 norm. We illustrate this in Figure 6 (a) for n = 2 and the Euclidean distance (based on the L 2 norm).
We see then in Figure 6 (b) that for n = 2, the midpoint x between two axially adjacent pixels p, q ∈ D must satisfy ∆r c (x) = {p, q}. This means that the discrete traces ∆ r H (F) (x) of balls B r H (F) (x) will not necessarily be thick, so in view of Corollary 26, for a connected F ∈ F ′ (E), ∆ H (F) will not necessarily be thick.
In the case of the L 1 norm, since r H (F) ≤ rc [1] = n 2 , Corollary 20 gives the following: Corollary 28. Let N be the L 1 norm. Let F ∈ F ′ (E) be non-separated and let r ≥ r H (F). Then any Hausdorff discretization of F is connected for the graph with vertex set D and with an edge linking any two distinct p, q ∈ D such that ‖q − p‖ 1 ≤ 2r. In particular, it is connected for the graph with r = rc [1] = n 2 , that is, with an edge linking two distinct p, q ∈ D when ‖q − p‖ 1 ≤ n.
In order to deal with the case when the norm N is not proportional to the L 1 norm, we need the following:  We can now state the counterpart of Corollary 28 in the case of a coordinate-homogeneous norm that is not proportional to the L 1 norm: Proposition 30. Let N be a coordinate-homogeneous norm that is not proportional to the L 1 norm. Let F ∈ F ′ (E) be non-separated and let r ≥ r H (F). Then any Hausdorff discretization of F is connected for the graph with vertex set D and with an edge linking any two distinct p, q ∈ D such that either p and q differ in at least two coordinates and ‖q − p‖ 1 ≤ nr rc , or p and q differ in exactly one coordinate and ‖q − p‖ 1 < nr rc . In particular, it is connected for the graph with r = rc, that is, with an edge linking two distinct p, q ∈ D when either p and q differ in at least two coordinates and ‖q − p‖ 1 ≤ n, or p and q differ in exactly one coordinate and ‖q − p‖ 1 < n. N(1, . . . , 1) = 2rc. Take two distinct p, q ∈ D such that Br(p)∩ Br(q) ≠ ∅; then N(q − p) ≤ 2r. Suppose first that p and q differ in at least two coordinates; by (18) we have

Proof. Recall that
so p and q are joined by an edge in the graph. Suppose next that p and q differ in exactly one coordinate; thus there is some j ∈ {1, . . . , n} and some x ∈ R such that q − p = xe j . By Lemma 29, we have N(e j ) > N(1,...,1) thus ‖q − p‖ 1 < nr rc , so p and q are joined by an edge in the graph. Hence any two distinct p, q ∈ D such that Br(p) ∩ Br(q) ≠ ∅ are joined by an edge in the graph; thus by Corollary 20 S is connected in that graph. Since r H (F) ≤ rc, we have the result with r = rc.
Note that for p, q ∈ D we have ‖q − p‖∞ < ‖q − p‖ 1 when p and q differ in at least two coordinates, but ‖q − p‖∞ = ‖q − p‖ 1 when p and q differ in exactly one coordinate. Thus the condition "either p and q differ in at least two coordinates and ‖q − p‖ 1 ≤ nr rc , or p and q differ in exactly one coordinate and ‖q − p‖ 1 < nr rc " can be written as the conjunction:  Let us apply the results of this section to the case where n = 2. We assume a coordinate-homogeneous norm N on R 2 , which means that for any (x 1 , x 2 ) ∈ R 2 , N(±x 1 , ±x 2 ) = N(±x 2 , ±x 1 ) = N(x 1 , x 2 ). Let F be a non-empty connected closed subset of R 2 . By Corollary 26, the greatest Hausdorff discretization ∆ H (F) will be 4-connected, as shown in [22]. Consider now an arbitrary Hausdorff discretization S of F. We have two cases: 1. N is not proportional to the L 1 norm. By Proposition 30, S will be connected for the graph linking any two distinct p, q ∈ D such that ‖q − p‖ 1 ≤ 2 and ‖q − p‖∞ < 2; this is simply the 8-adjacency graph, see Figure 7 (b), thus S will be 8-connected. We obtain thus the result of [22], namely that all Hausdorff discretizations of F are 8-connected. 2. N is the L 1 norm. By Corollary 28, S will be connected for the graph with vertex set D and with an edge linking any two distinct p, q ∈ D such that ‖q − p‖ 1 ≤ 2, see Figure 7 (a). As shown in Figure 8, S will not always be 8-connected, since we require the adjacency to include the case of two pixels that differ by 2 in exactly one coordinate. This case was overlooked in [22].

Conclusion
This paper is part of a series on Hausdorff discretization, the approach to discretization which associates to a closed set F in a "continuous" metric space E any subset S of a "discrete" subspace D which minimizes the Hausdorff distance between S and F [12][13][14][15]22]. Here we investigated the relation between the topological connectivity of that closed set F and the "discrete" connectivity of its discretization S. As in our previous papers, we have presented our theory in the most general framework possible: E is an arbitrary metric space and D is a boundedly finite subset of E with finite covering radius rc, see (11) and Axiom 1. Then the Hausdorff distance between a closed set F and any of its Hausforff discretizations is equal to its Hausdorff radius r H (F), see (12). We have r H (F) ≤ rc.
For some results we required that E is boundedly compact and that it satisfies the middle point property, so that it satisfies the interval property, see Definition 4 and Corollary 11. Of course, all these properties are indeed satisfied for E = R n , D = Z n and a distance based on a norm.
For connectivity, we have not restricted ourselves to the topological one (in E) and the graph-theoretical one (in D). We rather based ourselves on the theory of connections and partial connections [10,11,18].
We first consider a connected proximinal set F (when E is boundedly compact, every closed set is proximinal). The greatest Hausdorff discretization of F belongs to the partial connection generated by the traces of all balls with radius equal to r H (F). On the other hand, any Hausdorff discretization S of F will be connected in the graph G r H (F) (F) where two distinct points of D are joined by an edge if their closed balls of radius r H (F) intersect. In particular, it will be connected in the graph where two distinct points of D at distance ≤ 2rc are joined by an edge.
We next generalize topological connectivity by considering non-separated closed sets, see Definition 5. They include all connected sets, but also non-compact disconnected sets where the connected components are "asymptotic" to each other, see Example 2. The family of non-separated closed sets constitutes a connection on the lattice of closed sets. Given a non-separated closed set F, for all r > r H (F), any Hausdorff discretization S of F will be connected in the graph Gr(S) where two distinct points of D are joined by an edge if their closed balls of radius r intersect.
When the closed set F is separated, connectivity is not preserved when rc is small enough. More precisely, if F is s-separated for some s > 0, and rc < s/4, then there exists r > rc such that the graph Gr(S) will be disconnected.
We consider then discretization in multiple subspaces D(ρ) with resolution ρ tending to zero, see Axiom 2; then Hausdorff discretizations of F will be connected at all resolutions if and only F is non-separated.
In a second part of the paper, we consider the case where E = R n , D = Z n and the metric is induced by a coordinate-symmetrical norm (for instance, the Lp norm, 1 ≤ p ≤ ∞). Then the greatest Hausdorff discretization of a connected closed subset of E will be axially connected. Note that this result, see Corollary 26, relies on the theory of partial connections and chainings, and it could not be obtained by more conventional methods based on adjacency graphs or Voronoi tesselations.
When the norm is coordinate-homogeneous (again, this property holds for the Lp norm) the Hausdorff discretization of a non-separated closed set will be connected for a particular adjacency graph on D. When the norm is proportional to the L 1 norm, two distinct p, q ∈ D will be adjacent when ‖q − p‖ 1 ≤ n. When it is not proportional to the L 1 norm, p and q will be adjacent when either p and q differ in at least two coordinates and ‖q − p‖ 1 ≤ n, or p and q differ in exactly one coordinate and ‖q − p‖ 1 < n; an equivalent condition is that ‖q − p‖ 1 ≤ n and ‖q − p‖∞ < n.
For n = 2, this gives the result of [22]: the greatest Hausdorff discretization of a connected closed subset F of R 2 is 4-connected, and when the norm is coordinate-homogeneous but not proportional to the L 1 norm, every Hausdorff discretization of F will be 8-connected.
For n > 2, this graph based on the L 1 norm relies on neighbourhoods that are generally too large, so it is better to use the general theory, giving connectivity for the digital graph with an edge between two digital points p, q such that d(p, q) ≤ 2rc.
Our study shows the interest of the recent notion of a partial connection [11]. In particular, for E = R n , D = Z n and a coordinate-symmetrical norm, it would be interesting to investigate the partial connection generated by digital traces of balls of a given Hausdorff radius. Indeed, for n > 2 it might well be more restricted than the set of all axially connected digital sets, in other words, such sets may have some "thickness" (this is not the case for n = 2, see Figure 6 (b)). It would be interesting to extend our results to related forms of discretization, such as the discretization by dilation considered in [12,13,15].
Another possible approach would be to associate to every digital set S ⊆ Z n its Euclidean representation as the union of cells of its points, C(S) = ⋃︀ p∈S C(p), and to associate to a closed set F ⊆ R n a discretization S ⊆ Z n such that F can be homotopically deformed into C(S).