RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS

GRIGORY IVANOV AND MARIANA LOPUSHANSKI

Abstract. We provide an algorithm of constructing a rectifiable curve between two sufficiently close points of a proximally smooth set in a uniformly convex and uniformly smooth Banach space. Our algorithm returns a reasonably short curve between two sufficiently close points of a proximally smooth set, is iterative and uses a certain modification of the metric projection. We estimate the length of a constructed curve and its deviation from the segment with the same endpoints. These estimates coincide up to a constant factor with those for the geodesics in a proximally smooth set in a Hilbert space.

1. Introduction Weakly convex sets have been studied in non-smooth analysis for several decades. Several established mathematicians proposed their own definition of a weakly convex set, among them are Federer [Fed59], Efimov and Stechkin [ES58], Vial [Via83], Rockafellar (for akin classes of functions) [Roc81]. In this paper we stick to the most convenient in our opinion definition of a weakly convex set due to Clarke, Stern and Wolenski [CSW95]. A closed set in a Banach space is called proximally smooth with constant 푅 if distance to it from a point of the space is continuously differentiable in the open 푅-neighborhood of this set excluding the set itself (this and other definitions used in the introduction are formally given below in Section2). As for other definitions, each of them characterizes weakly convex sets as the set with a certain property, e.g. differentiability of the distance function, supporting by balls, hypomonotonicity of the normal cone, etc. It turns out that many of these definitions are equivalent in Hilbert space, which allows using weakly convex sets in different applications. See, for example, [Bal17] and [Iva06b]. However, everything is a bit trickier in Banach spaces. Different classes of weakly convex sets in Banach spaces were studied in [BTZ06],[BTZ11], [Ali12]. As was shown in [BI09], some of the definitions are still equivalent in uniformly smooth and uniformly convex Banach spaces, but some not. The first author [Iva17] showed that the hypomonotonicity of the normal cone of a closed set fails to be equivalent to the proximally smoothness in any Banach space that is not isomorphic to a Hilbert space (in a Hilbert space these two properties are equivalent, see [Iva06a, Theorem 1.9.1] and [PRT00, Corollary 2.2]). Another equivalent to the proximal smoothness property of a closed set in a Hilbert space was given in [Iva06a, Theorem 1.14.2]. We formulate it as follows. Proposition 1.1. Let 퐴 be a closed set in a Hilbert space and 푅 > 0. The following conditions arXiv:2012.10691v1 [math.FA] 19 Dec 2020 are equivalent: (1) The set 퐴 is proximally smooth with constant 푅. (2) For any two different points 푥0, 푥1 ∈ 퐴 with |푥0 − 푥1| < 2푅, there exists a curve 훾 in 퐴 with endpoints 푥0 and 푥1, whose length is at most (︂|푥 − 푥 |)︂ 2푅 arcsin 0 1 . 2푅 This result plays a crucial role in proofs of many other important results related to the properties and applications of proximally smooth sets. For example, it implies the existence and uniqueness of the shortest path connecting two sufficiently close points of a proximally smooth set in a

Date: September 13, 2021. Supported by Russian Science Foundation, project N 19-11-00087. 1 RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 2

Hilbert space, and that a “locally” proximally smooth set in a Hilbert space is proximally smooth. However, even in a sufficiently smooth and convex Banach space, the existence of rectifiable curve between to sufficiently close points of a proximally smooth set has been unknown yet. One might argue that the definition of a proximally smooth set implies locally connectedness of a proximally smooth set since the metric projection of a sufficiently short segment with endpoints in a proximally smooth set onto the set itself has to be continuous (to be more precise, this argument works in a uniformly smooth and uniformly smooth spaces for segments strictly shorter than 2푅). It is not clear whether the curve constructed in such a way is rectifiable since the metric projection onto a proximally smooth set is H¨oldercontinuous [Iva15, Theorem 3.2], but not Lipschitz continuous in a Banach space not isomorphic to a Hilbert space. Even if one can show that such a curve is rectifiable, the same rather unfortunate property of the metric projection implies that this curve has quite nasty behavior. In this paper we provide an algorithm of constructing a rectifiable curve between two suf- ficiently close points of a proximally smooth set in a uniformly convex and uniformly smooth Banach space. Our Algorithm1 returns a reasonably short curve between two sufficiently close points of a proximally smooth sets, is iterative and uses a certain modification of the metric projection. We collect two important properties of the curve constructed with the use of our algorithm in the two following Theorems. LABEL Theorem 1. Let 푋 be a uniformly convex and uniformly smooth Banach space whose modulus of smoothness is of power type 푠. Then there are positive constants 훽퐿 and 퐿 satisfying the following property. Let 퐴 ⊂ 푋 be a proximally smooth set with positive constant 푅, and let ‖푥0−푥1‖ 푥0, 푥1 ∈ 퐴 with 푅 < 훽퐿. Then Algorithm1 returns the curve 훾 in 퐴 with endpoints 푥0 and 푥1 such that inequality (︃ 푠(푠−1))︃ LABEL (︂‖푥 − 푥 ‖)︂ (1) length (훾) ≤ ‖푥 − 푥 ‖ 1 + 퐿 0 1 0 1 푅 holds. LABEL Theorem 2. Let 푋 be a uniformly convex and uniformly smooth Banach space whose modulus of smoothness is of power type 푠. Then there are positive constants 훽퐼 and 퐿퐼 satisfying the following property. Let 퐴 ⊂ 푋 be a proximally smooth with positive constant 푅 set, and let ‖푥0−푥1‖ 푥0, 푥1 ∈ 퐴 with 푅 < 훽퐼 . Then Algorithm1 returns the curve 훾 in 퐴 with endpoints 푥0 and 푥1 such that inclusion LABEL {︂ (︂푥 + 푥 )︂ }︂ (2) 훾 ⊂ conv 푥 , B 0 1 , 푥 0 푟 2 1

(︀ 푥0+푥1 )︀ 푥0+푥1 holds, where B푟 2 is the closed ball centered at 2 of radius (︂‖푥 − 푥 ‖)︂푠−1 푟 = 퐿 ‖푥 − 푥 ‖ 0 1 . 퐼 0 1 푅

We note that we will estimate constants 훽퐿, 퐿, 훽퐼 , 퐿퐼 using constants related to the smooth- ness of Banach space 푋. Moreover, we do not use the uniform convexity of 푋 directly; the reason for using this condition on a space is being able to use different definitions of a weakly convex set, which are equivalent to the proximally smoothness in a uniformly convex and uniformly smooth Banach space. This condition might be relaxed, for example, in a finite dimensional space, but it will add complications and little to the ideas. Also, since the complement of the interior of the unit ball is proximally smooth with constant one and by basic properties of the modulus of smoothness, the bound on 푟 in Theorem2 is asymptotically tight. The bound on the length of the constructed curve in Theorem1 coincides up to a constant factor with the bound on the shortest path between two points on a fixed distance in a proximally smooth set RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 3

in uniformly smooth space whose modulus of smoothness is of power type 2 (for example, in a Hilbert space and in 퐿푝 spaces with 푝 ≥ 2). Since the metric projection onto a proximally smooth set in a uniformly convex and uniformly smooth Banach space is H¨oldercontinuous [Iva15, Theorem 3.2], we get the following result as an immediate consequence of Theorem1. Corollary 1.1. Let 푋 be a uniformly convex and uniformly smooth Banach space whose moduli of smoothness and convexity are of power type. Let 퐴 ⊂ 푋 be a proximally smooth set with ‖푥0−푥1‖ positive constant 푅, and let 푥0, 푥1 ∈ 퐴 with 푅 < 2. Then there is a rectifiable curve 훾 in 퐴 with endpoints 푥0 and 푥1. The rest of the paper is organized as follows. In the next Section2 we give the standard terminology related to weakly convex sets and to the geometry of the unit ball of a Banach space. In Section3 we study the distance function to a proximally smooth set restricted to a segment with endpoints in this set. In Section4 we describe our Algorithm1 for the construction of the rectifiable curve with endpoints in a proximally smooth set, and also summarize the assumptions needed to show the correctness of the algorithm. Then in Section5 we estimate the length of curve returned by Algorithm1 and prove Theorem1. In Section6 we prove Theorem2. Finally, in the last section Section7 we prove several purely technical results used in the proofs.

2. Terminology and basic properties LABEL 2.1. Properties of the unit ball. Let 푋 be a real Banach space, and 푋* be its conjugate space. We use ⟨푝, 푥⟩ to denote the value of a functional 푝 ∈ 푋* at a vector 푥 ∈ 푋. For 푟 > 0 and 푐 ∈ 푋 we denote by B푟(푐) the closed ball with center 푐 and radius 푟. We will use the notation [푥푦] for the segment with endpoints 푥 and 푦. Define {︂ ‖푥 + 푦‖ }︂ 훿 (휀) = inf 1 − : 푥, 푦 ∈ B (0), ‖푥 − 푦‖ ≥ 휀 푋 2 1 and {︂‖푥 + 푦‖ + ‖푥 − 푦‖ }︂ 휚 (휏) = sup − 1 : ‖푥‖ = 1, ‖푦‖ = 휏 . 푋 2 + + Functions 훿푋 (·) : [0, 2] → [0, 1] and 휚푋(·): R → R are referred to as the moduli of convexity and smoothness of 푋, respectively. The modulus of convexity is of power type 푠 > 0 if for some constant 퐶푐푣 inequality LABEL 푠 (3) 훿푋 (휀) ≥ 퐶푐푣휏 holds for any 휏 ∈ [0, 2). A Banach space 푋 is called uniformly convex if 훿푋 (휀) > 0 for all 휀 > 0, and uniformly smooth 휚푋(휏) if 휏 → 0 as 휏 → 0. We refer the reader to the book [Die75] as a comprehensive survey on these moduli and their geometric properties. In what follows, we consider only uniformly smooth Banach spaces. In such spaces, for any non-zero vector 푥, there is a unique unit functional 푝 attaining its norm on 푥. Let 푥 be a non- zero vector of 푋 and 푝 be a unit functional attaining its norm on 푥, we use 퐻푥 to denote the hyperplane {푦 ∈ 푋 : ⟨푦, 푝⟩ = 0}. We will say that 푦 is quasi-orthogonal to vector 푥 ∈ 푋 ∖ {0} and write 푦q푥 if 푦 ∈ 퐻푥. Note that the following conditions are equivalent: – 푦 is quasi-orthogonal to 푥; – for any 휆 ∈ R vector 푥 + 휆푦 lies in the supporting hyperplane to the ball B‖푥‖(0) at 푥; – for any 휆 ∈ R inequality ‖푥 + 휆푦‖ ≥ ‖푥‖ holds; – 푥 is orthogonal to 푦 in the sense of Birkhoff–James (see [Die75, Chapter 2] and [AMW12]). RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 4

2.2. Modulus of smoothness and related functions. The modulus of smoothness of a Banach space is a strictly increasing convex function satisfying the following inequality of Day– Nordlander type (see [Die75, Chapter 3]) √ LABEL 2 (4) 1 + 휏 − 1 ≤ 휚푋(휏) ≤ 휏 for all 휏 ∈ [0, +∞).

The modulus of smoothness is of power type 푠 if for some constant 퐶푠푚, LABEL 푠 (5) 휚푋(휏) ≤ 퐶푠푚휏 for all 휏 ∈ [0, +∞). It follows that the modulus of smoothness of a uniformly smooth Banach space might be of power type 푠 only for some 푠 in (1, 2]. In our computations we will use two functions related to the modulus of smoothness of a Banach space. Define function 휔푋 : [0, +∞) → [0, +∞) by 휚 (휏) 휔 (휏) = 푋 . 푋 휏 Since the modulus of smoothness of a uniformly smooth Banach space is a strictly increasing convex function, we conclude that 휔푋 (·) is a strictly increasing function. Thus, the inverse −1 function 휔푋 (·) is also strictly increasing. + The second function 휁푋 : [0, +∞) → [0, +∞) is defined by + 휁푋 (휀) = sup {‖푥 + 휀푦‖ : ‖푥‖ = ‖푦‖ = 1, 푦q푥} . + Thus, 휁푋 (·) − 1 bounds the deviation of a point in a supporting hyperplane from the unit ball. This modulus of a Banach space was studied in [IM17], where it was shown that it is equivalent to the modulus of smoothness near zero. LABEL Proposition 2.1. Let 푋 be an arbitrary Banach space. Then (︂ 휀 )︂ [︂ 1]︂ 휚 ≤ 휁+(휀) − 1 ≤ 휚 (2휀) , 휀 ∈ 0, . 푋 2(1 + 휀) 푋 푋 2

+ (︀ +)︀−1 It is not hard to see that 휁푋 is strictly increasing, and hence, its inverse function 휁푋 is well-defined and is strictly increasing. 2.3. Weakly convex sets. The distance from a point 푥 ∈ 푋 to a set 퐴 ⊂ 푋 is defined as dist(푥, 퐴) = inf ‖푥 − 푎‖ . 푎∈퐴 The metric projection of a point 푥 onto a set 퐴 is defined as any element of the set

푃퐴(푥) = {푎 ∈ 퐴 : ‖푎 − 푥‖ = dist(푥, 퐴)}. We call the set {푥 ∈ 푋 : 0 < dist(푥, 퐴) < 푅} the open 푅-neighborhood of a set 퐴. Definition 2.1. A set 퐴 ⊂ 푋 is called proximally smooth with constant 푅 if it is closed and the distance function 푥 ↦→ dist(푥, 퐴) is continuously differentiable on the open 푅-neighborhood of 퐴. The geometric properties of proximally smooth sets are hidden in the definition. To clarify these geometrical properties, which are very useful in this paper, we introduce two equivalent (in certain spaces) to the proximal smoothness properties. LABEL Proposition 2.2 ([BI09]). Let 푋 be a uniformly convex and uniformly smooth Banach space, let 퐴 ⊂ 푋 be a closed set, and let 푅 > 0. The following assertions are equivalent: (1) 퐴 is proximally smooth with constant 푅. LABEL (2) the projection map 푥 → 푃퐴(푥) is single valued and continuous on the open 푅-neighborhood of 퐴. RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 5

LABEL (3) for any 푢 in the open 푅-neighborhood of 퐴 and any 푥 ∈ 푃퐴(푢) inequality (︂ 푅 )︂ dist 푥 + (푢 − 푥), 퐴 ≥ 푅 ‖푢 − 푥‖ holds. Roughly speaking, the last property here implies that the set can be supported by a ball of fixed radius 푅 at a point of its boundary. LABEL 2.4. Auxiliary geometric constructions.

Definition 2.2. Let 퐴 be proximally smooth with constant 푅, let 푥0 and 푥1 be two distinct ‖푥0−푥1‖ points of 퐴 with 푅 < 2, we say that an arbitrary point of the set (︂ 푥 + 푥 )︂ 푃 ([푥 , 푥 ]) ∩ 퐻 + 0 1 퐴 0 1 푥0−푥1 2

푥0+푥1 is a slice-projection of the midpoint 2 of the segment [푥0푥1] onto 퐴. Proposition 2.2 and the separation lemma imply that the slice-projection is non-empty in a uniformly smooth and uniformly convex Banach space. Given a point 푥 ∈ 푋 and a set 퐷 ⊂ 푋, we denote the cone {︃ 푛 }︃ ∑︁ 푥 + 휆푖(푠푖 − 푥): 푛 ∈ N; 휆푖 ≥ 0 and 푠푖 ∈ 퐷 for all 푖 ∈ [푛] 1 as cone(푥, 퐷). Note that cone(푥, 퐷) is a convex cone for any nonempty set 퐷. In our computations we will extensively use the following quantity, which describes in a certain way the distortion of the distance function. By definition put (︀ 휏 )︀ (︀ 휏 )︀ LABEL ′ 8 휚푋 푅 8휔푋 푅 (6) 푅 (휏, 푅) = 푅 (︀ 휏 )︀ = 휏 (︀ 휏 )︀. 1 − 8휔푋 푅 1 − 8휔푋 푅 3. Distance to a proximally smooth set LABEL In this section we bound the distance between the midpoint of a segment with endpoints in a proximally smooth set 퐴 and its slice-projection onto 퐴. LABEL Lemma 3.1. Let 푋 be a uniformly convex and uniformly smooth Banach space, let 푅 > 0 ‖푥0−푥1‖ and 퐴 ⊂ 푋 be a proximally smooth with constant 푅 set. Let 푥0, 푥1 ∈ 퐴 with 푅 < 2 and 휆 ∈ [0, 1]. Then the following bound on the distance from point 푥휆 = (1 − 휆)푥0 + 휆푥1 to set 퐴 holds LABEL (︂‖푥 − 푥 ‖)︂ (7) dist(푥 , 퐴) ≤ 8푅휆(1 − 휆) 휚 0 1 . 휆 푋 푅

Proof. Proposition 2.2 implies that 푃퐴(푥휆) is nonempty for all 휆 ∈ [0, 1]. Fix an arbitrary 휆 ∈ (0, 1) and consider 푦 ∈ 푃퐴(푥휆). Using assertion3 of Proposition 2.2, we get that (︁ )︁ ⃦ ⃦ ⃦ ⃦ 푥휆−푦 푥휆−푦 푥휆−푦 푦−푥0 dist 푦 + 푅 , 퐴 ≥ 푅. Hence ⃦푦 + 푅 − 푥0⃦ ≥ 푅, and ⃦ − ⃦ ≥ 1. By ‖푥휆−푦‖ ⃦ ‖푥휆−푦‖ ⃦ ⃦ ‖푥휆−푦‖ 푅 ⃦ the definition of the modulus of smoothness, we get that (︂⃦ ⃦)︂ ⃦ ⃦ ⃦ ⃦ ⃦푦 − 푥0 ⃦ ⃦ 푥휆 − 푦 푦 − 푥0 ⃦ ⃦ 푥휆 − 푦 푦 − 푥0 ⃦ 2 휚푋 ⃦ ⃦ ≥ ⃦ + ⃦ + ⃦ − ⃦ − 2 ≥ ⃦ 푅 ⃦ ⃦‖푥휆 − 푦‖ 푅 ⃦ ⃦‖푥휆 − 푦‖ 푅 ⃦ ⃦ ⃦ ⃦ 푥휆 − 푦 푦 − 푥0 ⃦ ⃦ + ⃦ − 1. ⃦‖푥휆 − 푦‖ 푅 ⃦ Let 푝 be the unit functional attaining its norm on vector 푥휆 − 푦. Then RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 6

⃦ ⃦ ⟨ ⟩ ⃦ 푥휆 − 푦 푦 − 푥0 ⃦ 푥휆 − 푦 푦 − 푥0 1 ⃦ + ⃦ − 1 ≥ 푝, + − 1 = ⟨푝, 푦 − 푥0⟩ . ⃦‖푥휆 − 푦‖ 푅 ⃦ ‖푥휆 − 푦‖ 푅 푅 Therefore, we obtain that LABEL (︂‖푥 − 푦‖)︂ (8) ⟨푝, 푥 − 푦⟩ ≤ 2푅 휚 0 . 0 푋 푅

Since 푦 ∈ 푃퐴(푥휆) and 푥0, 푥1 ∈ 퐴,

‖푦 − 푥휆‖ = dist(푥휆, 퐴) ≤ min{‖푥휆 − 푥0‖ , ‖푥휆 − 푥1‖} ≤

min{‖푥휆 − 푥0‖ , ‖푥휆 − 푥1‖} ≤ min {휆, (1 − 휆)}‖푥0 − 푥1‖ . Therefore,

‖푦 − 푥0‖ ≤ ‖푦 − 푥휆‖ + ‖푥휆 − 푥0‖ ≤ min{2휆, 1}‖푥0 − 푥1‖ . This, the monotonicity of the modulus of smoothness and inequality (8) yield inequality LABEL (9) ⟨푝, 푥0 − 푦⟩ ≤ Φ(휆), where LABEL (︂min {2휆, 1}‖푥 − 푥 ‖)︂ (10) Φ(휆) = 2푅 휚 0 1 . 푋 푅 Similarly, LABEL (11) ⟨푝, 푥1 − 푦⟩ ≤ Φ(1 − 휆). Multiplying inequalities (9) and (11) by (1 − 휆) and 휆 respectively, and then summing them, one has

⟨푝, 푥휆 − 푦⟩ ≤ (1 − 휆)Φ(휆) + 휆Φ(1 − 휆).

This and the inequality dist(푥휆, 퐴) ≤ ‖푥휆 − 푦‖ = ⟨푝, 푥휆 − 푦⟩ imply that LABEL (12) dist(푥휆, 퐴) ≤ (1 − 휆)Φ(휆) + 휆Φ(1 − 휆).

Using once more the convexity of function 휚푋(·), and the identities 휚푋(0) = 0 and (10), we get that (︂‖푥 − 푥 ‖)︂ (︂‖푥 − 푥 ‖)︂ Φ(휆) ≤ 2푅 min{2휆, 1} 휚 0 1 ≤ 4휆푅 휚 0 1 . 푋 푅 푋 푅 Thus, inequality (12) implies that (︂‖푥 − 푥 ‖)︂ dist(푥 , 퐴) ≤ 8푅휆(1 − 휆) 휚 0 1 . 휆 푋 푅  Proposition 2.2 and the separation lemma imply the following. LABEL Lemma 3.2. Let 푋 be a uniformly convex and uniformly smooth Banach space, let 퐴 ⊂ 푋 ‖푥0−푥1‖ be a proximally smooth with constant 푅 set. Then for any 푥0, 푥1 ∈ 퐴 with 푅 < 2, there exists a slice-projection of the midpoint of [푥0푥1] onto 퐴. LABEL Lemma 3.3. Let 푋 be a uniformly convex and uniformly smooth Banach space, let 퐴 ⊂ 푋 be ‖푥0−푥1‖ −1 a proximally smooth with constant 푅 set. Let 푥0, 푥1 ∈ 퐴 with 푅 < 휔푋 (1/8) , the following inclusion holds {︂ (︂ )︂ (︂ )︂ }︂ 푥0 + 푥1 푥0 + 푥1 푃 ([푥 푥 ]) ⊂ conv 푥 , B ′ ∩ 퐻 + , 푥 , 퐴 0 1 0 푟 2 푥1−푥0 2 1 ′ ′ where 푟 = 푅 (‖푥0 − 푥1‖ , 푅) (see (6)). RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 7

Proof. We fix a unit vector 푦 quasi-orthogonal to 푥1 −푥0 and consider the two-dimensional space (︁ ‖푥0−푥1‖ )︁ 푋2 = Lin{푦, 푥1 −푥0} with the induced norm. Fix 휆 ∈ (0, 1) and set 푟 = 8푅휆(1−휆) 휚푋 푅 ⋃︀ and 푥휆 = 휆푥0 + (1 − 휆)푥1. Note that Lemma 3.1 implies that 푃퐴([푥0, 푥1]) ⊂ B푟(푥휆). 휆∈[0,1] First, we will show that 푥1 ∈/ B푟(푥휆) . That is, we need to verify the following inequality 푟 < ‖푥휆 − 푥1‖ = 휆 ‖푥1 − 푥0‖ , which is equivalent to (︂ )︂ 푅 ‖푥0 − 푥1‖ 8(1 − 휆) 휚푋 < 1. ‖푥0 − 푥1‖ 푅

(︁ ‖푥0−푥1‖ )︁ Since 1 − 휆 ∈ (0, 1), this inequality holds whenever 휔푋 푅 ≤ 1/8. Thus, 푥1 ∈/ B푟(푥휆) . Denote the intersection point of ray 푥휆푥1 with the boundary of the ball B푟(푥휆) by 푣 and let ℓ be one of the two lines passing through 푥1 supporting B푟(푥휆). The tangent point of ℓ and B푟(푥휆) is denoted by 푤. Note that 푦 is the directional vector of the line supporting B푟(푥휆) at 푣. Therefore, the lines ℓ and 푥1/2 + Lin{푦} are not parallel and their intersection point, denoted by 푧, lies in the same half-plane with the point 푤. By similarity, it suffices to set 푟′ equal to ⃦ ⃦ any upper bound on ⃦푥1/2 − 푧⃦ that does not depend on 휆 and 푦. ⃦ ⃦ Let us estimate ⃦푥1/2 − 푧⃦ . Denote the intersection point of the ray 푥1푥휆 and the line 푤 + Lin{푦} by 푣′. By similarity, we get LABEL ⃦ ⃦ 1 ‖푤 − 푣′‖ (13) ⃦푥1/2 − 푧⃦ = ‖푥0 − 푥1‖ ′ . 2 ‖푣 − 푥1‖ ′ Since 푦q(푥1 − 푥0), we have that 푣 ∈ B푟(푥휆) . Hence, we get ′ ′ ‖푤 − 푣 ‖ ≤ 2푟 and ‖푣 − 푥1‖ ≥ ‖푥휆 − 푥1‖ − 푟 = 휆 ‖푥1 − 푥0‖ − 푟. Combining these inequalities with inequality (13), we get

(︁ ‖푥0−푥1‖ )︁ (︁ ‖푥0−푥1‖ )︁ 8푅(1 − 휆) 휚푋 8푅 휚푋 ⃦ ⃦ 푟 푅 푅 ′ ⃦푥1/2 − 푧⃦ = = < = 푟 . 푟 (︁ ‖푥 −푥 ‖ )︁ (︁ ‖푥 −푥 ‖ )︁ 휆 − ‖푥 −푥 ‖ 0 1 0 1 0 1 1 − 8(1 − 휆)휔푋 푅 1 − 8휔푋 푅 This completes the proof.  As an immediate corollary, we get. LABEL Corollary 3.1. Let 푋 be a uniformly convex and uniformly smooth Banach space, let 퐴 ⊂ 푋 be ‖푥0−푥1‖ −1 a proximally smooth with constant 푅 set. Let 푥0, 푥1 ∈ 퐴 with 푅 < 휔푋 (1/8) , fix 휆 ∈ [0, 1] and set 푥휆 = 휆푥0 + (1 − 휆)푥1. Then there is a point 푧휆 ∈ 푃퐴([푥0, 푥1]) such that ′ ‖푧휆 − 푥휆‖ ≤ 4휆(1 − 휆)푅 (‖푥0 − 푥1‖ , 푅)

푥0+푥1 and (푧휆−푥휆)q(푥1−푥0). Moreover, the distance between 2 and any point of the slice projection 푥0+푥1 ′ of the midpoint 2 of the segment [푥0, 푥1] onto 퐴 is at most 푅 (‖푥0 − 푥1‖ , 푅). 4. Construction of a curve LABEL 4.1. Assumptions on the distance between the endpoints. Our algorithm of curve con- struction between two distinct points 푥0 and 푥1 of a proximally smooth set works when points are sufficiently close. Moreover, we need different bounds to prove the convergence of the al- gorithm and, for example, to prove the inclusion in Theorem2. We have decided to collect all the assumptions on the distance between the two starting points. By definition put (︂ ′ )︂ LABEL + 2푅 (‖푥0 − 푥1‖ , 푅) (14) 휇 = 휁푋 ‖푥0 − 푥1‖ and recall the definition of 푅′(휏, 푅) (see (6)). ‖푥0−푥1‖ Assumptions on 푅 : RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 8

LABEL ‖푥0−푥1‖ −1 (1) 푅 < 휔푋 (1/8) . LABEL (2) 휇 < 2. LABEL 휇푠 (3) 2푠−1 < 1. In the next lemma, we show that all these assumptions are fulfilled for a sufficiently small ‖푥0−푥1‖ 푅 . LABEL +(︁ 2푅′(휏,푅) )︁ [︀ −1 )︀ Lemma 4.1. Function 휁푋 휏 is increasing in 휏 on 0, 푅 · 휔푋 (1/8) , and 휇 → 1 as ‖푥0−푥1‖ 푅 → 0. Proof. By Proposition 2.1, it suffices to show that

(︁ ‖푥0−푥1‖ )︁ ′ 8휔푋 푅 (‖푥0 − 푥1‖ , 푅) (6) 푅 = (︁ )︁ ‖푥0 − 푥1‖ ‖푥0−푥1‖ 1 − 8휔푋 푅

‖푥0−푥1‖ tends to zero as 푅 tends to zero, which is an immediate consequence of the uniformly smoothness of 푋. The monotonicity follows from the monotonicity of 휔푋 (·) .  As for the first two of the assumptions, we can bound the corresponding constants using characteristics of a Banach space. The following is a purely technical result, we formulate it as a separate statement and prove it later in Section7. LABEL Claim 4.1. Set LABEL (︂ (휁+)−1(2) )︂ (15) 훽 = 휔−1 푋 . 퐿 푋 + −1 8(2 + (휁푋 ) (2)) −1(︀ 1 )︀ 휏 Then 훽퐿 ≤ 휔푋 8 < 2, and for any positive constants 휏 and 푅 satisfying 푅 < 훽퐿, inequality +(︁ 2푅′(휏,푅) )︁ 휁푋 휏 < 2 holds. 4.2. Algorithm for the construction of a curve. Algorithm 1: Construction of a curve in a proximally smooth set LABEL Data: A proximally smooth with constant 푅 set 퐴 ⊂ 푋, two distinct points 푥0, 푥1 in 퐴 ‖푥0−푥1‖ with 푅 < 훽퐿, where 훽퐿 is given by (15). Result: A rectifiable curve 푓([0, 1]), where 푓 : [0, 1] → 퐴 is a continuous function with 푓(0) = 푥0 and 푓(1) = 푥1 푗 푖 Set 푆0 = {0, 1} and 푆푖 = { 2푖 | 푗 ∈ [2 ]}} ∪ {0} for 푖 ∈ N. (1) Define 푓 at points of 푆0 as follows: 푓(0) = 푥0 and 푓(1) = 푥1. (2) For every 푖 ∈ N, we extend the domain of 푓 to the set 푆푖 ∖ 푆푖−1 as follows: 2푗−1 set the value of 푓 at 2푖 to be a slice-projection of the midpoint of the segment [︀ (︀ 푗−1 )︀ (︀ 푗 )︀]︀ 푖−1 푓 2푖−1 푓 2푖−1 for all 푗 ∈ [2 ] on 퐴. (3) Continuously extend 푓 on [0, 1].

4.3. Problems needed to be justified. To show the correctness of Algorithm1, one needs to check: (1) For every 푖 ∈ N and 푗 ∈ [2푖−1], there exists a slice-projection of the midpoint of a [︀ (︀ 푗−1 )︀ (︀ 푗 )︀]︀ segment 푓 2푖−1 푓 2푖−1 onto the set 퐴. (2) 푓 can be continuously extended from the rational numbers of [0, 1] to the whole segment. (3) Curve 푓([0, 1]) is rectifiable. According to Lemma 3.2 to show the existence of a slice-projection at each step, it suffices [︀ (︀ 푗−1 )︀ (︀ 푗 )︀]︀ −1 to show that the length of segment 푓 2푖−1 푓 2푖−1 is less than 푅휔푋 (1/8) . We will justify these questions for 푥0, 푥1 and 푅 satisfying assumption2. RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 9

푥1 푥1

푥0+푥1 2 (︀ 3 )︀ 푥 푥 (︀ 1 )︀ 푓 0 (︀ 1 )︀ 0 푓 4 (︀ 1 )︀ 4 푓 2 푓 2

Figure 1. The first two iterations of the algorithm.

5. Bound on length LABEL Theorem1 is an immediate consequence of the following theorem. LABEL Theorem 3. Under the condition of Theorem1, additionally let the modulus of smoothness of 푋 satisfy inequality (5) and ‖푥 − 푥 ‖ (︂ (휁+)−1(2) )︂ 0 1 < 훽 = 휔−1 푋 . 퐿 푋 + −1 푅 8(2 + (휁푋 ) (2)) Then Algorithm1 returns curve 훾 satisfying the following inequality: [︃(︂ )︂푠 (︂ )︂푠(푠−1)]︃ 16 푠+1 1 ‖푥0 − 푥1‖ length (훾) ≤ ‖푥0 − 푥1‖ exp 퐶 , 5 푠푚 (︀ 휇 )︀푠(푠−1) 2푅 1 − 2 where 휇 is given by (14).

Proof. We denote by 훾푖 the polygonal curve with consecutive vertices of {푓(푡) | 푡 ∈ 푆푖} and by ∆푖 the largest length of a segment of 훾푖, 푖 ∈ N ∪ {0}. We start with an upper bound on ∆푖. By construction, we have that (︂ ′ )︂ LABEL ∆푖−1 + 2푅 (∆푖−1, 푅) (16) ∆푖 ≤ 휁푋 . 2 ∆푖−1 + To have a meaningful bound, one needs to guarantee that the argument of 휁푋 (·) is less than one. Using Claim 4.1, we see that ∆푖 < ∆푖−1 starting with 푖 = 1. Hence, we have Lemma 5.1. LABEL 휇 (︁휇)︁푖 (17) ∆ ≤ ∆ ≤ ∆ → 0 as 푖 → ∞. 푖 2 푖−1 2 0 +(︁ 2푅′(휏,푅) )︁ Denote 휓(휏) = 휁푋 휏 − 1. Using (17) in (16), we obtain ∆ (︂ (︂(︁휇)︁푖 )︂)︂ ∆ [︂ (︂(︁휇)︁푖 )︂]︂ ∆ ≤ 푖−1 1 + 휓 ∆ ≤ 푖−1 exp 휓 ∆ . 푖 2 2 0 2 2 0 Therefore, we have [︃ 푖 ]︃ (︂ 푖 )︂ ∆0 ∑︁ (︁휇)︁ ∆ ≤ exp 휓 ∆ . 푖 2푖 2 0 0 Finally, [︃ 푖 ]︃ [︃ ∞ ]︃ LABEL ∑︁ (︂(︁휇)︁푖 )︂ ∑︁ (︂(︁휇)︁푖 )︂ (18) length (훾 ) ≤ 2푖∆ ≤ ∆ exp 휓 ∆ ≤ ∆ exp 휓 ∆ . 푖 푖 0 2 0 0 2 0 0 0 We need to bound the series in the rightmost part in (18). It is a purely technical task that involves only routine computations. We formulate the following inequality and prove it later in Section7. RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 10

LABEL Claim 5.1. Let 0 < ∆0 < 푅 · 훽퐿. Then [︃ ∞ ]︃ [︃ ]︃ (︂ 푖 )︂ (︂ )︂푠 ∑︁ (︁휇)︁ 16 푠+1 1 푠(푠−1) exp 휓 ∆0 ≤ exp 퐶 (∆0) . 2 5 푠푚 (︀ 휇 )︀푠(푠−1) 0 1 − 2 Thus, passing to the limit in (18) as 푖 tends to infinity and using Claim 5.1, we get that the lengths of 훾푖 are uniformly bounded. Therefore, function 푓 constructed above is uniformly continuous on the rational numbers of interval [0, 1]. By routine, it can be extended to the continuous function on the whole interval with values in 퐴. Thus, 푓 defines a continuous curve in 퐴. Using (18) again, one sees that the first variation of 훾 = 푓([0, 1]) is bounded. Thus, 훾 is rectifiable.  6. Inclusion LABEL In this Section we prove Theorem2. The proof consists of several steps. Firstly, we show that the curve returned by Algorithm1 is in a cylinder of a certain radius around line 푥0푥1. Then we show that the second part of the curve, that is 푓([1/2, 1]), belongs to a certain convex 푘 푘−1 cone with apex at 푥0. Finally, we prove that all parts of the curve of the form 푓([1/2 , 1/2 ]) are in a certain cone with apex at 푥0. Again, we denote by 훾푖 the polygonal curve with consecutive vertices of {푓(푡) | 푡 ∈ 푆푖} and by ∆푖 the largest length of a segment of 훾푖, 푖 ∈ N ∪ {0}. Denote an intersection point of the

hyperplane 퐻푥0−푥1 + 푓(푡) and the line 푥0푥1 by 푔(푡). Define 푔2 : [0, 1] → [0, ∞) and 푔1 : [0, 1] → [0, ∞) by 푔2(푡) = ‖푥0 − 푔(푡)‖ and 푔1(푡) = ‖푓(푡) − 푔(푡)‖ . ‖푥0−푥1‖ We choose 훽퐼 in such a way that inequality 푅 < 훽퐼 implies assumption3. By Lemma 4.1 such a constant exists. To bound 푔1 from above and 푔2 from below, we need the following purely technical result, which we prove in the next section. LABEL Claim 6.1. Under the conditions of Theorem2, additionally let the modulus of smoothness of ‖푥0−푥1‖ 푋 satisfy inequality (5) and let 푅 satisfy assumption3, then inequality 푘 ′ (︂ )︂푠−1 ∑︁ 푅 (∆푗, 푅) 24퐶푠푚 ∆0 ∆0 < · · 2푘−푗 1 − 휇푠 푅 2푘 푗=0 2푠−1 holds. LABEL Lemma 6.1. Under the condition of Theorem2, additionally let the modulus of smoothness of ‖푥0−푥1‖ 푋 satisfy inequality (5) and 푅 satisfy assumption3. For any 푡 ∈ [0, 1], inequality (︂ )︂푠−1 48퐶푠푚 ∆0 푔1(푡) ≤ 휇푠 · ∆0 1 − 2푠−1 푅 holds.

Proof. We will greedy estimate 푔1(푡), 푡 ∈ (0, 1) from above. We proceed by induction on 푘 and will prove the following 푘 LABEL (︂2푗 − 1)︂ ∑︁ (19) 푔 ≤ 2 푅′(∆ , 푅) ∀ 푘 ∈ , 푗 ∈ [2푘−1]. 1 2푘 푖 N 푖=0 Lemma 3.3 yields the case 푘 = 1. Suppose inequality (19) holds for 푘 − 1. Let us prove it for 푘. Fix 푗 ∈ [2푘−1] and denote 푓(︀ 2푗−2 )︀ + 푓(︀ 2푗 )︀ 푎 = 2푘 2푘 and 푏 = 푥 푥 ∩ (퐻 + 푎). 2 0 1 푥0−푥1 RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 11

Then, by the triangle inequality, we get (︂ )︂ ⃦ (︂ )︂ (︂ )︂⃦ ⃦ (︂ )︂ ⃦ ⃦ (︂ )︂⃦ 2푗 − 1 ⃦ 2푗 − 1 2푗 − 1 ⃦ ⃦ 2푗 − 1 ⃦ ⃦ 2푗 − 1 ⃦ 푔1 = ⃦푔 − 푓 ⃦ ≤ ⃦푔 − 푏⃦+‖푏 − 푎‖+⃦푎 − 푓 ⃦ . 2푘 ⃦ 2푘 2푘 ⃦ ⃦ 2푘 ⃦ ⃦ 2푘 ⃦ Again, by the triangle inequality, 푔 (︀ 2푗−2 )︀ + 푔 (︀ 2푗 )︀ ‖푏 − 푎‖ ≤ 1 2푘 1 2푘 , 2 and by construction, we have ⃦ (︂ )︂ ⃦ ⃦ (︂ )︂ ⃦ ⃦ 2푗 − 1 ⃦ ⃦ 2푗 − 1 ⃦ ′ ⃦푔 − 푏⃦ ≤ ⃦푓 − 푎⃦ ≤ 푅 (∆푘, 푅). ⃦ 2푘 ⃦ ⃦ 2푘 ⃦ Thus, (︂ )︂ ⃦ (︂ )︂ ⃦ (︀ 2푗−2 )︀ (︀ 2푗 )︀ 푘−1 2푗 − 1 ⃦ 2푗 − 1 ⃦ 푔1 2푘 + 푔1 2푘 ′ ∑︁ ′ 푔1 ≤ 2 ⃦푓 − 푎⃦ + ≤ 2푅 (∆푘, 푅) + 2 푅 (∆푖, 푅). 2푘 ⃦ 2푘 ⃦ 2 푖=0 Inequality (19) is proven. Thus, by Claim 6.1 and continuity, we have that (︂ )︂푠−1 48퐶푠푚 ∆0 푔1(푡) ≤ 휇푠 · ∆0 1 − 2푠−1 푅 for any 푡 ∈ [0, 1].  Lemma 6.1 says that the curve returned by the algorithm lies in a certain cylinder around line 푥0푥1. To understand curve’s behavior near endpoints, we need a more subtle argument. LABEL Lemma 6.2. Under the condition of Theorem2, additionally let the modulus of smoothness ‖푥0−푥1‖ [︀ 1 ]︀ of 푋 satisfy inequality (5) and 푅 satisfy assumption3. For any 푡 ∈ 2 , 1 , inequality Δ0 푔2(푡) ≥ 4 holds. (︀ 1 )︀ Proof. We will greedy estimate 푔2(푡), 푡 ∈ 2 , 1 from below. We proceed by induction on 푘 and will prove the following (︂ )︂ 푘 LABEL 2푗 − 1 ∆0 ∑︁ (20) 푔 ≥ − 푅′(∆ , 푅) ∀ 푘 ∈ , 푗 ∈ [2푘−1], 푗 > 2푘−2. 2 2푘 2 푖 N 푖=0

The construction of the curve and definition of 푔2(·) yield that (︂1)︂ ∆ ∆ 푔 = 0 ≥ 0 − 푅′(∆ , 푅). 2 2 2 2 0 Thus, we have the induction basis. Suppose now that (20) holds for some 푘 − 1. Let us now prove that it holds for 푘. Fix 푗 ∈ [2푘−1], 푗 > 2푘−2 and denote 푓(︀ 2푗−2 )︀ + 푓(︀ 2푗 )︀ 푎 = 2푘 2푘 and 푏 = 푥 푥 ∩ (퐻 + 푎). 2 0 1 푥0−푥1

Note that 푏 and 푔(푡) lie on the line 푥0푥1.

By the triangle inequality, we obtain that (︂ )︂ ⃦ (︂ )︂⃦ ⃦ (︂ )︂ ⃦ 2푗 − 1 ⃦ 2푗 − 1 ⃦ ⃦ 2푗 − 1 ⃦ 푔2 = ⃦푥0 − 푔 ⃦ ≥ ‖푥0 − 푏‖ − ⃦푔 − 푏⃦ . 2푘 ⃦ 2푘 ⃦ ⃦ 2푘 ⃦ RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 12

퐻푥0−푥1 + 푎 푥 푔(︀ 3 )︀ 1 푏 4 푎 푥0 (︀ 3 )︀ 푓 4

(︀ 1 )︀ 푓 2 Figure 2. Illustration for the proof of Lemma 6.1. Here, 푘 = 2 and 푗 = 2. (︀ 1 )︀ (︀ 1 )︀ Thus, 2푎 = 푓 2 + 푓(1) = 푓 2 + 푥1. The dashed lines denotes hyperplanes

parallel to 퐻푥0−푥1 .

By the definition of 푏, we have 푔 (︀ 2푗−2 )︀ + 푔 (︀ 2푗 )︀ ‖푏 − 푥 ‖ = 2 2푘 2 2푘 . 0 2

As 푓(푡) − 푔(푡) is quasi-perpendicular to 푥0 − 푥1, we have that ⃦ (︂ )︂ ⃦ ⃦ (︂ )︂ ⃦ ⃦ 2푗 − 1 ⃦ ⃦ 2푗 − 1 ⃦ ⃦푔 − 푏⃦ ≤ ⃦푓 − 푎⃦ . ⃦ 2푘 ⃦ ⃦ 2푘 ⃦ Thus, (︂ )︂ (︀ 2푗−2 )︀ (︀ 2푗 )︀ ⃦ (︂ )︂ ⃦ 푘−1 2푗 − 1 푔2 2푘 + 푔2 2푘 ⃦ 2푗 − 1 ⃦ ∆0 ′ ∑︁ ′ 푔2 ≥ − ⃦푓 − 푎⃦ ≥ − 푅 (∆푘, 푅) − 푅 (∆푖, 푅). 2푘 2 ⃦ 2푘 ⃦ 2 푖=0 Inequality (20) is proven. Δ0 Claim 6.1, the assumptions on ∆0, and the continuity of 푔(·) imply that 푔2(푡) ≥ 4 for all [︀ 1 ]︀ 푡 ∈ 2 , 1 .  LABEL Corollary 6.1. Under the condition of Theorem2, additionally let the modulus of smoothness ‖푥0−푥1‖ [︀ 1 ]︀ of 푋 satisfy inequality (5) and 푅 satisfy assumption3. For any 푡 ∈ 2 , 1 , the set 1 {푓(푡): 푡 ∈ [ 2 , 1]} is a subset of

cone (푥0, B퐿1Δ0 (푥1) ∩ (퐻푥0−푥1 + 푥1)) , where (︂ )︂푠−1 400퐶푠푚 ∆0 퐿1 = 휇푠 푡 . 1 − 2푠−1 푅 Define (︂ )︂푠−1 (︂ (︂ (︂ )︂)︂)︂ 1 400퐶푠푚 ∆푖 1 푟푖 = 휇푠 · 휇푠 · ∆푖 and 퐺푖 = cone 푥0, B푟푖 푓 푖−1 , 푖 ∈ N. 1 − 2푠−1 1 − 2푠−1 푅 2 LABEL Lemma 6.3. Under the condition of Theorem2, additionally let the modulus of smoothness of ‖푥0−푥1‖ 푋 satisfy inequality (5) and 푅 satisfy assumption3. Then 푓([0, 1]) ⊂ 퐺1. ′ Proof. By inequality (17) and assumption (3), we have that ∆푖 ≤ ∆푖−1. Hence, 푅 (∆푖, 푅) ≤ ′ 푖 푖−1 푅 (∆푖−1, 푅). Thus, Corollary 6.1 implies that 푓([1/2 , 1/2 ]) ⊂ 퐺푖 for every 푖 ∈ N. Thus, to prove the lemma, it suffices to show that 퐺푖 ⊂ 퐺1 for all 푖 ∈ N. By construction, the curve 푖 푖 푓([0, 1/2 ]) coincides with the curve returned by Algorithm1 applied to points 푥0 and 푓(1/2 ). Let us prove that 퐺푖 ⊂ 퐺푖−1. By assumption (3) and by inequality (17), LABEL 1 휇푠 (21) 푟 ≤ · 푟 . 푖 2 2푠−1 푖−1 RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 13

푖−1 Denote the midpoint of a segment [푥0, 푓(1/2 )] by 푎푖. By the triangle inequality,

푖 퐺푖+1 ⊂ cone(푥0, B푟푖+‖푓(1/2 )−푎푖‖(푎푖)) for all 푖 ∈ N. By similarity, (︀ 푖−1 )︀ 푖 푖 cone(푥0, B푟푖+‖푓(1/2 )−푎푖‖(푎푖)) = cone(푥0, B2푟푖+2‖푓(1/2 )−푎푖‖ 푓(1/2 ) ) for all 푖 ∈ N. 푖 Applying Corollary 3.1 with 푥0 = 푥0 and 푥1 = 푓(1/2 ), we get that ⃦ 푖 ⃦ ′ ⃦ 푖−1 ⃦ ′ ⃦푓(1/2 ) − 푎푖⃦ ≤ 푅 (⃦푓(1/2 − 푥0)⃦ , 푅) ≤ 푅 (∆푖, 푅).

Since ∆푖 ≤ ∆0 ≤ 훽퐼 and by Claim 6.1, (︂ )︂푠−1 (︂ 푠 )︂ ⃦ 푖 ⃦ 24퐶푠푚 ∆푖−1 휇 ⃦푓(1/2 ) − 푎푖⃦ ≤ 휇푠 ∆푖 ≤ 1 − 푠−1 푟푖−1. 1 − 2푠−1 푅 2 This and inequality (21) imply that ⃦ 푖 ⃦ 2푟푖 + 2 ⃦푓(1/2 ) − 푎푖⃦ ≤ 푟푖−1. Consequently, by the triangle inequality, one has

푖 퐺푖+1 ⊂ cone(푥0, B푟푖+‖푓(1/2 )−푎푖‖(푎푖)) ⊂ 퐺푖 for all 푖 ∈ N.

Hence, we conclude that 퐺푖 ⊂ 퐺1 for all 푖 ∈ N, completing the proof of the lemma.  By symmetry and by Lemma 6.3, we get the following result which implies Theorem2. Theorem 4. Under the condition of Theorem2, additionally let the modulus of smoothness of 푋 satisfy inequality (5) and 훽퐼 satisfy inequality 푠 (︂ ′ )︂ 휈 + 2푅 (훽퐼 , 푅) 푠−1 < 1, where 휈 = 휁푋 . 2 훽퐼 Then the curve 훾 returned by Algorithm1 satisfies inclusion {︂ (︂푥 + 푥 )︂ }︂ 훾 ⊂ conv 푥 , B 0 1 , 푥 , 0 푟 2 1 where (︂ )︂푠−1 1 400퐶푠푚 ‖푥0 − 푥1‖ 푟 = 휇푠 · 휇푠 ‖푥0 − 푥1‖ . 1 − 2푠−1 1 − 2푠−1 푅

7. Proofs of technical results LABEL −1 (︀ +)︀−1 + Proof of Claim 4.1. Denote 휁 = 휁푋 (2). The definition of 휁푋 (·) implies that LABEL + (22) 1 + 휏 ≥ 휁푋 (휏) ≥ 1. −1 −1 −1 Since 휔푋 (·) is an increasing function, one has 휔푋 (3/40) < 휔푋 (1/8) . Thus, to show that −1 휁−1 3 훽퐿 ≤ 휔푋 (1/8) , and it suffices to show that 2+휁−1 ≤ 5 . By (22) and by monotonicity, we obtain that −1 ⃒ 휁 휏 + 1 ⃒ 3 −1 ≤ ⃒ = . 2 + 휁 2 + 휏 + 1⃒휏=2 5 8휔 휏 ′ 푋( 푅 ) Since 푅 (휏, 푅) = 휏 휏 (see (6)), we have 1−8휔푋( 푅 ) (︃ )︃ (︂2푅′(휏, 푅))︂ 16휔 (︀ 휏 )︀ 휁+ = 휁+ 푋 푅 . 푋 휏 푋 (︀ 휏 )︀ 1 − 8휔푋 푅 RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 14

휏 + Since 푅 < 훽퐿 and functions 휁푋 (·) and 휔푋 (·) are increasing, we obtain that 16휁−1(2) (︃ )︃ ⎛ 푋 ⎞ (︀ 휏 )︀ (︂ )︂ −1 16휔푋 16휔 (훽 ) 8(2+휁 (2)) 휁+ 푅 < 휁+ 푋 퐿 = 휁+ 푋 = 2 푋 (︀ 휏 )︀ 푋 푋 ⎝ 휁−1(2) ⎠ 1 − 8휔푋 1 − 8휔푋 (훽퐿) 푋 푅 1 − −1 (2+휁푋 (2)) completing the proof of Claim 4.1. −1 −1 Let us prove inequality 휔푋 (1/8) < 2. By monotonicity of 휔푋 (·) and by the definition of 휔푋 (·) , we have the following chain −1 휔푋 (1/8) < 2 ⇔ 1/8 < 휔푋 (2) ⇔ 1/4 < 휚푋(2) . The last inequality follows from (4).  Proof of Claim 5.1. By Proposition 2.1 and the definition of 푅′(휏, 푅) (see (6)), we get (︃ )︃ (︂2푅′(휏, 푅))︂ 32휔 (︀ 휏 )︀ 휓(휏) = 휁+ − 1 ≤ 휚 (4푅′(휏, 푅)휏) = 휚 푋 푅 . 푋 휏 푋 푋 (︀ 휏 )︀ 1 − 8휔푋 푅 푠 Claim 4.1 and inequality 휚푋(휏) ≤ 퐶푠푚휏 imply that

⎛ 휚 휏 ⎞ (︃ )︃ 푋( 푅 ) (︃ 푠−1 )︃ (︀ 휏 )︀ 32 휏 (︀ 휏 )︀ (︂ )︂푠 32휔푋 32퐶푠푚 16 (︁ 휏 )︁푠(푠−1) 휚 푅 ≤ 휚 ⎜ 푅 ⎟ ≤ 휚 푅 ≤ 퐶푠+1 . 푋 (︀ 휏 )︀ 푋⎝1 − 8휔 (퐶 )⎠ 푋 3 5 푠푚 푅 1 − 8휔푋 푅 푋 2 1 − 5

Thus, ∞ ∞ (︂ 푖 )︂ (︂ )︂푠 푖(푠(푠−1)) (︂ )︂푠 ∑︁ (︁휇)︁ 16 푠+1 (푠−1)푠 ∑︁ (︁휇)︁ 16 푠+1 (푠−1)푠 1 휓 ∆0 ≤ 퐶 ∆ = 퐶 ∆ . 2 5 푠푚 0 2 5 푠푚 0 (︀ 휇 )︀푠(푠−1) 0 0 1 − 2  Proof of Claim 6.1. Denote

(︁ Δ푗 )︁ 푘 ′ 푘 휚 1 ∑︁ 푅 (∆푗, 푅) ∑︁ 푋 푅 푆 = = . 푘 푘−푗 (︁ (︁ )︁)︁ 8푅 2 푘−푗 Δ푗 푗=0 푗=0 2 1 − 8휔푋 푅

(︁ Δ푗 )︁ 2 Taking into account that 1 − 8휔푋 푅 ≥ 5 , we obtain that 푘 (︂ )︂ 5 ∑︁ 1 ∆푗 푆 ≤ · 휚 . 푘 2 2푘−푗 푋 푅 푗=0 (︀ 휇 )︀푗 푠 Hence, considering that ∆푗 ≤ 2 ∆0 and 휚푋(휏) ≤ 퐶푠푚휏 , we get

푘 푘 ∞ (︂ 푗 )︂푠 (︂ )︂푠 푗푠 (︂ )︂푠 (︂ 푠 )︂푗 5 ∑︁ 퐶푠푚 (︁휇)︁ ∆0 5 퐶푠푚 ∆0 ∑︁ 휇 5 퐶푠푚 ∆0 ∑︁ 휇 푆 ≤ · = · < · . 푘 2 2푘−푗 2 푅 2 2푘 푅 2푗푠−푗 2 2푘 푅 2푠−1 푗=0 푗=0 푗=0 휇푠 Assumption (3) ( 2푠−1 < 1) yields that ∞ 푗 ∑︁ (︂ 휇푠 )︂ 1 = < ∞. 2푠−1 1 − 휇푠 푗=0 2푠−1 Finally, we obtain that (︂ )︂푠 3퐶푠푚 ∆0 1 푆푘 < 휇푠 · · 푘 , 1 − 2푠−1 푅 2 completing the proof of Claim 6.1.  RECTIFIABLE CURVES IN PROXIMALLY SMOOTH SETS 15

References [Ali12] A. R. Alimov. Monotone path-connectedness of 푟-weakly convex sets in spaces with linear embedding. Mathematical Notes, 3(2):21–30, 2012. [AMW12] J. Alonso, H. Martini, and S. Wu. On Birkhoff orthogonality and isosceles orthogonality in normed linear spaces. Aequationes Math., 83(1-2):153–189, 2012. [Bal17] M. V. Balashov. About the gradient projection algorithm for a strongly convex function and a proximally smooth set. J. Convex Analysis, 24(2):493–500, 2017. [BI09] M. V. Balashov and G. E. Ivanov. Weakly convex and proximally smooth sets in Banach spaces. Izv. RAN. Ser. Mat., 73(3):23–66, 2009. [BTZ06] F. Bernard, L. Thibault, and N. Zlateva. Characterizations of prox-regular sets in uniformly convex banach spaces. J. Convex Anal., 13:525–559, 2006. [BTZ11] F. Bernard, L. Thibault, and N. Zlateva. Prox-regular sets and epigraphs in uniformly convex Banach spaces: Various regularities and other properties. Trans. Amer. Math. Soc., 363:2211–2247, 2011. [CSW95] F. H. Clarke, R. J. Stern, and P. R. Wolenski. Proximal Smoothness and Lower–푐2 Property. J. Convex Anal., 2(1):117–144, 1995. [Die75] J. Diestel. Geometry of Banach Spaces - Selected Topics, volume 485. Springer-Verlag Berlin Heidel- berg, 1975. [ES58] Nikolai Vladimirovich Efimov and Sergei Borisovich Stechkin. Some properties of chebyshev sets. In Doklady Akademii Nauk, volume 118, pages 17–19. Russian Academy of Sciences, 1958. [Fed59] Herbert Federer. Curvature measures. Transactions of the American Mathematical Society, 93(3):418–491, 1959. [IM17] G. M. Ivanov and Horst Martini. New moduli for Banach spaces. Annals of Functional Analysis, 8(3):350–365, 2017. [Iva06a] G. E. Ivanov. Weakly Convex Sets and Functions. Theory and Applications. (in Russian). Moscow, 2006. [Iva06b] G. E. Ivanov. Weakly convex sets and their properties. Mathematical Notes, 79:55–78, 2006. [Iva15] G. E. Ivanov. Sharp estimates for the moduli of continuity of metric projections onto weakly convex sets. Izvestiya: Mathematics, 79(4):668, 2015. [Iva17] G. M. Ivanov. Hypomonotonicity of the normal cone and proximal smoothness. Journal Of Convex Analysis, 24(4):27. 1313–1339, 2017. [PRT00] R. Poliquin, R. Rockafellar, and L. Thibault. Local differentiability of distance functions. Trans. Amer. Math. Soc., 352(11):5231–5249, 2000. [Roc81] R. T. Rockafellar. Favorable classes of Lipschitz continuous functions in subgradient optimization. 1981. [Via83] J.-P. Vial. Strong and weak convexity of sets and functions. Math. Ops. Res., 8(2):231–259, 1983.

GRIGORY IVANOV: INSTITUTE OF SCIENCE AND TECHNOLOGY AUSTRIA (IST AUS- TRIA), KLEUSTENEUBURG, 3400, AUSTRIA; LABORATORY OF COMBINATORIAL AND GE- OMETRICAL STRUCTURES, MOSCOW INSTITUTE OF PHYSICS AND TECHNOLOGY, MOSCOW, 141701, RUSSIA Email address: [email protected]

STEKLOV MATHEMATICAL INSTITUTE OF THE RUSSIAN ACADEMY OF SCIENCES, MOSCOW, RUSSIA Email address: [email protected]