Diametric completions
Rolf Schneider
Universit ¨at Freiburg
Szeged Workshop in Convex and Discrete Geometry May 21–23, 2012
LetMbe a nonempty bounded set in a metric space(X, ρ).
ThediameterofMis
diamM:=sup{ρ(x,y) :x,y ∈M}.
Misdiametrically completeif
diam(M∪ {x})>diamM ∀x ∈X \M.
A(diametric) completionofMis any diametrically complete set containingMand with the same diameter.
Every nonempty bounded set has a completion (many, in general); for example, by Zorn’s lemma.
E. Akin: Maximalr-diameter sets and solids of constant width.
arXiv:1003.5824v2
InEuclidean spaces,Meissner (1911)proved:
K is diametrically complete ⇐⇒ K is of constant width.
Recall that a convex bodyK in Euclidean spaceRnisof
constant widthd if any two parallel supporting planes ofK have distanced.
Equivalent:
The support function ofK satisfies
h(K,u) +h(K,−u) =d ∀u.
Equivalent:
K + (−K) =B(o,d), ball of radiusd.
The space of bodies of constant widthd inRnis an infinite-dimensional closedconvexset inKn, the space of convex bodies:
IfK1,K2are bodies of constant widthd, then (1−λ)K1+λK2 (0≤λ≤1) is a body of constant widthd.
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
What about diametrically complete sets in Minkowski spaces?
Some answers are given in:
The space of bodies of constant widthd inRnis an infinite-dimensional closedconvexset inKn, the space of convex bodies:
IfK1,K2are bodies of constant widthd, then (1−λ)K1+λK2 (0≤λ≤1) is a body of constant widthd.
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
What about diametrically complete sets in Minkowski spaces?
Some answers are given in:
Joint work withJos ´e Pedro Moreno:
•Local Lipschitz continuity of the diametric completion mapping.
Houston J. Math.
•Diametrically complete sets in Minkowski spaces.
Israel J. Math.
•The structure of the space of diametrically complete sets in a Minkowski space.
Discrete Comput. Geom.
•Canonical diametric completions in Minkowski spaces.
(work in progress)
Minkowski spaces
X = (Rn,k · k) Minkowski space:
a finite-dimensional real normed space
The normk · kdefinesdistanceρ(x,y) :=kx−yk,width, diameter, unit ballB :={x ∈Rn:kxk ≤1},ballsλB+z.
Bodies ofconstant widthand(diametrically) complete setsare defined as before. Facts:
•Every set of diameterd is contained in a complete set of diameterd.
•K is of constant width =⇒ K is diametrically complete
Meissner stated the converse, and this was believed for more than 50 years (and ‘reproved’), untilEggleston (1965)gave counterexamples. The situation is worse:
Minkowski spaces
X = (Rn,k · k) Minkowski space:
a finite-dimensional real normed space
The normk · kdefinesdistanceρ(x,y) :=kx−yk,width, diameter, unit ballB :={x ∈Rn:kxk ≤1},ballsλB+z.
Bodies ofconstant widthand(diametrically) complete setsare defined as before. Facts:
•Every set of diameterd is contained in a complete set of diameterd.
•K is of constant width =⇒ K is diametrically complete Meissner stated the converse, and this was believed for more than 50 years (and ‘reproved’), untilEggleston (1965)gave counterexamples. The situation is worse:
Theorem 1.For a Minkowski space X , the following are equivalent:
•Every complete set is of constant width.
•The set of complete sets is convex.
•The set of completions of any given set is convex.
Two-dimensional spaces have these properties.
Theorem 2.(Yost 1991, M–S 2010)
Let n≥3. In the space of all n-dimensional Minkowski spaces, a dense open set of Minkowski spaces has the following properties:
•The only bodies of constant width are balls.
•The sum of a complete body and a ball need not be complete.
•The set of completions of a given set need not be convex.
Theorem 1.For a Minkowski space X , the following are equivalent:
•Every complete set is of constant width.
•The set of complete sets is convex.
•The set of completions of any given set is convex.
Two-dimensional spaces have these properties.
Theorem 2.(Yost 1991, M–S 2010)
Let n≥3. In the space of all n-dimensional Minkowski spaces, a dense open set of Minkowski spaces has the following properties:
•The only bodies of constant width are balls.
•The sum of a complete body and a ball need not be complete.
•The set of completions of a given set need not be convex.
Description of complete bodies
K,M ∈ Kn, dimK =n,d >0 Supporting slabofK:
set bounded by two parallel supporting hyperplanes ofK A supporting slab ofK isM-regularif at least one of the bounding hyperplanes of the parallel supporting slab ofM contains a smooth boundary point ofM.
Theorem 3.K is a diametrically complete body of diameter d if and only if(a)and(b)hold:
(a)Every B-regular supporting slab of K has width≤d . (b)Every K -regular supporting slab of K has width=d .
G
On the space of diametrically complete sets
LetDX be the space of translation classes of diametrically complete sets of diameter 2 inX.
Theorem 4.If X = (Rn,k · k)is polyhedral (i.e., the unit ball B is a polytope), thenDX is the union set of a finite polytopal
complex.
The proof uses representations of polyhedral sets introduced byMcMullen (1973)(a variant of the Gale diagram technique).
Corollary.If X is polyhedral, thenDX has only finitely many extreme points.
Open problem:Does this characterize polyhedral norms?
(Yes, ifn=2)
LetD2be the space of diametrically complete sets of diameter 2 inX.
In Euclidean space,D2is convex.
In a typical Minkowski space (in the sense of Baire category), D2is not even starshaped.
A positive result:
Theorem 5.The space D2is contractible.
LetD2be the space of diametrically complete sets of diameter 2 inX.
In Euclidean space,D2is convex.
In a typical Minkowski space (in the sense of Baire category), D2is not even starshaped.
A positive result:
Theorem 5.The space D2is contractible.
γ(K) :=set of completions ofK vmax(K) :=max{V(M) :M∈γ(K)}
γmv(K) :={M ∈γ(K) :V(M) =vmax(K)}
Groemer (1986):γmv(K)consists of translates of one body τ(K) :=M−s(M)for anyM∈γmv(K),sSteiner point loc. Lip. continuity ofγ (M–S 2010)⇒ vmaxis continuous
⇒ τ is continuous
ForK ∈D2andλ∈[0,1], letKλ := (1−λ)K +λBand F(K, λ) :=τ
2 diamKλKλ
+ (1−λ)s(K) +λs(B).
ThenF :D2×[0,1]→D2is continuous andF(K,0) =K, F(K,1) =B. Hence,D2is contractible.
Canonical completions
Is ‘completion’ continuous?
Letγ(K)be the set of all completions ofK ∈ Kn.
Theorem 6.The mappingγ :Kn→ C(Kn)is locally Lipschitz continuous, with respect to the Hausdorff metricδ induced by the norm onKn and the Hausdorff metric∆induced byδ on C(Kn), the space of nonempty compact subsets ofKn. In general,γ is many-valued.
For example, ifK is a segment of lengthd in Euclidean space, then a suitable translate of any body of constant widthd is a completion ofK.
Doesγ have a continuous selection?
The usual constructions of completions involve many arbitrary choices and hence cannot yield continuous completions.
This raises the question for ‘canonical’ completions.
A construction byMaehara (1984)can be slightly generalized.
Definition.ForK ∈ Knof diameterd, let η(K) := \
x∈K
B(x,d), θ(K) := \
x∈η(K)
B(x,d).
Then
µ(K) := 1
2[η(K) +θ(K)] is theMaehara setofK.
Doesγ have a continuous selection?
The usual constructions of completions involve many arbitrary choices and hence cannot yield continuous completions.
This raises the question for ‘canonical’ completions.
A construction byMaehara (1984)can be slightly generalized.
Definition.ForK ∈ Knof diameterd, let η(K) := \
x∈K
B(x,d), θ(K) := \
x∈η(K)
B(x,d).
Then
µ(K) := 1
2[η(K) +θ(K)]
is theMaehara setofK.
Always true:µ(K)is atight coverofK (i.e., containsK and has diameterd).
In Euclidean spaces:µ(K)is of constant width, and hence a completionofK.
Where does this work?
Definition.The normk · kwith unit ballB has thes-propertyif B∩(B+x)is a summand ofB, for eachx withkxk ≤1.
Maehara (1984), Sallee (1987), Balashov & Polovinkin (2000), Karas ¨ev (2001):
Theorem.The Maehara setµ(K), for any K ∈ Kn, is of constant width and hence a completion of K , if and only if the norm of X has the s-property.
Theorem 7.Suppose that the norm of X has the s-property, and let2C denote the Jung constant of X .
Let K,L∈ Knbe convex bodies with
δ(K,L)≤≤ 1
3(1−C)min{dK,dL}.
Then
δ(µ(K), µ(L))≤ 7−C
1−C≤ 13
2 (n+1). In particular, if X is a Euclidean space, then
δ(µ(K), µ(L))≤20,
and if X is a two-dimensional Minkowski space, then
δ(µ(K), µ(L))≤ 15 2 .
Two new properties of the Maehara completion in Euclidean spaces:
Recall thatγ(K)denotes the set of all completions ofK. Theorem 8.The Maehara completion of K is a metric centre of γ(K), that is, it minimizes the maximal Hausdorff distance from the elements ofγ(K).
The Maehara completion of a convex body is at least as smooth as the body itself:
Theorem 9.Every normal cone of the Maehara completion µ(K)is contained in some normal cone of K .
Back to Minkowski spaces:
The only known examples of Minkowski spaces with the s-property are Euclidean spaces, two-dimensional Minkowski spaces, and⊕∞sums of such spaces.
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
Therefore, forgeneral Minkowski spaces, a different canonical completion procedure is needed.
For this, we extend a Euclidean method ofReinhardt (1922) (n=2) andB ¨uckner (1936)(n=3).
In a first step, we replace eachK by 1
2[η(K) +K],
which is a tight cover ofK and has inradius at least (2(n+1))−1(diamK).
The only known examples of Minkowski spaces with the s-property are Euclidean spaces, two-dimensional Minkowski spaces, and⊕∞sums of such spaces.
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Therefore, forgeneral Minkowski spaces, a different canonical completion procedure is needed.
For this, we extend a Euclidean method ofReinhardt (1922) (n=2) andB ¨uckner (1936)(n=3).
In a first step, we replace eachK by 1
2[η(K) +K],
which is a tight cover ofK and has inradius at least (2(n+1))−1(diamK).
The generalized B ¨uckner completion
A convex bodyK of diameterd is complete if and only if K = \
x∈K
B(x,d)
(thespherical intersection property).
Aiming at completing a convex bodyK of diameterd, one is therefore tempted to consider
η(K) := \
x∈K
B(x,d),
thewide spherical hullofK.
However, in general,diamη(K)>d.
B ¨ucknerhad the idea to consider a ‘one-sided’ version of the wide spherical hull, namely
Cu(K) :=η(K)∩Z+(K,u) for givenu6=o, with
Z+(K,u) :={x+λu :x ∈K, λ≥0}.
Cu(K)is a tight cover ofK!
ButCu(K)is generally not complete.
However, it is‘partially complete’:
Each ‘upper’ boundary point (w.r.t. u) is the endpoint of a diameter segment ofCu(K).
{,t(
§g+NI
9{U
Finitely many iterations, formfixed directionsu1, . . . ,um (where mdepends only onn) yield a completionC(K)ofK.
We callC thegeneralized B ¨uckner completion.
Theorem 10.The generalized B ¨uckner completion is locally Lipschitz continuous.
Perfect norms
Recall:
Theorem 1.For a Minkowski space X , the following are equivalent:
•Every complete set is of constant width.
•The set of complete sets is convex.
•The set of completions of any given set is convex.
Two-dimensional spaces have these properties.
Definition.(Karas ¨ev) A Minkowski space with these properties and its norm are calledperfect.
Eggleston (1965)andChakerian & Groemer (1983)have asked for a determination of all perfect Minkowski spaces.
This is still open.
We know fromMaeharaandKaras ¨ev:
Theorem.If the norm of X has the s-property, then X is perfect.
Conjectures.LetK be a convex body of dimension≥3 with the s-property. IfK is either smooth (R. S. 1974) or strictly convex (Karas ¨ev 2001), then it is an ellipsoid.
Theorem(Karas ¨ev 2001).A Minkowski space with a strictly convex norm is perfect if and only if its norm has the s-property.
For general norms, this is not true.
Example:
A new necessary condition:
Theorem 11.If B is the unit ball of a perfect norm, then 1
2 B∩(B+x) is a summand of B for all x withkxk ≤1.
D ¨urer’s (1514)octahedron shows that the constant 12 is best possible.
The proof of Theorem 11:
Lemma.Let K,L∈ Kn. If to each supporting hyperplane H of K there are a point x∈H∩K and a vector t ∈Rnsuch that x ⊂L+t ⊂K , then L is a summand of K .
The steps to prove Theorem 11:
kuk ≤1,Harbitrary support plane toB∩(B+u) y ∈H∩B∩(B+u)
S:= 12((B∩(B+u)−y) +y ⇒ diamS∪ {o,u} ≤1 C:=completion ofS∪ {o,u} ⇒ C ⊂B∩(B+u)
⇒ HsupportsC aty
H0:=support plane ofCparallel toH k · kis perfect ⇒ dist(H,H0) =1 Lety0∈C∩H0 ⇒C⊂B+y0
HsupportsB+y0, sincedist(H,y0) =1
SinceH was arbitrary, the lemma shows that 12(B∩(B+u))is a summand ofB.