Distance Distributions and Inverse Problems For Metric Measure
Distance Distributions and Inverse Problems For Metric Measure
Spaces
Facundo Mémoli1 and Tom Needham2
1
Department of Mathematics and Department of Computer Science and Engineering, The
Ohio State University, [email protected]
2
Department of Mathematics, Florida State University, [email protected]
arXiv:1810.09646v3 [math.MG] 1 Jan 2021
January 5, 2021
Abstract
Applications in data science, shape analysis and object classification frequently require com-
parison of probability distributions defined on different ambient spaces. To accomplish this, one
requires a notion of distance on a given class of metric measure spaces—that is, compact metric
spaces endowed with probability measures. Such distances are typically defined as comparisons
between metric measure space invariants, such as distance distributions (also referred to as shape
distributions, distance histograms or shape contexts in the literature). Generally, distances defined in
terms of distance distributions are actually pseudometrics, in that they may vanish when comparing
nonisomorphic spaces. The goal of this paper is to set up a formal framework for assessing the
discrimininative power of distance distributions, i.e., the extend to which these pseudometrics fail to
define proper metrics. We formulate several precise inverse problems in terms of these invariants and
answer them in several categories of metric measure spaces, including the category of plane curves,
where we give a counterexample to the Curve Histogram Conjecture of Brinkman and Olver, the
categories of embedded and Riemannian manifolds, where we obtain sphere rigidity results, and the
category of metric graphs, where we obtain a local injectivity result along the lines of classical work
of Boutin and Kemper on point cloud configurations. The inverse problems are further contextual-
ized by the introduction of a variant of the well-known Gromov-Wasserstein distance on the space
of metric measure spaces, which is inspired by the original Monge formulation of optimal transport.
Contents
1 Introduction 2
1.1 Overview of Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Distance Distributions 7
2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Distance Distributions and Associated Pseudometrics . . . . . . . . . . . . . . . . . . . 10
2.3 Inverse Problems for Metric Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . 15
3 Gromov-Monge Quasi-Metrics 16
3.1 Gromov-Monge Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Subcategories of Metric Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Pseudometric Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1
4 Characterization Results for Embedded Manifolds 20
4.1 Plane Curves and the Curve Histogram Conjecture . . . . . . . . . . . . . . . . . . . . 21
4.2 Rigidity of the Unit Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Injectivity of the Local Distribution for Curves . . . . . . . . . . . . . . . . . . . . . . 25
4.4 Non-Injectivity for Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.5 Sphere Characterization for Embedded Manifolds . . . . . . . . . . . . . . . . . . . . . 28
7 Discussion 59
1 Introduction
Classical optimal transport (OT) problems deal with finding an optimal way to match two probability
measures µ and ν defined on the same ambient metric space Z. There are two historical formulations
of OT: the original Monge formulation [73] and the Kantorovich formulation [58]. The former involves
minimizing a certain cost over all measure preserving transformations between the measures, whereas
the latter introduces a convex relaxation which enlarges the set of admissible mappings to those proba-
bility measures on the product space Z × Z whose marginals coincide with µ and ν; such a measure is
called a coupling of µ and ν. Optimal transport has long been an active field of study in pure mathemat-
ics (standard references are [84,94]) and has recently become popular for machine learning applications,
as it is a natural way to compare data distributions [62, 79].
The convex relaxation corresponding to the Kantorovich formulation of OT has been adapted to the
setting when one wishes to compare not just two probability measures defined on the same ambient
space, but to the more general case when one wishes to compare triples of the form X = (X, dX , µX )
where (X, dX ) is a compact metric space and µX is a Borel probability measure on X. These triples
are called metric measure spaces (mm-spaces for short) or weighted metric spaces, and the resulting
notion of dissimilarity between pairs of such triples is referred to as Gromov-Wasserstein (GW) distance
[72] (also referred to as distortion distance in [88]). Gromov-Wasserstein distance has gained recent
popularity in the machine learning community as a way to compare datasets which do not live in the
same ambient space [8, 28, 43, 56, 80, 96–98].
We use the notation Mw for the space of isomorphism classes of weighted metric spaces (an isomor-
phism of mm-spaces being a measure-preserving isometry). The Gromov-Wasserstein distance defines
a metric on Mw ; however, the distance is not tractable to compute exactly, even in the setting of dis-
crete mm-spaces. One therefore relies on pseudometric estimates of GW distance given by comparing
certain distributional invariants of mm-spaces. These invariants, defined in terms of volume growth in
2
a mm-space, are useful in their own right, as they can be used to vectorize mm-space data in a princi-
pled manner. Informally, the goal of this paper is to characterize how badly these pseudometrics fail
to define proper metrics. From a more mathematical perspective, we address the inverse problem of
characterizing the extent to which volume growth data determines a space.
To make the goals of the paper more precise, define the local shape measure of an mm-space X =
(X, dX , µX ) tobe the function dhX from
X into the set P(R) of probability measures on R defined by
x 7→ dX (x, ·) # µX , with dX (x, ·) # µX denoting the pushforward of µX by the function dX (x, ·) :
X → R. The global shape measure dHX of X is the average of the local shape measure dhX ; that is,
for every measurable A ⊂ R,
Z
dHX (A) := dhX (x)(A) µX (dx). (1)
X
One can then pose any number of problems on the discriminatory power of these invariants in classes of
mm-spaces; e.g.:
1. For plane curves X and Y , considered as mm-spaces with extrinsic Euclidean distance and nor-
malized arclength measure, does dHX = dHY imply X and Y differ by a rigid motion?
3. For compact metric graphs X and Y with uniform measures, does the existence of a map φ : X →
Y such that dhX = dhY ◦ φ imply X and Y are isomorphic?
Questions regarding the use of volume-growth information to characterize a space have been studied
frequently in the Riemannian setting [11, 12, 22, 30, 31, 40, 48, 88], where infinitesimal volume growth is
related to scalar curvature, as well as for certain mm-spaces embedded in Euclidean space and endowed
with extrinsic distance [24,27]. Inverse problems with a similar flavor have also been studied recently in
the topological data analysis literature, where the goal is to characterize rougher spaces such as metric
graphs via families of topological signatures [16, 41, 44, 77, 78]. In this paper we bridge the gap between
these perspectives and establish an overarching framework for studying the recovery of general metric
measure spaces from volume growth data.
We provide a counterexample to Problem 1, thereby resolving the Curve Histogram Conjecture of
Brinkman and Olver [27]. Problem 2 can be viewed as a problem on sphere rigidity, in the tradition of
similar problems in differential geometry. We answer this question, and a similar question for embedded
manifolds with extrinsic distance, under additional curvature assumptions. Problem 3 is reminiscent of
the classical Graph Reconstruction Conjecture of Kelly [60] and Ulam [90], and is also of applied interest
as functions similar to dhX have recently been incorporated into algorithms for graph classification and
analysis [85, 89]. The answer to this problem is more subtle and depends on assumptions about the
regularity of the map φ and the topologies of the graphs. The following subsection provides more
details on the main results of the paper.
At a high level, one of the main goals of this paper is to formalize the study of inverse problems
in various categories of mm-spaces; e.g., Riemannian manifolds, embedded manifolds, metric graphs
equipped with measure-preserving morphisms of various prescribed regularities. To this end, we for-
mulate precisely several specific inverse problems of this type in Section 2. We contextualize these
category-specific inverse problems by introducing a novel variant of Gromov-Wasserstein distance,
dubbed Gromov-Monge distance, in Section 3. This variant is inspired by the original Monge formula-
tion of the classical optimal transport problem and can be restricted to specified classes of morphisms
between mm-space objects. Our inverse problems are resolved for several categories of mm-spaces in
Sections 4 and 5, which cover smooth manifolds, and 6, which covers metric graphs. Although these
sections treat similar problems, the techniques used are fairly distinct. Sections 4 and 5 primarily rely
on manipulations of Taylor expansions of distributional invariants and other results from the differential
3
geometry literature. The results of Section 6 are obtained by employing more combinatorial tools and
some results from the field of computational topology. These sections can be read independently. We
elucidate numerous open problems along the way, providing many directions for future research on the
inverse problems laid out in this paper.
Section 2. This section outlines the basic notation, concepts and main problems considered in the rest
of the paper. Here we introduce the notion of subcategories of mm-spaces. A subcategory of mm-spaces
C is a category whose object set ObjC is a collection of isomorphism classes of mm-spaces and whose
morphism set MorC consists of measure-preserving maps, considered up to natural equivalence. For
example, one could consider the subcategory of Riemannian manifolds, whose objects are equivalence
classes of Riemannian manifolds endowed with geodesic distance and normalized Riemannian volume
and whose morphisms are equivalence classes of smooth measure-preserving maps. Throughout the
paper, we abuse terminology slightly and use mm-spaces as representatives for elements of ObjC (which
are technically equivalence classes thereof), and likewise use maps to represent elements of MorC .
Given a subcategory of mm-spaces C, we pose the following inverse problems, stated here somewhat
informally:
(IP1) Global Injectivity. For X , Y ∈ ObjC , does the existence of φ ∈ MorC such that dhX = dhY ◦ φ
imply X ≈ Y? More strongly, does dHX = dHY imply X ≈ Y?
(IP2) Homotopy Type Characterization. Does the existence of φ ∈ MorC such that dhX = dhY ◦ φ
(more strongly, does dHX = dHY ) imply the underlying spaces X and Y are homotopy equiva-
lent?
(IP3) Sphere Rigidity. If ObjC has a notion of a unit sphere S, does dHX = dHS imply X ≈ S?
(IP4) Local Injectivity. For X ∈ ObjC , does there exist X > 0 such that the existence of φ ∈ MorC
with dhX = dhY ◦φ implies X ≈ Y, provided Y is X -close to X in Gromov-Hausdorff distance?
Problem (IP1) is an inverse problem in the sense that it asks whether the map taking a mm-space from a
prescribed class to its global shape measure is injective, i.e., whether it has a left inverse. Problems (IP2)
and (IP3) are relaxations of the first problem, generalizing problems studied classically in differential
geometry to handle categories whose objects are not necessarily smooth. In particular, (IP3) can be seen
as a generalization of the sphere rigidity problems which are the focus of much research in classical
differential geometry (e.g., Aleksandrov’s Theorem [6] and Cheng’s Rigidity Theorem [37]). Problem
(IP4) asks a localized version of (IP1).
The precise statements of these inverse problems, provided in Section 2.3, are given in terms of
pseudometrics LC C
h and LH on ObjC , defined using the local and global shape measures (or rather cumu-
lative versions thereof). The notation here is chosen because we show that these pseudometrics provide
(tractably computable) lower bounds on Gromov-Wasserstein distance when the category C is all of
Mw .
dC
GM,p (X , Y) = inf costp (φ), (2)
φ
4
Figure 1: Counterexample to the Curve Histogram Conjecture of Brinkman and Olver. These curves
address the inverse problem on global injectivity (IP1) in the category of simple closed plane curves.
That is, they have the same global shape measures, but are nonisomorphic as mm-spaces.
where the infimum is over measure-preserving maps φ : X → Y belonging to MorC and costp is an
Lp -type cost function measuring the overall geometric distortion incurred by φ (e.g., φ is an isometry if
and only if costp (φ) = 0). The main theorem of this section describes the sense in which each dC GM,p is
a distance.
Theorem (Theorem 1 and Remark 3.5). For any p ≥ 1 the function dCGM,p defines an extended quasi-
metric on ObjC .
This means that dCGM,p is sometimes infinite and that it is not necessarily symmetric. Categories
of mm-spaces where finiteness and/or symmetry are guaranteed are provided in Section 3.2. It is then
shown in Section 3.3 that LC C C
H and Lh provide lower bounds for dGW,1 .
Remark 1.1. The main purpose of Section 3 is to provide context for our inverse problems by establish-
ing these lower bounds. In fact, the results of Sections 3, 4, 5 and 6 are essentially disjoint and these
sections can be read independently.
Further theoretical results on Gromov-Monge distance are included in an appendix. These results
are not directly related to the rest of the paper, but they are included to provide more context for how
Gromov-Monge distances fit into the existing literature.
Sections 4 and 5. Here we begin our study of inverse problems (IP1)–(IP4) concerning the recovery
of mm-spaces from volume growth data. These sections are focused on smooth manifolds, considered in
Section 4 as mm-spaces via ambient metric structure coming from an embedding into Euclidean space
and in Section 5 via geodesic distance with respect to a choice of Riemannian metric. We state our main
results here somewhat informally and in each case refer back to one of the inverse problem to which the
result is relevant.
Section 4 begins with plane curves (considered as mm-spaces with extrinsic distance and normalized
arclength measure). The Curve Histogram Conjecture of Brinkman and Olver [27] proposes that the
global distance distribution is a complete invariant of a (sufficiently regular) plane curve X in the sense
that dHX = dHY implies that X and Y differ by a rigid motion; i.e., that (IP1) has is answered in
the afirmative. We adapt a counterexample of Mallows and Clarke to a conjecture by Blaschke [67] to
resolve the Curve Histogram Conjecture and prove several refinements.
Theorem. 1. The curves depicted in Figure 1 give a counterexample to the Curve Histogram Con-
jecture. Moreover, for any > 0 there exists a pair of non-isometric curves X and Y which are
both -close to the unit circle in Hausdorff distance for which HX = HY (Proposition 4.2 and
Corollary 4.5) (IP1).
2. The unit circle is uniquely determined by its global distance distribution among twice differen-
tiable plane curves (Proposition 4.7) (IP3). On the other hand, there exists a one-parameter
5
Figure 2: Examples of metric graphs. (Left) Neuron M1KO28 from the NeuroMorpho.Org database
[10, 13]. Neuron structures are frequently represented as metric graphs embedded in R3 . (Right) Road
network from [4], naturally represented as a metric graph embedded in R2 .
family of distributions on the circle whose global shape measures are the same (Example 5.8),
giving a negative answer to (IP4) in this setting.
3. If X and Y are twice differentiable curves and there exists a twice differentiable “infinitesimally
measure-preserving” map φ : X → Y with dhX = dhY ◦ φ, then X and Y are isometric
(Proposition 4.10) (IP1).
Leveraging Taylor expansions of volume growth functions, we derive rigidity results for higher
dimensional spheres.
Theorem. The unit sphere in Rn+1 is uniquely determined by its global distance distribution among:
1. smooth hypersurfaces of Rn+1 (with extrinsic distance, normalized uniform measure) satisfying
certain curvature conditions (Theorem 4) (IP3),
In the Riemannian setting, this result can be pushed further to a characterization result on other
constant curvature manifolds.
Theorem (Theorem 7). A 2-dimensional Riemannian manifold with constant Gauss curvature κ is de-
termined by its global distance distribution up to diffeomorphism (IP2). Moreover, if κ > 0 then the
manifold is determined up to isometry (IP3).
In the κ = 0 (flat torus) case of the above theorem, we are able to obtain a complete characterization
of nonisometric surfaces with the same distance distributions (Proposition 5.16). We also show that the
result can be refined further for local distance distributions: generic Riemannian metrics are determined
up to isometry by their local shape measures (Proposition 5.10) (IP1).
Section 6. Here we consider volume growth inverse problems for the class of metric graphs; i.e.,
geometric realizations of compact 1-dimensional simplicial complexes endowed with geodesic distance
and uniform measure. Metric graphs serve as models for road networks [4], neurons [63] (see Figure
2), blood vessel systems [35, 36] and generally any application dealing with intersecting filamentary
structures (see [1] for many more examples). While global distance distributions cannot distinguish
metric graphs in general (Example 6.12), we show that local distance distributions determine metric
graphs in the following sense.
6
Theorem (Theorem 9). For metric graphs G and H, with G not isomorphic to the circle graph, if there
exists a continuous measure-preserving map φ : G → H with dhG = dhH ◦ φ, then φ is an isomorphism
(IP1).
Theorem (Theorem 8). If a metric graph G satisfies dHG = dHC , where C is a circle endowed with
arclength distance and normalized length measure, then G ≈ C (IP3).
Theorem (Theorem 10). Let G and H be metric graphs. If there exists a measure preserving map
φ : G → H such that dhG = dhH ◦ φ then G and H are homotopy equivalent (IP2).
A more technical local injectivity result is obtained in the subcategory of metric trees (contractible
metric graphs).
Theorem (Theorem 11). Local distance distributions distinguish trees locally in the sense of (IP4): for
every metric tree T there exists T > 0 such that if T 0 is another metric tree which is T -close to T in
Gromov-Hausdorff distance and such that dhT 0 = dhT ◦ φ for some measure-preserving map φ, then
T0≈T.
The proof establishes a connection to a distance between metric trees first considered in the com-
putational topology community called the interleaving distance [3, 74]. Counterexamples to stronger
versions of this theorem are also provided.
Section 7. The paper concludes with a discussion of remaining open questions related to Gromov-
Monge distance, volume growth invariants and inverse problems for mm-spaces. We draw the reader’s
attention to this section, as the various flavors of injectivity results we are able to obtain in several cate-
gories of mm-spaces are conveniently summarized in Table 1. There are many open questions naturally
arising from our work, which are pointed out in “Question” environments throughout the paper. We hope
that this paper serves as a road map for continued research in the types of inverse problems formalized
here.
2 Distance Distributions
2.1 Preliminaries
Metric Measure Spaces
Let M denote the collection of isometry classes of compact metric spaces (X, dX )—we will refer to
a metric space by its underlying set X when it is clear that a metric has been fixed. In this paper,
we primarily consider metric spaces which have been endowed with extra data in the form of a prob-
ability measure. A triple (X, dX , µX ) where (X, dX ) is a compact metric space and µX is a Borel
probability measure on X is referred to as a metric measure space (mm-space) and will be denoted by
X = (X, dX , µX ). Metric measure spaces are also referred to as weighted metric spaces in the liter-
ature, and we will use this terminology interchangeably. For convenience, we always assume that the
support of µX is all of X.
Let P(X) denote the set of all fully-supported Borel probability measures on X. For a Borel mea-
surable map φ : X → Y between metric spaces and a measure µX on X, we will use the notation φ# µX
for the pushforward of µX by φ. This is defined on a Borel subset A ⊂ Y by
7
An isomorphism between mm-spaces X and Y is a measure-preserving map φ : X → Y which is also a
metric isometry; i.e., φ# µX = µY and dX = dY ◦ (φ × φ). When such an isomorphism exists, we write
X ≈ Y. The collection of isomorphism classes of weighted metric spaces will be denoted Mw .
Let πX : X × Y → X and πY : X × Y → Y be coordinate projection maps. For mm-spaces X and
Y, define
U(µX , µY ) := {µ ∈ P(X × Y ) | (πX )# µ = µX and (πY )# µ = µY }
to be the set of all probability measures on X ×Y with marginals µX and µY , which are called couplings
between µX and µY . The set of couplings is never empty, as it contains the product measure µX ⊗ µY ,
which is uniquely determined by the property
µX ⊗ µY (A × B) = µX (A)µY (B)
T (µX , µY ) := {φ : X → Y | φ# µX = µY }
denote the set of all measure-preserving maps between X and Y . Notice that it could be that T = ∅; for
example, there is never a measure-preserving map from the mm-space containing a single point to an
mm-space with larger cardinality. Given any φ ∈ T (µX , µY ), one can consider the probability measure
µφ on X × Y given by
µφ := (idX × φ)# µX , (3)
where idX is the identity map on X. It is straightforward to check that µφ ∈ U(µX , µY ).
Optimal Transport
In this paper, we consider comparisons between mm-spaces. To build up methods for doing so, we begin
with a classical notion of comparing probability distributions relative to some auxilliary cost function.
Let µX and µY be Borel probability measures on Polish spaces X and Y , respectively, and let c :
X×Y → R be some cost function. The Monge optimal transport problem seeks to solve the optimization
problem Z
inf c(x, φ(x)) µX (dx).
φ∈T (µX ,µY ) X
The problem was originally formulated by Monge in the 18th century [73] in the context of transporting
mass from one location to another, where X = Y = Rn and c(x, y) is `1 distance.
The Monge optimal transport problem is potentially ill-posed; as was observed above, the set
T (µX , µY ) can be empty. This problem can be relaxed to the more well-behaved Kantorovich-
Rubinstein-Wasserstein optimal transport problem, which seeks to solve
Z
inf c(x, y) µ(dx × dy). (4)
µ∈U (µX ,µY ) X×Y
8
Gromov-Hausdorff Distance
The classical notion of Hausdorff distance between compact subsets of an ambient metric space Z can
be vastly generalized to give a metric on M. The Gromov-Hausdorff distance dGH is defined by
where the infimum is taken over all ambient metric spaces (Z, dZ ) and isometric embeddings φX : X →
Z and φY : Y → Z, and where dZ H denotes Hausdorff distance in Z.
The Gromov-Hausdorff distance can be reformulated in several ways, and we will make particular
use of a reformulation in terms of the distortion of a correspondence. For compact metric spaces X and
Y , let ΓX,Y : X × Y × X × Y → R denote the distortion map, defined by
A correspondence between sets X and Y is a subset R ⊂ X × Y such that the coordinate projection
maps πX : X × Y → X and πY : X × Y → Y give surjections when restricted to R. The set of all
correspondences between X and Y is denoted R(X, Y ). We then have the following reformulation of
Gromov-Hausdorff distance [29]:
1
dGH (X, Y ) = inf sup ΓX,Y (x, y, x0 , y 0 ). (7)
2 R∈R(X,Y ) (x,y),(x0 ,y0 )∈R
One might notice that the righthand side of (7) takes the form of an infimum of instances of an L∞ norm
of the function ΓX,Y , suggesting a relaxation to an Lp norm. Before formulating such a relaxation, we
review some basic ideas from optimal transport.
Gromov-Wasserstein Distance
The Gromov-Wasserstein p-distance dGW,p gives a Kantorovich-style relaxation of the formulation (7)
of Gromov-Hausdorff distance. It is defined on mm-spaces X and Y by
9
• a pseudometric if it satisfies the metric axioms, except that we allow dX (x, y) = 0 for some pairs
x 6= y;
• a quasimetric if it satisfies the metric axioms, except that we allow dX (x, y) 6= dX (y, x) for some
pairs x and y;
• an extended metric if it satisfies the metric axioms, except that we allow dX (x, y) = ∞ for some
pairs x and y. We can similarly define extended pseudometrics and extended quasimetrics.
For a metric space (X, dX ), BX (x, r) denotes the metric ball centered at x ∈ X with radius r. We use
U to denote the closure of a subset U of a topological space.
HX : R≥0 → R≥0
r 7→ µX ⊗ µX {(x, x0 ) ∈ X × X | dX (x, x0 ) ≤ r} .
Distance distributions (sometimes referred to as shape distributions or distance histograms) are a stan-
dard tool for summarizing metric spaces. They have been used for classification of geometric ob-
jects [23, 76] and their mathematical properties have been studied in a variety of contexts [18, 24, 27].
The global distance distribution is related to the global shape measure dHX defined in the introduction
by the formula Z r
HX (r) = dHX (ds).
0
For p ≥ 1, define the function LH,p : Mw × Mw → R by
Since the global shape measures dHX and dHY are (compactly supported) probability distributions on
R, and Kantorovich optimal transport has a closed form solution in the 1-dimensional case given in terms
of generalized inverses of cumulative distributions (see [93, Remark 2.19]), yielding the more explicit
formula Z 1 1/p
−1 −1 p
LH,p (X , Y) = HX (u) − HY (u) du ,
0
The p = 1 case plays a special role, since one can show [72] that in this case the formula simplifies to
Z ∞
LH (X , Y) := LH,1 (X , Y) = |HX (u) − HY (u)| du.
0
10
The quantity LH,p (X , Y) (particularly in the p = 1 case) gives an intuitive notion of distance be-
tween two mm-spaces. To be precise, it is easy to check that LH,p is symmetric in its arguments, that
LH,p (X , Y) ≥ 0 for all X , Y, that if X and Y are isomorphic then LH,p (X , Y) = 0, and that LH,p
satisfies the triangle inequality. The obstruction to LH,p defining a proper metric on Mw is that it can
vanish for nonisomorphic mm-spaces.
Example 2.2. A simple example of nonisomorphic mm-spaces with the same distance distributions
appears in [20] in the context of difference sets of integers. Define mm-spaces X , Y by taking
X = {0, 1, 4, 10, 12, 17} and Y = {0, 1, 8, 11, 13, 17} and endowing each set with Euclidean distance
and uniform measures. One can easily check that X 6≈ Y, yet LH (X , Y) = 0.
The L in our notation LH,p is chosen because these pseudometrics serve as lower bounds for
Gromov-Wasserstein p-distances. The following is Corollary 6.2 of [72].
hX : X × R+ → [0, 1]
(x, r) 7→ µX BX (x, r) .
Local distance distributions (and closely related invariants, such as shape contexts) have also appeared
frequently in the shape analysis literature (e.g., [15,47,86]). For each x ∈ X, the function r 7→ hX (x, r)
measures volume growth, localized at x, and inverse problems about these volume growth functions
similar to those considered in the present paper have appeared previously in [48, 88]. The local distance
distribution is related to the the global distance distribution via
Z
HX (r) = hX (x, r) µX (dx)
X
As in the case of global distance distributions, local distance distributions can be used to define a
family of pseudometrics on Mw . These pseudometrics necessarily have a more complicated structure,
since comparison of local distributions requires some registration between elements of X and Y . For
each p ≥ 1, we define LUh,p : Mw × Mw → R by
Z 1/p
LUh,p (X , Y) := inf dRW,p (dhX (x), dhY (y))p µ(dx × dy) .
µ∈U (µX ,µY ) X×Y
Remark 2.5. This function is well-defined, since one can show that the function X × Y → R defined by
11
is continuous, hence Borel measurable. This follows from the observation that the map (X, dX ) →
(P(R), dRW,p ) : x 7→ dhX (x) is continuous. Indeed, by triangle inequality, if dX (x, x0 ) ≤ δ then
Since Wasserstein ∞-distance upper bounds Wasserstein p-distances for p < ∞, we conclude
Once again using the special structure of optimal transport on the real line, one obtains the more
explicit formula
Z Z 1 1/p
p
LUh,p (X , Y) = inf h−1
X (x, u) − h−1
Y (y, u) du µ(dx × dy) .
µ∈U (µX ,µY ) X×Y 0
This Kantorovich-style optimization over measure couplings suggests an alternative Monge-style for-
mulation over measure preserving maps, and we accordingly define
Z Z 1 1/p
p
LTh,p (X , Y) := inf h−1
X (x, u) − h−1
Y (φ(x), u) du µX (dx) .
φ∈T (µX ,µY ) X 0
Similar to the global distribution setting, the p = 1 versions of LUh,p and LTh,p can be simplified. This
special case has the added benefit that it provides a more direct connection to optimal transport. Given
X , Y ∈ Mw , one defines the following local distribution cost function cX ,Y : X × Y → R+ by
Z ∞
cX ,Y (x, y) := hX (x, t) − hY (y, t) dt.
0
Following [72], we can then show that
Z
LUh (X , Y) := LUh,1 (X , Y) = inf cX ,Y (x, y) µ(dx, dy)
µ∈U (µX ,µY ) X×Y
and Z
LTh (X , Y) := LTh,1 (X , Y) = inf cX ,Y (x, φ(x)) µX (dx).
φ∈T (µX ,µY ) X
Thus the p = 1 functions compare mm-spaces by searching for a coupling or a transport map, respec-
tively, which optimally preserve mm-space structure at the level of local volume growth.
It is straightfoward to show that LUh,p (respectively, LTh,p ) satisfies the axioms of a metric (respec-
tively, extended quasimetric) on Mw except that it is nonvanishing for nonisomorphic spaces. Indeed,
we the counterexample below shows that this is not the case in general.
Example 2.6. The spaces in X and Y with X = {0, 1, 4, 10, 12, 17} and Y = {0, 1, 8, 11, 13, 17}
from Example 2.2, which confound the pseudometric LH , are distinguished by both LUh and LTh . This
is easy to conclude from the observation that hX (0, r) is piecewise constant, with discontinuities at
r = 1, 3, 10, 12, 17, and that this pattern is not replicated by hY (y, r) for any y ∈ Y .
Example 2.7. The following example from [72] shows that LTh (and hence LUh ) does not distinguish
nonisomorphic mm-spaces in general. Let X = Y = {a1 , . . . , a9 } with the same metric given in matrix
12
form by
0 1 1 2 2 2 2 2 2
1 0 1 2 2 2 2 2 2
1 1 0 2 2 2 2 2 2
2 2 2 0 1 1 2 2 2
dX 2
= dY = 2 2 1 0 1 2 2 2
2 2 2 1 1 0 2 2 2
2 2 2 2 2 2 0 1 1
2 2 2 2 2 2 1 0 1
2 2 2 2 2 2 1 1 0
and probability measures given in vector form by
23 1 67 2 1 2 4 1 3
µX = , , , , , , , ,
140 105 420 15 15 15 21 28 28
and
23 1 67 2 1 2 4 1 3
µY = , , , , , , , , ,
140 105 420 15 15 15 21 28 28
respectively. One can check that LTh (X , Y) = 0 is realized by the unique measure preserving permuta-
tion of X = Y (i.e., the permutation taking the values of µX to µY ), but that this map is not an isometry,
from which we conclude X 6≈ Y (see [72, Example 5.6] for details).
Proposition 2.8. For all p ≥ 1, LUh,p defines a pseudometric and LTh,p defines an extended pseudometric
on Mw .
As in the global distance distribution setting, our choice of notation reflects the fact that LUh,p defines
a lower bound on Gromov-Wasserstein distance. The following result is implicit in Corollary 6.3 of [72].
In the following section we will extend this result to treat the extended pseudometric LTh,p .
Remark 2.10. This is actually a slightly tighter bound than the one presented in [72], where LUh,p is
further lower bounded using Jensen’s inequality.
We can further show that the lower bounds LUh,p and LH,p are themselves comparable. The following
result is novel—it was not obtained in the original paper [72].
Proof. The second inequality is Proposition 2.4, so it remains to prove the first inequality. For each
(x, y) ∈ X × Y , via [93, Theorem 2.18], choose γx,y ∈ U(µX , µY ) realizing dRW,p (dhX (x), dhY (y))
explicitly as that probability measure on R2 with cumulative distribution function equal to Fγx,y (s, t) :=
min{hX (x, s), hY (y, t)} for each (s, t) ∈ R2 . Also, using Lemma 2.1, choose µ ∈ U(µX , µY ) realizing
Z p
inf dRW,p (dhX (x), dhY (y)) µ(dx × dy).
µ∈U (µX ,µY ) X×Y
13
Then we have
Z Z Z
R
p
dW,p (dhX (x), dhY (y)) µ(dx × dy) = |s − t|p γx,y (ds × dt) µ(dx × dy)
X×Y X×Y R×R
Z
= |s − t|p µF (ds × dt), (9)
R×R
Since the integrand is continuous, as follows from (8), and µ is Borel, this function is well-defined. We
then define µF to be the probability measure associated to F . We claim that this is given by the formula
Z
µF (A) = γx,y (A) µ(dx × dy).
X×Y
for measurable A ⊂ R × R, hence (9) holds. One can check this claim easily on rectangles of type
(s, s0 ] × (t, t0 ] using the fact that the integrand of F (s, t) is equal to the cumulative distribution Fγx,y for
γx,y and the general formula holds by standard arguments.
Next we claim that µF ∈ U(dHX , dHY ). Indeed, for measurable S ⊂ R, we have
Z Z
µF (S × R) = γx.y (S × R)µ(dx × dy) = dhX (x)(S) µ(dx × dy)
X×Y X×Y
Z
= dhX (x)(S) µX (dx) = dHX (S),
X
where we have applied marginal constraints for γx,y and π. A similar computation shows that the other
necessary marginal constraint for π is satisfied. We conclude that
Z
U p
p
Lh,p (X , Y) = dRW,p (dhX (x), dhY (y)) µ(dx × dy)
ZX×Y
= |s − t|p µF (ds × dt)
R×R
≥ dW,p (dHX , dHY )p = LH,p (X , Y)p .
14
Example 2.12. The category C of all mm-spaces with MorC containing continuous measure-preserving
maps.
Example 2.13. The category C of all mm-spaces whose morphisms MorC are bijective measure-
preserving mappings. In this case, the set of morphisms from X to Y is nonempty if and only if X
and Y have the same cardinality.
Similarly, one could consider the category C whose objects ObjC are uncountable, separable,
nonatomic mm-spaces and whose morphisms MorC are bijective measure-preserving mappings. The
set of morphisms between any pair of objects in ObjC is nonempty. The conditions on ObjC imply that
for any morphism φ, φ−1 is also a morphism [66, Proposition A.4].
Throughout the paper, we use the term full subcategory to denote a subcategory with no extra re-
strictions placed on its morphisms.
Example 2.14. The full subcategory C whose objects ObjC are finite mm-spaces, perhaps with a fixed
cardinality. For a fixed metric space (X, dX ), one could similarly consider the full category of finite
subsets of X endowed with the restricted distance dX and uniform measure.
Example 2.15. The category C whose objects ObjC are compact Riemannian manifolds, perhaps with
fixed dimension, endowed with geodesic distance and normalized Riemannian volume, and whose mor-
phisms MorC are measure-preserving maps with fixed regularity C 0 , C 1 , C ∞ , et cetera.
A motivation for considering the Monge-style pseudometric LTh,p (as opposed to only considering
the Kantorvich-style LUh,p ) is that it naturally restricts to a given subcategory C of mm-spaces. Indeed,
define
LCh,p : ObjC × ObjC → R
by
Z Z 1 1/p
p
LC
h,p (X , Y) := inf h−1
X (x, u) − h−1
Y (φ(x), u) du µX (dx) .
φ∈MorC (µX ,µY ) X 0
In the p = 1 case, this reduces to
Z
LC C
h (X , Y) := Lh,1 (X , Y) = inf cX ,Y (x, φ(x)) µX (dx).
φ∈MorC (µX ,µY ) X
We similarly define LC C C
H,p to be the restriction of LH,p to ObjC × ObjC and use the notation LH := LH,1 .
Note that, since LH,p does not depend on any registration between mm-spaces, the formulas for LC H,p and
LH,p are exactly the same—we only use the notation LC H,p to emphasize restriction to the subcategory
and for the sake of symmetry with the notation LC h,p .
and LC C
h = Lh,1 . The main problem we consider in this paper is:
Main Problem. For a fixed subcategory of mm-spaces C, give conditions which guarantee that mm-
spaces X , Y ∈ ObjC satisfy LC C
H (X , Y) = 0 ⇔ X ≈ Y or Lh (X , Y) = 0 ⇔ X ≈ Y.
We specifically aim to address the following inverse problems, which can be viewed as refinements
of the Main Problem, for various subcategories of mm-spaces C:
15
(IP1) Global Injectivity. For X , Y ∈ ObjC , does LC
h (X , Y) = 0 imply X ≈ Y? More strongly, does
C
LH (X , Y) = 0 imply X ≈ Y?
(IP2) Homotopy Type Characterization. For X , Y ∈ ObjC , does LC h (X , Y) = 0 (or, more strongly,
C
LH (X , Y) = 0) imply that the metric spaces (X, dX ) and (Y, dY ) are homotopy equivalent?
(IP3) Sphere Rigidity. Suppose that C includes an object which can be characterized as a unit sphere
(for example, the category of Riemannian manifolds of fixed dimension), denoted S. Does
LC C
h (X , S) = 0 (or, more strongly, LH (X , S) = 0) imply X ≈ S?
(IP4) Local Injectivity. For X ∈ ObjC , does there exist X > 0 such that for all Y ∈ ObjC with
dGH (X, Y ) < X , LC C
h (X , Y) = 0 (or, more strongly, LH (X , Y) = 0) implies X ≈ Y?
These problems, and several related problems, are treated for embedded and abstract Riemannian
manifolds in Sections 4 and 5, respectively, and for certain stratified spaces (metric graphs) in Section 6.
For many results throughout the paper, we indicate which of these problems (IP1)–(IP4) to which the
result is relevant. Before considering these questions, we provide more context for the pseudometrics
LCh,p in the following section by showing that they lower bound a Monge-style variant of Gromov-
Wasserstein distance in Section 3. We remark that Sections 3, 4, 5 and 6 are essentially independent of
one another.
3 Gromov-Monge Quasi-Metrics
3.1 Gromov-Monge Distance
To contextualize the category-restricted pseudometrics LC C
H,p and Lh,p as lower bounds in their own
right, we introduce a variant of Gromov-Wasserstein which restricts the constraint set to only consider
measure-preserving mappings. The Gromov-Monge p-distance between mm-spaces X and Y is the
quantity
16
is lower than that of Gromov-Wasserstein—reducing the complexity from O(n3 log(n)) to O(n2 ) in
the p = 2 case [92]. At a theoretical level, it was observed in [88] that the geodesics (in the sense of
metric geometry) of Mw with respect to dGW,2 have a simpler structure when optimal couplings are
realized by measure-preserving maps; i.e., when the distances dGW,2 and dGM,2 coincide. This leads to
the following open question.
Question 3.2. For which classes of mm-spaces is it possible to guarantee that dGW,2 = dGM,2 ?
Remark 3.3. We remark here on progress on and further motivation for this question. The question of
realizing an optimal coupling through an optimal mapping was already raised by Sturm in [88, Chal-
lenge 3.6], where it is specified to smooth manifolds. In this setting, an answer to the problem would be
a quadratic version of famous results of Brenier [26] and McCann [69] in the classical optimal trans-
port setting. Sturm is able to show that optimal Gromov-Wasserstein couplings between rotationally
symmetric distributions in a Euclidean space are always realized by a transport map which is unique up
to composition with an isometry.
This question was recently tackled for discrete spaces in [92], with a view toward applications
in data science. In this work the authors consider the special category of uniformly weighted n-point
configurations on the real line. Even in this simple situation, the proof that Gromov-Monge and Gromov-
Wasserstein coincide is nontrivial. A proof for the general version of the discrete problem, as stated
in [88, Challenge 5.27], is still open.
It was recently observed empirically in [38] that numerically approximated optimal couplings be-
tween discrete spaces tend to at least be relatively sparse, which can be leveraged to implement the
geodesic formula in Gromov-Wasserstein space described by Sturm in [88]. A theoretical sparsity
bound was derived in [39] in a slightly different setting, where heat kernel representations of graphs,
rather than mm-spaces, are being compared. A better theoretical understanding of the optimal coupling
landscape would guide principled improvements to computation of Gromov-Wasserstein geodesics and
barycenters in applications.
The following basic result clarifies the sense in which dGM,p is a “distance”.
Proof. Let X , Y and Z be mm-spaces. From the definition of dGM,p , Remark 3.1 and the fact that dGW,p
is a metric on Mw , we easily see that dGM,p (X , Y) ≥ 0, with dGM,p (X , Y) = 0 if and only if X ≈ Y.
It only remains to show that dGM,p satisfies the triangle inequality dGM,p (X , Z) ≤ dGM,p (X , Y) +
dGM,p (Y, Z). If dGM,p (X , Z) = ∞, then T (µX , µZ ) = ∅ and it follows that either T (µX , µY ) = ∅
or T (µY , µZ ) = ∅, whence one of dGM,p (X , Y) or dGM,p (Y, Z) is infinity. The triangle inequality
therefore holds in this case. The triangle inequality follows trivially if dGM,p (X , Y) or dGM,p (Y, Z) is
infinite, so let us assume that all distances are finite. In this case we have
= inf kΓX ,Y kLp (µφ1 ⊗µφ1 ) + inf kΓY,Z kLp (µφ2 ⊗µφ2 )
φ1 ∈T (µX ,µY ) φ2 ∈T (µY ,µZ )
kΓX ,Z kLp (µφ ⊗µφ ) ≤ kΓX ,Y kLp (µφ1 ⊗µφ1 ) + kΓY,Z kLp (µφ2 ⊗µφ2 )
17
for any fixed φ = φ2 ◦ φ1 with φ1 ∈ T (µX , µY ) and φ2 ∈ T (µY , µZ ). This holds by the definition of
µφ , the Minkowski inequality and the fact that
for all x, x0 , y, y 0 , z, z 0 .
Example 3.4 (dGM,p is Not Symmetric). Let X = {u, v} be endowed with empirical measure δu =
δv = 12 and metric determined by dX (u, v) = 1. Let Y = {y} with measure δy = 1. Then the constant
map X → Y is measure-preserving, but neither of the two possible maps Y → X preserves measure
(we cannot split mass). Thus dGM,p (X , Y) = 1/21/p while dGM,p (X , Y) = ∞
Even if finite mm-spaces have the same cardinality, it is possible that the set of measure-preserving
transformations between them is empty. For example, take X as above and let Z = {Z, dZ , µZ } denote
the space with Z = {y, z}, dZ (y, z) = 1 and µZ the measure with weights δy = 1/4 and δz = 3/4.
Remark 3.5. The proof of Theorem 1 is easily modified to show that dCGM,p satisfies the triangle in-
equality for any subcategory of mm-spaces C.
Remark 3.6. As usual, we are slightly abusing terminology here and conflating elements of ObjC , which
are technically isomorphism classes of mm-spaces, with representatives of the classes.
We now explore properties of dC
GM,p for a few simple subcategories of mm-spaces.
Finite Spaces
In applications, mm-spaces under consideration are generally finite. For each natural number n, let
Cn denote the full subcategory of mm-spaces whose objects are (isomorphism classes of) mm-spaces
X = (X, dX , µX ) with |X| = n.
Proposition 3.7. dCGM,p
n
induces an extended metric on Cn .
Proof. Let δxj denote the weight assigned to xj ∈ X under µX and let δyj the weight assigned to
yj ∈ Y under µY . The sets T (µX , µY ) and T (µY , µX ) are both empty or both non-empty, and in the
former case we have dGM,p (X , Y) = dGM,p (Y, X ) = ∞. Supposing that a measure-preserving map
φ : X → Y exists, it is a bijection such that δφ(xj ) = δxj . A straightforward calculation then shows that
the inverse of the minimizer realizing dGM,p (X , Y) is the minimizer realizing dGM,p (Y, X ).
Let CN denote the full subcategory of finite mm-spaces, without restriction to a given cardinality.
That is, the object set of CN is G
ObjCN := ObjCn .
n∈N
Proposition 3.8. Let X and Y represent objects in ObjCN . If |X| = 6 |Y |, one of dGM,p (X , Y) or
dGM,p (Y, X ) must be infinity and symmetry only holds when dGM,p (X , Y) = dCGM,p
CN CN N
(Y, X ) = ∞.
Proof. A measure-preserving map φ : X → Y between finite mm-spaces with fully supported measures
must be surjective. Suppose that |X| = n and |Y | = m and assume without loss of generality that
n ≥ m. If n > m, then there is no measure-preserving map Y → X, hence dC
GM,p (Y, X ) = ∞.
N
18
Nonatomic Spaces
At the other extreme, dGM,p induces a true metric on a broad class of mm-spaces with infinite cardinality.
Let C∞ denote the full subcategory of mm-spaces whose objects are isomorphism classes of mm-spaces
X = (X, dX , µX ) such that X is uncountable, (X, dX ) is separable and µX has no atoms. Classical
results from measure theory (e.g., [82, Chapter 5, Theorem 16]) imply that for any X , Y representing
elements of C∞ , T (µX , µY ) is nonempty. It follows that dGM,p is finite on C∞ and one can symmetrize
it to obtain the following result.
n o
Proposition 3.9. The symmetrization (X , Y) 7→ max dCGM,p ∞
(X , Y), dCGM,p
∞
(Y, X ) of dCGM,p
∞
induces
a metric on C∞ .
Let C0∞ denote the subcategory of mm-spaces with ObjC0∞ = ObjC∞ whose morphisms MorC0∞
are bijective measure preserving maps.
C0
∞
Proposition 3.10. dGM,p induces a metric on C0∞ .
Proof. It follows once again from basic measure theory that MorC0∞ (µX , µY ) is nonempty, as both
spaces X and Y are isomorphic at the level of measure spaces to [0, 1] with Lebesgue measure [82,
Chapter 5, Theorem 16]. Moreover, if φ : X → Y is a measure preserving bijection, then ψ := φ−1 is
as well [66, Proposition A.4] and the change of variables formula applies to show that
kΓX ,Y kLp (µψ ⊗µψ ) = kΓX ,Y (ψ(·), ψ(·), ·, ·)kLp (µY ⊗µY ) = kΓX ,Y (ψ(·), ψ(·), ·, ·)kLp (φ# µX ⊗φ# µX )
= kΓX ,Y (ψ ◦ φ(·), ψ ◦ φ(·), φ(·), φ(·))kLp (µX ⊗µX )
= kΓX ,Y (·, ·, φ(·), φ(·))kLp (µX ⊗µX ) = kΓX ,Y kLp (µφ ⊗µφ ) .
∞ C0 ∞ C0
Taking a infimum over φ then yields dGM,p (Y, X ) ≤ dGM,p (X , Y) and a symmetric argument shows
C0
∞
the reverse inequality, hence dGM,p is symmetric.
Manifolds
Here we describe two examples of metrics from the literature which can be viewed as instances of dC
GM,p .
Example 3.11. Let C be the category whose objects are smooth embedded surfaces in R3 endowed
with normalized surface area measure and geodesic or Euclidean distance and whose morphisms are
continuous measure-preserving maps. The continuous Procrustes distance between objects in C was
introduced in [25], where it was used to classify anatomical surfaces (in particular, shapes of primate
teeth). The effectiveness at classification of this metric was shown to be roughly on par with a that of a
trained morphologist. The idea of the continuous Procrustes distance is to compare surfaces by simul-
taneously registering over rigid motions while looking for optimal measure-preserving maps between
them. Theoretical aspects of this metric are studied in [5], where it is shown that optimal mappings
are close to being conformal. The continuous Procrustes distance can be viewed as dCGM,2 . Work in the
Appendix (in particular, Sections A.3 and A.4) makes the comparison between these metrics precise.
Example 3.12. Let C denote the category of Euclidean mm-spaces in R2 whose morphisms are restric-
tions of measure-preserving diffeomorphisms of R2 . A metric similar to the continuous Procrustes metric
on this category is studied in [53] for applications to 2D image registration. Results in Sections A.3 and
A.4 show that (a variant of) dCGM,p is a generalization of this metric.
19
Proposition 3.13. For p ≥ 1 and mm-spaces X and Y representing objects of C,
Recall that computation of LCH,p does not require any registration of the spaces via a morphism, so it
is simply a restriction of LH,p to ObjC . On the other hand, LCh,p does involve registration of the spaces.
In this setting, we get a more novel result.
The proof follows the proof of Proposition 2.9 given originally in [72], replacing couplings with
maps as necessary. We omit the details here. Similarly, the proof of Proposition 2.11 can be adapted to
obtain the following refinement.
Point Clouds
While this section is generally concerned with addressing inverse problems for manifolds, we begin by
reviewing a result on arguably the simplest subclass of mm-spaces: point clouds in Euclidean space. Let
CN,k denote the full subcategory of mm-spaces whose objects are isomorphism classes of mm-spaces
X = (X, dX , µX ) such that X is a set of N points in Rk , dX is Euclidean distance and µX is uniform
measure. The problem of reconstructing X from its collection of interpoint distances (i.e., from its
distance distribution HX ) is classical and has applications to DNA sequencing and X-ray crystallography
[65].
C
It is well known that LHN,k does not distinguish elements of CN,k for any (N, k) with N > 3.
C
Example 2.2 gave an example of nonisomorphic mm-spaces X , Y in C1,6 with LH1,6 (X , Y) = 0. Coun-
terexamples in higher dimensions were given by Boutin and Kemper [24, Section 1.1], who refined the
question and proved the following theorem (which we state using our terminology).
C
Theorem 2 ( [24]). The pseudometric LHN,k is locally injective on CN,k . That is, for every point
cloud X , there exists X > 0 such that if a point cloud Y is X -close to X in Hausdorff distance
C
and LHN,k (X , Y) = 0 then X and Y differ by a rigid motion.
This result gives an affirmative answer for the category of Euclidean point clouds to the local injec-
tivity question (IP4) posed in Section 2.3.
20
S2 S1 T10
T3 0
S2
S3
T30
0
S3
T4 T2
S4
T40 T20
0 0
T1 S1 S4
21
The next theorem then follows immediately.
Proposition 4.2 (Noninjectivity of Global Distance Distributions for Plane Curves (IP1)). The curves
in Figure 3 are not isomorphic, but satisfy the hypotheses of Theorem 3 and therefore have the same
distance distribution.
Proof. Denote the curve on the left by X and the curve on the right by Y . We partition each curve into
triangular pieces Ti (respectively, Ti0 ) and straight pieces Sj (respectively, Sj0 ) according to the figure.
For each pair (Xi , Xj ) of pieces X (here Xj stands as a placeholder for either an Sj or a Tj ), we find
a corresponding pair (Yk , Y` ) in Y so that there is a rigid motion taking the first pair onto the second.
Since the functions HXi,j and HYk,` are invariant under rigid isometries, the hypotheses of Theorem 3 are
then satisfied.
Pairs of the form (Si , Si ) and (Ti , Ti ) are matched with (Si0 , Si0 ) and (Ti0 , Ti0 ), respectively. Moreover,
note that curve Y is obtained form curve X by “swapping” S1 with T1 , it is clear that (S1 , T1 ) should be
matched with (S10 , T10 ). It also follows that we have obvious matchings (Si , Sj ) ↔ (Si0 , Sj0 ), (Ti , Tj ) ↔
(Ti0 , Tj0 ) and (Si , Tj ) ↔ (Si0 , Tj0 ) for i, j 6= 1. It remains to show that the desired matchings exist for
pairs which include S1 or T1 . These are given by
(S1 , S2 ) ↔ (S10 , S40 ) (S1 , T2 ) ↔ (S10 , T20 ) (T1 , T2 ) ↔ (T10 , T20 ) (T1 , S2 ) ↔ (T10 , S40 )
(S1 , S3 ) ↔ (S10 , S30 ) (S1 , T3 ) ↔ (S10 , T40 ) (T1 , T3 ) ↔ (T10 , T40 ) (T1 , S3 ) ↔ (T10 , S30 )
(S1 , S4 ) ↔ (S10 , S20 ) (S1 , T4 ) ↔ (S10 , T30 ) (T1 , T4 ) ↔ (T10 , T30 ) (T1 , S4 ) ↔ (T10 , S20 ).
Remark 4.3. By smoothing corners for the curves in our example in a symmetric manner, we could
produce a pair of C ∞ curves satisfying the hypotheses of Theorem 3. This would therefore produce
a counterexample to the Curve Histogram Conjecture involving C ∞ curves. Moreover, making the
heights of the triangles (or smoothed triangles) small enough, the curves can be made convex while still
providing a counterexample to the conjecture.
Remark 4.4. Consider the point y at the tip of pyramid T10 in curve Y on the right hand side of Figure
3. It is easy to see that the local distance distribution hY (y, ·) is different from the local distribution
hX (x, ·) of any point x in the curve X on the left hand side of Figure 3. It follows that these curves
are distinguished by the lower bound Lh . We show below in Proposition 4.10 that, in an appropriately
restricted category, plane curves are always distinguished by local distributions.
Pushing this construction further, we obtain the following corollary demonstrating a lack of distin-
guishing power of LH locally (cf. the main result of [14] demonstrating injectivity in a neighborhood of
the unit circle for the circular integral invariant for plane curves).
Corollary 4.5. Let S1 denote the unit circle endowed with extrinsic Euclidean distance and normalized
arclength measure. For any > 0, there exist simple closed planar curves X1 and X2 such that
2. X1 is not isometric to X2 ;
3. HX1 = HX2 .
22
Figure 4: Counterexample to the Curve Histogram Conjecture constructed from 14-gons.
Proof. The construction of the pair of curves in Figure 3 can be generalized to produce counterexamples
to the curve histogram conjecture starting with any regular polygon with an even number of sides.
Indeed, starting from a regular 2n-gon, partition its edges into two sets of equal size such that there
is no isometry taking one of the sets onto the other. We then construct a pair of curves by appending
congruent isosceles triangles to each edge partition set—see Figure 4. One is able to show that the
resulting curves satisfy the hypotheses of Theorem 3. We note that a similar idea was used recently
in [46] to construct more counterexamples to Blaschke’s conjecture; it is shown there that a valid choice
of edges always exists. By starting with a regular 2n-gon with n sufficiently large and by appending
isosceles triangles with sufficiently small height, we can construct pairs of convex curves in this manner
which are arbitrarily close to a circle in Hausdorff distance. By smoothing corners in this construction,
we are also able to produce smooth convex curves satisfying the properties.
Proposition 4.2 shows that the global distance distribution is not globally injective on the category of
embedded plane curves, answering inverse problem (IP1) in the negative for plane curves. The question
of whether it is locally injective remains open (i.e., the answer to (IP4) for plane curves), and is stated
here more formally.
Question 4.6. If Xτ is a smooth one-parameter family of embedded plane curves such that the family
of global distance distributions HXτ is constant in τ , must each pair of curves in the family differ by a
rigid motion?
In the sequel we are able to show that the local distance distribution is injective on the category
of embedded curves (Proposition 4.10). Taking a different perspective, we will demonstrate below a
one-parameter family of nonisomorphic densities on the unit circle with constant distance distributions
(Example 5.8).
23
Proof. By [27, Section 7], one has for r > 0 small enough the Taylor expansion
r3
I
2r
HX (r) = + κ2 (s) ds + O(r5 ), (11)
`(X) 12`(X)2 X
where κ denotes curvature, ds denotes measure with respect to arc length and `(X) denotes the length
of X. For the curve S1 this yields
r r3
HS1 (r) = + + O(r5 ). (12)
π 24π
Equating these two Taylor expansions gives `(X) = 2π and X κ2 (s) ds = 2π. By assumption, since
H
We also exhibit below some examples of infinite subcategories of plane curves containing the unit
circle whose elements are completely distinguished from one another by their distance distributions.
Example 4.8 (Ellipses). For given A ≥ B ≥ 0, let XA,B be an ellipse with semi-major axis A and semi-
minor axis B. We treat each such XA,B as a mm-space XA,B by (as usual) endowing it with extrinsic
Euclidean distance and normalized arclength measure.
We claim that if ellipses XA,B and XA0 ,B 0 have the same distance distributions HXA,B = HXA0 ,B0
then A = A0 and B = B 0 , whence we conclude that the curves are related by a rigid motion. Indeed, if
the ellipses have the same distance distributions, then they have the same diameter, as this can be read
off from the distance distribution by the formula
diam(XA,B ) = inf{r ≥ 0 | HXA,B (r) = 1}.
The diameter of XA,B is just 2A and this implies that A = A0 . Using the degree-1 term of (11), we
also see that the ellipses have the same length. For fixed A, the length of an ellipse XA,B is a strictly
increasing function of B. Thus A = A0 and `(XA,B ) = `(XA0 ,B 0 ) together imply that B = B 0 .
Example 4.9 (Bumpy Circles). Below by x(s) for s ∈ [0, 2π] we denote an arclength parametrization
of the unit circle. For each A ∈ [0, 1) and n ∈ N consider the density function (w.r.t. the length measure
1
len on S1 ) fA,n : S1 → R+ given by fA,n (s) := 2π (1 + A sin(n · s)), for s ∈ [0, 2π]. Consider the
mm-space SA,n = (S , k · − · k, fA,n · len) (so S0,n is isomorphic to S1 for any n). By direct calculation
1
(following the same steps as in [27]) one can prove that for each s ∈ [0, 2π] and t > 0 small,
1 + A sin(n · s) 1 + A − 4An2
Z
hSA,n (s, t) = fA,n (s) ds = t+ sin(n · s) t3 + O(t5 ).
S1 ∩Bt (x(s)) π 24π
It follows that
2 + A2 2 + A2 − 4A2 n2 3
HSA,n (t) = t+ t + O(t5 ).
2π 48π 2
Note that this formula for HS0,n (t) reduces to the formula for HS1 (t) given by (12).
We immediately conclude that the global distribution of distances is able to discriminate these
“bumpy circles”. More precisely, if for some A, A0 ∈ [0, 1) and n, n0 ∈ N it holds that HSA,n = HSA0 ,n0 ,
then it must be that A = A0 and (if A, A0 6= 0) n = n0 .
24
Figure 5: A = 0.1 and n = 5. The blue curve Figure 6: A = 0.05 and n = 40. The blue
represents the density fA,n over S1 . The red curve represents the density fA,n over S1 . The
curve shows the uniform distribution. red curve shows the uniform distribution.
κ1 (s)2 = κ2 (s)2 ,
25
Figure 7: General picture of the local behavior of curves X1 and X2 from the proof of Proposition 4.10.
where δ > 0 is sufficiently small and f1 , f2 are both convex. The curve X2 is then expressed near γ2 (s0 )
as the graph of
f1 (x) x ∈ (−δ, 0]
g(x) =
−f2 (x) x ∈ (0, δ).
(See Figure 7 for an example.) Now fix x0 ∈ (0, δ). It follows from the definitions of these functions
that for any x ∈ (−δ, 0),
k(x, f (x)) − (x0 , f (x0 ))k < k(x, g(x)) − (x0 , g(x0 ))k. (13)
Choose x0 small enough so that B2x0 (x0 ) only intersects the curves Xj in the graphs of f and g, respec-
tively. Consider `(B2x0 (x0 ) ∩ Xj ). For j = 1, this length breaks into two pieces: ` (B2x0 (x0 ) ∩ X1 ) =
`− + − +
1 + `1 , where `1 is the length of the intersection of the ball with the graph of f1 and `1 is the length
of the intersection with the graph of f2 . We also have `(B2x0 (x0 ) ∩ X2 ) = `− + ±
2 + `2 , with `2 defined
+ + − −
similarly. By construction, we have `1 = `2 . On the other hand (13) implies that `1 > `1 . This implies
that the local distance distributions of X1 and X2 are not the same, and we have finally arrived at the
desired contradiction.
Corollary 4.11 (Global Injectivity of Local Distance Distributions for Plane Curves (IP4)). Let C be
the category whose objects are generic simple closed plane curves and whose morphisms are twice
continuously differentiable infinitesimally measure-preserving maps. Then LCh (X , Y) = 0 implies X ≈
Y.
Proof. Suppose there is a sequence φn of morphisms such that
Z
lim cX ,Y (x, φn (x)) µX (dx) = 0.
n→∞ X
Fix an arbitrary point x0 ∈ X. Observe that any infinitesimally measure-preserving map is completely
determined by the image of x0 under the map, together with the induced orientation on Y . The space of
infinitesimally measure-preserving maps is therefore homeomorphic to the compact space Y × {±1}.
The sequence {φn } then gives rise to a sequence {φn (x0 )} of points in Y . Since Y is compact, there
is a subsequence {φni } such that {φnj (x0 )} converges to some point y0 ∈ Y and such that each φni
induces the same orientation on Y . Let φ be the infinitesimally measure-preserving map determined by
y0 and this common orientation. Then the subsequence {φni } converges uniformly to φ and it follows
that hX (x, r) = hY (φ(x), r) for all (x, r) ∈ X × R≥0 . Proposition 4.10 then implies that X is isometric
to Y , which in turn implies that X ≈ Y.
26
Figure 8: Nonisometric polyhedral surfaces with the same distributions of distance. The surfaces are
constructed by adding pyramids according to the partition described by the net on the left.
Corollary 4.12. Let S2 denote the unit sphere. For any > 0, there exist closed smooth surfaces X1
and X2 such that
2. X1 is not isometric to X2 ;
3. HX1 = HX2 .
Proof. We begin by demonstrating nonisomorphic piecewise linear surfaces with the same distance
distributions. Starting with a dodecahedron or icosahedron, we partition the faces into two sets so that
there is no isometry mapping one set onto the other. For each set, we construct a new polyhedral surface
by attaching a symmetric pyramid along the faces in the set. The resulting pair of polyhedral surfaces
are nonisometric by construction, but one can show that they satisfy the hypotheses of Theorem 3. An
example is shown in Figure 8. Similar constructions are also explored in [46] as counterexamples to a
higher dimensional analogue of Blaschke’s conjecture.
Next we construct smooth surfaces in an arbitrarily small Gromov-Hausdorff neighborhood of the
unit sphere. Begin by circumscribing the unit sphere in an dodecahedron. Partition the faces of the
dodecahedron as in Figure 8. This gives a partition of the 12 points on the sphere which are tangent
to the dodecahedron, say into sets A and B. To construct X1 , smoothly perturb S2 by adding a small
radially symmetric bump over each point in set A. Similarly, construct X2 by adding a bump over each
point in set B. Then our surfaces can be made arbitrarily close to S2 in Hausdorff distance, but they are
nonisomorphic and have the same distance distribution by construction.
Corollary 4.12 shows that the global distance distribution is not globally injective on the space of
embedded surfaces (answering (IP1) in the negative). As in the case of curves, it is an open question
whether it is locally injective in the sense of (IP4).
Question 4.13. If Fτ : S2 → R3 is a smooth one-parameter family of embeddings such that the global
distance distributions of the induced mm-spaces (using extrinsic Euclidean distance on their images)
are constant in τ , must the images of Fτ all be isomorphic?
A similar question could be asked of one-parameter families of surfaces whose local distance dis-
tributions are constant in some sense. Stronger than this would be to prove an analogue of Proposition
4.10 in this setting.
Question 4.14. If X1 and X2 are smooth embedded surfaces and φ : X1 → X2 is a smooth map
satisfying hX1 (p, r) = hX2 (φ(p), r) for all p and r, must X1 and X2 be related by a rigid motion?
27
4.5 Sphere Characterization for Embedded Manifolds
Proposition 4.7 can be generalized to all dimensions, at least with assumed curvature bounds, giving a
partial answer to (IP3). Here we consider an embedded, smooth, closed hypersurface X in Rd+1 as a
mm-space X with extrinsic Euclidean distance and normalized d-dimensional Hausdorff measure. The
unit d-sphere Sd is considered as a mm-space in this manner. Let κj denote the principal curvature
functions of X with respect to a chosen orientation, labeled so that κ1 (p) ≤ κ2 (p) ≤ · · · ≤ κd (p) for all
p ∈ X. That is, κd (p) is defined to be the maximum of all principal curvatures at point p, with the rest
of the functions defined similarly. Without loss of generality, one can assume that κd (p) ≥ 0 for some
point p ∈ X.
Theorem 4. Let X be a hypersurface in Rd+1 with principal curvature functions κj as defined above.
1. In the d = 2 case, if HX (r) = HS2 (r) for all r ≥ 0, then X is isometric to S2 .
2. In the case d > 2, further assume that κj (p) ≤ 1 for all j = 1, . . . , d and p ∈ X and κd (p) ≥ 0
for all p ∈ X. If HX (r) = HSd (r) for all r ≥ 0 then X is isometric to Sd .
The proof will use the following result of Karp and Pinsky, which we express using our notation.
For an embedded hypersurface X, let volX denote its induced volume form and let
Z
Vol(X) = volX (dp)
X
denote its total volume.
Theorem 5 ( [59]). Let p ∈ X and let κ1 , . . . , κd denote the principal curvatures of X at p. Then for
sufficiently small r ≥ 0, we have
1 σd−1 d σd−1 d+1
d+2
hX (p, r) = r + (2A − B) r +O r , (14)
Vol(X) d 8d(d + 2)
where 2
d
X d
X
A := κ2j , B := κj
j=1 j=1
Then, on the domain (−∞, 1] × (−∞, 1] × · · · × (−∞, 1] × [0, 1], G has a unique minimum at the point
(1, 1, . . . , 1).
28
Proof. For any point (x1 , . . . , xd ) in the domain of G with xi 6= xj , we have that the point
xi + xj xi + xj
x1 , . . . , ,..., , . . . , xd
2 2
xi +xj xi +xj
is still in the domain of G and G(x1 , . . . , xd ) > G(x1 , . . . , 2 ,..., 2 , . . . , xd ). Indeed, writing
G(x1 , . . . , xd ) = 2 · P (x1 , . . . , xd ) − Q(x1 , . . . , xd )2 ,
the term Q does not change under this substitution but the term P becomes smaller due to the inequality
x2 + y 2 > 2( x+y 2 2
2 ) , which is equivalent to (x − y) > 0, valid for all x, y real and different. Now, this
means that we need to consider the minimum of g(x) := G(x, . . . , x) = (2d − d2 )x2 in (−∞, 1]. But
since d > 2, the unique minimizer is x = 1.
Proof of Theorem 4. Since the principal curvatures of Sd are all constantly equal to one, we derive from
(15) that, for sufficiently small r,
1 σd−1 d 1 σd−1
HSd (r) = r + rd+1 (2d − d2 ) + O rd+2 .
σd d σd 8d(d + 1)
Combining (15) with the assumption HX (r) = HSd (r), we see that volX (X) = σd and
Z
2A − B volX = (2d − d2 )σd .
X
We now specialize to the two cases.
1. First consider the d = 2 case. Here we have 2d − d2 = 0, whence
Z Z
(κ1 − κ2 )2 volX .
0= 2A − B volX =
X X
This implies κ1 (p) = κ2 (p) for all p ∈ X. A closed surface in which all points are umbillic is a
sphere [42, Section 3-2, Proposition 4], and Vol(X) = σ2 implies that X is isometric to S2 .
2. Now suppose d > 2. With G denoting the polynomial function from Lemma 4.17, we have
Z
2
(2d − d )σd = G(κ1 , . . . , κd ) volX ≥ min(G)Vol(X) = (2d − d2 )σd .
X
It follows that the inequality must actually be an equality, so that G is constantly equal to its
minimum value. By Lemma 4.17, this minimum is unique and we have κ1 = κ2 = · · · = κd = 1.
This implies that X is isometric to Sd .
Corollary 4.18 (Sphere Characterization for Embedded Hypersurfaces (IP3)). Let C be the category of
closed hypersurfaces in Rd+1 with curvature bounds as in Theorem 4. Then LCH (X , Sd ) = 0 implies X
is isomorphic to Sd .
29
5.1 Sphere Characterization for Riemannian Manifolds
We begin with a characterization of the unit sphere in the category of Riemannian mm-spaces. The state-
ment of the theorem uses notation introduced earlier: the global shape measure (1) and the Wasserstein
1-distance (5).
Theorem 6. Let (X, g) be a d-dimensional Riemannian manifold with Ricci curvatures bounded below
by d − 1. Let dHX and dHSd denote the global shape measure of X and the d-dimensional round sphere
with its standard Riemannian structure, respectively.
(1) There exists an = (d) > 0 such that if the Wasserstein 1-distance between the shape measures
satisfies
LH (X , Sd ) = dRW,1 (dHX , dHSd ) <
then X is diffeomorphic to Sd .
(2) If the Wasserstein distance is zero—equivalently, HX (r) = HSd (r) for all r ≥ 0, or LH (X , Sd ) =
0—then X is isometric to Sd .
The proof of the theorem will use the concept of 1-diameters, together with a lemma. For a mm-
space X , the 1-diameter of X diam1 (X ) is the number
ZZ
diam1 (X ) := dX (x, x0 )µX ⊗ µX (dx × dx0 ).
X×X
Properties of this invariant (and more generally of the p-diameters of a mm-space) are described in [72,
Section 5.2]. This invariant also appears in [61], where it is referred to as the mean distance of X and is
used in a sphere characterization theorem. We will appeal to this sphere characterization in the proof of
our theorem momentarily.
Lemma 5.1. For mm-spaces X and Y we have the lower bound
Proof. Let Z = ({z}, 0, δz ) denote the one point mm-space. One can check that dHZ = δ0 , the δ
probability measure on R supported on {0}. Thus the only coupling between the probability measures
dHX and dHZ is the product measure dHX ⊗ δ0 . From this observation, one is able to show
Proof of Theorem 6. To prove the first statement, we appeal to the previously mentioned sphere charac-
terization theorem of Kokkendorf [61, Theorem 4]. Translating Part 3 of this theorem into our terminol-
ogy, it says that there exists an = (d) such that, under our curvature assumptions,
implies that X is diffeomorphic to Sd . Together with Lemma 5.1, this proves the first claim of the
theorem.
The second claim of the theorem follows directly from Part 2 of [61, Theorem 4]. Alternatively, we
can observe that HX = HSd implies that X and Sd have the same diameter. It then follows immediately
from Cheng’s Rigidity Theorem [37] that X and Sd are isometric.
Corollary 5.2 (Sphere Characterization for Riemannian Manifolds (IP3)). Let C be the full subcategory
of closed Riemannian d-manifolds with curvature bounds as in Theorem 6. Then LCH (X , Sd ) = 0 implies
X is isomorphic to Sd .
30
In the special cases of 2 or 3-dimensional Riemannian manifolds, the second part of the above
theorem can be derived by a more elementary argument; i.e. without invoking Kokkendorf or Cheng’s
deeper results. This elementary proof relies on a Taylor expansion of the distance distributions. This
expansion will be useful later, so we introduce it here and provide the details of the elementary proof in
these special cases.
Lemma 5.3. Let X be a d-dimensional Riemannian manifold with its Riemannian mm-space structure.
The local distribution of distances of X admits the following Taylor expansion: for r > 0 small enough
and all p ∈ X,
ωd (r) Sg (p) 2 4
hX (p, r) = 1− r + O(r ) ,
Volg (X) 6(d + 2)
where ωd (r) is the volume of a ball of radius r in Rd and Sg (p) is the scalar curvature of X at the point
p. It then follows that
R
ωd (r) X Sg (p) volg (dp) 2 4
HX (r) = 1− r + O(r ) . (16)
Volg (X) 6(d + 2)Volg (X)
Proof. The local distance distribution expansion can be found in [72, p. 446]. The global distance
distribution expansion then follows by integrating.
Alternate Proof of Theorem 6, Part (2), for d = 2 and 3. By the Ricci curvature assumption, it holds
that for all p ∈ M the scalar curvature at that point satisfies Sg (p) ≥ d(d − 1). From the assump-
tion that HX = HSd and the Taylor expansion (16) that Volg (X) = Vol(Sd ) and
Z Z
Sg (p) volg (dp) = volSd (dp) = Vol(Sd )d(d − 1),
X Sd
where we use the fact that all scalar curvatures of the d-dimensional sphere are equal to 1. Since
Volg (X) = Vol(Sd ), this implies that
Z
Sg (p) − d(d − 1) volg (dp) = 0.
X
Then it must be that Sg (p) = d(d − 1) for all p ∈ X. But then, again by our assumed Ricci curvature
bounds, this implies that all Ricci curvatures must be equal to 1. This in turn implies that all sectional
curvatures are equal to 1: in the d = 2 case, Ricci curvatures are the same as Gauss curvatures and in
the d = 3 case constant Ricci curvatures imply that X is Einstein, and it is well known that this implies
X has constant sectional curvature [19]. In either case, this shows that X is a quotient of a unit sphere
by a discrete group of isometries. Any nontrivial quotient has smaller diameter than the unit sphere, but
HX = HSd implies that X has diameter π.
Observe that in both the embedded and Riemannian categories, our sphere characterization theorems
use assumptions on curvature for technical reasons. On the other hand, we do not have counterexamples
showing that these assumptions are necessary. The proofs of these results only use data about the distri-
butions coming from the infinitesimal radius regime (in the form of Taylor expansions) or rather coarse
long-range information (diameter of the spaces), so it is possible that using more detailed long-range
information from the distributions could allow one to drop the curvature assumptions.
31
the normalized Riemannian volume measure on X. A weighted Riemannian manifold is a metric mea-
sure space X = (X, dg , f · µg ), where f is a smooth probability density function with respect to µg .
Such spaces are frequently considered, for example, in the context of the concentration of measure phe-
nomenon [45, 51]. The next result shows that the normalized Riemannian volume is distinguished from
other densities by the global distance distribution. For this section, when a base Riemannian manifold
is fixed, we denote the global distance distribution of the weighted Riemannian manifold with density
f by Hf . We denote the global distance distribution with respect to normalized Riemannian volume by
Hg .
Proposition 5.5. Let (X, g) be a compact Riemannian manifold and let f be a probability density. If
Hf = Hg , then f is constantly equal to one.
Proof. The proof follows from the estimate
Z
f (x)2 µg (dx) ≤ 1. (17)
X
whence the Cauchy-Schwartz inequality must be an equality and f is linearly dependent on the constant
function 1. It follows that f = 1.
It remains to show (17). Let > 0. By the metric space Lebesgue Differentiation Theorem [55,
Theorem 1.8], for r sufficiently small we have
Z
1
f (x) < f (y)µg (dy) + ,
µg (BX (x, r)) BX (x,r)
for µg -almost every x ∈ X. This implies
Z Z Z !
2
f (x) µg (BX (x, r))µg (dx) < f (x) f (y)µg (dy) µg (dx)
X X BX (x,r)
Z
+ f (x)µg (BX (x, r))µg (dx).
X
Observing that !
Z Z
f (x) f (y)µg (dy) µg (dx) = Hf (r),
X BX (x,r)
we obtain
Z Z
2
f (x) µg (BX (x, r))µg (dx) < Hf (r) + f (x)µg (BX (x, r))µg (dx)
X ZX
= Hg (r) + f (x)µg (BX (x, r))µg (dx).
X
32
Densities on the Circle
In the case that X is the unit circle S1 , we are able to use tools from Fourier analysis to understand the
space of densities in more detail. In the following, let dS1 denote geodesic (arclength) distance on S1 and
let µS1 denote normalized arclength measure. For a density f with respect to µS1 , we use the simplified
notation Hf for the global distance distribution of the mm-space (S1 , dS1 , f µS1 ). We parameterize S1
as [0, 2π)/0 ∼ 2π with parameter s and use additive notation for the Lie group structure on S1 .
Proposition 5.6. Let f1 and f2 be two densities with respect to µS1 . The global distance distributions
Hf1 and Hf2 agree if and only if f1 and f2 have the same autocorrelation functions:
Z Z
f1 (s)f1 (s + t)µS1 (ds) = f2 (s)f2 (s + t)µS1 (ds)
S1 S1
for all t.
It follows that
Z Z
d
Hf (r) = fj (s) (fj (s + r) − fj (s − r)) µS1 (ds) = 2 fj (s)fj (s + r)µS1 (ds).
dr j S1 S1
d d
Since Hf (0) = 0 for any distribution f , Hf1 = Hf2 if and only if dr Hf1 = dr Hf2 , so this completes
the proof.
Corollary 5.7. Let f1 and f2 be densities with respect to µS1 with Fourier expansions
X X
f1 (s) = an eins , f2 (s) = bn eins
n n
By Proposition 5.6, Hf1 = Hf2 if and only if q1 = q2 . Letting q̂j denote the Fourier transform of qj , we
have that q1 = q2 if and only if q̂1 (n) = q̂2 (n) for all n, which in turn holds if and only if |an | = |bn |
for all n.
Example 5.8 (Global Distance Distribution is Not Locally Injective for Weighted S1 (IP4)). We now
show how to construct a one-parameter family of probability densities fτ on S1 such that Hfτ is constant
in τ but so that the corresponding mm-spaces are Pnot isomorphic for any distinct parameter pairs. Let
ins
f be any density with Fourier expansion f (s) = n an e . Because f is a real-valued density, a0 = 1
and a−n = an (complex conjugate of an ). Suppose that an is nonzero for at least two values of n > 0,
say n1 and n2 . Our one-parameter family of densities is then defined as
X
fτ (s) = 1 + eiτ an1 ein1 s + e−iτ a−n1 e−in1 s + an eins .
n6=0,±n1
For every value of τ ∈ [0, 2π), fτ is a real-valued density. Moreover, for each τ , the Fourier coef-
ficients of fτ agree in magnitude with the Fourier coefficients of f0 , by construction. It follows from
33
Figure 9: Examples from a one-parameter family of densities on the circle fτ such that Hfτ is constant
in τ , but no two distinct densities produce isomorphic mm-spaces (see the text for the formula for fτ
plotted here). The densities plotted are f0 , f π2 and fπ , from left to right. In each figure, the blue curve is
a graph of the density fτ and the red curve is a graph of the constant density, for reference.
Corollary 5.7 that Hfτ is constant in τ . Finally, we claim that the mm-spaces (S1 , dS1 , fτ1 µS1 ) and
(S1 , dS1 , fτ2 µS1 ) are not isomorphic for any τ1 6= τ2 in [0, 2π). Indeed, such an isomorphism would
consist of an isometry φ of S1 satisfying fτ2 ◦ φ−1 = fτ1 . If φ is a rotation, we can parametrically
express it as φ−1 (s) = s + θ for some θ ∈ [0, 2π). Comparing terms of fτ2 ◦ φ−1 (s) and fτ1 (s), one is
easily able to deduce from the assumption that there exist distinct nonzero coefficients an1 and an2 that
τ1 = τ2 . A similar analysis works in the case that φ is a reflection.
Examples from a specific one-parameter family are shown in Figure 9. The one-parameter family
shown there is specifically
fτ (s) = 1 + 0.1 eiτ eis + e−iτ e−is + ei3s + e−i3s + ei5s + e−i5s .
The densities plotted in the figure are f0 , f π2 and fπ , from left to right.
This example resolves (IP4) in the negative for the category of mm-spaces whose objects are
weighted S1 ’s.
Question 5.9. Let f be a density on S1 . When is it possible to find an embedding γ : S1 → R2 so that
(γ(S1 ), dEuc , len) (where dEuc is extrinsic Euclidean distance) has the same global distance distribution
as (S1 , dS1 , f µS1 )? Such a duality result would provide a way to port examples such as Example 5.8 to
the extrinsic setting considered in Section 4.
34
Proof. For a generic metric g, consider the 4 functions I1 , I2 , I3 , I4 : M → R:
I1 := Kg
I2 := k∇Kg k2
I3 := g(∇Kg , ∇I2 )
I4 := k∇I2 k2 .
(1) if X and X 0 are two surfaces such that the their corresponding I1 and I2 have everywhere inde-
pendent differentials, and
(2) f : X → X 0 is a map between the two surfaces which preserves the functions I1 , . . . , I4 ,
then f is an isometry. Indeed, this follows because the metric g can be recovered explicitly as
Corollary 5.11 (Global Injectivity for Local Distance Distributions of Riemannian Structures on a Fixed
Surface (IP1)). Fix a closed smooth 2-manifold X and let C denote the subcategory of mm-spaces
whose objects are given by different choices of Riemannian metric on X and whose only morphisms are
obtained from the identity map on X. Then LCh induces a metric on ObjC .
Remark 5.12. We relate the above corollary to the global injectivity inverse problem (IP1), but we
observe that it does not fit exactly into that framework: the identity map is not necessarily measure-
preserving when comparing different choices of metric on the same surface. The category C covered
by the corollary is therefore not actually a subcategory of mm-spaces. Nonetheless, the result can be
considered as addressing a variant of (IP1).
Lemma 5.13. Let X be any closed 2-dimensional Riemannian manifold. For r > 0 small enough and
all p ∈ X,
π r2
Kg (p) 2 1 2
4 6
hX (p, r) = 1− r + 2Kg (p) − 3∆Kg (p) r + O(r ) . (18)
Volg (X) 12 720
35
It then follows that
πr2
Z
1
HX (r) = 1− Kg (p) volg (dp) r2
Volg (X) 12 Volg (X) X
Z
1 2 4 6
+ K (p) volg (dp) r + O(r )
360 Volg (X) X g
and, in particular,
π r2 π2 r4
HX (r) = − χ(X) + O(r6 ).
Volg (X) 6 (Volg (X))2
Proof. Equation (18) follows from [48, Corollary 3.4]. Now, by integratingRthis expression against
the area measure volg (·) of X and using the fact that X is closed (so that X ∆Kg (p) volg (dp) =
0) we obtain the second Taylor expansion. Finally, the last formula follows from the Gauss-Bonnet
theorem.
We have the following immediate corollary, which affirmatively answers a strengthening of inverse
problem (IP2).
Corollary 5.14 (Diffeotype Characterization for Riemannian Surfaces (IP2)). Let C denote the full
subcategory of mm-spaces whose objects are closed Riemannian manifolds. Then LCH (X , Y) implies X
and Y are diffeomorphic.
The distinguishing power of global distance distributions can be pushed farther to also give an affir-
mative answer to a refinement of inverse problem (IP3).
Theorem 7 (Constant Curvature Surface Characterization for Riemannian Surfaces (IP3)). Let (X0 , g0 )
and (X, g) be any two oriented 2-dimensional closed Riemannian manifolds such that X0 has constant
Gaussian curvature κ ∈ R and HX = HX0 . Then X is diffeomorphic to X0 and also has constant
curvature κ. In particular, if κ > 0 then X and X0 are both isometric to a round sphere of radius κ−1/2 .
Remark 5.15. When X0 sits isometrically in R3 , a classical rigidity theorem of Liebmann from 1899
[64, page 189] constant κ implies that κ > 0 and that therefore X0 is isometric to the round sphere of
radius κ−1/2 .
Proof of Theorem 7. ByR Proposition R5.13, the Rcondition that hX0 = hX gives that Volg (X) =
Volg0 (X0 ), X Kg = X0 Kg0 , and X Kg2 = X0 Kg20 . It follows that χ(X) = χ(X0 ), whence X
R
where (19) and (21) follow by the equalities observed above, (20) follows by Cauchy-Schwarz and the
fact that Kg0 is constant and (22) once again follows
R by Cauchy-Schwarz.
R The inequality in (22) is then
forced to be equality, so that Kg is constant and X Kg = X0 Kg0 forces Kg = Kg0 = κ. If κ > 0, then
the Uniformization Theorem implies that X and X0 are isometric to a round sphere of radius κ−1/2 .
36
Figure 10: Each figure shows generating vectors for a lattice in R2 together with the lattice they generate.
In each figure, several fundamental domains for flat tori are shown. The fundamental domains in each
figure have the same lattice points. Proposition 5.16 implies that the flat tori generated within each figure
have the same distance distributions, but that the distance distributions from the tori in the left figure are
different than those from the flat tori in the right figure.
Flat Tori
We now observe that Theorem 7 is optimal in that κ > 0 is necessary in order to ensure the isometry
conclusion. We will work out the κ = 0 case in detail and leave the κ < 0 case as an open question.
The Uniformization Theorem tells us that a Riemannian surface X of constant curvature κ = 0
is isometric to a flat torus; that is, a quotient of Euclidean R2 by a lattice L, or a discrete group of
translations isomorphic to Z2 . Let a lattice L be generated by vectors u and v. A rigid motion of R2
induces an isomorphism of X = R2 /L, so we assume without loss of generality that u is aligned with the
positive x-axis. We use T to denote a fundamental domain for L; e.g., T is the parallelogram spanned
by u and v. We define the set of lattice points P (L) of a lattice L to be the orbit of some fixed point
(say, the origin) under the action of L on R2 . For a second lattice L0 , we likewise assume that one of its
generating vectors is aligned with the positive x-axis, we use T 0 to denote its fundamental domain and
P (L0 ) its set of lattice points.
Proposition 5.16. Flat tori R2 /L and R2 /L0 have the same global distance distribution if and only if
P (L) = P (L0 ).
Proof. We begin with the observation that, by homogeneity of the flat torus X = R2 /L, the local
distance distribution hX (x, r) is independent of basepoint x. We will therefore focus our attention on
the particular point x ∈ X whose fiber in R2 consists of the lattice points of L. Throughout the following,
let T and T 0 be fixed fundamental domains for lattices L and L0 , respectively. We also set X 0 = R2 /L0 .
From the definition of X it follows that the local distance distribution hX (x, r) satisfies
where µL is Euclidean area measure normalized by the Euclidean area of T . Proving one direction of
the proposition then reduces to verifying that if lattices L and L0 have the same set of lattice points
P (L) = P (L0 ) =: P then
To verify (24), we observe that the group SL2 (Z) of 2 × 2 integer valued matrices with unit deter-
minant acts transitively on the set of lattices with lattice point set P by considering elements of SL2 (Z)
as linear transformations written in the basis {u, v} for R2 consisting of generating vectors for L. This
implies that µL = µL0 =: µ (the fundamental domains T and T 0 have the same area). Moreover, SL2 (Z)
is generated by the two elements
1 1 1 0
S1 = and S2 = .
0 1 1 1
37
The action of S1 on the fundamental domain T is by shearing parallel to the u-direction (while S2 shears
parallel to v). The desired claim follows if we are able to show that the quantity (23) is preserved by
these shearing transformations.
Let us now show that (23) is preserved by shears from the action of S1 ; the S2 case is similar.
Consider a line segment ` in R2 which is parallel to u and has length kuk. Let `t be a translation of ` by
t units along the u-direction (so ` = `0 ). We claim that the quantity
is independent of t, where len denotes 1-dimensional Hausdorff measure. Indeeed, we first observe that
it suffices to prove this claim when 0 ≤ t < 1, by periodicity of the lattice. The claim then follows by a
geometrical argument. We can express the set `t as
`t = (` ∩ `t ) ∪ (` \ `t + u),
len (`t ∩ ∪p∈P BR2 (p, r)) = len ((` ∩ `t ) ∩ ∪p∈P BR2 (p, r) ∪ (` \ `t + u) ∩ ∪p∈P BR2 (p, r))
= len ((` ∩ `t ) ∩ ∪p∈P BR2 (p, r)) + len ((` \ `t + u) ∩ ∪p∈P BR2 (p, r))
= len ((` ∩ `t ) ∩ ∪p∈P BR2 (p, r)) + len (` \ `t ∩ ∪p∈P BR2 (p, r)) (25)
= len (` ∩ ∪p∈P BR2 (p, r)) ,
−1 r1
q
2 2
πr − 2r cos + 2r1 r2 − r12
r
for r = r1 + δ, δ ≥ 0 sufficently small (this is just area of a ball of radius r minus the area of the inter-
section of two such balls when their centers are separated by 2r1 ). In this case, r1 could be determined
from hX (x, r) as the first radius at which the volume growth changes. Similarly, one can work out how
volume growth changes as triple (or higher) intersections are introduced—the exact formulas are not
relevant; the point here is that the number of intersecting balls which appear can also be determined
from the volume growth function. If a triple intersection is introduced at r = r1 , the critical radius
r1 is counted with multiplicity two, et cetera. The first three critical radii (counted with multiplicity)
give the sidelengths of a triangle with its vertices on lattice points. Triangles with these side lengths are
those non-degenerate lattice triangles with the smallest possible perimeters. We construct a particular
example of such a triangle by taking one of its vertices at the origin and another on the positive x-axis.
38
Tesselating the plane with this triangle by reflections and translations, one recovers the full set of lattice
points P (L). Since P (L) is determined by local distance distribution data, it follows that if X and X 0
have the same global distance distribution then P (L) = P (L0 ).
This result shows that it is possible for non-isomorphic Riemannian surfaces to have the same global
distribution of distance. However, in the flat torus case, surfaces with the same global distribution are
not close to one another in the moduli space of flat tori (cf. a similar result of Wolpert on isospectral flat
tori [95]). An answer to (IP4) on local injectivity of global distance distributions holds in the category
of surfaces in general remains an open problem.
Question 5.17. If a one-parameter family of Riemannian surfaces (Xτ , gτ ) has global distance distri-
butions HXτ which are constant in τ , must all surfaces in the family be isomorphic?
While Proposition 5.10 provides a result on global injectivity of the local distance distribution when
the underlying manifold is fixed, local injectivity of the global distance distribution in this restricted
setting remains open.
Question 5.18. If gτ is a smooth one-parameter family of metrics on S2 such that the global distance
distribution of the induced mm-spaces is constant in τ , must all gτ be isometric?
We suspect that the local injectivity holds at least for constant negative curvature surfaces. It seems
likely that an exact classification of higher genus constant curvature surfaces with the same distribution
can be obtained, but we leave the details as an open question.
Question 5.19. Is there a natural characterization of higher genus constant curvature surfaces with the
same distribution of distance in terms of fundamental domains in the Poincaré disk?
Every point of a metric graph falls into exactly one of these classifications—see Figure 11 for an illus-
tration. An edge of a metric graph is the image of an isometric embedding of a closed interval whose
endpoints get mapped to nodes. For x ∈ G, we write deg(x) for the degree of x, and declare deg(x) = 2
if x is an interior edge point.
Remark 6.1. Since we assume that a metric graph is compact, it has finitely many edges and nodes [54,
Proposition A.1].
Remark 6.2. In Section 6.1, we will allow finitely many interior edge points of a metric graph to be
designated as degree-2 nodes. See below for details.
39
Figure 11: Metric Tree
Metric graphs and trees arise naturally in several applications, serving as models for systems of
earthquake fault lines, GPS traces of vehicles and stress cracks in materials (see [1]). Metric trees are of
particular interest, as they appear in geometric group theory as the simplest 0-hyperbolic spaces in the
sense of Gromov [50], in data science as targets for low-dimensional embeddings of datasets [57], in
computational anatomy as models for blood vessels [35,36] and in shape analysis as merge trees [52,74].
Theorem 8. If a metric graph G has the same global distance distribution as the length-L cycle graph
then it is isomorphic to the cycle graph.
Corollary 6.3 (Sphere Characterization for Metric Graphs (IP3)). Let C denote the full subcategory of
mm-spaces whose objects are metric graphs. Then LCH (G, CL ) = 0 implies G ≈ CL .
Remark 6.4. Sturm shows in [88, Proposition 8.5] that (using our terminology) the cycle graph is dis-
tinguished from all other “balanced” length spaces by its local distance distribution. The term balanced
means that the local distance distribution of the length space is basepoint independent. The theorem
(and its corollary) above considers a more restrictive category (metric graphs versus length spaces), but
does not assume the balanced condition and moreover uses the much weaker assumption of equality of
global distance distributions.
For this subsection, it will be convenient to relax the definition of definition of node given in the
previous subsection to allow a finite number of interior edge points to be designated as nodes with
degree-2. Once these synthetic nodes have been added to the node set, the edge set of the metric graph
is expanded accordingly by subdivision. Making this relaxation allows us to accurately refer to the
underlying combinatorial structure of a metric graph with self-loops. In particular, by adding at least
two degree-2 nodes, the length-L cycle graph CL has a well-defined underlying combinatorial graph
structure. Throughout the rest of this section, we use the notation Nk (G) to denote the set of degree-k
nodes of G and we use N (G) to denote the set of all nodes of G (i.e., the union of all Nk (G)).
The proof of Theorem 8 is handled in below in Propositions 6.9 and 6.10. We first collect some
auxiliary results.
40
r
δ
x
v
dG (x, v)
Figure 12: Explanation of (26). The figure shows a portion of a metric graph G, a point x in the graph
and the metric ball BG (x, r), highlighted in blue. The point x is within distance r from a node (split
point) v. The indicated distances, with δ = r − dG (x, v) show where the formula (26) comes from.
Preliminary Results
For a metric graph G = (G, dG , µG ) (whose node set has been expanded as necessary to provide a
well-defined underlying combinatorial graph structure), we define the combinatorial invariant q(G) by
X
q(G) := deg(v)2 − 4|E|,
v∈N (G)
where |E| denotes the number of edges in G. We have the following simple observation that this quantity
is invariant under the addition of degree-2 nodes. The proof is elementary and left to the reader.
Lemma 6.5. Let G be a metric graph with node set N (G) and let N 0 (G) and N 00 (G) be node sets of G
after adding finitely many degree-2 nodes to G. Let E 0 and E 00 be their corresponding edge sets. Then
X X
deg(v)2 − 4|E 0 | = deg(v)2 − 4|E 00 |.
v∈N 0 (G) v∈N 00 (G)
In this subsection, we use lenG to denote 1-dimensional Hausdorff (length) measure on G. Let
L = L(G) = lenG (G), so that the probability measure on G is given by µG = L1 lenG .
Lemma 6.6. Let G be a metric graph with shortest edge length r0 . For r < r0 , the global distance
distribution of G is given by
2r r2
HG (r) = + 2 q(G).
L L
Proof. The proof is a computation. If x ∈ G lies within distance r of a node v ∈ N (G), then this node
is unique. In this case
41
v1
v2
G2
G1 G = G1 ∨v G2
Figure 13: Example of a wedge product. (Left) A pair of metric graphs G1 and G2 with chosen nodes
vj ∈ N (Gj ). (Right) The wedge product metric graph G = G1 ∨v G2 .
and this completes the proof of the lemma, modulo some explanation of the computations. The first term
in (27) comes from the total length of the set {x ∈ G | dG (x, N (G))} and the second term comes from
the change of variables ρ = dG (x, v) for each fixed v; this allows us to rewrite the integral in the second
term as an integral over each segment emanating from the vertex v. The first term in (28) comes from
the basic graph theory fact that the sum of vertex degrees is equal to twice the number of edges in any
combinatorial graph. Finally, (29) follows by applying this fact again and simplifying.
We now introduce an operation on metric graphs called the wedge product, which is a case of the
metric gluing construction of [2]. Given two metric graphs G1 and G2 and choices of nodes vj ∈ N (Gj )
(generalized degree-2 nodes are allowed), we form a new graph G by identifying v1 = v2 =: v̄ and
gluing G1 and G2 along v̄ and endow it with shortest path distance and normalized length measure (see
Figure 13). We use the notation G = G1 ∨v̄ G2 . For wedge products we use deg(v) to denote the degree
of a node in G and degj (v) to denote the degree in either of the component graphs.
Lemma 6.7. For a metric graph G constructed as a wedge product G = G1 ∨v̄ G2 ,
42
Proof. This is a straightforward computation:
X
q(G) = deg(v)2 − 4|E|
v∈N (G)
X X
= deg(v̄)2 + deg1 (v)2 + deg2 (v)2 − 4|E1 ∪ E2 |
v∈N (G1 )\{v̄} v∈N (G2 )\{v̄}
2 2
= deg1 (v̄) + deg2 (v̄) + 2deg1 (v̄)deg2 (v̄)
X X
+ deg1 (v)2 + deg2 (v)2 − 4|E1 | − 4|E2 |
v∈N (G1 )\{v̄} v∈N (G2 )\{v̄}
X X
2 2
= deg1 (v) − 4|E1 | + deg2 (v) − 4|E2 | + 2deg1 (v̄)deg2 (v̄)
v∈N (G1 ) v∈N (G2 )
We now focus on the case of metric trees. There are two families of metric trees which play a special
role. A line graph is a metric tree which is isomorphic to a closed interval. A Y -graph is a metric tree
which can be expressed as T = T1 ∨v̄ T2 , where T1 and T2 are line graphs and v̄ is an interior edge point
in T1 and a leaf node in T2 .
Lemma 6.8. For a metric tree T ,
= −2 if T is a line graph,
q(T ) =0 if T is a Y -graph,
>0 otherwise.
Proof. First suppose that T is a line graph. Since q(T ) is invariant under addition of degree-2 interior
edge nodes, we can assume without loss of generality that T consists of two nodes joined by a single
edge. Then
q(T ) = 12 + 12 − 4 · 1 = −2.
Next we claim that if T is a metric tree then q(T ) ≥ −2, with q(T ) = −2 only if T is a line graph. We
prove this claim by induction on |N (T )|. The base cases are n = 2 and 3, where the only metric trees
are line graphs and we are done. Suppose that the claim holds for all metric trees with n − 1 or fewer
nodes and let T be a metric tree with n nodes. Choose a leaf of T and decompose T as T = T1 ∨v̄ T2 ,
where T1 consists of a leaf edge (and is therefore a line graph), v̄ is the attaching node of the leaf edge
and T2 is the remainder of T . Then T1 and T2 each have ≤ n − 1 nodes and we have, by Lemma 6.7,
q(T ) = q(T1 ) + q(T2 ) + 2deg1 (v̄)deg2 (v̄) ≥ −2 + −2 + 2 · 1 · 1 = −2,
with equality if and only if T2 is a line graph and v̄ is a leaf node in T2 , hence T is itself a line graph.
Next we consider the case that T = T1 ∨v̄ T2 is a Y graph. Then Lemma 6.7 implies
q(T ) = q(T1 ) + q(T )2) + 2 deg1 (v̄) deg2 (v̄) = −2 + −2 + 2 · 2 · 1 = 0.
By an inductive argument similar to the above, we can show that if T is neither a line graph nor a
Y -graph then q(T ) > 0.
Proof of Theorem 8
Recall that the first Betti number β1 (X) of a topological space X is the rank of its first singular homology
group. For a metric graph G the first Betti number has a simple formula:
β1 (G) = |E| − |N (G)| + 1.
It is easy to see that this quantity is invariant under adding finitely many degree-2 nodes to the node set.
We now prove Theorem 8 by checking two cases: β1 (G) ≥ 1 and β1 (G) = 0.
43
Proposition 6.9. If G is a metric graph with β1 (G) ≥ 1 and with the same global distance distribution
as the length-L cycle graph, then G is isomorphic to the cycle graph.
Proof. If HG = HCL then Lemma 6.6 implies that L(G) = L and q(G) = q(CL ) = 0. We have
P 2
X v∈N (G) deg(v)
0 = q(G) = deg(v)2 − 4|E| ≥ − 4|E| (30)
|N (G)|
v∈N (G)
4|E|2
|E|
= − 4|E| = 4|E| −1 (31)
|N (G)| |N (G)|
≥ 0. (32)
where (30) follows by Cauchy-Schwartz, (31) follows from the fact that sum of vertex degrees is equal
to twice the number of edges in any graph and (32) follows from the assumption that β1 (G) ≥ 1, whence
|E| ≥ |N (G)|. Equality in Cauchy-Schwartz is then forced, and we conclude that the sequence of vertex
degrees for G must be constant. Since all computations are invariant under adding synthetic degree-2
vertices to the node set, all node degrees in G must be equal to 2 and it follows that G ≈ CL .
Proposition 6.10. If G is a metric graph with β1 (G) = 0 then G does not have the same global distance
distribution as the length-L cycle graph.
Proof. Suppose on the contrary that HG = HCL and β1 (G) = 0. This implies that G is a metric tree.
By Lemma 6.6, L(G) = L and q(G) = q(CL ) = 0. Lemma 6.8, together with the latter observation,
immediately implies that G is a Y -graph. Suppose that G = T1 ∨v̄ T2 and let Lj = L(Tj ). We also
observe that G must have the same diameter as CL , namely L/2. Let w be the unattached leaf node of
T2 (i.e., w 6= v) and let u1 and u2 be the leaf nodes of T1 . Then the diameter of G is at least
max dG (w, uj ) = dG2 (w, v̄) + max dG1 (v̄, uj ) = L2 + max dG1 (v̄, uj ) ≥ L2 + L1 /2.
j=1,2 j=1,2 j=1,2
Thus
(L1 + L2 )/2 = L/2 ≥ L1 /2 + L2 ⇒ L1 + L2 ≥ L1 + 2L2 ⇒ L2 = 0,
hence T2 is a single node and T = T1 is a line graph, contradicting q(T ) = 0.
Theorem 9. Let G and H be metric graphs, with G not a cycle graph. Suppose that there exists a
continuous measure-preserving map φ : G → H such that
hG (x, r) = hH (φ(x), r)
We will provide examples below which show that it is not possible to weaken the assumptions of the
theorem.
Proof. First note that φ is surjective. Indeed, suppose that y ∈ H does not lie in the image of φ.
Since φ is continuous, it takes G to a compact subset of H, which is necessarily closed. Then there is
some open neighborhood U of y not contained in the image of φ. This open neighborhood has positive
measure, but φ−1 (U ) is empty, contradicting the measure-preserving assumption. We claim that φ is
also injective, which will imply that φ is a homeomorphism, since G and H are both compact and
Hausdorff [75, Theorem 26.6]. To obtain a contradiction, suppose that x, x0 ∈ G are distinct points with
φ(x) = φ(x0 ) =: y. There are two cases to consider, treated below.
44
Case 1: Interior Edge Points. First suppose that x is an interior edge point. By the assumption that φ
preserves local distance distributions, x0 and y must also be interior edge points. Let α : [0, a] → G be
the geodesic from α(0) = x to its nearest node α(a), at distance a from x. Let A ⊂ G denote the image
of α; observe that A is isomorphic to the interval [0, a]. Consider the path φ ◦ α : [0, a] → H. We have:
• φ ◦ α(0) = y,
• φ ◦ α(a) is a node,
• φ ◦ α is injective.
Indeed, the first point follows by definition, while the second and third follow immediately from preser-
vation of local distance distribution. The last point also follows from distance distribution preservation:
if φ ◦ α(t1 ) = φ ◦ α(t2 ) for some t1 , t2 ∈ [0, a], then
but α is injective and each point in A has a unique local distance distribution (the smallest radius r at
which h0G (α(t), r) 6= 2 is different at each point α(t) ∈ A). This implies that t1 = t2 .
So far we have shown that φ|A : A → φ(A) is a homeomorphism. The image φ(A) is also iso-
morphic to a closed interval [0, `] for some length `. The measure-preserving assumption on φ implies
` ≥ a, as
` = µH (φ(A)) = µG φ−1 (φ(A)) ≥ µG (A) = a.
On the other hand, we can show that ` > a gives a contradiction. Indeed, in this case choose a point y 00
in φ(A) at distance 0 < δ < ` − a from y. Its unique preimage x00 in A necessarily has h0G (x00 , r) 6= 2
at some r < a. We claim that h0H (y 00 , r) = 2 for all r ≤ a. To see this last claim, observe that the
distance to the neighboring node of y 00 which is contained in φ(A) is ` − δ > a. The distance to the
other neighboring node of y 00 must be at least a + δ; otherwise y had exactly one neighboring node at
distance strictly less than a, meaning that the slope of hH (y, r) is different from 2 at some r < a and
this violates preservation of local distance distribution.
Similar to the above, we define α0 : [0, a0 ] → G to be the geodesic from x0 to its nearest node at
distance a0 , with image A0 . While it is feasible that A ∩ A0 6= ∅, we do not have A = A0 since x and
x0 are distinct. Without loss of generality, assume that a ≥ a0 . Running the above arguments again,
we see that φ|A0 is also a homeomorphism onto its image φ(A0 ), which must be a segment in H with
y as one endpoint and a node at distance a0 from y as the other endpoint. There are two possibilities:
φ(A) = φ(A0 ) or not.
If φ(A) = φ(A0 ), then a = a0 and
≥ µ G A ∪ A0
= µG (A) + µG (A0 ) − µG A ∩ A0
= 2a − µG A ∩ A0 > a.
The latter inequality follows because A and A0 are both homeomorphic to closed intervals, so that A∩A0
is as well. Then A 6= A0 and µG (A) = µG (A0 ) = a together ensure µG (A ∩ A0 ) < a. It is therefore
impossible that φ(A) = φ(A0 ).
On the other hand, suppose φ(A) 6= φ(A0 ). Since φ(A) and φ(A0 ) are each homeomorphic to a
closed interval with y as an endpoint and a node as the other endpoint, it follows that φ(A)∩φ(A0 ) = {y}
and φ(A) ∪ φ(A0 ) is an edge of length a + a0 . In this case, we define β (respectively, β 0 ) to be the path
emanating from x (respectively, x0 ) of length a (respectively, a0 ) in the direction opposite that of α
(respectively α0 ). Let B and B 0 denote the images of these paths; note that they are isomorphic to
45
intervals and have no nodes in their interiors, by the assumption that a (respectively, a0 ) was the distance
from x (respectively, x0 ) to its nearest node. By preservation of local distance distributions, φ(B) and
φ(B 0 ) are both contained in the edge φ(A) ∪ φ(A0 ). We now derive a contradiction. Assume for the
sake of simplicity that a ≥ a0 . Then
2a ≥ a + a0 = µH φ(A) ∪ φ(A0 )
≥ µG (A ∪ B ∪ A0 ∪ B 0 )
= µG (A ∪ B) + µG (A0 ∪ B 0 ) − µG (A ∪ B) ∩ (A0 ∪ B 0 )
= 2a + 2a0 − µG (A ∪ B) ∩ (A0 ∪ B 0 )
> 2a.
To see the last inequality, first note that A0 ∪ B 0 is not contained in A ∪ B: equality would force x = x0 ,
while strict containment is impossible because the interior of A ∪ B contains no nodes. Thus
Therefore φ is a 1-Lipschitz map. Running the same argument on φ−1 , we see that it is also 1-Lipschitz.
It follows that φ is an isometry and this completes the proof.
Corollary 6.11 (Global Injectivity of Local Distance Distributions for Metric Graphs (IP1)). Let C be
the subcategory of mm-spaces whose objects are metric graphs and whose morphisms are continuous
measure preserving maps. Then LCh (G, H) = 0 implies G ≈ H.
Proof. Let G and H be metric graphs with LC h (G, H) = 0. If G is a cycle graph, we observe that
C C
Lh (G, H) = 0 implies LH (G, H) = 0, so that Corollary 6.3 implies G ≈ H. Suppose, then, that G is
not a cycle graph and let {φn } be a sequence of continuous measure-preserving maps φn : G → H such
that Z
lim cG,H (x, φn (x)) µG (dx) = 0.
n→∞ G
Since G and H are compact metric spaces, the family {φn } is obviously pointwise precompact. From
the proof of Theorem 9, we see that each φn is a 1-Lipshitz map and it follows that the family {φn }
is equicontinuous. The Arzelá-Ascoli theorem then implies that a subsequence of {φn } converges to
some continuous map φ : G → H. This map must also be measure-preserving: for any measurable set
A ⊂ H and any > 0, there exists n such that
46
Figure 14: A pair of non-isomorphic combinatorial trees with the same global distance distribution—
see [68].
We consider these as metric trees by declaring each edge to be isometric to a unit interval. The trees T1
and T2 are nonisomorphic. However, one can check explicitly that Lh (T1 , T2 ) = 0 by finding a measure
preserving mapping realizing this value. One such map is encoded by
A1 A2 A3 A1 B 3 C 1
B1 B2 B3 7→ B1 C2 A2
C1 C2 C3 A3 B 2 C 3
For example, all edges in lobe A2 of T1 are mapped isometrically to the edges in lobe B3 of T2 . The
mapping is the identity map on the interior edges. One can check that this map φ satisfies hT2 (φ(x), r) =
hT1 (x, r) for all r, with the key point being that the row sums of the matrices in (33) are all equal.
Finally, we provide examples which show that other assumptions of Theorem 9 cannot be dropped.
Example 6.14. Figure 16 shows trees T1 and T2 , which we take as metric trees by declaring dT1 (b, aj ) =
1/3 and dT2 (B, A) = 1. We define a map φ : T1 → T2 by φ(aj ) = A and φ(b) = B then extending lin-
early across edges. This map is continuous and measure-preserving, but obviously not an isomorphism.
The map does not preserve the local distance distribution; indeed, no point in T2 has the same local
distance distribution as point b in T1 .
47
A3 B3
A2 B2
A1 B1
C1 C3
C2
Figure 16: There is a continuous measure-preserving map from T1 to T2 which is not an isomorphism.
Example 6.15. Figure 17 shows a pair of graphs G1 and G2 . These become metric graphs by declaring
all edges of G1 to be isometric to the interval [0, 1/2] and all edges of G2 to be isometric to the unit
interval. We define a map φ : G1 → G2 by φ(aj ) = A, φ(bj ) = B, φ(cj ) = C and φ(dj ) = D,
j = 1, 2. This map is continuous and satisfies hG2 (φ(x), r) = hG1 (x, r) for all x ∈ G1 . Clearly φ is
not an isomorphism, and we see that the assumption that G is not isomorphic to a circle is necessary in
Theorem 9.
Figure 17: There is a continuous measure-preserving map from G1 to G2 which preserves local distance
distributions but which is not an isomorphism.
48
Figure 18: Distance distributions for metric graphs. The top row shows several examples of simple
metric graphs and the bottom row shows the right-sided derivatives of their global distance distribu-
tions.Observe that jump discontinuities correspond to geodesic loops in the graphs. The first two ex-
amples are homotopy equivalent and their distributions share the same number of jump discontinuities.
The last two examples are homotopy equivalent, but have a different number of jump continuities in
their distributions: the last example (which is homeomorphic to the 1-skeleton of a tetrahedron) has 4
geodesic loops, even though its first Betti number is 3. We see that counting discontinuities is not enough
to show that metric graphs are distinguished up to homotopy type by their global distance distributions.
Theorem 10. Let G and H be metric graphs. If LUh (G, H) = 0, then G and H are homotopy equivalent.
From the theorem, we immediately obtain the following.
Corollary 6.16 (Local Distance Distribution Characterizes Homotopy Type of Metric Graphs (IP2)).
Let C be the full subcategory of mm-spaces whose objects are metric graphs. Then LCh (G, H) = 0
implies G and H have the same homotopy type.
Remark 6.17. The corollary is stated in the language of inverse problem (IP2) for the sake of cohesive-
ness. However, Theorem 10 is stronger than the corollary, since it makes the weaker assumption that the
Kantorovich-style pseudometric LUh vanishes.
The above results show that the local distance distribution characterizes homotopy type of metric
graphs, but we leave it as an open question whether homotopy type is detected by global distance distri-
butions.
Question 6.18. Is the homotopy type of a metric graph characterized by its global distance distribution;
that is, if LH (G, H) = 0 for metric graphs G and H, must it be that G and H are homotopy equivalent?
Remark 6.19. We conjecture that the answer to the question is “yes”. This is supported by numerical
experiments and some feasible proof strategies:
1. Suppose HG = HH . One could construct Riemannian surfaces as tubular neighborhoods of G
and H after embedding them into some Euclidean space. These tubular neighborhood surfaces
would then have arbitrarily close (by taking the tube radii arbitrarily small) distance distributions
with respect to geodesic distance, and one could attempt to apply Corollary 5.14 to conclude that
the surfaces (hence the original metric graphs) have the same Euler characteristic.
2. One can show that dHG is piecewise linear, with jump discontinuities corresponding to geodesic
loops—or isometric embeddings of circles—in G. We can use this to show that if G and H both
have the property that their geodesic loops give a basis for their respective first homology vector
spaces over some field, then HG = HH implies G and H are homotopy equivalent. However, this
49
property is not enjoyed by arbitrary metric graphs, so the general result cannot be obtained from
this argument without more work. See Figure 18 for some examples.
We were unable to fully work out the technical details of these strategies and leave this question as a
direction for future research.
We refer to this collection as the node multiset of G. The node multiset is an isomorphism invariant of a
metric graph.
Proposition 6.20. Let G and H be metric graphs. If hN (G) 6= hN (H) as multisets of functions, then
LUh (G, H) 6= 0.
The proof will use some technical lemmas.
Lemma 6.21. Let X and Y be mm-spaces and let f : X → R and g : Y → R be continuous maps. For
each µ ∈ M(µX , µY ), we have
Z
|f (x) − g(y)| µ(dx × dy) ≥ dRW,1 (f# µX , g# µY ).
X×Y
Proof. Let f × g : X × Y → R × R be the product map (f × g)(x, y) = (f (x), g(y)) and consider the
measure (f × g)# µ ∈ P(R2 ). It is easy to check that (f × g)# µ is a measure coupling of f# µX and
g# µY . The change of variables formula then implies that
Z Z
|f (x) − g(y)| µ(dx × dy) = |u − v| (f × g)# µ (du × dv)
X×Y R×R
≥ dRW,1 (f# µX , g# µY ).
For a mm-space X and a fixed r > 0, let hrX : X → R denote the function hrX (x) = hX (x, r). We
will consider the one-parameter family of measures {(hrX )# µX }r≥0 ⊂ P(R). We note that this family
of measures appears in the definition of the modulus of mass distribution studied in [49].
Lemma 6.22. For mm-spaces X and Y, LUh (X , Y) = 0 implies that (hrX )# µX = (hrY )# µY for almost
every r > 0 (with respect to Lebesgue measure).
Proof. If LUh (X , Y) = 0 then for any > 0 there exists a coupling µ ∈ M(µX , µY ) such that
Z
> cX ,Y (x, y) µ(dx × dy)
X×Y
Z Z ∞
= |hX (x, r) − hY (y, r)| dr µ(dx × dy)
X×Y 0
Z ∞ Z
r r
= hX (x) − hY (y) µ(dx × dy) dr (34)
0 X×Y
Z ∞
dRW,1 (hrX )# µX , (hrY )# µY dr,
≥ (35)
0
50
where (34) follows by Fubini’s Theorem and (35) follows from Lemma 6.21. We deduce that
for almost every r > 0. This completes the proof, since Wasserstein distance is a metric on the space of
compactly supported probability measures on R.
We now prove our main technical lemma for Proposition 6.20. We use |A| to denote the cardinality
of a set A.
Lemma 6.23. Let G and H be metric graphs. If LUh (G, H) = 0 then |Nk (G)| = |Nk (H)| for all k. It
follows that LUh (G, H) = 0 implies |N (G)| = |N (H)|.
Proof. Suppose that LUh (G, H) = 0. Lemma 6.22 implies that we can choose an r which is less than
half the length of the shortest edge in G or H such that (hrG )# µG = (hrH )# µH .
We first show that the number of leaves of G can be recovered from (hrG )# µG :
r r r
· |Nk (G)| + · |Nk+1 (G)| + · · · + · |NkG (G)|, (36)
k−2 (k + 1) − 2 kG − 2
when k ≤ kG and that it is equal to zero otherwise. Assuming that the claim holds, the number of nodes
of each degree of G can be recovered recursively from (hrG )# µG , and this completes the proof.
It remains to derive (36). By definition,
By our choice of r, the maximum value of µG (BG (x, r)) is r · kG , when x is a node of degree kG . It
follows that hrG # µG ((k − 1) · r, k · r) = 0 when k ≥ kG + 1. We then consider k with 3 ≤ k ≤ kG .
A point x ∈ G satisfies (k − 1) · r < µG (BG (x, r)) < k · r if and only if it satisfies one of the following
mutually exclusive conditions, indexed by ` = k, k + 1, . . . , kT :
`−k `−k−1
r < dG (x, N` (G)) < r. (C` )
`−2 `−2
To see this, note that if µG (BG (x, r)) > 2r, then it must lie within distance r from a (unique) node of
degree at least 3. Suppose that x lies within distance < r to a node of degree `. Then
(k − 1) · r < ` · r − (` − 2) · < k · r,
and solving for shows that condition (C` ) must hold. The set of points satisfying this condition has
measure
`−k−1 `−k r
|N` (G)| · r− r = |N` (G)| · .
`−2 `−2 `−2
Adding up these measures for ` = k, k + 1, . . . , kG , we obtain (36).
51
Proof of Proposition 6.20. Suppose that hN (G) 6= hN (H). We first consider the case that there is some
v ∈ N (G) such that hG (v, ·) 6= hH (w, ·) for all w ∈ N (H). If this is the case, then hG (v, ·) 6= hH (y, ·)
for any point y in H, as hG (v, r) will differ from any hH (y, r) corresponding to a non-node y for small
values of r. It follows that the continuous function H → R≥0 on the compact space H defined by
y 7→ cG,H (v, y) achieves its minimum mv > 0. We claim that there exists an open neighborhood U of
v in T such that for any x ∈ U ,
mv
inf cG,H (x, y) ≥ .
y∈S 2
If this is not the case then there is a point x ∈ G such that cG,G (v, x) < mv /2 and cG,H (x, y) < mv /2
for some y ∈ H. It is not hard to see that the cost function satisfies a triangle inequality-like relation for
all v, x ∈ G and y ∈ H:
cG,H (v, y) ≤ cG,G (v, x) + cG,H (x, y).
Applying this to our particular points v, x and y, it follows that
which is a contradiction.
Now we observe that for any coupling µ ∈ U(µG , µH ),
Z Z
mv mv
cG,H (x, y) µ(dx × dy) ≥ cG,H (x, y) µ(dx × dy) ≥ · µ(U × H) = µG (U ).
G×H U ×H 2 2
Proof of Theorem 10. Let G be a metric graph. From the node multiset hN (G), one can determine the
number of nodes of G: it is simply the cardinality of hN (G) (with multiplicity). From a function hG (x, ·)
centered at a node x of G, one can read off the degree of x as limr→0+ hG (x, r). From hN (G), we can
therefore determine the sum of node degrees of G, which is equal to twice the number of edges of G. It
follows that knowledge of the node multiset hN (G) allows one to determine the Euler characteristic of
G.
Now let G and H be metric graphs with LUh (G, H) = 0. Then Proposition 6.20 implies hN (G) =
hN (H) as multisets. By the previous paragraph, G and H must have the same Euler characteristic. For
connected graphs, Euler characteristic determines homotopy type, so this completes the proof.
52
graphs up to isometry in the local sense (IP4). We are able to obtain such a result when restricting to
the subcategory of metric trees, where we have some extra tools at hand coming from the computational
topology literature. The last main result of the paper is the following.
Theorem 11. Local distance distributions distinguish metric trees
(a) locally (IP4): For every metric tree T , there exists T > 0 such that if a metric tree S satisfies
dGH (T , S) < T and LUh (T , S) = 0 then S is isomorphic to T ;
(b) densely and generically: For every metric tree T and every > 0, there exists a metric tree T 0
such that dGH (T , T 0 ) < and LUh distinguishes T 0 from all metric trees S not isomorphic to
T . Moreover, a metric tree is distinguished from all other metric trees by LUh almost surely with
respect to a natural measure on MTrees.
Remark 6.24. Theorem 11 result is analogous to Boutin and Kemper’s theorem on distance distributions
of point clouds, described above in Theorem 2. It is also similar to the main result of [77], where the
authors considered injectivity properties of a more complicated invariant of metric graphs based on
persistent homology.
This theorem allows us to promote LUh to a genuine metric on the space of metric trees, which
we denote MTrees.We define a distance on MTrees by setting it to be the length of the shortest path
between metric trees T and S in MTrees, measured according to LUh . More precisely, the length `(p) of
a continuous path p : [0, 1] → MTrees is given by
n
X
sup LUh (p(tj ), p(tj−1 )),
n,P j=1
with the supremum taken over natural numbers n and increasing partitions P = t0 < t1 < · · · < tn ,
t0 = 0, tn = 1. We define the intrinsic pseudo metric induced by LUh by
It is easy to check that dint satisfies the triangle inequality. Moreover, it can be symmetrized as in
Proposition 3.9. Finally, part (a) of the Theorem 11 implies that dint (T , S) = 0 if and only if T and S
are isomorphic. The proves the following corollary (cf. [32]).
Corollary 6.25. The symmetrization of dint is a metric on MTrees.
The height function f : epi(f ) → R is given by f (x, r) = r. One forms a topological space Mf =
epi(f )/ ∼, where (x, r) ∼ (x0 , r0 ) if and only if r = r0 and (x, r), (x0 , r0 ) ∈ f −1 (r) lie in the same
connected component of epi(f ). Then Mf has the structure of a tree and f induces a well-defined
function Mf → R by [x, r] 7→ r. We abuse notation and continue to denote this induced map by f .
Given a pair of merge trees Mf and Mg , one computes the interleaving distance di (Mf , Mg ) as
follows. For any > 0, there is a shift map σf : Mf → Mf which takes a point [x, r] at height r to
its unique ancestor [x0 , r + ] at height r + . An -morphism is a continuous map φ : Mf → Mg such
that g ◦ φ = f + . An -interleaving of Mf and Mg is a pair of -morphisms φ : Mf → Mg and
ψ : Mg → Mf such that
ψ ◦ φ = σf2 and φ ◦ ψ = σg2 .
53
When there exists such an -interleaving we say that Mf and Mg are -interleaved. We then define
Let T be a metric tree. For each x ∈ T , let fxT : T → R be defined by fxT (y) = −dT (x, y). We
can then form the merge tree MfxT , which we denote more succinctly as T x . In this case, the structure
of T is largely unchanged as we pass to the merge tree T x ; indeed, if x is not a leaf of T , then T x is a
rooting of T at x, with an extra edge extending to ∞ from x. If x is a leaf of T , then T x is a rooting at
the unique neighbor of x, once again with an infinite edge extending from the root.
For metric trees T and S, define
∆(T , S) := min di (T t , S s ).
t∈N (T ),s∈N (S)
Recent results of Agarwal et. al. [3] relate this quantity to the Gromov-Hausdorff distance between T
and S and give a short list of candidate values for ∆(T, S). These are summarized below.
Theorem 12 ( [3]). For metric trees T and S,
1
1. 2 dGH (T, S) ≤ ∆(T, S) ≤ 14 dGH (T, S);
for some sufficiently small > 0. For each of the rjx , ∂BT (x, r) necessarily contains a node. We
claim that for r 6∈ {r1x , . . . , rM
x }, ∂B (x, r) can only contain a node if it, in particular, contains a
x T
leaf. Indeed, if r lies in an interval (rjx , rj+1
x ), j = 0, 1, . . . , M (taking r x = 0 and r x
x 0 Mx +1 = ∞,
for notational convenience), and ∂BT (x, r) contains only non-leaf nodes, then, for sufficiently small
> 0, the number of non-leaf elements of ∂BT (x, r + ) is strictly greater than the number of non-leaf
elements of ∂BT (x, r). This contradicts the fact that the function h0T (x, ·) is constant on such intervals.
Now suppose that x is a node of T . The discussion of the previous paragraph allows us to distinguish
a finite list of candidate values r where ∂BT (x, r) can contain a node. In the lemma and in the rest of
the section, we use the term leaf edge to refer to an edge in T containing a leaf as an endpoint.
54
Lemma 6.27. Let T be a metric tree and let r1x , . . . , rM
x be as defined above. Let ` , . . . , ` denote the
x 1 N
lengths of all leaf edges of T . The sphere ∂BT (x, r) can only contain a node if r lies in the set
Proof. Suppose that ∂BT (x, r) contains a node and that r 6∈ {r1x , . . . , rM x }. By the above discussion,
∂BT (x, r) must contain a leaf v1 . Let γ1 denote the unique path joining x to v1 . Let x1 denote the
node lying on γ1 which immediately precedes v1 , let s1 = dT (x, x1 ) and let `j1 = dT (x1 , v1 ). Since
v1 is a leaf, `j1 is the length of its leaf edge. There are several cases to consider. If x1 = x, then
we are clearly finished, as r = `j1 . If s1 = rjx1 ∈ {r1x , . . . , rM x }, then we are finished because this
x
implies r = rjx1 + `j1 ∈ ΣT (x). We therefore assume that we are in neither of these situations and
iterate the process. That is, we have a node x1 ∈ ∂BT (x, s1 ), where s1 6∈ {r1x , . . . , rM
x }, and it follows
x
that ∂BT (x, s1 ) contains a leaf v2 . Let γ2 denote the path from x to v2 , let x2 denote the node on γ2
preceding v2 , let s2 = dT (x, x2 ) and let `j2 = dT (x2 , v2 ). The algorithm terminates at this stage if
x2 = x (in which case r = `2 + `1 ) or if s2 = rjx2 ∈ {r1x , . . . , rM
x } (in which case r = r x + ` + ` )
x j2 j2 j1
and otherwise iterates again. The algorithm must eventually terminate, since the distances sj in each
step are strictly decreasing.
Proof of Proposition 6.26. From the multiset hN (T ), we are able to extract all possible values of r for
which the function h0T (x, ·) is discontinuous, for each x ∈ N (T ). That is, using the notation of Lemma
6.27, we can determine the set
[
{r1 , . . . , rM } = {r1x , . . . , rM
x
x
}
x∈N (T )
Moreover, we can extract the set of lengths {`1 , . . . , `N } of all leaf edges of T from the multiset hN (T ).
Indeed, first note that a function in hN (T ) can be distinguished as the function of a leaf by the obser-
vation that h0T (x, 0) = 1 if and only if x is a leaf. Next use the fact that the length of the leaf edge of a
leaf x is the smallest value of r > 0 such that h0T (x, r) 6= 1. The multiset hN (T ) therefore determines
the set
ΣT = {λ1 r1 + · · · + λM rM + µ1 `1 + · · · + µN `N | λj , µk ∈ {0, 1}}.
In particular, ΣT contains the union of all ΣT (x) for x ∈ N (T ) (see (37)). In general, this containment
is strict.
Fix a node t ∈ N (T ) and consider the set
1 T T t
Λ1,1 = f (u) − f t (v) | u, v ∈ N (T )
2 t
defined in Theorem 12. Lemma 6.27 implies that for any node u ∈ N (T t ), the value −f tT (u) lies in
the set ΣT (t). It follows that
1 0 0
Λ1,1 ⊂ A − A | A, A ∈ ΣT (t)
2
1 0 0
⊂ A − A | A, A ∈ ΣT =: Σ1
2
If we assume that hN (T ) = hN (S), then the discussion from the previous paragraph implies that
ΣT = ΣS , and we have similar statements for the other sets Λi,j from Theorem 12:
1 0 0 1 0 0
Λ2,2 ⊂ B − B | B, B ∈ ΣS (t) ⊂ B − B | B, B ∈ ΣT = Σ1 ,
2 2
Λ1,2 ⊂ {|A − B| | A ∈ ΣT (t), B ∈ ΣS (s)} ⊂ {|A − B| | A, B ∈ ΣT } =: Σ2
55
Combining these inclusions with part 2 of Theorem 12, we have
∆(T , S) = min di (T t , S s ) ∈ Σ1 ∪ Σ2 .
t∈N (T ), s∈N (S)
Finally, let
1
T = min (Σ1 ∪ Σ2 \ {0}) .
14
If hN (T ) = hN (S) and dGH (T , S) < T , then part 1 of Theorem 12 implies that
By the discussion of the previous paragraph, ∆(T , S) ∈ Σ1 ∪ Σ2 , and it follows that ∆(T , S) = 0.
Applying the lower bound of Theorem 12, we conclude that dGH (T , S) = 0 and it follows that T and
S are isomorphic.
Proposition 6.28. The set of metric trees T which are determined up to isomorphism by the multiset
hN (T ) is dense in MTrees with respect to Gromov-Hausdorff distance and is full measure with respect
to ρ.
To prove the proposition, we will require some additional notation and preliminary lemmas. Let
T = (T, dT , µT ) be a metric tree. We will define a nested sequence of metric subforests Fk of T . That
is, each Fk = (Fk , dFk , µFk ) consists of a subset Fk ⊂ T such that each connected component of Fk is
a measured metric tree whose measure is allowed to have total volume less than one. The metric dFk is
geodesic distance in Fk with respect to the restriction of dT ; in particular, the distance between points
in distinct connected components of Fk is ∞, as there is no geodesic joining them. The measure µFk is
simply the restriction of µT .
The subset F0 defining the metric forest F0 consists of only the leaves of T . To define F1 , we
include the leaf edges of the leaves in F0 and their opposite endpoints. That is, F1 consists of leaves,
leaf edges and leaf parents—these are non-leaf nodes which are incident on an edge containing a leaf.
Let V1 denote the set of leaf parents v ∈ F1 with exactly one incident edge in T that is not a leaf edge.
The subforest F2 is formed from F1 by including these incident edges for each v ∈ V1 , together with
their endpoint nodes. We now continue our definition inductively. Assume that Fk has been defined.
Let
Vk = {v ∈ N (Fk ) | degT (v) = degFk (v) + 1},
where degT (v) denotes the degree of v in the full tree T and degFk (v) the degree of v in the subforest.
Then Fk+1 contains Fk , together with the single extra edge for each v ∈ Vk and endpoint nodes of these
edges. An example of this sequence shown in Figure 19.
Lemma 6.29. Let T be a metric tree with the sequence of subforests Fk defined as above. If Fk 6= T ,
then Vk is nonempty.
56
Figure 19: A nested subforest sequence F0 , F1 , F2 , F3 = T .
Proof. To obtain a contradiction, assume that Vk is empty. Enumerate the connected components of
Fk as T1 , . . . , TN . Since Fk 6= T and T is connected, there must be some node vj in each Tj with
degT (vj ) > degFk (vj ), whence degT (vj ) ≥ degFk (vj ) + 2 by our assumption that Vk = ∅. Choose
an edge incident with v1 which is not contained in T1 and consider any path which begins at v1 , passes
through this edge, and continues with increasing distance from v1 . Such a path must terminate in some
leaf, and said leaf cannot lie in T1 by the increasing distance condition. Assume without loss of general-
ity that the path terminates at a leaf contained in T2 . The path must then pass through v2 , by construction
of Fk . By our assumption, we can choose another edge which is incident with v2 , is not contained in
the path from the previous step, and is not contained in T2 . Now choose a new path which starts at v2 ,
passes through the newly chosen edge, and continues by increasing distance from v2 . The new path will
terminate in a leaf not contained in T2 . If its terminal leaf lies in T1 , then we concatenate with the path
from the first step to produce a non-trivial cycle in T , thereby obtaining a contradiction. We therefore
assume without loss of generality that the new path terminates at a leaf contained in T3 . Continuing with
this process, a nontrivial cycle must eventually be formed because there are finitely many Tj , and we
obtain a contradiction.
Corollary 6.30. With T and Fk defined as above, there exists some N such that Fk = T for all k ≥ N .
Proof. If Fk 6= T , then Vk 6= ∅, by Lemma 6.29. It follows that an edge is added to form Fk+1 . Since
there are finitely many edges in T , it must be that FN = T for some N .
We remark that it is possible to show that the number of trees in Fk+1 is strictly less than the number
of trees in Fk whenever Fk 6= T , and it follows that N is at most the number of leaves in T .
Proof. Since T is assumed to be connected, one direction is clear. Conversely, suppose that Fk 6= T .
Then we can choose some v ∈ Vk , by Lemma 6.29. Let e denote the single edge which is incident
with v but which is not contained in Fk and let v 0 denote the opposite node of e. If v 0 ∈ Fk , then Fk is
disconnected. Otherwise, choose a path which starts at v, passes through e and continues with increasing
distance from v. The path terminates in a leaf, and the leaf necessarily lies in a path component of Fk
which is different than that of v.
Lemma 6.32. If all functions in the multiset hN (T ) are distinct, then T is determined by hN (T ) up to
isomorphism.
Proof. Let T and S be metric trees and suppose that hN (T ) = hN (S) and that the functions contained
in this common set are distinct. Let Fk and Gk denote the subforests for T and S, respectively, as defined
above. Let Vk denote the subset of nodes of Fk as defined above, and let Wk ⊂ Gk be defined similarly.
57
We will construct an isomorphism φ : T → S by inductively defining isomorphisms φk : Fk → Gk on
the metric subforests—an isomorphism of subforests is a bijection whose restriction to each connected
component is an isomorphism of metric trees.
First consider the metric forests F0 and G0 containing only the leaves of their respective trees. For
each leaf v ∈ F0 , there is exactly one point wv ∈ S such that hT (v, ·) = hS (wv , ·), and wv is necessarily
in G0 , as a leaf can be recognized from its local distance distribution (i.e., h0T (x, 0) = 1 if and only if x
is a leaf of the metric tree T ). The isomorphism (in this case, simply a bijection) φ0 : F0 → G0 is the
map taking v to wv .
Next consider F1 and G1 . Let v ∈ N (F1 ) be a leaf parent with child leaves v1 , . . . , vn . Then
hT (vj , r + rj ) − (rj − r) r ≤ rj
hT (v, ·) = (38)
hT (vj , r + rj ) r ≥ rj
for each leaf vj , where rj is the length of its leaf edge. There is a unique vertex wv ∈ S such that
hT (v, ·) = hS (wv , ·). Moreover, it must be the leaf parent of φ0 (v1 ), . . . , φ0 (vn ), since the leaf parent
of each φ0 (vj ) is determined by an explicit equation of the form (38). We extend φ0 to a map φ1 :
N (F1 ) → N (G1 ) by sending each v to wv . The map extends to an isomorphism φ1 : F1 → G1 by
interpolation over edges. To see that this step is valid, consider a leaf vj with parent node v in F1 .
The length of the leaf edge of vj is given by the smallest value of r > 0 such that h0T (vj , r) 6= 1, and
it follows that the edge joining φ1 (v) and φ1 (vj ) must have the same length. We complete this step
by noting that there is also an explicit formula for the function hF1 (v, ·). Assuming without loss of
generality that the leaf edge lengths rj of the leaves vj are ordered increasingly as r1 < r2 < · · · < rn ,
the formula is determined by
n 0 ≤ r < r1
n − 1 r1 ≤ r < r2
0
hF1 (v, r) = .
.. ..
.
1 rn−1 ≤ r < rn
0 rn ≤ r.
As shown above, each vj maps under φk into Wk . It follows that for each such v there is a unique wv ∈
N (Gk+1 ) \ N (Gk ) with the same h-function. We extend φk to the map on nodes φk+1 : N (Fk+1 ) →
N (Gk+1 ) given by sending v to wv . Since a formula similar to (39) can be written for each φk (vj ) ∈ Wk ,
it must be that the elements of Wk which connect to wv are exactly φk (v1 ), . . . , φk (vn ). Finally, note
that the length of the edge connecting vj to v is the smallest r > 0 such that h0T (vj , r) − h0Fk (vj , r) 6= 1
and it follows that we can extend φk+1 to an isomorphism φk+1 : Fk+1 → Gk+1 by interpolating
over edges. Finally, we complete the inductive step in this case by deriving a formula for hFk+1 (v, ·).
58
Assuming without loss of generality that r1 ≤ r2 ≤ · · · ≤ rn , we have
n 0 ≤ r < r1
0
n − 1 + hFk (v1 , r − r1 ) r1 ≤ r < r2
0 .. ..
hFk+1 (v, r) = . .
0 0
1 + hFk (v1 , r − r1 ) + · · · + hFk (v1 , r − rn−1 ) rn−1 ≤ r < rn
h0Fk (v1 , r − r1 ) + · · · + h0Fk (v1 , r − rn ) rn ≤ r.
Proof of Proposition 6.28. Lemma 6.32 implies that the map T 7→ hN T ) is injective on the set of
isomorphism classes of metric trees T such that the functions in hN (T ) are distinct. This set contains,
in particular, the set of metric trees with the property that there are no nontrivial equalities between
edgelengths involving only addition and subtraction. This property is referred to as Property P in [77],
where the authors remark that the set of metric graphs lacking this property has zero measure with
respect to ρ. This proves the statement about generic injectivity.
To prove the statement about density, let T be a metric tree and let > 0. We wish to construct a new
tree T 0 which is -close to T in Gromov-Hausdorff distance and which has the property that hN (T 0 )
contains distinct functions. To do so, enumerate the elements of N (T ) as v1 , . . . , vN . Let r1 , . . . , rN be
a list of distinct real numbers which are also distinct from all edgelengths of T . At each vj we append
a leaf edge of length rj , and the resulting metric tree is T 0 . By rescaling the rj uniformly as needed,
we can ensure that dGH (T , T 0 ) < . It therefore remains to show that the functions in hN (T 0 ) are all
distinct.
Let v ∈ N (T 0 ) and let δ > 0 be less than the shortest edgelength of T 0 . If v is a leaf which was
appended to the vertex vj ∈ T in order to obtain T 0 , then for r ≤ rj + δ,
0 1 r ≤ rj ,
hT 0 (v, r) =
2 rj < r ≤ rj + δ.
By our choice of the rj , this is enough to characterize the function hT 0 (v, ·) uniquely amongst those in
hN (T 0 ). On the other hand, suppose that v = vj (that is, v corresponds to a node in the original tree T ).
Let d ≥ 2 denote the degree of vj , as a node in T 0 . Then for r ≤ rj + δ,
0 d r ≤ rj ,
hT 0 (v, r) =
d − 1 rj < r ≤ rj + δ,
7 Discussion
In this paper we formalized several inverse problems on recovering mm-spaces within a fixed subcat-
egory from invariants defined in terms of distributions of distances: (IP1) on global injectivity of the
invariants, (IP2) on the ability of the invariants to detect homotopy type, (IP3) on the ability of the in-
variants to distinguish spheres, and (IP4) on local injectivity of the invariants. We solved, at least in part,
several of these problems in categories of plane curves, embedded manifolds, Riemannian manifolds,
59
Table 1: Summary of inverse problem results.
Local Distribution Global Distribution
Class of Spaces
(IP1) (IP2) (IP4) (IP1) (IP2) (IP3) (IP4)
Plane Curves 3 n/a 3 7 n/a 3 7
Prop 4.10 – Prop 4.10 Prop 4.2 – Prop 4.7 Exmp 5.8
Emb. Manifolds ◦ ◦ ◦ 7 ◦ 3 ◦
– – – Cor 4.12 – Thm 4 –
Riem. Manifolds 3 3 3 7 3 3 ◦
Prop 5.10 Thm 7 Prop 5.10 Prop 5.16 Thm 7 Thm 6 –
Metric Graphs 3 3 3 7 ◦ 3 7
Thm 9 Thm 10 Thm 9 Exmp 6.13 – Thm 8 Exmp 5.8
Metric Trees 7 n/a 3 7 n/a n/a ◦
Exmp 6.13 – Thm 11 Exmp 6.13 – – –
metric graphs and metric trees. Our results are summarized in Table 1. In each category, for each inverse
problem, we mark 3if positive progress has been made and 7 if we have a counterexample and, in each
case, refer to the relevant result. Of course, these questions are not completely independent—e.g., global
injectivity implies local injectivity—and this is reflected in the Table. We mark with ◦ if the problem is
completely open in the category. The table should be read with the following disclaimer in mind: even
if an entry is marked with 3or 7, this does not mean that the problem has been completely solved in
full generality (far from it, in some cases). Future work will be to fill in the remaining gaps in Table 1
and to extend these results to new categories of mm-spaces. Starting points for obtaining results in these
directions are interspersed throughout Sections 4, 5 and 6 as Questions.
We also introduced Gromov-Monge distance as a variant of the Gromov-Wasserstein distance be-
tween mm-spaces with the stricter requirement that correspondences between mm-spaces are realized
by transport maps. The main open problem for Gromov-Monge distance is the existence of optimal
transport maps realizing the (coupling-based) Gromov-Wasserstein distance, as stated in Question 3.2.
The relatively few results in this direction, as summarized in Remark 3.3, show that there is a lot of
work to do in studying the problem for more general categories of mm-spaces. Based on the work in
this paper, it seems tractable to attack the question in the setting of metric graphs, where one might hope
that the simplicity of one dimensional optimal transport could provide insight on the quadratic problem.
Acknowledgements
We acknowledge funding from these sources: NSF CCF 1526513, NSF DMS 1723003, NSF CCF
1740761, NSF DMS 1547357. We also thank Zane Smith and Peter Olver for useful conversations on
some of the counterexamples presented in the paper.
References
[1] Mridul Aanjaneya, Frederic Chazal, Daniel Chen, Marc Glisse, Leonidas Guibas, and Dmitriy
Morozov. Metric graph reconstruction from noisy data. International Journal of Computational
Geometry & Applications, 22(04):305–325, 2012.
[2] Michal Adamaszek, Henry Adams, Ellen Gasparovic, Maria Gommel, Emilie Purvine, Radmila
Sazdanovic, Bei Wang, Yusu Wang, and Lori Ziegelmeier. On homotopy types of vietoris-rips
complexes of metric gluings, 2017.
60
[3] Pankaj K Agarwal, Kyle Fox, Abhinandan Nath, Anastasios Sidiropoulos, and Yusu Wang. Com-
puting the Gromov-Hausdorff distance for metric trees. In International Symposium on Algorithms
and Computation, pages 529–540. Springer, 2015.
[4] Mahmuda Ahmed, Brittany Terese Fasy, and Carola Wenk. Local persistent homology based dis-
tance between maps. In Proceedings of the 22nd ACM SIGSPATIAL International Conference on
Advances in Geographic Information Systems, pages 43–52, 2014.
[5] Reema Al-Aifari, Ingrid Daubechies, and Yaron Lipman. Continuous procrustes distance between
two surfaces. Communications on Pure and Applied Mathematics, 66(6):934–964, 2013.
[6] Aleksandr D Aleksandrov. Uniqueness theorems for surfaces in the large. Amer. Math. Soc.
Transl.(2), 21:341–354, 1962.
[7] P Alestalo, DA Trotsenko, and J Väisälä. Isometric approximation. Israel Journal of Mathematics,
125(1):61–82, 2001.
[8] David Alvarez-Melis and Tommi Jaakkola. Gromov-Wasserstein alignment of word embedding
spaces. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Pro-
cessing, pages 1881–1890, 2018.
[9] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Gradient flows: in metric spaces and in the
space of probability measures. Springer Science & Business Media, 2008.
[10] Giorgio A Ascoli, Duncan E Donohue, and Maryam Halavi. Neuromorpho.org: a central resource
for neuronal morphologies. Journal of Neuroscience, 27(35):9247–9251, 2007.
[11] Brenden Balch, Chris Peterson, and Clayton Shonkwiler. Distributions of distances and volumes
of balls in homogeneous lens spaces. arXiv preprint arXiv:2004.13196, 2020.
[12] Brenden Balch, Chris Peterson, and Clayton Shonkwiler. Expected distances on manifolds of
partially oriented flags. arXiv preprint arXiv:2001.07854, 2020.
[14] Martin Bauer, Thomas Fidler, and Markus Grasmair. Local uniqueness of the circular integral
invariant. Inverse Problems & Imaging, 7(1):107–122, 2013.
[15] Serge Belongie, Greg Mori, and Jitendra Malik. Matching with shape contexts. In Statistics and
Analysis of Shapes, pages 81–105. Springer, 2006.
[16] Robin Lynne Belton, Brittany Terese Fasy, Rostik Mertz, Samuel Micka, David L Millman, Daniel
Salinas, Anna Schenfisch, Jordan Schupbach, and Lucia Williams. Reconstructing embedded
graphs from persistence diagrams. Computational Geometry, page 101658, 2020.
[17] Marcel Berger. A panoramic view of Riemannian geometry. Springer Science & Business Media,
2012.
[18] José R Berrendero, Antonio Cuevas, and Beatriz Pateiro-López. Shape classification based on
interpoint distance distributions. Journal of Multivariate Analysis, 146:237–247, 2016.
[19] Arthur L Besse. Einstein manifolds. Springer Science & Business Media, 2007.
61
[21] Sergey Bobkov and Michel Ledoux. One-dimensional empirical measures, order statistics and
kantorovich transport distances. preprint, 2014.
[22] Neda Bokan, Mirjana Djorić, and Udo Simon. Geometric structures as determined by the volume
of generalized geodesic balls. Results in Mathematics, 43(3-4):205–234, 2003.
[23] Marco Bonetti and Marcello Pagano. The interpoint distance distribution as a descriptor of point
patterns, with an application to spatial disease clustering. Statistics in medicine, 24(5):753–773,
2005.
[24] Mireille Boutin and Gregor Kemper. On reconstructing n-point configurations from the distribution
of distances or areas. Advances in Applied Mathematics, 32(4):709–735, 2004.
[25] Doug M Boyer, Yaron Lipman, Elizabeth St Clair, Jesus Puente, Biren A Patel, Thomas
Funkhouser, Jukka Jernvall, and Ingrid Daubechies. Algorithms to automatically quantify the
geometric similarity of anatomical surfaces. Proceedings of the National Academy of Sciences,
2011.
[26] Yann Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Com-
munications on pure and applied mathematics, 44(4):375–417, 1991.
[27] Daniel Brinkman and Peter J Olver. Invariant histograms. American Mathematical Monthly,
119(1):4–24, 2012.
[28] Charlotte Bunne, David Alvarez-Melis, Andreas Krause, and Stefanie Jegelka. Learning generative
models across incomparable spaces. In International Conference on Machine Learning, pages
851–861, 2019.
[29] Dmitri Burago, Yuri Dmitrievich Burago, and Sergei Ivanov. A course in metric geometry, vol-
ume 33. American Mathematical Soc., 2001.
[30] G Calvaruso and L Vanhecke. Special ball-homogeneous spaces. Zeitschrift für Analysis und ihre
Anwendungen, 16(4):789–800, 1997.
[31] G Calvaruso and Lieven Vanhecke. Semi-symmetric ball-homogeneous spaces and a volume con-
jecture. Bulletin of the Australian Mathematical Society, 57(1):109–115, 1998.
[32] Mathieu Carrière and Steve Oudot. Local equivalence and intrinsic metrics between reeb graphs.
arXiv preprint arXiv:1703.02901, 2017.
[34] Felice Casorati. Mesure de la courbure des surfaces suivant l’idée commune. Acta mathematica,
14(1):95–110, 1890.
[35] Claire Chalopin, Gérard Finet, and Isabelle E Magnin. Modeling the 3d coronary tree for labeling
purposes. Medical image analysis, 5(4):301–315, 2001.
[36] Arnaud Charnoz, Vincent Agnus, Grégoire Malandain, Luc Soler, and Mohamed Tajine. Tree
matching applied to vascular system. In International Workshop on Graph-Based Representations
in Pattern Recognition, pages 183–192. Springer, 2005.
[37] Shiu-Yuen Cheng. Eigenvalue comparison theorems and its geometric applications. Mathematis-
che Zeitschrift, 143(3):289–297, 1975.
[38] Samir Chowdhury and Tom Needham. Gromov-Wasserstein averaging in a riemannian framework.
arXiv preprint arXiv:1910.04308, 2019.
62
[39] Samir Chowdhury and Tom Needham. Generalized spectral clustering via Gromov-Wasserstein
learning. arXiv preprint arXiv:2006.04163, 2020.
[40] Balázs Csikós, Márton Horváth, et al. A characterization of harmonic spaces. Journal of Differen-
tial Geometry, 90(3):383–389, 2012.
[41] Justin Curry, Sayan Mukherjee, and Katharine Turner. How many directions determine a shape and
other sufficiency results for two topological transforms. arXiv preprint arXiv:1805.09782, 2018.
[42] Manfredo P Do Carmo. Differential Geometry of Curves and Surfaces: Revised and Updated
Second Edition. Courier Dover Publications, 2016.
[43] Danielle Ezuz, Justin Solomon, Vladimir G Kim, and Mirela Ben-Chen. GWCNN: A metric
alignment layer for deep shape analysis. In Computer Graphics Forum, volume 36, pages 49–57.
Wiley Online Library, 2017.
[44] Brittany Terese Fasy, Samuel Micka, David L Millman, Anna Schenfisch, and Lucia
Williams. Persistence diagrams for efficient simplicial complex reconstruction. arXiv preprint
arXiv:1912.12759, 2019.
[45] Kei Funano and Takashi Shioya. Concentration, ricci curvature, and eigenvalues of laplacian.
Geometric and Functional Analysis, 23(3):888–936, 2013.
[46] Ricardo Garcıa-Pelayo. Pairs of subsets of regular polyhedra with the same distribution of distance.
Applied Mathematical Sciences, 10(26):1285–1297, 2016.
[47] Natasha Gelfand, Niloy J Mitra, Leonidas J Guibas, and Helmut Pottmann. Robust global registra-
tion. In Symposium on geometry processing, volume 2, page 5, 2005.
[48] A Gray, Lieven Vanhecke, et al. Riemannian geometry as determined by the volumes of small
geodesic balls. Acta Mathematica, 142:157–198, 1979.
[49] Andreas Greven, Peter Pfaffelhuber, and Anita Winter. Convergence in distribution of random
metric measure spaces (λ-coalescent measure trees). Probability Theory and Related Fields, 145(1-
2):285–322, 2009.
[50] Mikhael Gromov. Hyperbolic groups. In Essays in group theory, pages 75–263. Springer, 1987.
[51] Mikhael Gromov and Vitali D Milman. A topological application of the isoperimetric inequality.
American Journal of Mathematics, 105(4):843–854, 1983.
[52] Charles Gueunet, Pierre Fortin, Julien Jomier, and Julien Tierny. Task-based augmented merge
trees with fibonacci heaps. In Large Data Analysis and Visualization (LDAV), 2017 IEEE 7th
Symposium on, pages 6–15. IEEE, 2017.
[53] Steven Haker, Lei Zhu, Allen Tannenbaum, and Sigurd Angenent. Optimal mass transport for
registration and warping. International Journal of computer vision, 60(3):225–240, 2004.
[55] Juha Heinonen. Lectures on analysis on metric spaces. Springer Science & Business Media, 2012.
[56] Reigo Hendrikson et al. Using Gromov-Wasserstein distance to explore sets of networks. Univer-
sity of Tartu, Master Thesis, 2016.
[57] Piotr Indyk, Jiřı́ Matoušek, and Anastasios Sidiropoulos. 8: Low-distortion embeddings of finite
metric spaces. In Handbook of discrete and computational geometry, pages 211–231. Chapman
and Hall/CRC, 2017.
63
[58] Leonid V Kantorovich. On the translocation of masses. In Dokl. Akad. Nauk. USSR (NS), vol-
ume 37, pages 199–201, 1942.
[59] Leon Karp and Mark Pinsky. Volume of a small extrinsic ball in a submanifold. Bulletin of the
London Mathematical Society, 21(1):87–92, 1989.
[60] Paul J Kelly et al. A congruence theorem for trees. Pacific Journal of Mathematics, 7(1):961–968,
1957.
[61] Simon L Kokkendorff. Characterizing the round sphere by mean distance. Differential Geometry
and its Applications, 26(6):638–644, 2008.
[62] Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepcev, and Gustavo K Rohde. Optimal
mass transport: Signal processing and machine-learning applications. IEEE signal processing
magazine, 34(4):43–59, 2017.
[63] Jee-Hyun Kong, Daniel R Fish, Rebecca L Rockhill, and Richard H Masland. Diversity of ganglion
cells in the mouse retina: unsupervised morphological classification and its limits. Journal of
Comparative Neurology, 489(3):293–310, 2005.
[64] W. Kühnel and B. Hunt. Differential Geometry: Curves - Surfaces - Manifolds. Student mathe-
matical library. American Mathematical Society, 2006.
[65] Paul Lemke, Steven S Skiena, and Warren D Smith. Reconstructing sets from interpoint distances.
In Discrete and computational geometry, pages 597–631. Springer, 2003.
[66] László Lovász. Large networks and graph limits, volume 60. American Mathematical Soc., 2012.
[67] CL Mallows and JMC Clark. Linear-intercept distributions do not characterize plane sets. Journal
of Applied Probability, 7(1):240–244, 1970.
[68] Jeremy L Martin, Matthew Morin, and Jennifer D Wagner. On distinguishing trees by their chro-
matic symmetric functions. Journal of Combinatorial Theory, Series A, 115(2):237–253, 2008.
[69] Robert J McCann. Polar factorization of maps on Riemannian manifolds. Geometric & Functional
Analysis GAFA, 11(3):589–608, 2001.
[70] Facundo Mémoli. On the use of Gromov-Hausdorff distances for shape comparison. In Proceed-
ings of Point Based Graphics, 2007.
[71] Facundo Mémoli. Gromov-Hausdorff distances in euclidean spaces. In Computer Vision and
Pattern Recognition Workshops, 2008. CVPRW’08. IEEE Computer Society Conference on, pages
1–8. IEEE, 2008.
[72] Facundo Mémoli. Gromov–Wasserstein distances and the metric approach to object matching.
Foundations of computational mathematics, 11(4):417–487, 2011.
[73] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l’Académie Royale
des Sciences de Paris, 1781.
[74] Dmitriy Morozov, Kenes Beketayev, and Gunther Weber. Interleaving distance between merge
trees. Discrete and Computational Geometry, 49(22-45):52, 2013.
[76] Robert Osada, Thomas Funkhouser, Bernard Chazelle, and David Dobkin. Shape distributions.
ACM Transactions on Graphics (TOG), 21(4):807–832, 2002.
64
[77] Steve Oudot and Elchanan Solomon. Barcode embeddings, persistence distortion, and inverse
problems for metric graphs. arXiv preprint arXiv:1712.03630, 2017.
[78] Steve Oudot and Elchanan Solomon. Inverse problems in topological persistence. arXiv preprint
arXiv:1810.10813, 2018.
[79] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport. Foundations and Trends® in
Machine Learning, 11(5-6):355–607, 2019.
[80] Gabriel Peyré, Marco Cuturi, and Justin Solomon. Gromov-Wasserstein averaging of kernel and
distance matrices. In International Conference on Machine Learning, pages 2664–2672, 2016.
[81] Gabriel Peyré, Marco Cuturi, and Justin Solomon. Gromov-wasserstein averaging of kernel and
distance matrices. In International Conference on Machine Learning, pages 2664–2672, 2016.
[82] Halsey Lawrence Royden. Real analysis. Macmillan New York, 1988.
[83] Luis Antonio Santaló. Introduction to integral geometry, volume 1198. Hermann, 1953.
[84] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkäuser, NY, 55(58-63):94,
2015.
[85] Ryoma Sato, Marco Cuturi, Makoto Yamada, and Hisashi Kashima. Fast and robust comparison
of probability measures in heterogeneous spaces. arXiv preprint arXiv:2002.01615, 2020.
[86] Yonggang Shi, Paul M Thompson, Greig I de Zubicaray, Stephen E Rose, Zhuowen Tu, Ivo Di-
nov, and Arthur W Toga. Direct mapping of hippocampal surfaces with intrinsic shape context.
NeuroImage, 37(3):792–807, 2007.
[87] Karl-Theodor Sturm. On the geometry of metric measure spaces. Acta mathematica, 196(1):65–
131, 2006.
[88] Karl-Theodor Sturm. The space of spaces: curvature bounds and gradient flows on the space of
metric measure spaces. arXiv preprint arXiv:1208.0434, 2012.
[89] Anton Tsitsulin, Davide Mottin, Panagiotis Karras, Alexander Bronstein, and Emmanuel Müller.
Netlsd: hearing the shape of a graph. In Proceedings of the 24th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining, pages 2347–2356. ACM, 2018.
[92] Titouan Vayer, Rémi Flamary, Romain Tavenard, Laetitia Chapel, and Nicolas Courty. Sliced
Gromov-Wasserstein. In Advances in Neural Information Processing Systems (NeurIPS), vol-
ume 32, 2019.
[93] Cédric Villani. Topics in optimal transportation. Number 58. American Mathematical Soc., 2003.
[94] Cédric Villani. Optimal transport: old and new, volume 338. Springer Science & Business Media,
2008.
[95] Scott Wolpert. The eigenvalue spectrum as moduli for flat tori. Transactions of the American
Mathematical Society, 244:313–321, 1978.
[96] Hongteng Xu, Dixin Luo, and Lawrence Carin. Scalable Gromov-Wasserstein learning for graph
partitioning and matching. arXiv preprint arXiv:1905.07645, 2019.
65
[97] Hongteng Xu, Dixin Luo, Ricardo Henao, Svati Shah, and Lawrence Carin. Learning autoencoders
with relational regularization. arXiv preprint arXiv:2002.02913, 2020.
[98] Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-Wasserstein learn-
ing for graph matching and node embedding. In International Conference on Machine Learning,
pages 6932–6941, 2019.
demb
GW,p (X , Y) := inf dZ
W,p ((φX )# µX , (φY )# µY ),
Z,φX ,φY
where the infimum is taken over all isometric embeddings φX : X → Z and φY : Y → Z into some
compact metric space Z and dZ W,p is the usual Wasserstein p-distance on P(Z).
On the other hand, the original Monge formulation of optimal transport between α, β ∈ P(Z) seeks
the quantity
Z 1/p
Z
p
dM,p (α, β) := inf dZ (z, φ(z)) α(dz) ,
φ∈T (α,β) Z
demb
GM,p (X , Y) := inf dZ
M,p ((φX )# µX , (φY )# µY ),
Z,φX ,φY
where the infimum is once again taken over isometric embeddings into a common ambient metric space.
These definitions extend to the case p = ∞ as usual.
66
φY : Y → Z into a metric space Z such that dZ M,p ((φX )# µX , (φY )# µY ) < r. We may as well assume
that X, Y ⊂ Z and that µX and µY are measures on Z with supp[µX ] = X and supp[µY ] = Y . By
definition of dZ
M,p , there exists φ ∈ T (µX , µY ) such that
where dZ (·, φ(·)) denotes the map from Z to R given by z 7→ dZ (z, φ(z)). Now note that for all
x, x0 ∈ X, the triangle inequality in Z implies that
|dZ (x, x0 ) − dZ (φ(x), φ(x0 ))| ≤ dZ (x, φ(x)) + dZ (x0 , φ(x0 )).
Putting this together with the triangle inequality for the Lp norm, we have
kΓX ,Y kLp (µφ ⊗µφ ) ≤ 2kdZ (·, φ(·))kLp (µX ) < 2r.
The claim follows if we are able to construct a metric space (Z, dZ ) and isometric embeddings φX and
φY such that
dZ
M,∞ ((φX )# µX , (φY )# µY ) ≤ r.
Let Z denote the disjoint union of X and Y and define a metric dZ on Z by setting dZ |X×X = dX ,
dZ |Y ×Y = dY and
dZ
M,∞ (µX , µY ) = inf sup dZ (x, φ(x)) ≤ sup dZ (x, φ0 (x))
φ∈T (µX ,µY ) x∈X x∈X
dX (x, x ) + r + dY (φ0 (x0 ), φ0 (x)) = r.
0
= sup inf
0
x∈X x ∈X
Example A.1 (dGM,p and demb GM,p are Not BiLipschitz Equivalent). Consider the family of mm-spaces
∆n . Each ∆n consists of the space Xn = {1, . . . , n} with metric dn (i, j) = 1 − δij and measure νn
defined by νn (i) = 1/n. For p < ∞,
1
demb emb
GM,p (∆2n , ∆n ) ≥ dGW,p (∆2n , ∆n ) ≥ ,
4
where the former bound holds generally and the latter is shown in [72, Claim 5.3]. On the other hand,
for p < ∞, dGM,p (∆2n , ∆n ) = 1/(2n)1/p . This follows from the fact that T (ν2n , νn ) is simply the set
of 2-to-1 maps from {1, . . . , 2n} to {1, . . . , n} and all such maps have the same Monge cost, so that the
quantity is obtained by a direct calculation. It follows that the quasi-metrics dGM,p and demb GM,p are not
bi-Lipschitz equivalent for any p ∈ [1, ∞).
67
A.2 Gromov-Wasserstein Distance as a Gromov-Monge Distance
A pseudo-metric space is a pair (X, dX ) such that the pseudo-metric dX : X × X → R satisfies all of
the axioms of a metric, except that it is possible for dX (x, x0 ) = 0 for distinct x, x0 ∈ X. The pseudo-
metric defines a topology on X by taking a basis of open balls; the topology is non-Hausdorff if dX is
not actually a metric. A pseudo-metric measure space (pmm-space) is a triple X = (X, dX , µX ), where
(X, dX ) is a compact pseudo-metric space and µX is a Borel probability measure with full support. The
definition of Gromov-Monge distance extends to pseudo-metric measure spaces without modification.
Example A.2 (Pullback Pseudometric). Let (X, dX ) be a metric space, let Z be a set and let f : Z → X
be a function. The pullback of dX by f is the map f ∗ dX : Z × Z → R defined by
X Y
Proof of Theorem 14. Given a measure coupling µ of µX and µY , define a pmm-space Z by setting
Z = X × Y , µZ = µ, π = πX (projection X × Y → X onto the first coordinate) and dZ = π ∗ dX .
Then φ = πY : Z → Y is a measure-preserving map with
Z
kΓZ,Y kpLp (µφ ⊗µφ ) = ΓZ,Y (z, φ(z), z 0 , φ(z 0 ))p µZ ⊗ µZ (dz × dz 0 )
ZZ×Z
= ΓZ,Y ((x, y), y, (x0 , y 0 ), y 0 )p µZ ⊗ µZ (dx × dy × dx0 × dy 0 )
X×Y ×X×Y
Z
= ΓX ,Y (x, y, x0 , y 0 )p µ ⊗ µ(dx × dy × dx0 × dy 0 ),
X×Y ×X×Y
∗ d . We conclude that
where the last line follows by dZ = πX X
µ = (π × φ)# µZ ,
68
where π × φ : Z → X × Y is the map taking z to (π(z), φ(z)). Then µ defines a measure coupling of
µX and µY . Indeed, for any set A ⊂ X,
−1
(A) = (π × φ)# µZ (A × Y ) = µZ (π × φ)−1 (A × Y )
(πX )# µ(A) = µ πX
= µZ π −1 (A) = π# µZ (A) = µX (A)
and (πY )# µ = µY follows by a similar argument. Consider the following calculation (we suppress all
function arguments to condense notation):
Z Z
p
(ΓX ,Y ) µ ⊗ µ = (ΓX ,Y )p (π × φ)# µZ ⊗ (π × φ)# µZ
(X×Y ) 2 (X×Y )2
Z
= (ΓX ,Y )p ((π × φ) × (π × φ))# µZ ⊗ µZ (42)
(X×Y )2
Z
= (ΓX ,Y ◦ ((π × φ) × (π × φ)))p µZ ⊗ µZ (43)
Z×Z
Z
= (ΓZ,Y )p µφ ⊗ µφ . (44)
Z×Z
Equality (42) follows from the fact that (π×φ)# µZ ⊗(π×φ)# µZ and ((π×φ)×(π×φ))# µZ ⊗µZ define
equivalent measures on X × Y × X × Y . Equality (43) follows from the change-of-variables property of
pushforward measures. The final equality (44) follows once again from the change-of-variables property
and the fact that
ΓX ,Y (π(z), φ(z), π(z 0 ), φ(z 0 )) = ΓZ,Y (z, φ(z), z 0 , φ(z 0 )),
since dZ = π ∗ dX . This calculation shows that
where we use E(n) to denote the group of Euclidean isometries of Rn and k · k for the Euclidean norm.
Theorem 15. Let X and Y be Euclidean mm-spaces in the same ambient space Rn , let p ≥ 1 and let
M = max{diam(X), diam(Y )}. Then
Rn
1/2
demb
GM,p (X , Y) ≤ dM,p,iso (X , Y) ≤ M
1/2
· cn · demb
GM,p (X , Y) ,
demb
GM,p (X , Y) ≤ max{diam(X), diam(Y )}.
The theorem is analogous to [71, Theorem 4], which compares Gromov-Wasserstein distance and
isometry-invariant Wasserstein distance for Euclidean spaces. The proof of the theorem is essentially
the same, with the main difference being the use of the following technical lemma, which is specific to
the Gromov-Monge formulation.
69
Lemma A.5. Let X and Y be mm-spaces and let M = max{diam(X), diam(Y )}. If
demb
GM,p (X , Y) ≤ · M
and
dGH (X , Y ) ≤ sup ΓX,Y (x, y, x0 , y 0 ) ≤ 1/2 M. (46)
(x,y),(x0 ,y 0 )∈R
Proof. Without loss of generality, suppose that X and Y are subsets of an ambient metric space (Z, dZ )
and let φ ∈ S such that Z
dZ (z, φ(z))p µX (dz) ≤ p M p .
Z
Define a subset R ⊂ X × Y by
We then define X = πX (R ) and Y = πY (R ), so that R ∈ R(X , Y ). Also note that φ(X ) = Y .
A short calculation using the triangle inequality shows that R satisfies (46). Moreover,
Z Z
p p p
M ≥ dZ (z, φ(z)) µX (dz) ≥ dZ (z, φ(z))p µX (dz) ≥ p/2 M p µX (X \ X ).
Z X\X
Proof of Theorem 15. The inequality on the left is obvious, so let us consider the inequality on the right.
If either quantity is infinite, then both are, so assume not. By Remark A.4, we suppose without loss of
generality that demb
GM,p (X , Y) = M for some ≤ 1. Let X , Y , φ and R be as in Lemma A.5 and let
R = (X × Y ) \ R . Then dGH (X , Y ) ≤ 1/2 M and [29, Corollary 7.3.28] implies that there exists a
c
and such that ψ(X ) is a 21/2 M -net for Y . We apply [7, Theorem 2.2] to conclude that there is an
isometry T ∈ E(n) such that supx∈X kT (x) − ψ(x)k ≤ 1/2 · an · M , where an is a constant depending
only on n.
Applying the triangle inequality and the general inequality (a + b)p ≤ 2p−1 (ap + bp ) (for a, b ≥ 0),
we have
Z
Rn p
dM,p,iso (X , Y) ≤ kT (x) − ykp µφ (dx × dy) (47)
c
R
Z
p−1
+2 kT (x) − ψ(x)kp µφ (dx × dy) (48)
R
Z
+ 2p−1 kψ(x) − ykp µφ (dx × dy). (49)
R
70
We bound each term separately. First note that
p p−1 p p p p p
kT (x) − yk ≤ 2 (kT (x)k + kyk ) ≤ 2 max max kT (x)k , max kyk = (2M )p ,
x y
n
where we have used isometry invariance of dRM,p,iso to assume without loss of generality that the circum-
centers of X and Y are at the origin in order to obtain the last equality. This implies that
(47) ≤ (2M )p µφ (Rc ) ≤ 2p · M p · p/2 .
The bounds on (48) and (49) follow from our assumptions on ψ:
(48) ≤ 2p−1 · max kT (x) − ψ(x)kp · µφ (R ) ≤ 2p−1 · p/2 · apn · M p
x∈X
and
(49) ≤ 2p−1 · max kψ(x) − ykp µφ (R ) ≤ 2p−1 · 2p · p/2 · M p .
(x,y)∈R
Combining these estimates, we conclude
n
dRM,p,iso (X , Y)p ≤ 2p M p p/2 (1 + 2p−1 + apn ) ≤ 4p M p p/2 max{2, an }p ,
so that
n
dRM,p,iso (X , Y) ≤ M · cn · 1/2 = M 1/2 · cn · dGM,p (X , Y)1/2 ,
with cn = 4 max{2, an }.
where the infimum is taken over metric spaces (Z, dZ ) and isometric embeddings φX and φY with the
property that φX (X) and φY (Y ) are still objects in ObjC , and where
Z 1/p
Z,C p
dM,p (α, β) = inf (dZ (z, φ(z))) α(dz) ,
φ∈S Z
with S = S(supp[α], supp[β]) defined for measures supported on subsets of Z which are objects of C.
n
Finally, we extend the definition of dRM,p,iso analogously to obtain the isometry-invariant quasi-metric
n
dRM,p,iso
,C
; that is, this quasi-metric measures distance between Euclidean isometry orbits of subsets of Rn
which are also objects of C.
Going through the proofs, one observes that it is possible to extend most of our results to the re-
stricted quasi-metrics. For example, Theorem 13 can be adapted to give the general bound
emb,C
dC
GM,p (X , Y) ≤ dGM,p (X , Y).
Likewise, rewriting the arguments and replacing T with S, Theorem 15 can be extended to show that
Euclidean mm-spaces satisfy
1/2
Rn ,C
demb,C
GM,p (X , Y) ≤ dM,p,iso (X , Y) ≤ M 1/2
· cn · demb,C
GM,p (X , Y) (50)
for Euclidean mm-spaces X and Y belonging to ObjC .
These observations justify the claims made in Examples 3.11 and 3.12 in the main text. To be
3 ,C
precise, the continuous Procrustes metric described in Example 3.11 is given in our notation by dRM,2,iso ,
where C is the category whose objects are embedded smooth surfaces whose morphisms are continuous
measure-preserving maps. The bound (50) shows that demb,C GM,p can be seen as a generalization of the
2
continuous Procrustes metric. Similarly, the metric described in Example 3.12 is dRM,2,iso
,C
, where C is the
category whose objects are planar regions and whose morphisms are restrictions of measure-preserving
diffeomorphisms of R2 . Once again, (50) shows that this is generalized by a variant of Gromov-Monge
distance.
71