Average Minimum Distance (AMD) Overview

• AMD is a multifaceted concept defined as the average of shortest or nearest distances under various optimization and measurement settings, spanning combinatorial, geometric, and statistical domains.
• In computational geometry, AMD in triangulations minimizes the sum of shortest-path distances, with results showing NP-completeness under weighted settings and optimality of fan triangulations in one-point-visible regimes.
• AMD also appears in graph theory, Euclidean variational problems, coding theory, and phylogenetics, linking local decision-making to global distance minimization across diverse applied scenarios.

Average Minimum Distance (AMD) is a family of optimization and measurement notions built from averages of shortest or nearest distances. In the cited literature, the term does not denote a single invariant: in planar triangulations it is the average shortest-path distance induced by a triangulation; in spanning-tree problems it is the minimum average distance over all spanning trees of a graph; in Euclidean variational problems it is an integral of point-to-set distances under a length budget; and in several applied settings it appears as the average distance to the nearest selected leaf, the average of subsetwise nearest distances, or a directed nearest-neighbor score between embedding sets (Kozma, 2011, Du et al., 6 May 2026, O'Brien et al., 29 Mar 2025, Matsen et al., 2012, Garg et al., 2014, Goworek et al., 17 Feb 2026). Despite this terminological heterogeneity, the recurring theme is optimization of a global distance functional under combinatorial, geometric, or statistical constraints.

1. Terminological scope and canonical definitions

A useful first distinction is between pairwise formulations, which average shortest-path distances between pairs of points or vertices, and point-to-set formulations, which average the distance from a point sampled from a measure to its nearest admissible representative set. A third variant, prominent in lexical semantic change detection, averages nearest-neighbor distances between two finite point clouds and may be directional.

Setting	AMD quantity	Admissible object
Triangulations	$$\operatorname{AMD}(T)=\frac{2}{n(n-1)}\sum_{1\le i<j\le n}\operatorname{dist}_T(p_i,p_j)$$	triangulation $T$ of a point set or polygon
Spanning trees	$$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$	spanning tree $T$
Euclidean average-distance problem	$$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$	compact connected $\Sigma$ with $\mathcal{H}^1(\Sigma)\le l$
Phylogenetic trees	$$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$	subset of leaves $S$
LSCD	$$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$	cross-period embedding correspondence

In the triangulation formulation, the paper optimizes

$T$0

which is equivalent to minimizing $T$1 because the prefactor $T$2 is independent of $T$3 (Kozma, 2011). In connected graphs, the average distance is

$T$4

and AMD may then mean the minimum of $T$5 over spanning trees $T$6 of $T$7 (Du et al., 6 May 2026). In continuum optimization, the average-distance problem minimizes

$T$8

over compact $T$9 with a length constraint, or, in the probability-measure formulation,

$$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$0

over compact connected $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$1 with $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$2 (Basok et al., 2022, O'Brien et al., 29 Mar 2025).

This multiplicity of meanings is substantive rather than merely notational. Some formulations are discrete and combinatorial, some are geometric and variational, and others are algorithmic statistics on finite samples. The technical content of AMD therefore depends crucially on the ambient space, the admissible class, and the averaging measure.

2. Triangulations in computational geometry

The paper "Minimum Average Distance Triangulations" formalizes AMD for a planar point set $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$3 by assigning a symmetric positive weight function $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$4 and defining $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$5 as the shortest-path length between $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$6 and $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$7 along edges of a triangulation $$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$8. The objective is

$$\operatorname{AMD}(G)=\min\{\mu(T):T\text{ spanning tree of }G\}$$9

with an analogous polygon variant in which one triangulates the interior of a simple polygon. For arbitrary positive semimetric weights, the decision problem is in NP and is shown to be strongly NP-complete via a reduction from Planar3SAT using wire, variable, clause, and bridge gadgets with weights from $T$0 (Kozma, 2011).

The hardness proof depends on a gap construction. Pure versus impure triangulations are separated by a structural lemma; satisfying and non-satisfying assignments produce a provable gap in $T$1; and the parameters are chosen so that the threshold

$T$2

lies strictly between all satisfying and all non-satisfying triangulations. The paper states this formally for point sets, while noting that the approach extends in commentary to polygon triangulations as well (Kozma, 2011).

For equal weights, where link distance replaces weighted shortest paths, the situation changes sharply. If a point set is one-point-visible or a polygon is one-vertex-visible, every fan triangulation is optimal. The key observation is that all triangulations have the same number of edges, so a fan simultaneously maximizes the number of distance-$T$3 pairs and forces all remaining pairs to have link distance $T$4. The paper’s Theorem 1 states that every fan triangulation attains the minimum average link distance in these visibility regimes (Kozma, 2011).

For general simple polygons in the equal-weight case, AMD remains nonlocal. The paper introduces special indices $T$5 relative to a triangle $T$6 and an extended objective

$T$7

which carries cross-subproblem information. A decomposition lemma expresses every cross-side distance through local distances plus a constant offset, and the resulting dynamic program runs in $T$8 time. The Euclidean and, more generally, metric-weight versions remain open, as does the unit-weight point-set case without a one-point-visible vertex (Kozma, 2011).

3. Graph-theoretic shortest-path formulations

In graph theory, AMD often refers to minimization of average pairwise shortest-path distance under graph-theoretic constraints. One formulation fixes a connected graph $T$9 and asks for a spanning tree $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$0 minimizing

$$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$1

The problem is known to be NP-hard. Mukwembi proved that if $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$2 has order $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$3 and independence number $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$4, then for $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$5 there exists a spanning tree with $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$6; the 2026 improvement shows that for $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$7 one can always find a spanning tree with $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$8, and for $$\mathcal{J}_p(\Sigma)=\int \operatorname{dist}^p(x,\Sigma)\,d\mu(x)$$9 the sharper bound

$\Sigma$0

holds. The proof proceeds by showing the existence of a spanning tree in a class $\Sigma$1 consisting of a small core tree with many pendent vertices, where $\Sigma$2 (Du et al., 6 May 2026).

A second formulation fixes the order $\Sigma$3 and size $\Sigma$4 of a connected simple graph and minimizes the average graph distance

$\Sigma$5

The exact minimum is

$\Sigma$6

and it is attained precisely when every pair of non-adjacent vertices is at distance $\Sigma$7, equivalently when the graph has diameter at most $\Sigma$8. The same paper shows that the graph with the largest average clustering is usually unique and simultaneously attains this minimum average distance. The associated extremal architectures are hub-like: complete or almost-complete modules sharing a single cut vertex, with universal vertices providing the required two-hop connectivity (Barmpoutis et al., 2010).

A continuum analogue appears on compact metric graphs, where the mean distance is defined by

$\Sigma$9

This is a continuum average over all points of the graph rather than a vertex average. Among all metric graphs of fixed total length $\mathcal{H}^1(\Sigma)\le l$0, the path maximizes AMD with $\mathcal{H}^1(\Sigma)\le l$1, and among doubly connected graphs the loop is extremal with $\mathcal{H}^1(\Sigma)\le l$2. For fixed length $\mathcal{H}^1(\Sigma)\le l$3 and number of edges $\mathcal{H}^1(\Sigma)\le l$4, the equilateral flower graph minimizes AMD: $\mathcal{H}^1(\Sigma)\le l$5 The same paper establishes the scale-invariant spectral relation

$\mathcal{H}^1(\Sigma)\le l$6

and asks whether $\mathcal{H}^1(\Sigma)\le l$7 admits absolute upper and lower bounds on all compact metric graphs (Baptista et al., 2023).

4. Hamming-space, coding-theoretic, and Fourier formulations

For binary codes, AMD is the minimum average pairwise Hamming distance. If $\mathcal{H}^1(\Sigma)\le l$8 has size $\mathcal{H}^1(\Sigma)\le l$9, then

$$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$0

This problem admits a linear-programming treatment through the Hamming association scheme, Krawtchouk polynomials, and Delsarte positivity constraints on the distance distribution. In particular, the average distance is controlled by the first dual coefficient $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$1 via

$$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$2

so LP upper bounds on $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$3 become AMD lower bounds (0706.3295).

The Boolean-function formulation makes the Fourier connection explicit. For a set $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$4 of density $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$5 with indicator $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$6, the average Hamming distance

$$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$7

satisfies the exact identity

$$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$8

where $$\operatorname{ADCL}(S)=\frac{1}{\mu(T)}\int_T\min_{s\in S}d_T(x,s)\,d\mu(x)$$9 is the level-$S$0 Fourier weight. Maximizing level-$S$1 weight is therefore equivalent to minimizing average distance. This permits direct translation of Chang-type bounds into AMD bounds. The 2025 paper improves classical estimates by proving piecewise upper envelopes $S$2 and $S$3 for $S$4, yielding lower bounds on $S$5; it also shows that Hamming balls maximize the dimension of the span of large Fourier coefficients in $S$6 (Yu, 3 Apr 2025).

A related LP improvement for binary codes with density $S$7 proves

$S$8

where

$S$9

This improves the earlier Fu–Wei–Yeung bound for $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$0 and matches it for $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$1. The same paper records that subcubes attain equality at $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$2 and $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$3, and translates the average-distance bound into the Fourier inequality

$$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$4

The 2007 LP paper proves additional closed-form lower bounds for sparse regimes and, notably, establishes the asymptotically exact result

$$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$5

for binary codes of size $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$6 (Yu et al., 2019, 0706.3295).

These code-theoretic formulations are mathematically close to discrete isoperimetry and harmonic analysis on the hypercube. They also show that AMD can be studied either directly as a distance problem or indirectly through transformed quantities such as Krawtchouk spectra and low-degree Fourier weights.

5. Variational average-distance minimizers in Euclidean space

In Euclidean variational problems, AMD is a length-constrained optimization over connected one-dimensional sets. One formulation fixes a bounded open set $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$7, a length budget $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$8, and a non-decreasing function $$\operatorname{AMD}(A\to B)=\frac{1}{|A|}\sum_{a\in A}\min_{b\in B}\delta(a,b)$$9, and minimizes

$T$00

over compact $T$01 with $T$02. The paper "Inverse maximal and average distance minimizer problems" relates this to maximal distance minimizers and generalizes Tilli’s planar results to $T$03. If a rectifiable curve $T$04 satisfies the unique nearest-point property in $T$05, then $T$06 solves the AMD problem on $T$07 for any non-decreasing $T$08. Sufficient conditions include curvature radius at least $T$09 and length at most $T$10, and every simple $T$11 curve is shown to be an AMD minimizer for sufficiently small tubular radius (Basok et al., 2022).

A second, more recent formulation fixes a compactly supported probability measure $T$12 on $T$13, a power $T$14, and a budget $T$15, and minimizes

$T$16

over

$T$17

A central tool is the barycentre field

$T$18

together with the first-variation formula for continuous deformations of $T$19. Minimizers satisfy vanishing net barycentre field, the ambiguous locus has $T$20-measure zero, and under $T$21 every minimizer lies in $T$22 (O'Brien et al., 29 Mar 2025).

The main structural theorem holds for $T$23 and for

$T$24

Under these hypotheses, optimal sets are finite trees: they contain no cycles, have finitely many noncut points and branching points, and every branching point has order $T$25. The paper does not assert explicit angle conditions, but it provides the first complete topological description in arbitrary dimension for this parameter range (O'Brien et al., 29 Mar 2025).

Taken together, these papers place AMD within the calculus of variations and geometric measure theory. The admissible set is no longer a graph chosen from a finite combinatorial class, but a connected compact subset of $T$26 constrained only by one-dimensional Hausdorff measure. Existence, first variation, regularity, and topology then replace combinatorial enumeration.

6. Applied AMD: phylogenetics, density estimation, and lexical semantic change

In phylogenetics, AMD appears as the Average Distance to the Closest Leaf (ADCL). Given a phylogenetic tree $T$27, a Borel mass measure $T$28 on the tree, and a budget $T$29, one chooses a leaf subset $T$30 with $T$31 to minimize

$T$32

or its discrete weighted analogue. This is the tree-metric $T$33-median problem with centers restricted to leaves. The paper shows that greedy pruning is not globally optimal, that a PAM-style heuristic can get trapped in local minima, and that an exact dynamic program is possible by decomposing the tree into bubbles and maintaining RMD and RMP state families encoding how mass crosses subtree roots. The exact method returns globally optimal solutions for all $T$34 in a single run (Matsen et al., 2012).

In nonparametric density estimation, AMD is a statistic rather than an optimization target. A sample $T$35 is partitioned into $T$36 disjoint subsets of size $T$37, and for a query point $T$38 one defines

$T$39

The estimator is

$T$40

Under the one-dimensional assumptions in the paper,

$T$41

The authors recommend $T$42 as a robust default when $T$43, with $T$44 preferable in smoother regimes where the bias behaves like $T$45 (Garg et al., 2014).

In lexical semantic change detection, AMD is a nearest-neighbor correspondence metric between contextual embedding clouds from two periods. If $T$46 and $T$47 are usage embeddings for a target word and $T$48 is cosine distance, the directed score is

$T$49

and the paper’s symmetric AMD is the average of the two directions. A stricter variant, SAMD, downsamples to equal cardinality and greedily constructs a one-to-one matching over the pairwise distance matrix. Empirically, AMD is reported as especially robust with non-specialized encoders and under dimensionality reduction, while SAMD is strongest with specialized encoders such as XL-LEXEME and under PCA (Goworek et al., 17 Feb 2026).

These applications illustrate three distinct computational interpretations of AMD: facility location on a tree metric, a first-order-statistic estimator built from minimum sample distances, and a local correspondence score for diachronic embedding geometry.

7. Structural themes and open problems

Across these literatures, AMD objectives are consistently global and nonlocal. In triangulations, local edge choices alter shortest paths between far-separated vertices; this is precisely why the polygon algorithm requires the extended objective $T$50 and special-index bookkeeping (Kozma, 2011). In Euclidean average-distance minimization, the barycentre field and the nullity of the ambiguous locus play the analogous role of encoding global transport geometry into local first-variation data (O'Brien et al., 29 Mar 2025). In Hamming-space formulations, AMD is controlled indirectly through global spectral data such as Krawtchouk transforms and level-$T$51 Fourier weight (Yu, 3 Apr 2025).

A second recurring theme is the emergence of shortcut-rich extremal structures. Fan triangulations collapse all non-edge pairs to distance $T$52 in one-point-visible settings (Kozma, 2011). Graphs of fixed order and size minimize average distance exactly when every non-adjacent pair is at distance $T$53 (Barmpoutis et al., 2010). On metric graphs, equilateral flowers minimize continuum AMD at fixed total length and edge count (Baptista et al., 2023). This suggests that hub-mediated or highly redundant routing structures repeatedly arise when the admissible class permits them.

Several open problems remain central. For minimum average distance triangulations, the metric and Euclidean weight cases are open, as is the unit-weight point-set case without a one-point-visible vertex; for the strongly NP-hard semimetric variant, an FPTAS is ruled out unless $T$54, while the existence of a PTAS is open (Kozma, 2011). In continuum optimization, the barycentre-field proof leaves the range

$T$55

open for the general-dimensional topological characterization (O'Brien et al., 29 Mar 2025). On compact metric graphs, the existence of absolute constants $T$56 such that

$T$57

for all compact metric graphs is explicitly posed as an open question (Baptista et al., 2023). In Hamming-space problems, exact extremizers at intermediate densities remain unresolved despite the new upper envelopes for level-$T$58 Fourier weight and the improved LP bounds for binary codes (Yu, 3 Apr 2025, Yu et al., 2019).

AMD is therefore best understood not as a single definition but as a stable research motif. Whether the admissible objects are triangulations, spanning trees, compact connected sets, code subsets, phylogenetic representatives, or embedding clouds, the central problem is to compress geometry, connectivity, or semantics while keeping typical shortest or nearest distances as small as possible.