 Research
 Open Access
Ultrametric networks: a new tool for phylogenetic analysis
 Alberto Apostolico^{1},
 Matteo Comin^{2}Email author,
 Andres Dress^{3} and
 Laxmi Parida^{4}
https://doi.org/10.1186/1748718887
© Apostolico et al.; licensee BioMed Central Ltd. 2013
 Received: 15 October 2012
 Accepted: 18 February 2013
 Published: 5 March 2013
Abstract
Background
The large majority of optimization problems related to the inference of distance‐based trees used in phylogenetic analysis and classification is known to be intractable. One noted exception is found within the realm of ultrametric distances. The introduction of ultrametric trees in phylogeny was inspired by a model of evolution driven by the postulate of a molecular clock, now dismissed, whereby phylogeny could be represented by a weighted tree in which the sum of the weights of the edges separating any given leaf from the root is the same for all leaves. Both, molecular clocks and rooted ultrametric trees, fell out of fashion as credible representations of evolutionary change. At the same time, ultrametric dendrograms have shown good potential for purposes of classification in so far as they have proven to provide good approximations for additive trees. Most of these approximations are still intractable, but the problem of finding the nearest ultrametric distance matrix to a given distance matrix with respect to the L_{ ∞ } distance has been long known to be solvable in polynomial time, the solution being incarnated in any minimum spanning tree for the weighted graph subtending to the matrix.
Results
This paper expands this subdominant ultrametric perspective by studying ultrametric networks, consisting of the collection of all edges involved in some minimum spanning tree. It is shown that, for a graph with n vertices, the construction of such a network can be carried out by a simple algorithm in optimal time O(n^{2}) which is faster by a factor of n than the direct adaptation of the classical O(n^{3}) paradigm by Warshall for computing the transitive closure of a graph. This algorithm, called UltraNet, will be shown to be easily adapted to compute relaxed networks and to support the introduction of artificial points to reduce the maximum distance between vertices in a pair. Finally, a few experiments will be discussed to demonstrate the applicability of subdominant ultrametric networks.
Keywords
 Phylogenetic network
 Ultrametric distance
 STR data analysis
Background
As is well known, most optimization problems related to the inference of distance‐based trees used in phylogenetic analysis and classification are intractable (see [1, 2] for a pertinent discussion). One notable exception is found within the realm of ultrametric distances (cf. [3]). The introduction of such distances in phylogeny was inspired by a model of evolution, now largely abandoned, driven by the postulate of a molecular clock whereby the amount of phylogenetic change observable between any two extant species is directly related to the amount of time that elapsed since their last common ancestor roamed this planet, implying that phylogenetic distances could simply be represented by a weighted tree in which the sum of the weights of the edges separating any given leaf from the root is the same for all leaves.
Both molecular clocks and rooted ultrametric trees fell out of fashion as credible representations of evolutionary change. At the same time, a rooted dated tree is still the “object of desire” in taxonomy and Tree‐of‐Life research, and ultrametric dendrograms have shown good potential for purposes of classification in so far as they have proven to provide good approximations for additive trees. While finding the “best” such approximation is, in most cases, still intractable, the problem of finding an ultrametric distance matrix that is closest to a given distance matrix with respect to the L_{ ∞ } distance has long been known to be solvable in polynomial time, its solution being incarnated in any minimum spanning tree for the weighted graph subtending to the matrix.
Applications of minimum spanning trees in connection with problems of population classification and genetics are as old as any other of their numerous applications. An application to taxonomic problems related to species interrelationship dates back to [4]. And as early as 1964, Edwards and Cavalli Sforza [5] used MSTs to approximate evolutionary trees reconstructed from gene frequencies in blood groups from fifteen contemporary human populations.
Most approximation problems arising in this context fall within the framework of the following
Closest Metric Problem: Given a set$\mathcal{M}$of metrics C defined on a set V, an V×V−m a t r i x M, and a distance function$D:({M}^{\prime},{M}^{\u2033})\to {\mathbb{R}}_{\ge 0}$defined on the set${\mathbb{R}}^{\leftV\right\times \leftV\right}$of all V×V−m a t r i c e s, find a metric$C\in \mathcal{M}$with minimum distance to M relative to D.
Basic facts for the closest metric problem
C∖ D_{ M }  L _{ 1 }  L _{ 2 }  L _{ ∞ } 

Additive  NP‐Hard^{+}  NP‐Hard  NP‐Hard^{*} 
Ultrametric  NP‐Hard^{+}  NP‐Hard  P^{§} 
Subdominant ultrametrics have been traditionally applied to many problems of physics and optimization theory [9]‐[12]. More recently, implications of this theory in the analysis of financial markets, stock exchange, and evolutionary biology have attracted new interest in the topic.
Phylogenetic networks are increasingly featured in modeling of molecular evolution, as evidence of reticulate events such as hybridization, horizontal gene transfer and recombination becomes more prominent. Traditionally, the use of binary data and, in particular, the notion of splits gave rise to a number of alternative models. In the literature, several definitions of networks have been proposed to model parallel events. Popular examples are consensus networks [13], reticulate networks, recombination networks, median networks [14], Neighbour Nets [15], QNets [16] etc. In order to control the degree of connectivity of a network, each model optimizes an objective function; examples are Bayesian methods, maximum likelihood methods, and maximum parsimony [17, 18], calculated that the number of equally parsimonious trees for a data set of just 56 haplotypes exceeded one billion. This estimate was computed through resort of the notion of Minimum Spanning Network. In a different context, they proposed a counting procedure based on the Prim’s algorithm that is analogous to the work presented in this paper.
Anther popular framework is the statistical parsimony analysis [19]. Hart and Sunday [20] found empirically that subnetworks, as implemented in the TCS program [21], coincided significantly with taxonomy names. The TCS program calculates the maximum number of mutational steps constituting a parsimonious connection between two haplotypes with the probability of 95%. Although Hart and Sunday’s [20] results suggest that statistical parsimony analysis could be used in practice to differentiate species, this methodology is not mathematically well‐founded.
In this paper, we extend the approach based on the construction of subdominant ultrametric trees by studying ultrametric networks, consisting of the collection of all edges involved in some minimum spanning tree. This can be viewed as a network of kinship between the extant sequences that embodies the least‐resistant paths in terms of bottlenecks, where a bottleneck is simply the worst possible transition between two intermediate states. We show that, for a graph with n vertices, the construction of such a network can be carried out by a simple algorithm in optimal time O(n^{2}), which is faster by a factor of n than the more straightforward O(n^{3}) closure performed by the classical Floyd‐Warshall paradigm. We show that our algorithm can easily be adapted to compute relaxed networks and to support the introduction of artificial points when it is desirable to reduce maximum distance between vertices. Finally, we discuss a few experiments demonstrating the applicability of this method.
The ultrametric network
that associates a positive weight W(v,u) to every 2‐subset $\{v,u\}\in \left(\genfrac{}{}{0.0pt}{}{V}{2}\right)$ of V that we imagine to be deduced, in one way or the other, from the given sequences, and to represent, for every $\{v,u\}\in \left(\genfrac{}{}{0.0pt}{}{V}{2}\right)$, the observed degree of dissimilarity between u and v.
It is well known (cf. [3]) and easy to see (cf. [8] for a review) that there exists a unique largest ultrametric defined on V and denoted by, say, W^{∗} that is dominated by W, i.e., the (necessarily unique and symmetric) largest map from V×V into $\mathbb{R}$ for which
 (i)
W^{∗}(v,v)=0 and W^{∗}(v,u)≤ max(W^{∗}(v,w),W^{∗}(u,w)) holds for all u,v,w in V, and
 (ii)
W ^{∗}(v,u)≤W(v,u) for all u,v∈V.
W^{∗} must indeed be an ultrametric, called the subdominant ultrametric for W.
In this paper, we will study the ultrametric network G(VW) associated with W, i.e. the graph G(VW):=(V, E(VW)) with vertex set V and edge set $E\left(V\rightW):=\left\{\{u,v\}\in \left(\genfrac{}{}{0.0pt}{}{V}{2}\right):W(u,v)={W}^{\ast}(u,v)\right\}$.
It is easy to see that E(VW) is actually the union of the edge sets of all minimum spanning trees with vertex set V relative to W, considered as a weighting of the complete graph G(V)=(V,E(V)) with vertex set V and edge set $E\left(V\right):=\left(\genfrac{}{}{0.0pt}{}{V}{2}\right)$.
Any path for which this minimum is attained represents a minimum‐bottleneck path (for u and v). Clearly, W^{∗}(u,v) is the lowest weight possible for the highest weight in any path leading from u to v. It can be computed by a straightforward adaptation of the Floyd‐Warshall algorithm in O(V^{3}) time.
Remarkably, E⊆E(VW) holds for every subset E of E(V) for which the graph (V,E) is connected and minimizes the sum $\leftE\right:=\sum _{e\in E}\lefte\right$, that is, for the edge set of any minimum spanning tree T=(X,E) for W. This follows from a result generally credited to [3], that we formalize as follows:
Theorem 1
With V and W as above, the edge set of every minimum spanning tree for W is contained in E(VW) while, conversely, there exists, for any edge e∈E(VW), a minimum spanning tree for W whose edge set contains e. In particular, the network G(VW) is always connected.
Proof
Indeed, given any such subset E⊆E(V) and any edge e={u,v}∈E, we may denote by Π(e)=Π_{ E }(e) the bi‐partition of V given by the (vertex sets of the) two connected components of the graph (V,E−{e}), and by A(w)=A_{ E }(w), for any w∈V, the unique component A(w)∈Π(e) with w∈A(e). Clearly, we have Π(e)={A(u),A(v)} for every edge e={u,v}∈E.
Now, assume that there exists some e={u,v}∈E with e∉E(VW). Then, we could find some P=v_{0}v_{1}...v_{k−1}v_{ k } from v_{0}:=u to v_{ k }:=v in G(V) with B(P)<e. Furthermore, as A(v_{0})=A(u)≠A(v)=A(v_{ k }) must hold, there must be some $i\in \{1,\dots ,k\}$ with A(v_{i−1})≠A(v_{ i }), eg the smallest i in $\{1,\dots ,k\}$ with A(u)≠A(v_{ i }). Consequently, exchanging the edge e with the edge e_{ i }:={v_{i−1},v_{ i }} in E would also give rise to a spanning tree for G(V), and we would have E^{′}=E+e_{ i }−e<E in view of e_{ i }≤B(P)<e, thus contradicting our choice of E. So, E⊆E(VW) must hold, as claimed.
To establish the converse, assume that e={u,v}∈E(VW) is not contained in any minimum spanning tree. Then, given any such tree, let P=v_{0}v_{1}...v_{k−1}v_{ k } denote the unique path from v_{0}:=u to v_{ k }:=v in that tree. Then, exchanging any edge e^{′} in the support of P with the edge e will produce a spanning tree for G(V) of larger weight, implying that e^{′}<e must hold for every such edge e^{′} implying that also B(P)<e must hold. This, however, would clearly contradict our assumption e∈E(VW). □
Optimal computation of the ultrametric network
Clearly, given V and W as above, the ultrametric network can be produced in time O(V^{3}) by a straightforward adaptation of the Floyd‐Warshall all‐pairs shortest‐path algorithm [22]. In view of Theorem 1, this network could be produced in time O(V^{3}) also by first computing one MST by, say, Prim’s algorithm, and then computing W^{∗} using the paths in this tree. We present here an algorithm to compute the entire network in time O(V^{2}). This is optimal since any algorithm must produce Θ(V^{2}) values at the outset.
The main idea is that the computation can be cast within a control structure that is strongly reminiscent of Prim’s MST ‐ or Dijkstra’s single‐source shortest‐path algorithm (refer to, e.g., [22]): starting with an arbitrary vertex r, a subset $\stackrel{\u0304}{V}$ of V is progressively expanded by annexing, at each step, the one vertex u in $V\stackrel{\u0304}{V}$ that is connected to $\stackrel{\u0304}{V}$ by an edge (v,u) that minimizes cost. As is well known, in Prim’s MST the cost to be minimized is the weight of the partial tree over the vertices in $\stackrel{\u0304}{V}$, whence the edge to be chosen is one of minimum weight. In Dijkstra’s algorithm, the cost to be minimized is the sum of weights on the arcs connecting u to r, whence the edge to be chosen is the one minimizing this sum. Note that in both cases there can be more than one vertex that minimizes the cost, however they will all produce the same global minimum. One important point of our algorithm is that (see Theorem 1) choosing (v,u) as in Prim’s MST computes the ultrametric distance not only between u and r but between u and any other vertex in $\stackrel{\u0304}{V}$. Moreover, it can be seen that the pairwise ultrametric distances between any pair of vertices in $\stackrel{\u0304}{V}$ are not affected by the introduction of u in this set. This last circumstance yields the speedup from O(V^{3}) to O(V^{2}).
Lemma 1
At each iteration, d coincides with${\left(W{}_{\stackrel{\u0304}{V}}\right)}^{\ast}$, the subdominant ultrametric of the restriction$W{}_{\stackrel{\u0304}{V}}$of W to$\stackrel{\u0304}{V}$, where the edge {v^{′},v^{″}}∈Ē if and only if:

d(v^{′},v^{″})=W(v^{′},v^{″}) and d(v^{′},v^{″})≤m a x(W(v^{′},u),W(v^{″},u)) holds for any two vertices v^{′},v^{″} in $\stackrel{\u0304}{V}$ and all u∈Q,

$\mathit{\text{key}}\left[u\right]=W\left(\mathit{\text{prec}}\right[u],u)=min\left(W\right(u,v):v\in \stackrel{\u0304}{V})$ holds for every $u\in Q=V\stackrel{\u0304}{V}$.
Lemma 2
The algorithm constructs the subdominant ultrametric and its associated ultrametric network in optimal time O(V^{2}).
Proof
All the initialization steps, inclusive of the insertion of (V−1) initial values in the queue Q take time O(V^{2}). Following this, each of the (V−1) iterations of the while loop contains two cascaded for cycles of O(V) elementary steps each. The first for computes the ultrametric network for the vertices in $\stackrel{\u0304}{V}$, whereas the second one updates the queue Q, which stores all vertices $v\notin \stackrel{\u0304}{V}$ according to the index k e y[v]. All the operations in each for take trivially constant time, except for the queue updates. If the queue is implemented as a Fibonacci heap, we can extract the minimum element in amortized O(logV) and update the queue in amortized O(1), when k e y[v] is decreased. There are (V−1)extractmin (at the beginning of every iteration of the while), which thus charge O(V logV) overall. There are O(V^{2}) constant‐time updates throughout all the executions of the second for loop. Hence the total cost of the algorithm is O(V^{2}). The subdominant ultrametric requires Θ(V^{2}) entries, and an ultrametric network contains at most V^{2} edges, so that O(V^{2}) time is optimal. □
Ultrametric network relaxations
The algorithm of the preceding section lends itself naturally to variants that accommodate some tolerance in the ultrametric distance and relax the notion of ultrametric network. We outline here these two variants, respectively leading to Δ‐ultrametric networks and to the introduction of new artificial vertices.
Δ‐Ultrametric extension
We define the Δ ‐ultrametric network in which edges are inserted if their weights do not deviate more that a given threshold Δ from the corresponding ultrametric distance.
Formally the Δ ‐ultrametric network consists of the graph G_{ Δ }(VW)=(V, E_{ Δ }(VW)) with vertex set V and edge set ${E}_{\Delta}\left(V\rightW):=\left\{\right.\{u,v\}\in \left(\genfrac{}{}{0.0pt}{}{V}{2}\right):W(u,v)\le {W}^{\ast}(u,v)+\Delta \left\}\right.$, where W^{∗}(u,v) is the standard ultrametric distance. Intuitively, the Δ‐ultrametric network is thus a relaxation of the ultrametric network, resulting in increased connectivity. More precisely the graph G_{ Δ }(VW) coincides with the map min(W,W^{∗}+Δ).
The Δ‐ultrametric network can be computed as a postprocessing by adding all such edges E_{ Δ }(VW) to the exact ultrametric network. In summary at first we run the algorithm on the original weights W(u,v) to compute W^{∗}. Then, we apply the postprocess that includes all edges (u,v) with weight W(u,v) that deviates at most Δ from the corresponding ultrametric value W^{∗}(u,v). Similarly to the main algorithm, this postprocess takes optimal O(V^{2}) time. Other alternatives relaxations can be explored, like the subdominant Δ‐ultrametrics for which analogous results can be established. The subdominant Δ‐ultrametrics relaxation will be addressed in a future paper.
Artificial vertices
In applications such as phylogeny on biological data of extant species/individuals, the topology must account for missing data points. In other words, there is a need to reduce the distance between a pair of vertices by introducing a new artificial vertex in the network.
In the phylogeny construction problem, the given data points are the terminals and the artificial vertices correspond to missing (or ancestral) data points. The traditional Steiner tree problem [23] involves the minimization of the sum of the lengths of all edges used after introducing artificial vertices, as opposed to the sum of the pairwise distances of all the terminals. For different metrics the Steiner tree problem is known to be NP‐Hard [24]. Thus in our context, given the graph induced by the ultrametric W^{∗}, the problem of introducing new artificial vertices that minimize the sum of the weights of all edges is still NP‐Hard.
In our case the input graph G is not just any graph, but it can be characterized as follow. Suppose we are given a (big) metric space $\mathcal{R}=(V,D)$ (could be the metric space consisting of the vertex set R of a connected weighted graph G with the “induced" metric, i.e., the largest metric D on V with D(u,v)≤w(u,v) for all edges u,v in G), and a finite subset R of V.
of (V,D). It is natural in this case as any Steiner tree for V can be mapped into T(D) by a non‐expanding map that preserves the distances between the points in V[25]. We don’t know how to efficiently search for the best Steiner tree within T(D), also known as the optimal realization, without an exhaustive enumeration [26, 27]. Instead starting from a given graph ${G}_{{W}^{\ast}}$ over R we look at the “neighborhood" of this seed in T(D). The input graph ${G}_{{W}^{\ast}}$ over R is the ultrametric network computed in the previous section and we are interested in the “neighborhood" of this network such that the sum of all edges is smaller and that the pairwise distances in R are preserved.
To control the number of artificial vertices, the new vertex x is created only if the sum of pairwise distances of the triangle among u, w and y exceeds the threshold Δ. Note that if two triangles share an edge we need to select where to insert the new artificial vertex. A canonical order can be established by ranking all candidate triangles by the sum of pairwise distances. This ensures that, at least generically, the introduction of new artificial vertices is unique and does not depend on the input order.
Experimental results
To conclude our presentation, we report two examples of inference of Human Y‐chromosome phylogeny from Short Tandem Repeats. This can be based on the study of Human migration and the associated relationships among different populations. In typical experiments, we are interested in constructing a network from the STRs information of various individuals and in comparing the results with known paths of migrations. An interesting example of such a phylogeny reconstruction can be found in [28], which discusses the significance of STRs data as markers for human evolution, but also highlights the difficulties that the analysis of this data derives from the lack of an appropriate methodology.
Using the same data of [28, 29], we study migration histories within two different scenarios. In the first experiment, we analyze a very broad spectrum of populations: Africans, Europeans, Asians and Australians. In the second, we concentrate on Native Americans spanning North America (Navajo, Zuni, Sioux), Central America (Maya), and South America (Ticuna, Wichi, Toba, Chorote, Tehuelche, Susque, Humahuaqueño).
For both experiments the data available include a number of different STRs, specifically, 12 in the first experiment and 7 in second. The first step is to establish for all STRs a weighting scheme reflecting the different mutation rates. To this end, we use the three weights 1,2 and 4, and assign to each STR a weight proportional to its mutation rate.
Conclusions
In conclusion, this paper expands the subdominant ultrametric perspective by studying ultrametric networks. We shown that, for a graph with n vertices, the construction of such a network can be carried out by a simple algorithm in optimal time O(n^{2}). This algorithm can be easily adapted to compute relaxed networks, such as Δ‐ultrametric networks and to support the introduction of artificial points to reduce the maximum distance between vertices in a pair. Finally, we discussed a few experiments to demonstrate the applicability of subdominant ultrametric networks.
Declarations
Acknowledgements
Enrico Guariento implemented and tested the software Ultranet. M. Comin was partially supported by the Ateneo Project CPDA110239.
Authors’ Affiliations
References
 Agarwala R, Bafna V, Farach M, Narayanan B, Paterson M, Thorup M: On the approximability of numerical taxonomy: Fitting distances by tree metrics. Proceedings of the 7th Annual ACM‐SIAM Symposium on Discrete Algorithms. 1996, 28 (3): 10731085.Google Scholar
 Farach M, Kannan S, Warnow T: A Robust model for finding optimal evolutionary trees. Algorithmica, special issue on Computational Biology. 1996, 13: 155179.Google Scholar
 Gower J, Ross G: Minimum spanning trees and single linkage cluster analysis. Appl Stat. 1969, 18: 5464. 10.2307/2346439.View ArticleGoogle Scholar
 Florek K, Lukaszewickz J, Perkal H, Steinhaus H, Zubrzycki S: Sur la Liaison et la Division des Points d’un Ensemble Fini. Colloq Matematicum. 1951, 2: 282285.Google Scholar
 Edwards A, Sforza LC: Reconstruction of evolutionary trees. Phenetic Phylogenet Classif. 1964, 6: 6776.Google Scholar
 Gromov M: Hyperbolic groups, essays in group theory, MSRI series vol. 8, S.Gersten, ed., Springer‐Verlag (1988). trees and single linkage cluster analysis. Appl Stat. 1969, 18: 5464. 10.2307/2346439.View ArticleGoogle Scholar
 Bowditch B: Notes on Gromov’s hyperbolicity criterion for path metric spaces. E. Ghys et al.: E. Ghys et al., Proceedings of Group Theory from a Geometric Viewpoint, World Scientific. Singapore 1991, 64167.Google Scholar
 Dress A, Huber K, Moulton V: Some uses of the Farris Transform in Mathematics and Phylogenetics – A Review. Ann Combinatorics, Special Volume Biomathematics. 2007, 11: 137.View ArticleGoogle Scholar
 Bayod JM, Martinez‐Maurica J: Subdominant ultrametrics. Proc Am Math Soc. 1990, 109 (3): 829834. 10.1090/S00029939199010156761.View ArticleGoogle Scholar
 Parisi G: Spin glasses and fragile glasses: Static, dynamics and complexity. PNAS. 2006, 103 (21): 79487955.PubMed CentralView ArticlePubMedGoogle Scholar
 Rammal R, Toulouse G, Virasoso MA: Ultrametricity for Physicists. Rev Mod Phys. 1986, 58: 765788. 10.1103/RevModPhys.58.765.View ArticleGoogle Scholar
 Soete GD: Ultrametric tree representations of incomplete dissimilarity data. J Classif. 1984, 1: 235242. 10.1007/BF01890124.View ArticleGoogle Scholar
 Huson D, Nettles S, Warnow T: Obtaining accurate topology estimates of evolutionary trees from very short sequences. Proc RECOMB. 1999, 198207.Google Scholar
 Bandelt HJ, Forster P, Rühl A: Median‐joining networks for inferring intraspecific phylogenies. Mol Biol Evol. 1999, 16: 3748.View ArticlePubMedGoogle Scholar
 Bryant D, Moulton V: Neighbor‐Net: An agglomerative method for the construction of phylogenetic networks. Mol Biol Evol. 2004, 21: 255265.View ArticlePubMedGoogle Scholar
 Grünewald S, Forslund K, Dress A, Moulton V: QNet: an agglomerative method for the construction of phylogenetic networks from weighted quartets. Mol Biol Evol. 2007, 24: 532538.View ArticlePubMedGoogle Scholar
 Saitou N, Imanishi T: Relative effciencies of the fitch‐margoliash, maximum parsimony, maximum likelihood, minimum evolution, and neighbor‐joining methods of phylogenetic tree construction in obtaining the correct tree. Mol Biol Evol. 1989, 6: 514525.Google Scholar
 Excoffier L, Smouse P: Using allele frequencies and geographic subdivision to reconstruct gene genealogies within a species: Molecular variance parsimony. Genetics. 1994, 136: 343359.PubMed CentralPubMedGoogle Scholar
 Templeton A, Crandall K, Sing C: A cladistic analysis of phenotypic associations with haplotypes inferred from restriction endonuclease mapping and DNA sequence data. III. Cladogram estimation. Genetics. 1992, 132: 619633.PubMed CentralPubMedGoogle Scholar
 Hart M, Sunday J: Things fall apart: biological species form unconnected parsimony networks. Biol Lett. 2007, 3: 509512.PubMed CentralView ArticlePubMedGoogle Scholar
 Clement M, Posada D, Crandall K: TCS: a computer program to estimate gene genealogies. Mol Ecol. 2000, 9: 16571659.View ArticlePubMedGoogle Scholar
 Cormen TH, Leiserson CE, Rivest RL, Stein C: Introduction to Algorithms. Cambridge: MIT Press 2001.Google Scholar
 Hwang FK, Richards DS, Winter P: The Steiner Tree Problem, Volume Annals of Discrete Mathematics. Amderdam: Elsevier Science Publishes 1992.Google Scholar
 Bern M, Eppstein D: Hardness of approximations algorithms for geometric problems. Approximation algorithms for NP‐Hardness Problems edited by D. S. Hochbaum. Boston, MA: PWS Publishing Company 1997, 296345.Google Scholar
 Dress A, Huber KT, Moulton V: Metric spaces in pure and applied mathematics. Documenta Mathematica (Proceedings Quadratic Forms LSU). 2001, 121139.Google Scholar
 Dress A, Huber KT, Koolen JH, Moultonm V: Block realizations of finite metrics and the tight‐span construction I: The embedding theorem. Discrete Appl Math. 2008, 21 (12): 13061309.Google Scholar
 Dress A, Wu T, Xu X: A note on single‐linkage equivalence. Appl Math Lett. 2010, 23: 432435. 10.1016/j.aml.2009.11.012.View ArticleGoogle Scholar
 Forster P, Rohl A, Lunnemann P, Brinkmann C, Zerjal T, Tyler‐Smith C, Brinkmann B: A short tandem repeat‐based phylogeny for the human Y chromosome. Am J Hum Genet. 2000, 67: 182196.PubMed CentralView ArticlePubMedGoogle Scholar
 Bianchi N, Catanesi C, Bailliet G, Martinez‐Marignac V, Bravi C, Vidal‐Rioja L, Herrera R, Lopez‐Camelo J: Characterization of ancestral and derived Y‐chromosome haplotypes of new world native populations. Am J Hum Genet. 1998, 63 (6): 18621871.PubMed CentralView ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.