Phylogenetic Vignettes: Owen-Provan algorithm

On September 19, 2022September 20, 2022 By In math, Phylogenetics1 Comment

Important announcement

I am currently deciding on the direction this post series will take after next week. Please check out the What’s next section and let me know if you have any preferences among described routes.

Recap

In the previous post (following Gavryushkin and Drummond, 2016) we have defined a geometry on the space of phylogenetic ultrametric trees and called the resulting metric space 𝛕-space. We also briefly mentioned the geometry on the space of trees defined by Billera, Holmes and Vogtmann, 2001, and referred to the resulting construction as BHV space.

In this post we will provide a brief overview of the algorithm presented in Owen and Provan, 2010, which allows us to compute the geodesics in 𝛕-space and BHV space in polynomial time (in the number of the tree leaves/taxa). For the ease of exposition we will delegate detailed overview of the BHV space to a later post. Hence, in what follows we will assume without the proof that BHV space is a CAT(0) complex and trees in BHV space are parametrized by the internal branch lengths with different orthants corresponding to different tree topologies.

Goal

As we mentioned before one of the key properties we want from a metric space structure on the tree space is the efficient computation of geodesics. Clearly given any finite metric space (complexes constructed in BHV or 𝛕-spaces are not finite, but the topologies are and the problem of distance within a fixed orthant determined by topology is easy to solve) we can simply consider all possible paths and then pick the one minimizing the length function. However, such approach yields an algorithm that will fail to finish computations even for modestly sized spaces.

Thus, our goal is to have an efficient algorithm that can find shortest path in polynomial time with respect to the number of leaves our trees can have.

CAT(0) properties

Since both BHV and 𝛕-spaces are CAT(0) we can leverage nice properties of the non-positively curved geometries to get an idea for an efficient algorithm. The key idea behind the algorithm comes from the following lemma (Lemma 2.1 in Owen and Provan, 2010, and Proposition 1.4, p. 160 in ).

Lemma 1. In CAT(0) space, every local geodesic is a geodesic.

Proof. Consider a local geodesic $\Gamma_\varepsilon$ between points $P$ and $Q$ . Choose a set of points $P=P_0,P_1,...,P_k=Q$ such that $L(\Gamma_\varepsilon(P_i, P_{i+1}))\leq \varepsilon/2$ . Note that it follows from the definition of a local geodesic that $\Gamma_\varepsilon(P_{i-1}, P_{i+1})$ is a geodesic. Now, assume by induction that $\Gamma_\varepsilon(P_0, P_j)$ is a geodesic for all $j\leq i$. Hence, we have $L(\Gamma_\varepsilon(P_0, P_{i-1}))=d(P_0, P_{i-1})$ by the induction hypothesis and $L(\Gamma_varepsilon(P_{i-1},P_{i+1}))=d(P_{i-1}, P_{i+1})$ as noted above. Consider a corresponding model triangle $p_0p_{i-1}p_{i+1}$ in Euclidean space (if you forgot what a model triangle is check: Bridson and Haefliger, 2013 p. 158 or previous post on CAT spaces). Let $p_i$ be a point on $p_{i-1}p_{i+1}$ such that $|p_{i-1}p_i|=d(P_{i-1}, P_i)$ . Since we are working in a CAT(0) space it follows that $d(P_0, P_i)\leq |p_0p_i|$ . On the other hand, by induction hypothesis we have $d(P_0, P_i) = d(P_0, P_{i-1}) + d(P_{i-1}, P_i) = |p0p_{i-1}|+|p_{i-1}p_i|$ . Combining this result with triangle inequality gives us $|p0p_{i-1}|+|p_{i-1}p_i|=|p_0p_i|$ , and hence $p_{i-1}\in p_0p_i$ and therefore $p_{i-1}\in p_0p_{i+1}$ . From here we can conclude that $\Gamma_\varepsilon(P_0, P_{i+1}$ is a geodesic and hence $\Gamma_\varepsilon$ is a geodesic.

Thus, the main idea used in the paper is to start with any path between two trees (consider the path from a tree to the star tree and then back to target tree, this path called cone path is a common idea in the analyses of tree spaces; recall: Robinson-Foulds metric) and iteratively check whether it’s a local geodesic refining the path in cases when that’s not the case.

Path space geodesics

Following Owen and Provan, 2010 we will assume that the trees $T, T^\prime$ between which we are computing geodesics have distinct edges. We will also call two sets of edges $A\subseteq E(T)$ and $B\subseteq E(T^\prime)$ compatible if all pairs of splits associated with these edges are compatible, or equivalently if the union $A\cup B$ defines a tree. In other words, compatible sets of edges produce “intermediate” trees between $T$ and $T^\prime$ . Now, for a pair of partitions $\mathcal{A}=\{A_1,...,A_k\}$ , $\mathcal{B}=\{B_1,...,B_k\}$ of $E(T)$ and $E(T^\prime)$ , respectively, we can define an associated path space given that the following property holds: for every $i>j$ , $A_i$ and $B_j$ are compatible. To construct the path space consider the collection of orthants $\mathcal{P}(T, T^\prime)=\cup_{i=1}^k O_i$ given by $O_i=O(B_1\cup ...\cup B_i\cup A_{i+1}\cup ...\cup A_k)$ , where $O(T)$ is the mapping of a tree to the corresponding orthant. We will call the corresponding partition pair $(\mathcal{A}, \mathcal{B})$ the support, and the shortest path between $T$ and $T^\prime$ in $\mathcal{P}$ the path space geodesic. These definitions make the geodesic in the tree space a more concrete and constructable object, which of course is important for any algorithmic approach. Furthermore, the following theorem from Billera, Holmes and Vogtmann, 2001 (see Proposition 4.1) guarantees that the path space geodesics are recovering the true geodesics.

Theorem 1. For disjoint trees $T, T^\prime$ the geodesic is a path space geodesic for some choice of $\mathcal{P}(T, T^\prime)$ .

We will also need the following two results, the first of which is proven in Owen, 2011 and the second in Owen and Provan, 2010 (Theorem 2.5).

Theorem 2. Let $\Gamma$ be a geodesic between $T$ and $T^\prime$ . Then there exist partitions $\mathcal{A}=\{A_1,...,A_k\} and$ latex \mathcal{B}=\{B_1,…,B_k\} of $E(T)$ and $E(T^\prime)$ , respectively, such that $(\mathcal{A}, \mathcal{B})$ (a) form a support of path space in which $\Gamma$ is a path space geodesic, and (b) the following condition is satisfied:

$\frac{\parallel A_1\parallel}{\parallel B_1\parallel}\leq \frac{\parallel A_2\parallel}{\parallel B_2\parallel}\leq\cdots\frac{\parallel A_k\parallel}{\parallel B_k\parallel}$ .

A path space satisfying the above condition for its support is referred to as a proper path space, and the corresponding path space geodesic is called a proper path.

Theorem 3. A proper $(T, T^\prime)$ -path $\Gamma$ with support $(\mathcal{A}, \mathcal{B})$ is a geodesic iff for each pair of edge sets $(A_i, B_i)$ in the support there is no nontrivial partition $C_1\cup C_2=A_i$ , $D_1\cup D_2=B_i$ such that $C_2$ is compatible with $D_1$ and $\frac{\parallel C_1\parallel}{\parallel D_1\parallel}\leq \frac{\parallel C_2\parallel}{\parallel D_2\parallel}$ .

Theorem 3 gives us the final ingredient needed to build out an algorithm for finding BHV space geodesics. Namely, partitions of edge sets give us a way to construct proper paths and the condition on local improvement of such proper paths guarantees that we achieve a geodesic.

Algorithm overview

As we mentioned earlier the natural choice for the starting path is the cone path between the two trees $T$ and $T^\prime$ , more formally it’s a proper path space defined by the support $(\mathcal{A}=\{E(T)\}, \mathcal{B}=\{E(T^\prime)\})$ and the corresponding path is given by uniform contraction of all edges which results in star tree and then a follow up uniform decontraction. Now, the algorithm will proceed iteratively at each step checking whether the condition in Theorem 3 is satisfied and if not proposing a new support and a new proper path.

In order to facilitate an efficient solution for the iterative step, the problem of checking the condition of Theorem 3 and finding a new path can be recasted as an independent set problem over a bipartite graph. Namely, we will construct the incompatibility graph $G(A, B) between two edge sets$ latex A\subseteq E(T), B\subseteq E(T^\prime)$ as a bipartite graph with vertices given by $A\cup B$ and edges given by the pairs $e\in A, f\in B$ where the splits associated with $e$ and $f$ are incompatible. It is easy to see that for any $A\subseteq E(T), B\subseteq E(T^\prime)$ compatibility of $A$ and $B$ is equivalent to $A\cup B$ forming an independent set in $G(E(T), E(T^\prime))$.

Thus, the condition from Theorem 3, can now be recasted as follows: Given two edge sets $A\subseteq E(T), B\subseteq E(T^\prime)$ does there exist a pair of partitions $A=C_1\cup C_2$ , $B=D_1\cup D_2$ such that the vertex set $C_2\cup D_1$ is independent in $G(A, B)$ and

$\frac{\parallel C_1\parallel}{\parallel D_1\parallel}\leq \frac{\parallel C_2\parallel}{\parallel D_2\parallel}$ .

In particular, a proper path with support $(\mathcal{A}, \mathcal{B})$ is geodesic iff no edge set pair $(A_i\in\mathcal{A},B_i\in\mathcal{B})$ satisfies the above condition.

Now, since the complement of an independent set is a vertex cover we can reformulate the above problem as a minimum weight (by transforming the inequality condition) vertex cover problem, which can be solved on bipartite graphs in $O(n^3)$ . Further analysis of the algorithm steps yields that the total runtime to find the geodesic is $O(n^4)$ . For the exact details of the proof of correctness of the algorithm and the runtime I suggest reading the original manuscript by Owen and Provan, 2010.

Data/code availability

This post has no associated custom code or data. Those looking for an implementation of the algorithm described should check out Megan Owen’s webpage or GitHub repo.

What’s next

This post concludes the beginning of our exploration of the geometry of the space of phylogenetic trees. There are several ways we can proceed from here.

First, we can finally start tackling the foundational paper of Billera, Holmes and Vogtmann, 2001. This path will likely require us to spend a few posts dissecting original BHV space and some associated results from metric geometry and CAT(0) spaces. We then can return to the paper of Gavryushkin and Drummond, 2016 and explore the t-space construction in a way similar to our exploration of the 𝛕-space. Finally, we can then continue our journey by catching up to more recent work by Megan Owen et al. on defining and computing statistics in BHV space (Brown and Owen, 2020), comparing phylogenetic trees with different leaf sets (Grindstaff and Owen, 2018), and extending the framework to phylogenetic cactuses (a type of phylogenetic network, Huber et al., 2021).

Second, we can instead pivot completely and ask a general question of what other geometries one can consider on the space of phylogenetic trees. In particular, we can take a look at algebraic geometry inspired view of the phylogenetic tree space as a tropical geometry (Monod et al., 2018), and an information geometry angle which gives rise to wald space (pun credit to the authors) of the phylogenetic trees (Garba et al., 2021). I will most likely skip the category theoretic angle for the topic (Baez and Otter, 2015), since frankly speaking I have no desire to introduce the notion of an operad.

Third, we can temporarily halt the process of investigating deeper into the various geometries on the tree space and ask more broadly what are geometries of interest in the case of phylogenetic networks. This route will require us to go back to some definitions of the phylogenetic network and related concepts, and will likely demand at least some amount of rehashing of the biological context.

Since, I am still debating which route is of the most interest to me at the moment, next week’s post will likely be sparser than usual, as I am figuring out which rabbit hole to pick as the next destination. However, if you actually would prefer me to pick one of these three directions do not hesitate to comment or reach out to me via other means of contact.

References

Baez, John C., and Nina Otter. “Operads and phylogenetic trees.” arXiv preprint arXiv:1512.03337 (2015). arXiv:1512.03337

Billera, Louis J., Susan P. Holmes, and Karen Vogtmann. “Geometry of the space of phylogenetic trees.” Advances in Applied Mathematics 27, no. 4 (2001): 733-767. https://doi.org/10.1006/aama.2001.0759

Bridson, Martin R., and André Haefliger. Metric spaces of non-positive curvature. Vol. 319. Springer Science & Business Media, 2013. https://doi.org/10.1007/978-3-662-12494-9

Brown, Daniel G., and Megan Owen. “Mean and variance of phylogenetic trees.” Systematic biology 69, no. 1 (2020): 139-154. https://doi.org/10.1093/sysbio/syz041, arXiv:1708.00294

Garba, Maryam Kashia, Tom MW Nye, Jonas Lueg, and Stephan F. Huckemann. “Information geometry for phylogenetic trees.” Journal of Mathematical Biology 82, no. 3 (2021): 1-39. https://doi.org/10.1007/s00285-021-01553-x, arXiv:2003.13004

Gavryushkin, Alex, and Alexei J. Drummond. “The space of ultrametric phylogenetic trees.” Journal of theoretical biology 403 (2016): 197-208. arXiv:1410.3544

Grindstaff, Gillian, and Megan Owen. “Geometric comparison of phylogenetic trees with different leaf sets.” arXiv preprint arXiv:1807.04235 (2018). arXiv:1807.04235

Huber, Katharina T., Vincent Moulton, Megan Owen, Andreas Spillner, and Katherine St John. “The space of equidistant phylogenetic cactuses.” arXiv preprint arXiv:2111.06115 (2021). arXiv:2111.06115

Monod, Anthea, Bo Lin, Ruriko Yoshida, and Qiwen Kang. “Tropical geometry of phylogenetic tree space: a statistical perspective.” arXiv preprint arXiv:1805.12400 (2018). arXiv:1805.12400

Owen, Megan. “Computing geodesic distances in tree space.” SIAM Journal on Discrete Mathematics 25, no. 4 (2011): 1506-1529. arXiv:0903.0696

Owen, Megan, and J. Scott Provan. “A fast algorithm for computing geodesic distances in tree space.” IEEE/ACM Transactions on Computational Biology and Bioinformatics 8, no. 1 (2010): 2-13. https://doi.org/10.1109/TCBB.2010.3

Phylogenetic Vignettes: 𝛕-space

On September 12, 2022September 19, 2022 By In math, Phylogenetics1 Comment

In the previous post we have introduced the notion of a CAT(0) space, and briefly talked about some of the properties of CAT(0) spaces, and in particular cubical complexes. In this post, we will take this knowledge a step further into the applied direction of phylogeny. Following Gavryushkin and Drummond, 2016 we will define a metric on the space of ultrametric phylogenetic trees, show that the resulting space is CAT(0), and then discuss some of the implications of these results. Similarly to the previous post, we will have quite a bit of mathematical notation to deal with, so I will attempt to use visual aids whenever possible to improve exposition.

Motivation

One way of defining a metric on a space of phylogenetic trees is to embed the tree space $\mathcal{T}$ into some metric space $\mathcal{M}$ . Thus, by associating every tree in the tree space with some point in the chosen metric space, we induce a pseudometric on tree space. However, such embeddings/parameterizations are not guaranteed to be “nice” with respect to the questions we aim to answer. Namely, not all embeddings are injective (multiple distinct trees can end up being mapped to the same point in $M$ , hence in general we induce a pseudometric rather than a metric via embedding) although in the case of Gavryushkin and Drummond, 2016 the embedding has to be injective by definition. Additionally, not all embeddings are surjective (meaning that a path in $M$ might contain non-tree associated points) which creates a plethora of problems ranging from potential multiple distance minimizing solutions to the lack of tree space midpoints.

Thus, it is natural to formulate a list of desiderata for the embedding $p:\mathcal{T}\to\mathcal{M}$ that would enable fruitful analyses. From now onwards (unless otherwise specified) we will assume that the embedding $p$ is injective.

In order to make continuous distributions behave correctly under the pullback into the tree space, we have to require the image of the embedding to be path-connected. Note, that having a path connected image is also intuitively “nice” as the continuous tree transformations would result in a path being traced out in the model metric space. Hence, we have:

(D1) The set $\mathrm{Image}(p)$ is path-connected in $\mathcal{M}$ .

However, being path-connected does not guarantee that the shortest path between the two points will lie within the image. Hence, we also want:

(D2) The set $\mathrm{Image}(p)$ is convex in $\mathcal{M}$ .

We note that the injectivity criterion forces a lower bound on the dimension of $\mathcal{M}$ . However, if we want to define probability measure on $\mathcal{M}$ that can be meaningfully pulled back onto $\mathcal{T}$ we need to ensure that:

(D3) $\mathrm{Image}(p)$ has the same dimension as $\mathcal{M}$ .

Next, for statistical analyses to stay sound we need to ensure uniqueness of the shortest paths. This is the point in our desiderata list where you might have a brief flashback to the CAT(0) space discussion from previous week.

(D4) The space $\mathcal{M}$ is uniquely geodesic.

Finally, since we are ultimately tackling these questions from the computer science viewpoint, we have the last desiderata that ties this into the computational realm. Namely, we want the following:

(D5) Geodesics in $\mathcal{M}$ are computable.

In reality we probably want an even stronger version of (D5), which is:

(D5′) Geodesics in $\mathcal{M}$ are efficiently (i.e. polynomial time) computable.

It’s worth remarking that in a general metric space we do not necessarily have (D5) with a simple example being the halting problem reduction to shortest paths in an infinite graph.

It is also important to realize that these desiderata do not guarantee existence of such embeddings, and are rather criteria for discerning more or less useful parameterizations of the tree space.

A quick aside

In general, we can attempt to parametrize the whole tree space (that’s precisely what Billera, Holmes and Vogtmann, 2001 do in their paper, which lies at the foundation of the connection between the geometry of non-positive curvature spaces and phylogenetic trees; we will refer to this space as the BHV space going forward) or we might focus on a specific subclass of trees such as ultrametric phylogenetic trees (recall that a tree is ultrametric is the distance from root to every leaf is the same). The caveat is that a parametrization that enjoys our desiderata (D1)-(D5′) for the whole tree space can fail to do so when restricted to the space of ultrametric trees. This is precisely the case for the BHV space parametrization. Hence, it can be of interest (motivated by the evolutionary models context) to explore parametrizations designed specifically for the space of the ultrametric trees. This is precisely what is done in Gavryushkin and Drummond, 2016, and what we are aiming to do in this post.

𝛕-space

We now will construct one version of a space of ultrametric trees called the 𝛕-space. To proceed consider an ultrametric tree T on n taxa, with internal nodes ordered by their time from the extant taxa. We will then parametrize the tree by mapping it to its ranked topology and a n-1-dimensional vector $\overline{\tau}=(\tau_1,...,\tau_{n-1})$ , where $\tau_i$ is the time difference between the i-th and i+1-th nodes. The picture below taken from Gavryushkin and Drummond, 2016 represents this parametrization for a tree on 5 taxa.

Hence, formally we have the mapping given by $p(T)=(\mathrm{rt}(T), \overline{\tau}(T)$ that embeds the space of ultrametric phylogenetic trees into a disjoint union of $m=\frac{n!(n-1)!}{2^{n-1}}$ (Semple and Steel, 2003) non-negative real n-1-dimensional orthants (i.e. sets of the form $\mathbb{R}^{n-1}_0=\{(x_1,...,x_{n-1}\in\mathbb{R}^n|x_i\geq 0\}$ ). We will impose an upper bound on all orthants to turn our construction into a cubical complex for the ease of the exposition. However, the construction of 𝛕-space with orthants will still yield a CAT(0) space with the key properties and desiderata (D1)-(D5′) preserved.

Note that the ranked topology in this case is a stricter condition on the shape of the tree than just the topology constraint. In particular, the two trees pictured below have distinct ranked topologies, and hence will belong to two distinct orthants in the 𝛕-space.

**Figure 2.** An example of two distinct ranked topologies.

Finally, to properly turn this into a cubical complex we will need to identify the isometries that glue the faces of our cubes together. However, this is rather obvious by construction, since whenever any 𝛕 coordinate becomes 0 we observe a collapse of two nodes in the ranked hierarchy to the same level. Hence, for example the two trees shown above will belong to the cubes that share the $\tau_2=0$ face. Since the resulting space is a cubical complex, we can naturally consider the Euclidean metric within each cube with paths for joining points in different cubes being the sums of the respective within cube paths.

The clip below shows how a tree can vary in 𝛕-space while being restricted to a single orthant determined by its ranked topology. Note that just as we discussed above in order to stay within the orthant in 𝛕-space we must keep the ranked topology constant.

For comparison, we show how a similar ultrametric tree can vary within a single fixed orthant of a BHV-space. In this case, the topology of the tree is fixed, but the ranked topology can vary.

Properties

After defining the 𝛕-space it is natural to ask whether it manages to achieve our desiderata. In order to check these properties we need to first figure out how many common faces two cubes in our complex can share (recall from the previous post that a theorem of M. Gromov gives characterization of CAT(0) cubical complexes in terms of face-sharing properties).

In general, similarly to the previous post by face of a cube we mean any sub-cube, thus it is useful to have a separate term for faces of codimension 1 (i.e. faces that are 1 less dimensional than the cubes). Hence, we will call faces of codimension 1 facets. In particular, any facet can be shared by 1, 2 or 3 cubes in total. If we let $t_1=0$ , then the resulting facet is only contained in the cube of corresponding rooted topology. If setting $t_i=0$ does not result in a multifurcating topology then the corresponding facet will be shared by exactly two cubes (see example in Figure 2 and/or Figure 2 of Gavryushkin and Drummond, 2016). Finally, if we get a multifurcation then the total number of cubes sharing the facet is 3.

Now, we are ready to take a stab at the critical result about the 𝛕-space, namely that it is a CAT(0) space, and hence uniquely geodesic. We will recall the theorem of M. Gromov here to reiterate the key result we need to prove.

Theorem 1. A cubical complex $K$ with the intrinsic Euclidean metric is CAT(0) if and only if $K$ is connected, simply connected, and for all natural $k$ : if three $(k+2)$ -cubes of $K$ share a common $k$ -cube and pairwise share common distinct $(k+1)$ -cubes, then they are contained within a $(k+3)$ -cube in $K$ .

We start by noting that the metric we introduced on our cubical complex is precisely the intrinsic Euclidean metric, and it is easy to see from our construction that the resulting complex is connected and simply connected.

To finalize the proof we need to introduce the concept of the link of a vertex v in a complex. We say that the graph G with nodes given by facets containing v and edges given by cubes that contain the two incident nodes as facets is the link of the vertex v.

First, we need to show that the cubes of the dimension (k+2) in the theorem cannot be the highest dimensional cubes in the complex. If that was the case, then the link of origin would have to contain a 3-cycle (given by the pairwise distinct (k+1)-cubes) which contradicts the result that the nearest-neighbor interchange graph does not contain 3-cycles. Thus, it follows that each of the (k+2)-cubes has at least one coordinate equal to 0. For the sake of clarity of the presentation we will assume that this is unique coordinate for each of the cubes, although the similar argument can be made in the general case. Let i, j, and r denote the respective coordinates in three cubes. We have three cases to analyze:

(i=j=r) This is impossible due to no 3-cycle property of the link of the origin.

(i=j) In this case the first two cubes must share a (k+1)-cube, and since all shared (k+1)-cubes must be pairwise distinct it cannot be the cube obtained via setting $\tau_r=0$ . Hence, there exists $\tau_s$ s.t. it is greater than zero in both of the first two cubes, and is zero in their shared (k+1)-cube. Now, if the first and third cube share a (k+1)-cube then it follows that their i and r coordinates both have to be zero in the shared cube, implying that the s coordinate has to be resolved in the same way between the first and third cube. However, the same exact argument can be repeated for the second and third cube implying that the first and second cubes had to be indetical.

(all distinct) In this case the left out coordinate (r for the first and second cube, i for the second and third, j for the first and third) has to be resolved the same way in the two cubes under consideration. Now, we can construct a (k+3)-cube that contains all three cubes by taking the first cube and resolving its zero coordinate (i) the same way as it is resolved in the remaining two cubes.

This analysis concludes the proof of the property required to claim that the cubical complex we defined is indeed CAT(0). It immediately follows that the geodesics in 𝛕-space are unique. At this point we are essentially done with (D1)-(D4) and the only remaining piece is the (efficient) computability of the said geodesics.

Efficiently computing CAT(0) geodesics

In their manuscript Gavryushkin and Drummond, 2016 indicate that the algorithm proposed by Owen and Provan, 2010 will work in 𝛕-space, and provide a link to a Java implementation hosted here:

Due to the constraints on the length and scope of the post we will not dive into how does the algorithm of Owen and Provan, 2010 work in the 𝛕-space. Instead, we will discuss the algorithm in its own dedicated post.

Data and code availability

Code used to generate the 𝛕-space and BHV-space tree examples, as well as HTML outputs from Bokeh are provided in the following archive.

Phylogenetic-vignettes-tau-space Download

What’s next

We spent a noticeable amount of time discussing mathematical properties of different metrics that can be imposed on tree spaces. As our next step, we will switch the lens to a computational perspective and explore the algorithm proposed by Owen and Provan, 2010. We might also touch upon more recent work in the area of the efficient computation of geodesics in CAT(0) cubical complexes. Once we are satisfied with our algorithmic solutions, we will continue the journey by exploring the BHV-space and t-space of phylogenetic trees.

References

Bridson, Martin R., and André Haefliger. Metric spaces of non-positive curvature. Vol. 319. Springer Science & Business Media, 2013. https://doi.org/10.1007/978-3-662-12494-9

Gavryushkin, Alex, and Alexei J. Drummond. “The space of ultrametric phylogenetic trees.” Journal of theoretical biology 403 (2016): 197-208. arXiv:1410.3544

Semple, Charles, and Mike Steel. Phylogenetics. Vol. 24. Oxford University Press on Demand, 2003.

Phylogenetic Vignettes: CATs you can’t pet

On September 5, 2022September 19, 2022 By In math, Phylogenetics2 Comments

Unlike the previous two posts it will not be obvious how what we will discuss connects to phylogeny. Furthermore, we will be diving into a noticeably more mathematically dense content this time. I will try to use some analogies and visual aids through out this post in order to make the material slightly more accessible.

Introduction

Geometry is a wonderful subject (even if I struggled with it in high school) that makes appearances in many aspects of various sciences (including social ones). However, musing about geometry at large is better left to people who are (a) more proficient in the study of it, and (b) have an acumen for writing more lengthy works, hence I will simply point you in the direction of Jordan Ellenberg’s Shape. What we will concern ourselves with today is a more specific slice of the geometry in which we will look at metric spaces of non-positive curvature. I will not define the notion of a metric space again, as you can simply navigate either to the previous post in the series or Wikipedia. The bulk of this post will consist of defining non-positive curvature, and then we will follow up with a few properties of such spaces, and some examples.

Most of the content that follows will adopt the definitions and notation from Bridson and Haefliger, 2013.

Geodesics

Geodesics are a natural notion extending the concept of the “shortest” path curve to general metric space framework. More precisely, given a metric space $(X, d)$ a geodesic from $x$ to $y$ is the map $c:[0, l]\to X$ s.t. $c(0)=x,\, c(l)=y$ and $d(c(t), c(t^\prime))=|t-t^\prime|$ . We will call a metric space a geodesic space if for any $x, y\in X$ there exists a geodesic from $x$ to $y$ . Furthermore, if for all pairs of points such geodesic is unique we will call such space uniquely geodesic (see Bridson and Haefliger, 2013, pp. 4-8 for additional details and examples).

For example, the classic Euclidean space is uniquely geodesic with the geodesics given by straight lines, i.e. for any two points $x,y\in\mathbb{R}^n$ we have the geodesic given by $c(t)=(1-t)x+ty$ . On a sphere geodesic between two points $x, y$ is the arc segment obtained by intersecting a plane through $x, y$ , and the center of the sphere with the surface (i.e. the great circle arc). An example of a geodesic triangle on a sphere is given below (image courtesy of Wikipedia).

Note, that a sphere is not a uniquely geodesic space, since any pair of diametrically opposed points has an infinite set of possible geodesics between them.

Triangles

Just like the notion of a geodesic extends our naïve understanding of the shortest paths, the notion of a geodesic triangle (Bridson and Haefliger, 2013, p. 158) extends our notion of a triangle. Given a metric space $(X, d)$ a geodesic triangle $\Delta$ is a set of tree points $p, q, r\in X$ called vertices and a choice of three geodesic segments $[p, q], [q, r], [r, p]$ joining them called sides, we will denote such a triangle $\Delta([p, q], [q, r], [r, p])$ or more briefly $\Delta(p, q, r)$ . Note, that the later notation is not precise, as in the case of a non-uniquely geodesic space there is not necessarily a unique choice of a geodesic between two vertices. Finally, we will write $x\in\Delta$ to indicate that $x$ belongs to the union $[p, q]\cup [q, r]\cup [r, p]$ .

In order to define a CAT(k) space we also need the notion of a comparison triangle. In order to avoid a lengthy (and somewhat involved) discussion of the nuances about the model spaces and corresponding notation, we will focus on three general types of model spaces: $M_0^n=\mathbb{E}^n$ Euclidean n-dimensional space, $M_{1}^n=\mathbb{S}^n$ the n-sphere, and $M_{-1}^n=\mathbb{H}^n$ hyperbolic n-space. We already defined geodesics in the case of $\mathbb{E}^n$ . For the n-sphere we will define the metric via cosine distance $\cos d(x, y) = x^\top y$ where the inner product is in the corresponding embedding $\mathbb{S}^n\to \mathbb{E}^{n+1}$ and the corresponding geodesics are given by minimal great arcs (for full definition and description see Bridson and Haefliger, 2013, pp. 16-17). For the hyperbolic n-space we will use the same construction as Bridson and Haefliger, 2013 (p. 18-20) by considering the space $\mathbb{E^{n, 1}}$ which consists of $\mathbb{R}^{n+1}$ with bilinear form $\langle u|v\rangle = -u_{n+1}v_{n+1}+\sum_{i=1}^n u_iv_i$ and defining $\mathbb{H}^n=\{u\in\mathbb{E^{n, 1}}|\langle u|u\rangle = -1, u_{n+1}> 0\}$ . The distance on this space will be given by $\cosh d(x, y) = -\langle x|y\rangle$ (similarly to the sphere case, see Bridson and Haefliger, 2013, pp. 18-23 for details).

Finally, a comparison triangle $\overline{\Delta}=\Delta(\overline{p}, \overline{q}, \overline{r})$ is a triangle in the model space $M^2_k$ that satisfies $d(\overline{p}, \overline{q}) = d(p, q)$ , $d(\overline{q}, \overline{r}) = d(q, r)$ , and $d(\overline{r}, \overline{p}) = d(r, p)$ . A point $\overline{x}\in[\overline{p}, \overline{q}]$ for $x\in[p, q]$ is called a comparison point if $d(\overline{p}, \overline{x}) = d(p, x)$ . Note, that for $k>0$ we need an additional condition on the perimeter of a triangle to guarantee existence of a comparison triangle, for the purposes of this post we will state the condition, but not elaborate on it in detail.

CAT(k)

So what’s a cat? It’s a gorgeous animal that many humans have as a pet. Many cats are fluffy, and so is mine. Also cats are silly and generally cool to have around. However, this post is not about these kinds of cats.

The term “CAT(k)” space was coined by Mikhail Gromov and consists of three initials honoring: Élie Cartan, Alexander D. Alexandrov, and Victor A. Toponogov. Now, we will define a CAT(k) space. Let $X$ be a metric space and let $k$ be a real number. Let $\Delta$ be a geodesic triangle in $X$ with perimeter less than $2D_k$ (the diameter of the $M^2_k$ space, this is the condition need to guarantee the existence of comparison triangle). Then we say that $Delta$ satisfies the CAT(k) inequality if for the corresponding comparison triangle $\overline{\Delta}$ and all $x, y\in\Delta$ with corresponding comparison points $\overline{x}, \overline{y}\in\overline{\Delta}$ the inequality $d(x, y)\leq d(\overline{x}, overline{y})$ is satisfied. Thus, for a $k\leq 0$ we will call $X$ a CAT(k) space if $X$ is geodesic and all its geodesic triangles satisfy the CAT(k) inequality. For $k>0$ , we relax definition by only requiring $X$ to be $D_k$ -geodesic and only requiring the inequality condition for triangles of perimeter bounded by $2D_k$ .

Finally, we will call a space to be of curvature at most $k$ if it is locally a CAT(k) space. Hence, any space that is (locally) CAT(0) is a space on non-positive curvature (in the Alexandrov sense).

Intuitively speaking the triangles in a CAT(k) space have to be thinner than the ones in the corresponding model space. In particular, triangles in CAT(0) spaces are thinner than those in the regular Euclidean space (namely given two points on the sides of a triangle the distance between them in CAT(0) space is less than or equal to the distance between the corresponding comparison points in the Euclidean space). The illustration below shows three model spaces (sphere, plane, and hyperbolical paraboloid) and a cat drawn in each space. You can see that sphere cat is chubbier than the flat cat, which is chubbier than the hyperbolic cat, turns out triangles in these spaces are also of different chubbiness.

The cat icon used here is taken from Flaticon (https://www.flaticon.com/free-icon/black-cat_2179088#) provided by Victoruler.

For the rest of the post we will focus specifically on the case of the CAT(0) spaces, as they present the most interest to us in the follow up posts.

An immediate, but important consequence of the definition of a CAT(k) space is that any CAT(0) space is uniquely geodesic. Note, that being uniquely geodesic also implies that a space is contractible (the converse does not hold, there are contractible spaces that are not uniquely geodesic, a simple example is the closed upper hemisphere which is clearly contractible, and also not uniquely geodesic for the same reason a sphere isn’t).

Next, we will briefly describe some interesting spaces which under certain conditions turn out to be CAT(0).

My CAT is a cubical complex!

Let $I=[0, 1]$ be the unit interval, then we call the n-fold product $I^n$ the unit cube. We will let $I^0$ denote a point by convention. Since we will be mainly operating in $n\geq 3$ dimensions, we will use the term “face” to describe any-dimensional face of the cube. Thus, the faces of $I=[0, 1]$ are $\{0\}, \{1\}$ (the 0-dimensional faces), and $[0, 1]$ (the 1-dimensional face). Analogously, for $I^n$ a face is a subset $S\subseteq I^n$ which can be written as $\prod_{i=1}^n S_i$ where each $S_i$ is a face of $I$ . The dimension of the face will be the sum of dimensions of $S_i$ , and we also note that a k-dimensional face will be isometric to $I^k$ .

Now, we are ready to define a cubical complex (Def. 7.32 from Bridson and Haefliger, 2013) in a similar vein to the definition of a simplicial complex (for those familiar with the latter construction). A cubical complex $K$ is the quotient of a disjoint union of cubes $X=\amalg_\Lambda I^{n_\lambda}$ by an equivalence relation $\sim$ with the restrictions $p_{\lambda}: I^{n_\lambda}\to K$ of the natural projection $p:X\to K$ satisfying

for every $\lambda\in\Lambda$ the map $p_\lambda$ is injective;
if $p_{\lambda}(I^{n_\lambda})\cap p_{\lambda^\prime}(I^{n_{\lambda^\prime}})\neq\emptyset$ then there is an isometry $h_{\lambda, \lambda^\prime}$ from a face $T_\lambda\subseteq I^{n_\lambda}$ onto a face $T_{\lambda^\prime}\subseteq I^{n_{\lambda^\prime}}$ such that $p_\lambda (x)=p_{\lambda^\prime}(x^\prime)\iff x^\prime=h_{\lambda, {\lambda^\prime}}(x)$ .

While not precisely a cubical complex, the above rendition of a cat is reminiscent of one. The picture was taken from: https://www.boredpanda.com/i-made-animal-cube. All credit for creating those goes to the original author(s), which as far I found is: Aditya Aryanto.

Note, that in general a cubical complex needs not be a CAT(0) space. However, the following theorem of Gromov, 1987 gives the necessary and sufficient conditions for a cubical complex $K$ to be CAT(0).

A bit of perspective

In many cases the underlying geometry of the solution space can make or break our ability to solve problems efficiently. For example, many interesting discrete problems are NP-hard, and often do not even have good approximations (for a deeper dive Google “inapproximability results”, Khot, 2010 gives a great overview). In general, computing geodesics is a hard problem, but in certain spaces we can still efficiently compute (meaning in polynomial time) or approximate geodesics. CAT(0) cubical complexes are one particular class of “nice” spaces, meaning that geodesics can be computed or approximated well enough in polynomial time (Owen, and Provan, 2010, Hayashi, 2021). Implications of these algorithmic results mean that in certain phylogenetic (and robotics) problems we are able to efficiently compute distance between two points in the space, and reconstruct the optimal path between them (or at least do so up to $\varepsilon$ error).

Code and data availability

All code used in this post is available in the following Jupyter notebook.

Phylogenetic-vignettes-CAT-spaces.ipynb_Download

What’s next

The next post in the series will take us back to the world of phylogeny, but this time armed with the knowledge about CAT(0) spaces. We will try to make sense of defining nice metrics on the space of phylogenetic trees with branch lengths, and learn a thing or two in the process. Essentially, the idea for the remainder of this stretch of the series will be to work our way through the paper of Gavryushkin and Drummond, 2016 on the space of ultrametric phylogenetic trees.

References

Bridson, Martin R., and André Haefliger. Metric spaces of non-positive curvature. Vol. 319. Springer Science & Business Media, 2013. https://doi.org/10.1007/978-3-662-12494-9

Gavryushkin, Alex, and Alexei J. Drummond. “The space of ultrametric phylogenetic trees.” Journal of theoretical biology 403 (2016): 197-208. arXiv:1410.3544

Gromov, Mikhael. “Hyperbolic groups.” In Essays in group theory, pp. 75-263. Springer, New York, NY, 1987. https://doi.org/10.1007/978-1-4613-9586-7_3

Hayashi, Koyo. “A polynomial time algorithm to compute geodesics in CAT (0) cubical complexes.” Discrete & Computational Geometry 65, no. 3 (2021): 636-654. arXiv:1710.09932

Khot, Subhash. “Inapproximability of NP-complete problems, discrete Fourier analysis, and geometry.” In Proceedings of the International Congress of Mathematicians 2010 (ICM 2010) (In 4 Volumes) Vol. I: Plenary Lectures and Ceremonies Vols. II–IV: Invited Lectures, pp. 2676-2697. 2010. https://cs.nyu.edu/~khot/papers/icm-khot.pdf

Nicolae's Homepage

Tag: metric geometry

Phylogenetic Vignettes: Owen-Provan algorithm

Important announcement

Recap

Goal

CAT(0) properties

Path space geodesics

Algorithm overview

Data/code availability

What’s next

References

Phylogenetic Vignettes: 𝛕-space

Motivation

A quick aside

𝛕-space

Properties

Efficiently computing CAT(0) geodesics

Data and code availability

What’s next

References

Phylogenetic Vignettes: CATs you can’t pet

Introduction

Geodesics

Triangles

CAT(k)

My CAT is a cubical complex!

A bit of perspective

Code and data availability

What’s next

References