SciencePediaWhat is the "shape" of an object? While intuitively simple, this question poses a significant challenge for scientists seeking to compare biological forms objectively. A fossil jaw in one museum and its counterpart in another may differ in size, position, and orientation, but their fundamental shapes hold the key to understanding evolutionary change. The problem lies in developing a rigorous method to isolate these essential geometric properties from what are considered "nuisance" variables. Kendall's shape space provides the elegant mathematical solution to this very problem.
This article explores the theory and application of Kendall's shape space, a cornerstone of the field known as geometric morphometrics. It bridges the gap between the abstract world of high-dimensional geometry and the tangible questions of biological science. Across the following chapters, you will discover the complete process of translating physical forms into analyzable data. The first chapter, "Principles and Mechanisms," will unpack the mathematical foundation, explaining how information is systematically discarded to define shape, how objects are aligned using Procrustes analysis, and how the resulting "shape space" is constructed and navigated. Following this, the "Applications and Interdisciplinary Connections" chapter will demonstrate how this powerful framework is used to answer real-world questions in evolution, genetics, and paleontology, transforming abstract points in space into profound insights about the living world.
What is "shape"? The question seems almost childishly simple, yet the journey to a precise answer takes us to the heart of modern geometry and statistics. Think about a photograph of your hand. If you move closer to it, the image of your hand gets bigger. If you rotate the photo, its orientation changes. If you slide it across the table, its location is different. In all these cases, something has changed, but the essential "hand-ness"—its shape—has not.
Science, at its best, is about identifying what is fundamental by learning what to ignore. In the study of shape, we are interested in the geometry of an object that is invariant under these everyday transformations: location, size, and orientation. To study shape is to master the art of purposefully throwing away information.
Let's make this concrete. Imagine a biologist studying the evolution of a fish jaw. They can't compare two fossil jaws if one is in a drawer in London and the other in a museum in Tokyo. They can't meaningfully compare them if one is photographed from twice as far away as the other, or if one is simply rotated 90 degrees. To compare their shapes, we must first strip away these "nuisance" variables.
The modern approach, called geometric morphometrics, does this using a set of corresponding points, or homologous landmarks. For the fish jaw, these might be the tip of a particular tooth, the hinge point, and so on. We represent an object not as a continuous outline, but as a collection of these landmark coordinates, a configuration of points in space. Let's say we have landmarks in -dimensional space (for a 2D image, ; for a 3D scan, ). Our object is now a list of numbers.
Now, the "throwing away" becomes a precise mathematical procedure:
Removing Location (Translation): This is the easiest step. We find the geometric center of all the landmarks—their average position, called the centroid—and we shift the whole object so that its centroid sits at the origin . Now, every object we study is centered in the same place. This standardization removes pieces of information (the original coordinates of the centroid).
Removing Size (Scale): How do we define the "size" of a cloud of points? A natural way is to measure their overall spread around the center. We can calculate the centroid size, which is the square root of the sum of squared distances of each landmark from the centroid. We then simply scale the entire object—shrinking or expanding it—until this centroid size is equal to 1. Now, every object has the same, standard size. This removes one more piece of information: the original scale.
After these two steps, we have what is called a preshape. It's a version of our object that has been standardized for location and size. But the most interesting part of the puzzle remains: how do we handle orientation?
There's a figure in Greek mythology named Procrustes, an infamous innkeeper who had an iron bed. He would invite travelers to stay the night and then force them to fit the bed: if they were too short, he would stretch them on a rack; if they were too tall, he would chop off their legs. It's a gruesome tale, but it gives its name to the elegant mathematical procedure for standardizing orientation: Procrustes superimposition.
Unlike the mythical innkeeper, our goal isn't to distort the object. Instead, given two preshapes (say, a reference jaw and a target jaw ), we want to find the best possible alignment by rotating without changing its shape. What does "best" mean? It means rotating until the sum of the squared distances between its landmarks and the corresponding landmarks of is as small as possible. We are looking for the orientation that minimizes the mismatch.
Imagine you have two transparent star charts. You've already aligned their centers and scaled them to be the same size. Now, you lay one on top of the other and rotate it until the patterns of stars match up as closely as possible. The slight differences that remain—where a star on one chart doesn't quite line up with its counterpart on the other—represent the true difference in their "shape."
This optimization problem, finding the one rotation that creates the best fit, has a unique and beautiful mathematical solution. It uses a powerful tool from linear algebra called Singular Value Decomposition (SVD) to instantly find the optimal angle of rotation. The result of this process is that both objects are now in a common coordinate system, optimally aligned to one another.
Once we have performed this optimal alignment, we can finally give a number to the "shape difference." The Procrustes distance is simply the remaining mismatch after we've done our absolute best to align two shapes. It's the square root of that minimized sum of squared distances.
This single number is incredibly powerful:
This distance is the fundamental metric of shape analysis. It allows us to quantify what our eyes perceive intuitively. But what kind of space are these distances being measured in?
This is where the story takes a turn from the practical to the profound. The collection of all possible shapes is not a simple, flat, "Euclidean" space. Instead, it forms a curved, high-dimensional manifold known as Kendall's shape space.
Let's build this fascinating universe step-by-step. We start with our preshapes—objects already centered and scaled to unit size. The set of all possible preshapes forms the surface of a high-dimensional sphere, called the preshape sphere. Every point on this sphere's surface is a specific configuration of landmarks with a particular orientation.
But we decided we don't care about orientation. All rotated versions of a single preshape represent the same shape. On the preshape sphere, these equivalent configurations form a path, or an "orbit," as you apply all possible rotations. To get to shape space, we must consider this entire orbit to be a single point. This mathematical process of "gluing" together equivalent points is called taking a quotient. Kendall's shape space is the quotient of the preshape sphere by the group of rotations.
Think of a simpler analogy. Imagine the globe. The lines of longitude all converge at the North and South poles. What if we decided that we don't care about longitude, only latitude? We could conceptually "squish" every line of longitude down into a single point. The entire circle of the equator would become one point. The arctic circle would become another. The result of this quotient would be a simple line segment running from the North Pole to the South Pole. We have created a new, one-dimensional "latitude space" from the two-dimensional surface of a sphere.
Shape space is a far more complex version of this idea. But the result is a space where each point is a unique shape. And the Procrustes distance we discussed earlier is the geodesic distance in this space—the shortest path between two points along its curved surface.
The dimensionality of this space, the number of independent variables needed to define a shape, can be calculated by simple counting. We start with variables, and we subtract the number of degrees of freedom we threw away: for translation, 1 for scale, and for rotation. The dimension of shape space is thus .
This abstract construction leads to some astonishingly beautiful results. For the simplest possible shape—a triangle in a 2D plane ()—the vast, complicated machinery of shape theory simplifies miraculously. The shape space for triangles turns out to be the surface of a sphere! Every possible triangle shape, from long and skinny to equilateral, corresponds to a unique point on this sphere. The "distance" between the shape of a perfect equilateral triangle and a right-angled isosceles triangle is simply an angle on this sphere: a mere 15 degrees, or radians.
As elegant as curved spaces are, they are difficult places to do statistics. How do you calculate an "average" on the surface of a sphere? Or a line of best fit? For this, we borrow a brilliant trick from mapmakers and physicists: we make a flat map. We approximate the curved shape space with a flat tangent space.
Imagine you're in the middle of a large field. For all practical purposes, the ground looks flat. You don't need to worry about the curvature of the Earth to walk from one side to the other. This "flat field" is your tangent plane to the spherical Earth.
In shape analysis, we do the same thing. For a population of, say, 100 fish jaws, we first calculate the Procrustes mean shape. This is the central point of our data cloud in the curved shape space. Then, we "lay a flat sheet of paper" on the space, just touching it at that mean shape. This sheet is our tangent space. We then project every one of our 100 jaw shapes from the curved space onto this flat sheet.
This process, called linearization, transforms our shape data into a set of points in a standard, flat Euclidean vector space. Now, all the familiar tools of statistics are at our disposal. We can perform a Principal Component Analysis (PCA) to find the major axes of shape variation, run regressions to see how shape changes with size (allometry), and perform tests to compare groups.
You might worry that this approximation introduces errors. It does, but they are remarkably small. The "flat-map distance" in the tangent space is an excellent approximation of the true geodesic Procrustes distance. For shapes that are close to the mean, the error in the distance is of the order of the distance cubed ()—an exceptionally good approximation.
This explains the "morphospace" plots you so often see in papers on evolution and development. They are typically 2D views (PC1 vs. PC2) of this high-dimensional flat tangent space. Each point is a specimen, and the distance between points on the plot represents, to a good approximation, the true shape difference between them. This also comes with a warning: since PCA plots are often a lower-dimensional "shadow" of the full tangent space, the distances you see on the plot are an underestimate of the full tangent-space distance, which is itself a slight underestimate of the true Procrustes distance. The relationship is always: .
The Procrustes superimposition is an incredibly powerful tool, but it leaves behind a subtle signature, a mathematical ghost. By forcing every object to conform to a set of global constraints—centroid at the origin, unit size, common orientation—we inadvertently create statistical dependencies among the landmarks.
Think of it this way. Imagine you have a set of independent landmarks, each with some random biological "jitter." Now, you apply the centering rule: the sum of all their position vectors must be zero. If you move one landmark a little to the right, the others must collectively shift a little to the left to keep the centroid at the origin. Their movements are no longer independent. The Procrustes alignment acts like a set of invisible strings, linking the landmarks together.
This means that even if the underlying biological variation were completely random and independent for each landmark, the coordinates after Procrustes alignment will show weak, artifactual correlations. Does this invalidate our analyses? Fortunately, no. Morphometricians are aware of this ghost in the machine. They account for it by using clever statistical methods. For example, when testing a hypothesis about shape integration, they can run simulations where they generate random shapes, apply the exact same Procrustes procedure, and see what level of correlation is produced simply as a mathematical artifact. They then compare their real biological result to this artifactual baseline.
This final point encapsulates the spirit of the field. The journey into shape space begins with a simple, intuitive question, leads us through elegant geometry and powerful algorithms, and ends with the statistical sophistication needed to distinguish true biological signal from the subtle echoes of our own mathematical tools. It is a perfect microcosm of the scientific process itself: a constant dialogue between the world we observe and the language we invent to describe it.
So, we have journeyed into a strange and beautiful new land: a mathematical sphere where every point is not a place, but a shape. We have learned how to navigate this space, to measure distances, and to understand its local geometry. But what is the point of all this abstraction? Why go to the trouble of building this "Kendall's shape space"? The answer, and this is where the real fun begins, is that this peculiar globe is not just a mathematical curiosity. It is a powerful lens through which we can ask—and answer—some of the most profound questions about the living world. By translating the messy, complex reality of biological form into the clean, rigorous language of geometry, we unlock a new kind of understanding.
Let us start with one of the most classic problems in human evolution: how different is a Neanderthal skull from our own? You can hold two skulls, one Neanderthal and one modern human, and see they are different. But how different? In what specific ways? Is the brow ridge the main story, or is it the chin, or the overall roundness of the cranium? To answer this, we must first agree on a fair basis for comparison. The skulls might be of different sizes, and they certainly won't be sitting in the same orientation on the museum shelf.
This is where the first, most fundamental application comes into play: Procrustes superimposition. Think of it as a sophisticated way of saying, "Let's be fair." We take our landmark data from each skull and, mathematically, we perform three simple-sounding operations: we slide them so their centers of mass align, we shrink or expand one so they are the same overall size, and finally, we rotate one to match the other as closely as possible. The criterion for "closest possible match" is wonderfully simple: we minimize the sum of the squared distances between all corresponding landmarks. What remains after this democratic alignment process—the residual mismatch between the landmarks—is the pure, unadulterated shape difference. We have filtered out the "noise" of size, position, and orientation, leaving only the signal of shape. This single procedure is the gateway to all that follows; it is the act of placing our two specimens onto the shape globe to see where they lie.
Once we have aligned not just two, but a whole collection of shapes—say, hundreds of leaves from different plant genotypes—we find ourselves with a cloud of points on our shape globe. What is the structure of this cloud? This is no longer a question of simple comparison but of variation. Here, we borrow a classic tool from statistics, Principal Component Analysis (PCA), but we apply it in the flat "tangent space" that locally approximates our curved globe.
Imagine the cloud of leaf shapes. PCA is like finding a new set of coordinate axes for this cloud. But they aren't just any axes. The first axis, PC1, is the direction of the greatest variation in the cloud—the "main knob" of shape change. The second axis, PC2, is the direction of the next greatest variation, and it's perpendicular to the first. What does this mean in practice? For our leaves, we might discover that PC1 corresponds to a "complexity" axis, with simple, smooth-edged leaves at one end and deeply lobed, intricate leaves at the other. PC2 might correspond to an "aspect ratio" axis, ranging from short, wide leaves to long, narrow ones.
The true magic happens when we connect this to other fields, like developmental genetics. Suppose we have leaves from a wild-type plant, a plant engineered to overexpress a gene known to promote complexity (like KNOX), and another with a gene promoting simple blades (like ARP). When we plot our leaves in the morphospace defined by PC1 and PC2, we might see a beautiful separation. The KNOX leaves will cluster at the "high complexity" end of PC1, the ARP leaves at the "low complexity" end, and the wild-type in the middle. Suddenly, our abstract geometric axis, PC1, has a tangible biological meaning: it reflects the antagonistic action of these developmental genes! We have forged a direct link from gene to geometry.
This leads us to the next level of inquiry: hypothesis testing. Seeing that two clouds of points for two species of lizards look separate in our morphospace is one thing. Proving it statistically is another. This is the domain of Procrustes ANOVA. Just like a standard ANOVA partitions variance in a single variable (like height), Procrustes ANOVA partitions the total shape variance into components. How much of the variation in lizard head shape is due to differences between species? How much is due to differences between males and females? And how much is due to the interaction between the two? By analyzing the distances between points in our tangent space, we can assign statistical significance to these effects, turning a visual impression into a rigorous conclusion.
But what if our goal is not just to see if groups differ, but to find the single best way to tell them apart? PCA finds the axes of greatest variation overall, but that variation might be common to all groups. We need a tool that specifically seeks out what makes the groups distinct. Enter Canonical Variates Analysis (CVA). CVA is designed to find the linear axes in shape space that maximize the ratio of between-group variance to within-group variance. It's like a finely tuned filter that ignores the shape variation common to all species and highlights the precise features that best discriminate them. The first canonical variate might represent a shape change that perfectly separates species A from B and C, even if that change isn't the single largest source of variation in the entire dataset. It is the ultimate tool for characterization and diagnosis.
Of course, shape does not exist in a biological vacuum. It is intimately tied to other attributes, most notably size. As an animal grows, its shape changes. A baby's head is not just a scaled-down adult head. This relationship between shape and size is called allometry. With our geometric framework, we can study it with stunning precision. We can perform a multivariate regression, modeling the vector of shape coordinates as a function of the logarithm of size. The result is a "vector" in shape space, , that represents the allometric trajectory—the specific path of shape change that accompanies growth.
We can extend this idea from static, within-population allometry to the dynamic changes that occur through an organism's entire life, its ontogeny. Imagine tracking the shape of a jaw and the shape of the braincase in an animal from birth to adulthood. We can compute the ontogenetic trajectory for each part as a path in its own shape space. Are these paths parallel? Do they point in the same direction? The angle between these trajectory vectors becomes a precise, quantitative measure of their developmental coupling. We can then ask if different parts of an organism (modules) follow their own developmental paths or if they are tightly integrated. We can even test whether the strength of their connection—their integration, measured by a method like Partial Least Squares (PLS)—changes as the animal grows.
Finally, let us zoom out from the scale of an individual lifetime to the grand tapestry of evolutionary history. A paleontologist unearths fossils from a lineage over millions of years. A key question is: how has the diversity of form changed over time? This is the concept of morphological disparity. In our framework, disparity is simply the measure of how "spread out" the cloud of shape-points is in morphospace for a given time period. Did a key innovation, like the evolution of wings, lead to an "explosion" of new shapes? We would see this as a rapid increase in the volume occupied by the shape-cloud. And our geometric tools give us robust ways to measure this volume, for instance, by summing the variances along the principal axes. This is far more reliable than just looking at the range of shapes, which can be misleading if the fossil record is sparse. We can directly visualize and quantify the great radiations and contractions in the history of life, written in the geometry of shape.
From comparing two skulls to mapping the effects of genes, testing evolutionary hypotheses, and charting the course of life over eons, the applications are as vast as biology itself. The abstract space of shapes, once a purely mathematical construct, becomes a laboratory for the natural historian. It reveals the underlying unity of form and process, demonstrating with breathtaking clarity that, as the great D'Arcy Thompson suspected over a century ago, the study of biological form is, in the end, a science of geometry.