Simplex Geometry: The Hidden Rules of Compositional Data

In many scientific disciplines, from biology to economics, we are less interested in absolute quantities and more in the relative proportions of components that make up a whole. This type of information, known as compositional data, is everywhere, but it comes with a hidden catch: because all parts must sum to a constant, they are not independent. This fundamental constraint traps the data in a unique geometric space called a simplex, where our intuitive, everyday statistical tools based on straight lines and flat planes begin to break down, leading to paradoxes and incorrect conclusions.

This article addresses the critical knowledge gap between how we often analyze proportional data and how we should. It serves as a guide to a new geometric perspective that resolves these paradoxes and provides a more powerful and truthful way to understand the world. Across the following chapters, you will learn the core principles of this approach and witness its transformative impact. First, the "Principles and Mechanisms" section will explain why our intuition fails with compositional data and introduce the elegant solution of Aitchison geometry and log-ratios. Following that, "Applications and Interdisciplinary Connections" will demonstrate how this single geometric idea unifies our understanding of seemingly disparate fields, from the microbes in our gut to the design of advanced materials.

In science, as in life, we are often interested in not just one thing, but the makeup of a whole system. Think of a baker's recipe: what matters is not the absolute amount of flour, but its proportion relative to sugar, eggs, and butter. Or consider a nation's economy: we track the market share of different industries as parts of the total economic output. This kind of data, representing parts of a whole, is called ​​compositional data​​.

Perhaps one of the most exciting frontiers for this way of thinking is in biology, particularly in the study of the ​​microbiota​​—the vast communities of microorganisms living in and on our bodies. When scientists sequence the gut microbiota, they don't get an absolute count of every single bacterium. The sequencing machine gives them a huge collection of genetic snippets, and from these, they can only determine the relative abundance of each species. You might find that 20% of the community is Bacteroides, 15% is Prevotella, and so on.

No matter how large or small the total microbial population is, these relative abundances must always add up to a constant, usually normalized to $1$ (or $100\%$). This is the fundamental rule of compositional data, the ​​closure constraint​​. This seemingly innocent constraint has profound and often counter-intuitive consequences. It means that the numbers are not free to vary independently. If the proportion of one component goes up, the proportion of at least one other component must go down. They are all tied together in a zero-sum game.

This constraint forces the data to live in a specific mathematical space called a ​​simplex​​. What is a simplex? It’s the simplest possible geometric shape in any given dimension. For a composition with three parts (say, bacteria A, B, and C), the possible proportions must lie on the surface of a triangle. For four parts, they lie on the surface of a tetrahedron, a four-faced pyramid. For the millions of components in a microbiome, the data lives on a multi-million-dimensional version of such a shape. This is the stage on which the drama of compositional data unfolds.

For centuries, we have developed powerful statistical tools—correlation, regression, analysis of variance—based on the principles of Euclidean geometry, the familiar world of flat planes and straight lines taught in high school. But the simplex is not a flat, open space. It is a constrained surface. And applying our usual tools in this new space is like trying to navigate the curved Earth with a flat map; you get distorted results and can end up hopelessly lost.

The first illusion in this hall of mirrors is ​​spurious correlation​​. Imagine a simple gut ecosystem where an increase in one bacterium has no biological effect on another. Yet, if the first bacterium blooms and its relative abundance increases from $10\%$ to $30\%$, the total "pie" is still only $100\%$. That extra $20\%$ has to come from somewhere. The relative abundances of other bacteria must decrease, creating a negative correlation in the data even where no real biological antagonism exists. We see an effect that isn't real; it's just a mathematical ghost created by the closure constraint.

The situation is even more treacherous, leading to a breakdown of logic known as ​​subcompositional incoherence​​. Let's see this with a startlingly simple example. Imagine we track the absolute abundances of three bacteria, $A$, $B$, and $C$, in three samples, and find that $A$ and $B$ are perfectly, positively correlated.

Sample	Absolute A	Absolute B	Absolute C
1	1	2	7
2	2	3	5
3	3	4	3

Now, let's do what a sequencing machine does: convert these to relative abundances. The totals for each sample happen to be $10$, so the proportions are:

Sample	Relative A	Relative B	Relative C
1	0.1	0.2	0.7
2	0.2	0.3	0.5
3	0.3	0.4	0.3

If we calculate the Pearson correlation between the relative abundances of $A$ and $B$, it's still a perfect $+1$. So far, so good. But what if we were only interested in the relationship between $A$ and $B$, and decided to ignore $C$? We would take their absolute abundances, form a subcomposition, and re-normalize them to sum to $1$.

Sample	Absolute A	Absolute B	Sub-Total	Rel. A'	Rel. B'
1	1	2	3	$1/3$	$2/3$
2	2	3	5	$2/5$	$3/5$
3	3	4	7	$3/7$	$4/7$

Now, if we calculate the correlation between these new relative abundances, Rel. A' and Rel. B', we get a perfect $-1$. The relationship completely inverted!. This is absurd. The true relationship between $A$ and $B$ cannot possibly depend on whether or not we are paying attention to $C$. This demonstrates that standard correlation is fundamentally broken for compositional data. Our statistical ruler gives a different measurement every time we change which parts of the system we look at.

This brings us to the core problem: we are using the wrong ruler. The standard ​​Euclidean distance​​, $\sqrt{\sum (x_i - y_i)^2}$, measures the length of a straight line between two points. This is fine in an open, flat space. But on the constrained surface of the simplex, this "straight line" goes right through the middle of the shape, outside the space where our data can even exist. The change from a proportion of $0.1$ to $0.2$ (a $100\%$ increase) is treated the same as a change from $0.8$ to $0.9$ (a $12.5\%$ increase). Our ruler is blind to the relative nature of the data.

The solution to this puzzle came from a Scottish mathematician named John Aitchison in the 1980s. He proposed a radical and beautiful idea: if the data's structure is the problem, let's change our perspective. He realized that the fundamental, stable information in a composition is not in the values of the parts themselves, but in their ​​ratios​​.

Why ratios? Because ratios are ​​scale-invariant​​. If you have a sample with $10$ units of microbe A and $20$ units of microbe B, the ratio is $10/20 = 1/2$. If, due to better experimental methods, you get double the total material and now measure $20$ units of A and $40$ of B, the absolute amounts have changed, the proportions might have changed (depending on what happened to other microbes), but the ratio $20/40$ is still $1/2$. The ratio captures the intrinsic relationship, independent of the arbitrary total size.

Aitchison's genius was to build an entire geometry, now called ​​Aitchison geometry​​, based on these ratios. To do this, he used a classic mathematical tool: the logarithm. Logarithms have a wonderful property—they transform multiplication and division into addition and subtraction. By taking the logarithm of the ratios, we can take the multiplicative, constrained world of the simplex and "unfold" it into a standard, additive, unconstrained Euclidean space.

One of the most important ways to do this is the ​​Centered Log-Ratio (CLR) transformation​​. The idea is to find a "center" for the composition and express everything relative to that center. The natural center for a set of positive numbers is their ​​geometric mean​​, $g(\mathbf{x}) = (x_1 \times x_2 \times \dots \times x_D)^{1/D}$. The CLR transformation for each part $x_i$ is then simply:

For example, for a simple three-part composition like $[0.2, 0.3, 0.5]$, the geometric mean is $(0.2 \times 0.3 \times 0.5)^{1/3} \approx 0.3107$. The CLR coordinates would then be $[\ln(0.2/0.3107), \ln(0.3/0.3107), \ln(0.5/0.3107)]$, which calculates to approximately $[-0.441, -0.035, 0.476]$. We have transformed the three constrained numbers that must sum to $1$ into three unconstrained numbers that now sum to $0$. These new coordinates live in a standard Euclidean space where our familiar statistical tools can finally be used correctly.

Other, more sophisticated transformations like the ​​Isometric Log-Ratio (ILR) transformation​​ also exist. They offer further advantages, such as providing a set of coordinates that can be constructed to guarantee the property of subcompositional coherence, which is crucial for building reliable predictive models. The key insight remains the same: analyze log-ratios, not raw proportions.

Now that we have a map from the simplex to a familiar Euclidean space, we can finally define a proper ruler. The ​​Aitchison distance​​ between two compositions, $\mathbf{x}$ and $\mathbf{y}$, is simply the standard Euclidean distance between their CLR-transformed coordinates:

This distance measures the "relative difference" between two compositions in a way that is coherent and meaningful. Let's revisit the two microbiome profiles from an earlier thought experiment: $\mathbf{x} = (0.5, 0.2, 0.2, 0.1)$ and $\mathbf{y} = (0.4, 0.3, 0.2, 0.1)$. The naive Euclidean distance between them is a mere $0.141$. However, the true Aitchison distance, calculated in the space of log-ratios, is approximately $0.454$—more than three times larger!. The naive ruler vastly underestimated the true extent of the relative change between the two communities because it was blind to the doubling of the ratio of part 2 to part 1 in sample $\mathbf{y}$ compared to sample $\mathbf{x}$. The Aitchison distance captures this essential information.

What began as a statistical puzzle has blossomed into a universal language for understanding systems made of parts. This is not just a tool for microbiologists studying dysbiosis in the gut. Geologists use it to analyze the elemental composition of rocks. Ecologists use it to study the species composition of forests. Economists use it to model market shares.

The principles of simplex geometry are even revolutionizing materials science. In the design of ​​High-Entropy Alloys​​—complex metals made from nearly equal proportions of five or more elements—the precise compositional balance is key to creating materials with extraordinary properties. Aitchison geometry provides the right framework to measure the distance between different alloy compositions and build machine learning models that predict their strength, corrosion resistance, or other features.

The connections run even deeper, into the heart of computer science and artificial intelligence. In ​​online optimization​​, algorithms that learn from a stream of data must make decisions within a constrained set. The performance of these algorithms fundamentally depends on the geometry of that set. An algorithm designed for the open space of a Euclidean ball behaves very differently from one designed for the constrained space of a simplex. Understanding the geometry is essential for designing efficient learning algorithms.

This is the beauty of a profound scientific idea. The same principles that help us understand the balance of our inner microbial world can guide the creation of futuristic materials and the design of intelligent machines. By learning to see the world not as a collection of absolute amounts but as a symphony of relative parts, we gain a deeper and more unified understanding of its intricate structure.

We have spent some time exploring the peculiar world of the simplex, a geometric space where the parts of a whole must sum to a constant. We've learned its strange rules of addition and distance, based not on absolute amounts but on relative ratios. At first glance, this might seem like a mathematical curiosity, a detour from the familiar Euclidean space of our everyday intuition. But what if I told you that this is not a detour at all? What if this is the native language of a vast array of natural phenomena?

The moment you start looking for things made of parts—a diet made of nutrients, an ecosystem of species, an alloy of metals, a budget of expenditures—you begin to see simplexes everywhere. The principles we've uncovered are not just abstract rules; they are the key to unlocking a deeper understanding of the world. Let us now take a journey through some of these seemingly disconnected fields and witness how the single, beautiful idea of simplex geometry provides a unifying thread.

Nowhere has the impact of this geometric perspective been more profound than in the life sciences. For decades, biologists and medical researchers have been collecting data on the "parts" of a system—genes, proteins, metabolites, cells—often expressed as proportions or percentages. Yet, we have often analyzed this data as if it lived in a flat, Euclidean world, a mistake akin to using a flat map of the Earth to plan an intercontinental flight. Recognizing the data's true home on the simplex has sparked a revolution.

Consider the bustling metropolis of microbes in your gut. Your microbiome is a composition of hundreds of bacterial species, and its balance is intimately linked to your health. Suppose we want to know if a new diet affects the microbiome. The old way might be to ask, "Did the amount of Lactobacillus increase?" But this is a misleading question. An increase in Lactobacillus must be accompanied by a decrease in something else. The simplex geometry teaches us to ask a better, more meaningful question: "Did the balance between different groups of bacteria change?" For instance, we can now precisely formulate and test a hypothesis about the ratio of fiber-digesting bacteria to protein-digesting bacteria, a concept captured elegantly by an isometric log-ratio (ILR) coordinate. This is not just a statistical correction; it is a fundamental shift in our scientific questioning, from "how much?" to "in relation to what?"

This new perspective extends to how we compare individuals. Imagine a doctor wanting to know if your microbiome profile resembles that of a healthy person or someone with a particular disease. A simple ruler—a Euclidean distance—would measure the absolute differences in percentages, a flawed metric. The proper tool is the Aitchison distance, which measures the "distance" between the internal log-ratios of two compositions. It tells us how different the two microbial ecosystems are in their structure. Using this distance, we can cluster patients into clinically relevant groups, perhaps identifying subgroups of a disease or predicting treatment response, all based on the geometry of their cellular makeup.

The story culminates in our ability to trace complex causal pathways. We know diet affects our health, but how? Much of the effect might be mediated through the gut microbiome. We can now build a statistical model that follows the causal chain: a dietary change (say, from a high-fat to a low-fat diet) first alters the composition of the microbiome. This change in composition, properly represented using log-ratio coordinates, then causes a change in a host phenotype, like the level of a key metabolite in the blood. By respecting the simplex geometry at the crucial intermediate step, we can untangle these effects and quantify exactly how much of a diet's benefit comes from its influence on our microbial partners. This allows us to investigate deep questions, such as whether a dietary intervention can produce a "phenocopy"—an environmentally induced trait that mimics a known genetic one. This level of sophistication is simply impossible without the right geometric tools. Similarly, when studying how the composition of dietary fats—saturated, monounsaturated, polyunsaturated—affects inflammatory markers, log-ratio analysis allows us to interpret our findings correctly, not as the effect of adding one fat, but as the health consequence of substituting one type for another.

Let's zoom out from the microscopic world within us to the world seen from above. When a satellite orbiting Earth takes a picture of a forest or a coastline, each pixel in its image is rarely a single, pure substance. A pixel over a landscape is a mixture of soil, water, rock, and vegetation. The spectrum of light reflected from that pixel is a combination of the pure spectra of its components, called "endmembers."

In the simplest model, the measured spectrum is a point lying inside a high-dimensional simplex whose vertices are the unknown spectra of the pure endmembers. The task of "spectral unmixing" is a fascinating geometric puzzle: given a cloud of thousands of data points (the pixels), find the vertices of the simplex that contains them! It’s like being given a thousand different paint swatches and having to deduce the primary colors used to create them. Algorithms like N-FINDR tackle this by searching for the set of data points that themselves form the simplex of the largest possible volume, based on the beautiful idea that the true endmembers must form the container for all other mixtures. Other methods, like SISAL, take a different approach, searching for the smallest possible simplex that can enclose the entire data cloud.

Of course, the real world is rarely so simple. What happens when a mountainside is cast in shadow? The mixture of rock and grass is the same, but the reflected light is dimmer. This uniform dimming acts as a multiplicative scaling factor. Geometrically, this has a dramatic effect: our neat, bounded simplex is stretched into an infinite, unbounded cone with its apex at the origin. All the points on a single ray of this cone correspond to the same material composition, just under different lighting. Understanding this geometric transformation is the key to solving it. By normalizing each pixel's spectrum—for example, by dividing it by its total brightness—we can project all the points on each ray back onto a single point on a plane. This act of normalization collapses the cone back into a simplex-like object, allowing our vertex-finding algorithms to work once more. This is a triumph of geometric reasoning: by understanding how our model was broken, we found the precise mathematical operation to fix it.

The influence of the simplex extends into the world of engineering and technology, shaping how we create new materials and how we build intelligent machines.

In materials science, researchers are designing "high-entropy alloys" by mixing five, six, or even more metallic elements in nearly equal proportions. The exact local composition at any point in the material is a point on a high-dimensional "Gibbs simplex." When this alloy is heated, the atoms jiggle and diffuse, and the local composition changes over time. This is a dynamic process—a flow on the surface of the simplex. The total flux of atoms must sum to zero; an atom cannot appear from nowhere. This physical constraint means that the "velocity" of any compositional change must be tangent to the simplex. The famous Maxwell-Stefan equations for multicomponent diffusion are, when viewed through a geometric lens, a set of laws for constrained motion on a manifold. Physicists and engineers use mathematical tools called projection operators to ensure their simulations of this process respect the geometry, keeping the dynamics confined to the simplex, just as a train is confined to its tracks.

Finally, as we build increasingly complex machine learning models, we face the challenge of making them understandable. If a model predicts a patient's risk based on the composition of their blood cells, how can we trust it? Methods like SHAP (Shapley Additive Explanations) aim to explain a prediction by assigning an importance score to each input feature. But a naive application to compositional data can be misleading, as it fails to recognize the sum-to-one constraint. If one cell type's percentage goes up, its SHAP value might increase, but this ignores the fact that other cell types must have gone down. A truly "compositional" explanation method must be built on the logic of log-ratios. By designing explainability tools that operate in the natural geometry of the data, we can ensure that our interpretations are not just plausible, but principled. We can teach our machines not just to see the numbers, but to understand the ratios that give them meaning.

From the balance of life in our bodies to the mixture of minerals on distant planets, from the flow of atoms in an alloy to the logic of an artificial mind, the geometry of the simplex is a deep and unifying principle. It reveals a hidden layer of structure in our world. By learning its language, we have not only found better ways to solve existing problems, but we have also equipped ourselves to ask—and answer—questions we never before knew how to formulate. The simplex is not a prison of constraints, but a rich canvas for discovery.