Imagine a city. Not an abstract one, but a living one – with people meeting in cafes, arguing in marketplaces, whispering in doorways. If you want to understand how rumors or diseases spread, you could draw a diagram: dots for people, lines for acquaintances. This is a graph. A simple, beautiful, almost poetic map of connections.
But life is rarely pairwise. A rumor isn't born when one person tells another – it ignites in a group, at a table where three or five people hear the same thing at once and react to each other. A group is not just a sum of pairs. It's something more, and this 'more' changes everything.
It's precisely here that the mathematics of complex systems hits a wall. The equations describing such interactions become non-linear – that is, they behave unpredictably, not obeying the simple rules of addition. And for a long time, it seemed nothing could be done about it: non-linearity was the price of realism.
But what if non-linearity could be moved? Not eliminated, but transferred – from the rules of behavior into the very architecture of the network? This is the very question that scientists explored, and their findings form the basis of this text. And the answer turned out to be unexpectedly beautiful.
Graphs, Hypergraphs, and a Touch of Structural Magic
Let's start with the basics, because without them, things will get foggy ahead.
A standard graph is dots and lines. Dots are objects (people, neurons, molecules), and lines are the connections between them. Each line connects exactly two dots. It's convenient, it's familiar, it works – but only for pairwise interactions.
A hypergraph is like a graph that has 'grown up'. Here, a single link can connect three, five, or ten participants at once. Such a link is called a hyperedge. Imagine not a thread between two beads, but a web that captures a whole group at once. Hypergraphs emerged as a more honest way to reflect reality: after all, in chemistry, a reaction can involve several molecules at once; in ecology, several species; in sociology, entire communities.
Now for the key question: when we describe the behavior of such a system, non-linearity can 'live' in two different places.
The first place is in the rules of behavior. For example: 'a neuron activates if the product of signals from its neighbors exceeds a threshold.' The product – that's already non-linearity. The rule is complex, the equation is complex.
The second place is in the structure of the network. We can keep the rules simple (linear), but complicate the map of connections itself so that it 'carries' all the complexity of the interactions. Simply put, instead of tricky rules – a tricky architecture.
This is what's called delegating non-linearity to the structure. And this isn't just a philosophical exercise – it's a mathematically rigorous transformation that opens up entirely new tools for analysis.
Why Linearity Is So Valuable
I'll digress for a moment to explain why we strive for linearity at all.
A linear system is one where doubling the input signal doubles the output. Where two independent processes can simply be added together. Where the future can be predicted, the past reconstructed, and stability checked using the spectrum of a matrix. Over several centuries, mathematicians have built an entire palace of tools for linear systems: linear algebra, stability theory, optimal control, signal filtering.
Non-linear systems are a jungle. Beautiful, alive, but difficult to navigate. A small change in initial conditions can lead to catastrophically different outcomes. This is precisely why physicists, mathematicians, and engineers have always dreamed of a way to 'straighten out' non-linearity – to turn the jungle into, at least, a managed park.
If a non-linear dynamic can be rewritten as a linear one – even in a larger space – we gain access to that entire palace. This is what makes the described approach so appealing.
Multilinearity: A Non-linearity That Can Be Tamed
Not all non-linearity is the same. There is a special class: multilinear functions. These are functions that are non-linear as a whole, but linear in each argument separately. It sounds like a play on words, but in reality, it's an important property.
A simple example: the area of a rectangle is the product of its two sides. If you double one side, the area doubles. If you double the other, it doubles too. But if you double both, the area quadruples. This is multilinearity: linear in each argument individually, but non-linear in combination.
Precisely these kinds of interactions are often found in nature. Two neurons whose joint activity strengthens a third. Three molecules that form a complex only when present simultaneously. A group of agents making a collective decision.
And here is the key discovery, which researchers proved rigorously: a multilinear dynamic on a standard graph admits an exact, finite, linear representation on a hypergraph.
What does this mean in practice? Imagine you have a system where the state of each element depends on pairwise and triplet products of its neighbors' states. Instead of working with these products directly, you can introduce new 'variables' – one for each pair, one for each triplet – and write their equations of motion. These new variables will become the 'states of the hyperedges' of your hypergraph.
The trick is that when you expand the equations for these new variables, everything remains within the same finite set. The system 'closes on itself'. No infinite growth, no unforeseen terms – just a large, but strictly linear, matrix.
It's as if you were to describe planetary motion not by their positions and velocities, but by all possible pairwise and triplet angular momenta – and suddenly discovered that in this expanded description, the laws of motion become simpler.
What Happens to the State Space
Let's talk about the price of this transformation – because there is one, and it's not small.
When you introduce new variables for pairs, triplets, and higher-order combinations, the system's state space grows dramatically. If you initially have, say, a hundred elements, the number of pairs among them is almost five thousand, and the number of triplets is over one hundred fifty thousand. With each order, the size grows rapidly, like an avalanche.
If the original dynamic had a non-linearity of order M (i.e., involved products of at most M variables), then an exact linear representation will require a space including products up to order 2M − 1. For a quadratic non-linearity (M = 2), this means triplet products. For a cubic one (M = 3), fifth-order products. And so on.
This is not a problem of principle; it is a practical one. Mathematically, the transformation is exact. Computationally, it's resource-intensive. But this is where it's important to understand: we are talking about equivalence in principle, not an algorithm for tomorrow's supercomputer. The understanding that such a transformation exists already changes how we think about the nature of complexity.
There is another subtle point. For the linear dynamics in the expanded space to accurately reproduce the original non-linear dynamics, the initial conditions for the 'hyperedge states' must be consistent with the initial conditions for the vertices. That is, if a hyperedge variable describes the product of two vertices, at the initial moment it must be exactly equal to that product – no more, no less. If this condition is violated, the two descriptions diverge, like paths at a fork in the road.
When Multilinearity Isn't Enough: The Carleman Method
But not everything in nature is multilinear. Many real systems are described by far more capricious functions – sine, exponent, logarithm. Recall the model for the synchronization of pendulums or neural oscillators: it features the sine of the phase difference. This is not multilinearity; it's something more general.
Analytic functions – those that can be expanded into an infinite series in powers of variables – can be represented as an infinite sum of multilinear terms. The sine is a difference of odd powers. The exponent is a sum of all powers. If we consider each power separately, we once again find ourselves in the world of variable products.
This is precisely what the method proposed by mathematician Torsten Carleman back in the 1930s does. The idea is elegant: take a non-linear system, expand the right-hand sides of the equations into Taylor series, and then introduce new variables for each power and combination of variables. The result is an infinite-dimensional linear system that exactly reproduces the original non-linear dynamics.
It sounds like a fairy tale. And in a sense, it is – with a caveat. To accurately represent a general analytic non-linearity, an infinite-dimensional state space is required. This means an infinite number of hyperedges of all possible orders. The hypergraph ceases to be a finite object – it becomes something far more exotic.
The study's authors show that for such cases, a standard hypergraph is no longer sufficient. A richer combinatorial architecture is needed – a so-called hb-graph (from 'hyperbolic graph' or 'higher-order bond graph,' depending on the interpretation). This is a structure that can encode not just sets of vertices, but also their ordered combinations and hierarchical relationships – everything needed to accurately reproduce the infinite sum of non-linear terms.
In practice, of course, no one builds infinite structures. Instead, the series is truncated at some order M: we consider products of no more than M variables and obtain an approximate linearization. The larger the M, the more accurate the approximation – and the more massive the hypergraph. This is the classic trade-off between accuracy and complexity, familiar to anyone who has worked with numerical methods.
Two Sides of the Same Coin
The main idea behind all this mathematics is duality. The same physical reality can be described in two completely different ways, and both are equally valid.
On one side, a simple network with tricky rules. A graph with non-linear dynamics. Familiar language, a complex equation.
On the other side, a tricky network with simple rules. A hypergraph (or an even richer structure) with linear dynamics. Unfamiliar language, a simple equation.
The transition between these two descriptions is not just a mathematical trick. It is a shift in perspective, after which the same phenomena begin to look different. Much like the shift from describing planetary motion through forces to describing it through the geometry of spacetime: the physics is the same, the language is different, and each language opens up its own possibilities.
When non-linearity 'lives' in the dynamics, we see complex trajectories, bifurcations, chaos. When that same non-linearity 'lives' in the structure, we see a branching network of hyperedges – but one governed by linear equations. And then the entire arsenal of linear analysis is at our disposal: eigenvalues, stability, controllability, observability.
Why It Matters: From Theory to Practice
The abstract beauty of duality is wonderful, but science does not live on beauty alone. What practical significance do these results hold?
First, stability analysis. Understanding whether a non-linear system is stable is a non-trivial task. Understanding whether a linear system is stable is a solved problem. If we know how to translate the former into the latter, we gain a powerful tool for studying when and why complex systems 'break down' into instability.
Second, control. Engineers often need not just to observe a system, but to guide its behavior. The theory of linear control is developed in great detail and depth. Applying it directly to a non-linear system is difficult. Applying it to its linear equivalent is significantly easier.
Third, prediction and modeling. The neural networks of the brain, the spread of infections, the collective behavior of bird flocks or fish schools – all are systems where higher-order interactions play a key role. The ability to translate them into a linear language opens up the possibility of building more accurate and interpretable models.
Finally, there is a profound conceptual conclusion: what we call 'complexity,' is not an inherent property of the dynamics as such. Complexity can be shifted into the structure. Or, to put it another way, complexity is not a sentence, but a choice of descriptive language.
Open Horizons
The research also raises questions that do not yet have answers – and that, perhaps, is the most interesting part of any scientific work.
What do the infinite-order hyperedges that arise in exact linearization mean physically? Can they be assigned a meaningful interpretation – or are they a purely mathematical artifact? How should one choose the truncation order of the Carleman series to minimize error while keeping the structure reasonably sized? How does this approach relate to other methods of 'straightening out' non-linear systems – for example, the Koopman operator, which translates non-linear dynamics into linear ones in the space of observables?
The connection to the Koopman operator is particularly intriguing here. Both approaches expand the state space to make the dynamics linear. But they do so in different ways, and understanding their relationship could offer a new perspective on the nature of non-linearity in general.
The universe is full of systems where many things depend on many other things simultaneously – and simple pairwise connections don't capture the essence. The brain. The ecosystem. The climate. The social fabric. They all live in the space of higher-order interactions. And if we have a way to converse with this complexity in the language of linear mathematics – even at the cost of expanding the space – it changes a great deal.
Non-linearity doesn't disappear. It just changes its address. And when you know where to look for it, you can start the conversation.
