The Geometry of Collapse: Why Mathematical Systems Forget Where They Started

The Shape of Time's Arrow

Imagine watching two rivers merge. At first, the water moves in complex, unpredictable patterns — swirls, eddies, countercurrents. But as you wait, something strange happens: no matter where you start a cork floating on the surface, it eventually gets caught in the same dominant flow, riding the same invisible tracks carved by the river's geometry. The specific initial position matters less and less. What matters is the river itself — its shape, its banks, the forces that guide water toward particular paths.

This is, in essence, what Eduardo Sontag describes in his new paper: a rigorous mathematical explanation for why certain dynamical systems — systems governing everything from chemical reactions to ecosystems to the firing of neurons — seem to forget where they started and collapse onto remarkably simple tracks. Cooperative systems, he argues, "look one-dimensional." Two-cooperative systems "look two-dimensional." This isn't a trick of perception. It's baked into the mathematics.

The finding matters because it connects several deep threads in dynamical systems theory — threads that have been known to produce similar results but whose underlying unity has remained obscured. Sontag's contribution is to reveal the common engine: a geometric contraction that pulls the tangent geometry of trajectories toward a handful of dominant directions, regardless of initial conditions.

The Science

Sontag works in the tradition of monotone dynamical systems, a field that dates back to the work of Hirsch and Hal Smith in the 1980s. The core idea is this: some systems have a property called monotonicity with respect to a cone — roughly, if you start from two different initial conditions, one always stays "above" the other in a geometric sense, and this ordering persists forever.

Cooperative systems are the canonical example. In the language of nonlinear dynamics, a system $\dot{x} = f(x)$ evolving in $\mathbb{R}^n$ is cooperative if the Jacobian matrix $Df(x)$ — which captures how the system responds to small perturbations — is what mathematicians call a Metzler matrix: all its off-diagonal entries are nonnegative. (The diagonal entries can be anything; the off-diagonal entries represent interactions between different components of the system. Nonnegative off-diagonal entries mean that increasing one coordinate never decreases any other coordinate — a natural condition in many biological and chemical systems.)

This condition has remarkable consequences. In the 1980s, Morris Hirsch proved what became known as the generic convergence theorem: in a cooperative system on $\mathbb{R}^n$, almost every trajectory — every one except a set of measure zero — eventually converges to an equilibrium. The system might wander for a while, exploring different states. But in the long run, it settles down. Think of it as a mathematical version of the rivers: no matter where you start, you end up in the same place.

What's striking is that this convergence happens in systems of arbitrary dimension. A cooperative system in $\mathbb{R}^{100}$ — with 100 coordinates changing simultaneously — still exhibits the same generic one-dimensional collapse. You might expect that more dimensions would introduce more possibilities, more degrees of freedom for the system to explore. But the Metzler condition constrains things so strongly that, in the asymptotic limit, the dynamics collapse onto a single line.

Sontag's paper tackles a deceptively simple question: why? Why should a high-dimensional cooperative system behave like a one-dimensional one? And why should the generalization — two-cooperative systems — behave like a two-dimensional one?

The answer, Sontag argues, lies in a classical theorem from linear operator theory, combined with the geometry of projective metrics. Specifically, the tool is Birkhoff's theorem (from Garrett Birkhoff's work in the 1950s), which states that any linear map sending a cone into its interior contracts the Hilbert projective metric on that cone.

To understand what this means, picture the positive orthant — the cone of all vectors with nonnegative coordinates. The Hilbert projective metric is a way of measuring the "distance" between two rays (directions) inside this cone. If you have two different directions, you can ask how far apart they are in this metric. Birkhoff's theorem says that applying a positive linear map (one that keeps everything inside the cone) brings any two rays closer together — it contracts the distance. The contraction isn't just a little bit; it's exponential, with a rate that depends on how strongly positive the map is.

This is the engine that drives the collapse.

The connection to dynamical systems comes through the linearized flow. Consider a trajectory $x(t) = \varphi_t(x_0)$ of the system, starting from some initial point $x_0$. Small perturbations $v_0$ along the trajectory evolve according to the variational equation $\dot{v} = Df(x(t)) v$. The solution to this linear time-varying system is given by the derivative of the flow map, $D\varphi_t(x_0)$. This matrix tells you how an infinitesimal perturbation at $x_0$ gets stretched and rotated as it travels along the trajectory.

Now, if the system is cooperative, the Jacobian $Df(x(t))$ is Metzler for every $t$. This means the variational equation stays in the cooperative family at every moment. Over any finite time window, the resulting linear map $D\varphi_\tau(x_0)$ has a crucial property: it sends the positive orthant into its interior. It doesn't just preserve the cone — it pushes things strictly inside, away from the boundary. (Sontag makes this precise with a uniform positivity condition: the entries of the compound Jacobian matrix must lie in a fixed band $["delta", \Delta]$, and the support graph must be strongly connected.)

This uniform positivity, as Sontag puts it, is what yields the exponential contraction. The finite-window map $\Psi(t+h_0, t)$ becomes entrywise strictly positive, with entries bounded above and below by constants that depend on the band parameters and the graph diameter. For an entrywise positive matrix $P$, Birkhoff's bound on the contraction coefficient is:

$$\kappa (P)\le \tanh \left(\frac{1}{4}\mathrm{diam}(P)\right)$$

where the projective diameter $\operatorname{diam}(P)$ measures the spread of the matrix entries. Composed over many windows, this gives exponential decay:

$$d_H({\wedge }^{k}D{\phi }_{\tau }\xi ,{\wedge }^{k}D{\phi }_{\tau }\eta )\le {\kappa }_0^{\lfloor \tau /h_0\rfloor }{d}_H(\xi ,\eta )\le {\kappa }_0^{-1}{e}^{-c\tau }{d}_H(\xi ,\eta )$$

The rate $c = -h_0^{-1} \log \kappa_0 > 0$ is uniform in time. This exponential contraction is the mathematical heartbeat of the paper.

Sontag emphasizes that this argument is heuristic in its presentation but rigorous in its conclusions. The alignment statements — that generic tangent vectors align with a dominant direction, that generic 2-planes align with a dominant 2-plane — are statements about directions in the tangent bundle. Their clean rigorous form (given as Theorem 1 in the paper) is a finite-time statement: once $T \geq h_0$, the images of interior cone directions under the compound cocycle have collapsed to within $e^{-cT}$ of one another. No compactness assumptions, no limiting bundles required.

What They Found

The core result runs like this. In a cooperative system, pick any initial point $x_0$ and any initial perturbation $v_0$ that isn't in the "slow bundle" (the measure-zero set of directions that don't align). As you follow the dynamics forward in time, the direction of $v(t) = D\varphi_t(x_0) v_0$ gets pulled inexorably toward the direction of the velocity field $\dot{x}(t) = f(x(t))$ itself. Formally:

$$\frac{D{\phi }_{\tau }(x_0)v_0}{|D{\phi }_{\tau }(x_0)v_0|}\approx \pm \frac{\dot{x}(\tau )}{|\dot{x}(\tau )|}$$

In words: if you're patient enough, every perturbation points in the same direction as the flow. The alignment is exponential, decaying at rate $c > 0$.

There's a beautiful geometric consequence. Take a nearby initial state $y_0 = x_0 + v_0$ — close enough that the linear approximation is valid. Then:

$$y(\tau )-x(\tau )\approx {\alpha }_{\tau }(v_0)\dot{x}(\tau )$$

The perturbed state, at time $\tau$, has landed on the reference trajectory — only at a slightly shifted time $\alpha_\tau(v_0)$. This is a one-dimensional picture: the high-dimensional system has collapsed onto a one-dimensional curve, and nearby trajectories are just time-shifted versions of each other.

The story for two-cooperative systems is analogous but one dimension up. A system is two-cooperative if the second additive compound of the Jacobian, $Df(x)^{[2]}$, is Metzler for every $x$. This is a stronger condition — it requires that not just individual vectors but pairs of vectors (2-planes) maintain certain positivity properties. The technical language involves exterior powers: the wedge product $\wedge^2 \mathbb{R}^n$ is a new vector space of dimension $\binom{n}{2}$, and the linear map $D\varphi_\tau(x_0)$ induces a map $\wedge^2 D\varphi_\tau(x_0)$ on this space.

Uniform 2-positivity means this induced map sends the positive orthant of $\wedge^2 \mathbb{R}^n$ into its interior. The same Birkhoff contraction applies, but now it contracts 2-planes, not just lines. Any two 2-planes in the positive Plücker region get pulled exponentially close to each other. The dominant object is always a genuine 2-plane — not just an abstract element of the exterior algebra — because the compound operator carries decomposable bivectors to decomposable bivectors.

Sontag constructs a concrete anchor for this dominant 2-plane using a frozen-eigenvector argument. Since $Df(x_0)^{[2]}$ is Metzler and irreducible, Perron-Frobenius gives a simple dominant eigen-bivector $\xi_0 = a_0 \wedge b_0$, which determines the 2-plane $\Pi_{anc}(\tau) = \wedge^2 D\varphi_\tau(x_0) \xi_0$. This transported plane serves as a stand-in for the true dominant bundle, becoming exponentially aligned with it.

The geometric picture is this: the tangent bundle collapses not to a line but to a 2-plane. The dynamics look two-dimensional. This is why two-cooperative systems satisfy Poincaré-Bendixson-type theorems — results that, in classical dynamical systems theory, hold only in two dimensions. A compact $\omega$-limit set containing no equilibrium must be a periodic orbit. The system can't produce chaos or strange attractors; the asymptotic geometry is too constrained.

Sontag's precise Theorem 1 states the mechanism cleanly: for a $k$-positive cocycle satisfying the uniform band and connectivity hypotheses, the compound flow contracts the Hilbert metric on the appropriate cone at rate $c > 0$. The finite-time version doesn't require compactness or limiting bundles — it follows directly from the Birkhoff argument applied to the finite-window map.

Why This Changes Things

The significance of Sontag's paper isn't primarily in the theorems — the core results (Hirsch's generic convergence, the Poincaré-Bendixson property for two-cooperative systems) were already known. The significance is in the explanation.

For decades, mathematicians have known that cooperative systems converge to equilibria in ways that look one-dimensional. They've known that two-cooperative systems admit Poincaré-Bendixson behavior that looks two-dimensional. But the unifying geometric picture — the sense in which these are two instances of the same phenomenon — has remained opaque.

Sontag's contribution is to show that the engine is the same in both cases: Birkhoff's contraction of the Hilbert projective metric. The cone contraction pulls the tangent geometry toward a dominant structure (a line for $k=1$, a plane for $k=2$) simply because positive linear maps contract projective distances. This isn't a new result; it's a new lens, illuminating why the known results hold.

The power of this lens is that it makes the dimensionality intuition precise. When we say a cooperative system "looks one-dimensional," we now have a geometric mechanism: the positive flow contracts the Hilbert metric on the tangent cone, so any two directions in the cone get pulled toward each other exponentially fast. The dominant direction turns out to be represented by the velocity field itself — the natural in-cone marker along the trajectory. Perturbed initial conditions, at any finite time, land near the reference trajectory at a slightly shifted time.

For two-cooperative systems, the same story holds in the wedge-squared space. The positive flow contracts the metric on $\wedge^2 \mathbb{R}^n$, pulling 2-planes toward each other. The dominant 2-plane is always a genuine plane (never an abstract bivector), because the wedge compound carries decomposables to decomposables. This is why the asymptotic geometry can look two-dimensional: the collapse is to a plane, not a line.

The paper has practical implications for anyone studying monotone systems. The explicit contraction rate $c = -h_0^{-1} \log \kappa_0$ is computable from the band parameters $\delta, \Delta$, the global bound $\Lambda$, and the graph diameter $r$. You can estimate how fast alignment happens from the system itself. The rate isn't just an existence theorem — it's a bound you can work with.

More broadly, the paper connects to the theory of exponential dichotomies and dominated splittings in dynamical systems. The dominant line bundle $E_1(\tau)$ and complementary slow bundle $F_1(\tau)$ in the cooperative case, and the dominant plane bundle $E_2(\tau)$ with slow bundle $F_2(\tau)$ in the two-cooperative case, are instances of a more general phenomenon studied in smooth dynamical systems. Sontag's contribution is to show that, for positive systems, these splittings arise naturally from cone contraction.

What's Next

Sontag is careful to scope what the paper does and doesn't prove. The alignment argument explains why the tangent geometry collapses — it accounts for the collapse to two dominant directions in the two-cooperative case. But it isn't itself a proof that every bounded trajectory converges to an equilibrium or periodic orbit. That convergence requires additional steps: compactness assumptions, existence of an invariant measure, passage from tangent-bundle statements to nonlinear statements. The paper provides the skeleton; fleshing it out connects to the full theory of monotone systems.

Several open questions remain. Can the finite-time contraction estimates be sharpened? The current bound involves the graph diameter $r \leq N-1$ and the window length $h_0$; can one prove tighter estimates for specific classes of systems? The paper focuses on $k=1$ and $k=2$ as the cases of primary interest, but the framework works for arbitrary $k$. What does the geometry look like for $k=3$ or higher? The dominant object would be a $k$-plane, but the connection to classical theorems (Poincaré-Bendixson generalizations) is less developed.

There's also the question of what happens when the uniform positivity condition is relaxed. The paper assumes a fixed connected support for the graph and uniform bounds on the Jacobian entries. Real-world systems often exhibit varying interaction strengths, switching topologies, or other perturbations. How robust is the contraction picture to such modifications? Early work on switched cooperative systems suggests that the alignment can persist, but the rates may degrade.

The Perron-Frobenius anchors — the frozen dominant eigenvector $p_0$ and frozen dominant bivector $\xi_0$ — are concrete stand-ins for the true dynamical bundles. They serve as computable markers. But they drift with time, and the true invariant bundles $E_1(\tau)$ and $E_2(\tau)$ need not converge to fixed limits. Understanding the asymptotic behavior of these anchors is a natural next step.

Finally, there's the question of applications. Cooperative and two-cooperative systems arise naturally in biology (population dynamics, neural networks, gene regulation), chemistry (reaction networks with monotone kinetics), and economics (models with complementary goods). Sontag's geometric intuition could inform the design of control strategies for such systems, the analysis of transient behavior, or the classification of asymptotic attractors.

The river, it turns out, always finds its shape. Sontag's paper is a map of the terrain — a geometric account of why, in certain systems, time erases the details and leaves only the essential structure. Whether you're watching water, neurons, or chemical concentrations, the mathematics of positivity pulls everything toward the same invisible tracks. The convergence isn't an accident. It's geometry.

---

References within the paper: The core theory of monotone dynamical systems derives from Hirsch (1985) and Smith (1987). The $k$-cooperative framework and Poincaré-Bendixson theorems for $k=2$ build on work of Sánchez and others (1990s–2000s). Birkhoff's contraction theorem dates to Birkhoff (1957) with refinements by others. The specific formulation of the contraction rate and finite-window estimates appear in the appendix of Sontag's paper. Exponential separation (dominated splitting) language follows Poláčik and Tereščák (1991).